staging → main: auto-promote f1b72af

staging → main: auto-promote e39d818

.github/workflows/auto-promote-staging.yml

@@ -111,7 +111,60 @@ jobs:
       all_green: ${{ steps.gates.outputs.all_green }}
       head_sha: ${{ steps.gates.outputs.head_sha }}
     steps:
+      # Skip empty-tree promotes (the perpetual auto-promote↔auto-sync cycle
+      # observed 2026-05-03). Sequence: auto-promote merges via the staging
+      # merge-queue's MERGE strategy, creating a merge commit on main that
+      # staging doesn't have. auto-sync then merges main back into staging
+      # via another merge commit (the queue's MERGE strategy applies on
+      # the staging side too, even when the workflow's local FF would
+      # have sufficed). Now staging has a new merge-commit SHA whose
+      # tree == main's tree — but auto-promote sees "staging ahead of
+      # main by 1" and opens YET another empty promote PR. Each round
+      # costs ~30-40 min wallclock, ~2 manual approvals, and burns a
+      # full CodeQL Go run (~15 min). Without this guard the cycle
+      # repeats indefinitely.
+      #
+      # Long-term fix is to switch the merge_queue ruleset's
+      # `merge_method` away from MERGE so FF-able PRs land cleanly,
+      # but that's a broader change affecting every staging PR's
+      # commit shape. This guard is the one-line surgical fix that
+      # breaks the cycle without touching merge-queue config.
+      #
+      # Fail-open: if `git diff` errors for any reason, fall through
+      # to the gate check (preserve existing behavior). Only skip
+      # when the diff is DEFINITIVELY empty.
+      - name: Checkout for tree-diff check
+        uses: actions/checkout@de0fac2e4500dabe0009e67214ff5f5447ce83dd # v6.0.2
+        with:
+          fetch-depth: 0
+          ref: staging
+      - name: Skip if staging tree == main tree (perpetual-cycle break)
+        id: tree-diff
+        env:
+          HEAD_SHA: ${{ github.event.workflow_run.head_sha || github.sha }}
+        run: |
+          set -eu
+          git fetch origin main --depth=50 || { echo "::warning::git fetch main failed — proceeding (fail-open)"; exit 0; }
+          # Compare staging tip's tree against main's tree. `git diff
+          # --quiet` exits 0 if no differences, 1 if there are.
+          if git diff --quiet origin/main "$HEAD_SHA" -- 2>/dev/null; then
+            {
+              echo "## ⏭ Skipped — no code to promote"
+              echo
+              echo "staging tip (\`${HEAD_SHA:0:8}\`) and \`main\` have identical trees."
+              echo "This is the auto-promote↔auto-sync merge-commit cycle: staging has a"
+              echo "new SHA (a sync-back merge commit) but the underlying file tree is"
+              echo "already on main, so there's no real code to ship."
+              echo
+              echo "Skipping to avoid opening an empty promote PR. Cycle terminates here."
+            } >> "$GITHUB_STEP_SUMMARY"
+            echo "::notice::auto-promote: staging tree == main tree — no code to promote, skipping"
+            echo "skip=true" >> "$GITHUB_OUTPUT"
+          else
+            echo "skip=false" >> "$GITHUB_OUTPUT"
+          fi
       - name: Check all required gates on this SHA
+        if: steps.tree-diff.outputs.skip != 'true'
         id: gates
         env:
           GH_TOKEN: ${{ secrets.GITHUB_TOKEN }}
@@ -209,10 +262,25 @@ jobs:
             exit 0
           fi
 
+      # Mint the App token BEFORE the promote-PR step so the auto-merge
+      # call can use it. GITHUB_TOKEN-initiated merges suppress the
+      # downstream `push` event on main, breaking the
+      # publish-workspace-server-image → canary-verify → redeploy-tenants
+      # chain (issue #2357). Using the App token here means the
+      # merge-queue-landed merge IS able to fire the cascade naturally;
+      # the polling tail below stays as defense-in-depth.
+      - name: Mint App token for promote-PR + downstream dispatch
+        if: ${{ vars.AUTO_PROMOTE_ENABLED == 'true' || github.event.inputs.force == 'true' }}
+        id: app-token
+        uses: actions/create-github-app-token@1b10c78c7865c340bc4f6099eb2f838309f1e8c3 # v3.1.1
+        with:
+          app-id: ${{ secrets.MOLECULE_AI_APP_ID }}
+          private-key: ${{ secrets.MOLECULE_AI_APP_PRIVATE_KEY }}
+
       - name: Open (or reuse) staging → main promote PR + enable auto-merge
         if: ${{ vars.AUTO_PROMOTE_ENABLED == 'true' || github.event.inputs.force == 'true' }}
         env:
-          GH_TOKEN: ${{ secrets.GITHUB_TOKEN }}
+          GH_TOKEN: ${{ steps.app-token.outputs.token }}
           REPO: ${{ github.repository }}
           TARGET_SHA: ${{ needs.check-all-gates-green.outputs.head_sha }}
         run: |
@@ -267,52 +335,34 @@ jobs:
           echo "promote_pr_num=${PR_NUM}" >> "$GITHUB_OUTPUT"
         id: promote_pr
 
-      # Mint a short-lived GitHub App installation token for the dispatch
-      # step below. We CANNOT use `secrets.GITHUB_TOKEN` to dispatch the
-      # downstream publish chain — workflow runs created by GITHUB_TOKEN
-      # do not fire `workflow_run` triggers on completion (the
-      # documented "no recursion" rule —
-      # https://docs.github.com/en/actions/using-workflows/triggering-a-workflow#triggering-a-workflow-from-a-workflow).
-      #
-      # Symptom this caused (root-caused on 2026-04-30): publish-image
-      # ran successfully twice (21313dc 14:41Z, 59dec57 15:21Z) but
-      # canary-verify and redeploy-tenants-on-main never chained,
-      # because the publish run's `triggering_actor` was
-      # `github-actions[bot]` (i.e. GITHUB_TOKEN). A manual dispatch
-      # earlier in the day with the operator's PAT (d850ec7 06:52Z) did
-      # chain — same workflow file, only the actor differed.
-      #
-      # An App token's triggering_actor is the App user (e.g.
-      # `molecule-ai[bot]`), which IS allowed to fire downstream
-      # workflow_run cascades.
-      - name: Mint App token for downstream dispatch
-        if: steps.promote_pr.outputs.promote_pr_num != ''
-        id: app-token
-        uses: actions/create-github-app-token@1b10c78c7865c340bc4f6099eb2f838309f1e8c3 # v3.1.1
-        with:
-          app-id: ${{ secrets.MOLECULE_AI_APP_ID }}
-          private-key: ${{ secrets.MOLECULE_AI_APP_PRIVATE_KEY }}
-
+      # The App token minted above (before the promote-PR step) is
+      # also used by the polling tail below. Defense-in-depth: with
+      # the merge-queue-landed merge now using the App token, the
+      # main-branch push event SHOULD fire the publish/canary/redeploy
+      # cascade naturally — but if for any reason it doesn't (e.g. an
+      # unrelated event-suppression edge case), the explicit dispatches
+      # below still wake the chain.
       - name: Wait for promote merge, then dispatch publish + redeploy (#2357)
-        # GITHUB_TOKEN-initiated merges suppress downstream `push` events
-        # (https://docs.github.com/en/actions/using-workflows/triggering-a-workflow#triggering-a-workflow-from-a-workflow).
-        # Result: when the merge queue lands the promote PR, the resulting
-        # main-branch push DOES NOT fire publish-workspace-server-image,
-        # so canary-verify and redeploy-tenants-on-main never run and
-        # tenants stay on stale code (issue #2357).
+        # Defense-in-depth dispatch. With the auto-merge call above
+        # now using the App token (this commit), the merge-queue-landed
+        # merge SHOULD fire publish-workspace-server-image naturally
+        # via on:push:[main] — App-token-initiated pushes DO trigger
+        # workflow_run cascades, unlike GITHUB_TOKEN-initiated ones
+        # (the documented "no recursion" rule —
+        # https://docs.github.com/en/actions/using-workflows/triggering-a-workflow#triggering-a-workflow-from-a-workflow).
         #
-        # Workaround: poll for the merge to land, then explicitly
-        # `gh workflow run` publish-workspace-server-image. The dispatch
-        # MUST authenticate as the molecule-ai App (App token minted
-        # above) — not GITHUB_TOKEN — so that the resulting publish
-        # run's completion event can fire the workflow_run cascade
-        # into canary-verify + redeploy-tenants-on-main. See the prior
-        # step's comment for the GITHUB_TOKEN no-recursion details.
+        # This explicit dispatch stays as belt-and-suspenders for any
+        # edge case where the natural cascade misfires. If it never
+        # observably fires after this token swap (i.e. the publish
+        # workflow has already started by the time we get here), the
+        # second dispatch is a harmless no-op (publish-workspace-server-image
+        # has its own concurrency group that dedupes).
         #
-        # Long-term fix: switch the auto-merge call above to use the
-        # same App token, so the merge's push event fires
-        # publish-workspace-server-image naturally and this polling tail
-        # becomes unnecessary. Tracked in #2357.
+        # See PR for #2357: pre-fix the merge action was via
+        # GITHUB_TOKEN, suppressing the cascade and forcing this tail
+        # to be the SOLE chain trigger. With the auto-merge token swap
+        # the tail becomes redundant in the happy path; keep until
+        # we've observed >=10 successful natural cascades, then drop.
         if: steps.promote_pr.outputs.promote_pr_num != ''
         env:
           GH_TOKEN: ${{ steps.app-token.outputs.token }}

.github/workflows/canary-staging.yml

@@ -50,19 +50,35 @@ jobs:
     env:
       MOLECULE_CP_URL: https://staging-api.moleculesai.app
       MOLECULE_ADMIN_TOKEN: ${{ secrets.MOLECULE_STAGING_ADMIN_TOKEN }}
-      # Without an LLM key the test_staging_full_saas.sh script provisions
-      # the workspace with empty secrets, hermes derive-provider.sh resolves
-      # `openai/gpt-4o` to PROVIDER=openrouter, no OPENROUTER_API_KEY is
-      # found in env, and A2A returns "No LLM provider configured" at
-      # request time (canary step 8/11). The full-lifecycle workflow
-      # (e2e-staging-saas.yml) has carried this secret since launch — the
-      # canary regressed when it was first split out and lost the env
-      # block. Issue #1500 had ~30 consecutive failures before this was
-      # spotted; do NOT remove without re-reading the script's secrets-
-      # injection block.
+      # MiniMax is the canary's PRIMARY LLM auth path post-2026-05-04.
+      # Switched from hermes+OpenAI after #2578 (the staging OpenAI key
+      # account went over quota and stayed dead for 36+ hours, taking
+      # the canary red the entire time). claude-code template's
+      # `minimax` provider routes ANTHROPIC_BASE_URL to
+      # api.minimax.io/anthropic and reads MINIMAX_API_KEY at boot —
+      # ~5-10x cheaper per token than gpt-4.1-mini AND on a separate
+      # billing account, so OpenAI quota collapse no longer wedges the
+      # canary. Mirrors the migration continuous-synth-e2e.yml made on
+      # 2026-05-03 (#265) for the same reason. tests/e2e/test_staging_
+      # full_saas.sh branches SECRETS_JSON on which key is present —
+      # MiniMax wins when set.
+      E2E_MINIMAX_API_KEY: ${{ secrets.MOLECULE_STAGING_MINIMAX_API_KEY }}
+      # Direct-Anthropic alternative for operators who don't want to
+      # set up a MiniMax account (priority below MiniMax — first
+      # non-empty wins in test_staging_full_saas.sh's secrets-injection
+      # block). See #2578 PR comment for the rationale.
+      E2E_ANTHROPIC_API_KEY: ${{ secrets.MOLECULE_STAGING_ANTHROPIC_API_KEY }}
+      # OpenAI fallback — kept wired so an operator-dispatched run with
+      # E2E_RUNTIME=hermes overridden via workflow_dispatch can still
+      # exercise the OpenAI path without re-editing the workflow.
       E2E_OPENAI_API_KEY: ${{ secrets.MOLECULE_STAGING_OPENAI_KEY }}
       E2E_MODE: canary
-      E2E_RUNTIME: hermes
+      E2E_RUNTIME: claude-code
+      # Pin the canary to a specific MiniMax model rather than relying
+      # on the per-runtime default (which could resolve to "sonnet" →
+      # direct Anthropic and defeat the cost saving). M2.7-highspeed
+      # is "Token Plan only" but cheap-per-token and fast.
+      E2E_MODEL_SLUG: MiniMax-M2.7-highspeed
       E2E_RUN_ID: "canary-${{ github.run_id }}"
 
     steps:
@@ -75,13 +91,47 @@ jobs:
             exit 2
           fi
 
-      - name: Verify OpenAI key present
+      - name: Verify LLM key present
         run: |
-          if [ -z "$E2E_OPENAI_API_KEY" ]; then
-            echo "::error::MOLECULE_STAGING_OPENAI_KEY secret not set — A2A will fail at request time with 'No LLM provider configured'"
+          # Per-runtime key check — claude-code uses MiniMax; hermes /
+          # langgraph (operator-dispatched only) use OpenAI. Hard-fail
+          # rather than soft-skip per the lesson from synth E2E #2578:
+          # an empty key silently falls through to the wrong
+          # SECRETS_JSON branch and the canary fails 5 min later with
+          # a confusing auth error instead of the clean "secret
+          # missing" message at the top.
+          case "${E2E_RUNTIME}" in
+            claude-code)
+              # Either MiniMax OR direct-Anthropic works — first
+              # non-empty wins in the test script's secrets-injection
+              # priority chain. Operators only need to set ONE of these
+              # secrets; we don't force a choice between them.
+              if [ -n "${E2E_MINIMAX_API_KEY:-}" ]; then
+                required_secret_name="MOLECULE_STAGING_MINIMAX_API_KEY"
+                required_secret_value="${E2E_MINIMAX_API_KEY}"
+              elif [ -n "${E2E_ANTHROPIC_API_KEY:-}" ]; then
+                required_secret_name="MOLECULE_STAGING_ANTHROPIC_API_KEY"
+                required_secret_value="${E2E_ANTHROPIC_API_KEY}"
+              else
+                required_secret_name="MOLECULE_STAGING_MINIMAX_API_KEY or MOLECULE_STAGING_ANTHROPIC_API_KEY"
+                required_secret_value=""
+              fi
+              ;;
+            langgraph|hermes)
+              required_secret_name="MOLECULE_STAGING_OPENAI_KEY"
+              required_secret_value="${E2E_OPENAI_API_KEY:-}"
+              ;;
+            *)
+              echo "::warning::Unknown E2E_RUNTIME='${E2E_RUNTIME}' — skipping LLM-key check"
+              required_secret_name=""
+              required_secret_value="present"
+              ;;
+          esac
+          if [ -n "$required_secret_name" ] && [ -z "$required_secret_value" ]; then
+            echo "::error::${required_secret_name} secret not set for runtime=${E2E_RUNTIME} — A2A will fail at request time with 'No LLM provider configured'"
             exit 2
           fi
-          echo "OpenAI key present ✓ (len=${#E2E_OPENAI_API_KEY})"
+          echo "LLM key present ✓ (runtime=${E2E_RUNTIME}, key=${required_secret_name}, len=${#required_secret_value})"
 
       - name: Canary run
         id: canary
@@ -231,10 +281,34 @@ jobs:
                         and o.get('status') not in ('purged',)]
           print('\n'.join(candidates))
           " 2>/dev/null)
+          # Per-slug DELETE with HTTP-code verification. The previous
+          # `... >/dev/null || true` swallowed every failure, so a 5xx
+          # or timeout from CP looked identical to "successfully cleaned
+          # up" and the tenant kept eating ~2 vCPU until the hourly
+          # stale sweep caught it (up to 2h later). Now we capture the
+          # response code and surface non-2xx as a workflow warning, so
+          # the run page shows which slug leaked. We still don't `exit 1`
+          # on cleanup failure — a single-canary cleanup miss shouldn't
+          # fail-flag the canary itself when the actual smoke check
+          # passed. The sweep-stale-e2e-orgs cron (now every 15 min,
+          # 30-min threshold) is the safety net for whatever slips past.
+          # See molecule-controlplane#420.
+          leaks=()
           for slug in $orgs; do
-            curl -sS -X DELETE "$MOLECULE_CP_URL/cp/admin/tenants/$slug" \
+            code=$(curl -sS -o /tmp/canary-cleanup.out -w "%{http_code}" \
+              -X DELETE "$MOLECULE_CP_URL/cp/admin/tenants/$slug" \
               -H "Authorization: Bearer $ADMIN_TOKEN" \
               -H "Content-Type: application/json" \
-              -d "{\"confirm\":\"$slug\"}" >/dev/null || true
+              -d "{\"confirm\":\"$slug\"}" \
+              || echo "000")
+            if [ "$code" = "200" ] || [ "$code" = "204" ]; then
+              echo "[teardown] deleted $slug (HTTP $code)"
+            else
+              echo "::warning::canary teardown for $slug returned HTTP $code — sweep-stale-e2e-orgs will catch it within ~45 min. Body: $(head -c 300 /tmp/canary-cleanup.out 2>/dev/null)"
+              leaks+=("$slug")
+            fi
           done
+          if [ ${#leaks[@]} -gt 0 ]; then
+            echo "::warning::canary teardown left ${#leaks[@]} leak(s): ${leaks[*]}"
+          fi
           exit 0

.github/workflows/ci.yml

@@ -272,6 +272,18 @@ jobs:
           find tests/e2e infra/scripts -type f -name '*.sh' -print0 \
             | xargs -0 shellcheck --severity=warning
 
+      - if: needs.changes.outputs.scripts == 'true'
+        name: Run E2E bash unit tests (no live infra)
+        # Pure-bash unit tests for E2E helper libs (lib/*.sh). These pin
+        # behavior of dispatch logic that — when broken — silently masks as
+        # "Could not resolve authentication method" only after a successful
+        # tenant + workspace provision (PR #2571 incident, 2026-05-03). Add
+        # new self-contained unit tests here as the lib/ directory grows;
+        # tests requiring live CP/tenant credentials belong in the dedicated
+        # e2e-staging-* workflows, not this job.
+        run: |
+          bash tests/e2e/test_model_slug.sh
+
   canvas-deploy-reminder:
     name: Canvas Deploy Reminder
     runs-on: ubuntu-latest
@@ -346,6 +358,72 @@ jobs:
       - if: needs.changes.outputs.python == 'true'
         run: python -m pytest --tb=short
 
+      - if: needs.changes.outputs.python == 'true'
+        name: Per-file critical-path coverage (MCP / inbox / auth)
+        # MCP-critical Python files have a per-file floor on top of the
+        # 86% total floor in pytest.ini. Rationale (issue #2790, after
+        # the PR #2766 → PR #2771 cycle): the total floor averages ~6000
+        # lines, so a single MCP file could regress to ~50% with no
+        # complaint as long as other modules compensate. These five
+        # files handle multi-tenant routing + auth + inbox dispatch —
+        # a coverage drop here is the same risk shape as a Go-side
+        # workspace-server token/secrets file dropping below 10%.
+        #
+        # Floor 75% sits below current actuals (80-96%) so this gate is
+        # strictly additive — no existing PR fails. Ratchet plan in
+        # COVERAGE_FLOOR.md.
+        run: |
+          set -e
+          PER_FILE_FLOOR=75
+          CRITICAL_FILES=(
+            "a2a_mcp_server.py"
+            "mcp_cli.py"
+            "a2a_tools.py"
+            "inbox.py"
+            "platform_auth.py"
+          )
+
+          # pytest already wrote .coverage; emit a JSON view scoped to
+          # the critical files so jq/python can read the per-file pct
+          # without parsing tabular text. --include uses fnmatch, and
+          # the leading "*" allows the file to live anywhere under the
+          # workspace root (today they sit at workspace/<name>.py).
+          INCLUDES=$(printf '*%s,' "${CRITICAL_FILES[@]}")
+          INCLUDES="${INCLUDES%,}"
+          python -m coverage json -o /tmp/critical-cov.json --include="$INCLUDES"
+
+          FAILED=0
+          for f in "${CRITICAL_FILES[@]}"; do
+            # Match by top-level path key (e.g. "a2a_tools.py", not
+            # "builtin_tools/a2a_tools.py" — different file at 100%).
+            # The keys in coverage.json are paths relative to the run
+            # cwd (workspace/), so the critical-path entry sits at the
+            # bare basename.
+            pct=$(jq -r --arg f "$f" '.files | to_entries | map(select(.key == $f)) | .[0].value.summary.percent_covered // "MISSING"' /tmp/critical-cov.json)
+            if [ "$pct" = "MISSING" ]; then
+              echo "::error file=workspace/$f::No coverage data — file may have moved or test exclusion mis-set."
+              FAILED=$((FAILED+1))
+              continue
+            fi
+            echo "$f: ${pct}%"
+            if awk "BEGIN{exit !($pct < $PER_FILE_FLOOR)}"; then
+              echo "::error file=workspace/$f::${pct}% < ${PER_FILE_FLOOR}% per-file floor (MCP critical path). See COVERAGE_FLOOR.md."
+              FAILED=$((FAILED+1))
+            fi
+          done
+
+          if [ "$FAILED" -gt 0 ]; then
+            echo ""
+            echo "$FAILED MCP critical-path file(s) below the ${PER_FILE_FLOOR}% per-file floor."
+            echo "These paths handle multi-tenant routing, auth tokens, and inbox dispatch."
+            echo "A coverage drop here is the same risk shape as Go-side tokens/secrets files"
+            echo "dropping below 10% (see COVERAGE_FLOOR.md). Either:"
+            echo "  (a) add tests to raise coverage back above ${PER_FILE_FLOOR}%, or"
+            echo "  (b) if this is unavoidable historical debt, file an issue and propose"
+            echo "      adjusting the floor with rationale in COVERAGE_FLOOR.md."
+            exit 1
+          fi
+
       # SDK + plugin validation moved to standalone repo:
       # github.com/Molecule-AI/molecule-sdk-python
 

.github/workflows/continuous-synth-e2e.yml

@@ -32,16 +32,41 @@ name: Continuous synthetic E2E (staging)
 
 on:
   schedule:
-    # Every 20 minutes, on the :00 :20 :40. Offsets the existing :15
-    # sweep-cf-orphans and :45 sweep-cf-tunnels so the three
-    # operations don't all hit Cloudflare/AWS at the same minute.
-    - cron: '0,20,40 * * * *'
+    # Every 10 minutes, on :02 :12 :22 :32 :42 :52. Three constraints:
+    #   1. Stay off the top-of-hour. GitHub Actions scheduler drops
+    #      :00 firings under high load (own docs:
+    #      https://docs.github.com/en/actions/using-workflows/events-that-trigger-workflows#schedule).
+    #      Prior history: cron was '0,20,40' (2026-05-02) — only :00
+    #      ever survived. Bumped to '10,30,50' (2026-05-03) on the
+    #      theory that further-from-:00 wins. Empirically 2026-05-04
+    #      that ALSO dropped to ~60 min effective cadence (only ~1
+    #      schedule fire per hour — see molecule-core#2726). Detection
+    #      latency was claimed 20 min, actual 60 min.
+    #   2. Avoid colliding with the existing :15 sweep-cf-orphans
+    #      and :45 sweep-cf-tunnels — both hit the CF API and we
+    #      don't want to fight for rate-limit tokens.
+    #   3. Avoid the :30 heavy slot (canary-staging /30, sweep-aws-
+    #      secrets, sweep-stale-e2e-orgs every :15) — multiple
+    #      overlapping cron registrations on the same minute is part
+    #      of what GH drops under load.
+    # Solution: bump fires-per-hour 3 → 6 AND keep all slots in clean
+    # lanes (1-3 min away from any other cron). Even with empirically-
+    # observed ~67% GH drop ratio, 6 attempts/hour yields ~2 effective
+    # fires = ~30 min cadence; closer to the 20-min target than the
+    # current shape and provides a real degradation alarm if drops
+    # get worse.
+    - cron: '2,12,22,32,42,52 * * * *'
   workflow_dispatch:
     inputs:
       runtime:
-        description: "Runtime to provision (langgraph = fastest, default; hermes = slower but covers SDK-native path; claude-code = needs OAUTH token in tenant env)"
+        description: "Runtime to provision (claude-code = default + cheapest via MiniMax; langgraph = OpenAI-only; hermes = SDK-native path, slower)"
         required: false
-        default: "langgraph"
+        default: "claude-code"
+        type: string
+      model_slug:
+        description: "Model id to provision the workspace with (default MiniMax-M2.7-highspeed; e.g. 'sonnet' to test direct Anthropic, 'openai/gpt-4o' for hermes)"
+        required: false
+        default: "MiniMax-M2.7-highspeed"
         type: string
       keep_org:
         description: "Skip teardown for post-mortem debugging (only manual dispatch — never set this for cron runs)"
@@ -68,15 +93,36 @@ jobs:
   synth:
     name: Synthetic E2E against staging
     runs-on: ubuntu-latest
-    timeout-minutes: 12
+    # Bumped from 12 → 20 (2026-05-04). Tenant user-data install phase
+    # (apt-get update + install docker.io/jq/awscli/caddy + snap install
+    # ssm-agent) runs from raw Ubuntu on every boot — none of it is
+    # pre-baked into the tenant AMI. Empirical fetch_secrets/ok timing
+    # across today's canaries: 51s → 82s → 143s → 625s. apt-mirror tail
+    # latency drives the boot-to-fetch_secrets phase from ~1min to >10min.
+    # A 12min budget leaves only ~2min for the workspace (which needs
+    # ~3.5min for claude-code cold boot) on slow-apt days, blowing the
+    # budget. 20min absorbs the worst tenant tail so the workspace probe
+    # gets the full ~7min it needs even on a slow apt day. Real fix:
+    # pre-bake caddy + ssm-agent into the tenant AMI (controlplane#TBD).
+    timeout-minutes: 20
     env:
-      # langgraph default keeps cold-start under 5 min on staging EC2.
-      # hermes is slower (~7-10 min) and isn't needed for the
-      # regression class this gate exists to catch (deployment-pipeline
-      # + schema-drift + integration). Operators can pick hermes via
-      # workflow_dispatch when they need to exercise the SDK-native
-      # session path.
-      E2E_RUNTIME: ${{ github.event.inputs.runtime || 'langgraph' }}
+      # claude-code default: cold-start ~5 min (comparable to langgraph),
+      # but uses MiniMax-M2.7-highspeed via the template's third-party-
+      # Anthropic-compat path (workspace-configs-templates/claude-code-
+      # default/config.yaml:64-69). MiniMax is ~5-10x cheaper than
+      # gpt-4.1-mini per token AND avoids the recurring OpenAI quota-
+      # exhaustion class that took the canary down 2026-05-03 (#265).
+      # Operators can pick langgraph / hermes via workflow_dispatch
+      # when they specifically need to exercise the OpenAI or SDK-
+      # native paths.
+      E2E_RUNTIME: ${{ github.event.inputs.runtime || 'claude-code' }}
+      # Pin the canary to a specific MiniMax model rather than relying
+      # on the per-runtime default ("sonnet" → routes to direct
+      # Anthropic, defeats the cost saving). Operators can override
+      # via workflow_dispatch by setting a different E2E_MODEL_SLUG
+      # input if they need to exercise a specific model. M2.7-highspeed
+      # is "Token Plan only" but cheap-per-token and fast.
+      E2E_MODEL_SLUG: ${{ github.event.inputs.model_slug || 'MiniMax-M2.7-highspeed' }}
       # Bound to 10 min so a stuck provision fails the run instead of
       # holding up the next cron firing. 15-min default in the script
       # is for the on-PR full lifecycle where we have more headroom.
@@ -88,37 +134,79 @@ jobs:
       E2E_KEEP_ORG: ${{ github.event.inputs.keep_org == 'true' && '1' || '' }}
       MOLECULE_CP_URL: ${{ vars.STAGING_CP_URL || 'https://staging-api.moleculesai.app' }}
       MOLECULE_ADMIN_TOKEN: ${{ secrets.CP_STAGING_ADMIN_API_TOKEN }}
-      # Provisioned tenant's default model (langgraph: openai:gpt-4.1-mini)
-      # needs OPENAI_API_KEY at first call. Sibling workflows
-      # e2e-staging-saas.yml + canary-staging.yml use the same secret;
-      # without this wire-up the tenant boots, accepts a2a messages,
-      # then returns "Could not resolve authentication method" — masked
-      # earlier by the a2a-sdk task-mode contract bugs PR #2558+#2563
-      # fixed. tests/e2e/test_staging_full_saas.sh:325 reads this and
-      # persists it as a workspace_secret on tenant create.
+      # MiniMax key is the canary's PRIMARY auth path. claude-code
+      # template's `minimax` provider routes ANTHROPIC_BASE_URL to
+      # api.minimax.io/anthropic and reads MINIMAX_API_KEY at boot.
+      # tests/e2e/test_staging_full_saas.sh branches SECRETS_JSON on
+      # which key is present — MiniMax wins when set.
+      E2E_MINIMAX_API_KEY: ${{ secrets.MOLECULE_STAGING_MINIMAX_API_KEY }}
+      # Direct-Anthropic alternative for operators who don't want to
+      # set up a MiniMax account (priority below MiniMax — first
+      # non-empty wins in test_staging_full_saas.sh's secrets-injection
+      # block). See #2578 PR comment for the rationale.
+      E2E_ANTHROPIC_API_KEY: ${{ secrets.MOLECULE_STAGING_ANTHROPIC_API_KEY }}
+      # OpenAI fallback — kept wired so operators can dispatch with
+      # E2E_RUNTIME=langgraph or =hermes and still have a working
+      # canary path. The script picks the right blob shape based on
+      # which key is non-empty.
       E2E_OPENAI_API_KEY: ${{ secrets.MOLECULE_STAGING_OPENAI_KEY }}
     steps:
       - uses: actions/checkout@de0fac2e4500dabe0009e67214ff5f5447ce83dd # v6.0.2
 
-      - name: Verify required secret present
+      - name: Verify required secrets present
         run: |
-          # Schedule-vs-dispatch hardening (mirrors the sweep-cf-* and
-          # redeploy-tenants-on-* workflows): hard-fail on missing secret
-          # for cron firing so a misconfigured-repo doesn't silently
-          # report green while doing nothing. Soft-skip on operator
-          # dispatch — operators can dispatch ad-hoc to verify a fix
-          # without setting up the secret first.
+          # Hard-fail on missing secret REGARDLESS of trigger. Previously
+          # this step soft-skipped on workflow_dispatch via `exit 0`, but
+          # `exit 0` only ends the STEP — subsequent steps still ran with
+          # the empty secret, the synth script fell through to the wrong
+          # SECRETS_JSON branch, and the canary failed 5 min later with a
+          # confusing "Agent error (Exception)" instead of the clean
+          # "secret missing" message at the top. Caught 2026-05-04 by
+          # dispatched run 25296530706: claude-code + missing MINIMAX
+          # silently used OpenAI keys but kept model=MiniMax-M2.7, then
+          # the workspace 401'd against MiniMax once it tried to call.
+          # Fix: exit 1 in both cron and dispatch paths. Operators who
+          # want to verify a YAML change without setting up the secret
+          # can read the verify-secrets step's stderr — the failure is
+          # itself the verification signal.
           if [ -z "${MOLECULE_ADMIN_TOKEN:-}" ]; then
-            if [ "${{ github.event_name }}" = "workflow_dispatch" ]; then
-              echo "::warning::CP_STAGING_ADMIN_API_TOKEN not set — synth E2E cannot run"
-              echo "::warning::Set it at Settings → Secrets and Variables → Actions"
-              exit 0
-            fi
             echo "::error::CP_STAGING_ADMIN_API_TOKEN secret missing — synth E2E cannot run"
             echo "::error::Set it at Settings → Secrets and Variables → Actions; pull from staging-CP's CP_ADMIN_API_TOKEN env in Railway."
             exit 1
           fi
 
+          # LLM-key requirement is per-runtime: claude-code accepts
+          # EITHER MiniMax OR direct-Anthropic (whichever is set first),
+          # langgraph + hermes use OpenAI (MOLECULE_STAGING_OPENAI_KEY).
+          case "${E2E_RUNTIME}" in
+            claude-code)
+              if [ -n "${E2E_MINIMAX_API_KEY:-}" ]; then
+                required_secret_name="MOLECULE_STAGING_MINIMAX_API_KEY"
+                required_secret_value="${E2E_MINIMAX_API_KEY}"
+              elif [ -n "${E2E_ANTHROPIC_API_KEY:-}" ]; then
+                required_secret_name="MOLECULE_STAGING_ANTHROPIC_API_KEY"
+                required_secret_value="${E2E_ANTHROPIC_API_KEY}"
+              else
+                required_secret_name="MOLECULE_STAGING_MINIMAX_API_KEY or MOLECULE_STAGING_ANTHROPIC_API_KEY"
+                required_secret_value=""
+              fi
+              ;;
+            langgraph|hermes)
+              required_secret_name="MOLECULE_STAGING_OPENAI_KEY"
+              required_secret_value="${E2E_OPENAI_API_KEY:-}"
+              ;;
+            *)
+              echo "::warning::Unknown E2E_RUNTIME='${E2E_RUNTIME}' — skipping LLM-key check"
+              required_secret_name=""
+              required_secret_value="present"
+              ;;
+          esac
+          if [ -n "$required_secret_name" ] && [ -z "$required_secret_value" ]; then
+            echo "::error::${required_secret_name} secret missing — runtime=${E2E_RUNTIME} cannot authenticate against its LLM provider"
+            echo "::error::Set it at Settings → Secrets and Variables → Actions, OR dispatch with a different runtime"
+            exit 1
+          fi
+
       - name: Install required tools
         run: |
           # The script depends on jq + curl (already on ubuntu-latest)

.github/workflows/e2e-staging-canvas.yml

@@ -184,8 +184,23 @@ jobs:
             exit 0
           fi
           echo "Deleting orphan tenant: $slug"
-          curl -sS -X DELETE "$MOLECULE_CP_URL/cp/admin/tenants/$slug" \
+          # Verify HTTP 2xx instead of `>/dev/null || true` swallowing
+          # failures. A 5xx or timeout previously looked identical to
+          # success, leaving the tenant alive for up to ~45 min until
+          # sweep-stale-e2e-orgs caught it. Surface failures as
+          # workflow warnings naming the slug. Don't `exit 1` — a single
+          # cleanup miss shouldn't fail-flag the canvas test when the
+          # actual smoke check passed; the sweeper is the safety net.
+          # See molecule-controlplane#420.
+          code=$(curl -sS -o /tmp/canvas-cleanup.out -w "%{http_code}" \
+            -X DELETE "$MOLECULE_CP_URL/cp/admin/tenants/$slug" \
             -H "Authorization: Bearer $ADMIN_TOKEN" \
             -H "Content-Type: application/json" \
-            -d "{\"confirm\":\"$slug\"}" >/dev/null || true
+            -d "{\"confirm\":\"$slug\"}" \
+            || echo "000")
+          if [ "$code" = "200" ] || [ "$code" = "204" ]; then
+            echo "[teardown] deleted $slug (HTTP $code)"
+          else
+            echo "::warning::canvas teardown for $slug returned HTTP $code — sweep-stale-e2e-orgs will catch it within ~45 min. Body: $(head -c 300 /tmp/canvas-cleanup.out 2>/dev/null)"
+          fi
           exit 0

.github/workflows/e2e-staging-external.yml

@@ -153,12 +153,28 @@ jobs:
           if [ -n "$orgs" ]; then
             echo "Safety-net sweep: deleting leftover orgs:"
             echo "$orgs"
+            # Per-slug verified DELETE — see molecule-controlplane#420.
+            # `>/dev/null 2>&1` previously hid every failure; surface
+            # non-2xx as workflow warnings so the run page names what
+            # leaked. Sweeper catches the rest within ~45 min.
+            leaks=()
             for slug in $orgs; do
-              curl -sS -X DELETE "$MOLECULE_CP_URL/cp/admin/tenants/$slug" \
+              code=$(curl -sS -o /tmp/external-cleanup.out -w "%{http_code}" \
+                -X DELETE "$MOLECULE_CP_URL/cp/admin/tenants/$slug" \
                 -H "Authorization: Bearer $ADMIN_TOKEN" \
                 -H "Content-Type: application/json" \
-                -d "{\"confirm\":\"$slug\"}" >/dev/null 2>&1
+                -d "{\"confirm\":\"$slug\"}" \
+                || echo "000")
+              if [ "$code" = "200" ] || [ "$code" = "204" ]; then
+                echo "[teardown] deleted $slug (HTTP $code)"
+              else
+                echo "::warning::external teardown for $slug returned HTTP $code — sweep-stale-e2e-orgs will catch it within ~45 min. Body: $(head -c 300 /tmp/external-cleanup.out 2>/dev/null)"
+                leaks+=("$slug")
+              fi
             done
+            if [ ${#leaks[@]} -gt 0 ]; then
+              echo "::warning::external teardown left ${#leaks[@]} leak(s): ${leaks[*]}"
+            fi
           else
             echo "Safety-net sweep: no leftover orgs to clean."
           fi

.github/workflows/e2e-staging-saas.yml

@@ -48,9 +48,9 @@ on:
   workflow_dispatch:
     inputs:
       runtime:
-        description: "Runtime to test (hermes | claude-code | langgraph)"
+        description: "Runtime to test (claude-code [default, MiniMax] | hermes [OpenAI] | langgraph [OpenAI])"
         required: false
-        default: "hermes"
+        default: "claude-code"
       keep_org:
         description: "Skip teardown for debugging (only use via manual dispatch!)"
         required: false
@@ -83,11 +83,32 @@ jobs:
       # retrieval + teardown. Configure in
       # Settings → Secrets and variables → Actions → Repository secrets.
       MOLECULE_ADMIN_TOKEN: ${{ secrets.MOLECULE_STAGING_ADMIN_TOKEN }}
-      # OpenAI key for workspace LLM calls (section 8 A2A). Without it,
-      # Hermes runtime crashes at boot with "No provider API key found".
-      # Configure at Settings → Secrets → Actions → MOLECULE_STAGING_OPENAI_KEY.
+      # MiniMax is the PRIMARY LLM auth path post-2026-05-04. Switched
+      # from hermes+OpenAI default after #2578 (the staging OpenAI key
+      # account went over quota and stayed dead for 36+ hours, taking
+      # the full-lifecycle E2E red on every provisioning-critical push).
+      # claude-code template's `minimax` provider routes
+      # ANTHROPIC_BASE_URL to api.minimax.io/anthropic and reads
+      # MINIMAX_API_KEY at boot — separate billing account so an
+      # OpenAI quota collapse no longer wedges the gate. Mirrors the
+      # canary-staging.yml + continuous-synth-e2e.yml migrations.
+      E2E_MINIMAX_API_KEY: ${{ secrets.MOLECULE_STAGING_MINIMAX_API_KEY }}
+      # Direct-Anthropic alternative for operators who don't want to
+      # set up a MiniMax account (priority below MiniMax — first
+      # non-empty wins in test_staging_full_saas.sh's secrets-injection
+      # block). See #2578 PR comment for the rationale.
+      E2E_ANTHROPIC_API_KEY: ${{ secrets.MOLECULE_STAGING_ANTHROPIC_API_KEY }}
+      # OpenAI fallback — kept wired so an operator-dispatched run with
+      # E2E_RUNTIME=hermes or =langgraph via workflow_dispatch can still
+      # exercise the OpenAI path.
       E2E_OPENAI_API_KEY: ${{ secrets.MOLECULE_STAGING_OPENAI_KEY }}
-      E2E_RUNTIME: ${{ github.event.inputs.runtime || 'hermes' }}
+      E2E_RUNTIME: ${{ github.event.inputs.runtime || 'claude-code' }}
+      # Pin the model when running on the default claude-code path —
+      # the per-runtime default ("sonnet") routes to direct Anthropic
+      # and defeats the cost saving. Operators can override via the
+      # workflow_dispatch flow (no input wired here yet — runtime
+      # override is enough for ad-hoc).
+      E2E_MODEL_SLUG: ${{ github.event.inputs.runtime == 'hermes' && 'openai/gpt-4o' || github.event.inputs.runtime == 'langgraph' && 'openai:gpt-4o' || 'MiniMax-M2.7-highspeed' }}
       E2E_RUN_ID: "${{ github.run_id }}-${{ github.run_attempt }}"
       E2E_KEEP_ORG: ${{ github.event.inputs.keep_org && '1' || '0' }}
 
@@ -102,13 +123,45 @@ jobs:
           fi
           echo "Admin token present ✓"
 
-      - name: Verify OpenAI key present
+      - name: Verify LLM key present
         run: |
-          if [ -z "$E2E_OPENAI_API_KEY" ]; then
-            echo "::error::MOLECULE_STAGING_OPENAI_KEY secret not set — workspaces will fail at boot with 'No provider API key found'"
+          # Per-runtime key check — claude-code uses MiniMax; hermes /
+          # langgraph (operator-dispatched only) use OpenAI. Hard-fail
+          # rather than soft-skip per #2578's lesson — empty key
+          # silently falls through to the wrong SECRETS_JSON branch and
+          # produces a confusing auth error 5 min later instead of the
+          # clean "secret missing" message at the top.
+          case "${E2E_RUNTIME}" in
+            claude-code)
+              # Either MiniMax OR direct-Anthropic works — first
+              # non-empty wins in the test script's secrets-injection
+              # priority chain.
+              if [ -n "${E2E_MINIMAX_API_KEY:-}" ]; then
+                required_secret_name="MOLECULE_STAGING_MINIMAX_API_KEY"
+                required_secret_value="${E2E_MINIMAX_API_KEY}"
+              elif [ -n "${E2E_ANTHROPIC_API_KEY:-}" ]; then
+                required_secret_name="MOLECULE_STAGING_ANTHROPIC_API_KEY"
+                required_secret_value="${E2E_ANTHROPIC_API_KEY}"
+              else
+                required_secret_name="MOLECULE_STAGING_MINIMAX_API_KEY or MOLECULE_STAGING_ANTHROPIC_API_KEY"
+                required_secret_value=""
+              fi
+              ;;
+            langgraph|hermes)
+              required_secret_name="MOLECULE_STAGING_OPENAI_KEY"
+              required_secret_value="${E2E_OPENAI_API_KEY:-}"
+              ;;
+            *)
+              echo "::warning::Unknown E2E_RUNTIME='${E2E_RUNTIME}' — skipping LLM-key check"
+              required_secret_name=""
+              required_secret_value="present"
+              ;;
+          esac
+          if [ -n "$required_secret_name" ] && [ -z "$required_secret_value" ]; then
+            echo "::error::${required_secret_name} secret not set for runtime=${E2E_RUNTIME} — workspaces will fail at boot with 'No provider API key found'"
             exit 2
           fi
-          echo "OpenAI key present ✓ (len=${#E2E_OPENAI_API_KEY})"
+          echo "LLM key present ✓ (runtime=${E2E_RUNTIME}, key=${required_secret_name}, len=${#required_secret_value})"
 
       - name: CP staging health preflight
         run: |
@@ -164,11 +217,27 @@ jobs:
                         and o.get('instance_status') not in ('purged',)]
           print('\n'.join(candidates))
           " 2>/dev/null)
+          # Per-slug verified DELETE (was `>/dev/null || true` — see
+          # molecule-controlplane#420). Surface non-2xx as a workflow
+          # warning naming the leaked slug; don't exit 1 (sweeper is
+          # the safety net within ~45 min).
+          leaks=()
           for slug in $orgs; do
             echo "Safety-net teardown: $slug"
-            curl -sS -X DELETE "$MOLECULE_CP_URL/cp/admin/tenants/$slug" \
+            code=$(curl -sS -o /tmp/saas-cleanup.out -w "%{http_code}" \
+              -X DELETE "$MOLECULE_CP_URL/cp/admin/tenants/$slug" \
               -H "Authorization: Bearer $ADMIN_TOKEN" \
               -H "Content-Type: application/json" \
-              -d "{\"confirm\":\"$slug\"}" >/dev/null || true
+              -d "{\"confirm\":\"$slug\"}" \
+              || echo "000")
+            if [ "$code" = "200" ] || [ "$code" = "204" ]; then
+              echo "[teardown] deleted $slug (HTTP $code)"
+            else
+              echo "::warning::saas teardown for $slug returned HTTP $code — sweep-stale-e2e-orgs will catch it within ~45 min. Body: $(head -c 300 /tmp/saas-cleanup.out 2>/dev/null)"
+              leaks+=("$slug")
+            fi
           done
+          if [ ${#leaks[@]} -gt 0 ]; then
+            echo "::warning::saas teardown left ${#leaks[@]} leak(s): ${leaks[*]}"
+          fi
           exit 0

.github/workflows/e2e-staging-sanity.yml

@@ -143,10 +143,25 @@ jobs:
                         and o.get('status') not in ('purged',)]
           print('\n'.join(candidates))
           " 2>/dev/null)
+          # Per-slug verified DELETE — see molecule-controlplane#420.
+          # Failures surface as workflow warnings; the sweeper is the
+          # safety net within ~45 min.
+          leaks=()
           for slug in $orgs; do
-            curl -sS -X DELETE "$MOLECULE_CP_URL/cp/admin/tenants/$slug" \
+            code=$(curl -sS -o /tmp/sanity-cleanup.out -w "%{http_code}" \
+              -X DELETE "$MOLECULE_CP_URL/cp/admin/tenants/$slug" \
               -H "Authorization: Bearer $ADMIN_TOKEN" \
               -H "Content-Type: application/json" \
-              -d "{\"confirm\":\"$slug\"}" >/dev/null || true
+              -d "{\"confirm\":\"$slug\"}" \
+              || echo "000")
+            if [ "$code" = "200" ] || [ "$code" = "204" ]; then
+              echo "[teardown] deleted $slug (HTTP $code)"
+            else
+              echo "::warning::sanity teardown for $slug returned HTTP $code — sweep-stale-e2e-orgs will catch it within ~45 min. Body: $(head -c 300 /tmp/sanity-cleanup.out 2>/dev/null)"
+              leaks+=("$slug")
+            fi
           done
+          if [ ${#leaks[@]} -gt 0 ]; then
+            echo "::warning::sanity teardown left ${#leaks[@]} leak(s): ${leaks[*]}"
+          fi
           exit 0

.github/workflows/publish-runtime.yml

@@ -327,13 +327,19 @@ jobs:
             echo "::error::publish job did not expose a version output — cascade cannot fan out"
             exit 1
           fi
-          # Source of truth: manifest.json workspace_templates (PR #2536 pruned
-          # to 4 actively-supported runtimes: claude-code, hermes, openclaw, codex).
-          # Removed langgraph/crewai/autogen/deepagents/gemini-cli (deprecated, no
-          # shipping images); added codex (had been missing since #2512).
-          # Long-term: derive this list from manifest.json so the cascade can't
-          # drift again — tracked in RFC #388 as a Phase-1 invariant.
-          TEMPLATES="claude-code hermes openclaw codex"
+          # All 9 active workspace template repos. The PR #2536 pruning
+          # ("deprecated, no shipping images") was empirically wrong:
+          # continuous-synth-e2e.yml defaults to langgraph as its primary
+          # canary (line 44), and every excluded template had successful
+          # publish-image runs as of 2026-05-03 — none were dormant.
+          # Symptom of the prune: today's a2a-sdk strict-mode fix
+          # (#2566 / commit e1628c4) cascaded to 4 templates but never
+          # reached langgraph, so the synth-E2E correctly canary'd a fix
+          # that had landed but not deployed. Re-added the 5 templates.
+          # Long-term: derive this list from manifest.json so cascade
+          # scope can't drift from E2E scope — tracked in RFC #388 as a
+          # Phase-1 invariant.
+          TEMPLATES="claude-code hermes openclaw codex langgraph crewai autogen deepagents gemini-cli"
           FAILED=""
           for tpl in $TEMPLATES; do
             REPO="Molecule-AI/molecule-ai-workspace-template-$tpl"

.github/workflows/redeploy-tenants-on-main.yml

@@ -17,7 +17,7 @@ name: redeploy-tenants-on-main
 #   1. publish-workspace-server-image completes → new :latest in GHCR.
 #   2. This workflow fires via workflow_run, waits 30s for GHCR's
 #      CDN to propagate the new tag to the region the tenants pull from.
-#   3. Calls redeploy-fleet with canary_slug=hongmingwang and a 60s
+#   3. Calls redeploy-fleet with canary_slug=hongming and a 60s
 #      soak. Canary proves the image boots; batches follow.
 #   4. Any failure aborts the rollout and leaves older tenants on the
 #      prior image — safer default than half-and-half state.
@@ -56,7 +56,12 @@ on:
         description: 'Tenant slug to deploy first + soak (empty = skip canary, fan out immediately).'
         required: false
         type: string
-        default: 'hongmingwang'
+        # Must be an actual prod tenant slug (current: hongming,
+        # chloe-dong, reno-stars). The previous default 'hongmingwang'
+        # didn't match any tenant — CP soft-skipped the missing canary
+        # and the fleet rolled out without the soak gate, defeating the
+        # whole point of canary-first.
+        default: 'hongming'
       soak_seconds:
         description: 'Seconds to wait after canary before fanning out.'
         required: false
@@ -148,7 +153,7 @@ jobs:
           CP_URL: ${{ vars.CP_URL || 'https://api.moleculesai.app' }}
           CP_ADMIN_API_TOKEN: ${{ secrets.CP_ADMIN_API_TOKEN }}
           TARGET_TAG: ${{ steps.tag.outputs.target_tag }}
-          CANARY_SLUG: ${{ inputs.canary_slug || 'hongmingwang' }}
+          CANARY_SLUG: ${{ inputs.canary_slug || 'hongming' }}
           SOAK_SECONDS: ${{ inputs.soak_seconds || '60' }}
           BATCH_SIZE: ${{ inputs.batch_size || '3' }}
           DRY_RUN: ${{ inputs.dry_run || false }}

.github/workflows/retarget-main-to-staging.yml

@@ -26,11 +26,22 @@ jobs:
     runs-on: ubuntu-latest
     # Only fire for bot-authored PRs. Human CEO PRs (staging→main promotion)
     # are intentional and pass through.
+    #
+    # Head-ref guard: never retarget a PR whose head IS `staging` — those
+    # are the auto-promote staging→main PRs (opened by molecule-ai[bot]
+    # since #2586 switched to an App token, which now passes the bot
+    # filter below). Retargeting head=staging onto base=staging fails
+    # with HTTP 422 "no new commits between base 'staging' and head
+    # 'staging'", which used to surface as a noisy red workflow run on
+    # every auto-promote (caught 2026-05-03 on PR #2588).
     if: >-
-      github.event.pull_request.user.type == 'Bot'
-      || endsWith(github.event.pull_request.user.login, '[bot]')
-      || github.event.pull_request.user.login == 'app/molecule-ai'
-      || github.event.pull_request.user.login == 'molecule-ai[bot]'
+      github.event.pull_request.head.ref != 'staging'
+      && (
+        github.event.pull_request.user.type == 'Bot'
+        || endsWith(github.event.pull_request.user.login, '[bot]')
+        || github.event.pull_request.user.login == 'app/molecule-ai'
+        || github.event.pull_request.user.login == 'molecule-ai[bot]'
+      )
     steps:
       - name: Retarget PR base to staging
         id: retarget

.github/workflows/sweep-stale-e2e-orgs.yml

@@ -25,16 +25,23 @@ name: Sweep stale e2e-* orgs (staging)
 
 on:
   schedule:
-    # Every hour on the hour. E2E orgs are short-lived (~10-25 min wall
-    # clock from create to teardown). Anything older than the
-    # MAX_AGE_MINUTES threshold below is presumed dead.
-    - cron: '0 * * * *'
+    # Every 15 min. E2E orgs are short-lived (~8-25 min wall clock from
+    # create to teardown — canary is ~8 min, full SaaS ~25 min). The
+    # previous hourly + 120-min stale threshold meant a leaked tenant
+    # could keep an EC2 alive for up to 2 hours, eating ~2 vCPU per
+    # leak. Tightening the cadence + threshold reduces the worst-case
+    # leak window from 120 min to ~45 min (15-min sweep cadence + 30-min
+    # threshold) without risk of catching in-progress runs (the longest
+    # e2e run is the 25-min canary, well under the 30-min threshold).
+    # See molecule-controlplane#420 for the leak-class accounting that
+    # motivated this tightening.
+    - cron: '*/15 * * * *'
   workflow_dispatch:
     inputs:
       max_age_minutes:
-        description: "Delete e2e-* orgs older than N minutes (default 120)"
+        description: "Delete e2e-* orgs older than N minutes (default 30)"
         required: false
-        default: "120"
+        default: "30"
       dry_run:
         description: "Dry run only — list what would be deleted"
         required: false
@@ -58,7 +65,7 @@ jobs:
     env:
       MOLECULE_CP_URL: https://staging-api.moleculesai.app
       ADMIN_TOKEN: ${{ secrets.MOLECULE_STAGING_ADMIN_TOKEN }}
-      MAX_AGE_MINUTES: ${{ github.event.inputs.max_age_minutes || '120' }}
+      MAX_AGE_MINUTES: ${{ github.event.inputs.max_age_minutes || '30' }}
       DRY_RUN: ${{ github.event.inputs.dry_run || 'false' }}
       # Refuse to delete more than this many orgs in one tick. If the
       # CP DB is briefly empty (or the admin endpoint goes weird and

COVERAGE_FLOOR.md

@@ -1,7 +1,7 @@
 # Coverage Floor
 
-CI enforces three coverage gates on `workspace-server` (Go). All defined in
-`.github/workflows/ci.yml` → `platform-build` job.
+CI enforces coverage gates on two surfaces — `workspace-server` (Go) and
+`workspace/` (Python). All defined in `.github/workflows/ci.yml`.
 
 ## Current floors (2026-04-23)
 
@@ -76,3 +76,51 @@ This gate makes "no untested critical paths merged" a mechanical property of
 the CI, not a behavioural property of QA agents or individual reviewers —
 which is the only way to make it survive fleet outages, agent rotations, or
 QA process changes.
+
+## Python (workspace/) — added 2026-05-04 from #2790
+
+The Python side has its own gates in the `python-lint` job:
+
+| Gate | Threshold | Where |
+|---|---|---|
+| **Total floor** | `86%` | `workspace/pytest.ini` `--cov-fail-under=86` (issue #1817) |
+| **Critical-path per-file floor** | `75%` | Inline shell step after the pytest run |
+
+### Critical-path Python files
+
+These handle multi-tenant routing, auth tokens, and inbox dispatch. A
+coverage drop here is the same risk shape as a Go-side `tokens*` /
+`secrets*` file regressing below 10%.
+
+- `workspace/a2a_mcp_server.py` — MCP dispatcher (PR #2766 / #2771)
+- `workspace/mcp_cli.py` — molecule-mcp standalone CLI entry
+- `workspace/a2a_tools.py` — workspace-scoped tool implementations
+- `workspace/inbox.py` — multi-workspace inbox + per-workspace cursors
+- `workspace/platform_auth.py` — per-workspace token resolver
+
+### Why 75% (vs 86% total)
+
+The total floor averages ~6000 lines across `workspace/`. A single MCP
+file could drop to ~50% with no CI complaint as long as other modules
+compensate. The per-file floor closes that distribution gap. 75% sits
+below current actuals (80–96% as of 2026-05-04) — strictly additive,
+no existing PR fails.
+
+### Python ratchet plan
+
+| Date | Total | Per-file critical | Notes |
+|---|---|---|---|
+| 2026-05-04 | 86% | 75% | Initial gate (this file). |
+| 2026-06-04 | 86% | 80% | First ratchet — at-floor files must catch up. |
+| 2026-07-04 | 88% | 85% | |
+| 2026-08-04 | 90% | 90% | Target steady-state. |
+
+### Why this Python gate exists
+
+Issue #2790, after the PR #2766 → PR #2771 cycle. PR #2766 added
+multi-workspace routing through `a2a_tools.py` + `a2a_mcp_server.py`,
+shipped to main with green CI, but the dispatcher silently dropped a
+load-bearing kwarg for 4 of 9 tools — caught only by post-merge code
+review. The structural drift gate (`test_dispatcher_schema_drift.py`,
+PR #2791) catches the schema↔dispatcher mismatch class; this floor
+catches the broader "MCP-critical file regressed" class.

canvas/e2e/staging-setup.ts

@@ -169,7 +169,17 @@ export default async function globalSetup(_config: FullConfig): Promise<void> {
         orgID = row.id;
         return true;
       }
-      if (row.instance_status === "failed") throw new Error(`provision failed: ${slug}`);
+      if (row.instance_status === "failed") {
+        // Dump every diagnostic field the admin row carries — boot stage,
+        // last error, terraform/SSM state, etc. The bare slug message used
+        // to surface ZERO context, so triaging a failed provision meant
+        // re-running locally to repro. Now the failure log carries enough
+        // to point at the right subsystem (CP/AWS/SSM/runtime) without a
+        // second round-trip.
+        throw new Error(
+          `provision failed: ${slug} — admin-orgs row: ${JSON.stringify(row)}`,
+        );
+      }
       return null;
     },
     PROVISION_TIMEOUT_MS,
@@ -249,7 +259,17 @@ export default async function globalSetup(_config: FullConfig): Promise<void> {
       if (r.status !== 200) return null;
       if (r.body?.status === "online") return true;
       if (r.body?.status === "failed") {
-        throw new Error(`Workspace failed: ${r.body.last_sample_error || ""}`);
+        // last_sample_error is often empty when the failure happens before
+        // the agent emits a sample (e.g. boot crash, image pull error,
+        // missing PYTHONPATH, OpenAI quota at startup). Dumping the full
+        // body gives triage the boot_stage / last_error / image fields it
+        // needs without a second probe. Otherwise this propagates as a
+        // bare "Workspace failed: " — the exact useless message that
+        // sent #2632 to the issue tracker.
+        const detail = r.body.last_sample_error
+          ? r.body.last_sample_error
+          : `(no last_sample_error) full body: ${JSON.stringify(r.body)}`;
+        throw new Error(`Workspace failed: ${detail}`);
       }
       return null;
     },

canvas/src/app/globals.css

@@ -1,6 +1,15 @@
 @import "tailwindcss";
 @plugin "@tailwindcss/typography";
 
+/*
+ * Tailwind v4 defaults the `dark:` variant to `prefers-color-scheme: dark`.
+ * Our theme switcher writes `data-theme="dark"` on <html> instead (so user
+ * choice via the toggle wins over OS preference). Re-bind `dark:` to that
+ * attribute so component classes like `dark:bg-zinc-800` track the same
+ * source of truth as the `[data-theme="dark"]` token overrides below.
+ */
+@custom-variant dark (&:where([data-theme="dark"], [data-theme="dark"] *));
+
 /*
  * Load order:
  *   1. Tailwind core (v4) — provides preflight + utility generation.

canvas/src/app/page.tsx

@@ -54,7 +54,7 @@ export default function Home() {
   if (hydrating) {
     return (
       <div className="fixed inset-0 flex items-center justify-center bg-surface">
-        <div className="flex flex-col items-center gap-3">
+        <div role="status" aria-live="polite" className="flex flex-col items-center gap-3">
           <Spinner size="lg" />
           <span className="text-xs text-ink-soft">Loading canvas...</span>
         </div>

canvas/src/components/A2ATopologyOverlay.tsx

@@ -138,14 +138,37 @@ export function A2ATopologyOverlay() {
   // Stable Zustand action reference — safe to call inside effects
   const setA2AEdges = useCanvasStore((s) => s.setA2AEdges);
 
-  // Read the nodes array as a primitive ref; derive visible IDs outside the selector
-  const nodes = useCanvasStore((s) => s.nodes);
+  // Subscribe to a STABLE STRING KEY of visible workspace IDs, not the
+  // nodes array itself. Zustand returns a new array reference on every
+  // store update (status flips, position drags, peer-discovery writes,
+  // workspace-tab opens, etc.) — even when the set of visible IDs is
+  // unchanged. Selecting a sorted-CSV string makes Zustand's default
+  // shallow-equal short-circuit the re-render unless the actual ID set
+  // changes.
+  //
+  // Why this matters: previously visibleIds was useMemo'd on `nodes`, so
+  // the array reference recreated on every store mutation. fetchAndUpdate
+  // (useCallback'd on visibleIds) then recreated, the useEffect re-fired,
+  // it tore down the 60s setInterval and immediately re-ran the fan-out.
+  // With ~5 store updates/second from heartbeats + polling, the canvas
+  // hammered /workspaces/<id>/activity?type=delegation 5×N requests/sec
+  // until edge rate-limit kicked in with HTTP 429. The recursive React
+  // render trace in the original bug report (uE → ux → uE → ux ...) is
+  // the symptom of this re-render storm.
+  //
+  // The fix is purely the dependency-stability change here; the fetch
+  // logic is unchanged.
+  const visibleIdsKey = useCanvasStore((s) =>
+    s.nodes
+      .filter((n) => !n.hidden)
+      .map((n) => n.id)
+      .sort()
+      .join(",")
+  );
 
-  // IDs of visible (non-nested, non-hidden) workspace nodes.
-  // Recomputed only when the nodes array reference changes.
   const visibleIds = useMemo(
-    () => nodes.filter((n) => !n.hidden).map((n) => n.id),
-    [nodes]
+    () => (visibleIdsKey ? visibleIdsKey.split(",") : []),
+    [visibleIdsKey]
   );
 
   // Fetch delegation activity for all visible workspaces and rebuild overlay edges.

canvas/src/components/ApprovalBanner.tsx

@@ -73,14 +73,19 @@ export function ApprovalBanner() {
                 <button
                   type="button"
                   onClick={() => handleDecide(approval, "approved")}
-                  className="px-3 py-1.5 bg-emerald-600 hover:bg-emerald-500 text-xs rounded-lg text-white font-medium transition-colors"
+                  // Hover DARKER not lighter — emerald-500 on white text
+                  // drops contrast vs emerald-700.
+                  className="px-3 py-1.5 bg-emerald-600 hover:bg-emerald-700 text-xs rounded-lg text-white font-medium transition-colors focus:outline-none focus-visible:ring-2 focus-visible:ring-offset-2 focus-visible:ring-offset-amber-950 focus-visible:ring-emerald-400/70"
                 >
                   Approve
                 </button>
                 <button
                   type="button"
                   onClick={() => handleDecide(approval, "denied")}
-                  className="px-3 py-1.5 bg-surface-card hover:bg-surface-card text-xs rounded-lg text-ink-mid transition-colors"
+                  // Was a no-op hover (`bg-surface-card hover:bg-surface-card`).
+                  // Lift to surface-elevated on hover so the button visibly
+                  // responds before a destructive deny.
+                  className="px-3 py-1.5 bg-surface-card hover:bg-surface-elevated hover:text-ink text-xs rounded-lg text-ink-mid transition-colors focus:outline-none focus-visible:ring-2 focus-visible:ring-offset-2 focus-visible:ring-offset-amber-950 focus-visible:ring-amber-400/70"
                 >
                   Deny
                 </button>

canvas/src/components/BatchActionBar.tsx

@@ -30,6 +30,24 @@ export function BatchActionBar() {
     if (count === 0 && hasFailedBatch) setHasFailedBatch(false);
   }, [count, hasFailedBatch]);
 
+  // Esc clears selection — the deselect button title has been promising
+  // "(Escape)" since the bar shipped, but no handler was wired. Skip when
+  // the confirm dialog is open (`pending !== null`) so the dialog's own
+  // Esc-cancels takes precedence and we don't double-handle the keystroke.
+  // Also skip during a busy in-flight action so the user can't accidentally
+  // strand a partial-failure mid-flight.
+  useEffect(() => {
+    if (count === 0 || pending !== null || busy) return;
+    const onKey = (e: KeyboardEvent) => {
+      if (e.key === "Escape") {
+        e.stopPropagation();
+        clearSelection();
+      }
+    };
+    window.addEventListener("keydown", onKey);
+    return () => window.removeEventListener("keydown", onKey);
+  }, [count, pending, busy, clearSelection]);
+
   // Hide when nothing is selected. Hide for single-node selection UNLESS a
   // partial-failure left a survivor awaiting retry.
   if (count === 0) return null;
@@ -129,7 +147,7 @@ export function BatchActionBar() {
         onClick={clearSelection}
         aria-label="Clear selection"
         title="Clear selection (Escape)"
-        className="p-1.5 rounded-lg text-[12px] text-ink-mid hover:text-ink hover:bg-surface-card/50 transition-colors disabled:opacity-50 focus-visible:outline-none focus-visible:ring-2 focus-visible:ring-zinc-500/70"
+        className="p-1.5 rounded-lg text-[12px] text-ink-mid hover:text-ink hover:bg-surface-card/50 transition-colors disabled:opacity-50 focus-visible:outline-none focus-visible:ring-2 focus-visible:ring-accent/50"
       >
         ✕
       </button>

canvas/src/components/BundleDropZone.tsx

@@ -117,9 +117,11 @@ export function BundleDropZone() {
         📦 Import bundle
       </button>
 
-      {/* Visual overlay when dragging */}
+      {/* Visual overlay when dragging — was hardcoded blue-950/blue-400
+          which doesn't flip with theme. accent colors stay visually
+          consistent with the rest of the canvas in both modes. */}
       {isDragging && (
-        <div className="fixed inset-0 z-20 flex items-center justify-center bg-blue-950/40 backdrop-blur-sm border-2 border-dashed border-blue-400/50 pointer-events-none">
+        <div className="fixed inset-0 z-20 flex items-center justify-center bg-accent/15 backdrop-blur-sm border-2 border-dashed border-accent/40 pointer-events-none">
           <div className="bg-surface-sunken/95 border border-accent/50 rounded-2xl px-8 py-6 shadow-2xl text-center">
             <div className="text-3xl mb-2" aria-hidden="true">📦</div>
             <div className="text-sm font-semibold text-ink">Drop Bundle to Import</div>
@@ -128,10 +130,21 @@ export function BundleDropZone() {
         </div>
       )}
 
-      {/* Importing spinner */}
+      {/* Importing indicator — role=status + aria-live so SR users hear
+          "Importing bundle..." while the API call is in flight, not just
+          the result toast that fires after. motion-safe:animate-spin
+          respects prefers-reduced-motion (Tailwind's motion-safe variant
+          gates animation on the user's OS setting). */}
       {importing && (
-        <div className="fixed bottom-6 left-1/2 -translate-x-1/2 z-50 bg-surface-sunken/95 border border-line/60 rounded-xl px-5 py-3 shadow-2xl flex items-center gap-3">
-          <div className="w-4 h-4 border-2 border-sky-400 border-t-transparent rounded-full animate-spin" />
+        <div
+          role="status"
+          aria-live="polite"
+          className="fixed bottom-6 left-1/2 -translate-x-1/2 z-50 bg-surface-sunken/95 border border-line/60 rounded-xl px-5 py-3 shadow-2xl flex items-center gap-3"
+        >
+          <div
+            aria-hidden="true"
+            className="w-4 h-4 border-2 border-accent border-t-transparent rounded-full motion-safe:animate-spin"
+          />
           <span className="text-sm text-ink">Importing bundle...</span>
         </div>
       )}

canvas/src/components/CommunicationOverlay.tsx

@@ -32,11 +32,18 @@ export function CommunicationOverlay() {
 
   const fetchComms = useCallback(async () => {
     try {
-      // Fetch activity from all online workspaces
+      // Fan-out cap: each polled workspace = 1 round-trip. The platform
+      // rate limits at 600 req/min/IP; combined with heartbeats + other
+      // canvas polling, every workspace polled here costs ~6 req/min
+      // (1 every 30s × 1 per workspace). Capping at 3 keeps this
+      // overlay's footprint at 18 req/min worst case — well under
+      // budget even with 8+ workspaces visible. Caught 2026-05-04 when
+      // a user with 8+ workspaces (Design Director + 6 sub-agents +
+      // 3 standalones) saw sustained 429s in canvas console.
       const onlineNodes = nodesRef.current.filter((n) => n.data.status === "online");
       const allComms: Communication[] = [];
 
-      for (const node of onlineNodes.slice(0, 6)) {
+      for (const node of onlineNodes.slice(0, 3)) {
         try {
           const activities = await api.get<Array<{
             id: string;
@@ -91,10 +98,20 @@ export function CommunicationOverlay() {
   }, []);
 
   useEffect(() => {
+    // Gate polling on visibility — when the user collapses the overlay
+    // the data isn't being read, so the per-workspace fan-out becomes
+    // pure rate-limit overhead. Pre-fix this overlay polled regardless
+    // of whether the panel was shown, costing ~36 req/min from a
+    // hidden surface.
+    if (!visible) return;
     fetchComms();
-    const interval = setInterval(fetchComms, 10000);
+    // 30s cadence (was 10s). At 3-workspace fan-out that's 6 req/min
+    // worst case from this overlay. Combined with heartbeats (~30/min)
+    // and other canvas polling, leaves ample headroom under the 600/
+    // min/IP server-side rate limit even at 8+ workspace tenants.
+    const interval = setInterval(fetchComms, 30000);
     return () => clearInterval(interval);
-  }, [fetchComms]);
+  }, [fetchComms, visible]);
 
   if (!visible || comms.length === 0) {
     return (

canvas/src/components/ConfirmDialog.tsx

@@ -91,12 +91,15 @@ export function ConfirmDialog({
 
   if (!open || !mounted) return null;
 
+  // Hover goes DARKER, not lighter — lighter shades on white text drop
+  // contrast below AA on the accent and red ramps. Darker hovers stay
+  // readable in both light and dark themes.
   const confirmColors =
     confirmVariant === "danger"
-      ? "bg-red-600 hover:bg-red-500 text-white"
+      ? "bg-red-600 hover:bg-red-700 text-white"
       : confirmVariant === "warning"
-        ? "bg-amber-600 hover:bg-amber-500 text-white"
-        : "bg-accent-strong hover:bg-accent text-white";
+        ? "bg-amber-600 hover:bg-amber-700 text-white"
+        : "bg-accent hover:bg-accent-strong text-white";
 
   // Render via Portal so the fixed-position dialog escapes any containing block
   // (e.g. parents with transform, filter, will-change that break position:fixed).
@@ -123,7 +126,7 @@ export function ConfirmDialog({
             <button
               type="button"
               onClick={onCancel}
-              className="px-3.5 py-1.5 text-[13px] text-ink-mid hover:text-ink bg-surface-card hover:bg-surface-card border border-line rounded-lg transition-colors"
+              className="px-3.5 py-1.5 text-[13px] text-ink-mid hover:text-ink bg-surface-card hover:bg-surface-elevated border border-line hover:border-line-soft rounded-lg transition-colors focus:outline-none focus-visible:ring-2 focus-visible:ring-accent/40"
             >
               Cancel
             </button>
@@ -131,7 +134,7 @@ export function ConfirmDialog({
           <button
             type="button"
             onClick={onConfirm}
-            className={`px-3.5 py-1.5 text-[13px] rounded-lg transition-colors ${confirmColors}`}
+            className={`px-3.5 py-1.5 text-[13px] rounded-lg transition-colors focus:outline-none focus-visible:ring-2 focus-visible:ring-offset-2 focus-visible:ring-offset-surface-sunken focus-visible:ring-accent/60 ${confirmColors}`}
           >
             {confirmLabel}
           </button>

canvas/src/components/ConsoleModal.tsx

@@ -113,7 +113,10 @@ export function ConsoleModal({ workspaceId, workspaceName, open, onClose }: Prop
             ref={closeButtonRef}
             onClick={onClose}
             aria-label="Close"
-            className="text-ink-mid hover:text-ink text-sm px-2"
+            // 24x24 touch target (was ~10x16, well under WCAG 2.5.5).
+            // Hover bg makes the area visible; focus-visible ring matches
+            // the rest of the canvas chrome.
+            className="w-6 h-6 inline-flex items-center justify-center rounded text-sm text-ink-mid hover:text-ink hover:bg-surface-card/40 focus:outline-none focus-visible:ring-2 focus-visible:ring-accent/60 transition-colors"
           >
             ✕
           </button>
@@ -150,12 +153,19 @@ export function ConsoleModal({ workspaceId, workspaceName, open, onClose }: Prop
               type="button"
               onClick={() => {
                 if (navigator.clipboard) {
-                  navigator.clipboard.writeText(output);
+                  // Add success feedback — without it, clicking Copy
+                  // looked like a no-op since the previous hover bg was
+                  // also a no-op (`hover:bg-surface-card` on top of the
+                  // same base). Toast confirms the write actually fired.
+                  navigator.clipboard
+                    .writeText(output)
+                    .then(() => showToast("Console output copied", "success"))
+                    .catch(() => showToast("Copy failed", "error"));
                 } else {
                   showToast("Copy requires HTTPS — please select and copy manually", "info");
                 }
               }}
-              className="px-3 py-1.5 text-[11px] text-ink-mid hover:text-ink bg-surface-card hover:bg-surface-card border border-line rounded-lg transition-colors"
+              className="px-3 py-1.5 text-[11px] text-ink-mid hover:text-ink bg-surface-card hover:bg-surface-elevated border border-line hover:border-line-soft rounded-lg transition-colors focus:outline-none focus-visible:ring-2 focus-visible:ring-accent/60 focus-visible:ring-offset-2 focus-visible:ring-offset-surface"
             >
               Copy
             </button>
@@ -163,7 +173,10 @@ export function ConsoleModal({ workspaceId, workspaceName, open, onClose }: Prop
           <button
             type="button"
             onClick={onClose}
-            className="px-3 py-1.5 text-[11px] text-ink-mid bg-surface-card hover:bg-surface-card border border-line rounded-lg transition-colors"
+            // Was hover:bg-surface-card (same as base — silent no-op).
+            // Lift to surface-elevated so the button visibly responds,
+            // matching the Cancel button in ConfirmDialog.
+            className="px-3 py-1.5 text-[11px] text-ink-mid hover:text-ink bg-surface-card hover:bg-surface-elevated border border-line hover:border-line-soft rounded-lg transition-colors focus:outline-none focus-visible:ring-2 focus-visible:ring-accent/60 focus-visible:ring-offset-2 focus-visible:ring-offset-surface"
           >
             Close
           </button>

canvas/src/components/ContextMenu.tsx

@@ -29,15 +29,38 @@ export function ContextMenu() {
   const setPendingDelete = useCanvasStore((s) => s.setPendingDelete);
   const ref = useRef<HTMLDivElement>(null);
   const [actionLoading, setActionLoading] = useState(false);
+  // Clamped position — (left, top) from contextMenu may overflow when the
+  // user right-clicks near the right/bottom viewport edge. We measure the
+  // rendered menu and shift it back inside on the same frame the cursor
+  // opens it, so it never visibly clips. Falls back to the raw cursor
+  // coords until the rAF runs.
+  const [clamped, setClamped] = useState<{ x: number; y: number } | null>(null);
 
-  // Auto-focus first enabled item when menu opens
+  // Auto-focus first enabled item when menu opens, AND clamp position.
+  // Both run together in a single rAF so we avoid two synchronous layout
+  // reads + a paint between them.
   useEffect(() => {
     if (!contextMenu) return;
-    requestAnimationFrame(() => {
-      const first = ref.current?.querySelector<HTMLButtonElement>("button:not(:disabled)");
+    setClamped(null);
+    const raf = requestAnimationFrame(() => {
+      const node = ref.current;
+      if (!node) return;
+      const first = node.querySelector<HTMLButtonElement>("button:not(:disabled)");
       first?.focus();
+      // 8px viewport margin so the menu doesn't kiss the edge — matches
+      // the floating-tooltip top-edge clamp in Tooltip.tsx.
+      const margin = 8;
+      const rect = node.getBoundingClientRect();
+      const vw = window.innerWidth;
+      const vh = window.innerHeight;
+      let x = contextMenu.x;
+      let y = contextMenu.y;
+      if (x + rect.width + margin > vw) x = Math.max(margin, vw - rect.width - margin);
+      if (y + rect.height + margin > vh) y = Math.max(margin, vh - rect.height - margin);
+      if (x !== contextMenu.x || y !== contextMenu.y) setClamped({ x, y });
     });
-  }, [contextMenu?.nodeId]);
+    return () => cancelAnimationFrame(raf);
+  }, [contextMenu?.nodeId, contextMenu?.x, contextMenu?.y]);
 
   // Close on click outside or Escape
   useEffect(() => {
@@ -288,7 +311,7 @@ export function ContextMenu() {
       aria-label={`Actions for ${contextMenu.nodeData.name}`}
       onKeyDown={handleMenuKeyDown}
       className="fixed z-[60] min-w-[200px] bg-surface/95 backdrop-blur-xl border border-line/60 rounded-xl shadow-2xl shadow-black/60 py-1 overflow-hidden"
-      style={{ left: contextMenu.x, top: contextMenu.y }}
+      style={{ left: clamped?.x ?? contextMenu.x, top: clamped?.y ?? contextMenu.y }}
     >
       {/* Header */}
       <div className="px-3.5 py-2 border-b border-line/40 mb-0.5">
@@ -314,7 +337,7 @@ export function ContextMenu() {
             onClick={item.action}
             disabled={item.disabled}
             aria-disabled={item.disabled}
-            className={`w-full px-3.5 py-1.5 flex items-center gap-2.5 text-left text-[11px] transition-colors focus:outline-none focus:ring-1 focus:ring-inset focus:ring-zinc-600 disabled:opacity-25 disabled:cursor-not-allowed ${
+            className={`w-full px-3.5 py-1.5 flex items-center gap-2.5 text-left text-[11px] transition-colors focus:outline-none focus-visible:ring-1 focus-visible:ring-inset focus-visible:ring-accent/50 disabled:opacity-25 disabled:cursor-not-allowed ${
               item.danger
                 ? "text-bad hover:bg-red-950/40 hover:text-bad"
                 : "text-ink-mid hover:bg-surface-card/40 hover:text-ink"

canvas/src/components/CookieConsent.tsx

@@ -98,9 +98,17 @@ export function CookieConsent() {
   };
 
   return (
-    <div
-      role="dialog"
-      aria-modal="true"
+    // role="region" + aria-label, NOT role="dialog" + aria-modal. The
+    // banner is informational — it never blocks the page, never traps
+    // focus, and the user can keep using the canvas while it's up.
+    // Claiming aria-modal="true" without a focus trap is genuinely
+    // harmful for screen-reader users: they get told the rest of the
+    // page is inert, jump into the banner, and then can't escape.
+    // Region semantics let assistive tech navigate around it normally.
+    // (Also: forcing a modal cookie banner would be a dark pattern —
+    // GDPR explicitly discourages it.)
+    <section
+      role="region"
       aria-labelledby="cookie-consent-title"
       aria-describedby="cookie-consent-body"
       className="fixed bottom-0 left-0 right-0 z-[9999] border-t border-line bg-surface/95 backdrop-blur-sm p-4 shadow-[0_-4px_12px_rgba(0,0,0,0.4)]"
@@ -117,7 +125,7 @@ export function CookieConsent() {
             workspaces). See our{" "}
             <a
               href="https://moleculesai.app/legal/privacy"
-              className="text-accent underline hover:text-accent"
+              className="text-accent underline underline-offset-2 hover:text-accent-strong focus:outline-none focus-visible:ring-2 focus-visible:ring-accent/60 rounded-sm"
               target="_blank"
               rel="noreferrer"
             >
@@ -130,20 +138,20 @@ export function CookieConsent() {
           <button
             type="button"
             onClick={() => decide("rejected")}
-            className="rounded border border-line bg-surface-sunken px-4 py-2 text-sm text-ink hover:bg-surface-card"
+            className="rounded border border-line bg-surface-sunken px-4 py-2 text-sm text-ink hover:bg-surface-card focus:outline-none focus-visible:ring-2 focus-visible:ring-accent/60 focus-visible:ring-offset-2 focus-visible:ring-offset-surface"
           >
             Necessary only
           </button>
           <button
             type="button"
             onClick={() => decide("accepted")}
-            className="rounded border border-accent bg-accent-strong px-4 py-2 text-sm font-medium text-white hover:bg-accent"
+            className="rounded border border-accent bg-accent-strong px-4 py-2 text-sm font-medium text-white hover:bg-accent focus:outline-none focus-visible:ring-2 focus-visible:ring-accent/60 focus-visible:ring-offset-2 focus-visible:ring-offset-surface"
           >
             Accept all
           </button>
         </div>
       </div>
-    </div>
+    </section>
   );
 }
 

canvas/src/components/CreateWorkspaceDialog.tsx

@@ -310,7 +310,7 @@ export function CreateWorkspaceButton() {
   return (
     <Dialog.Root open={open} onOpenChange={setOpen}>
       <Dialog.Trigger asChild>
-        <button type="button" className="fixed bottom-6 right-6 z-40 px-5 py-2.5 bg-accent-strong hover:bg-accent active:bg-accent-strong text-sm font-medium rounded-xl text-white shadow-lg shadow-blue-600/20 hover:shadow-xl hover:shadow-blue-500/30 transition-all duration-200 flex items-center gap-2">
+        <button type="button" className="fixed bottom-6 right-6 z-40 px-5 py-2.5 bg-accent hover:bg-accent-strong active:bg-accent text-sm font-medium rounded-xl text-white shadow-lg shadow-accent/20 hover:shadow-xl hover:shadow-accent/30 transition-all duration-200 flex items-center gap-2 focus:outline-none focus-visible:ring-2 focus-visible:ring-accent/60 focus-visible:ring-offset-2 focus-visible:ring-offset-surface">
           <svg
             width="14"
             height="14"
@@ -502,7 +502,7 @@ export function CreateWorkspaceButton() {
                   placeholder="sk-…"
                   aria-label="Hermes API key"
                   autoComplete="off"
-                  className="w-full bg-surface-card/60 border border-line/50 rounded-lg px-3 py-2 text-sm text-ink placeholder-zinc-600 focus:outline-none focus:border-violet-500/60 focus:ring-1 focus:ring-violet-500/20 transition-colors font-mono"
+                  className="w-full bg-surface-card/60 border border-line/50 rounded-lg px-3 py-2 text-sm text-ink placeholder-ink-soft focus:outline-none focus:border-violet-500/60 focus:ring-1 focus:ring-violet-500/20 transition-colors font-mono"
                 />
               </div>
 
@@ -527,7 +527,7 @@ export function CreateWorkspaceButton() {
                   autoComplete="off"
                   spellCheck={false}
                   list="hermes-model-suggestions"
-                  className="w-full bg-surface-card/60 border border-line/50 rounded-lg px-3 py-2 text-sm text-ink placeholder-zinc-600 focus:outline-none focus:border-violet-500/60 focus:ring-1 focus:ring-violet-500/20 transition-colors font-mono"
+                  className="w-full bg-surface-card/60 border border-line/50 rounded-lg px-3 py-2 text-sm text-ink placeholder-ink-soft focus:outline-none focus:border-violet-500/60 focus:ring-1 focus:ring-violet-500/20 transition-colors font-mono"
                 />
                 <datalist id="hermes-model-suggestions">
                   {HERMES_PROVIDERS.find((p) => p.id === hermesProvider)?.models.map(
@@ -552,7 +552,7 @@ export function CreateWorkspaceButton() {
 
           <div className="flex justify-end gap-2.5 mt-6">
             <Dialog.Close asChild>
-              <button type="button" className="px-4 py-2 bg-surface-card hover:bg-surface-card text-sm rounded-lg text-ink-mid transition-colors">
+              <button type="button" className="px-4 py-2 bg-surface-card hover:bg-surface-elevated hover:text-ink text-sm rounded-lg text-ink-mid transition-colors focus:outline-none focus-visible:ring-2 focus-visible:ring-accent/40 focus-visible:ring-offset-1 focus-visible:ring-offset-surface">
                 Cancel
               </button>
             </Dialog.Close>
@@ -560,7 +560,7 @@ export function CreateWorkspaceButton() {
               type="button"
               onClick={handleCreate}
               disabled={creating}
-              className="px-5 py-2 bg-accent-strong hover:bg-accent active:bg-accent-strong text-sm rounded-lg text-white disabled:opacity-50 transition-colors"
+              className="px-5 py-2 bg-accent hover:bg-accent-strong active:bg-accent text-sm rounded-lg text-white disabled:opacity-50 transition-colors focus:outline-none focus-visible:ring-2 focus-visible:ring-accent/60 focus-visible:ring-offset-1 focus-visible:ring-offset-surface"
             >
               {creating ? "Creating..." : "Create"}
             </button>
@@ -623,7 +623,7 @@ function InputField({
         placeholder={placeholder}
         min={type === "number" ? "0" : undefined}
         step={type === "number" ? "0.01" : undefined}
-        className={`w-full bg-surface-card/60 border border-line/50 rounded-lg px-3 py-2 text-sm text-ink placeholder-zinc-500 focus:outline-none focus:border-accent/60 focus:ring-1 focus:ring-accent/20 transition-colors ${mono ? "font-mono text-xs" : ""}`}
+        className={`w-full bg-surface-card/60 border border-line/50 rounded-lg px-3 py-2 text-sm text-ink placeholder-ink-soft focus:outline-none focus:border-accent/60 focus:ring-1 focus:ring-accent/20 transition-colors ${mono ? "font-mono text-xs" : ""}`}
       />
       {helper && (
         <p className="mt-1 text-xs text-ink-soft">{helper}</p>

canvas/src/components/DeleteCascadeConfirmDialog.tsx

@@ -127,13 +127,16 @@ export function DeleteCascadeConfirmDialog({
             </p>
           </div>
 
-          {/* Checkbox guard */}
+          {/* Checkbox guard. Ring-offset color was zinc-900 — the dialog
+              actually sits on bg-surface-sunken, so the offset showed
+              the wrong color through the ring gap. Switched to the
+              real bg + a danger-tinted ring. */}
           <label className="flex items-start gap-2.5 cursor-pointer group select-none">
             <input
               type="checkbox"
               checked={checked}
               onChange={(e) => onCheckedChange(e.target.checked)}
-              className="mt-0.5 w-4 h-4 rounded border-line bg-surface-card text-bad focus:ring-red-500 focus:ring-offset-0 focus:ring-offset-zinc-900 cursor-pointer"
+              className="mt-0.5 w-4 h-4 rounded border-line bg-surface-card text-bad cursor-pointer focus:outline-none focus-visible:ring-2 focus-visible:ring-red-500/60 focus-visible:ring-offset-2 focus-visible:ring-offset-surface-sunken"
             />
             <span className="text-[12px] text-ink-mid group-hover:text-ink-mid leading-relaxed">
               I understand this will permanently delete all listed workspaces and their data
@@ -145,7 +148,11 @@ export function DeleteCascadeConfirmDialog({
           <button
             type="button"
             onClick={onCancel}
-            className="px-3.5 py-1.5 text-[13px] text-ink-mid hover:text-ink bg-surface-card hover:bg-surface-card border border-line rounded-lg transition-colors"
+            // Was hover:bg-surface-card (same as base — silent no-op).
+            // Lift to surface-elevated to match the Cancel pattern in
+            // ConfirmDialog. Added focus-visible ring so keyboard users
+            // see where focus lands.
+            className="px-3.5 py-1.5 text-[13px] text-ink-mid hover:text-ink bg-surface-card hover:bg-surface-elevated border border-line hover:border-line-soft rounded-lg transition-colors focus:outline-none focus-visible:ring-2 focus-visible:ring-accent/40 focus-visible:ring-offset-2 focus-visible:ring-offset-surface-sunken"
           >
             Cancel
           </button>
@@ -153,9 +160,12 @@ export function DeleteCascadeConfirmDialog({
             type="button"
             onClick={onConfirm}
             disabled={!checked}
-            className={`px-3.5 py-1.5 text-[13px] rounded-lg transition-colors
+            // Hover goes DARKER, not lighter — bg-red-500 on white text
+            // drops contrast below AA vs bg-red-700. Same trap fixed in
+            // ConfirmDialog and ApprovalBanner. focus-visible ring matches.
+            className={`px-3.5 py-1.5 text-[13px] rounded-lg transition-colors focus:outline-none focus-visible:ring-2 focus-visible:ring-red-500/60 focus-visible:ring-offset-2 focus-visible:ring-offset-surface-sunken
               ${checked
-                ? "bg-red-600 hover:bg-red-500 text-white cursor-pointer"
+                ? "bg-red-600 hover:bg-red-700 text-white cursor-pointer"
                 : "bg-red-900/30 text-bad/40 cursor-not-allowed"
               }`}
           >

canvas/src/components/ExternalConnectModal.tsx

@@ -18,6 +18,157 @@
 import { useCallback, useState } from "react";
 import * as Dialog from "@radix-ui/react-dialog";
 
+type Tab = "python" | "curl" | "claude" | "mcp" | "hermes" | "codex" | "openclaw" | "fields";
+
+// Per-tab help metadata: docs link, where-to-install link, common errors.
+// All URLs verified against repo content (docs/guides/* file paths map to
+// docs.molecule.ai/docs/guides/*; canonical hostname confirmed by existing
+// blog post canonical metadata) or against the snippet text the operator
+// just copied. Never linking to a URL that wasn't already in product —
+// dead links here defeat the purpose of "more comprehensive instructions."
+const TAB_HELP: Record<
+  Tab,
+  {
+    docsUrl?: string;
+    docsLabel?: string;
+    downloadUrl?: string;
+    downloadLabel?: string;
+    commonIssues?: { symptom: string; check: string }[];
+  }
+> = {
+  mcp: {
+    docsUrl: "https://docs.molecule.ai/docs/guides/mcp-server-setup",
+    docsLabel: "MCP server setup guide",
+    downloadUrl: "https://pypi.org/project/molecule-ai-workspace-runtime/",
+    downloadLabel: "molecule-ai-workspace-runtime on PyPI",
+    commonIssues: [
+      {
+        symptom: "Tools not appearing in your agent",
+        check:
+          "Run `claude mcp list` (or your runtime's equivalent) — the molecule entry should be listed. If missing, re-run the `claude mcp add` line.",
+      },
+      {
+        symptom: "ConnectionRefused / DNS error on first call",
+        check:
+          "PLATFORM_URL must include the scheme (https://) and have no trailing slash. Verify with `curl $PLATFORM_URL/healthz`.",
+      },
+    ],
+  },
+  python: {
+    docsUrl:
+      "https://docs.molecule.ai/docs/guides/external-agent-registration",
+    docsLabel: "External agent registration guide",
+    downloadUrl: "https://pypi.org/project/molecule-ai-workspace-runtime/",
+    downloadLabel: "molecule-ai-workspace-runtime on PyPI",
+    commonIssues: [
+      {
+        symptom: "401 from /heartbeat",
+        check:
+          "AUTH_TOKEN expired or wrong workspace_id. Tokens are shown only once at create time — re-create the workspace to get a fresh token.",
+      },
+      {
+        symptom: "AGENT_URL not reachable from platform",
+        check:
+          "Public HTTPS URL required for inbound A2A. Use ngrok or Cloudflare Tunnel if your agent is behind NAT.",
+      },
+    ],
+  },
+  claude: {
+    docsUrl:
+      "https://docs.molecule.ai/docs/guides/external-agent-registration",
+    docsLabel: "External agent registration guide",
+    downloadUrl: "https://claude.com/claude-code",
+    downloadLabel: "Claude Code (claude.com)",
+    commonIssues: [
+      {
+        symptom: "plugin not installed",
+        check:
+          "Run `/plugin marketplace add Molecule-AI/molecule-mcp-claude-channel` then `/plugin install molecule@molecule-mcp-claude-channel` inside Claude Code, then `/reload-plugins`.",
+      },
+      {
+        symptom: "not on the approved channels allowlist",
+        check:
+          "Custom channels need `--dangerously-load-development-channels` on the launch command. Team/Enterprise orgs need admin to set `channelsEnabled` + `allowedChannelPlugins` in claude.ai admin settings.",
+      },
+      {
+        symptom: "Inbound messages not arriving",
+        check:
+          "Check stderr for `molecule channel: connected — watching N workspace(s)`. Verify ~/.claude/channels/molecule/.env has the right PLATFORM_URL + token.",
+      },
+    ],
+  },
+  hermes: {
+    docsUrl:
+      "https://docs.molecule.ai/docs/guides/external-agent-registration",
+    docsLabel: "External agent registration guide",
+    downloadUrl: "https://github.com/NousResearch/hermes-agent",
+    downloadLabel: "hermes-agent (NousResearch)",
+    commonIssues: [
+      {
+        symptom: "Gateway start failure",
+        check:
+          "Tail ~/.hermes/gateway.log. YAML duplicate-key in config.yaml is the most common cause — `gateway:` block must appear exactly once.",
+      },
+      {
+        symptom: "Plugin not discovered after install",
+        check:
+          "Run `pip show hermes-channel-molecule` to confirm install. Some hermes builds need `hermes plugin reload` before the new platform_plugins entry takes effect.",
+      },
+    ],
+  },
+  codex: {
+    docsUrl: "https://docs.molecule.ai/docs/guides/mcp-server-setup",
+    docsLabel: "MCP server setup guide",
+    downloadUrl: "https://github.com/openai/codex",
+    downloadLabel: "openai/codex",
+    commonIssues: [
+      {
+        symptom: "[mcp_servers.molecule] not loaded",
+        check:
+          "Codex must be ≥ 0.57. Check with `codex --version`; upgrade via `npm install -g @openai/codex@latest`.",
+      },
+      {
+        symptom: "TOML parse error after re-running setup",
+        check:
+          "TOML rejects duplicate `[mcp_servers.molecule]` tables. Open ~/.codex/config.toml and remove the old block before pasting the new one.",
+      },
+    ],
+  },
+  openclaw: {
+    docsUrl: "https://docs.molecule.ai/docs/guides/mcp-server-setup",
+    docsLabel: "MCP server setup guide",
+    commonIssues: [
+      {
+        symptom: "Gateway not starting",
+        check:
+          "Tail ~/.openclaw/gateway.log. The loopback bind requires :18789 to be free — check with `lsof -iTCP:18789`.",
+      },
+      {
+        symptom: "openclaw mcp set rejected",
+        check:
+          "The heredoc generates JSON; verify it parsed by running `jq < ~/.openclaw/mcp/molecule.json`. Re-run `openclaw mcp set` if the file is malformed.",
+      },
+    ],
+  },
+  curl: {
+    docsUrl:
+      "https://docs.molecule.ai/docs/guides/external-agent-registration",
+    docsLabel: "External agent registration guide",
+    commonIssues: [
+      {
+        symptom: "401 / 403 on register",
+        check:
+          "WORKSPACE_AUTH_TOKEN must be the value shown at workspace create. Tokens are shown only once.",
+      },
+    ],
+  },
+  fields: {
+    docsUrl:
+      "https://docs.molecule.ai/docs/guides/external-agent-registration",
+    docsLabel: "External agent registration guide",
+  },
+};
+
 export interface ExternalConnectionInfo {
   workspace_id: string;
   platform_url: string;
@@ -40,6 +191,22 @@ export interface ExternalConnectionInfo {
   // + inbound. Optional for backward compat with platforms that
   // haven't shipped PR #2413 yet.
   universal_mcp_snippet?: string;
+  // Hermes channel snippet — for operators whose external agent IS a
+  // hermes-agent session. Routes A2A traffic into the hermes gateway
+  // via the molecule-channel plugin (Molecule-AI/hermes-channel-molecule).
+  // Long-poll based (no tunnel) — same UX shape as the Claude Code
+  // channel tab. Gives hermes true push parity. Optional for backward
+  // compat with platforms that haven't shipped this PR yet.
+  hermes_channel_snippet?: string;
+  // Codex MCP config snippet — wires the molecule MCP server into
+  // ~/.codex/config.toml so codex agents can call platform tools.
+  // Outbound-tools-only today (codex's MCP client doesn't route
+  // notifications/*); push parity would need a separate bridge daemon.
+  codex_snippet?: string;
+  // OpenClaw MCP config snippet — wires molecule MCP + starts the
+  // openclaw gateway on loopback. Outbound-tools-only today; push
+  // parity on an external openclaw needs a sessions.steer bridge.
+  openclaw_snippet?: string;
 }
 
 interface Props {
@@ -47,13 +214,19 @@ interface Props {
   onClose: () => void;
 }
 
-type Tab = "python" | "curl" | "claude" | "mcp" | "fields";
-
 export function ExternalConnectModal({ info, onClose }: Props) {
-  // Default to Claude Code when the platform offers it — that's the
-  // newest + simplest path (no tunnel needed). Falls back to Python
-  // for older platform builds that don't ship the snippet.
-  const initialTab: Tab = info?.claude_code_channel_snippet ? "claude" : "python";
+  // Default to Universal MCP when the platform offers it — runtime-
+  // agnostic outbound tool path that works for any MCP-aware runtime
+  // (Claude Code, hermes, codex, etc.) and lets operators inspect the
+  // primitives before picking a runtime-specific tab. Python SDK is
+  // the fallback for platforms predating the universal_mcp_snippet
+  // field. Pre-2026-05-03 the default was "claude" (Claude Code first)
+  // but operators using non-Claude runtimes opened to a tab they had
+  // to skip past — universal MCP works for everyone as a starting
+  // point and the runtime-specific tabs are still one click away.
+  const initialTab: Tab = info?.universal_mcp_snippet
+    ? "mcp"
+    : "python";
   const [tab, setTab] = useState<Tab>(initialTab);
   const [copiedKey, setCopiedKey] = useState<string | null>(null);
 
@@ -108,6 +281,24 @@ export function ExternalConnectModal({ info, onClose }: Props) {
     'MOLECULE_WORKSPACE_TOKEN="<paste from create response>"',
     `MOLECULE_WORKSPACE_TOKEN="${info.auth_token}"`,
   );
+  // Hermes channel snippet uses MOLECULE_WORKSPACE_TOKEN (same env-var
+  // name as Universal MCP). Stamp the auth_token in so the operator's
+  // copy-paste is fully ready-to-run.
+  const filledHermes = info.hermes_channel_snippet?.replace(
+    'MOLECULE_WORKSPACE_TOKEN="<paste from create response>"',
+    `MOLECULE_WORKSPACE_TOKEN="${info.auth_token}"`,
+  );
+  // Codex + OpenClaw snippets carry the placeholder inside the
+  // generated config block (TOML / JSON respectively). Stamp the
+  // token in so the copy-paste is one less manual edit.
+  const filledCodex = info.codex_snippet?.replace(
+    'MOLECULE_WORKSPACE_TOKEN = "<paste from create response>"',
+    `MOLECULE_WORKSPACE_TOKEN = "${info.auth_token}"`,
+  );
+  const filledOpenClaw = info.openclaw_snippet?.replace(
+    'WORKSPACE_TOKEN="<paste from create response>"',
+    `WORKSPACE_TOKEN="${info.auth_token}"`,
+  );
 
   return (
     <Dialog.Root open onOpenChange={(o) => !o && onClose()}>
@@ -135,10 +326,18 @@ export function ExternalConnectModal({ info, onClose }: Props) {
               // SDK second (full register+heartbeat+inbound); Universal
               // MCP third (any MCP-aware runtime, outbound-only); curl
               // for one-shot register; Fields for raw values.
+              // Tab order: Universal MCP first (default, runtime-
+              // agnostic primitives), then runtime-specific channel/
+              // SDK tabs, then curl + Fields. Each runtime tab only
+              // appears when the platform supplies the snippet — no
+              // dead "tab missing snippet" UX.
               const tabs: Tab[] = [];
-              if (filledChannel) tabs.push("claude");
-              tabs.push("python");
               if (filledUniversalMcp) tabs.push("mcp");
+              tabs.push("python");
+              if (filledChannel) tabs.push("claude");
+              if (filledHermes) tabs.push("hermes");
+              if (filledCodex) tabs.push("codex");
+              if (filledOpenClaw) tabs.push("openclaw");
               tabs.push("curl", "fields");
               return tabs;
             })().map((t) => (
@@ -156,6 +355,12 @@ export function ExternalConnectModal({ info, onClose }: Props) {
               >
                 {t === "claude"
                   ? "Claude Code"
+                  : t === "hermes"
+                  ? "Hermes"
+                  : t === "codex"
+                  ? "Codex"
+                  : t === "openclaw"
+                  ? "OpenClaw"
                   : t === "python"
                   ? "Python SDK"
                   : t === "mcp"
@@ -205,6 +410,33 @@ export function ExternalConnectModal({ info, onClose }: Props) {
                 onCopy={() => copy(filledUniversalMcp, "mcp")}
               />
             )}
+            {tab === "hermes" && filledHermes && (
+              <SnippetBlock
+                value={filledHermes}
+                label="Hermes channel — bridges this workspace's A2A traffic into your hermes-agent session as platform messages (push parity with Claude Code). Long-poll based; no tunnel needed."
+                copyKey="hermes"
+                copied={copiedKey === "hermes"}
+                onCopy={() => copy(filledHermes, "hermes")}
+              />
+            )}
+            {tab === "codex" && filledCodex && (
+              <SnippetBlock
+                value={filledCodex}
+                label="Codex MCP config — wires the molecule MCP server into ~/.codex/config.toml. Outbound tools today; inbound A2A push needs the Python SDK tab paired in (codex's MCP runtime doesn't route arbitrary notifications/* yet)."
+                copyKey="codex"
+                copied={copiedKey === "codex"}
+                onCopy={() => copy(filledCodex, "codex")}
+              />
+            )}
+            {tab === "openclaw" && filledOpenClaw && (
+              <SnippetBlock
+                value={filledOpenClaw}
+                label="OpenClaw MCP config — wires the molecule MCP server via openclaw mcp set + starts the gateway on loopback. Outbound tools today; inbound A2A push on an external openclaw needs the Python SDK tab paired in (a sessions.steer bridge daemon is future work)."
+                copyKey="openclaw"
+                copied={copiedKey === "openclaw"}
+                onCopy={() => copy(filledOpenClaw, "openclaw")}
+              />
+            )}
             {tab === "fields" && (
               <div className="space-y-2">
                 <Field label="workspace_id" value={info.workspace_id} onCopy={() => copy(info.workspace_id, "wsid")} copied={copiedKey === "wsid"} />
@@ -220,6 +452,7 @@ export function ExternalConnectModal({ info, onClose }: Props) {
                 <Field label="heartbeat_endpoint" value={info.heartbeat_endpoint} onCopy={() => copy(info.heartbeat_endpoint, "hb")} copied={copiedKey === "hb"} />
               </div>
             )}
+            <HelpBlock help={TAB_HELP[tab]} />
           </div>
 
           <div className="mt-5 flex justify-end gap-2">
@@ -268,6 +501,70 @@ function SnippetBlock({
   );
 }
 
+// HelpBlock — collapsible "Need help?" section under each tab's snippet.
+// Renders only the keys present in the per-tab help metadata (no empty
+// sections). Closed by default so the snippet stays the visual focus;
+// operators with a working setup never see this. Uses native <details>
+// for keyboard accessibility (Tab + Enter) without extra ARIA wiring.
+function HelpBlock({
+  help,
+}: {
+  help: (typeof TAB_HELP)[Tab] | undefined;
+}) {
+  if (!help) return null;
+  const { docsUrl, docsLabel, downloadUrl, downloadLabel, commonIssues } = help;
+  if (!docsUrl && !downloadUrl && !commonIssues?.length) return null;
+
+  return (
+    <details className="mt-3 border border-line rounded-lg bg-surface text-xs">
+      <summary className="cursor-pointer select-none px-3 py-2 text-ink-mid hover:text-ink">
+        Need help? — install link, docs, common errors
+      </summary>
+      <div className="px-3 pb-3 pt-1 space-y-2">
+        {downloadUrl && (
+          <div>
+            <span className="text-ink-soft">Where to install: </span>
+            <a
+              href={downloadUrl}
+              target="_blank"
+              rel="noopener noreferrer"
+              className="text-accent underline hover:text-accent-strong"
+            >
+              {downloadLabel || downloadUrl}
+            </a>
+          </div>
+        )}
+        {docsUrl && (
+          <div>
+            <span className="text-ink-soft">Documentation: </span>
+            <a
+              href={docsUrl}
+              target="_blank"
+              rel="noopener noreferrer"
+              className="text-accent underline hover:text-accent-strong"
+            >
+              {docsLabel || docsUrl}
+            </a>
+          </div>
+        )}
+        {commonIssues && commonIssues.length > 0 && (
+          <div>
+            <div className="text-ink-soft mb-1">Common errors:</div>
+            <ul className="space-y-1.5 pl-3">
+              {commonIssues.map((issue, i) => (
+                <li key={i}>
+                  <code className="text-warm font-mono">{issue.symptom}</code>
+                  <span className="text-ink-mid"> — {issue.check}</span>
+                </li>
+              ))}
+            </ul>
+          </div>
+        )}
+      </div>
+    </details>
+  );
+}
+
 function Field({
   label,
   value,

canvas/src/components/Legend.tsx

@@ -1,11 +1,23 @@
 "use client";
 
 import { useEffect, useState } from "react";
-import { STATUS_CONFIG } from "@/lib/design-tokens";
+import { STATUS_CONFIG, TIER_CONFIG } from "@/lib/design-tokens";
 import { useCanvasStore } from "@/store/canvas";
 
 const LEGEND_STATUSES = ["online", "provisioning", "degraded", "failed", "paused", "offline"] as const;
 
+// Tier descriptions kept in sync with CreateWorkspaceDialog.tsx (the
+// source of truth for what each tier means semantically). Colors come
+// from TIER_CONFIG so the legend swatch matches the badge actually
+// rendered on every WorkspaceNode — drift here misled users into
+// thinking the legend documented a different tier than the one shown.
+const LEGEND_TIERS: ReadonlyArray<{ tier: number; label: string }> = [
+  { tier: 1, label: "Sandboxed" },
+  { tier: 2, label: "Standard" },
+  { tier: 3, label: "Privileged" },
+  { tier: 4, label: "Full Access" },
+];
+
 // Persist the user's choice across sessions. Default is "open" so
 // first-time users still see the symbol key; once dismissed we
 // respect that until they explicitly reopen via the floating pill.
@@ -65,7 +77,7 @@ export function Legend() {
         onClick={openLegend}
         aria-label="Show legend"
         title="Show legend"
-        className={`fixed bottom-6 ${leftClass} z-30 flex items-center gap-1.5 rounded-full bg-surface-sunken/95 border border-line/50 px-3 py-1.5 text-[11px] font-semibold text-ink-mid uppercase tracking-wider shadow-xl shadow-black/30 backdrop-blur-sm hover:text-ink hover:border-line transition-[left,colors] duration-200`}
+        className={`fixed bottom-6 ${leftClass} z-30 flex items-center gap-1.5 rounded-full bg-surface-sunken/95 border border-line/50 px-3 py-1.5 text-[11px] font-semibold text-ink-mid uppercase tracking-wider shadow-xl shadow-black/30 backdrop-blur-sm hover:text-ink hover:border-line focus:outline-none focus-visible:ring-2 focus-visible:ring-accent/60 focus-visible:ring-offset-2 focus-visible:ring-offset-surface transition-[left,colors] duration-200`}
       >
         <span aria-hidden="true" className="text-[10px]">ⓘ</span>
         Legend
@@ -82,7 +94,10 @@ export function Legend() {
           onClick={closeLegend}
           aria-label="Hide legend"
           title="Hide legend"
-          className="-mt-0.5 -mr-1 px-1.5 text-[14px] leading-none text-ink-soft hover:text-ink transition-colors"
+          // 24×24 touch target (was ~10×16, well under WCAG 2.5.5 min).
+          // Negative margin keeps the visual position the same as before
+          // — only the hit area + focus ring are larger.
+          className="-mt-1.5 -mr-1.5 w-6 h-6 inline-flex items-center justify-center rounded text-[14px] leading-none text-ink-soft hover:text-ink hover:bg-surface-card/40 focus:outline-none focus-visible:ring-2 focus-visible:ring-accent/60 transition-colors"
         >
           ×
         </button>
@@ -102,9 +117,9 @@ export function Legend() {
       <div className="mb-2">
         <div className="text-[11px] text-ink-soft font-medium mb-1">Tier</div>
         <div className="flex flex-wrap gap-x-3 gap-y-1">
-          <TierItem tier={1} label="Sandboxed" color="text-sky-300 bg-sky-950/40 border-sky-700/30" />
-          <TierItem tier={2} label="Standard" color="text-violet-300 bg-violet-950/40 border-violet-700/30" />
-          <TierItem tier={3} label="Full Access" color="text-warm bg-amber-950/40 border-amber-700/30" />
+          {LEGEND_TIERS.map(({ tier, label }) => (
+            <TierItem key={tier} tier={tier} label={label} color={TIER_CONFIG[tier].border} />
+          ))}
         </div>
       </div>
 

canvas/src/components/OnboardingWizard.tsx

@@ -134,10 +134,12 @@ export function OnboardingWizard() {
       aria-label="Onboarding guide"
       className="fixed bottom-20 left-4 z-50 w-80 rounded-2xl border border-line/60 bg-surface-sunken/95 backdrop-blur-xl shadow-2xl shadow-black/40 overflow-hidden"
     >
-      {/* Progress bar */}
+      {/* Progress bar — was hardcoded from-blue-500 to-sky-400, neither
+          tone exists in warm-paper light theme. Switched to the accent
+          ramp so the gradient reads as brand color in both themes. */}
       <div className="h-1 bg-surface-card">
         <div
-          className="h-full bg-gradient-to-r from-blue-500 to-sky-400 transition-all duration-500"
+          className="h-full bg-gradient-to-r from-accent to-accent-strong transition-all duration-500"
           style={{ width: `${((currentStepIdx + 1) / STEPS.length) * 100}%` }}
         />
       </div>
@@ -155,14 +157,16 @@ export function OnboardingWizard() {
       <div className="p-4">
         {/* Step indicator */}
         <div className="flex items-center justify-between mb-2">
-          <span className="text-[9px] font-semibold uppercase tracking-widest text-sky-400/80">
+          {/* text-sky-400/80 was hardcoded; flip to text-accent so the
+              indicator stays brand-tinted in both themes. */}
+          <span className="text-[9px] font-semibold uppercase tracking-widest text-accent">
             Step {currentStepIdx + 1} of {STEPS.length}
           </span>
           <button
             type="button"
             onClick={dismiss}
             aria-label="Skip onboarding guide"
-            className="text-[10px] text-ink-mid hover:text-ink transition-colors"
+            className="text-[10px] text-ink-mid hover:text-ink transition-colors rounded-sm focus:outline-none focus-visible:ring-2 focus-visible:ring-accent/50"
           >
             Skip guide
           </button>
@@ -181,7 +185,11 @@ export function OnboardingWizard() {
           <button
             type="button"
             onClick={handleAction}
-            className="flex-1 px-3 py-1.5 bg-accent-strong/90 hover:bg-accent rounded-lg text-[11px] font-medium text-white transition-colors"
+            // Was bg-accent-strong/90 hover:bg-accent — accent is the
+            // LIGHTER variant, so this hovered lighter on white text and
+            // dropped contrast below AA. Same trap fixed in
+            // ConfirmDialog/ApprovalBanner. Hover the OTHER direction.
+            className="flex-1 px-3 py-1.5 bg-accent hover:bg-accent-strong rounded-lg text-[11px] font-medium text-white transition-colors focus:outline-none focus-visible:ring-2 focus-visible:ring-accent/60 focus-visible:ring-offset-2 focus-visible:ring-offset-surface-sunken"
           >
             {step === "welcome"
               ? "Create Workspace"
@@ -199,7 +207,10 @@ export function OnboardingWizard() {
                 if (next) setStep(next.id);
                 else dismiss();
               }}
-              className="px-3 py-1.5 bg-surface-card hover:bg-surface-card rounded-lg text-[11px] text-ink-mid transition-colors"
+              // Was hover:bg-surface-card on top of bg-surface-card —
+              // silent no-op hover. Lift to surface-elevated, matching
+              // the Cancel pattern in ConfirmDialog.
+              className="px-3 py-1.5 bg-surface-card hover:bg-surface-elevated hover:text-ink rounded-lg text-[11px] text-ink-mid transition-colors focus:outline-none focus-visible:ring-2 focus-visible:ring-accent/40 focus-visible:ring-offset-2 focus-visible:ring-offset-surface-sunken"
             >
               Next
             </button>

canvas/src/components/OrgImportPreflightModal.tsx

@@ -293,7 +293,7 @@ export function OrgImportPreflightModal({
           <button
             type="button"
             onClick={onCancel}
-            className="px-3 py-1.5 text-[11px] rounded bg-surface-card hover:bg-surface-card text-ink-mid"
+            className="px-3 py-1.5 text-[11px] rounded bg-surface-card hover:bg-surface-elevated hover:text-ink text-ink-mid transition-colors focus:outline-none focus-visible:ring-2 focus-visible:ring-accent/40 focus-visible:ring-offset-1 focus-visible:ring-offset-surface"
           >
             Cancel
           </button>
@@ -308,7 +308,7 @@ export function OrgImportPreflightModal({
               type="button"
               onClick={onProceed}
               disabled={!canProceed}
-              className="px-4 py-1.5 text-[11px] font-semibold rounded bg-accent-strong hover:bg-accent text-white disabled:bg-surface-card disabled:text-white-soft disabled:cursor-not-allowed"
+              className="px-4 py-1.5 text-[11px] font-semibold rounded bg-accent hover:bg-accent-strong text-white disabled:bg-surface-card disabled:text-white-soft disabled:cursor-not-allowed"
             >
               Import
             </button>
@@ -428,7 +428,7 @@ function StrictEnvRow({
             type="button"
             onClick={() => onSave(envKey)}
             disabled={d?.saving || !d?.value.trim()}
-            className="px-2 py-1 text-[10px] rounded bg-accent-strong hover:bg-accent text-white disabled:opacity-40 disabled:cursor-not-allowed"
+            className="px-2 py-1 text-[10px] rounded bg-accent hover:bg-accent-strong text-white disabled:opacity-40 disabled:cursor-not-allowed"
           >
             {d?.saving ? "…" : "Save"}
           </button>
@@ -520,7 +520,7 @@ function AnyOfEnvGroup({
                     type="button"
                     onClick={() => onSave(m)}
                     disabled={d?.saving || !d?.value.trim()}
-                    className="px-2 py-1 text-[10px] rounded bg-accent-strong hover:bg-accent text-white disabled:opacity-40 disabled:cursor-not-allowed"
+                    className="px-2 py-1 text-[10px] rounded bg-accent hover:bg-accent-strong text-white disabled:opacity-40 disabled:cursor-not-allowed"
                   >
                     {d?.saving ? "…" : "Save"}
                   </button>

canvas/src/components/SearchDialog.tsx

@@ -36,11 +36,6 @@ export function SearchDialog() {
     }
   }, [open]);
 
-  // Reset focused index when query changes
-  useEffect(() => {
-    setFocusedIndex(-1);
-  }, [query]);
-
   const filtered = nodes.filter((n) => {
     if (!query) return true;
     const q = query.toLowerCase();
@@ -51,6 +46,18 @@ export function SearchDialog() {
     );
   });
 
+  // Auto-highlight the first match while the user is typing, so Enter
+  // selects something instead of being a no-op. With an empty query we
+  // keep -1 so opening the dialog (which shows ALL workspaces) doesn't
+  // visually pin one row arbitrarily — only commit a highlight once the
+  // user has narrowed the list.
+  useEffect(() => {
+    setFocusedIndex(query && filtered.length > 0 ? 0 : -1);
+    // Re-running on filtered.length keeps the highlight pinned to the
+    // first row while the result set shrinks/grows; the effect handler
+    // above already short-circuits to -1 when results disappear.
+  }, [query, filtered.length]);
+
   const handleSelect = useCallback(
     (nodeId: string) => {
       selectNode(nodeId);
@@ -113,7 +120,7 @@ export function SearchDialog() {
             onChange={(e) => setQuery(e.target.value)}
             onKeyDown={handleInputKeyDown}
             placeholder="Search workspaces..."
-            className="flex-1 bg-transparent text-sm text-ink placeholder-zinc-400 focus-visible:outline-none focus-visible:ring-2 focus-visible:ring-accent focus:outline-none rounded"
+            className="flex-1 bg-transparent text-sm text-ink placeholder-ink-soft focus:outline-none focus-visible:outline-none focus-visible:ring-2 focus-visible:ring-accent rounded"
           />
           <kbd className="text-[9px] text-ink-mid bg-surface-card/60 px-1.5 py-0.5 rounded border border-line/40">ESC</kbd>
         </div>

canvas/src/components/SidePanel.tsx

@@ -202,7 +202,7 @@ export function SidePanel() {
       {/* Tabs — relative wrapper lets the fade gradient position against the scroll container */}
       <div className="relative border-b border-line/40">
         {/* Right-edge fade: signals more tabs are hidden off-screen when the bar overflows */}
-        <div className="pointer-events-none absolute inset-y-0 right-0 w-8 bg-gradient-to-l from-zinc-950 to-transparent z-10" aria-hidden="true" />
+        <div className="pointer-events-none absolute inset-y-0 right-0 w-8 bg-gradient-to-l from-surface to-transparent z-10" aria-hidden="true" />
       <div
         role="tablist"
         aria-label="Workspace panel tabs"
@@ -232,8 +232,8 @@ export function SidePanel() {
             onClick={() => setPanelTab(tab.id)}
             className={`shrink-0 px-3 py-2.5 text-[10px] font-medium tracking-wide transition-all rounded-t-lg mx-0.5 focus:outline-none focus-visible:ring-2 focus-visible:ring-accent/70 ${
               panelTab === tab.id
-                ? "text-ink bg-surface-card/40 border-b-2 border-accent"
-                : "text-ink-soft hover:text-ink hover:bg-surface-card/40"
+                ? "text-ink bg-surface-card border-b-2 border-accent"
+                : "text-ink-mid hover:text-ink hover:bg-surface-card/60"
             }`}
           >
             <span className="mr-1 opacity-50" aria-hidden="true">{tab.icon}</span>

canvas/src/components/TermsGate.tsx

@@ -1,6 +1,6 @@
 "use client";
 
-import { useEffect, useState } from "react";
+import { useEffect, useRef, useState } from "react";
 import { PLATFORM_URL } from "@/lib/api";
 
 // TermsGate blocks the page it wraps until the user has accepted the
@@ -73,39 +73,72 @@ export function TermsGate({ children }: { children: React.ReactNode }) {
     }
   };
 
+  // Move focus to the "I agree" button when the modal opens (WCAG 2.4.3).
+  // The dialog is a hard gate — no Esc dismiss — so we don't need a focus
+  // trap loop, just a one-shot focus move into the dialog.
+  const agreeButtonRef = useRef<HTMLButtonElement>(null);
+  useEffect(() => {
+    if (status !== "pending") return;
+    const raf = requestAnimationFrame(() => agreeButtonRef.current?.focus());
+    return () => cancelAnimationFrame(raf);
+  }, [status]);
+
   return (
     <>
       {children}
       {status === "pending" && (
-        <div aria-hidden="true" className="fixed inset-0 z-50 flex items-center justify-center bg-surface/80 backdrop-blur-sm">
+        // Backdrop is decorative — does NOT carry aria-hidden anymore.
+        // The earlier version put aria-hidden="true" on this wrapper,
+        // which hid the dialog AND its descendants from screen readers,
+        // making the entire terms-acceptance flow invisible to AT users.
+        // Backdrop click intentionally does nothing — this is a hard
+        // gate.
+        <div className="fixed inset-0 z-50 flex items-center justify-center bg-surface/80 backdrop-blur-sm">
           <div
             role="dialog"
             aria-modal="true"
             aria-labelledby="terms-dialog-title"
+            aria-describedby="terms-dialog-body"
             className="mx-4 max-w-lg rounded-lg border border-line bg-surface-sunken p-6 shadow-xl"
           >
             <h2 id="terms-dialog-title" className="text-lg font-semibold text-ink">Terms &amp; conditions</h2>
-            <p className="mt-3 text-sm text-ink-mid">
-              Before you create an organization, please review our{" "}
-              <a href="/legal/terms" className="text-sky-400 underline" target="_blank" rel="noreferrer">
-                Terms of Service
-              </a>{" "}
-              and{" "}
-              <a href="/legal/privacy" className="text-sky-400 underline" target="_blank" rel="noreferrer">
-                Privacy Policy
-              </a>
-              . Click agree to continue.
-            </p>
-            <p className="mt-3 text-xs text-ink-soft">
-              By agreeing you acknowledge that workspace data is stored in AWS us-east-2 (Ohio, United States).
-            </p>
+            <div id="terms-dialog-body">
+              <p className="mt-3 text-sm text-ink-mid">
+                Before you create an organization, please review our{" "}
+                <a
+                  href="/legal/terms"
+                  className="text-accent underline underline-offset-2 hover:text-accent-strong focus:outline-none focus-visible:ring-2 focus-visible:ring-accent/60 rounded-sm"
+                  target="_blank"
+                  rel="noreferrer"
+                >
+                  Terms of Service
+                </a>{" "}
+                and{" "}
+                <a
+                  href="/legal/privacy"
+                  className="text-accent underline underline-offset-2 hover:text-accent-strong focus:outline-none focus-visible:ring-2 focus-visible:ring-accent/60 rounded-sm"
+                  target="_blank"
+                  rel="noreferrer"
+                >
+                  Privacy Policy
+                </a>
+                . Click agree to continue.
+              </p>
+              <p className="mt-3 text-xs text-ink-soft">
+                By agreeing you acknowledge that workspace data is stored in AWS us-east-2 (Ohio, United States).
+              </p>
+            </div>
             {error && <p role="alert" className="mt-3 text-sm text-bad">{error}</p>}
             <div className="mt-5 flex justify-end gap-2">
               <button
                 type="button"
+                ref={agreeButtonRef}
                 onClick={accept}
                 disabled={submitting}
-                className="rounded bg-emerald-600 px-4 py-2 text-sm font-medium text-white hover:bg-emerald-500 disabled:opacity-50"
+                // Hover goes DARKER, not lighter — emerald-500 on white
+                // text drops contrast below AA vs emerald-700. Same trap
+                // I fixed in ApprovalBanner + ConfirmDialog.
+                className="rounded bg-emerald-600 hover:bg-emerald-700 px-4 py-2 text-sm font-medium text-white disabled:opacity-50 transition-colors focus:outline-none focus-visible:ring-2 focus-visible:ring-emerald-400/70 focus-visible:ring-offset-2 focus-visible:ring-offset-surface-sunken"
               >
                 {submitting ? "Saving…" : "I agree"}
               </button>

canvas/src/components/Toaster.tsx

@@ -38,6 +38,18 @@ export function Toaster() {
     };
   }, []);
 
+  // Esc dismisses the newest toast — keyboard parity with the × button.
+  // Errors never auto-expire, so without this a keyboard-only user has to
+  // tab through the entire app to reach the dismiss button on a stuck error.
+  useEffect(() => {
+    const onKey = (e: KeyboardEvent) => {
+      if (e.key !== "Escape") return;
+      setToasts((prev) => (prev.length === 0 ? prev : prev.slice(0, -1)));
+    };
+    window.addEventListener("keydown", onKey);
+    return () => window.removeEventListener("keydown", onKey);
+  }, []);
+
   const toastCls = (type: Toast["type"]) =>
     `flex items-center gap-2 pl-4 pr-2 py-2.5 rounded-xl shadow-2xl shadow-black/40 text-sm backdrop-blur-md animate-in slide-in-from-bottom duration-200 ${
       type === "success"
@@ -47,6 +59,17 @@ export function Toaster() {
         : "bg-surface-sunken/90 border border-line/40 text-ink"
     }`;
 
+  // Success/error toasts are intentionally dark in both themes (high-vis).
+  // Info uses the semantic surface that flips with theme — so the dismiss
+  // button needs a tint that stays visible on a light bg in light mode.
+  const dismissCls = (type: Toast["type"]) => {
+    const base =
+      "ml-1 w-7 h-7 inline-flex items-center justify-center text-base leading-none rounded transition-colors opacity-70 hover:opacity-100 focus-visible:opacity-100 focus:outline-none focus-visible:ring-2 shrink-0";
+    return type === "info"
+      ? `${base} hover:bg-ink/10 focus-visible:ring-accent/60`
+      : `${base} hover:bg-white/15 focus-visible:ring-white/70`;
+  };
+
   const pos =
     "fixed bottom-16 left-1/2 -translate-x-1/2 z-[80] flex flex-col gap-2 items-center";
 
@@ -66,7 +89,7 @@ export function Toaster() {
                 type="button"
                 onClick={() => dismiss(toast.id)}
                 aria-label="Dismiss notification"
-                className="ml-1 p-1 rounded hover:bg-surface-card/50 transition-colors opacity-70 hover:opacity-100 shrink-0"
+                className={dismissCls(toast.type)}
               >
                 ×
               </button>
@@ -94,7 +117,7 @@ export function Toaster() {
                 type="button"
                 onClick={() => dismiss(toast.id)}
                 aria-label="Dismiss notification"
-                className="ml-1 p-1 rounded hover:bg-surface-card/50 transition-colors opacity-70 hover:opacity-100 shrink-0"
+                className={dismissCls(toast.type)}
               >
                 ×
               </button>

canvas/src/components/Toolbar.tsx

@@ -154,10 +154,10 @@ export function Toolbar() {
         {counts.failed > 0 && (
           <StatusPill color={statusDotClass("failed")} count={counts.failed} label="failed" />
         )}
-        <span className="text-ink-soft" aria-hidden="true">·</span>
-        <span className="text-[10px] text-ink-soft whitespace-nowrap">
+        <span className="text-ink-mid" aria-hidden="true">·</span>
+        <span className="text-[10px] text-ink-mid whitespace-nowrap">
           {counts.roots} workspace{counts.roots !== 1 ? "s" : ""}
-          {counts.children > 0 && <span className="text-ink-soft"> + {counts.children} sub</span>}
+          {counts.children > 0 && <span className="text-ink-mid"> + {counts.children} sub</span>}
         </span>
       </div>
 
@@ -172,7 +172,7 @@ export function Toolbar() {
           type="button"
           onClick={stopAll}
           disabled={stopping}
-          className="flex items-center gap-1.5 px-2.5 py-1 bg-red-950/50 hover:bg-red-900/60 border border-red-800/40 rounded-lg transition-colors disabled:opacity-50"
+          className="flex items-center gap-1.5 px-2.5 py-1 bg-bad/10 hover:bg-bad/20 border border-bad/40 rounded-lg transition-colors disabled:opacity-50 focus:outline-none focus-visible:ring-2 focus-visible:ring-bad/40"
           title={`Stop all running tasks (${counts.activeTasks} active)`}
           aria-label={stopping ? "Stopping all running tasks" : `Stop all running tasks (${counts.activeTasks} active)`}
         >
@@ -191,7 +191,7 @@ export function Toolbar() {
           type="button"
           onClick={() => setRestartConfirmOpen(true)}
           disabled={restartingAll}
-          className="flex items-center gap-1.5 px-2.5 py-1 bg-amber-950/40 hover:bg-amber-900/50 border border-amber-800/40 rounded-lg transition-colors disabled:opacity-50"
+          className="flex items-center gap-1.5 px-2.5 py-1 bg-warm/10 hover:bg-warm/20 border border-warm/40 rounded-lg transition-colors disabled:opacity-50 focus:outline-none focus-visible:ring-2 focus-visible:ring-warm/40"
           title={`Restart ${needsRestartNodes.length} workspace${needsRestartNodes.length === 1 ? "" : "s"} that need to pick up config or secret changes`}
           aria-label={restartingAll ? "Restarting workspaces" : `Restart ${needsRestartNodes.length} workspace${needsRestartNodes.length === 1 ? "" : "s"} pending config or secret changes`}
         >
@@ -216,10 +216,10 @@ export function Toolbar() {
         aria-pressed={showA2AEdges}
         aria-label={showA2AEdges ? "Hide A2A edges" : "Show A2A edges"}
         title={showA2AEdges ? "Hide A2A delegation edges" : "Show A2A delegation edges (last 60 min)"}
-        className={`flex items-center justify-center w-7 h-7 border rounded-lg transition-colors ${
+        className={`flex items-center justify-center w-7 h-7 border rounded-lg transition-colors focus:outline-none focus-visible:ring-2 focus-visible:ring-accent/40 ${
           showA2AEdges
-            ? "bg-blue-950/50 hover:bg-blue-900/50 border-blue-800/40 text-accent"
-            : "bg-surface-card/50 hover:bg-surface-card/50 border-line/40 text-ink-soft hover:text-ink-mid"
+            ? "bg-accent/15 hover:bg-accent/25 border-accent/50 text-accent"
+            : "bg-surface-card hover:bg-surface-card/70 border-line text-ink-mid hover:text-ink"
         }`}
       >
         {/* Mesh / network icon */}
@@ -255,7 +255,7 @@ export function Toolbar() {
         }}
         aria-label="Open audit trail for selected workspace"
         title="Audit — view ledger for the selected workspace"
-        className="flex items-center justify-center w-7 h-7 bg-surface-card/50 hover:bg-surface-card/50 border border-line/40 rounded-lg transition-colors text-ink-soft hover:text-ink-mid"
+        className="flex items-center justify-center w-7 h-7 bg-surface-card hover:bg-surface-card/70 border border-line rounded-lg transition-colors text-ink-mid hover:text-ink focus:outline-none focus-visible:ring-2 focus-visible:ring-accent/40"
       >
         {/* Scroll / ledger icon */}
         <svg
@@ -277,7 +277,7 @@ export function Toolbar() {
         onClick={() => useCanvasStore.getState().setSearchOpen(true)}
         aria-label="Search workspaces"
         title="Search (⌘K)"
-        className="flex items-center justify-center w-7 h-7 bg-surface-card/50 hover:bg-surface-card/50 border border-line/40 rounded-lg transition-colors text-ink-soft hover:text-ink-mid"
+        className="flex items-center justify-center w-7 h-7 bg-surface-card hover:bg-surface-card/70 border border-line rounded-lg transition-colors text-ink-mid hover:text-ink focus:outline-none focus-visible:ring-2 focus-visible:ring-accent/40"
       >
         <svg width="14" height="14" viewBox="0 0 16 16" fill="none" aria-hidden="true">
           <circle cx="7" cy="7" r="5" stroke="currentColor" strokeWidth="1.5" />
@@ -290,7 +290,7 @@ export function Toolbar() {
         <button
           type="button"
           onClick={() => setHelpOpen((open) => !open)}
-          className="flex items-center justify-center w-7 h-7 bg-surface-card/50 hover:bg-surface-card/50 border border-line/40 rounded-lg transition-colors text-ink-soft hover:text-ink-mid"
+          className="flex items-center justify-center w-7 h-7 bg-surface-card hover:bg-surface-card/70 border border-line rounded-lg transition-colors text-ink-mid hover:text-ink focus:outline-none focus-visible:ring-2 focus-visible:ring-accent/40"
           aria-expanded={helpOpen}
           aria-label="Open quick help"
           title="Help — shortcuts & quick start"
@@ -308,7 +308,7 @@ export function Toolbar() {
               <button
                 type="button"
                 onClick={() => setHelpOpen(false)}
-                className="text-[10px] text-ink-soft hover:text-ink-mid transition-colors"
+                className="text-[10px] text-ink-mid hover:text-ink transition-colors focus:outline-none focus-visible:underline"
               >
                 Close
               </button>
@@ -358,7 +358,7 @@ function WsStatusPill({ status }: { status: "connected" | "connecting" | "discon
     return (
       <div className="flex items-center gap-1.5" title="Real-time updates: connected" aria-label="Real-time updates: connected">
         <div className={`w-1.5 h-1.5 rounded-full ${statusDotClass("online")}`} aria-hidden="true" />
-        <span className="text-[10px] text-ink-soft" aria-hidden="true">Live</span>
+        <span className="text-[10px] text-ink-mid" aria-hidden="true">Live</span>
       </div>
     );
   }
@@ -366,14 +366,14 @@ function WsStatusPill({ status }: { status: "connected" | "connecting" | "discon
     return (
       <div className="flex items-center gap-1.5" title="Real-time updates: reconnecting…" aria-label="Real-time updates: reconnecting">
         <div className="w-1.5 h-1.5 rounded-full bg-amber-400 motion-safe:animate-pulse" aria-hidden="true" />
-        <span className="text-[10px] text-ink-soft" aria-hidden="true">Reconnecting</span>
+        <span className="text-[10px] text-warm" aria-hidden="true">Reconnecting</span>
       </div>
     );
   }
   return (
     <div className="flex items-center gap-1.5" title="Real-time updates: disconnected" aria-label="Real-time updates: disconnected">
       <div className={`w-1.5 h-1.5 rounded-full ${statusDotClass("failed")}`} aria-hidden="true" />
-      <span className="text-[10px] text-ink-soft" aria-hidden="true">Offline</span>
+      <span className="text-[10px] text-bad" aria-hidden="true">Offline</span>
     </div>
   );
 }
@@ -384,7 +384,7 @@ function HelpRow({ shortcut, text }: { shortcut: string; text: string }) {
       <span className="shrink-0 rounded-md border border-line/60 bg-surface/70 px-2 py-0.5 text-[9px] font-medium uppercase tracking-[0.18em] text-ink-mid">
         {shortcut}
       </span>
-      <p className="text-[11px] leading-relaxed text-ink-soft">{text}</p>
+      <p className="text-[11px] leading-relaxed text-ink-mid">{text}</p>
     </div>
   );
 }

canvas/src/components/Tooltip.tsx

@@ -22,6 +22,24 @@ export function Tooltip({ text, children }: Props) {
 
   useEffect(() => () => clearTimeout(timerRef.current), []);
 
+  // WCAG 1.4.13 (Content on Hover or Focus) — Dismissible: a mechanism
+  // is available to dismiss the additional content WITHOUT moving
+  // pointer hover or keyboard focus. Esc dismisses while the trigger
+  // stays focused/hovered, so a screen-magnifier user can read what
+  // the tooltip was covering without losing their place.
+  useEffect(() => {
+    if (!show) return;
+    const onKey = (e: KeyboardEvent) => {
+      if (e.key === "Escape") {
+        e.stopPropagation();
+        clearTimeout(timerRef.current);
+        setShow(false);
+      }
+    };
+    window.addEventListener("keydown", onKey, true);
+    return () => window.removeEventListener("keydown", onKey, true);
+  }, [show]);
+
   const enter = useCallback(() => {
     timerRef.current = setTimeout(() => {
       if (triggerRef.current) {

canvas/src/components/WorkspaceNode.tsx

@@ -3,6 +3,7 @@
 import { useCallback, useMemo } from "react";
 import { Handle, NodeResizer, Position, type NodeProps, type Node } from "@xyflow/react";
 import { useCanvasStore, type WorkspaceNodeData } from "@/store/canvas";
+import { getConfigurationError, getConfigurationStatus } from "@/store/canvas-topology";
 import { showToast } from "@/components/Toaster";
 import { Tooltip } from "@/components/Tooltip";
 import { STATUS_CONFIG, TIER_CONFIG } from "@/lib/design-tokens";
@@ -35,8 +36,28 @@ function EjectIcon(props: React.SVGProps<SVGSVGElement>) {
 }
 
 export function WorkspaceNode({ id, data }: NodeProps<Node<WorkspaceNodeData>>) {
-  const statusCfg = STATUS_CONFIG[data.status] || STATUS_CONFIG.offline;
-  const tierCfg = TIER_CONFIG[data.tier] || { label: `T${data.tier}`, color: "text-ink-soft bg-surface-card" };
+  // Configuration-status overlay (PR #2756 / #467 chain). When the
+  // workspace is reachable but adapter.setup() failed (typically a
+  // missing/rotated LLM credential), the agent_card carries
+  // configuration_status: "not_configured". Surface this as a distinct
+  // tile state so the operator sees a useful error instead of an
+  // ambiguous "online but silent" workspace.
+  //
+  // The override only applies when the underlying status is "online" —
+  // a workspace that's actually offline / failed / provisioning gets
+  // its own treatment. "online + not_configured" is the gap PR #2756
+  // introduced; everything else was already covered.
+  const isMisconfigured =
+    data.status === "online" &&
+    getConfigurationStatus(data.agentCard) === "not_configured";
+  const configurationError = getConfigurationError(data.agentCard);
+  const effectiveStatus = isMisconfigured ? "not_configured" : data.status;
+  const statusCfg = STATUS_CONFIG[effectiveStatus] || STATUS_CONFIG.offline;
+  const tierCfg = TIER_CONFIG[data.tier] || { label: `T${data.tier}`, color: "text-ink-mid bg-surface-card border border-line" };
+  const tooltipExtra = isMisconfigured && configurationError
+    ? `Agent not configured: ${configurationError}`
+    : null;
+  void tooltipExtra; // wired in via aria-label below; reserved here for future tooltip surface.
   // Org-deploy context — four derived flags off one store subscription.
   // Drives the shimmer while provisioning, the dimmed/non-draggable
   // treatment on locked descendants, and the Cancel pill on the root.
@@ -75,7 +96,12 @@ export function WorkspaceNode({ id, data }: NodeProps<Node<WorkspaceNodeData>>)
     <div
       role="button"
       tabIndex={0}
-      aria-label={`${data.name} workspace — ${data.status}`}
+      aria-label={
+        isMisconfigured && configurationError
+          ? `${data.name} workspace — agent not configured: ${configurationError}`
+          : `${data.name} workspace — ${data.status}`
+      }
+      title={isMisconfigured && configurationError ? `Agent not configured: ${configurationError}` : undefined}
       aria-pressed={isSelected}
       onClick={(e) => {
         e.stopPropagation();
@@ -179,7 +205,7 @@ export function WorkspaceNode({ id, data }: NodeProps<Node<WorkspaceNodeData>>)
           </div>
           <div className="flex items-center gap-1.5 shrink-0">
             {hasChildren && (
-              <span className="text-[10px] font-mono text-violet-300 bg-violet-900/40 border border-violet-700/30 px-1.5 py-0.5 rounded-md">
+              <span className="text-[10px] font-mono text-accent bg-accent/15 border border-accent/40 px-1.5 py-0.5 rounded-md">
                 {descendantCount} sub
               </span>
             )}
@@ -207,13 +233,13 @@ export function WorkspaceNode({ id, data }: NodeProps<Node<WorkspaceNodeData>>)
             <div className="mb-1 flex items-center gap-1">
               {runtime === "external" ? (
                 <span
-                  className="text-[7px] font-mono px-1.5 py-0.5 rounded-md text-violet-200 bg-violet-900/50 border border-violet-500/40"
+                  className="text-[7px] font-mono px-1.5 py-0.5 rounded-md text-white bg-violet-600 border border-violet-700"
                   title="Phase 30 remote agent — runs outside this platform's Docker network. Lifecycle managed via heartbeat-based polling, not Docker exec."
                 >
                   ★ REMOTE
                 </span>
               ) : (
-                <span className="text-[7px] font-mono px-1.5 py-0.5 rounded-md text-ink-mid bg-surface-card/60 border border-line/30">
+                <span className="text-[7px] font-mono px-1.5 py-0.5 rounded-md text-ink-mid bg-surface-card border border-line">
                   {runtime}
                 </span>
               )}
@@ -237,15 +263,15 @@ export function WorkspaceNode({ id, data }: NodeProps<Node<WorkspaceNodeData>>)
                 key={skill}
                 className={`text-[10px] px-1.5 py-0.5 rounded-md border ${
                   isOnline
-                    ? "text-good/80 bg-emerald-950/30 border-emerald-800/30"
-                    : "text-ink-mid bg-surface-card/60 border-line/40"
+                    ? "text-good bg-good/15 border-good/40"
+                    : "text-ink-mid bg-surface-card border-line"
                 }`}
               >
                 {skill}
               </span>
             ))}
             {skills.length > 4 && (
-              <span className="text-[10px] text-ink-soft self-center">
+              <span className="text-[10px] text-ink-mid self-center">
                 +{skills.length - 4}
               </span>
             )}
@@ -274,21 +300,22 @@ export function WorkspaceNode({ id, data }: NodeProps<Node<WorkspaceNodeData>>)
               e.stopPropagation();
               useCanvasStore.getState().restartWorkspace(id).catch(() => showToast("Restart failed", "error"));
             }}
-            className="flex items-center gap-1.5 mt-1 w-full bg-sky-950/30 px-2 py-1 rounded-md border border-sky-800/30 hover:bg-sky-900/40 transition-colors text-left focus-visible:ring-2 focus-visible:ring-accent/70 focus-visible:outline-none"
+            className="flex items-center gap-1.5 mt-1 w-full bg-accent/10 px-2 py-1 rounded-md border border-accent/40 hover:bg-accent/20 transition-colors text-left focus-visible:ring-2 focus-visible:ring-accent/70 focus-visible:outline-none"
           >
-            <span className="text-[10px]">↻</span>
-            <span className="text-[10px] text-sky-300/80">Restart to apply changes</span>
+            <span className="text-[10px] text-accent">↻</span>
+            <span className="text-[10px] text-accent">Restart to apply changes</span>
           </button>
         )}
 
         {/* Bottom row: status / active tasks */}
         <div className="flex items-center justify-between mt-0.5">
-          {data.status !== "online" ? (
+          {effectiveStatus !== "online" ? (
             <div className={`text-[10px] uppercase tracking-widest font-medium ${
-              data.status === "failed" ? "text-bad" :
-              data.status === "degraded" ? "text-warm" :
-              data.status === "provisioning" ? "text-sky-400" :
-              "text-ink-soft"
+              effectiveStatus === "failed" ? "text-bad" :
+              effectiveStatus === "degraded" ? "text-warm" :
+              effectiveStatus === "not_configured" ? "text-warm" :
+              effectiveStatus === "provisioning" ? "text-accent" :
+              "text-ink-mid"
             }`}>
               {statusCfg.label}
             </div>
@@ -296,8 +323,8 @@ export function WorkspaceNode({ id, data }: NodeProps<Node<WorkspaceNodeData>>)
 
           {data.activeTasks > 0 && (
             <div className="flex items-center gap-1">
-              <div className="w-1 h-1 rounded-full bg-amber-400 motion-safe:animate-pulse" />
-              <span className="text-[10px] text-warm/80 tabular-nums">
+              <div className="w-1 h-1 rounded-full bg-warm motion-safe:animate-pulse" />
+              <span className="text-[10px] text-warm tabular-nums">
                 {data.activeTasks} task{data.activeTasks > 1 ? "s" : ""}
               </span>
             </div>
@@ -307,12 +334,25 @@ export function WorkspaceNode({ id, data }: NodeProps<Node<WorkspaceNodeData>>)
         {/* Degraded error preview */}
         {data.status === "degraded" && data.lastSampleError && (
           <div
-            className="text-[10px] text-warm/60 truncate mt-1 bg-amber-950/20 px-1.5 py-0.5 rounded border border-amber-800/20"
+            className="text-[10px] text-warm truncate mt-1 bg-warm/10 px-1.5 py-0.5 rounded border border-warm/40"
             title={data.lastSampleError}
           >
             {data.lastSampleError}
           </div>
         )}
+
+        {/* Configuration error preview — same visual as the degraded
+         *  error preview but keyed off the agent_card's configuration_status.
+         *  Tells the operator which env var is missing so they can fix it
+         *  without having to dig into the workspace logs. */}
+        {isMisconfigured && configurationError && (
+          <div
+            className="text-[10px] text-warm truncate mt-1 bg-warm/10 px-1.5 py-0.5 rounded border border-warm/40"
+            title={configurationError}
+          >
+            {configurationError}
+          </div>
+        )}
       </div>
 
       <Handle
@@ -357,7 +397,7 @@ function TeamMemberChip({
 }) {
   const { data } = node;
   const statusCfg = STATUS_CONFIG[data.status] || STATUS_CONFIG.offline;
-  const tierCfg = TIER_CONFIG[data.tier] || { label: `T${data.tier}`, color: "text-ink-soft bg-surface-card" };
+  const tierCfg = TIER_CONFIG[data.tier] || { label: `T${data.tier}`, color: "text-ink-mid bg-surface-card border border-line" };
   const isOnline = data.status === "online";
   const skills = getSkillNames(data.agentCard);
 
@@ -408,7 +448,7 @@ function TeamMemberChip({
           </div>
           <div className="flex items-center gap-1 shrink-0">
             {hasSubChildren && (
-              <span className="text-[7px] font-mono text-violet-300 bg-violet-900/40 border border-violet-700/30 px-1 py-0.5 rounded">
+              <span className="text-[7px] font-mono text-accent bg-accent/15 border border-accent/40 px-1 py-0.5 rounded">
                 {descendantCount}
               </span>
             )}
@@ -423,7 +463,7 @@ function TeamMemberChip({
                 e.stopPropagation();
                 onExtract(node.id);
               }}
-              className="opacity-0 group-hover/child:opacity-100 text-ink-soft hover:text-sky-400 transition-all focus-visible:ring-2 focus-visible:ring-accent/70 focus-visible:outline-none rounded"
+              className="opacity-0 group-hover/child:opacity-100 text-ink-mid hover:text-accent transition-all focus-visible:ring-2 focus-visible:ring-accent/70 focus-visible:outline-none rounded"
             >
               <EjectIcon aria-hidden="true" />
             </button>
@@ -432,7 +472,7 @@ function TeamMemberChip({
 
         {/* Role */}
         {data.role && (
-          <div className="text-[10px] text-ink-soft mb-1 leading-tight truncate">{data.role}</div>
+          <div className="text-[10px] text-ink-mid mb-1 leading-tight truncate">{data.role}</div>
         )}
 
         {/* Skills */}
@@ -443,8 +483,8 @@ function TeamMemberChip({
                 key={skill}
                 className={`text-[10px] px-1 py-0.5 rounded border ${
                   isOnline
-                    ? "text-good/70 bg-emerald-950/20 border-emerald-800/20"
-                    : "text-ink-soft bg-surface-card/40 border-line/30"
+                    ? "text-good bg-good/15 border-good/40"
+                    : "text-ink-mid bg-surface-card border-line"
                 }`}
               >
                 {skill}
@@ -462,8 +502,8 @@ function TeamMemberChip({
             <span className={`text-[10px] uppercase tracking-widest font-medium ${
               data.status === "failed" ? "text-bad" :
               data.status === "degraded" ? "text-warm" :
-              data.status === "provisioning" ? "text-sky-400" :
-              "text-ink-soft"
+              data.status === "provisioning" ? "text-accent" :
+              "text-ink-mid"
             }`}>
               {statusCfg.label}
             </span>

canvas/src/components/__tests__/A2ATopologyOverlay.test.tsx

@@ -296,4 +296,75 @@ describe("A2ATopologyOverlay component", () => {
     // setA2AEdges should still be called with an empty array
     expect(mockStoreState.setA2AEdges).toHaveBeenCalled();
   });
+
+  // Regression for the 2026-05-04 render-loop incident:
+  // tenant heartbeats / status flips / peer-discovery writes mutated
+  // canvas store .nodes ~5x/sec. Previously visibleIds was useMemo'd on
+  // [nodes] so the array reference recreated on every store mutation,
+  // causing fetchAndUpdate to recreate, the useEffect to re-fire, and
+  // the 60-second polling fan-out to fire on EVERY store update. With
+  // 5 visible workspaces and 5 store updates/sec, the canvas hammered
+  // /workspaces/<id>/activity?type=delegation 25×/sec until edge rate
+  // -limit returned 429 (per browser console captured by user).
+  //
+  // Fix: select a stable string key (sorted CSV of IDs) from Zustand
+  // so the selector's shallow-equal short-circuit prevents re-renders
+  // when the actual ID set hasn't changed.
+  //
+  // This test verifies the fetch fires ONCE on mount + only re-fires
+  // when the visible ID set actually changes, NOT on every nodes[]
+  // reference change.
+  it("does not re-fetch when nodes[] reference changes but visible IDs are the same", async () => {
+    // eslint-disable-next-line @typescript-eslint/no-explicit-any
+    mockGet.mockResolvedValue([] as any);
+    const { rerender } = render(<A2ATopologyOverlay />);
+    await act(async () => { await Promise.resolve(); await Promise.resolve(); });
+
+    const callsAfterMount = mockGet.mock.calls.length;
+    // Sanity: 2 visible nodes (ws-a, ws-b) → 2 fan-out requests on mount
+    expect(callsAfterMount).toBe(2);
+
+    // Simulate a store mutation that changes the nodes array reference
+    // (e.g. status flip on a node) WITHOUT changing the set of visible
+    // IDs. Pre-fix: this triggered a re-fetch storm. Post-fix: the
+    // sorted-CSV selector returns the same key, Zustand's shallow-equal
+    // short-circuits, useMemo keeps the same visibleIds, fetchAndUpdate
+    // keeps the same identity, useEffect does NOT re-fire.
+    mockStoreState.nodes = [
+      { id: "ws-a", hidden: false, data: { newStatus: "online" } },  // mutated
+      { id: "ws-b", hidden: false, data: {} },
+      { id: "ws-hidden", hidden: true, data: {} },
+    ];
+    rerender(<A2ATopologyOverlay />);
+    await act(async () => { await Promise.resolve(); await Promise.resolve(); });
+
+    // No additional fetches should have fired.
+    expect(mockGet.mock.calls.length).toBe(callsAfterMount);
+  });
+
+  it("re-fetches when the visible ID set actually changes", async () => {
+    // eslint-disable-next-line @typescript-eslint/no-explicit-any
+    mockGet.mockResolvedValue([] as any);
+    const { rerender } = render(<A2ATopologyOverlay />);
+    await act(async () => { await Promise.resolve(); await Promise.resolve(); });
+
+    const callsAfterMount = mockGet.mock.calls.length;
+    expect(callsAfterMount).toBe(2);
+
+    // Add a new visible workspace — the visible-ID-set actually changed.
+    mockStoreState.nodes = [
+      { id: "ws-a", hidden: false, data: {} },
+      { id: "ws-b", hidden: false, data: {} },
+      { id: "ws-c", hidden: false, data: {} }, // NEW
+      { id: "ws-hidden", hidden: true, data: {} },
+    ];
+    rerender(<A2ATopologyOverlay />);
+    await act(async () => { await Promise.resolve(); await Promise.resolve(); });
+
+    // Should have fetched the additional workspace + the existing two
+    // (the effect re-fires once with the new ID set). Total: 2 + 3 = 5.
+    expect(mockGet.mock.calls.length).toBe(callsAfterMount + 3);
+    const allPaths = mockGet.mock.calls.map(([p]) => p as string);
+    expect(allPaths.some((p) => p.includes("ws-c"))).toBe(true);
+  });
 });

canvas/src/components/__tests__/BatchActionBar.test.tsx

@@ -130,6 +130,26 @@ describe("BatchActionBar", () => {
     const toolbar = screen.getByRole("toolbar");
     expect(toolbar.getAttribute("aria-label")).toBe("Batch workspace actions");
   });
+
+  it("Esc clears the selection — matches the deselect button title", () => {
+    // The deselect button has been promising "Clear selection (Escape)"
+    // since the bar shipped, but no handler was wired. This pins the
+    // contract.
+    mockSelectedNodeIds = new Set(["ws-1", "ws-2"]);
+    render(<BatchActionBar />);
+    fireEvent.keyDown(window, { key: "Escape" });
+    expect(mockClearSelection).toHaveBeenCalled();
+  });
+
+  it("Esc is a no-op when nothing is selected", () => {
+    mockSelectedNodeIds = new Set<string>();
+    render(<BatchActionBar />);
+    fireEvent.keyDown(window, { key: "Escape" });
+    // The early-return at count===0 prevents the bar from mounting at all,
+    // so the keydown listener never registers. clearSelection must NOT be
+    // called.
+    expect(mockClearSelection).not.toHaveBeenCalled();
+  });
 });
 
 /**

canvas/src/components/__tests__/CommunicationOverlay.test.tsx

@@ -0,0 +1,178 @@
+// @vitest-environment jsdom
+/**
+ * CommunicationOverlay tests — pin the rate-limit fix shipped 2026-05-04.
+ *
+ * The overlay polls /workspaces/:id/activity?limit=5 for each online
+ * workspace. Pre-fix it (a) polled regardless of visibility and (b)
+ * fanned out to 6 workspaces every 10s. With 8+ workspaces a user
+ * triggered sustained 429s (server-side rate limit is 600 req/min/IP).
+ *
+ * These tests pin:
+ *  1. Fan-out cap of 3 — even with 6 online nodes, only 3 fetches
+ *  2. Visibility gate — when collapsed, no polling
+ *
+ * If a future refactor pushes either dial back up, CI fails before
+ * the regression hits a paying tenant.
+ */
+import { describe, it, expect, vi, beforeEach, afterEach } from "vitest";
+import { render, cleanup, act, fireEvent } from "@testing-library/react";
+
+// ── Mocks (hoisted before imports) ────────────────────────────────────────────
+
+vi.mock("@/lib/api", () => ({
+  api: { get: vi.fn() },
+}));
+
+// Six online nodes — enough to verify the cap of 3.
+const mockStoreState = {
+  selectedNodeId: null as string | null,
+  nodes: [
+    { id: "ws-1", data: { status: "online", name: "ws-1" } },
+    { id: "ws-2", data: { status: "online", name: "ws-2" } },
+    { id: "ws-3", data: { status: "online", name: "ws-3" } },
+    { id: "ws-4", data: { status: "online", name: "ws-4" } },
+    { id: "ws-5", data: { status: "online", name: "ws-5" } },
+    { id: "ws-6", data: { status: "online", name: "ws-6" } },
+    { id: "ws-offline", data: { status: "offline", name: "off" } },
+  ],
+};
+
+vi.mock("@/store/canvas", () => ({
+  useCanvasStore: vi.fn(
+    (selector: (s: typeof mockStoreState) => unknown) =>
+      selector(mockStoreState)
+  ),
+}));
+
+// design-tokens has named exports — keep the shape minimal.
+vi.mock("@/lib/design-tokens", () => ({
+  COMM_TYPE_LABELS: {
+    a2a_send: "→",
+    a2a_receive: "←",
+    task_update: "✓",
+  },
+}));
+
+// ── Imports (after mocks) ─────────────────────────────────────────────────────
+
+import { api } from "@/lib/api";
+import { CommunicationOverlay } from "../CommunicationOverlay";
+
+const mockGet = vi.mocked(api.get);
+
+// ── Setup ─────────────────────────────────────────────────────────────────────
+
+beforeEach(() => {
+  vi.useFakeTimers();
+  mockGet.mockReset();
+  mockGet.mockResolvedValue([]);
+});
+
+afterEach(() => {
+  cleanup();
+  vi.useRealTimers();
+});
+
+// ── Tests ─────────────────────────────────────────────────────────────────────
+
+describe("CommunicationOverlay — fan-out cap", () => {
+  it("polls at most 3 of 6 online workspaces (rate-limit floor)", async () => {
+    await act(async () => {
+      render(<CommunicationOverlay />);
+    });
+    // Mount fires the first poll synchronously (no interval tick yet).
+    // Pre-fix: 6 calls. Post-fix: 3.
+    expect(mockGet).toHaveBeenCalledTimes(3);
+    // Verify the calls are for the FIRST 3 online nodes (slice order).
+    expect(mockGet).toHaveBeenCalledWith("/workspaces/ws-1/activity?limit=5");
+    expect(mockGet).toHaveBeenCalledWith("/workspaces/ws-2/activity?limit=5");
+    expect(mockGet).toHaveBeenCalledWith("/workspaces/ws-3/activity?limit=5");
+  });
+
+  it("never polls offline workspaces", async () => {
+    await act(async () => {
+      render(<CommunicationOverlay />);
+    });
+    expect(mockGet).not.toHaveBeenCalledWith(
+      "/workspaces/ws-offline/activity?limit=5",
+    );
+  });
+});
+
+describe("CommunicationOverlay — cadence", () => {
+  it("uses 30s interval cadence (was 10s pre-fix)", async () => {
+    await act(async () => {
+      render(<CommunicationOverlay />);
+    });
+    expect(mockGet).toHaveBeenCalledTimes(3); // initial mount poll
+
+    // Advance 10s — pre-fix this would fire another poll. Post-fix: silent.
+    await act(async () => {
+      vi.advanceTimersByTime(10_000);
+    });
+    expect(mockGet).toHaveBeenCalledTimes(3);
+
+    // Advance to 30s — interval fires.
+    await act(async () => {
+      vi.advanceTimersByTime(20_000);
+    });
+    expect(mockGet).toHaveBeenCalledTimes(6); // +3 from second tick
+  });
+});
+
+describe("CommunicationOverlay — visibility gate", () => {
+  // The visibility gate is the dial that drops collapsed-panel polling
+  // to ZERO. The cadence test above can't catch its removal — if a
+  // refactor dropped `if (!visible) return`, the cadence test would
+  // still pass because the effect would still fire every 30s.
+  //
+  // Direct probe: render with comms-returning mock so the panel
+  // actually renders (close button only exists in the expanded panel,
+  // not the collapsed button-state). Click close, advance the clock,
+  // assert no further fetches.
+  it("stops polling after the user collapses the panel", async () => {
+    // Mock returns one a2a_send so comms.length > 0 → panel renders →
+    // close button accessible.
+    mockGet.mockResolvedValue([
+      {
+        id: "act-1",
+        workspace_id: "ws-1",
+        activity_type: "a2a_send",
+        source_id: "ws-1",
+        target_id: "ws-2",
+        summary: "test",
+        status: "completed",
+        duration_ms: 100,
+        created_at: new Date().toISOString(),
+      },
+    ]);
+
+    const { getByLabelText } = await act(async () => {
+      return render(<CommunicationOverlay />);
+    });
+    // Drain pending microtasks (resolves the await in fetchComms) so
+    // setComms lands and the panel renders. Don't advance time — that
+    // would fire the next interval tick and pollute the assertion.
+    await act(async () => {
+      await Promise.resolve();
+      await Promise.resolve();
+      await Promise.resolve();
+    });
+    // Initial mount polled 3 workspaces.
+    expect(mockGet).toHaveBeenCalledTimes(3);
+    mockGet.mockClear();
+
+    // Click the close button. Synchronous getByLabelText avoids
+    // findBy's internal setTimeout (deadlocks under useFakeTimers).
+    const closeBtn = getByLabelText("Close communications panel");
+    await act(async () => {
+      fireEvent.click(closeBtn);
+    });
+
+    // Advance well past the 30s cadence — gate should suppress the tick.
+    await act(async () => {
+      vi.advanceTimersByTime(60_000);
+    });
+    expect(mockGet).not.toHaveBeenCalled();
+  });
+});

canvas/src/components/__tests__/CookieConsent.test.tsx

@@ -40,7 +40,7 @@ afterEach(() => {
 describe("CookieConsent", () => {
   it("renders the banner when no decision is stored", () => {
     render(<CookieConsent />);
-    expect(screen.getByRole("dialog")).toBeTruthy();
+    expect(screen.getByRole("region")).toBeTruthy();
     expect(screen.getByRole("button", { name: "Accept all" })).toBeTruthy();
     expect(screen.getByRole("button", { name: "Necessary only" })).toBeTruthy();
   });
@@ -48,7 +48,7 @@ describe("CookieConsent", () => {
   it("stores 'accepted' and hides the banner when user clicks Accept all", () => {
     render(<CookieConsent />);
     fireEvent.click(screen.getByRole("button", { name: "Accept all" }));
-    expect(screen.queryByRole("dialog")).toBeNull();
+    expect(screen.queryByRole("region")).toBeNull();
 
     const raw = window.localStorage.getItem(STORAGE_KEY);
     expect(raw).not.toBeNull();
@@ -61,7 +61,7 @@ describe("CookieConsent", () => {
   it("stores 'rejected' and hides the banner when user clicks Necessary only", () => {
     render(<CookieConsent />);
     fireEvent.click(screen.getByRole("button", { name: "Necessary only" }));
-    expect(screen.queryByRole("dialog")).toBeNull();
+    expect(screen.queryByRole("region")).toBeNull();
 
     const parsed = JSON.parse(window.localStorage.getItem(STORAGE_KEY)!);
     expect(parsed.decision).toBe("rejected");
@@ -73,7 +73,7 @@ describe("CookieConsent", () => {
       JSON.stringify({ decision: "accepted", decidedAt: new Date().toISOString(), version: 1 }),
     );
     render(<CookieConsent />);
-    expect(screen.queryByRole("dialog")).toBeNull();
+    expect(screen.queryByRole("region")).toBeNull();
   });
 
   it("re-prompts when the stored decision is on an older policy version", () => {
@@ -82,13 +82,13 @@ describe("CookieConsent", () => {
       JSON.stringify({ decision: "accepted", decidedAt: new Date().toISOString(), version: 0 }),
     );
     render(<CookieConsent />);
-    expect(screen.getByRole("dialog")).toBeTruthy();
+    expect(screen.getByRole("region")).toBeTruthy();
   });
 
   it("re-prompts when localStorage contains invalid JSON", () => {
     window.localStorage.setItem(STORAGE_KEY, "{not json");
     render(<CookieConsent />);
-    expect(screen.getByRole("dialog")).toBeTruthy();
+    expect(screen.getByRole("region")).toBeTruthy();
   });
 
   it("exposes a privacy-policy link with target='_blank'", () => {
@@ -99,11 +99,19 @@ describe("CookieConsent", () => {
     expect(link.getAttribute("rel")).toContain("noreferrer");
   });
 
-  it("uses role=dialog with aria-labelledby and aria-describedby for screen readers", () => {
+  it("uses role=region (NOT dialog) with aria-labelledby/describedby — banner is informational, not modal", () => {
+    // Regression guard: an earlier version claimed role="dialog"
+    // aria-modal="true" without a focus trap. That falsely told screen
+    // readers the rest of the page was inert, trapping AT users in a
+    // banner they couldn't escape. role="region" lets assistive tech
+    // navigate around it normally; the banner stays informational.
     render(<CookieConsent />);
-    const dialog = screen.getByRole("dialog");
-    expect(dialog.getAttribute("aria-labelledby")).toBe("cookie-consent-title");
-    expect(dialog.getAttribute("aria-describedby")).toBe("cookie-consent-body");
+    const banner = screen.getByRole("region");
+    expect(banner.getAttribute("aria-labelledby")).toBe("cookie-consent-title");
+    expect(banner.getAttribute("aria-describedby")).toBe("cookie-consent-body");
+    // No aria-modal claim — explicit guard against regression.
+    expect(banner.getAttribute("aria-modal")).toBeNull();
+    expect(screen.queryByRole("dialog")).toBeNull();
   });
 
   it("does NOT render on local dev (non-SaaS hostname)", () => {
@@ -116,7 +124,7 @@ describe("CookieConsent", () => {
       value: { ...window.location, hostname: "localhost" },
     });
     render(<CookieConsent />);
-    expect(screen.queryByRole("dialog")).toBeNull();
+    expect(screen.queryByRole("region")).toBeNull();
   });
 
   it("does NOT render on a LAN hostname (192.168.*, *.local)", () => {
@@ -125,7 +133,7 @@ describe("CookieConsent", () => {
       value: { ...window.location, hostname: "192.168.1.74" },
     });
     render(<CookieConsent />);
-    expect(screen.queryByRole("dialog")).toBeNull();
+    expect(screen.queryByRole("region")).toBeNull();
   });
 });
 

canvas/src/components/__tests__/SearchDialog.keyboard.test.tsx

@@ -155,18 +155,31 @@ describe("SearchDialog — keyboard accessibility", () => {
     expect(selectNode).not.toHaveBeenCalled();
   });
 
-  it("typing a new query resets focusedIndex to -1", () => {
+  it("typing a query that matches auto-highlights the first result", () => {
+    // Replaces the older "resets to -1" assertion. New behavior: a query
+    // with at least one match pins the highlight to row 0 so Enter picks
+    // a result instead of being a no-op. Empty-query case is covered by
+    // "Enter at focusedIndex=-1 does not select anything" above.
+    render(<SearchDialog />);
+    const input = screen.getByRole("combobox");
+    fireEvent.change(input, { target: { value: "Alpha" } });
+    const options = screen.getAllByRole("option");
+    expect(options[0].getAttribute("aria-selected")).toBe("true");
+    // Enter on the auto-highlighted match should select it without
+    // needing a manual ArrowDown first.
+    fireEvent.keyDown(input, { key: "Enter" });
+    expect(selectNode).toHaveBeenCalledWith("ws-1");
+  });
+
+  it("typing a query that matches NOTHING resets focusedIndex to -1", () => {
     render(<SearchDialog />);
     const input = screen.getByRole("combobox");
     fireEvent.keyDown(input, { key: "ArrowDown" }); // focusedIndex → 0
-    // Verify selection before reset
-    expect(screen.getAllByRole("option")[0].getAttribute("aria-selected")).toBe("true");
-    // Change query — triggers the useEffect that resets focusedIndex
-    fireEvent.change(input, { target: { value: "Alpha" } });
-    // After reset all options must have aria-selected="false"
-    screen.getAllByRole("option").forEach((opt) => {
-      expect(opt.getAttribute("aria-selected")).toBe("false");
-    });
+    fireEvent.change(input, { target: { value: "zzz-no-match" } });
+    // No options remain, so nothing to assert on aria-selected directly —
+    // the empty-state message takes over. But Enter should be a no-op.
+    fireEvent.keyDown(input, { key: "Enter" });
+    expect(selectNode).not.toHaveBeenCalled();
   });
 
   it("aria-activedescendant matches the focused option's id", () => {

canvas/src/components/__tests__/Toaster.test.tsx

@@ -0,0 +1,82 @@
+// @vitest-environment jsdom
+import { describe, it, expect, afterEach, beforeEach, vi } from "vitest";
+import { render, screen, fireEvent, cleanup, act } from "@testing-library/react";
+import { Toaster, showToast } from "../Toaster";
+
+beforeEach(() => {
+  vi.useFakeTimers();
+});
+
+afterEach(() => {
+  cleanup();
+  vi.useRealTimers();
+});
+
+describe("Toaster keyboard a11y", () => {
+  it("Esc dismisses the most recent toast", () => {
+    render(<Toaster />);
+    act(() => {
+      showToast("first", "info");
+      showToast("second", "info");
+    });
+    expect(screen.getByText("first")).toBeTruthy();
+    expect(screen.getByText("second")).toBeTruthy();
+
+    act(() => {
+      fireEvent.keyDown(window, { key: "Escape" });
+    });
+    expect(screen.queryByText("second")).toBeNull();
+    expect(screen.getByText("first")).toBeTruthy();
+  });
+
+  it("Esc dismisses persistent error toasts", () => {
+    render(<Toaster />);
+    act(() => {
+      showToast("boom", "error");
+    });
+    expect(screen.getByText("boom")).toBeTruthy();
+
+    act(() => {
+      fireEvent.keyDown(window, { key: "Escape" });
+    });
+    expect(screen.queryByText("boom")).toBeNull();
+  });
+
+  it("Esc with no toasts is a no-op", () => {
+    render(<Toaster />);
+    act(() => {
+      fireEvent.keyDown(window, { key: "Escape" });
+    });
+    // no throw, nothing rendered
+    expect(screen.queryAllByRole("button", { name: "Dismiss notification" })).toHaveLength(0);
+  });
+
+  it("dismiss button has accessible label and is keyboard reachable", () => {
+    render(<Toaster />);
+    act(() => {
+      showToast("hi", "info");
+    });
+    const btn = screen.getByRole("button", { name: "Dismiss notification" });
+    expect(btn).toBeTruthy();
+    // Native <button> defaults to keyboard-focusable; explicit assertion guards
+    // against a future regression where someone adds tabindex=-1.
+    expect(btn.getAttribute("tabindex")).not.toBe("-1");
+  });
+
+  it("dismiss button click removes that specific toast", () => {
+    render(<Toaster />);
+    act(() => {
+      showToast("a", "info");
+      showToast("b", "info");
+    });
+    const buttons = screen.getAllByRole("button", { name: "Dismiss notification" });
+    expect(buttons).toHaveLength(2);
+
+    // Click the first dismiss → "a" goes away, "b" stays
+    act(() => {
+      fireEvent.click(buttons[0]);
+    });
+    expect(screen.queryByText("a")).toBeNull();
+    expect(screen.getByText("b")).toBeTruthy();
+  });
+});

canvas/src/components/settings/OrgTokensTab.tsx

@@ -182,7 +182,7 @@ export function OrgTokensTab() {
 
       {/* Token list */}
       {loading ? (
-        <div className="flex items-center justify-center gap-2 py-6 text-ink-soft text-xs">
+        <div role="status" aria-live="polite" className="flex items-center justify-center gap-2 py-6 text-ink-soft text-xs">
           <Spinner /> Loading keys...
         </div>
       ) : tokens.length === 0 ? (

canvas/src/components/settings/TokensTab.tsx

@@ -129,7 +129,7 @@ export function TokensTab({ workspaceId }: TokensTabProps) {
 
       {/* Token list */}
       {loading ? (
-        <div className="flex items-center justify-center gap-2 py-6 text-ink-soft text-xs">
+        <div role="status" aria-live="polite" className="flex items-center justify-center gap-2 py-6 text-ink-soft text-xs">
           <Spinner /> Loading tokens...
         </div>
       ) : tokens.length === 0 ? (

canvas/src/components/tabs/ActivityTab.tsx

@@ -110,8 +110,11 @@ export function ActivityTab({ workspaceId }: Props) {
               Full Trace
             </button>
             <button
+              type="button"
               onClick={loadActivities}
-              className="px-2 py-1 bg-surface-card hover:bg-surface-card text-[11px] rounded text-ink-mid"
+              // hover:bg-surface-card on top of itself was a no-op;
+              // lift to surface-elevated + focus-visible ring.
+              className="px-2 py-1 bg-surface-card hover:bg-surface-elevated hover:text-ink text-[11px] rounded text-ink-mid transition-colors focus:outline-none focus-visible:ring-2 focus-visible:ring-accent/50"
             >
               Refresh
             </button>

canvas/src/components/tabs/ChannelsTab.tsx

@@ -365,8 +365,12 @@ export function ChannelsTab({ workspaceId }: Props) {
             <p className="text-[10px] text-bad">{formError}</p>
           )}
           <button
+            type="button"
             onClick={handleCreate}
-            className="w-full text-xs py-1.5 rounded bg-accent-strong hover:bg-accent text-white transition"
+            // Was bg-accent-strong hover:bg-accent — accent is the
+            // LIGHTER variant; same AA contrast trap fixed in
+            // ScheduleTab/MemoryTab/OnboardingWizard.
+            className="w-full text-xs py-1.5 rounded bg-accent hover:bg-accent-strong text-white transition focus:outline-none focus-visible:ring-2 focus-visible:ring-accent/60 focus-visible:ring-offset-2 focus-visible:ring-offset-surface"
           >
             Connect Channel
           </button>

canvas/src/components/tabs/ChatTab.tsx

@@ -177,10 +177,10 @@ export function ChatTab({ workspaceId, data }: Props) {
           aria-controls="chat-panel-my-chat"
           tabIndex={subTab === "my-chat" ? 0 : -1}
           onClick={() => setSubTab("my-chat")}
-          className={`px-3 py-1.5 text-[10px] font-medium transition-colors ${
+          className={`px-3 py-1.5 text-[10px] font-medium transition-colors focus:outline-none focus-visible:ring-2 focus-visible:ring-accent/40 ${
             subTab === "my-chat"
               ? "text-ink border-b-2 border-accent"
-              : "text-ink-soft hover:text-ink-mid"
+              : "text-ink-mid hover:text-ink"
           }`}
         >
           My Chat
@@ -192,10 +192,10 @@ export function ChatTab({ workspaceId, data }: Props) {
           aria-controls="chat-panel-agent-comms"
           tabIndex={subTab === "agent-comms" ? 0 : -1}
           onClick={() => setSubTab("agent-comms")}
-          className={`px-3 py-1.5 text-[10px] font-medium transition-colors ${
+          className={`px-3 py-1.5 text-[10px] font-medium transition-colors focus:outline-none focus-visible:ring-2 focus-visible:ring-accent/40 ${
             subTab === "agent-comms"
               ? "text-ink border-b-2 border-accent"
-              : "text-ink-soft hover:text-ink-mid"
+              : "text-ink-mid hover:text-ink"
           }`}
         >
           Agent Comms
@@ -773,14 +773,39 @@ function MyChatPanel({ workspaceId, data }: Props) {
             <div
               className={`max-w-[85%] rounded-lg px-3 py-2 text-xs ${
                 msg.role === "user"
-                  ? "bg-accent-strong/30 text-blue-100 border border-accent/20"
+                  // Solid blue-600 in both modes — `bg-accent` themes
+                  // lighter in dark, dropping white-text contrast to
+                  // ~3:1 (fails AA). blue-600 keeps ~5:1 against white
+                  // on both warm-paper and dark-slate panels.
+                  ? "bg-blue-600 text-white border border-blue-700 dark:bg-blue-500 dark:border-blue-400 shadow-sm"
                   : msg.role === "system"
-                    ? "bg-red-900/30 text-red-200 border border-red-800/30"
-                    : "bg-surface-card/80 text-ink border border-line/30"
+                    // Bump the system bubble's opacity in dark — /10
+                    // overlay was nearly invisible against the dark
+                    // panel bg.
+                    ? "bg-bad/10 text-bad border border-bad/40 dark:bg-bad/25 dark:text-bad dark:border-bad/60"
+                    // Agent bubble in dark: surface-card (#1a1d23) is
+                    // only ~7% lighter than the panel bg-surface
+                    // (#0e1014). Bump to zinc-700 for a clearly
+                    // elevated bubble; light mode keeps the warm
+                    // surface-card tint.
+                    : "bg-surface-card text-ink border border-line dark:bg-zinc-700 dark:text-zinc-100 dark:border-zinc-600 shadow-sm"
               }`}
             >
               {msg.content && (
-                <div className="prose prose-sm prose-invert max-w-none [&>p]:mb-1 [&>p:last-child]:mb-0">
+                <div
+                  className={`prose prose-sm max-w-none [&>p]:mb-1 [&>p:last-child]:mb-0 ${
+                    msg.role === "user"
+                      ? "prose-invert"
+                      // Agent bubbles in dark mode: invert prose AND brighten
+                      // the body/heading/bold/code tokens. prose-invert's
+                      // default `--tw-prose-invert-body: zinc-300` lands at
+                      // ~5.3:1 against bg-zinc-700 — passes AA but reads
+                      // washed out next to the user bubble's crisp
+                      // white-on-blue (~10:1). Push body to zinc-100 so the
+                      // agent text matches that crispness.
+                      : "dark:prose-invert dark:[--tw-prose-invert-body:theme(colors.zinc.100)] dark:[--tw-prose-invert-headings:theme(colors.white)] dark:[--tw-prose-invert-bold:theme(colors.white)] dark:[--tw-prose-invert-code:theme(colors.zinc.100)]"
+                  }`}
+                >
                   <ReactMarkdown remarkPlugins={[remarkGfm]}>{msg.content}</ReactMarkdown>
                 </div>
               )}
@@ -796,7 +821,7 @@ function MyChatPanel({ workspaceId, data }: Props) {
                   ))}
                 </div>
               )}
-              <div className="text-[9px] text-ink-soft mt-1">
+              <div className={`text-[9px] mt-1 ${msg.role === "user" ? "text-white/70" : "text-ink-mid"}`}>
                 {new Date(msg.timestamp).toLocaleTimeString()}
               </div>
             </div>
@@ -896,7 +921,7 @@ function MyChatPanel({ workspaceId, data }: Props) {
             placeholder={agentReachable ? "Send a message... (Shift+Enter for new line, paste images to attach)" : `Agent is ${data.status}`}
             disabled={!agentReachable || sending}
             rows={1}
-            className="flex-1 bg-surface-card border border-line rounded-lg px-3 py-2 text-xs text-ink placeholder-zinc-500 focus:outline-none focus:border-accent resize-none disabled:opacity-50"
+            className="flex-1 bg-surface-card border border-line rounded-lg px-3 py-2 text-xs text-ink placeholder-ink-soft dark:bg-zinc-800 dark:border-zinc-600 dark:placeholder-zinc-500 focus:outline-none focus:border-accent focus-visible:ring-2 focus-visible:ring-accent/40 resize-none disabled:opacity-50"
           />
           <button
             onClick={sendMessage}

canvas/src/components/tabs/ConfigTab.tsx

@@ -65,11 +65,11 @@ function AgentCardSection({ workspaceId }: { workspaceId: string }) {
           {error && <div className="px-2 py-1 bg-red-900/30 border border-red-800 rounded text-[10px] text-bad">{error}</div>}
           <div className="flex gap-2">
             <button type="button" onClick={handleSave} disabled={saving}
-              className="px-2 py-1 bg-accent-strong hover:bg-accent text-[10px] rounded text-white disabled:opacity-50">
+              className="px-2 py-1 bg-accent hover:bg-accent-strong text-[10px] rounded text-white disabled:opacity-50 transition-colors focus:outline-none focus-visible:ring-2 focus-visible:ring-accent/60 focus-visible:ring-offset-1 focus-visible:ring-offset-surface">
               {saving ? "Saving..." : "Save"}
             </button>
             <button type="button" onClick={() => setEditing(false)}
-              className="px-2 py-1 bg-surface-card hover:bg-surface-card text-[10px] rounded text-ink-mid">Cancel</button>
+              className="px-2 py-1 bg-surface-card hover:bg-surface-elevated hover:text-ink text-[10px] rounded text-ink-mid transition-colors focus:outline-none focus-visible:ring-2 focus-visible:ring-accent/40 focus-visible:ring-offset-1 focus-visible:ring-offset-surface">Cancel</button>
           </div>
         </div>
       ) : (
@@ -655,7 +655,8 @@ export function ConfigTab({ workspaceId }: Props) {
                 >
                   <option value={1}>T1 — Sandboxed</option>
                   <option value={2}>T2 — Standard</option>
-                  <option value={3}>T3 — Full Access</option>
+                  <option value={3}>T3 — Privileged</option>
+                  <option value={4}>T4 — Full Access</option>
                 </select>
               </div>
             </div>
@@ -889,7 +890,6 @@ export function ConfigTab({ workspaceId }: Props) {
             <TagList label="Skills" values={config.skills || []} onChange={(v) => update("skills", v)} placeholder="e.g. code-review" />
             <TagList label="Tools" values={config.tools || []} onChange={(v) => update("tools", v)} placeholder="e.g. web_search, filesystem" />
             <TagList label="Prompt Files" values={config.prompt_files || []} onChange={(v) => update("prompt_files", v)} placeholder="e.g. system-prompt.md" />
-            <TagList label="Shared Context" values={config.shared_context || []} onChange={(v) => update("shared_context", v)} placeholder="e.g. architecture.md" />
           </Section>
 
           <Section title="A2A Protocol" defaultOpen={false}>
@@ -955,7 +955,8 @@ export function ConfigTab({ workspaceId }: Props) {
           type="button"
           onClick={() => handleSave(true)}
           disabled={!isDirty || saving}
-          className="px-3 py-1.5 bg-accent-strong hover:bg-accent text-xs rounded text-white disabled:opacity-30 transition-colors"
+          // Same accent-LIGHTER fix shipped on every other tab.
+          className="px-3 py-1.5 bg-accent hover:bg-accent-strong text-xs rounded text-white disabled:opacity-30 transition-colors focus:outline-none focus-visible:ring-2 focus-visible:ring-accent/60 focus-visible:ring-offset-1 focus-visible:ring-offset-surface"
         >
           {saving ? "Restarting..." : "Save & Restart"}
         </button>

canvas/src/components/tabs/DetailsTab.tsx

@@ -166,7 +166,10 @@ export function DetailsTab({ workspaceId, data }: Props) {
                 type="button"
                 onClick={handleSave}
                 disabled={saving}
-                className="px-3 py-1 bg-accent-strong hover:bg-accent text-xs rounded text-white disabled:opacity-50"
+                // Was bg-accent-strong hover:bg-accent — accent is the
+                // LIGHTER variant; flipped + focus-visible ring (same
+                // trap fix shipped on every other tab).
+                className="px-3 py-1 bg-accent hover:bg-accent-strong text-xs rounded text-white disabled:opacity-50 transition-colors focus:outline-none focus-visible:ring-2 focus-visible:ring-accent/60 focus-visible:ring-offset-1 focus-visible:ring-offset-surface"
               >
                 {saving ? "Saving..." : "Save"}
               </button>
@@ -322,7 +325,10 @@ export function DetailsTab({ workspaceId, data }: Props) {
               <button
                 type="button"
                 onClick={handleDelete}
-                className="px-3 py-1 bg-red-600 hover:bg-red-500 text-xs rounded text-white"
+                // hover:bg-red-500 LIGHTER on white text drops AA;
+                // flipped to bg-red-700 + focus-visible danger ring,
+                // matching the ConfirmDialog/DeleteCascade pattern.
+                className="px-3 py-1 bg-red-600 hover:bg-red-700 text-xs rounded text-white transition-colors focus:outline-none focus-visible:ring-2 focus-visible:ring-red-500/60 focus-visible:ring-offset-1 focus-visible:ring-offset-surface"
               >
                 Confirm Delete
               </button>
@@ -334,7 +340,9 @@ export function DetailsTab({ workspaceId, data }: Props) {
                   // Return focus to the trigger so keyboard users aren't stranded
                   deleteButtonRef.current?.focus();
                 }}
-                className="px-3 py-1 bg-surface-card hover:bg-surface-card text-xs rounded text-ink-mid"
+                // Was hover:bg-surface-card on top of itself (no-op);
+                // lift to surface-elevated.
+                className="px-3 py-1 bg-surface-card hover:bg-surface-elevated hover:text-ink text-xs rounded text-ink-mid transition-colors focus:outline-none focus-visible:ring-2 focus-visible:ring-accent/40 focus-visible:ring-offset-1 focus-visible:ring-offset-surface"
               >
                 Cancel
               </button>

canvas/src/components/tabs/EventsTab.tsx

@@ -15,14 +15,20 @@ interface EventEntry {
   created_at: string;
 }
 
+// Use semantic warm-paper tokens so colors flip with theme. Earlier
+// the table referenced text-yellow-400 / text-purple-400 (Tailwind
+// raw colors, no theme variant), which read fine in dark mode but
+// washed out in the warm-paper light theme. text-warm covers the
+// "degraded" amber tone in both modes; AGENT_CARD_UPDATED is informational
+// metadata, so reuse text-accent for theme-consistency.
 const EVENT_COLORS: Record<string, string> = {
   WORKSPACE_ONLINE: "text-good",
   WORKSPACE_OFFLINE: "text-ink-mid",
-  WORKSPACE_DEGRADED: "text-yellow-400",
+  WORKSPACE_DEGRADED: "text-warm",
   WORKSPACE_PROVISIONING: "text-accent",
   WORKSPACE_REMOVED: "text-bad",
   WORKSPACE_PROVISION_FAILED: "text-bad",
-  AGENT_CARD_UPDATED: "text-purple-400",
+  AGENT_CARD_UPDATED: "text-accent",
 };
 
 export function EventsTab({ workspaceId }: Props) {
@@ -64,8 +70,12 @@ export function EventsTab({ workspaceId }: Props) {
       <div className="flex items-center justify-between mb-2">
         <span className="text-xs text-ink-mid">{events.length} events</span>
         <button
+          type="button"
           onClick={loadEvents}
-          className="px-2 py-1 bg-surface-card hover:bg-surface-card text-[10px] rounded text-ink-mid"
+          // Was hover:bg-surface-card on top of bg-surface-card — silent
+          // no-op hover. Lift to surface-elevated, matching the Cancel
+          // pattern from ConfirmDialog.
+          className="px-2 py-1 bg-surface-card hover:bg-surface-elevated hover:text-ink text-[10px] rounded text-ink-mid transition-colors focus:outline-none focus-visible:ring-2 focus-visible:ring-accent/50"
         >
           Refresh
         </button>
@@ -81,39 +91,51 @@ export function EventsTab({ workspaceId }: Props) {
         <p className="text-xs text-ink-soft text-center py-4">No events yet</p>
       ) : (
         <div className="space-y-1">
-          {events.map((event) => (
-            <div key={event.id} className="bg-surface-card rounded border border-line">
-              <button
-                onClick={() => setExpanded(expanded === event.id ? null : event.id)}
-                className="w-full flex items-center gap-2 px-3 py-2 text-left"
-              >
-                <span
-                  className={`text-xs font-mono ${
-                    EVENT_COLORS[event.event_type] || "text-ink-mid"
-                  }`}
+          {events.map((event) => {
+            const isOpen = expanded === event.id;
+            const panelId = `events-payload-${event.id}`;
+            return (
+              <div key={event.id} className="bg-surface-card rounded border border-line">
+                <button
+                  type="button"
+                  onClick={() => setExpanded(isOpen ? null : event.id)}
+                  // aria-expanded + aria-controls so screen readers
+                  // announce the open/closed state and link the row to
+                  // its payload panel. Without these, AT users hear
+                  // a generic "button" with no indication that it
+                  // toggles or what it controls.
+                  aria-expanded={isOpen}
+                  aria-controls={panelId}
+                  className="w-full flex items-center gap-2 px-3 py-2 text-left rounded-t hover:bg-surface-elevated/40 focus:outline-none focus-visible:ring-2 focus-visible:ring-inset focus-visible:ring-accent/50 transition-colors"
                 >
-                  {event.event_type}
-                </span>
-                <span className="text-[9px] text-ink-soft ml-auto">
-                  {formatTime(event.created_at)}
-                </span>
-                <span className="text-[10px] text-ink-soft">
-                  {expanded === event.id ? "▼" : "▶"}
-                </span>
-              </button>
+                  <span
+                    className={`text-xs font-mono ${
+                      EVENT_COLORS[event.event_type] || "text-ink-mid"
+                    }`}
+                  >
+                    {event.event_type}
+                  </span>
+                  <span className="text-[9px] text-ink-soft ml-auto">
+                    {formatTime(event.created_at)}
+                  </span>
+                  <span aria-hidden="true" className="text-[10px] text-ink-soft">
+                    {isOpen ? "▼" : "▶"}
+                  </span>
+                </button>
 
-              {expanded === event.id && (
-                <div className="px-3 pb-2">
-                  <pre className="text-[10px] text-ink-mid bg-surface-sunken rounded p-2 overflow-x-auto max-h-40">
-                    {JSON.stringify(event.payload, null, 2)}
-                  </pre>
-                  <div className="mt-1 text-[9px] text-ink-soft font-mono">
-                    ID: {event.id}
+                {isOpen && (
+                  <div id={panelId} className="px-3 pb-2">
+                    <pre className="text-[10px] text-ink-mid bg-surface-sunken rounded p-2 overflow-x-auto max-h-40">
+                      {JSON.stringify(event.payload, null, 2)}
+                    </pre>
+                    <div className="mt-1 text-[9px] text-ink-soft font-mono">
+                      ID: {event.id}
+                    </div>
                   </div>
-                </div>
-              )}
-            </div>
-          ))}
+                )}
+              </div>
+            );
+          })}
         </div>
       )}
     </div>

canvas/src/components/tabs/FilesTab.tsx

@@ -162,25 +162,29 @@ export function FilesTab({ workspaceId }: Props) {
       />
 
       {showDeleteAll && (
-        <div className="mx-3 mt-2 px-3 py-2 bg-red-950/30 border border-red-800/40 rounded space-y-1.5">
-          <p className="text-xs text-bad">Delete all {files.filter((f) => !f.dir).length} files? This cannot be undone.</p>
+        // role=alertdialog so SR users hear this destructive prompt
+        // immediately. Delete-All hovers DARKER (bg-red-700) — same AA
+        // contrast trap that bit ConfirmDialog/ApprovalBanner. Cancel
+        // lifts to surface-elevated instead of the prior no-op hover.
+        <div role="alertdialog" aria-labelledby="files-delete-all-msg" className="mx-3 mt-2 px-3 py-2 bg-red-950/30 border border-red-800/40 rounded space-y-1.5">
+          <p id="files-delete-all-msg" className="text-xs text-bad">Delete all {files.filter((f) => !f.dir).length} files? This cannot be undone.</p>
           <div className="flex gap-2">
-            <button type="button" onClick={() => { handleDeleteAll(); setShowDeleteAll(false); }} className="px-2 py-0.5 bg-red-600 hover:bg-red-500 text-[10px] rounded text-white">Delete All</button>
-            <button type="button" onClick={() => setShowDeleteAll(false)} className="px-2 py-0.5 bg-surface-card hover:bg-surface-card text-[10px] rounded text-ink-mid">Cancel</button>
+            <button type="button" onClick={() => { handleDeleteAll(); setShowDeleteAll(false); }} className="px-2 py-0.5 bg-red-600 hover:bg-red-700 text-[10px] rounded text-white transition-colors focus:outline-none focus-visible:ring-2 focus-visible:ring-red-500/60 focus-visible:ring-offset-1 focus-visible:ring-offset-surface">Delete All</button>
+            <button type="button" onClick={() => setShowDeleteAll(false)} className="px-2 py-0.5 bg-surface-card hover:bg-surface-elevated hover:text-ink text-[10px] rounded text-ink-mid transition-colors focus:outline-none focus-visible:ring-2 focus-visible:ring-accent/40 focus-visible:ring-offset-1 focus-visible:ring-offset-surface">Cancel</button>
           </div>
         </div>
       )}
 
       {error && (
-        <div className="mx-3 mt-2 px-3 py-1.5 bg-red-900/30 border border-red-800 rounded text-xs text-bad">{error}</div>
+        <div role="alert" className="mx-3 mt-2 px-3 py-1.5 bg-red-900/30 border border-red-800 rounded text-xs text-bad">{error}</div>
       )}
 
       {confirmDelete && (
-        <div className="mx-3 mt-2 px-3 py-2 bg-amber-950/30 border border-amber-800/40 rounded space-y-1.5">
-          <p className="text-xs text-warm">Delete <span className="font-mono">{confirmDelete}</span>{files.find((f) => f.path === confirmDelete && f.dir) ? " and all its contents" : ""}?</p>
+        <div role="alertdialog" aria-labelledby="files-delete-one-msg" className="mx-3 mt-2 px-3 py-2 bg-amber-950/30 border border-amber-800/40 rounded space-y-1.5">
+          <p id="files-delete-one-msg" className="text-xs text-warm">Delete <span className="font-mono">{confirmDelete}</span>{files.find((f) => f.path === confirmDelete && f.dir) ? " and all its contents" : ""}?</p>
           <div className="flex gap-2">
-            <button type="button" onClick={confirmDeleteFile} className="px-2 py-0.5 bg-red-600 hover:bg-red-500 text-[10px] rounded text-white">Delete</button>
-            <button type="button" onClick={() => setConfirmDelete(null)} className="px-2 py-0.5 bg-surface-card hover:bg-surface-card text-[10px] rounded text-ink-mid">Cancel</button>
+            <button type="button" onClick={confirmDeleteFile} className="px-2 py-0.5 bg-red-600 hover:bg-red-700 text-[10px] rounded text-white transition-colors focus:outline-none focus-visible:ring-2 focus-visible:ring-red-500/60 focus-visible:ring-offset-1 focus-visible:ring-offset-surface">Delete</button>
+            <button type="button" onClick={() => setConfirmDelete(null)} className="px-2 py-0.5 bg-surface-card hover:bg-surface-elevated hover:text-ink text-[10px] rounded text-ink-mid transition-colors focus:outline-none focus-visible:ring-2 focus-visible:ring-accent/40 focus-visible:ring-offset-1 focus-visible:ring-offset-surface">Cancel</button>
           </div>
         </div>
       )}

canvas/src/components/tabs/MemoryTab.tsx

@@ -10,6 +10,7 @@ interface Props {
 interface MemoryEntry {
   key: string;
   value: unknown;
+  version?: number;
   expires_at: string | null;
   updated_at: string;
 }
@@ -28,6 +29,10 @@ export function MemoryTab({ workspaceId }: Props) {
   const [newValue, setNewValue] = useState("");
   const [newTTL, setNewTTL] = useState("");
   const [error, setError] = useState<string | null>(null);
+  const [editingKey, setEditingKey] = useState<string | null>(null);
+  const [editValue, setEditValue] = useState("");
+  const [editTTL, setEditTTL] = useState("");
+  const [editError, setEditError] = useState<string | null>(null);
 
   const awarenessUrl = useMemo(() => {
     try {
@@ -109,6 +114,69 @@ export function MemoryTab({ workspaceId }: Props) {
     }
   };
 
+  const beginEdit = (entry: MemoryEntry) => {
+    setEditError(null);
+    setEditingKey(entry.key);
+    // Stringify objects/arrays as pretty JSON; render plain strings raw so the
+    // editor doesn't surprise users with surrounding quotes.
+    setEditValue(
+      typeof entry.value === "string"
+        ? entry.value
+        : JSON.stringify(entry.value, null, 2),
+    );
+    if (entry.expires_at) {
+      const remainingMs = new Date(entry.expires_at).getTime() - Date.now();
+      const ttl = Math.max(0, Math.floor(remainingMs / 1000));
+      setEditTTL(ttl > 0 ? String(ttl) : "");
+    } else {
+      setEditTTL("");
+    }
+  };
+
+  const cancelEdit = () => {
+    setEditingKey(null);
+    setEditValue("");
+    setEditTTL("");
+    setEditError(null);
+  };
+
+  const handleEditSave = async (entry: MemoryEntry) => {
+    setEditError(null);
+
+    let parsedValue: unknown;
+    try {
+      parsedValue = JSON.parse(editValue);
+    } catch {
+      parsedValue = editValue;
+    }
+
+    // if_match_version closes the silent-overwrite hole when two writers
+    // race. The handler returns 409 with the current version on mismatch
+    // — surface that as a retry hint and reload to pick up the new state.
+    const body: Record<string, unknown> = { key: entry.key, value: parsedValue };
+    if (typeof entry.version === "number") {
+      body.if_match_version = entry.version;
+    }
+    if (editTTL) {
+      const ttl = parseInt(editTTL);
+      if (!Number.isNaN(ttl) && ttl > 0) body.ttl_seconds = ttl;
+    }
+
+    try {
+      await api.post(`/workspaces/${workspaceId}/memory`, body);
+      cancelEdit();
+      loadMemory();
+    } catch (e) {
+      const message = e instanceof Error ? e.message : "Failed to save";
+      if (message.includes("409") || /if_match_version mismatch/i.test(message)) {
+        setEditError("This entry changed since you opened it. Reloading.");
+        loadMemory();
+      } else {
+        setEditError(message);
+      }
+    }
+  };
+
   const openAwareness = () => {
     window.open(awarenessUrl, "_blank", "noopener,noreferrer");
   };
@@ -137,14 +205,14 @@ export function MemoryTab({ workspaceId }: Props) {
             <button
               type="button"
               onClick={() => setShowAwareness((prev) => !prev)}
-              className="shrink-0 px-2 py-1 bg-surface-card hover:bg-surface-card text-[10px] rounded text-ink"
+              className="shrink-0 px-2 py-1 bg-surface-card hover:bg-surface-elevated text-[10px] rounded text-ink"
             >
               {showAwareness ? "Collapse" : "Expand"}
             </button>
             <button
               type="button"
               onClick={openAwareness}
-              className="shrink-0 px-2 py-1 bg-surface-card hover:bg-surface-card text-[10px] rounded text-ink"
+              className="shrink-0 px-2 py-1 bg-surface-card hover:bg-surface-elevated text-[10px] rounded text-ink"
             >
               Open
             </button>
@@ -177,7 +245,7 @@ export function MemoryTab({ workspaceId }: Props) {
             <button
               type="button"
               onClick={() => setShowAwareness(true)}
-              className="shrink-0 px-2 py-1 bg-accent-strong hover:bg-accent text-[10px] rounded text-white"
+              className="shrink-0 px-2 py-1 bg-accent hover:bg-accent-strong text-[10px] rounded text-white"
             >
               Expand
             </button>
@@ -212,21 +280,21 @@ export function MemoryTab({ workspaceId }: Props) {
             <button
               type="button"
               onClick={() => setShowAdvanced((prev) => !prev)}
-              className="px-2 py-1 bg-surface-card hover:bg-surface-card text-[10px] rounded text-ink-mid"
+              className="px-2 py-1 bg-surface-card hover:bg-surface-elevated text-[10px] rounded text-ink-mid"
             >
               {showAdvanced ? "Hide Advanced" : "Advanced"}
             </button>
             <button
               type="button"
               onClick={loadMemory}
-              className="px-2 py-1 bg-surface-card hover:bg-surface-card text-[10px] rounded text-ink-mid"
+              className="px-2 py-1 bg-surface-card hover:bg-surface-elevated text-[10px] rounded text-ink-mid"
             >
               Refresh
             </button>
             <button
               type="button"
               onClick={() => { setShowAdd(!showAdd); if (!showAdd) setShowAdvanced(true); }}
-              className="px-2 py-1 bg-accent-strong hover:bg-accent text-[10px] rounded text-white"
+              className="px-2 py-1 bg-accent hover:bg-accent-strong text-[10px] rounded text-white"
             >
               + Add
             </button>
@@ -262,7 +330,7 @@ export function MemoryTab({ workspaceId }: Props) {
               <button
                 type="button"
                 onClick={handleAdd}
-                className="px-3 py-1 bg-accent-strong hover:bg-accent text-xs rounded text-white"
+                className="px-3 py-1 bg-accent hover:bg-accent-strong text-xs rounded text-white"
               >
                 Save
               </button>
@@ -272,7 +340,7 @@ export function MemoryTab({ workspaceId }: Props) {
                   setShowAdd(false);
                   setError(null);
                 }}
-                className="px-3 py-1 bg-surface-card hover:bg-surface-card text-xs rounded text-ink-mid"
+                className="px-3 py-1 bg-surface-card hover:bg-surface-elevated text-xs rounded text-ink-mid"
               >
                 Cancel
               </button>
@@ -308,20 +376,71 @@ export function MemoryTab({ workspaceId }: Props) {
 
                   {expanded === entry.key && (
                     <div className="px-3 pb-2 space-y-2">
-                      <pre className="text-[10px] text-ink-mid bg-surface-sunken rounded p-2 overflow-x-auto max-h-40">
-                        {JSON.stringify(entry.value, null, 2)}
-                      </pre>
+                      {editingKey === entry.key ? (
+                        <div className="space-y-2">
+                          <textarea
+                            value={editValue}
+                            onChange={(e) => setEditValue(e.target.value)}
+                            rows={4}
+                            aria-label={`Edit value for ${entry.key}`}
+                            className="w-full bg-surface-sunken border border-line rounded px-2 py-1 text-xs font-mono text-ink focus:outline-none focus:border-accent resize-none"
+                          />
+                          <input
+                            value={editTTL}
+                            onChange={(e) => setEditTTL(e.target.value)}
+                            placeholder="TTL in seconds (blank = no expiry)"
+                            aria-label={`Edit TTL for ${entry.key}`}
+                            className="w-full bg-surface-sunken border border-line rounded px-2 py-1 text-xs text-ink focus:outline-none focus:border-accent"
+                          />
+                          {editError && (
+                            <div role="alert" className="text-[10px] text-bad">
+                              {editError}
+                            </div>
+                          )}
+                          <div className="flex gap-2">
+                            <button
+                              type="button"
+                              onClick={() => handleEditSave(entry)}
+                              className="px-3 py-1 bg-accent hover:bg-accent-strong text-xs rounded text-white"
+                            >
+                              Save
+                            </button>
+                            <button
+                              type="button"
+                              onClick={cancelEdit}
+                              className="px-3 py-1 bg-surface-card hover:bg-surface-elevated text-xs rounded text-ink-mid"
+                            >
+                              Cancel
+                            </button>
+                          </div>
+                        </div>
+                      ) : (
+                        <pre className="text-[10px] text-ink-mid bg-surface-sunken rounded p-2 overflow-x-auto max-h-40">
+                          {JSON.stringify(entry.value, null, 2)}
+                        </pre>
+                      )}
                       <div className="flex items-center justify-between">
                         <span className="text-[9px] text-ink-soft">
                           Updated: {new Date(entry.updated_at).toLocaleString()}
                         </span>
-                        <button
-                          type="button"
-                          onClick={() => handleDelete(entry.key)}
-                          className="text-[10px] text-bad hover:text-bad"
-                        >
-                          Delete
-                        </button>
+                        <div className="flex items-center gap-2">
+                          {editingKey !== entry.key && (
+                            <button
+                              type="button"
+                              onClick={() => beginEdit(entry)}
+                              className="text-[10px] text-ink-mid hover:bg-surface-elevated rounded px-1 transition-colors focus:outline-none focus-visible:ring-2 focus-visible:ring-accent/60"
+                            >
+                              Edit
+                            </button>
+                          )}
+                          <button
+                            type="button"
+                            onClick={() => handleDelete(entry.key)}
+                            className="text-[10px] text-bad hover:bg-red-950/40 rounded px-1 transition-colors focus:outline-none focus-visible:ring-2 focus-visible:ring-red-500/60"
+                          >
+                            Delete
+                          </button>
+                        </div>
                       </div>
                     </div>
                   )}
@@ -340,7 +459,7 @@ export function MemoryTab({ workspaceId }: Props) {
             <button
               type="button"
               onClick={() => setShowAdvanced(true)}
-              className="shrink-0 px-2 py-1 bg-accent-strong hover:bg-accent text-[10px] rounded text-white"
+              className="shrink-0 px-2 py-1 bg-accent hover:bg-accent-strong text-[10px] rounded text-white"
             >
               Show
             </button>

canvas/src/components/tabs/ScheduleTab.tsx

@@ -269,15 +269,23 @@ export function ScheduleTab({ workspaceId }: Props) {
           {error && <div className="text-[10px] text-bad">{error}</div>}
           <div className="flex gap-2">
             <button
+              type="button"
               onClick={handleSubmit}
               disabled={!formCron || !formPrompt}
-              className="text-[11px] px-3 py-1 bg-accent-strong text-white rounded hover:bg-accent disabled:opacity-40 transition-colors"
+              // Was bg-accent-strong hover:bg-accent — accent is the
+              // LIGHTER variant, so this hovered lighter on white text
+              // and dropped contrast below AA. Same trap fixed in
+              // OnboardingWizard, ConfirmDialog, ApprovalBanner.
+              className="text-[11px] px-3 py-1 bg-accent text-white rounded hover:bg-accent-strong disabled:opacity-40 transition-colors focus:outline-none focus-visible:ring-2 focus-visible:ring-accent/60 focus-visible:ring-offset-1 focus-visible:ring-offset-surface"
             >
               {editId ? "Update" : "Create"}
             </button>
             <button
+              type="button"
               onClick={resetForm}
-              className="text-[11px] px-3 py-1 bg-surface-card text-ink-mid rounded hover:bg-surface-card transition-colors"
+              // Was hover:bg-surface-card on top of bg-surface-card —
+              // silent no-op hover. Lift to surface-elevated.
+              className="text-[11px] px-3 py-1 bg-surface-card text-ink-mid rounded hover:bg-surface-elevated hover:text-ink transition-colors focus:outline-none focus-visible:ring-2 focus-visible:ring-accent/40 focus-visible:ring-offset-1 focus-visible:ring-offset-surface"
             >
               Cancel
             </button>

canvas/src/components/tabs/SkillsTab.tsx

@@ -403,7 +403,7 @@ export function SkillsTab({ workspaceId, data }: Props) {
                   }}
                   placeholder="e.g. github://owner/repo#v1.0"
                   spellCheck={false}
-                  className="flex-1 rounded border border-line bg-surface px-2 py-1 text-[10px] text-ink placeholder:text-ink-soft focus:border-violet-600 focus:outline-none"
+                  className="flex-1 rounded border border-line bg-surface px-2 py-1 text-[10px] text-ink placeholder:text-ink-soft focus:outline-none focus:border-violet-600 focus-visible:ring-2 focus-visible:ring-violet-600/50"
                 />
                 <button
                   onClick={handleInstallCustom}

canvas/src/components/tabs/TerminalTab.tsx

@@ -123,15 +123,18 @@ export function TerminalTab({ workspaceId }: Props) {
 
   return (
     <div className="flex flex-col h-full">
-      {/* Status bar — role="status" so connection state changes are announced politely */}
+      {/* Status bar — role="status" so connection state changes are announced politely.
+          Terminal body stays dark unconditionally (Canvas v4 design rule), but the
+          chrome wrapping it now uses semantic status colors so the dot/text stay
+          readable in both themes. */}
       <div role="status" aria-live="polite" className="flex items-center justify-between px-3 py-1.5 border-b border-zinc-700 bg-zinc-800/50">
         <div className="flex items-center gap-2">
           <div className={`w-2 h-2 rounded-full ${
-            status === "connected" ? "bg-green-500" :
-            status === "connecting" ? "bg-yellow-500 motion-safe:animate-pulse" :
-            status === "error" ? "bg-red-500" : "bg-zinc-500"
+            status === "connected" ? "bg-good" :
+            status === "connecting" ? "bg-warm motion-safe:animate-pulse" :
+            status === "error" ? "bg-bad" : "bg-ink-soft"
           }`} />
-          <span className="text-[10px] text-zinc-400">
+          <span className="text-[10px] text-zinc-300">
             {status === "connected" ? "Shell active" :
              status === "connecting" ? "Connecting..." :
              status === "error" ? "Connection failed" : "Disconnected"}
@@ -139,8 +142,13 @@ export function TerminalTab({ workspaceId }: Props) {
         </div>
         {(status === "disconnected" || status === "error") && (
           <button
+            type="button"
             onClick={reconnect}
-            className="text-[10px] text-blue-400 hover:text-blue-300"
+            // Accent over hardcoded blue. text-accent + hover-strong stays
+            // readable on the dark terminal chrome and matches the rest
+            // of the canvas semantic palette. Focus-visible ring added so
+            // keyboard users see where focus lands on a recovery button.
+            className="text-[10px] text-accent hover:text-accent-strong rounded-sm px-1 transition-colors focus:outline-none focus-visible:ring-2 focus-visible:ring-accent/60"
           >
             Reconnect
           </button>
@@ -149,7 +157,7 @@ export function TerminalTab({ workspaceId }: Props) {
 
       {/* Error message — role="alert" announces immediately via assertive live region */}
       {errorMsg && (
-        <div role="alert" className="mx-3 mt-2 px-3 py-1.5 bg-red-900/30 border border-red-800 rounded text-xs text-red-400">
+        <div role="alert" className="mx-3 mt-2 px-3 py-1.5 bg-red-900/30 border border-red-800 rounded text-xs text-bad">
           {errorMsg}
         </div>
       )}

canvas/src/components/tabs/TracesTab.tsx

@@ -55,7 +55,13 @@ export function TracesTab({ workspaceId }: Props) {
     <div className="p-4 space-y-2">
       <div className="flex items-center justify-between mb-2">
         <span className="text-xs text-ink-mid">{traces.length} traces</span>
-        <button type="button" onClick={loadTraces} className="text-[10px] text-ink-soft hover:text-ink-mid">
+        <button
+          type="button"
+          onClick={loadTraces}
+          // Added focus-visible ring; previous version was hover-only,
+          // invisible to keyboard users.
+          className="text-[10px] text-ink-soft hover:text-ink-mid rounded-sm px-1 transition-colors focus:outline-none focus-visible:ring-2 focus-visible:ring-accent/50"
+        >
           Refresh
         </button>
       </div>
@@ -79,66 +85,79 @@ export function TracesTab({ workspaceId }: Props) {
         </div>
       ) : (
         <div className="space-y-1">
-          {traces.map((trace) => (
-            <div key={trace.id} className="bg-surface-card/40 border border-line/40 rounded-lg overflow-hidden">
-              <button
-                onClick={() => setExpanded(expanded === trace.id ? null : trace.id)}
-                className="w-full px-3 py-2 flex items-center gap-2 text-left hover:bg-surface-card/60 transition-colors"
-              >
-                <div className={`w-1.5 h-1.5 rounded-full shrink-0 ${
-                  trace.status === "ERROR" ? "bg-red-400" : "bg-emerald-400"
-                }`} />
-                <div className="flex-1 min-w-0">
-                  <div className="text-[11px] text-ink truncate">{trace.name || "trace"}</div>
-                  <div className="text-[9px] text-ink-soft">{formatTime(trace.timestamp)}</div>
-                </div>
-                <div className="flex items-center gap-2 shrink-0">
-                  {trace.latency != null && (
-                    <span className="text-[9px] text-ink-soft tabular-nums">
-                      {trace.latency > 1000 ? `${(trace.latency / 1000).toFixed(1)}s` : `${trace.latency}ms`}
-                    </span>
-                  )}
-                  {trace.usage?.total != null && (
-                    <span className="text-[9px] text-ink-soft tabular-nums">
-                      {trace.usage.total} tok
-                    </span>
-                  )}
-                  <span className="text-[9px] text-ink-soft">
-                    {expanded === trace.id ? "▼" : "▶"}
-                  </span>
-                </div>
-              </button>
-
-              {expanded === trace.id && (
-                <div className="px-3 pb-2 space-y-2 border-t border-line/30">
-                  {trace.input && (
-                    <div>
-                      <div className="text-[9px] text-ink-soft uppercase tracking-wider mt-2 mb-1">Input</div>
-                      <pre className="text-[9px] text-ink-mid bg-surface-sunken rounded p-2 overflow-x-auto max-h-32">
-                        {String(typeof trace.input === "string" ? trace.input : JSON.stringify(trace.input, null, 2))}
-                      </pre>
-                    </div>
-                  )}
-                  {trace.output && (
-                    <div>
-                      <div className="text-[9px] text-ink-soft uppercase tracking-wider mb-1">Output</div>
-                      <pre className="text-[9px] text-ink-mid bg-surface-sunken rounded p-2 overflow-x-auto max-h-32">
-                        {String(typeof trace.output === "string" ? trace.output : JSON.stringify(trace.output, null, 2))}
-                      </pre>
-                    </div>
-                  )}
-                  {trace.totalCost != null && (
-                    <div className="text-[9px] text-ink-soft">
-                      Cost: ${trace.totalCost.toFixed(6)}
-                    </div>
-                  )}
-                  <div className="text-[8px] text-ink-soft font-mono select-all">
-                    {trace.id}
+          {traces.map((trace) => {
+            const isOpen = expanded === trace.id;
+            const panelId = `trace-detail-${trace.id}`;
+            return (
+              <div key={trace.id} className="bg-surface-card/40 border border-line/40 rounded-lg overflow-hidden">
+                <button
+                  type="button"
+                  onClick={() => setExpanded(isOpen ? null : trace.id)}
+                  // aria-expanded + aria-controls so SR announces the
+                  // open/closed state and links the row to its detail
+                  // panel. Same pattern shipped on EventsTab.
+                  aria-expanded={isOpen}
+                  aria-controls={panelId}
+                  className="w-full px-3 py-2 flex items-center gap-2 text-left hover:bg-surface-card/60 focus:outline-none focus-visible:ring-2 focus-visible:ring-inset focus-visible:ring-accent/50 transition-colors"
+                >
+                  {/* Status dot uses semantic bad/good tokens — was hardcoded
+                      bg-red-400 / bg-emerald-400 which doesn't pin to the
+                      canvas-wide ramp. */}
+                  <div className={`w-1.5 h-1.5 rounded-full shrink-0 ${
+                    trace.status === "ERROR" ? "bg-bad" : "bg-good"
+                  }`} />
+                  <div className="flex-1 min-w-0">
+                    <div className="text-[11px] text-ink truncate">{trace.name || "trace"}</div>
+                    <div className="text-[9px] text-ink-soft">{formatTime(trace.timestamp)}</div>
                   </div>
-                </div>
-              )}
-            </div>
-          ))}
+                  <div className="flex items-center gap-2 shrink-0">
+                    {trace.latency != null && (
+                      <span className="text-[9px] text-ink-soft tabular-nums">
+                        {trace.latency > 1000 ? `${(trace.latency / 1000).toFixed(1)}s` : `${trace.latency}ms`}
+                      </span>
+                    )}
+                    {trace.usage?.total != null && (
+                      <span className="text-[9px] text-ink-soft tabular-nums">
+                        {trace.usage.total} tok
+                      </span>
+                    )}
+                    <span aria-hidden="true" className="text-[9px] text-ink-soft">
+                      {isOpen ? "▼" : "▶"}
+                    </span>
+                  </div>
+                </button>
+
+                {isOpen && (
+                  <div id={panelId} className="px-3 pb-2 space-y-2 border-t border-line/30">
+                    {trace.input && (
+                      <div>
+                        <div className="text-[9px] text-ink-soft uppercase tracking-wider mt-2 mb-1">Input</div>
+                        <pre className="text-[9px] text-ink-mid bg-surface-sunken rounded p-2 overflow-x-auto max-h-32">
+                          {String(typeof trace.input === "string" ? trace.input : JSON.stringify(trace.input, null, 2))}
+                        </pre>
+                      </div>
+                    )}
+                    {trace.output && (
+                      <div>
+                        <div className="text-[9px] text-ink-soft uppercase tracking-wider mb-1">Output</div>
+                        <pre className="text-[9px] text-ink-mid bg-surface-sunken rounded p-2 overflow-x-auto max-h-32">
+                          {String(typeof trace.output === "string" ? trace.output : JSON.stringify(trace.output, null, 2))}
+                        </pre>
+                      </div>
+                    )}
+                    {trace.totalCost != null && (
+                      <div className="text-[9px] text-ink-soft">
+                        Cost: ${trace.totalCost.toFixed(6)}
+                      </div>
+                    )}
+                    <div className="text-[8px] text-ink-soft font-mono select-all">
+                      {trace.id}
+                    </div>
+                  </div>
+                )}
+              </div>
+            );
+          })}
         </div>
       )}
     </div>

canvas/src/components/tabs/__tests__/MemoryTab.edit.test.tsx

@@ -0,0 +1,220 @@
+// @vitest-environment jsdom
+//
+// Pins the Edit affordance added to MemoryTab. Until this PR the Memory tab
+// was Add+Delete only; an entry that needed correction had to be deleted and
+// re-added — losing the version-counter and any in-flight optimistic-locking
+// invariants other writers depend on.
+//
+// Each test pins one branch of the new flow. If any fails, the bug is back.
+
+import { describe, it, expect, vi, afterEach, beforeEach } from "vitest";
+import { render, screen, cleanup, waitFor, fireEvent } from "@testing-library/react";
+import React from "react";
+
+afterEach(cleanup);
+
+const apiGet = vi.fn();
+const apiPost = vi.fn();
+const apiDel = vi.fn();
+vi.mock("@/lib/api", () => ({
+  api: {
+    get: (path: string) => apiGet(path),
+    post: (path: string, body: unknown) => apiPost(path, body),
+    del: (path: string) => apiDel(path),
+    patch: vi.fn(),
+    put: vi.fn(),
+  },
+}));
+
+import { MemoryTab } from "../MemoryTab";
+
+const sampleEntries = [
+  {
+    key: "team_brief",
+    value: { goal: "ship v2" },
+    version: 3,
+    expires_at: null,
+    updated_at: "2026-05-04T10:00:00Z",
+  },
+  {
+    key: "plain_note",
+    value: "raw text note",
+    version: 1,
+    expires_at: "2099-01-01T00:00:00Z",
+    updated_at: "2026-05-04T10:01:00Z",
+  },
+];
+
+beforeEach(() => {
+  apiGet.mockReset();
+  apiPost.mockReset();
+  apiDel.mockReset();
+  apiGet.mockImplementation((path: string) => {
+    if (path === "/workspaces/ws-test/memory") {
+      return Promise.resolve(sampleEntries);
+    }
+    return Promise.reject(new Error(`unmocked api.get: ${path}`));
+  });
+});
+
+async function renderAndExpand(key: string) {
+  render(<MemoryTab workspaceId="ws-test" />);
+  await waitFor(() => expect(apiGet).toHaveBeenCalled());
+  // Reveal the Advanced section that hosts the entry list.
+  const showAdvanced = await screen.findByRole("button", { name: "Show" });
+  fireEvent.click(showAdvanced);
+  // Expand the row.
+  const row = await screen.findByRole("button", { name: new RegExp(key) });
+  fireEvent.click(row);
+}
+
+describe("MemoryTab Edit affordance", () => {
+  it("Edit button appears once a row is expanded", async () => {
+    await renderAndExpand("team_brief");
+    expect(screen.getAllByRole("button", { name: "Edit" }).length).toBeGreaterThan(0);
+  });
+
+  it("clicking Edit on a JSON-valued entry pre-fills the textarea with pretty JSON", async () => {
+    await renderAndExpand("team_brief");
+    fireEvent.click(screen.getAllByRole("button", { name: "Edit" })[0]);
+    const textarea = (await screen.findByLabelText(
+      "Edit value for team_brief",
+    )) as HTMLTextAreaElement;
+    expect(textarea.value).toBe('{\n  "goal": "ship v2"\n}');
+  });
+
+  it("clicking Edit on a string-valued entry pre-fills raw (no surrounding quotes)", async () => {
+    await renderAndExpand("plain_note");
+    fireEvent.click(screen.getAllByRole("button", { name: "Edit" })[0]);
+    const textarea = (await screen.findByLabelText(
+      "Edit value for plain_note",
+    )) as HTMLTextAreaElement;
+    expect(textarea.value).toBe("raw text note");
+  });
+
+  it("Save POSTs with if_match_version + parsed value, then reloads", async () => {
+    apiPost.mockResolvedValue({ status: "ok", key: "team_brief", version: 4 });
+    await renderAndExpand("team_brief");
+    fireEvent.click(screen.getAllByRole("button", { name: "Edit" })[0]);
+    const textarea = await screen.findByLabelText("Edit value for team_brief");
+    fireEvent.change(textarea, { target: { value: '{"goal":"ship v3"}' } });
+    fireEvent.click(screen.getByRole("button", { name: "Save" }));
+
+    await waitFor(() => expect(apiPost).toHaveBeenCalledTimes(1));
+    expect(apiPost).toHaveBeenCalledWith("/workspaces/ws-test/memory", {
+      key: "team_brief",
+      value: { goal: "ship v3" },
+      if_match_version: 3,
+    });
+    // Reload after save → second GET.
+    await waitFor(() => expect(apiGet).toHaveBeenCalledTimes(2));
+  });
+
+  it("Save with non-JSON text falls back to plain string", async () => {
+    apiPost.mockResolvedValue({ status: "ok" });
+    await renderAndExpand("team_brief");
+    fireEvent.click(screen.getAllByRole("button", { name: "Edit" })[0]);
+    const textarea = await screen.findByLabelText("Edit value for team_brief");
+    fireEvent.change(textarea, { target: { value: "free-form note" } });
+    fireEvent.click(screen.getByRole("button", { name: "Save" }));
+
+    await waitFor(() => expect(apiPost).toHaveBeenCalledTimes(1));
+    expect(apiPost.mock.calls[0][1].value).toBe("free-form note");
+  });
+
+  it("TTL field is forwarded as ttl_seconds when set", async () => {
+    apiPost.mockResolvedValue({ status: "ok" });
+    await renderAndExpand("team_brief");
+    fireEvent.click(screen.getAllByRole("button", { name: "Edit" })[0]);
+    const ttlInput = await screen.findByLabelText("Edit TTL for team_brief");
+    fireEvent.change(ttlInput, { target: { value: "3600" } });
+    fireEvent.click(screen.getByRole("button", { name: "Save" }));
+
+    await waitFor(() => expect(apiPost).toHaveBeenCalledTimes(1));
+    expect(apiPost.mock.calls[0][1].ttl_seconds).toBe(3600);
+  });
+
+  it("blank/zero/non-numeric TTL is omitted from the payload", async () => {
+    apiPost.mockResolvedValue({ status: "ok" });
+    await renderAndExpand("team_brief");
+    fireEvent.click(screen.getAllByRole("button", { name: "Edit" })[0]);
+    const ttlInput = await screen.findByLabelText("Edit TTL for team_brief");
+    // Junk + zero both must drop out — payload must not contain ttl_seconds.
+    fireEvent.change(ttlInput, { target: { value: "abc" } });
+    fireEvent.click(screen.getByRole("button", { name: "Save" }));
+    await waitFor(() => expect(apiPost).toHaveBeenCalledTimes(1));
+    expect(apiPost.mock.calls[0][1]).not.toHaveProperty("ttl_seconds");
+  });
+
+  it("Cancel discards edits and restores the rendered value", async () => {
+    await renderAndExpand("team_brief");
+    fireEvent.click(screen.getAllByRole("button", { name: "Edit" })[0]);
+    const textarea = await screen.findByLabelText("Edit value for team_brief");
+    fireEvent.change(textarea, { target: { value: '{"goal":"discarded"}' } });
+    fireEvent.click(screen.getByRole("button", { name: "Cancel" }));
+
+    expect(apiPost).not.toHaveBeenCalled();
+    // Editor is gone; the JSON pre-block is back.
+    expect(screen.queryByLabelText("Edit value for team_brief")).toBeNull();
+    expect(screen.getAllByText(/"goal": "ship v2"/i).length).toBeGreaterThan(0);
+  });
+
+  it("409 response surfaces a retry hint and reloads", async () => {
+    apiPost.mockRejectedValueOnce(
+      new Error("HTTP 409: if_match_version mismatch"),
+    );
+    await renderAndExpand("team_brief");
+    fireEvent.click(screen.getAllByRole("button", { name: "Edit" })[0]);
+    const textarea = await screen.findByLabelText("Edit value for team_brief");
+    fireEvent.change(textarea, { target: { value: '{"goal":"ship v3"}' } });
+    fireEvent.click(screen.getByRole("button", { name: "Save" }));
+
+    await waitFor(() => expect(apiPost).toHaveBeenCalledTimes(1));
+    const alert = await screen.findByRole("alert");
+    expect(alert.textContent).toMatch(/changed since you opened it/i);
+    // Initial mount load + post-conflict reload.
+    await waitFor(() => expect(apiGet).toHaveBeenCalledTimes(2));
+  });
+
+  it("non-409 error surfaces the message and does not reload", async () => {
+    apiPost.mockRejectedValueOnce(new Error("boom"));
+    await renderAndExpand("team_brief");
+    fireEvent.click(screen.getAllByRole("button", { name: "Edit" })[0]);
+    fireEvent.click(screen.getByRole("button", { name: "Save" }));
+
+    const alert = await screen.findByRole("alert");
+    expect(alert.textContent).toBe("boom");
+    // Only the initial mount load — no retry reload.
+    expect(apiGet).toHaveBeenCalledTimes(1);
+  });
+
+  it("entry with no version omits if_match_version (back-compat with older shape)", async () => {
+    // Pre-version-counter shape: drop the `version` field from the row.
+    apiGet.mockReset();
+    apiGet.mockImplementation((path: string) => {
+      if (path === "/workspaces/ws-test/memory") {
+        return Promise.resolve([
+          {
+            key: "old_entry",
+            value: "legacy",
+            expires_at: null,
+            updated_at: "2026-05-04T10:00:00Z",
+          },
+        ]);
+      }
+      return Promise.reject(new Error(`unmocked: ${path}`));
+    });
+    apiPost.mockResolvedValue({ status: "ok" });
+
+    await renderAndExpand("old_entry");
+    fireEvent.click(screen.getAllByRole("button", { name: "Edit" })[0]);
+    const textarea = await screen.findByLabelText("Edit value for old_entry");
+    fireEvent.change(textarea, { target: { value: "updated" } });
+    fireEvent.click(screen.getByRole("button", { name: "Save" }));
+
+    await waitFor(() => expect(apiPost).toHaveBeenCalledTimes(1));
+    const payload = apiPost.mock.calls[0][1];
+    expect(payload).not.toHaveProperty("if_match_version");
+    expect(payload.value).toBe("updated");
+  });
+});

canvas/src/components/tabs/chat/AgentCommsPanel.tsx

@@ -472,6 +472,7 @@ function GroupedCommsView({
             <NormalMessage key={msg.id} msg={msg} />
           ),
         )}
+        <WaitingBubbles visible={visible} />
         <div ref={bottomRef} />
       </div>
     </div>
@@ -560,6 +561,83 @@ function PeerTabButton({
   );
 }
 
+/** WaitingBubbles renders one "typing" bubble per peer that has an
+ *  in-flight outbound delegation — i.e., the most recent outbound
+ *  message to that peer is still pending or queued and no later inbound
+ *  reply has arrived. Mirrors the bouncing-dots indicator in ChatTab so
+ *  the operator sees the same visual cue regardless of whether they're
+ *  watching their own chat or a peer thread.
+ *
+ *  Why "per peer" not "one global": when multiple delegations are in
+ *  flight to different peers (common during fan-out), one shared
+ *  spinner under-reports — the user can't tell whether ALL peers are
+ *  still working or only the visible ones. Per-peer matches Slack-style
+ *  typing indicators and keeps the signal honest.
+ *
+ *  Why we look at the LAST per-peer message: once a peer replies (an
+ *  "in" bubble lands), the corresponding "out" bubble is no longer the
+ *  tail — even if status hasn't been mutated to "completed", the inbound
+ *  reply means the wait is over. Looking at the tail collapses both
+ *  cases into one rule.
+ */
+function WaitingBubbles({ visible }: { visible: CommMessage[] }) {
+  // Group by peer, keep only the chronologically-last message per peer,
+  // emit a bubble when that tail is an outbound pending/queued.
+  const tailByPeer = new Map<string, CommMessage>();
+  for (const m of visible) {
+    const prev = tailByPeer.get(m.peerId);
+    if (!prev || m.timestamp > prev.timestamp) tailByPeer.set(m.peerId, m);
+  }
+  const waitingPeers = Array.from(tailByPeer.values()).filter(
+    (m) => m.flow === "out" && (m.status === "pending" || m.status === "queued"),
+  );
+  if (waitingPeers.length === 0) return null;
+  return (
+    <>
+      {waitingPeers.map((m) => (
+        <div
+          key={`waiting-${m.peerId}`}
+          className="flex justify-end"
+          // Outbound thread → right-justified to match the "out" bubble
+          // alignment, so the dots feel like they belong to the message
+          // they're replying to.
+        >
+          <div
+            className="max-w-[85%] rounded-lg px-3 py-2 text-xs bg-cyan-900/30 border border-cyan-700/20"
+            // role+aria-label so screen readers announce the wait;
+            // matches the announcing pattern used by Toaster.
+            role="status"
+            aria-label={`Waiting for reply from ${m.peerName}`}
+          >
+            <div className="text-[9px] text-ink-soft mb-1">→ To {m.peerName}</div>
+            <span className="flex items-center gap-2 text-ink-mid">
+              <span className="flex gap-0.5" aria-hidden="true">
+                <span
+                  className="w-1.5 h-1.5 bg-cyan-300/70 rounded-full motion-safe:animate-bounce"
+                  style={{ animationDelay: "0ms" }}
+                />
+                <span
+                  className="w-1.5 h-1.5 bg-cyan-300/70 rounded-full motion-safe:animate-bounce"
+                  style={{ animationDelay: "150ms" }}
+                />
+                <span
+                  className="w-1.5 h-1.5 bg-cyan-300/70 rounded-full motion-safe:animate-bounce"
+                  style={{ animationDelay: "300ms" }}
+                />
+              </span>
+              <span className="text-[10px]">
+                {m.status === "queued"
+                  ? `${m.peerName} is busy — reply will arrive when they're free`
+                  : `Waiting for ${m.peerName}…`}
+              </span>
+            </span>
+          </div>
+        </div>
+      ))}
+    </>
+  );
+}
+
 function NormalMessage({ msg }: { msg: CommMessage }) {
   return (
     <div className={`flex ${msg.flow === "out" ? "justify-end" : "justify-start"}`}>
@@ -574,12 +652,22 @@ function NormalMessage({ msg }: { msg: CommMessage }) {
           {msg.flow === "out" ? `→ To ${msg.peerName}` : `← From ${msg.peerName}`}
         </div>
         {msg.text ? (
-          <MarkdownBody className="text-ink-mid">{msg.text}</MarkdownBody>
+          // Outgoing bubble (cyan-900) is dark in both themes → prose-invert default.
+          // Incoming bubble (surface-card) themes light → only invert in dark.
+          <MarkdownBody
+            className="text-ink-mid"
+            invert={msg.flow === "out" ? "always" : "dark-only"}
+          >
+            {msg.text}
+          </MarkdownBody>
         ) : (
           <div className="text-ink-mid">(no message text)</div>
         )}
         {msg.responseText && (
-          <MarkdownBody className="mt-1.5 pt-1.5 border-t border-line/30 text-ink-mid">
+          <MarkdownBody
+            className="mt-1.5 pt-1.5 border-t border-line/30 text-ink-mid"
+            invert={msg.flow === "out" ? "always" : "dark-only"}
+          >
             {msg.responseText}
           </MarkdownBody>
         )}
@@ -706,17 +794,29 @@ function ErrorMessage({ msg }: { msg: CommMessage }) {
  *  prose tweaks that keep paragraphs tight inside a small bubble.
  *  Code blocks get an `overflow-x-auto` so a long line of code doesn't
  *  blow out the bubble's max-width — agent-to-agent replies routinely
- *  ship code samples and JSON. */
+ *  ship code samples and JSON.
+ *
+ *  `invert` controls the prose color flip:
+ *  - "always": container bg is dark in BOTH themes (cyan-900, red-950),
+ *    so prose always wants light body text.
+ *  - "dark-only": container bg uses a theming token that goes light in
+ *    light mode (e.g. bg-surface-card). Prose only inverts in dark
+ *    mode; light mode keeps default dark prose colors against the
+ *    light bg. Without this, light mode rendered light text on light
+ *    bg = invisible markdown. */
 function MarkdownBody({
   children,
   className,
+  invert = "always",
 }: {
   children: string;
   className?: string;
+  invert?: "always" | "dark-only";
 }) {
+  const proseInvert = invert === "always" ? "prose-invert" : "dark:prose-invert";
   return (
     <div
-      className={`prose prose-sm prose-invert max-w-none [&>p]:mb-1 [&>p:last-child]:mb-0 [&_pre]:overflow-x-auto [&_table]:block [&_table]:overflow-x-auto ${className ?? ""}`}
+      className={`prose prose-sm ${proseInvert} max-w-none [&>p]:mb-1 [&>p:last-child]:mb-0 [&_pre]:overflow-x-auto [&_table]:block [&_table]:overflow-x-auto ${className ?? ""}`}
     >
       <ReactMarkdown remarkPlugins={[remarkGfm]}>{children}</ReactMarkdown>
     </div>

canvas/src/components/tabs/config/form-inputs.tsx

@@ -22,7 +22,6 @@ export interface ConfigData {
   // task_budget maps to output_config.task_budget.total (requires beta header task-budgets-2026-03-13)
   task_budget?: number;
   prompt_files: string[];
-  shared_context: string[];
   skills: string[];
   tools: string[];
   a2a: { port: number; streaming: boolean; push_notifications: boolean };
@@ -40,7 +39,6 @@ export const DEFAULT_CONFIG: ConfigData = {
   effort: "",
   task_budget: 0,
   prompt_files: [],
-  shared_context: [],
   skills: [],
   tools: [],
   a2a: { port: 8000, streaming: true, push_notifications: true },

canvas/src/components/tabs/config/yaml-utils.ts

@@ -120,7 +120,6 @@ export function toYaml(config: ConfigData): string {
   if (config.effort) { lines.push(""); simple("effort", config.effort); }
   if (config.task_budget && config.task_budget > 0) { simple("task_budget", config.task_budget); }
   if (config.prompt_files?.length) { lines.push(""); list("prompt_files", config.prompt_files); }
-  if (config.shared_context?.length) { lines.push(""); list("shared_context", config.shared_context); }
   lines.push(""); list("skills", config.skills);
   if (config.tools?.length) { list("tools", config.tools); }
   lines.push(""); obj("a2a", config.a2a as unknown as Record<string, unknown>);

canvas/src/lib/design-tokens.ts

@@ -5,6 +5,13 @@ export const STATUS_CONFIG: Record<string, { dot: string; glow: string; label: s
   degraded: { dot: "bg-amber-400", glow: "shadow-amber-400/50", label: "Degraded", bar: "from-amber-500/20 to-transparent" },
   failed: { dot: "bg-red-400", glow: "shadow-red-400/50", label: "Failed", bar: "from-red-500/20 to-transparent" },
   provisioning: { dot: "bg-sky-400 motion-safe:animate-pulse", glow: "shadow-sky-400/50", label: "Starting", bar: "from-sky-500/20 to-transparent" },
+  // not_configured: derived state from agent_card.configuration_status (PR #2756 chain).
+  // Workspace is reachable (heartbeating, /agent-card serves) but adapter.setup()
+  // failed — typically a missing/rotated LLM credential. Amber to differentiate from
+  // online (green) and failed (red) — the workspace itself is healthy, just needs
+  // configuration. Hover renders agent_card.configuration_error in the tooltip so
+  // the operator sees the exact env var to set.
+  not_configured: { dot: "bg-amber-300", glow: "shadow-amber-300/50", label: "Not configured", bar: "from-amber-400/20 to-transparent" },
 };
 
 export function statusDotClass(status: string): string {
@@ -12,10 +19,10 @@ export function statusDotClass(status: string): string {
 }
 
 export const TIER_CONFIG: Record<number, { label: string; color: string; border: string }> = {
-  1: { label: "T1", color: "text-ink-soft bg-surface-card/80", border: "text-ink-mid border-line/60" },
-  2: { label: "T2", color: "text-sky-400 bg-sky-950/50", border: "text-sky-400 border-sky-500/30" },
-  3: { label: "T3", color: "text-violet-400 bg-violet-950/50", border: "text-violet-400 border-violet-500/30" },
-  4: { label: "T4", color: "text-warm bg-amber-950/50", border: "text-warm border-amber-500/30" },
+  1: { label: "T1", color: "text-ink-mid bg-surface-card border border-line", border: "text-ink-mid border-line" },
+  2: { label: "T2", color: "text-white bg-accent border border-accent-strong", border: "text-accent border-accent" },
+  3: { label: "T3", color: "text-white bg-violet-600 border border-violet-700", border: "text-violet-600 border-violet-500" },
+  4: { label: "T4", color: "text-white bg-warm border border-warm", border: "text-warm border-warm" },
 };
 
 export const COMM_TYPE_LABELS: Record<string, string> = {

canvas/src/lib/tenant.ts

@@ -59,8 +59,8 @@ export function getTenantSlug(): string {
  * isSaaSTenant reports whether the canvas is running as the UI for a
  * SaaS tenant (served at <slug>.moleculesai.app). Use for client-side
  * UX branches that should behave differently on SaaS vs self-hosted —
- * e.g. the workspace tier picker hides T1/T2 sandbox tiers because every
- * SaaS workspace gets its own EC2 VM (inherently T3 Full Access).
+ * e.g. the workspace tier picker hides T1/T2/T3 sandbox tiers because
+ * every SaaS workspace gets its own EC2 VM (inherently T4 Full Access).
  *
  * SSR-safe: returns false on the server to avoid hydration drift; call
  * sites should tolerate a flip from false→true on first client render.

canvas/src/store/__tests__/canvas-topology-configuration-status.test.ts

@@ -0,0 +1,103 @@
+import { describe, it, expect } from "vitest";
+import {
+  getConfigurationStatus,
+  getConfigurationError,
+} from "../canvas-topology";
+
+// Tests for the getConfigurationStatus / getConfigurationError helpers
+// (issue #467 / PR #2756 chain). Surfacing the workspace's
+// `agent_card.configuration_status` is the user-visible payoff of
+// PR #2756's decoupling — without it, a misconfigured workspace looks
+// identical to a healthy one in the canvas tile.
+
+describe("getConfigurationStatus", () => {
+  it("returns null when agentCard is null", () => {
+    expect(getConfigurationStatus(null)).toBe(null);
+  });
+
+  it("returns null when agentCard has no configuration_status", () => {
+    expect(getConfigurationStatus({ name: "x" })).toBe(null);
+  });
+
+  it("returns 'ready' when agent reports configuration ok", () => {
+    expect(
+      getConfigurationStatus({ configuration_status: "ready" }),
+    ).toBe("ready");
+  });
+
+  it("returns 'not_configured' when agent reports setup failed", () => {
+    expect(
+      getConfigurationStatus({ configuration_status: "not_configured" }),
+    ).toBe("not_configured");
+  });
+
+  it("ignores unknown values defensively", () => {
+    // A future agent reporting a status string we don't yet recognise
+    // shouldn't crash the canvas — we treat it as 'no info' (null).
+    expect(
+      getConfigurationStatus({ configuration_status: "starting" }),
+    ).toBe(null);
+    expect(
+      getConfigurationStatus({ configuration_status: 42 }),
+    ).toBe(null);
+    expect(
+      getConfigurationStatus({ configuration_status: null }),
+    ).toBe(null);
+  });
+});
+
+describe("getConfigurationError", () => {
+  it("returns null when agentCard is null", () => {
+    expect(getConfigurationError(null)).toBe(null);
+  });
+
+  it("returns null when status is 'ready' even if error string present", () => {
+    // Defensive: if the agent somehow ships configuration_status=ready
+    // alongside a stale configuration_error from a previous boot, we
+    // trust the live status flag and don't surface the stale error.
+    expect(
+      getConfigurationError({
+        configuration_status: "ready",
+        configuration_error: "stale: was unset",
+      }),
+    ).toBe(null);
+  });
+
+  it("returns the error string when status is 'not_configured'", () => {
+    expect(
+      getConfigurationError({
+        configuration_status: "not_configured",
+        configuration_error:
+          "RuntimeError: Neither OPENAI_API_KEY nor MINIMAX_API_KEY is set",
+      }),
+    ).toBe(
+      "RuntimeError: Neither OPENAI_API_KEY nor MINIMAX_API_KEY is set",
+    );
+  });
+
+  it("returns null when status is 'not_configured' but error is missing", () => {
+    expect(
+      getConfigurationError({ configuration_status: "not_configured" }),
+    ).toBe(null);
+  });
+
+  it("returns null when error is empty string", () => {
+    // Empty string isn't actionable for the operator — treat same as
+    // missing.
+    expect(
+      getConfigurationError({
+        configuration_status: "not_configured",
+        configuration_error: "",
+      }),
+    ).toBe(null);
+  });
+
+  it("returns null when error is non-string", () => {
+    expect(
+      getConfigurationError({
+        configuration_status: "not_configured",
+        configuration_error: { reason: "object" },
+      }),
+    ).toBe(null);
+  });
+});

canvas/src/store/canvas-topology.ts

@@ -564,3 +564,42 @@ export function extractSkillNames(agentCard: Record<string, unknown> | null): st
     .map((skill: Record<string, unknown>) => String(skill.name || skill.id || ""))
     .filter(Boolean);
 }
+
+/**
+ * Returns the configuration status reported by the workspace, or null
+ * when the agent card doesn't carry one (older runtime, or pre-PR #2756
+ * worker).
+ *
+ * Pairs with molecule-core PR #2756: when adapter.setup() fails, the
+ * runtime mounts a not-configured handler AND advertises the failure
+ * via agent_card.configuration_status = "not_configured" +
+ * configuration_error = "<reason>". Canvas reads both to render a
+ * "needs config" tile instead of a confused "online but silent" state.
+ *
+ * Returns null (not undefined) so callers can distinguish "no info"
+ * from explicit values via a strict equality check.
+ */
+export function getConfigurationStatus(
+  agentCard: Record<string, unknown> | null,
+): "ready" | "not_configured" | null {
+  if (!agentCard) return null;
+  const raw = agentCard.configuration_status;
+  if (raw === "ready" || raw === "not_configured") return raw;
+  return null;
+}
+
+/**
+ * Returns the configuration error string from the agent card when
+ * configuration_status is "not_configured", or null otherwise.
+ *
+ * Already redacted server-side via secret_redactor (PR #2778) — safe to
+ * render in the UI verbatim.
+ */
+export function getConfigurationError(
+  agentCard: Record<string, unknown> | null,
+): string | null {
+  if (!agentCard) return null;
+  if (getConfigurationStatus(agentCard) !== "not_configured") return null;
+  const raw = agentCard.configuration_error;
+  return typeof raw === "string" && raw.length > 0 ? raw : null;
+}

docs/agent-runtime/config-format.md

@@ -27,11 +27,11 @@ prompt_files:
 # AGENTS.md-style example:
 #   prompt_files: [AGENTS.md]
 
-# Files to share with direct children (1-level inheritance)
-# Children fetch these at startup via GET /workspaces/:id/shared-context
-shared_context:
-  - architecture.md
-  - conventions.md
+# NOTE: `shared_context` (parent → child file injection at boot) was removed.
+# To share knowledge across a team, use memory v2's team:<id> namespace via
+# the recall_memory MCP tool — the agent pulls it on demand instead of
+# paying for it at every boot. For large blob-shaped artefacts, see RFC
+# #2789 (platform-owned shared file storage).
 
 # Skills to load -- folder names under skills/
 skills:
@@ -123,7 +123,6 @@ env:
 | `runtime` | No | Adapter to use: `langgraph` (default), `claude-code`, `crewai`, `autogen`, `deepagents`, `openclaw`. See [Agent Runtime Adapters](./cli-runtime.md). |
 | `model` | Yes | LangChain-compatible provider string (e.g. `anthropic:claude-sonnet-4-6`). Overridden by `MODEL_PROVIDER` env var if set. |
 | `prompt_files` | No | Ordered list of markdown files to load as system prompt. Defaults to `["system-prompt.md"]` if omitted. `MEMORY.md` and `USER.md` are auto-appended when present so frozen memory snapshots do not need to be duplicated here. Supports any agent framework's file structure (OpenClaw, Claude Code, etc.) |
-| `shared_context` | No | Files from this workspace's config dir to share with direct children. Children fetch these at startup and inject into their system prompt as `## Parent Context`. 1-level inheritance only (grandchildren don't see grandparent's context). |
 | `skills` | Yes | List of skill folder names to load from `skills/` |
 | `tools` | No | Built-in tools from workspace-template |
 | `memory` | No | Memory backend config (defaults to filesystem) |
@@ -157,7 +156,6 @@ The file watcher monitors the entire config directory. When `config.yaml` change
 | `name`, `description`, `version` | Yes | Rebuild Agent Card with new metadata |
 | `a2a` | **No** | Port and protocol changes require container restart |
 | `delegation` | Yes | Retry/timeout defaults take effect on next delegation call |
-| `shared_context` | Yes | Children fetch on next prompt rebuild; no restart needed |
 | `sub_workspaces` | **No** | Team structure changes go through `POST /workspaces/:id/expand` |
 
 See [Skills — Live Reload](./skills.md#live-reload) for the full file watcher flow.

docs/agent-runtime/system-prompt-structure.md

@@ -24,21 +24,19 @@ When you receive a task, break it into sub-tasks and delegate to your team.
 Always review work before reporting completion to the caller.
 ```
 
-### 2. Parent Context (if child workspace)
+### 2. Team-shared knowledge (on demand)
 
-If this workspace was created via team expansion (has a `PARENT_ID` env var), it fetches its parent's shared context files at startup via `GET /workspaces/{parent_id}/shared-context`. The parent declares which files to share in its `config.yaml`:
+Team-scoped knowledge is no longer injected at boot. The previous
+`shared_context` field + `GET /workspaces/{parent_id}/shared-context`
+fetch was removed; agents now pull team-shared knowledge on demand via
+memory v2's `team:<id>` namespace using the `recall_memory` MCP tool.
 
-```yaml
-shared_context:
-  - architecture.md
-  - conventions.md
-```
-
-These files are injected as a `## Parent Context` section, with each file rendered under a `### {filename}` heading. This gives children the parent's project knowledge (architecture, conventions, API schemas) without exposing the parent's system prompt or full config.
-
-**1-level inheritance only:** A grandchild sees its direct parent's shared context, not its grandparent's. This mirrors the L2 Team Memory scope.
-
-**Graceful degradation:** If the parent is offline or the endpoint returns an error, the child starts normally without parent context.
+This shifts cost from "every boot, always" to "only when the agent
+asks", and lets team members write to the shared store from anywhere
+that can resolve the namespace (canvas Memory tab, agent
+`commit_memory`, admin import). For large blob-shaped artefacts (full
+architecture docs, brand assets, PDFs) see RFC #2789 (platform-owned
+shared file storage).
 
 ### 3. Skill Instructions
 

docs/api-protocol/memory-plugin-v1.yaml

@@ -0,0 +1,358 @@
+openapi: 3.0.3
+info:
+  title: Molecule Memory Plugin v1
+  version: 1.0.0
+  description: |
+    Contract between workspace-server and a memory backend plugin. The
+    plugin owns its own storage; workspace-server is the security
+    perimeter (secret redaction, namespace ACL, GLOBAL audit/wrap).
+
+    Defined in RFC #2728. See docs/rfc/memory-v2-rationale.md for design
+    rationale.
+
+    Auth: none. Plugins MUST be reachable only on a private network or
+    unix socket — workspace-server is the only sanctioned client.
+servers:
+  - url: http://localhost:9100
+    description: Built-in postgres-backed plugin (default)
+
+paths:
+  /v1/health:
+    get:
+      summary: Liveness + capability probe
+      operationId: getHealth
+      responses:
+        '200':
+          description: Plugin healthy
+          content:
+            application/json:
+              schema: { $ref: '#/components/schemas/HealthResponse' }
+        '503':
+          description: Plugin unhealthy (e.g., backing store down)
+          content:
+            application/json:
+              schema: { $ref: '#/components/schemas/Error' }
+
+  /v1/namespaces/{name}:
+    parameters:
+      - $ref: '#/components/parameters/NamespaceName'
+    put:
+      summary: Upsert a namespace (idempotent)
+      operationId: upsertNamespace
+      requestBody:
+        required: true
+        content:
+          application/json:
+            schema: { $ref: '#/components/schemas/NamespaceUpsert' }
+      responses:
+        '200': { $ref: '#/components/responses/Namespace' }
+        '400': { $ref: '#/components/responses/BadRequest' }
+    patch:
+      summary: Update namespace metadata or TTL
+      operationId: patchNamespace
+      requestBody:
+        required: true
+        content:
+          application/json:
+            schema: { $ref: '#/components/schemas/NamespacePatch' }
+      responses:
+        '200': { $ref: '#/components/responses/Namespace' }
+        '404': { $ref: '#/components/responses/NotFound' }
+    delete:
+      summary: Delete namespace and all its memories (operator action)
+      operationId: deleteNamespace
+      responses:
+        '204':
+          description: Deleted
+        '404': { $ref: '#/components/responses/NotFound' }
+
+  /v1/namespaces/{name}/memories:
+    parameters:
+      - $ref: '#/components/parameters/NamespaceName'
+    post:
+      summary: Write a memory to a namespace
+      description: |
+        `content` MUST already be secret-redacted by the workspace-server.
+        Plugin does not run additional redaction.
+      operationId: commitMemory
+      requestBody:
+        required: true
+        content:
+          application/json:
+            schema: { $ref: '#/components/schemas/MemoryWrite' }
+      responses:
+        '201':
+          description: Memory persisted
+          content:
+            application/json:
+              schema: { $ref: '#/components/schemas/MemoryWriteResponse' }
+        '400': { $ref: '#/components/responses/BadRequest' }
+        '404': { $ref: '#/components/responses/NotFound' }
+
+  /v1/search:
+    post:
+      summary: Search memories across one or more namespaces
+      description: |
+        workspace-server MUST intersect the requested `namespaces` with
+        the caller's currently-readable set BEFORE invoking this
+        endpoint. The plugin treats the list as authoritative.
+      operationId: searchMemories
+      requestBody:
+        required: true
+        content:
+          application/json:
+            schema: { $ref: '#/components/schemas/SearchRequest' }
+      responses:
+        '200':
+          description: Search results
+          content:
+            application/json:
+              schema: { $ref: '#/components/schemas/SearchResponse' }
+        '400': { $ref: '#/components/responses/BadRequest' }
+
+  /v1/memories/{id}:
+    parameters:
+      - in: path
+        name: id
+        required: true
+        schema: { type: string, format: uuid }
+    delete:
+      summary: Forget a memory by id
+      description: |
+        `requested_by_namespace` is the namespace the caller has write
+        access to; the plugin SHOULD reject if the memory doesn't belong
+        to that namespace.
+      operationId: forgetMemory
+      requestBody:
+        required: true
+        content:
+          application/json:
+            schema: { $ref: '#/components/schemas/ForgetRequest' }
+      responses:
+        '204':
+          description: Forgotten
+        '403': { $ref: '#/components/responses/Forbidden' }
+        '404': { $ref: '#/components/responses/NotFound' }
+
+components:
+  parameters:
+    NamespaceName:
+      in: path
+      name: name
+      required: true
+      schema:
+        type: string
+        minLength: 1
+        maxLength: 256
+        pattern: '^[a-z]+:[A-Za-z0-9_:.\-]+$'
+        example: 'workspace:550e8400-e29b-41d4-a716-446655440000'
+
+  responses:
+    Namespace:
+      description: Namespace state
+      content:
+        application/json:
+          schema: { $ref: '#/components/schemas/Namespace' }
+    BadRequest:
+      description: Invalid input
+      content:
+        application/json:
+          schema: { $ref: '#/components/schemas/Error' }
+    NotFound:
+      description: Resource not found
+      content:
+        application/json:
+          schema: { $ref: '#/components/schemas/Error' }
+    Forbidden:
+      description: Caller lacks write access to the requested namespace
+      content:
+        application/json:
+          schema: { $ref: '#/components/schemas/Error' }
+
+  schemas:
+    HealthResponse:
+      type: object
+      required: [status, version, capabilities]
+      properties:
+        status: { type: string, enum: [ok, degraded] }
+        version: { type: string, example: "1.0.0" }
+        capabilities:
+          type: array
+          items:
+            type: string
+            enum: [embedding, fts, ttl, pin, propagation]
+          description: |
+            Optional features this plugin supports. workspace-server
+            adapts MCP responses based on this list (e.g., agents can
+            request semantic search only when `embedding` is present).
+
+    NamespaceKind:
+      type: string
+      enum: [workspace, team, org, custom]
+
+    Namespace:
+      type: object
+      required: [name, kind, created_at]
+      properties:
+        name: { type: string }
+        kind: { $ref: '#/components/schemas/NamespaceKind' }
+        expires_at:
+          type: string
+          format: date-time
+          nullable: true
+        metadata:
+          type: object
+          additionalProperties: true
+          nullable: true
+        created_at: { type: string, format: date-time }
+
+    NamespaceUpsert:
+      type: object
+      required: [kind]
+      properties:
+        kind: { $ref: '#/components/schemas/NamespaceKind' }
+        expires_at: { type: string, format: date-time, nullable: true }
+        metadata:
+          type: object
+          additionalProperties: true
+          nullable: true
+
+    NamespacePatch:
+      type: object
+      properties:
+        expires_at: { type: string, format: date-time, nullable: true }
+        metadata:
+          type: object
+          additionalProperties: true
+          nullable: true
+
+    MemoryKind:
+      type: string
+      enum: [fact, summary, checkpoint]
+
+    MemorySource:
+      type: string
+      enum: [agent, runtime, user]
+
+    MemoryWrite:
+      type: object
+      required: [content, kind, source]
+      properties:
+        id:
+          type: string
+          format: uuid
+          nullable: true
+          description: |
+            Optional idempotency key. When supplied, the plugin MUST
+            treat the write as upsert keyed on this id (re-running
+            the same write does not duplicate). When omitted, the
+            plugin generates a fresh UUID. Used by the backfill CLI.
+        content:
+          type: string
+          minLength: 1
+          description: Already secret-redacted by workspace-server.
+        kind: { $ref: '#/components/schemas/MemoryKind' }
+        source: { $ref: '#/components/schemas/MemorySource' }
+        expires_at: { type: string, format: date-time, nullable: true }
+        propagation:
+          type: object
+          additionalProperties: true
+          nullable: true
+          description: |
+            Opaque metadata the plugin stores and returns. Reserved for
+            future cross-namespace propagation semantics.
+        pin: { type: boolean, default: false }
+        embedding:
+          type: array
+          items: { type: number }
+          nullable: true
+          description: |
+            Optional pre-computed embedding. Plugins reporting the
+            `embedding` capability MAY ignore this and recompute.
+
+    MemoryWriteResponse:
+      type: object
+      required: [id, namespace]
+      properties:
+        id: { type: string, format: uuid }
+        namespace: { type: string }
+
+    Memory:
+      type: object
+      required: [id, namespace, content, kind, source, created_at]
+      properties:
+        id: { type: string, format: uuid }
+        namespace: { type: string }
+        content: { type: string }
+        kind: { $ref: '#/components/schemas/MemoryKind' }
+        source: { $ref: '#/components/schemas/MemorySource' }
+        expires_at: { type: string, format: date-time, nullable: true }
+        propagation:
+          type: object
+          additionalProperties: true
+          nullable: true
+        pin: { type: boolean }
+        created_at: { type: string, format: date-time }
+        score:
+          type: number
+          nullable: true
+          description: Relevance score from search (semantic + FTS).
+
+    SearchRequest:
+      type: object
+      required: [namespaces]
+      properties:
+        namespaces:
+          type: array
+          items: { type: string }
+          minItems: 1
+          description: |
+            Already intersected with the caller's readable set by
+            workspace-server.
+        query: { type: string }
+        kinds:
+          type: array
+          items: { $ref: '#/components/schemas/MemoryKind' }
+        limit:
+          type: integer
+          minimum: 1
+          maximum: 100
+          default: 20
+        embedding:
+          type: array
+          items: { type: number }
+          nullable: true
+
+    SearchResponse:
+      type: object
+      required: [memories]
+      properties:
+        memories:
+          type: array
+          items: { $ref: '#/components/schemas/Memory' }
+
+    ForgetRequest:
+      type: object
+      required: [requested_by_namespace]
+      properties:
+        requested_by_namespace:
+          type: string
+          description: Namespace the caller has write access to.
+
+    Error:
+      type: object
+      required: [code, message]
+      properties:
+        code:
+          type: string
+          enum:
+            - bad_request
+            - not_found
+            - forbidden
+            - internal
+            - unavailable
+        message: { type: string }
+        details:
+          type: object
+          additionalProperties: true
+          nullable: true

docs/api-protocol/platform-api.md

@@ -199,7 +199,6 @@ Install safeguards bound the cost of a single install (env-tunable via `PLUGIN_I
 | `GET` | `/templates` | List available templates. **Requires AdminAuth** (PR #701). |
 | `GET` | `/org/templates` | List available org templates. **Requires AdminAuth** (PR #701). |
 | `POST` | `/templates/import` | Import an agent folder as a new template |
-| `GET` | `/workspaces/:id/shared-context` | Read parent shared-context files |
 | `GET` | `/workspaces/:id/files` | List files under an allowed root |
 | `GET` | `/workspaces/:id/files/*path` | Read a file |
 | `PUT` | `/workspaces/:id/files/*path` | Write a file |

docs/api-reference.md

@@ -68,7 +68,6 @@ Full contract: `docs/runbooks/admin-auth.md`.
 | GET | /channels/adapters | channels.go (list available platforms) |
 | POST | /channels/discover | channels.go (auto-detect chats for a bot token) |
 | POST | /webhooks/:type | channels.go (incoming social webhook) |
-| GET | /workspaces/:id/shared-context | templates.go |
 | GET/PUT/DELETE | /workspaces/:id/files[/*path] | templates.go |
 | GET | /canvas/viewport | viewport.go — open, no auth required (cosmetic, bootstrap-friendly) |
 | PUT | /canvas/viewport | viewport.go — `CanvasOrBearer` middleware; accepts bearer OR Origin matching `CORS_ORIGINS`. Cosmetic-only route — worst case viewport corruption, recovered by page refresh. |

docs/architecture/molecule-technical-doc.md

@@ -523,7 +523,8 @@ runtime_config:                            # Runtime-specific settings
 skills: ["skill1", "skill2"]               # Folder names under skills/
 tools: ["web_search", "filesystem"]        # Built-in tool names
 prompt_files: ["system-prompt.md"]         # Additional prompt text files
-shared_context: []                         # Files from parent workspace
+# `shared_context` was removed; team-shared knowledge now lives in memory v2's
+# team:<id> namespace (recall_memory MCP tool). See RFC #2789 for shared files.
 
 a2a:
   port: 8000

docs/memory-plugins/CHANGELOG.md

@@ -0,0 +1,113 @@
+# Memory Plugin Contract — Changelog
+
+Every breaking or operationally-relevant change to the v1 plugin
+contract or the workspace-server-side wiring lands here. Plugin
+authors should subscribe to PRs touching this file.
+
+## [Unreleased] — fixup wave 1 (post-RFC-#2728 self-review)
+
+A self-review of the initial 11-PR rollout (PRs #2729-#2742) flagged
+two correctness bugs and three operational hazards. This wave fixes
+all of them. Order matches operator-impact severity.
+
+### Critical: backfill idempotency via `MemoryWrite.id` (#2744)
+
+**The bug.** The backfill CLI claimed idempotent on re-run, but
+`gen_random_uuid()` in the plugin's INSERT meant every retry created
+a fresh row. Operators retrying a failed `-apply` would silently
+double their memory count.
+
+**The fix.** Optional `id` field on `MemoryWrite`. When supplied,
+plugins MUST upsert. The backfill now forwards `agent_memories.id`
+to `MemoryWrite.id`, so retries update in place.
+
+**Plugin author action.** If your plugin uses
+`INSERT INTO ... DEFAULT gen_random_uuid()`, switch to
+`INSERT ... ON CONFLICT (id) DO UPDATE` when `id` is set. The wire
+contract is forward-compatible — plugins that ignore the field still
+work for production agent commits (which leave `id` empty), but they
+will silently corrupt backfill retries.
+
+### Critical: `memory-backfill -verify` mode (#2747)
+
+**The miss.** The original PR-7 task spec called for a parity-check
+mode but it never landed. Operators had no way to confirm a
+migration succeeded short of "no errors logged."
+
+**The fix.** New `-verify` flag samples N workspaces, queries
+`agent_memories` direct, runs an equivalent plugin search via the
+namespace resolver, multiset-compares contents. Reports mismatches
+to stdout and exits non-zero so CI can gate the cutover.
+
+```bash
+memory-backfill -verify                        # default sample 50
+memory-backfill -verify -verify-sample=200     # bigger
+memory-backfill -verify -workspace=<uuid>      # one workspace
+```
+
+### Important: `expires_at` validation (#2746)
+
+**The bug.** `commit_memory_v2` silently dropped malformed
+`expires_at` strings. Agent passes `expires_at: "tomorrow"`, gets a
+200, memory has no TTL — agent thinks it set a TTL, didn't.
+
+**The fix.** Returns
+`fmt.Errorf("invalid expires_at: must be RFC3339")` on parse
+failure. Plugin is not called in this case.
+
+**Plugin author action.** None — this is a workspace-server-side
+fix. But: if your plugin advertises the `ttl` capability, make sure
+you actually evict expired rows on read (not just on a janitor cron
+that runs once a day). The harness in `testing-your-plugin.md` has
+a TTL-eviction test you should run.
+
+### Important: audit log JSON via `json.Marshal` (#2746)
+
+**The bug.** `auditOrgWrite` built `activity_logs.metadata` via
+`fmt.Sprintf` with `%q`. For ASCII (today's UUID + hex digest) this
+coincidentally produces valid JSON; for unicode or control bytes it
+silently produces non-JSON.
+
+**The fix.** Replaced with `json.Marshal(map[string]string{...})`.
+Same wire shape today, won't regress when metadata grows.
+
+**Plugin author action.** None — workspace-server-internal.
+
+### Operator action: staging verification (#292)
+
+**Status.** Tracked as task #292. PR-merged ≠ verified. Operator
+must:
+1. Provision a staging tenant, set `MEMORY_PLUGIN_URL`
+2. Run real `commit_memory_v2` from a workspace
+3. `memory-backfill -dry-run` against staging data
+4. `memory-backfill -apply`, then `-verify`
+5. Set `MEMORY_V2_CUTOVER=true`, verify admin export still works
+6. Run a legacy `commit_memory` from a workspace, verify it lands
+   in plugin storage via the PR-6 shim
+
+### Other follow-ups still open
+
+- **#289**: admin export O(workspaces) → O(namespaces) — N+1 pattern
+  in `exportViaPlugin` (1000-workspace tenants run 1000× resolver
+  CTEs + 1000× plugin searches today).
+- **#291**: workspace deletion must call `DELETE
+  /v1/namespaces/{name}` — orphans accumulate today.
+- **#293**: real-subprocess boot E2E — current PR-11 is integration
+  (httptest + sqlmock), not E2E.
+
+These are tracked but deferred; they're operationally annoying, not
+incident-shaped.
+
+## [v1.0.0] — initial release (RFC #2728, PRs #2729-#2742)
+
+Initial plugin contract + 11-PR rollout. See
+[issue #2728](https://github.com/Molecule-AI/molecule-core/issues/2728)
+for the full RFC.
+
+Endpoints: `/v1/health`, `/v1/namespaces/{name}` (PUT/PATCH/DELETE),
+`/v1/namespaces/{name}/memories` (POST), `/v1/search` (POST),
+`/v1/memories/{id}` (DELETE).
+
+Capabilities: `embedding`, `fts`, `ttl`, `pin`, `propagation`.
+
+Operator runbook: see [README.md § Replacing the built-in plugin](README.md#replacing-the-built-in-plugin).

docs/memory-plugins/README.md

@@ -0,0 +1,191 @@
+# Writing a Memory Plugin
+
+This document is for operators and ecosystem authors who want to
+replace the built-in postgres-backed memory plugin (the default
+implementation that ships with workspace-server) with their own.
+
+The contract was introduced by RFC #2728. The shipped binary is
+`cmd/memory-plugin-postgres/`; reading its source is the fastest way
+to see a complete reference implementation.
+
+## What the contract is
+
+The plugin is an HTTP server that workspace-server talks to via the
+OpenAPI v1 spec at [`docs/api-protocol/memory-plugin-v1.yaml`](../api-protocol/memory-plugin-v1.yaml).
+
+Six endpoints:
+
+| Endpoint | Method | Purpose |
+|---|---|---|
+| `/v1/health` | GET | Liveness probe + capability list |
+| `/v1/namespaces/{name}` | PUT | Idempotent upsert |
+| `/v1/namespaces/{name}` | PATCH | Update TTL or metadata |
+| `/v1/namespaces/{name}` | DELETE | Remove namespace and its memories |
+| `/v1/namespaces/{name}/memories` | POST | Write a memory |
+| `/v1/search` | POST | Multi-namespace search |
+| `/v1/memories/{id}` | DELETE | Forget a memory |
+
+The wire types are defined in
+`workspace-server/internal/memory/contract/contract.go`. Run-time
+validation is built into the Go bindings via `Validate()` methods —
+your plugin SHOULD perform equivalent validation.
+
+## What workspace-server takes care of
+
+You do **not** implement these in the plugin; workspace-server is the
+security perimeter:
+
+- **Secret redaction** (SAFE-T1201). All `content` you receive is
+  already scrubbed. Don't run additional redaction; it's pointless.
+- **Namespace ACL**. workspace-server intersects the caller's
+  readable namespaces against the requested list before sending you
+  the search request. The list you receive is authoritative.
+- **GLOBAL audit**. Org-namespace writes are recorded in
+  `activity_logs` server-side; you don't see them.
+- **Prompt-injection wrap**. Org memories returned to agents get a
+  `[MEMORY id=... scope=ORG ns=...]:` prefix added at the
+  workspace-server layer. Your `content` field is plain text.
+
+## What you implement
+
+- Storage of `memory_namespaces` and `memory_records` (or whatever
+  shape you want — Pinecone vectors, an in-memory map, etc.)
+- The 7 endpoints above with the request/response shapes the spec
+  defines
+- `/v1/health` reporting your supported capabilities (see below)
+- Idempotency on namespace upsert (PUT semantics, not POST)
+- Idempotency on memory commit when `MemoryWrite.id` is supplied
+  (see "Memory idempotency" below)
+
+## Memory idempotency
+
+`MemoryWrite.id` is optional. Two contracts to honor:
+
+| Caller passes | Plugin MUST |
+|---|---|
+| `id` omitted | Generate a fresh UUID, return it in the response |
+| `id` set | Upsert keyed on this id — if a row with that id already exists, UPDATE it in place rather than inserting a duplicate |
+
+The backfill CLI (`memory-backfill`) relies on the upsert behavior
+so retries don't duplicate rows. Production agent commits leave `id`
+empty and rely on the plugin's UUID generator — the hot path is
+unchanged.
+
+The built-in postgres plugin implements this with `INSERT ... ON
+CONFLICT (id) DO UPDATE`. A vector-DB plugin (e.g., Pinecone) would
+use the database's native upsert primitive on the same id.
+
+## Capability negotiation
+
+Your `/v1/health` response declares what features you support:
+
+```json
+{
+  "status": "ok",
+  "version": "1.0.0",
+  "capabilities": ["embedding", "fts", "ttl", "pin", "propagation"]
+}
+```
+
+| Capability | What it gates |
+|---|---|
+| `embedding` | Agents may ask for semantic search; you receive `embedding: [...]` in search bodies |
+| `fts` | Agents may pass a query string; you decide how to match (FTS, ILIKE, regex) |
+| `ttl` | Agents may set `expires_at`; you must not return expired rows |
+| `pin` | Agents may set `pin: true`; you should rank pinned rows first |
+| `propagation` | Agents may set `propagation: {...}`; you must store it as opaque JSON and return it on read |
+
+A capability you DON'T list is fine — workspace-server adapts the MCP
+tool surface to match. E.g., a Pinecone-only plugin that lists only
+`embedding` will silently ignore agents' `query` strings.
+
+## Deployment models
+
+Three common shapes:
+
+1. **Same machine, different process**: workspace-server boots, then
+   `MEMORY_PLUGIN_URL=http://localhost:9100` points at your plugin
+   running on a unix socket or localhost port. This is what the
+   built-in postgres plugin does.
+
+2. **Separate container**: deploy your plugin as its own service on
+   the private network. Set `MEMORY_PLUGIN_URL` to its DNS name.
+
+3. **Self-managed**: customer-owned plugin running on customer-owned
+   infrastructure, accessed over a tunnel. Same env-var wiring.
+
+Auth is **none** — the plugin must be reachable only on a private
+network. workspace-server is the only sanctioned client.
+
+## Replacing the built-in plugin
+
+This is the canonical operator runbook for swapping the default
+plugin out. The same sequence applies whether you're swapping for
+another postgres plugin variant, Pinecone, Letta, or a custom
+implementation.
+
+1. **Stand up the new plugin.** Deploy the binary/container, confirm
+   it boots, confirm `/v1/health` returns `ok` with the capability
+   list you expect.
+
+2. **Run the backfill in dry-run mode** to scope the migration:
+   ```bash
+   DATABASE_URL=postgres://... \
+   MEMORY_PLUGIN_URL=http://your-plugin:9100 \
+   memory-backfill -dry-run
+   ```
+   Reports row count + namespace mapping per workspace, no writes.
+
+3. **Apply the backfill:**
+   ```bash
+   memory-backfill -apply
+   ```
+   Idempotent on retry — the backfill passes each `agent_memories.id`
+   to `MemoryWrite.id`, so partial-then-full re-runs upsert in place.
+
+4. **Verify parity** before flipping the cutover flag:
+   ```bash
+   memory-backfill -verify -verify-sample=200
+   ```
+   Random-samples N workspaces, diffs `agent_memories` direct query
+   against plugin search via the workspace's readable namespaces.
+   Reports mismatches and exits non-zero if any are found — wire
+   into your CI to gate the cutover.
+
+5. **Flip the cutover flag.** Set `MEMORY_V2_CUTOVER=true` on
+   workspace-server and restart. Admin export/import now route
+   through the plugin; legacy `agent_memories` becomes read-only.
+
+6. **Existing data in the old plugin's tables is NOT auto-dropped.**
+   Deliberate safety property — operator drops manually after the
+   ~60-day grace window. If you switch back later, old data comes
+   back into use (no loss).
+
+If `-verify` reports mismatches, do NOT set `MEMORY_V2_CUTOVER` —
+inspect the output, re-run `-apply` to backfill missing rows (it
+upserts, so this is safe), and re-verify.
+
+## Worked examples
+
+- [`pinecone-example/`](pinecone-example/) — full Pinecone-backed plugin
+- [`testing-your-plugin.md`](testing-your-plugin.md) — running the
+  contract test harness against your implementation
+
+## When to write one vs. fork the default
+
+Fork the default postgres plugin if:
+- You want different SQL (Materialized views? Different vector index?)
+- You want extra auth on top
+- You want server-side metrics emission
+
+Write a fresh plugin if:
+- The storage backend is fundamentally different (vector DB, KV store,
+  in-memory, file-based)
+- You're integrating an existing memory service (Letta, Mem0, etc.)
+
+## See also
+
+- [`CHANGELOG.md`](CHANGELOG.md) — contract revisions and fixup waves
+- RFC #2728 — design rationale
+- [`cmd/memory-plugin-postgres/`](../../workspace-server/cmd/memory-plugin-postgres/) — reference implementation
+- [`docs/api-protocol/memory-plugin-v1.yaml`](../api-protocol/memory-plugin-v1.yaml) — full OpenAPI spec

docs/memory-plugins/pinecone-example/README.md

@@ -0,0 +1,124 @@
+# Pinecone-backed Memory Plugin (worked example)
+
+A working sketch of a memory plugin that delegates storage to
+[Pinecone](https://www.pinecone.io/) instead of postgres.
+
+This is **example code, not a production binary**. It demonstrates
+how to map the v1 contract onto a vector database. Operators who
+want to ship this would harden auth, add retries, batch the
+commit path, etc.
+
+## Why Pinecone is interesting
+
+The default postgres plugin's pgvector index works for ~10M memories
+on a single node. Beyond that, semantic search becomes painful. A
+managed vector database can handle 1B+ memories, but the trade-offs
+are different:
+
+- **Capabilities**: Pinecone is great at `embedding` (its core
+  feature) but has no first-class FTS. So the plugin reports
+  `["embedding"]` and ignores the `query` field.
+- **TTL**: Pinecone supports per-vector metadata with deletion via
+  metadata filter — TTL becomes a periodic janitor task, not a
+  per-row property.
+- **Cost**: per-vector billing, so the plugin should batch writes
+  and dedup before posting.
+
+## Wire mapping
+
+| Contract field | Pinecone shape |
+|---|---|
+| `namespace` | `namespace` (Pinecone's first-class concept) |
+| `id` (caller-supplied) | `id` (Pinecone vector id; plugin upserts on this) |
+| `id` (omitted) | Plugin generates `uuid.NewString()` before upsert |
+| `content` | metadata.text |
+| `embedding` | `values` |
+| `kind` / `source` / `pin` / `expires_at` | `metadata.{kind, source, pin, expires_at}` |
+| `propagation` (opaque JSON) | `metadata.propagation` (also opaque) |
+
+The contract's `expires_at` becomes a metadata field; a separate
+janitor cron periodically queries `expires_at < now` and deletes.
+
+Pinecone's native upsert is the right fit for the idempotency-key
+contract: passing the same `id` twice updates in place. So a
+Pinecone plugin gets idempotent backfill retries "for free" if it
+just forwards `MemoryWrite.id` (or its generated UUID) to the
+upsert call.
+
+## Skeleton
+
+```go
+package main
+
+import (
+    "context"
+    "encoding/json"
+    "log"
+    "net/http"
+    "os"
+
+    "github.com/pinecone-io/go-pinecone/pinecone"
+)
+
+type pineconePlugin struct {
+    client *pinecone.Client
+    index  string
+}
+
+func main() {
+    apiKey := os.Getenv("PINECONE_API_KEY")
+    if apiKey == "" {
+        log.Fatal("PINECONE_API_KEY required")
+    }
+    client, err := pinecone.NewClient(pinecone.NewClientParams{ApiKey: apiKey})
+    if err != nil {
+        log.Fatal(err)
+    }
+    p := &pineconePlugin{client: client, index: os.Getenv("PINECONE_INDEX")}
+
+    http.HandleFunc("/v1/health", p.health)
+    http.HandleFunc("/v1/search", p.search)
+    // ... rest of the routes ...
+
+    log.Fatal(http.ListenAndServe(":9100", nil))
+}
+
+func (p *pineconePlugin) health(w http.ResponseWriter, r *http.Request) {
+    w.Header().Set("Content-Type", "application/json")
+    json.NewEncoder(w).Encode(map[string]interface{}{
+        "status":       "ok",
+        "version":      "1.0.0",
+        "capabilities": []string{"embedding"}, // no FTS, no TTL out-of-box
+    })
+}
+
+func (p *pineconePlugin) search(w http.ResponseWriter, r *http.Request) {
+    // Parse contract.SearchRequest
+    // Build Pinecone QueryByVectorValuesRequest with body.Embedding
+    // For each Pinecone namespace in body.Namespaces, call Query
+    // Map results to contract.Memory
+    // ...
+}
+```
+
+## What's missing from this sketch
+
+A production-ready Pinecone plugin would add:
+
+- **Batch commits**: bulk upsert N memories in a single Pinecone call
+- **TTL janitor**: periodic deletion of expired vectors
+- **Connection pooling**: keep one Pinecone client alive across requests
+- **Retry + circuit breaker**: Pinecone occasionally returns 5xx
+- **Metrics**: latency histograms per endpoint, write/read counters
+- **Idempotency-key handling**: when `MemoryWrite.id` is supplied,
+  forward it as the Pinecone vector id verbatim; otherwise generate
+  one. Pinecone's `Upsert` is naturally idempotent on id match.
+
+But the mapping above is the load-bearing part — the rest is
+operational hardening, not contract-specific.
+
+## See also
+
+- [Pinecone Go SDK docs](https://docs.pinecone.io/reference/go-sdk)
+- [Memory plugin contract spec](../../api-protocol/memory-plugin-v1.yaml)
+- [Default postgres plugin source](../../../workspace-server/cmd/memory-plugin-postgres/) — for comparison

docs/memory-plugins/testing-your-plugin.md

@@ -0,0 +1,181 @@
+# Testing Your Memory Plugin
+
+Once you have a plugin implementing the v1 contract, you can validate
+it against the spec without booting workspace-server.
+
+## The contract test harness
+
+Workspace-server ships typed Go bindings + round-trip tests in
+`workspace-server/internal/memory/contract/`. The simplest way to
+gain confidence in your plugin's wire compatibility is to point those
+tests at it.
+
+A minimal contract suite:
+
+```go
+package myplugin_test
+
+import (
+    "context"
+    "testing"
+
+    mclient "github.com/Molecule-AI/molecule-monorepo/platform/internal/memory/client"
+    "github.com/Molecule-AI/molecule-monorepo/platform/internal/memory/contract"
+)
+
+func TestMyPlugin_FullRoundTrip(t *testing.T) {
+    // Start your plugin somehow (subprocess, in-process, etc.)
+    pluginURL := startMyPlugin(t)
+    cl := mclient.New(mclient.Config{BaseURL: pluginURL})
+
+    // 1. Health
+    hr, err := cl.Boot(context.Background())
+    if err != nil {
+        t.Fatalf("Boot: %v", err)
+    }
+    if hr.Status != "ok" {
+        t.Errorf("status = %q", hr.Status)
+    }
+
+    // 2. Namespace upsert
+    if _, err := cl.UpsertNamespace(context.Background(), "workspace:test-1",
+        contract.NamespaceUpsert{Kind: contract.NamespaceKindWorkspace}); err != nil {
+        t.Fatalf("UpsertNamespace: %v", err)
+    }
+
+    // 3. Commit memory
+    resp, err := cl.CommitMemory(context.Background(), "workspace:test-1",
+        contract.MemoryWrite{
+            Content: "hello",
+            Kind:    contract.MemoryKindFact,
+            Source:  contract.MemorySourceAgent,
+        })
+    if err != nil {
+        t.Fatalf("CommitMemory: %v", err)
+    }
+    if resp.ID == "" {
+        t.Errorf("plugin must return a non-empty memory id")
+    }
+
+    // 4. Search
+    sresp, err := cl.Search(context.Background(), contract.SearchRequest{
+        Namespaces: []string{"workspace:test-1"},
+        Query:      "hello",
+    })
+    if err != nil {
+        t.Fatalf("Search: %v", err)
+    }
+    if len(sresp.Memories) == 0 {
+        t.Errorf("plugin returned no memories for the query we just wrote")
+    }
+
+    // 5. Forget
+    if err := cl.ForgetMemory(context.Background(), resp.ID,
+        contract.ForgetRequest{RequestedByNamespace: "workspace:test-1"}); err != nil {
+        t.Errorf("ForgetMemory: %v", err)
+    }
+}
+```
+
+## Testing idempotency
+
+The contract requires that `MemoryWrite.id`, when supplied, behaves
+as an upsert key. The backfill CLI relies on this — without it,
+operator retries silently duplicate every memory.
+
+```go
+func TestMyPlugin_IDIsIdempotencyKey(t *testing.T) {
+    pluginURL := startMyPlugin(t)
+    cl := mclient.New(mclient.Config{BaseURL: pluginURL})
+    if _, err := cl.UpsertNamespace(context.Background(), "workspace:test-1",
+        contract.NamespaceUpsert{Kind: contract.NamespaceKindWorkspace}); err != nil {
+        t.Fatal(err)
+    }
+
+    fixedID := "11111111-2222-3333-4444-555555555555"
+
+    // First write with a specific id.
+    resp1, err := cl.CommitMemory(context.Background(), "workspace:test-1",
+        contract.MemoryWrite{
+            ID:      fixedID,
+            Content: "first version",
+            Kind:    contract.MemoryKindFact,
+            Source:  contract.MemorySourceAgent,
+        })
+    if err != nil {
+        t.Fatalf("first commit: %v", err)
+    }
+    if resp1.ID != fixedID {
+        t.Errorf("plugin must echo the supplied id, got %q", resp1.ID)
+    }
+
+    // Second write with the same id — must update, not insert.
+    if _, err := cl.CommitMemory(context.Background(), "workspace:test-1",
+        contract.MemoryWrite{
+            ID:      fixedID,
+            Content: "second version (updated)",
+            Kind:    contract.MemoryKindFact,
+            Source:  contract.MemorySourceAgent,
+        }); err != nil {
+        t.Fatalf("second commit: %v", err)
+    }
+
+    // Search must return exactly one row, with the updated content.
+    sresp, _ := cl.Search(context.Background(), contract.SearchRequest{
+        Namespaces: []string{"workspace:test-1"},
+    })
+    matches := 0
+    for _, m := range sresp.Memories {
+        if m.ID == fixedID {
+            matches++
+            if m.Content != "second version (updated)" {
+                t.Errorf("upsert didn't update content: got %q", m.Content)
+            }
+        }
+    }
+    if matches != 1 {
+        t.Errorf("upsert produced %d rows for id=%s, want 1", matches, fixedID)
+    }
+}
+```
+
+## What the harness does NOT cover
+
+- **Capability accuracy**: if you list `embedding` you must actually
+  do semantic search. The harness can't tell you whether ranking is
+  meaningful — only that you don't crash.
+- **TTL eviction**: write a memory with `expires_at` 1 second in the
+  future, sleep 2 seconds, search — assert the memory is gone.
+- **Concurrency**: hit your plugin with 100 parallel writes; assert
+  no IDs collide.
+- **Recovery**: kill your plugin's storage backend, send a request,
+  assert your plugin returns 503 (not 200 with stale data).
+- **Backfill compatibility**: run the operator backfill against your
+  plugin twice in a row (`memory-backfill -apply`); assert the row
+  count doesn't double. The idempotency test above verifies the unit
+  contract; this checks the operational integration.
+- **Verify-mode parity**: after a backfill, run `memory-backfill
+  -verify`; assert it reports zero mismatches against
+  `agent_memories`.
+
+## Smoke test against workspace-server
+
+Once unit-level wire tests pass, run a real workspace-server with your
+plugin URL:
+
+```bash
+DATABASE_URL=postgres://... \
+MEMORY_PLUGIN_URL=http://localhost:9100 \
+./workspace-server
+```
+
+Then ask an agent to call `commit_memory_v2` and `search_memory`. If
+both round-trip cleanly, you're done.
+
+For the full E2E flow (including the namespace resolver, MCP layer,
+and security perimeter), see [PR-11's plugin-swap test](../../workspace-server/test/e2e/memory_plugin_swap_test.go).
+
+## Reporting bugs
+
+If you find a contract ambiguity or missing edge case, file an issue
+against `Molecule-AI/molecule-core` referencing RFC #2728.

manifest.json

@@ -28,7 +28,12 @@
     {"name": "claude-code-default", "repo": "Molecule-AI/molecule-ai-workspace-template-claude-code", "ref": "main"},
     {"name": "hermes", "repo": "Molecule-AI/molecule-ai-workspace-template-hermes", "ref": "main"},
     {"name": "openclaw", "repo": "Molecule-AI/molecule-ai-workspace-template-openclaw", "ref": "main"},
-    {"name": "codex", "repo": "Molecule-AI/molecule-ai-workspace-template-codex", "ref": "main"}
+    {"name": "codex", "repo": "Molecule-AI/molecule-ai-workspace-template-codex", "ref": "main"},
+    {"name": "langgraph", "repo": "Molecule-AI/molecule-ai-workspace-template-langgraph", "ref": "main"},
+    {"name": "crewai", "repo": "Molecule-AI/molecule-ai-workspace-template-crewai", "ref": "main"},
+    {"name": "autogen", "repo": "Molecule-AI/molecule-ai-workspace-template-autogen", "ref": "main"},
+    {"name": "deepagents", "repo": "Molecule-AI/molecule-ai-workspace-template-deepagents", "ref": "main"},
+    {"name": "gemini-cli", "repo": "Molecule-AI/molecule-ai-workspace-template-gemini-cli", "ref": "main"}
   ],
   "org_templates": [
     {"name": "molecule-dev", "repo": "Molecule-AI/molecule-ai-org-template-molecule-dev", "ref": "main"},

scripts/build_runtime_package.py

@@ -58,6 +58,8 @@ TOP_LEVEL_MODULES = {
     "adapter_base",
     "agent",
     "agents_md",
+    "boot_routes",
+    "card_helpers",
     "config",
     "configs_dir",
     "consolidation",
@@ -73,12 +75,14 @@ TOP_LEVEL_MODULES = {
     "main",
     "mcp_cli",
     "molecule_ai_status",
+    "not_configured_handler",
     "platform_auth",
     "platform_inbound_auth",
     "plugins",
     "preflight",
     "prompt",
     "runtime_wedge",
+    "secret_redactor",
     "shared_runtime",
     "smoke_mode",
     "transcript_auth",

tests/e2e/lib/model_slug.sh

@@ -0,0 +1,51 @@
+#!/usr/bin/env bash
+# Per-runtime model slug dispatch for E2E provisioning.
+#
+# Different runtimes parse the model slug differently (PR #2571 incident,
+# 2026-05-03):
+#
+#   hermes      → "openai/gpt-4o"  (slash-form: derive-provider.sh splits
+#                                    on the prefix to set
+#                                    HERMES_INFERENCE_PROVIDER. Bare
+#                                    "gpt-4o" falls through to Anthropic
+#                                    default + 401, see PR #1714.)
+#
+#   langgraph   → "openai:gpt-4o"  (colon-form: langchain init_chat_model
+#                                    requires "<provider>:<model>".
+#                                    Slash-form was misinterpreted as
+#                                    OpenRouter routing → fell through
+#                                    without auth, surfaced 2026-05-03
+#                                    after the a2a-sdk v1 contract bugs
+#                                    PR #2558+#2563+#2567 cleared the
+#                                    masking layers.)
+#
+#   claude-code → "sonnet"         (entry-id form: claude-code template's
+#                                    config.yaml uses bare model names,
+#                                    auth comes via CLAUDE_CODE_OAUTH_TOKEN
+#                                    or ANTHROPIC_API_KEY rather than the
+#                                    slug.)
+#
+# When E2E_MODEL_SLUG is set, it overrides this dispatch — useful when an
+# operator dispatches the workflow to test a specific slug.
+#
+# Unit tested by tests/e2e/test_model_slug.sh — every branch must stay
+# pinned because regressions silently mask as "Could not resolve
+# authentication method" + the synth-E2E gate goes red without naming
+# the slug-format mismatch.
+
+# Usage: pick_model_slug <runtime>
+#   stdout: the slug string
+#   E2E_MODEL_SLUG (env): if set + non-empty, used as-is (operator override)
+pick_model_slug() {
+  local runtime="${1:-}"
+  if [ -n "${E2E_MODEL_SLUG:-}" ]; then
+    printf '%s' "$E2E_MODEL_SLUG"
+    return 0
+  fi
+  case "$runtime" in
+    hermes)      printf 'openai/gpt-4o' ;;
+    langgraph)   printf 'openai:gpt-4o' ;;
+    claude-code) printf 'sonnet' ;;
+    *)           printf 'openai/gpt-4o' ;;  # safest fallback (matches hermes)
+  esac
+}

tests/e2e/test_model_slug.sh

@@ -0,0 +1,90 @@
+#!/usr/bin/env bash
+# Regression test for tests/e2e/lib/model_slug.sh.
+#
+# PR #2571 fixed a synth-E2E masking bug where MODEL_SLUG was hardcoded
+# to "openai/gpt-4o" (slash-form) but langgraph's init_chat_model needs
+# "openai:gpt-4o" (colon-form). Fix shipped as a per-runtime case
+# statement. Without this regression test, dropping any branch of the
+# case (or flipping a slug format) would silently revert behavior — the
+# E2E only fails as "Could not resolve authentication method" at the
+# very first message, after a successful tenant + workspace provision.
+#
+# Each branch must FAIL the test if the dispatch behavior changes, not
+# just produce some non-empty string.
+set -uo pipefail
+
+# Resolve to the lib relative to this test file so the test runs from
+# any cwd (CI, local invocation, repo root).
+SCRIPT_DIR="$(cd "$(dirname "${BASH_SOURCE[0]}")" && pwd)"
+# shellcheck source=lib/model_slug.sh
+source "$SCRIPT_DIR/lib/model_slug.sh"
+
+PASS=0
+FAIL=0
+
+assert_eq() {
+  local label="$1" got="$2" want="$3"
+  if [ "$got" = "$want" ]; then
+    echo "  ✓ $label"
+    PASS=$((PASS+1))
+  else
+    echo "  ✗ $label: got=$(printf %q "$got")  want=$(printf %q "$want")" >&2
+    FAIL=$((FAIL+1))
+  fi
+}
+
+run_test() {
+  local label="$1" runtime="$2" want="$3"
+  # Pin per-test isolation: explicitly unset the override so a leaked
+  # E2E_MODEL_SLUG from caller env can't poison the dispatch branches.
+  local got
+  got=$(unset E2E_MODEL_SLUG; pick_model_slug "$runtime")
+  assert_eq "$label" "$got" "$want"
+}
+
+echo "Test: pick_model_slug — per-runtime dispatch"
+echo
+
+# ── Per-runtime branches (the load-bearing ones for synth-E2E) ──
+run_test "hermes → slash-form (derive-provider.sh contract)"       hermes      "openai/gpt-4o"
+run_test "langgraph → colon-form (init_chat_model contract)"       langgraph   "openai:gpt-4o"
+run_test "claude-code → bare model name (entry-id form)"           claude-code "sonnet"
+
+# ── Fallback for unknown runtime ──
+# Picks slash-form (hermes-shaped) since hermes is the historical
+# default and most third-party runtimes behave hermes-like. Pinning
+# this so a future "smarter" fallback (e.g., empty string, error) is
+# a deliberate choice, not silent drift.
+run_test "unknown runtime → slash-form fallback"                   gemini      "openai/gpt-4o"
+run_test "empty runtime → slash-form fallback"                     ""          "openai/gpt-4o"
+
+# ── Override via E2E_MODEL_SLUG ──
+# When the operator sets E2E_MODEL_SLUG, the per-runtime dispatch is
+# bypassed. Used during workflow_dispatch to A/B specific slugs.
+echo
+echo "Test: pick_model_slug — E2E_MODEL_SLUG override"
+echo
+
+got=$(E2E_MODEL_SLUG="anthropic:claude-opus-4-7" pick_model_slug langgraph)
+assert_eq "override beats langgraph default"                      "$got" "anthropic:claude-opus-4-7"
+
+got=$(E2E_MODEL_SLUG="custom/whatever" pick_model_slug hermes)
+assert_eq "override beats hermes default"                         "$got" "custom/whatever"
+
+got=$(E2E_MODEL_SLUG="some-bare-id" pick_model_slug claude-code)
+assert_eq "override beats claude-code default"                    "$got" "some-bare-id"
+
+# Empty-string override does NOT activate (falls through to dispatch).
+# This is the historical bash idiom: -n "" → false → no override. Pin
+# it because changing this behavior (e.g. via -v test) would silently
+# break the dispatch when an operator passes "" to clear an inherited
+# env var.
+got=$(E2E_MODEL_SLUG="" pick_model_slug langgraph)
+assert_eq "empty-string override falls through to dispatch"       "$got" "openai:gpt-4o"
+
+echo
+echo "─────────────────────────────────────────────────"
+echo "PASSED: $PASS"
+echo "FAILED: $FAIL"
+echo "─────────────────────────────────────────────────"
+[ "$FAIL" -eq 0 ]

tests/e2e/test_secrets_dispatch.sh

@@ -0,0 +1,145 @@
+#!/usr/bin/env bash
+# Regression test for the SECRETS_JSON branching in
+# tests/e2e/test_staging_full_saas.sh (lines ~322-368).
+#
+# The synth-E2E canary picks one of two LLM auth paths based on which
+# E2E_*_API_KEY is set. The branch order is load-bearing:
+#
+#   E2E_MINIMAX_API_KEY first   → claude-code MiniMax path (cheap canary
+#                                  default since 2026-05-03; routes via
+#                                  workspace-configs-templates/claude-
+#                                  code-default/config.yaml's `minimax`
+#                                  provider entry).
+#
+#   E2E_OPENAI_API_KEY second  → langgraph + hermes legacy path (kept
+#                                  as fallback for operator dispatches
+#                                  that need the OpenAI-shaped
+#                                  HERMES_CUSTOM_* env block).
+#
+# Without this gate, a future "tidy up the if/elif" refactor could
+# silently flip the precedence (OpenAI wins when both are set →
+# claude-code workspace boots without MINIMAX_API_KEY → 401 at first
+# turn → canary red without any signal that the wrong key shape was
+# selected). The 2026-05-03 OpenAI-quota incident took ~16h to
+# diagnose for exactly this class of "looks like an LLM problem,
+# was actually a wiring problem" failure.
+
+set -uo pipefail
+
+SCRIPT_DIR="$(cd "$(dirname "${BASH_SOURCE[0]}")" && pwd)"
+SAAS_SCRIPT="$SCRIPT_DIR/test_staging_full_saas.sh"
+
+if [ ! -f "$SAAS_SCRIPT" ]; then
+  echo "FATAL: cannot locate test_staging_full_saas.sh at $SAAS_SCRIPT" >&2
+  exit 2
+fi
+
+PASS=0
+FAIL=0
+
+assert_eq() {
+  local label="$1" got="$2" want="$3"
+  if [ "$got" = "$want" ]; then
+    echo "  ✓ $label"
+    PASS=$((PASS+1))
+  else
+    echo "  ✗ $label" >&2
+    echo "      got:  $got" >&2
+    echo "      want: $want" >&2
+    FAIL=$((FAIL+1))
+  fi
+}
+
+# Extract just the SECRETS_JSON block from the saas script and source
+# it into a sub-shell so we can run the branching logic in isolation.
+# Anchor on the comment header so a structural refactor that moves the
+# block fails this test loudly rather than silently sourcing nothing.
+extract_block() {
+  awk '
+    /^# ─── 5\. Provision parent workspace/ {capture=1; next}
+    capture && /^MODEL_SLUG=/ {exit}
+    capture {print}
+  ' "$SAAS_SCRIPT"
+}
+
+BLOCK=$(extract_block)
+if [ -z "$BLOCK" ]; then
+  echo "FATAL: SECRETS_JSON block not found in $SAAS_SCRIPT — refactor anchor changed?" >&2
+  exit 2
+fi
+
+# Run the extracted block in a clean env, capturing SECRETS_JSON.
+run_block() {
+  # Caller passes vars on the command line, e.g.
+  #   run_block E2E_MINIMAX_API_KEY=mx-test
+  env -i PATH="$PATH" "$@" bash -c "
+    set -uo pipefail
+    $BLOCK
+    echo \"\$SECRETS_JSON\"
+  " 2>/dev/null | tail -1
+}
+
+# Resolve a JSON key from the captured payload using python3 (already
+# a hard dep of the saas script). Returns empty string on missing key.
+get_json_key() {
+  local payload="$1" key="$2"
+  python3 -c "
+import json, sys
+p = json.loads(sys.argv[1])
+print(p.get(sys.argv[2], ''))
+" "$payload" "$key"
+}
+
+list_json_keys() {
+  python3 -c "
+import json, sys
+p = json.loads(sys.argv[1])
+print(','.join(sorted(p.keys())))
+" "$1"
+}
+
+echo "Test: SECRETS_JSON branching in test_staging_full_saas.sh"
+echo
+
+# ── Branch 1: MiniMax wins when set ──
+SECRETS_JSON=$(run_block E2E_MINIMAX_API_KEY=mx-test)
+assert_eq "MiniMax key set → MINIMAX_API_KEY in payload" \
+  "$(get_json_key "$SECRETS_JSON" MINIMAX_API_KEY)" "mx-test"
+assert_eq "MiniMax-only payload contains exactly MINIMAX_API_KEY" \
+  "$(list_json_keys "$SECRETS_JSON")" "MINIMAX_API_KEY"
+
+# ── Branch 1 precedence: MiniMax beats OpenAI when both set ──
+# Critical: the 2026-05-03 incident shape was "two paths exist, wrong
+# one wins". The bash if/elif must keep MiniMax above OpenAI so the
+# claude-code default canary doesn't accidentally use the (more
+# expensive, quota-burnt) OpenAI key.
+SECRETS_JSON=$(run_block E2E_MINIMAX_API_KEY=mx-priority E2E_OPENAI_API_KEY=oai-loser)
+assert_eq "Both keys set → MiniMax wins" \
+  "$(get_json_key "$SECRETS_JSON" MINIMAX_API_KEY)" "mx-priority"
+assert_eq "Both keys set → OpenAI block NOT emitted" \
+  "$(get_json_key "$SECRETS_JSON" OPENAI_API_KEY)" ""
+assert_eq "Both keys set → no HERMES_* leakage from OpenAI branch" \
+  "$(get_json_key "$SECRETS_JSON" HERMES_INFERENCE_PROVIDER)" ""
+
+# ── Branch 2: OpenAI used when MiniMax absent ──
+SECRETS_JSON=$(run_block E2E_OPENAI_API_KEY=oai-test)
+assert_eq "Only OpenAI set → OPENAI_API_KEY in payload" \
+  "$(get_json_key "$SECRETS_JSON" OPENAI_API_KEY)" "oai-test"
+assert_eq "Only OpenAI set → HERMES_CUSTOM_API_KEY mirrors OpenAI key" \
+  "$(get_json_key "$SECRETS_JSON" HERMES_CUSTOM_API_KEY)" "oai-test"
+assert_eq "Only OpenAI set → MODEL_PROVIDER pinned to colon-form" \
+  "$(get_json_key "$SECRETS_JSON" MODEL_PROVIDER)" "openai:gpt-4o"
+assert_eq "Only OpenAI set → MINIMAX_API_KEY NOT emitted" \
+  "$(get_json_key "$SECRETS_JSON" MINIMAX_API_KEY)" ""
+
+# ── No keys: empty payload ──
+SECRETS_JSON=$(run_block)
+assert_eq "No keys set → SECRETS_JSON is empty object" \
+  "$SECRETS_JSON" "{}"
+
+echo
+echo "─────────────────────────────────────────────────"
+echo "PASSED: $PASS"
+echo "FAILED: $FAIL"
+echo "─────────────────────────────────────────────────"
+[ "$FAIL" -eq 0 ]

tests/e2e/test_staging_full_saas.sh

@@ -67,6 +67,12 @@ log()  { echo "[$(date +%H:%M:%S)] $*"; }
 fail() { echo "[$(date +%H:%M:%S)] ❌ $*" >&2; exit 1; }
 ok()   { echo "[$(date +%H:%M:%S)] ✅ $*"; }
 
+# Per-runtime model slug dispatch — see lib/model_slug.sh for the rationale.
+# Extracted so unit tests (tests/e2e/test_model_slug.sh) can pin every branch
+# without booting the full 11-step lifecycle.
+# shellcheck source=lib/model_slug.sh
+source "$(dirname "$0")/lib/model_slug.sh"
+
 CURL_COMMON=(-sS --fail-with-body --max-time 30)
 
 # ─── cleanup trap ───────────────────────────────────────────────────────
@@ -314,29 +320,68 @@ tenant_call() {
 }
 
 # ─── 5. Provision parent workspace ─────────────────────────────────────
-# Runtimes like hermes crash at boot with "No provider API key found"
-# if nothing in the standard env-var list is set. Inject the API key
-# from E2E_OPENAI_API_KEY so the runtime can actually start — it's
-# per-workspace secret, so it's persisted as a workspace_secret and
-# materialized into the container env. Missing key falls through to
-# an empty secrets map; workspace will still fail but the error is
-# expected and actionable.
+# Inject the LLM provider key so the runtime can authenticate at boot.
+# Branch by which secret is set so the script supports multiple paths
+# without forcing every dispatch to ship them all. Priority order
+# matters — first non-empty wins:
+#
+#   E2E_MINIMAX_API_KEY → claude-code MiniMax path. Cheapest, default
+#     for the cron canary post-2026-05-03. Routes via the claude-code
+#     template's `minimax` provider (workspace-configs-templates/
+#     claude-code-default/config.yaml:64-69) which sets
+#     ANTHROPIC_BASE_URL=https://api.minimax.io/anthropic at boot.
+#     MINIMAX_API_KEY is the vendor-specific env name the adapter
+#     reads (PR #244 — per-vendor envs prevent ANTHROPIC_AUTH_TOKEN
+#     collisions when a user runs MiniMax + Z.ai workspaces side-by-
+#     side).
+#
+#   E2E_ANTHROPIC_API_KEY → claude-code direct-Anthropic path (added
+#     2026-05-04 after #2578 left the operator with an awkward choice
+#     between paying OpenAI's billing top-up and registering a new
+#     MiniMax account). Lower friction than MiniMax for operators
+#     who already have an Anthropic API key for their own Claude
+#     Code session. Pricier per-token than MiniMax but billing is
+#     still independent of MOLECULE_STAGING_OPENAI_KEY. Pinned to the
+#     claude-code runtime — hermes/langgraph use OpenAI-shaped envs.
+#
+#   E2E_OPENAI_API_KEY → langgraph + hermes paths. Kept as fallback
+#     for operator dispatches that explicitly want to exercise the
+#     OpenAI path. The HERMES_* fields pin hermes-agent's bridge to
+#     api.openai.com (template-hermes' derive-provider.sh otherwise
+#     resolves openai/* → openrouter.ai and 401s). MODEL_PROVIDER
+#     follows workspace/config.py:258's 'provider:model' format.
+#
+# All empty → '{}' (workspace will fail at first turn with an
+# expected, actionable auth error rather than masking the test).
 SECRETS_JSON='{}'
-if [ -n "${E2E_OPENAI_API_KEY:-}" ]; then
-  # MODEL_PROVIDER is a full model slug in 'provider:model' format per
-  # workspace/config.py:258. Using just "openai" gets parsed as the
-  # model name → 404 model_not_found. Also set OPENAI_BASE_URL to
-  # OpenAI's own endpoint — default is openrouter.ai which would need
-  # a different key format.
-  #
-  # The HERMES_* fields below bypass template-hermes/scripts/derive-provider.sh
-  # — verified 2026-04-24 that even with template-hermes#19's fix in main,
-  # staging tenants sometimes resolve openai/* to PROVIDER=openrouter and
-  # emit {'message':'Missing Authentication header','code':401} (OpenRouter's
-  # shape) in the A2A reply. Setting HERMES_INFERENCE_PROVIDER=custom +
-  # HERMES_CUSTOM_{BASE_URL,API_KEY,API_MODE} pins the bridge deterministically
-  # so the test doesn't depend on every tenant EC2 having a freshly-cloned
-  # template-hermes.
+if [ -n "${E2E_MINIMAX_API_KEY:-}" ]; then
+  SECRETS_JSON=$(python3 -c "
+import json, os
+k = os.environ['E2E_MINIMAX_API_KEY']
+print(json.dumps({
+    'MINIMAX_API_KEY': k,
+}))
+")
+elif [ -n "${E2E_ANTHROPIC_API_KEY:-}" ]; then
+  # Direct Anthropic path — claude-code adapter reads ANTHROPIC_API_KEY
+  # natively when ANTHROPIC_BASE_URL is unset. Useful for operators
+  # who already have an Anthropic API key (e.g. for their own Claude
+  # Code session) and want to avoid setting up a separate MiniMax
+  # account just for E2E. Pricier per-token than MiniMax but billing
+  # is still independent of MOLECULE_STAGING_OPENAI_KEY, so an OpenAI
+  # quota collapse doesn't wedge this path. Pinned to the claude-code
+  # runtime: hermes/langgraph use OpenAI-shaped envs and won't honour
+  # ANTHROPIC_API_KEY without further wiring (out of scope for this
+  # branch; if you need a hermes/Anthropic path, dispatch with
+  # E2E_RUNTIME=hermes + E2E_OPENAI_API_KEY pointing at a working key).
+  SECRETS_JSON=$(python3 -c "
+import json, os
+k = os.environ['E2E_ANTHROPIC_API_KEY']
+print(json.dumps({
+    'ANTHROPIC_API_KEY': k,
+}))
+")
+elif [ -n "${E2E_OPENAI_API_KEY:-}" ]; then
   SECRETS_JSON=$(python3 -c "
 import json, os
 k = os.environ['E2E_OPENAI_API_KEY']
@@ -352,15 +397,7 @@ print(json.dumps({
 ")
 fi
 
-# Model slug MUST be provider-prefixed for hermes — the template's
-# derive-provider.sh parses the slug prefix (`openai/…`, `anthropic/…`,
-# `minimax/…`) to set HERMES_INFERENCE_PROVIDER at install time. A bare
-# "gpt-4o" has no prefix → provider falls back to hermes auto-detect →
-# picks Anthropic default → tries Anthropic API with the OpenAI key →
-# 401 on A2A. Same trap that trapped prod users in PR #1714. We pin
-# "openai/gpt-4o" here because the E2E's secret is always the OpenAI
-# key; non-hermes runtimes ignore the prefix.
-MODEL_SLUG="openai/gpt-4o"
+MODEL_SLUG=$(pick_model_slug "$RUNTIME")
 
 log "5/11 Provisioning parent workspace (runtime=$RUNTIME)..."
 PARENT_RESP=$(tenant_call POST /workspaces \
@@ -431,6 +468,99 @@ for wid in $WS_TO_CHECK; do
   ok "    $wid online"
 done
 
+# ─── 7b. Canvas-terminal diagnose (EIC chain probe) ────────────────────
+# This step exists because the canvas-terminal failure of 2026-05-03
+# was structurally invisible to local-dev (handleLocalConnect uses
+# docker exec; handleRemoteConnect uses EIC + ssh). The CP provisioner
+# shipped without the tcp/22 EIC ingress rule for ~6 months and nobody
+# noticed until a paying tenant clicked Terminal in canvas. Probing the
+# diagnose endpoint here at synth-E2E time means a regression in
+#   - tenantIngressRules / workspaceIngressRules (CP)
+#   - eicSSHIngressRule helper (CP)
+#   - AuthorizeIngress source-group support (CP awsapi)
+#   - EIC_ENDPOINT_SG_ID Railway env
+#   - handleRemoteConnect's send-ssh-public-key/open-tunnel/ssh chain
+# surfaces within ~20 min of merge instead of waiting for a user report.
+#
+# The diagnose endpoint runs the full EIC + ssh probe from inside the
+# tenant's workspace-server (which already has AWS creds via its IAM
+# profile) and reports per-step status. We only need to call it as the
+# tenant — no AWS creds needed on the GHA runner. Returns
+# {"ok": bool, "first_failure": "name", "steps": [...]}.
+#
+# Local-docker workspaces (instance_id NULL) get diagnoseLocal which
+# probes docker.Ping + container exec; we still expect ok=true there
+# since local-docker is the alternative production path.
+log "7b/11 Canvas-terminal EIC diagnose probe..."
+for wid in $WS_TO_CHECK; do
+  DIAG_JSON=$(tenant_call GET "/workspaces/$wid/terminal/diagnose" 2>/dev/null || echo '{}')
+  DIAG_OK=$(echo "$DIAG_JSON" | python3 -c "import json,sys; d=json.load(sys.stdin); print('true' if d.get('ok') else 'false')" 2>/dev/null || echo "false")
+  if [ "$DIAG_OK" = "true" ]; then
+    ok "    $wid terminal-reachable (canvas terminal will work)"
+  else
+    DIAG_FAIL=$(echo "$DIAG_JSON" | python3 -c "import json,sys; d=json.load(sys.stdin); print(d.get('first_failure','unknown'))" 2>/dev/null || echo "unknown")
+    DIAG_DETAIL=$(echo "$DIAG_JSON" | python3 -c "import json,sys; d=json.load(sys.stdin); s=[x for x in d.get('steps',[]) if not x.get('ok')]; print(s[0].get('error','') if s else '')" 2>/dev/null || echo "")
+    fail "Workspace $wid terminal diagnose failed at step '$DIAG_FAIL': $DIAG_DETAIL — check tenant SG has tcp/22 from EIC endpoint SG (sg-0785d5c6138220523), EIC_ENDPOINT_SG_ID set in Railway, and EIC endpoint health"
+  fi
+done
+
+# ─── 7c. Workspace files API config.yaml round-trip ────────────────────
+# Pin the config-save path that drives the Canvas Config tab's Save &
+# Restart. Two failure classes this gate catches in one shot:
+#
+#   1. Path map drift (PR #2769). Runtime falls through to the wrong
+#      base path (e.g. /opt/configs when user-data only created /configs)
+#      → SSH `install -D` fails with EACCES on a parent dir that doesn't
+#      exist. The user-visible 500 was unobservable without exercising
+#      this code path on a fresh workspace.
+#   2. Permission drift on /configs. The path is root-owned by cloud-init,
+#      so the SSH-as-ubuntu install needs `sudo -n`. Any future change
+#      that drops the sudo, switches to a non-passwordless-sudo OS user,
+#      or moves the path to a non-ubuntu-writable dir without sudo will
+#      regress this gate.
+#
+# Round-trip: PUT a known marker, GET it back, assert content matches.
+# Marker shape includes the run id so a stale file from a prior canary
+# can't false-pass.
+log "7c/11 Files API config.yaml round-trip..."
+CONFIG_MARKER="# molecule-synth-e2e: ${E2E_RUN_ID:-unknown} ${RUNTIME} $(date -u +%Y-%m-%dT%H:%M:%SZ)"
+CONFIG_PAYLOAD="${CONFIG_MARKER}
+name: synth-canary
+runtime: ${RUNTIME}
+"
+for wid in $WS_TO_CHECK; do
+  PUT_BODY=$(python3 -c "import json,sys; print(json.dumps({'content': sys.stdin.read()}))" <<< "$CONFIG_PAYLOAD")
+  # Capture body to a tempfile so curl's -w '%{http_code}' is the only
+  # thing on stdout. The first version used `-w '\n%{http_code}\n'` and
+  # parsed via `tail -n 2 | head -n 1`, which broke because bash $(...)
+  # strips the trailing newline → only 2 lines remain in the captured
+  # value → head -n 1 returned the body, not the status code. Caught
+  # post-merge by E2E Staging SaaS at 22:06 UTC: a 200-with-body got
+  # misreported as "PUT returned <body>".
+  PUT_TMP=$(mktemp -t synth_put.XXXXXX)
+  PUT_CODE=$(tenant_call PUT "/workspaces/$wid/files/config.yaml" \
+    -H "Content-Type: application/json" \
+    -d "$PUT_BODY" \
+    -o "$PUT_TMP" \
+    -w '%{http_code}' \
+    2>/dev/null || echo "000")
+  PUT_BODY_OUT=$(cat "$PUT_TMP" 2>/dev/null || echo "")
+  rm -f "$PUT_TMP"
+  if [ "$PUT_CODE" != "200" ] && [ "$PUT_CODE" != "204" ]; then
+    fail "Workspace $wid Files API PUT config.yaml returned $PUT_CODE: $PUT_BODY_OUT — likely a path-map or permission regression in workspace-server template_files_eic.go"
+  fi
+  # PUT-only check; the GET-back round-trip assertion was dropped
+  # 2026-05-04 because PUT (template_files_eic.go SSH-via-EIC →
+  # workspace EC2) and GET (templates.go ReadFile → docker exec on
+  # platform-tenant-local container) hit DIFFERENT paths and DIFFERENT
+  # hosts. The asymmetry is a separate latent bug — Canvas Config tab
+  # rendering reads workspace state via other endpoints, not via this
+  # GET, so the user-facing Save & Restart works (container reads
+  # /configs/config.yaml directly via bind-mount). When the read/write
+  # paths are unified, restore the GET-back marker check here.
+  ok "    $wid config.yaml PUT OK (HTTP $PUT_CODE)"
+done
+
 # ─── 8. A2A round-trip on parent ───────────────────────────────────────
 log "8/11 Sending A2A message to parent — expecting agent response..."
 # Smoke prompt phrasing — DO NOT trim back to the bare "Reply with exactly: PONG"
@@ -461,7 +591,17 @@ print(json.dumps({
     }
 }))
 ")
+# Override CURL_COMMON's --max-time 30 for THIS call only. Each canary
+# creates a fresh org → workspace, so the A2A POST hits a cold model:
+# claude-code adapter starts its event loop, opens TLS to the LLM
+# endpoint, ships the first prompt, waits for first token. With MiniMax
+# (which is the canary default since #2710) cold-call latency
+# routinely exceeds 30s on the first request after workspace boot.
+# 90s gives ~3x headroom over observed cold-call P95 (~25-30s).
+# Subsequent A2A turns hit the same workspace and are sub-second, so
+# this only widens the window for step 8/11 of the canary's first turn.
 A2A_RESP=$(tenant_call POST "/workspaces/$PARENT_ID/a2a" \
+  --max-time 90 \
   -H "Content-Type: application/json" \
   -d "$A2A_PAYLOAD")
 AGENT_TEXT=$(echo "$A2A_RESP" | python3 -c "
@@ -483,6 +623,7 @@ fi
 #   "Encrypted content is not supported" → hermes codex_responses API misroute (#14)
 #   "Unknown provider"               → bridge misconfigured PROVIDER= (regression of #13 fix)
 #   "hermes-agent unreachable"       → gateway process died
+#   "exceeded your current quota"    → MOLECULE_STAGING_OPENAI_KEY billing (NOT a platform regression — #2578)
 #
 # Fail LOUD with the specific pattern so CI log + alert channel makes the
 # regression unambiguous.
@@ -508,6 +649,16 @@ fi
 if echo "$AGENT_TEXT" | grep -qF "Invalid API key"; then
   fail "A2A — REGRESSION: tenant auth chain returned 'Invalid API key'. Likely CP boot-event 401 race (CP #238) or stale OPENAI_API_KEY in the runtime env. Raw: $AGENT_TEXT"
 fi
+# Provider quota exhausted — distinguish from a platform regression so
+# the canary alert names the operator action directly instead of falling
+# through to the generic "error-shaped response" message. Steps 0-7 having
+# passed means the platform itself is healthy (CP up, tenant provisioned,
+# workspace online, A2A delivery end-to-end). When the agent comes back
+# with a provider-side 429, that is a billing event on the configured
+# OpenAI key, not a platform regression. Tracked in #2578.
+if echo "$AGENT_TEXT" | grep -qiE "exceeded your current quota|insufficient_quota"; then
+  fail "A2A — PROVIDER QUOTA EXHAUSTED (NOT a platform regression). Operator action: top up MOLECULE_STAGING_OPENAI_KEY billing or rotate to a higher-quota org at Settings → Secrets and Variables → Actions. Tracked in #2578. Raw: $AGENT_TEXT"
+fi
 # Generic catch-all — falls through if none of the known regressions hit.
 if echo "$AGENT_TEXT" | grep -qiE "error|exception"; then
   fail "A2A returned an error-shaped response: $AGENT_TEXT"
@@ -555,8 +706,80 @@ print(json.dumps({
 d=json.load(sys.stdin)
 print(len(d if isinstance(d, list) else d.get('events', [])))" 2>/dev/null || echo 0)
   log "    Activity events observed: $ACTIVITY_COUNT"
+
+  # ─── 9c. Workspace KV memory Edit round-trip ─────────────────────────
+  # Pins the Edit affordance added to the canvas Memory tab. The UI calls
+  # POST /workspaces/:id/memory with if_match_version, so the contract is:
+  #   1. initial POST creates row at version 1
+  #   2. GET returns version 1 + value
+  #   3. POST with if_match_version=1 updates → version 2
+  #   4. POST with if_match_version=1 again → 409 (optimistic-lock enforcement)
+  # Without (3) there is no Edit; without (4) two concurrent writers can
+  # silently overwrite each other and the agent loses delegation-ledger state.
+  log "9c.  Memory KV Edit round-trip (Edit affordance + 409 gate)"
+  EDIT_KEY="e2e_edit_gate_$SLUG"
+
+  # 1. seed
+  tenant_call POST "/workspaces/$PARENT_ID/memory" \
+    -H "Content-Type: application/json" \
+    -d "{\"key\":\"$EDIT_KEY\",\"value\":{\"step\":1}}" >/dev/null \
+    || fail "memory KV seed POST failed"
+
+  # 2. read back, capture version
+  EDIT_GET=$(tenant_call GET "/workspaces/$PARENT_ID/memory/$EDIT_KEY")
+  EDIT_VER=$(echo "$EDIT_GET" | python3 -c "import json,sys; print(json.load(sys.stdin)['version'])" 2>/dev/null || echo "")
+  [ -z "$EDIT_VER" ] && fail "memory KV GET missing version field. Body: ${EDIT_GET:0:200}"
+
+  # 3. conditional update with matching version
+  tenant_call POST "/workspaces/$PARENT_ID/memory" \
+    -H "Content-Type: application/json" \
+    -d "{\"key\":\"$EDIT_KEY\",\"value\":{\"step\":2},\"if_match_version\":$EDIT_VER}" >/dev/null \
+    || fail "memory KV conditional Edit failed (if_match_version=$EDIT_VER)"
+
+  # 4. value flipped + version incremented?
+  EDIT_GET2=$(tenant_call GET "/workspaces/$PARENT_ID/memory/$EDIT_KEY")
+  EDIT_VAL2=$(echo "$EDIT_GET2" | python3 -c "import json,sys; print(json.load(sys.stdin)['value'].get('step'))" 2>/dev/null || echo "")
+  [ "$EDIT_VAL2" = "2" ] || fail "memory KV Edit did not persist new value. Body: ${EDIT_GET2:0:200}"
+
+  # 5. stale-version POST must 409 — pin the optimistic-lock contract.
+  #
+  # tenant_call uses CURL_COMMON which carries --fail-with-body, so an
+  # expected-409 makes curl exit 22. The previous shape
+  #   $(tenant_call ... -w "%{http_code}" || echo "000")
+  # concatenated the captured "409" with the fallback "000" giving a
+  # bogus "409000" value (caught on PR #2792's first E2E run, which is
+  # also why staging-saas E2E has been silent-failing this gate since
+  # PR #2787 merged). Fix: route the status code into its own tempfile
+  # so curl's exit code can't pollute the captured stdout. set +e/-e
+  # keeps the 22 from tripping the outer `set -e` pipeline.
+  set +e
+  tenant_call POST "/workspaces/$PARENT_ID/memory" \
+    -H "Content-Type: application/json" \
+    -d "{\"key\":\"$EDIT_KEY\",\"value\":{\"step\":3},\"if_match_version\":$EDIT_VER}" \
+    -o /tmp/memory_stale_resp.txt -w "%{http_code}" >/tmp/memory_stale_code.txt 2>/dev/null
+  set -e
+  EDIT_STALE_CODE=$(cat /tmp/memory_stale_code.txt 2>/dev/null || echo "000")
+  [ "$EDIT_STALE_CODE" = "409" ] || fail "memory KV stale Edit must 409 (optimistic-lock). Got '$EDIT_STALE_CODE': $(cat /tmp/memory_stale_resp.txt 2>/dev/null | head -c 200)"
+
+  # cleanup
+  tenant_call DELETE "/workspaces/$PARENT_ID/memory/$EDIT_KEY" >/dev/null 2>&1 || true
+  ok "Memory KV Edit round-trip + 409 gate passed"
+
+  # ─── 9d. shared_context removal gate ─────────────────────────────────
+  # Pin the deletion of GET /workspaces/:id/shared-context. The route + handler
+  # were removed; team-shared knowledge now flows through memory v2's
+  # team:<id> namespace. If anyone re-introduces a shared-context endpoint
+  # without going through RFC #2789, this gate fires.
+  set +e
+  SC_CODE=$(tenant_call GET "/workspaces/$PARENT_ID/shared-context" \
+    -o /dev/null -w "%{http_code}" 2>/dev/null || echo "000")
+  set -e
+  if [ "$SC_CODE" = "200" ]; then
+    fail "shared-context route should be gone but returned 200 — regression. See task #304."
+  fi
+  ok "shared-context route confirmed removed (HTTP $SC_CODE)"
 else
-  log "9/11 Canary mode — skipping HMA / peers / activity"
+  log "9/11 Canary mode — skipping HMA / peers / activity / memory-edit / shared-context-gone"
 fi
 
 # ─── 10. Delegation mechanics (full mode + child) ──────────────────────

tests/harness/replays/peer-discovery-404.sh

@@ -75,9 +75,14 @@ from unittest.mock import AsyncMock, MagicMock, patch
 # Stub platform_auth so a2a_client imports cleanly without requiring a
 # real workspace token file. The helper's auth_headers() only matters
 # when going through the network; we're feeding it a mock response.
+#
+# Both stubs accept *args, **kwargs because the multi-workspace work
+# (#2739, #2743) added optional ``workspace_id`` parameters to
+# ``auth_headers`` and made ``self_source_headers`` 1-arg-required.
+# The stubs need to accept whatever the helpers pass without caring.
 _pa = types.ModuleType("platform_auth")
-_pa.auth_headers = lambda: {}
-_pa.self_source_headers = lambda: {}
+_pa.auth_headers = lambda *a, **kw: {}
+_pa.self_source_headers = lambda *a, **kw: {}
 sys.modules.setdefault("platform_auth", _pa)
 
 sys.path.insert(0, sys.argv[1])

workspace-server/cmd/memory-backfill/main.go

@@ -0,0 +1,305 @@
+// memory-backfill is a one-shot CLI that copies rows from the legacy
+// agent_memories table into the v2 plugin via its HTTP API.
+//
+// Idempotent on re-run: the backfill passes each source row's UUID
+// to the plugin's MemoryWrite.ID field, and the plugin upserts on
+// conflict. Re-running the backfill (whole or partial) updates rows
+// in place rather than duplicating.
+//
+// Usage:
+//   memory-backfill -dry-run                    # count + diff
+//   memory-backfill -apply                      # actually copy
+//   memory-backfill -apply -limit=10000         # cap rows per run
+//   memory-backfill -apply -workspace=<uuid>    # one workspace only
+//
+// Required env:
+//   DATABASE_URL                — workspace-server DB (read agent_memories)
+//   MEMORY_PLUGIN_URL           — target plugin (write memory_records)
+package main
+
+import (
+	"context"
+	"database/sql"
+	"errors"
+	"flag"
+	"fmt"
+	"log"
+	"os"
+	"strings"
+	"time"
+
+	_ "github.com/lib/pq"
+
+	mclient "github.com/Molecule-AI/molecule-monorepo/platform/internal/memory/client"
+	"github.com/Molecule-AI/molecule-monorepo/platform/internal/memory/contract"
+	"github.com/Molecule-AI/molecule-monorepo/platform/internal/memory/namespace"
+)
+
+const defaultLimit = 1000000 // effectively unlimited; cap keeps SQL pageable
+
+func main() {
+	if err := run(os.Args[1:], os.Stdout, os.Stderr); err != nil {
+		log.Fatalf("memory-backfill: %v", err)
+	}
+}
+
+// run is extracted so tests can drive it with synthesized argv +
+// captured stdout/stderr. Returns nil on success.
+func run(argv []string, stdout, stderr *os.File) error {
+	fs := flag.NewFlagSet("memory-backfill", flag.ContinueOnError)
+	fs.SetOutput(stderr)
+	dryRun := fs.Bool("dry-run", false, "count + diff only, no writes")
+	apply := fs.Bool("apply", false, "actually copy rows to the plugin")
+	verify := fs.Bool("verify", false, "post-apply parity check: random-sample N workspaces, diff agent_memories vs plugin search")
+	verifySample := fs.Int("verify-sample", 50, "number of workspaces to sample in -verify mode")
+	workspace := fs.String("workspace", "", "limit to a single workspace UUID (empty = all)")
+	limit := fs.Int("limit", defaultLimit, "max rows to process this run")
+	if err := fs.Parse(argv); err != nil {
+		return err
+	}
+	modesPicked := 0
+	if *dryRun {
+		modesPicked++
+	}
+	if *apply {
+		modesPicked++
+	}
+	if *verify {
+		modesPicked++
+	}
+	if modesPicked != 1 {
+		return errors.New("specify exactly one of -dry-run, -apply, or -verify")
+	}
+
+	dbURL := os.Getenv("DATABASE_URL")
+	if dbURL == "" {
+		return errors.New("DATABASE_URL is required")
+	}
+	pluginURL := os.Getenv("MEMORY_PLUGIN_URL")
+	if pluginURL == "" {
+		return errors.New("MEMORY_PLUGIN_URL is required")
+	}
+
+	db, err := sql.Open("postgres", dbURL)
+	if err != nil {
+		return fmt.Errorf("open db: %w", err)
+	}
+	defer db.Close()
+	ctx, cancel := context.WithTimeout(context.Background(), 5*time.Second)
+	defer cancel()
+	if err := db.PingContext(ctx); err != nil {
+		return fmt.Errorf("ping db: %w", err)
+	}
+
+	plugin := mclient.New(mclient.Config{BaseURL: pluginURL})
+	resolver := namespace.New(db)
+
+	if *verify {
+		vcfg := verifyConfig{
+			DB:          db,
+			Plugin:      plugin,
+			Resolver:    namespaceResolverAdapter{resolver},
+			SampleSize:  *verifySample,
+			WorkspaceID: *workspace,
+		}
+		report, err := verifyParity(context.Background(), vcfg, stdout)
+		if err != nil {
+			return err
+		}
+		fmt.Fprintf(stdout, "\nVerify complete: workspaces_sampled=%d matches=%d mismatches=%d errors=%d\n",
+			report.WorkspacesSampled, report.Matches, report.Mismatches, report.Errors)
+		if report.Mismatches > 0 || report.Errors > 0 {
+			return fmt.Errorf("verify found %d mismatches and %d errors", report.Mismatches, report.Errors)
+		}
+		return nil
+	}
+
+	cfg := backfillConfig{
+		DB:          db,
+		Plugin:      plugin,
+		Resolver:    resolver,
+		WorkspaceID: *workspace,
+		Limit:       *limit,
+		DryRun:      *dryRun,
+	}
+	stats, err := backfill(context.Background(), cfg, stdout)
+	if err != nil {
+		return err
+	}
+	fmt.Fprintf(stdout, "\nBackfill complete: scanned=%d copied=%d skipped=%d errors=%d\n",
+		stats.Scanned, stats.Copied, stats.Skipped, stats.Errors)
+	return nil
+}
+
+// backfillStats accumulates the counters the CLI reports.
+type backfillStats struct {
+	Scanned int
+	Copied  int
+	Skipped int
+	Errors  int
+}
+
+// backfillConfig is the typed dependency bundle. Tests inject stubs
+// for Plugin and Resolver; production wires real client + resolver.
+type backfillConfig struct {
+	DB          *sql.DB
+	Plugin      backfillPlugin
+	Resolver    backfillResolver
+	WorkspaceID string
+	Limit       int
+	DryRun      bool
+}
+
+// backfillPlugin is the slice of memory-plugin client we call.
+type backfillPlugin interface {
+	UpsertNamespace(ctx context.Context, name string, body contract.NamespaceUpsert) (*contract.Namespace, error)
+	CommitMemory(ctx context.Context, namespace string, body contract.MemoryWrite) (*contract.MemoryWriteResponse, error)
+}
+
+// backfillResolver lets the backfill compute namespace strings the
+// same way the live MCP layer does.
+type backfillResolver interface {
+	WritableNamespaces(ctx context.Context, workspaceID string) ([]namespace.Namespace, error)
+}
+
+// backfill is the workhorse. Iterates agent_memories, maps each row's
+// scope to a v2 namespace via the resolver, and POSTs to the plugin.
+// Returns final stats. Stops after Limit rows.
+func backfill(ctx context.Context, cfg backfillConfig, stdout *os.File) (*backfillStats, error) {
+	stats := &backfillStats{}
+
+	query := `
+		SELECT id, workspace_id, content, scope, created_at
+		FROM agent_memories
+	`
+	args := []interface{}{}
+	if cfg.WorkspaceID != "" {
+		query += ` WHERE workspace_id = $1`
+		args = append(args, cfg.WorkspaceID)
+	}
+	query += ` ORDER BY created_at ASC LIMIT $` + fmt.Sprintf("%d", len(args)+1)
+	args = append(args, cfg.Limit)
+
+	rows, err := cfg.DB.QueryContext(ctx, query, args...)
+	if err != nil {
+		return stats, fmt.Errorf("query agent_memories: %w", err)
+	}
+	defer rows.Close()
+
+	for rows.Next() {
+		stats.Scanned++
+		var (
+			id, workspaceID, content, scope string
+			createdAt                       time.Time
+		)
+		if err := rows.Scan(&id, &workspaceID, &content, &scope, &createdAt); err != nil {
+			fmt.Fprintf(stdout, "scan: %v\n", err)
+			stats.Errors++
+			continue
+		}
+
+		ns, err := mapScopeToNamespace(ctx, cfg.Resolver, workspaceID, scope)
+		if err != nil {
+			fmt.Fprintf(stdout, "[skip] id=%s workspace=%s: %v\n", id, workspaceID, err)
+			stats.Skipped++
+			continue
+		}
+
+		if cfg.DryRun {
+			fmt.Fprintf(stdout, "[dry] id=%s scope=%s → ns=%s\n", id, scope, ns)
+			stats.Copied++ // would-have-copied
+			continue
+		}
+
+		// Ensure the namespace exists before posting memories. Plugin's
+		// UpsertNamespace is idempotent so calling per-row is wasteful
+		// but safe; for v1 we accept the chattiness.
+		if _, err := cfg.Plugin.UpsertNamespace(ctx, ns, contract.NamespaceUpsert{
+			Kind: namespaceKindFromString(scope),
+		}); err != nil {
+			fmt.Fprintf(stdout, "[err-ns] id=%s ns=%s: %v\n", id, ns, err)
+			stats.Errors++
+			continue
+		}
+
+		// Pass the source row's UUID as the idempotency key so re-runs
+		// upsert in place. Without this, retries would duplicate every
+		// memory.
+		if _, err := cfg.Plugin.CommitMemory(ctx, ns, contract.MemoryWrite{
+			ID:      id,
+			Content: content,
+			Kind:    contract.MemoryKindFact,
+			Source:  contract.MemorySourceAgent,
+		}); err != nil {
+			fmt.Fprintf(stdout, "[err-mem] id=%s ns=%s: %v\n", id, ns, err)
+			stats.Errors++
+			continue
+		}
+		stats.Copied++
+	}
+	if err := rows.Err(); err != nil {
+		return stats, fmt.Errorf("iterate rows: %w", err)
+	}
+	return stats, nil
+}
+
+// mapScopeToNamespace mirrors the legacy-shim translation. The
+// backfill needs the SAME mapping the runtime uses so reads work
+// after cutover.
+func mapScopeToNamespace(ctx context.Context, r backfillResolver, workspaceID, scope string) (string, error) {
+	writable, err := r.WritableNamespaces(ctx, workspaceID)
+	if err != nil {
+		return "", fmt.Errorf("resolve writable: %w", err)
+	}
+	wantKind := contract.NamespaceKindWorkspace
+	switch scope {
+	case "LOCAL":
+		wantKind = contract.NamespaceKindWorkspace
+	case "TEAM":
+		wantKind = contract.NamespaceKindTeam
+	case "GLOBAL":
+		wantKind = contract.NamespaceKindOrg
+	default:
+		return "", fmt.Errorf("unknown scope %q", scope)
+	}
+	for _, ns := range writable {
+		if ns.Kind == wantKind {
+			return ns.Name, nil
+		}
+	}
+	return "", fmt.Errorf("no writable namespace of kind %s for workspace %s", wantKind, workspaceID)
+}
+
+// namespaceKindFromString returns the contract.NamespaceKind for a
+// legacy scope value. Unknown scopes default to "workspace" so the
+// backfill never aborts on an unexpected row.
+func namespaceKindFromString(scope string) contract.NamespaceKind {
+	switch strings.ToUpper(scope) {
+	case "TEAM":
+		return contract.NamespaceKindTeam
+	case "GLOBAL":
+		return contract.NamespaceKindOrg
+	default:
+		return contract.NamespaceKindWorkspace
+	}
+}
+
+// namespaceResolverAdapter bridges *namespace.Resolver (which returns
+// []namespace.Namespace) to verify.go's verifyResolver interface
+// (which wants []ResolvedNamespace). Keeps verify.go independent of
+// the namespace-package dependency so its tests can stub easily.
+type namespaceResolverAdapter struct {
+	r *namespace.Resolver
+}
+
+func (a namespaceResolverAdapter) ReadableNamespaces(ctx context.Context, workspaceID string) ([]ResolvedNamespace, error) {
+	src, err := a.r.ReadableNamespaces(ctx, workspaceID)
+	if err != nil {
+		return nil, err
+	}
+	out := make([]ResolvedNamespace, len(src))
+	for i, ns := range src {
+		out[i] = ResolvedNamespace{Name: ns.Name}
+	}
+	return out, nil
+}

workspace-server/cmd/memory-backfill/main_test.go

@@ -0,0 +1,434 @@
+package main
+
+import (
+	"context"
+	"errors"
+	"os"
+	"strings"
+	"testing"
+	"time"
+
+	"github.com/DATA-DOG/go-sqlmock"
+
+	"github.com/Molecule-AI/molecule-monorepo/platform/internal/memory/contract"
+	"github.com/Molecule-AI/molecule-monorepo/platform/internal/memory/namespace"
+)
+
+// stubBackfillPlugin records calls for assertions.
+type stubBackfillPlugin struct {
+	upsertedNamespaces  []string
+	committedNamespaces []string
+	committedIDs        []string // captures MemoryWrite.ID per call
+	upsertErr           error
+	commitErr           error
+}
+
+func (s *stubBackfillPlugin) UpsertNamespace(_ context.Context, name string, _ contract.NamespaceUpsert) (*contract.Namespace, error) {
+	s.upsertedNamespaces = append(s.upsertedNamespaces, name)
+	if s.upsertErr != nil {
+		return nil, s.upsertErr
+	}
+	return &contract.Namespace{Name: name, Kind: contract.NamespaceKindWorkspace}, nil
+}
+func (s *stubBackfillPlugin) CommitMemory(_ context.Context, ns string, body contract.MemoryWrite) (*contract.MemoryWriteResponse, error) {
+	s.committedNamespaces = append(s.committedNamespaces, ns)
+	s.committedIDs = append(s.committedIDs, body.ID)
+	if s.commitErr != nil {
+		return nil, s.commitErr
+	}
+	id := body.ID
+	if id == "" {
+		id = "out-1"
+	}
+	return &contract.MemoryWriteResponse{ID: id, Namespace: ns}, nil
+}
+
+type stubBackfillResolver struct {
+	writable []namespace.Namespace
+	err      error
+}
+
+func (s *stubBackfillResolver) WritableNamespaces(_ context.Context, _ string) ([]namespace.Namespace, error) {
+	return s.writable, s.err
+}
+
+func rootBackfillResolver() *stubBackfillResolver {
+	return &stubBackfillResolver{
+		writable: []namespace.Namespace{
+			{Name: "workspace:root-1", Kind: contract.NamespaceKindWorkspace, Writable: true},
+			{Name: "team:root-1", Kind: contract.NamespaceKindTeam, Writable: true},
+			{Name: "org:root-1", Kind: contract.NamespaceKindOrg, Writable: true},
+		},
+	}
+}
+
+// --- mapScopeToNamespace ---
+
+func TestMapScopeToNamespace(t *testing.T) {
+	cases := []struct {
+		scope   string
+		want    string
+		wantErr string
+	}{
+		{"LOCAL", "workspace:root-1", ""},
+		{"TEAM", "team:root-1", ""},
+		{"GLOBAL", "org:root-1", ""},
+		{"WEIRD", "", "unknown scope"},
+	}
+	for _, tc := range cases {
+		t.Run(tc.scope, func(t *testing.T) {
+			got, err := mapScopeToNamespace(context.Background(), rootBackfillResolver(), "root-1", tc.scope)
+			if tc.wantErr != "" {
+				if err == nil || !strings.Contains(err.Error(), tc.wantErr) {
+					t.Errorf("err = %v, want %q", err, tc.wantErr)
+				}
+				return
+			}
+			if err != nil {
+				t.Fatalf("err: %v", err)
+			}
+			if got != tc.want {
+				t.Errorf("got %q, want %q", got, tc.want)
+			}
+		})
+	}
+}
+
+func TestMapScopeToNamespace_ResolverError(t *testing.T) {
+	r := &stubBackfillResolver{err: errors.New("dead")}
+	_, err := mapScopeToNamespace(context.Background(), r, "root-1", "LOCAL")
+	if err == nil {
+		t.Error("expected error")
+	}
+}
+
+func TestMapScopeToNamespace_NoMatchingKind(t *testing.T) {
+	r := &stubBackfillResolver{writable: []namespace.Namespace{
+		{Name: "workspace:x", Kind: contract.NamespaceKindWorkspace, Writable: true},
+	}}
+	_, err := mapScopeToNamespace(context.Background(), r, "root-1", "TEAM")
+	if err == nil || !strings.Contains(err.Error(), "no writable namespace") {
+		t.Errorf("err = %v", err)
+	}
+}
+
+// --- namespaceKindFromString ---
+
+func TestNamespaceKindFromString(t *testing.T) {
+	cases := []struct {
+		in   string
+		want contract.NamespaceKind
+	}{
+		{"LOCAL", contract.NamespaceKindWorkspace},
+		{"local", contract.NamespaceKindWorkspace},
+		{"TEAM", contract.NamespaceKindTeam},
+		{"team", contract.NamespaceKindTeam},
+		{"GLOBAL", contract.NamespaceKindOrg},
+		{"global", contract.NamespaceKindOrg},
+		{"weird", contract.NamespaceKindWorkspace}, // safe default
+		{"", contract.NamespaceKindWorkspace},
+	}
+	for _, tc := range cases {
+		if got := namespaceKindFromString(tc.in); got != tc.want {
+			t.Errorf("namespaceKindFromString(%q) = %q, want %q", tc.in, got, tc.want)
+		}
+	}
+}
+
+// --- backfill (the workhorse) ---
+
+// TestBackfill_PassesSourceUUIDAsIdempotencyKey pins the Critical-1
+// fix: backfill must forward agent_memories.id to MemoryWrite.ID so
+// re-runs upsert in place.
+func TestBackfill_PassesSourceUUIDAsIdempotencyKey(t *testing.T) {
+	db, mock, _ := sqlmock.New()
+	defer db.Close()
+	now := time.Now().UTC()
+	mock.ExpectQuery("SELECT id, workspace_id, content, scope, created_at").
+		WillReturnRows(sqlmock.NewRows([]string{"id", "workspace_id", "content", "scope", "created_at"}).
+			AddRow("source-uuid-A", "root-1", "fact 1", "LOCAL", now).
+			AddRow("source-uuid-B", "root-1", "fact 2", "LOCAL", now))
+
+	plugin := &stubBackfillPlugin{}
+	cfg := backfillConfig{DB: db, Plugin: plugin, Resolver: rootBackfillResolver(), Limit: 100}
+	devnull, _ := os.Open(os.DevNull)
+	defer devnull.Close()
+	if _, err := backfill(context.Background(), cfg, devnull); err != nil {
+		t.Fatalf("backfill: %v", err)
+	}
+	if len(plugin.committedIDs) != 2 {
+		t.Fatalf("commits = %d", len(plugin.committedIDs))
+	}
+	if plugin.committedIDs[0] != "source-uuid-A" || plugin.committedIDs[1] != "source-uuid-B" {
+		t.Errorf("committedIDs = %v; idempotency key not forwarded", plugin.committedIDs)
+	}
+}
+
+// TestBackfill_RerunIsIdempotent: same agent_memories rows backfilled
+// twice. Plugin sees the same UUIDs both times; without the fix the
+// plugin would generate fresh UUIDs and duplicate.
+func TestBackfill_RerunIsIdempotent(t *testing.T) {
+	db, mock, _ := sqlmock.New()
+	defer db.Close()
+	now := time.Now().UTC()
+	rows1 := sqlmock.NewRows([]string{"id", "workspace_id", "content", "scope", "created_at"}).
+		AddRow("uuid-1", "root-1", "fact", "LOCAL", now)
+	rows2 := sqlmock.NewRows([]string{"id", "workspace_id", "content", "scope", "created_at"}).
+		AddRow("uuid-1", "root-1", "fact", "LOCAL", now)
+	mock.ExpectQuery("SELECT id, workspace_id, content, scope, created_at").WillReturnRows(rows1)
+	mock.ExpectQuery("SELECT id, workspace_id, content, scope, created_at").WillReturnRows(rows2)
+
+	plugin := &stubBackfillPlugin{}
+	cfg := backfillConfig{DB: db, Plugin: plugin, Resolver: rootBackfillResolver(), Limit: 100}
+	devnull, _ := os.Open(os.DevNull)
+	defer devnull.Close()
+
+	if _, err := backfill(context.Background(), cfg, devnull); err != nil {
+		t.Fatal(err)
+	}
+	if _, err := backfill(context.Background(), cfg, devnull); err != nil {
+		t.Fatal(err)
+	}
+	if len(plugin.committedIDs) != 2 {
+		t.Errorf("commits = %d, want 2", len(plugin.committedIDs))
+	}
+	if plugin.committedIDs[0] != "uuid-1" || plugin.committedIDs[1] != "uuid-1" {
+		t.Errorf("ids = %v; both runs must pass uuid-1 (relies on plugin upsert for actual de-dup)", plugin.committedIDs)
+	}
+}
+
+func TestBackfill_HappyPath_Apply(t *testing.T) {
+	db, mock, _ := sqlmock.New()
+	defer db.Close()
+	now := time.Now().UTC()
+	mock.ExpectQuery("SELECT id, workspace_id, content, scope, created_at").
+		WillReturnRows(sqlmock.NewRows([]string{"id", "workspace_id", "content", "scope", "created_at"}).
+			AddRow("mem-1", "root-1", "fact x", "LOCAL", now).
+			AddRow("mem-2", "root-1", "team y", "TEAM", now).
+			AddRow("mem-3", "root-1", "org z", "GLOBAL", now))
+
+	plugin := &stubBackfillPlugin{}
+	cfg := backfillConfig{
+		DB:       db,
+		Plugin:   plugin,
+		Resolver: rootBackfillResolver(),
+		Limit:    100,
+		DryRun:   false,
+	}
+	devnull, _ := os.Open(os.DevNull)
+	defer devnull.Close()
+	stats, err := backfill(context.Background(), cfg, devnull)
+	if err != nil {
+		t.Fatalf("err: %v", err)
+	}
+	if stats.Scanned != 3 || stats.Copied != 3 || stats.Errors != 0 {
+		t.Errorf("stats = %+v", stats)
+	}
+	if len(plugin.committedNamespaces) != 3 {
+		t.Errorf("commits = %v", plugin.committedNamespaces)
+	}
+}
+
+func TestBackfill_DryRun_DoesNotCallPlugin(t *testing.T) {
+	db, mock, _ := sqlmock.New()
+	defer db.Close()
+	now := time.Now().UTC()
+	mock.ExpectQuery("SELECT id, workspace_id, content, scope, created_at").
+		WillReturnRows(sqlmock.NewRows([]string{"id", "workspace_id", "content", "scope", "created_at"}).
+			AddRow("mem-1", "root-1", "fact x", "LOCAL", now))
+
+	plugin := &stubBackfillPlugin{}
+	cfg := backfillConfig{DB: db, Plugin: plugin, Resolver: rootBackfillResolver(), Limit: 100, DryRun: true}
+	devnull, _ := os.Open(os.DevNull)
+	defer devnull.Close()
+	stats, err := backfill(context.Background(), cfg, devnull)
+	if err != nil {
+		t.Fatalf("err: %v", err)
+	}
+	if stats.Copied != 1 {
+		t.Errorf("copied = %d", stats.Copied)
+	}
+	if len(plugin.committedNamespaces) != 0 {
+		t.Errorf("plugin must not be called in dry-run mode")
+	}
+}
+
+func TestBackfill_WorkspaceFilter(t *testing.T) {
+	db, mock, _ := sqlmock.New()
+	defer db.Close()
+	mock.ExpectQuery("SELECT id, workspace_id, content, scope, created_at").
+		WithArgs("specific-ws", 100).
+		WillReturnRows(sqlmock.NewRows([]string{"id", "workspace_id", "content", "scope", "created_at"}))
+	cfg := backfillConfig{DB: db, Plugin: &stubBackfillPlugin{}, Resolver: rootBackfillResolver(), Limit: 100, WorkspaceID: "specific-ws"}
+	devnull, _ := os.Open(os.DevNull)
+	defer devnull.Close()
+	if _, err := backfill(context.Background(), cfg, devnull); err != nil {
+		t.Fatalf("err: %v", err)
+	}
+	if err := mock.ExpectationsWereMet(); err != nil {
+		t.Errorf("workspace filter not applied: %v", err)
+	}
+}
+
+func TestBackfill_QueryError(t *testing.T) {
+	db, mock, _ := sqlmock.New()
+	defer db.Close()
+	mock.ExpectQuery("SELECT id, workspace_id, content, scope, created_at").
+		WillReturnError(errors.New("dead"))
+	cfg := backfillConfig{DB: db, Plugin: &stubBackfillPlugin{}, Resolver: rootBackfillResolver(), Limit: 100}
+	devnull, _ := os.Open(os.DevNull)
+	defer devnull.Close()
+	_, err := backfill(context.Background(), cfg, devnull)
+	if err == nil {
+		t.Error("expected error")
+	}
+}
+
+func TestBackfill_ScanError(t *testing.T) {
+	db, mock, _ := sqlmock.New()
+	defer db.Close()
+	mock.ExpectQuery("SELECT id, workspace_id, content, scope, created_at").
+		WillReturnRows(sqlmock.NewRows([]string{"id"}). // wrong shape
+								AddRow("mem-1"))
+	cfg := backfillConfig{DB: db, Plugin: &stubBackfillPlugin{}, Resolver: rootBackfillResolver(), Limit: 100}
+	devnull, _ := os.Open(os.DevNull)
+	defer devnull.Close()
+	stats, err := backfill(context.Background(), cfg, devnull)
+	if err != nil {
+		t.Fatalf("err: %v", err)
+	}
+	if stats.Errors != 1 {
+		t.Errorf("errors = %d, want 1", stats.Errors)
+	}
+}
+
+func TestBackfill_RowsErr(t *testing.T) {
+	db, mock, _ := sqlmock.New()
+	defer db.Close()
+	mock.ExpectQuery("SELECT id, workspace_id, content, scope, created_at").
+		WillReturnRows(sqlmock.NewRows([]string{"id", "workspace_id", "content", "scope", "created_at"}).
+			AddRow("mem-1", "root-1", "x", "LOCAL", time.Now().UTC()).
+			RowError(0, errors.New("mid-iter")))
+	cfg := backfillConfig{DB: db, Plugin: &stubBackfillPlugin{}, Resolver: rootBackfillResolver(), Limit: 100}
+	devnull, _ := os.Open(os.DevNull)
+	defer devnull.Close()
+	_, err := backfill(context.Background(), cfg, devnull)
+	if err == nil || !strings.Contains(err.Error(), "iterate") {
+		t.Errorf("err = %v", err)
+	}
+}
+
+func TestBackfill_SkipsUnmappableRow(t *testing.T) {
+	db, mock, _ := sqlmock.New()
+	defer db.Close()
+	mock.ExpectQuery("SELECT id, workspace_id, content, scope, created_at").
+		WillReturnRows(sqlmock.NewRows([]string{"id", "workspace_id", "content", "scope", "created_at"}).
+			AddRow("mem-1", "root-1", "x", "WEIRD", time.Now().UTC()))
+	cfg := backfillConfig{DB: db, Plugin: &stubBackfillPlugin{}, Resolver: rootBackfillResolver(), Limit: 100}
+	devnull, _ := os.Open(os.DevNull)
+	defer devnull.Close()
+	stats, err := backfill(context.Background(), cfg, devnull)
+	if err != nil {
+		t.Fatalf("err: %v", err)
+	}
+	if stats.Skipped != 1 || stats.Copied != 0 {
+		t.Errorf("stats = %+v", stats)
+	}
+}
+
+func TestBackfill_PluginUpsertNamespaceError(t *testing.T) {
+	db, mock, _ := sqlmock.New()
+	defer db.Close()
+	mock.ExpectQuery("SELECT id, workspace_id, content, scope, created_at").
+		WillReturnRows(sqlmock.NewRows([]string{"id", "workspace_id", "content", "scope", "created_at"}).
+			AddRow("mem-1", "root-1", "x", "LOCAL", time.Now().UTC()))
+	cfg := backfillConfig{DB: db, Plugin: &stubBackfillPlugin{upsertErr: errors.New("ns dead")}, Resolver: rootBackfillResolver(), Limit: 100}
+	devnull, _ := os.Open(os.DevNull)
+	defer devnull.Close()
+	stats, err := backfill(context.Background(), cfg, devnull)
+	if err != nil {
+		t.Fatalf("err: %v", err)
+	}
+	if stats.Errors != 1 || stats.Copied != 0 {
+		t.Errorf("stats = %+v", stats)
+	}
+}
+
+func TestBackfill_PluginCommitMemoryError(t *testing.T) {
+	db, mock, _ := sqlmock.New()
+	defer db.Close()
+	mock.ExpectQuery("SELECT id, workspace_id, content, scope, created_at").
+		WillReturnRows(sqlmock.NewRows([]string{"id", "workspace_id", "content", "scope", "created_at"}).
+			AddRow("mem-1", "root-1", "x", "LOCAL", time.Now().UTC()))
+	cfg := backfillConfig{DB: db, Plugin: &stubBackfillPlugin{commitErr: errors.New("mem dead")}, Resolver: rootBackfillResolver(), Limit: 100}
+	devnull, _ := os.Open(os.DevNull)
+	defer devnull.Close()
+	stats, err := backfill(context.Background(), cfg, devnull)
+	if err != nil {
+		t.Fatalf("err: %v", err)
+	}
+	if stats.Errors != 1 || stats.Copied != 0 {
+		t.Errorf("stats = %+v", stats)
+	}
+}
+
+// --- run (CLI driver) ---
+
+func TestRun_RejectsBothModes(t *testing.T) {
+	stderr, _ := os.OpenFile(os.DevNull, os.O_WRONLY, 0)
+	defer stderr.Close()
+	stdout, _ := os.OpenFile(os.DevNull, os.O_WRONLY, 0)
+	defer stdout.Close()
+	err := run([]string{"-dry-run", "-apply"}, stdout, stderr)
+	if err == nil || !strings.Contains(err.Error(), "exactly one") {
+		t.Errorf("err = %v", err)
+	}
+}
+
+func TestRun_RejectsNeitherMode(t *testing.T) {
+	stderr, _ := os.OpenFile(os.DevNull, os.O_WRONLY, 0)
+	defer stderr.Close()
+	stdout, _ := os.OpenFile(os.DevNull, os.O_WRONLY, 0)
+	defer stdout.Close()
+	err := run([]string{}, stdout, stderr)
+	if err == nil || !strings.Contains(err.Error(), "exactly one") {
+		t.Errorf("err = %v", err)
+	}
+}
+
+func TestRun_RejectsMissingDatabaseURL(t *testing.T) {
+	t.Setenv("DATABASE_URL", "")
+	t.Setenv("MEMORY_PLUGIN_URL", "http://x")
+	stderr, _ := os.OpenFile(os.DevNull, os.O_WRONLY, 0)
+	defer stderr.Close()
+	stdout, _ := os.OpenFile(os.DevNull, os.O_WRONLY, 0)
+	defer stdout.Close()
+	err := run([]string{"-dry-run"}, stdout, stderr)
+	if err == nil || !strings.Contains(err.Error(), "DATABASE_URL") {
+		t.Errorf("err = %v", err)
+	}
+}
+
+func TestRun_RejectsMissingPluginURL(t *testing.T) {
+	t.Setenv("DATABASE_URL", "postgres://invalid")
+	t.Setenv("MEMORY_PLUGIN_URL", "")
+	stderr, _ := os.OpenFile(os.DevNull, os.O_WRONLY, 0)
+	defer stderr.Close()
+	stdout, _ := os.OpenFile(os.DevNull, os.O_WRONLY, 0)
+	defer stdout.Close()
+	err := run([]string{"-dry-run"}, stdout, stderr)
+	if err == nil || !strings.Contains(err.Error(), "MEMORY_PLUGIN_URL") {
+		t.Errorf("err = %v", err)
+	}
+}
+
+func TestRun_BadFlags(t *testing.T) {
+	stderr, _ := os.OpenFile(os.DevNull, os.O_WRONLY, 0)
+	defer stderr.Close()
+	stdout, _ := os.OpenFile(os.DevNull, os.O_WRONLY, 0)
+	defer stdout.Close()
+	err := run([]string{"-not-a-flag"}, stdout, stderr)
+	if err == nil {
+		t.Error("expected flag parse error")
+	}
+}

workspace-server/cmd/memory-backfill/verify.go

@@ -0,0 +1,200 @@
+package main
+
+// verify.go — post-apply parity check.
+//
+// After a backfill -apply, run with -verify to confirm the migration
+// actually produced equivalent data. Picks `SampleSize` random
+// workspaces, queries agent_memories direct + plugin search via the
+// caller's namespaces, and diffs the result sets by content.
+//
+// The diff is best-effort: pg's recent-first ordering and the plugin's
+// internal ordering may differ, so we compare as sets, not lists.
+// We do require strict 1:1 multiset equality (every legacy row maps
+// to exactly one plugin row, ignoring id since the backfill preserves
+// it via the C1 idempotency key).
+
+import (
+	"context"
+	"database/sql"
+	"fmt"
+	"math/rand"
+	"os"
+
+	"github.com/Molecule-AI/molecule-monorepo/platform/internal/memory/contract"
+)
+
+// verifyConfig is the typed dependency bundle for verifyParity.
+type verifyConfig struct {
+	DB          *sql.DB
+	Plugin      verifyPlugin
+	Resolver    verifyResolver
+	SampleSize  int
+	WorkspaceID string // optional: limit to one workspace
+	Rand        *rand.Rand
+}
+
+// verifyPlugin is the slice of memory-plugin client we call.
+type verifyPlugin interface {
+	Search(ctx context.Context, body contract.SearchRequest) (*contract.SearchResponse, error)
+}
+
+// verifyResolver mirrors namespace.Resolver. Same shape as
+// backfillResolver but kept distinct so verify isn't tied to
+// backfill's interface.
+type verifyResolver interface {
+	ReadableNamespaces(ctx context.Context, workspaceID string) ([]ResolvedNamespace, error)
+}
+
+// ResolvedNamespace is the minimum we need from the resolver — kept
+// separate so the verify code doesn't depend on the namespace package
+// (the live tests inject stubs, the binary uses an adapter).
+type ResolvedNamespace struct {
+	Name string
+}
+
+// verifyReport accumulates the per-workspace results.
+type verifyReport struct {
+	WorkspacesSampled int
+	Matches           int
+	Mismatches        int
+	Errors            int
+}
+
+// verifyParity is the workhorse. Returns a report; the CLI converts
+// any non-zero mismatches/errors into a non-zero exit so CI can gate
+// the cutover.
+func verifyParity(ctx context.Context, cfg verifyConfig, stdout *os.File) (*verifyReport, error) {
+	report := &verifyReport{}
+	rng := cfg.Rand
+	if rng == nil {
+		rng = rand.New(rand.NewSource(42)) //nolint:gosec // determinism > unpredictability for ops
+	}
+
+	wsIDs, err := pickWorkspaceSample(ctx, cfg.DB, cfg.WorkspaceID, cfg.SampleSize, rng)
+	if err != nil {
+		return report, fmt.Errorf("pick sample: %w", err)
+	}
+
+	for _, wsID := range wsIDs {
+		report.WorkspacesSampled++
+		legacy, err := queryLegacyMemories(ctx, cfg.DB, wsID)
+		if err != nil {
+			fmt.Fprintf(stdout, "[err] workspace=%s legacy query: %v\n", wsID, err)
+			report.Errors++
+			continue
+		}
+		readable, err := cfg.Resolver.ReadableNamespaces(ctx, wsID)
+		if err != nil {
+			fmt.Fprintf(stdout, "[err] workspace=%s resolve: %v\n", wsID, err)
+			report.Errors++
+			continue
+		}
+		nsList := make([]string, len(readable))
+		for i, ns := range readable {
+			nsList[i] = ns.Name
+		}
+		if len(nsList) == 0 {
+			// No readable namespaces — empty plugin result expected.
+			if len(legacy) == 0 {
+				report.Matches++
+			} else {
+				fmt.Fprintf(stdout, "[mismatch] workspace=%s legacy=%d plugin=0 (no readable namespaces)\n", wsID, len(legacy))
+				report.Mismatches++
+			}
+			continue
+		}
+		resp, err := cfg.Plugin.Search(ctx, contract.SearchRequest{Namespaces: nsList, Limit: 100})
+		if err != nil {
+			fmt.Fprintf(stdout, "[err] workspace=%s plugin search: %v\n", wsID, err)
+			report.Errors++
+			continue
+		}
+		pluginContents := make(map[string]int, len(resp.Memories))
+		for _, m := range resp.Memories {
+			pluginContents[m.Content]++
+		}
+		// Compare as multisets: each legacy content appears at least
+		// once in plugin output. We deliberately tolerate plugin
+		// having MORE rows (the namespace might include team-shared
+		// memories from sibling workspaces that aren't in this
+		// workspace's agent_memories rows).
+		matched := true
+		for _, c := range legacy {
+			if pluginContents[c] == 0 {
+				fmt.Fprintf(stdout, "[mismatch] workspace=%s missing-from-plugin content=%q\n", wsID, truncate(c, 80))
+				matched = false
+				break
+			}
+			pluginContents[c]--
+		}
+		if matched {
+			report.Matches++
+		} else {
+			report.Mismatches++
+		}
+	}
+	return report, nil
+}
+
+// pickWorkspaceSample returns up to N workspace UUIDs. If
+// WorkspaceID is set, returns only that one. Otherwise selects N
+// random workspaces from the workspaces table (TABLESAMPLE would be
+// nicer but SYSTEM/BERNOULLI sampling has surprising distribution
+// properties for small populations; we just ORDER BY random() LIMIT).
+func pickWorkspaceSample(ctx context.Context, db *sql.DB, workspaceID string, n int, _ *rand.Rand) ([]string, error) {
+	if workspaceID != "" {
+		return []string{workspaceID}, nil
+	}
+	rows, err := db.QueryContext(ctx, `
+		SELECT id::text
+		FROM workspaces
+		WHERE status != 'removed'
+		ORDER BY random()
+		LIMIT $1
+	`, n)
+	if err != nil {
+		return nil, err
+	}
+	defer rows.Close()
+	out := make([]string, 0, n)
+	for rows.Next() {
+		var id string
+		if err := rows.Scan(&id); err != nil {
+			return nil, err
+		}
+		out = append(out, id)
+	}
+	return out, rows.Err()
+}
+
+// queryLegacyMemories pulls all agent_memories rows for a workspace
+// (LOCAL + TEAM scopes — what the plugin search would return through
+// the resolver's readable list, mapped via PR-6 shim semantics).
+func queryLegacyMemories(ctx context.Context, db *sql.DB, workspaceID string) ([]string, error) {
+	rows, err := db.QueryContext(ctx, `
+		SELECT content
+		FROM agent_memories
+		WHERE workspace_id = $1
+		ORDER BY created_at DESC
+	`, workspaceID)
+	if err != nil {
+		return nil, err
+	}
+	defer rows.Close()
+	out := []string{}
+	for rows.Next() {
+		var c string
+		if err := rows.Scan(&c); err != nil {
+			return nil, err
+		}
+		out = append(out, c)
+	}
+	return out, rows.Err()
+}
+
+func truncate(s string, n int) string {
+	if len(s) <= n {
+		return s
+	}
+	return s[:n] + "…"
+}

workspace-server/cmd/memory-backfill/verify_test.go

@@ -0,0 +1,390 @@
+package main
+
+import (
+	"context"
+	"errors"
+	"os"
+	"strings"
+	"testing"
+
+	"github.com/DATA-DOG/go-sqlmock"
+
+	"github.com/Molecule-AI/molecule-monorepo/platform/internal/memory/contract"
+)
+
+// stubVerifyPlugin records search calls and returns canned results.
+type stubVerifyPlugin struct {
+	searchFn func(ctx context.Context, body contract.SearchRequest) (*contract.SearchResponse, error)
+}
+
+func (s *stubVerifyPlugin) Search(ctx context.Context, body contract.SearchRequest) (*contract.SearchResponse, error) {
+	if s.searchFn != nil {
+		return s.searchFn(ctx, body)
+	}
+	return &contract.SearchResponse{}, nil
+}
+
+// stubVerifyResolver returns a canned readable namespace list.
+type stubVerifyResolver struct {
+	namespaces []ResolvedNamespace
+	err        error
+}
+
+func (s *stubVerifyResolver) ReadableNamespaces(_ context.Context, _ string) ([]ResolvedNamespace, error) {
+	return s.namespaces, s.err
+}
+
+// --- pickWorkspaceSample ---
+
+func TestPickWorkspaceSample_SingleWorkspaceShortCircuit(t *testing.T) {
+	db, _, _ := sqlmock.New()
+	defer db.Close()
+	got, err := pickWorkspaceSample(context.Background(), db, "specific-ws", 50, nil)
+	if err != nil {
+		t.Fatalf("err: %v", err)
+	}
+	if len(got) != 1 || got[0] != "specific-ws" {
+		t.Errorf("got %v, want [specific-ws]", got)
+	}
+}
+
+func TestPickWorkspaceSample_RandomSample(t *testing.T) {
+	db, mock, _ := sqlmock.New()
+	defer db.Close()
+	mock.ExpectQuery("SELECT id::text FROM workspaces").
+		WithArgs(50).
+		WillReturnRows(sqlmock.NewRows([]string{"id"}).
+			AddRow("ws-1").
+			AddRow("ws-2").
+			AddRow("ws-3"))
+	got, err := pickWorkspaceSample(context.Background(), db, "", 50, nil)
+	if err != nil {
+		t.Fatalf("err: %v", err)
+	}
+	if len(got) != 3 {
+		t.Errorf("got len %d, want 3", len(got))
+	}
+}
+
+func TestPickWorkspaceSample_QueryError(t *testing.T) {
+	db, mock, _ := sqlmock.New()
+	defer db.Close()
+	mock.ExpectQuery("SELECT id::text FROM workspaces").
+		WillReturnError(errors.New("dead"))
+	_, err := pickWorkspaceSample(context.Background(), db, "", 50, nil)
+	if err == nil {
+		t.Error("expected error")
+	}
+}
+
+func TestPickWorkspaceSample_ScanError(t *testing.T) {
+	db, mock, _ := sqlmock.New()
+	defer db.Close()
+	mock.ExpectQuery("SELECT id::text FROM workspaces").
+		WillReturnRows(sqlmock.NewRows([]string{"id", "extra"}). // wrong shape
+								AddRow("ws-1", "extra"))
+	_, err := pickWorkspaceSample(context.Background(), db, "", 50, nil)
+	if err == nil {
+		t.Error("expected scan error")
+	}
+}
+
+// --- queryLegacyMemories ---
+
+func TestQueryLegacyMemories_HappyPath(t *testing.T) {
+	db, mock, _ := sqlmock.New()
+	defer db.Close()
+	mock.ExpectQuery("SELECT content FROM agent_memories").
+		WithArgs("ws-1").
+		WillReturnRows(sqlmock.NewRows([]string{"content"}).
+			AddRow("fact 1").
+			AddRow("fact 2"))
+	got, err := queryLegacyMemories(context.Background(), db, "ws-1")
+	if err != nil {
+		t.Fatalf("err: %v", err)
+	}
+	if len(got) != 2 || got[0] != "fact 1" {
+		t.Errorf("got %v", got)
+	}
+}
+
+func TestQueryLegacyMemories_QueryError(t *testing.T) {
+	db, mock, _ := sqlmock.New()
+	defer db.Close()
+	mock.ExpectQuery("SELECT content FROM agent_memories").
+		WillReturnError(errors.New("dead"))
+	_, err := queryLegacyMemories(context.Background(), db, "ws-1")
+	if err == nil {
+		t.Error("expected error")
+	}
+}
+
+// --- verifyParity (the workhorse) ---
+
+func TestVerifyParity_AllMatch(t *testing.T) {
+	db, mock, _ := sqlmock.New()
+	defer db.Close()
+	mock.ExpectQuery("SELECT id::text FROM workspaces").
+		WillReturnRows(sqlmock.NewRows([]string{"id"}).AddRow("ws-1"))
+	mock.ExpectQuery("SELECT content FROM agent_memories").
+		WithArgs("ws-1").
+		WillReturnRows(sqlmock.NewRows([]string{"content"}).
+			AddRow("fact A").
+			AddRow("fact B"))
+
+	plugin := &stubVerifyPlugin{
+		searchFn: func(_ context.Context, _ contract.SearchRequest) (*contract.SearchResponse, error) {
+			return &contract.SearchResponse{Memories: []contract.Memory{
+				{ID: "id-A", Content: "fact A"},
+				{ID: "id-B", Content: "fact B"},
+			}}, nil
+		},
+	}
+	resolver := &stubVerifyResolver{
+		namespaces: []ResolvedNamespace{{Name: "workspace:ws-1"}},
+	}
+	cfg := verifyConfig{DB: db, Plugin: plugin, Resolver: resolver, SampleSize: 50}
+	devnull, _ := os.Open(os.DevNull)
+	defer devnull.Close()
+	report, err := verifyParity(context.Background(), cfg, devnull)
+	if err != nil {
+		t.Fatalf("err: %v", err)
+	}
+	if report.Matches != 1 || report.Mismatches != 0 || report.Errors != 0 {
+		t.Errorf("report = %+v, want 1 match", report)
+	}
+}
+
+func TestVerifyParity_MismatchDetectsMissingFromPlugin(t *testing.T) {
+	db, mock, _ := sqlmock.New()
+	defer db.Close()
+	mock.ExpectQuery("SELECT id::text FROM workspaces").
+		WillReturnRows(sqlmock.NewRows([]string{"id"}).AddRow("ws-1"))
+	mock.ExpectQuery("SELECT content FROM agent_memories").
+		WillReturnRows(sqlmock.NewRows([]string{"content"}).
+			AddRow("fact A").
+			AddRow("fact-missing-from-plugin"))
+
+	plugin := &stubVerifyPlugin{
+		searchFn: func(_ context.Context, _ contract.SearchRequest) (*contract.SearchResponse, error) {
+			return &contract.SearchResponse{Memories: []contract.Memory{
+				{ID: "id-A", Content: "fact A"},
+			}}, nil
+		},
+	}
+	resolver := &stubVerifyResolver{
+		namespaces: []ResolvedNamespace{{Name: "workspace:ws-1"}},
+	}
+	cfg := verifyConfig{DB: db, Plugin: plugin, Resolver: resolver, SampleSize: 50}
+	devnull, _ := os.Open(os.DevNull)
+	defer devnull.Close()
+	report, err := verifyParity(context.Background(), cfg, devnull)
+	if err != nil {
+		t.Fatalf("err: %v", err)
+	}
+	if report.Mismatches != 1 {
+		t.Errorf("report = %+v, want 1 mismatch", report)
+	}
+}
+
+func TestVerifyParity_PluginExtraRowsTolerated(t *testing.T) {
+	db, mock, _ := sqlmock.New()
+	defer db.Close()
+	mock.ExpectQuery("SELECT id::text FROM workspaces").
+		WillReturnRows(sqlmock.NewRows([]string{"id"}).AddRow("ws-1"))
+	mock.ExpectQuery("SELECT content FROM agent_memories").
+		WillReturnRows(sqlmock.NewRows([]string{"content"}).
+			AddRow("fact A"))
+
+	// Plugin returns more rows (e.g., team-shared from a sibling).
+	// Verify treats this as a match — legacy is a subset of plugin.
+	plugin := &stubVerifyPlugin{
+		searchFn: func(_ context.Context, _ contract.SearchRequest) (*contract.SearchResponse, error) {
+			return &contract.SearchResponse{Memories: []contract.Memory{
+				{ID: "id-A", Content: "fact A"},
+				{ID: "id-team-1", Content: "team-shared content from sibling"},
+			}}, nil
+		},
+	}
+	resolver := &stubVerifyResolver{
+		namespaces: []ResolvedNamespace{{Name: "workspace:ws-1"}, {Name: "team:root"}},
+	}
+	cfg := verifyConfig{DB: db, Plugin: plugin, Resolver: resolver, SampleSize: 50}
+	devnull, _ := os.Open(os.DevNull)
+	defer devnull.Close()
+	report, err := verifyParity(context.Background(), cfg, devnull)
+	if err != nil {
+		t.Fatalf("err: %v", err)
+	}
+	if report.Matches != 1 || report.Mismatches != 0 {
+		t.Errorf("report = %+v, want 1 match (plugin-extra is OK)", report)
+	}
+}
+
+func TestVerifyParity_LegacyQueryError(t *testing.T) {
+	db, mock, _ := sqlmock.New()
+	defer db.Close()
+	mock.ExpectQuery("SELECT id::text FROM workspaces").
+		WillReturnRows(sqlmock.NewRows([]string{"id"}).AddRow("ws-1"))
+	mock.ExpectQuery("SELECT content FROM agent_memories").
+		WillReturnError(errors.New("dead"))
+
+	cfg := verifyConfig{
+		DB:       db,
+		Plugin:   &stubVerifyPlugin{},
+		Resolver: &stubVerifyResolver{namespaces: []ResolvedNamespace{{Name: "workspace:ws-1"}}},
+	}
+	devnull, _ := os.Open(os.DevNull)
+	defer devnull.Close()
+	report, err := verifyParity(context.Background(), cfg, devnull)
+	if err != nil {
+		t.Fatalf("err: %v", err)
+	}
+	if report.Errors != 1 {
+		t.Errorf("report = %+v, want 1 error", report)
+	}
+}
+
+func TestVerifyParity_ResolverError(t *testing.T) {
+	db, mock, _ := sqlmock.New()
+	defer db.Close()
+	mock.ExpectQuery("SELECT id::text FROM workspaces").
+		WillReturnRows(sqlmock.NewRows([]string{"id"}).AddRow("ws-1"))
+	mock.ExpectQuery("SELECT content FROM agent_memories").
+		WillReturnRows(sqlmock.NewRows([]string{"content"}).AddRow("x"))
+
+	cfg := verifyConfig{
+		DB:       db,
+		Plugin:   &stubVerifyPlugin{},
+		Resolver: &stubVerifyResolver{err: errors.New("dead")},
+	}
+	devnull, _ := os.Open(os.DevNull)
+	defer devnull.Close()
+	report, _ := verifyParity(context.Background(), cfg, devnull)
+	if report.Errors != 1 {
+		t.Errorf("report = %+v, want 1 error", report)
+	}
+}
+
+func TestVerifyParity_PluginSearchError(t *testing.T) {
+	db, mock, _ := sqlmock.New()
+	defer db.Close()
+	mock.ExpectQuery("SELECT id::text FROM workspaces").
+		WillReturnRows(sqlmock.NewRows([]string{"id"}).AddRow("ws-1"))
+	mock.ExpectQuery("SELECT content FROM agent_memories").
+		WillReturnRows(sqlmock.NewRows([]string{"content"}).AddRow("x"))
+
+	cfg := verifyConfig{
+		DB: db,
+		Plugin: &stubVerifyPlugin{
+			searchFn: func(_ context.Context, _ contract.SearchRequest) (*contract.SearchResponse, error) {
+				return nil, errors.New("plugin dead")
+			},
+		},
+		Resolver: &stubVerifyResolver{namespaces: []ResolvedNamespace{{Name: "workspace:ws-1"}}},
+	}
+	devnull, _ := os.Open(os.DevNull)
+	defer devnull.Close()
+	report, _ := verifyParity(context.Background(), cfg, devnull)
+	if report.Errors != 1 {
+		t.Errorf("report = %+v, want 1 error", report)
+	}
+}
+
+func TestVerifyParity_NoReadableNamespacesEmptyLegacy(t *testing.T) {
+	db, mock, _ := sqlmock.New()
+	defer db.Close()
+	mock.ExpectQuery("SELECT id::text FROM workspaces").
+		WillReturnRows(sqlmock.NewRows([]string{"id"}).AddRow("ws-1"))
+	mock.ExpectQuery("SELECT content FROM agent_memories").
+		WillReturnRows(sqlmock.NewRows([]string{"content"})) // empty
+
+	cfg := verifyConfig{
+		DB:       db,
+		Plugin:   &stubVerifyPlugin{},
+		Resolver: &stubVerifyResolver{namespaces: []ResolvedNamespace{}}, // empty
+	}
+	devnull, _ := os.Open(os.DevNull)
+	defer devnull.Close()
+	report, _ := verifyParity(context.Background(), cfg, devnull)
+	// Empty legacy + empty namespaces → match.
+	if report.Matches != 1 {
+		t.Errorf("report = %+v, want 1 match (both empty)", report)
+	}
+}
+
+func TestVerifyParity_NoReadableNamespacesNonEmptyLegacy(t *testing.T) {
+	db, mock, _ := sqlmock.New()
+	defer db.Close()
+	mock.ExpectQuery("SELECT id::text FROM workspaces").
+		WillReturnRows(sqlmock.NewRows([]string{"id"}).AddRow("ws-1"))
+	mock.ExpectQuery("SELECT content FROM agent_memories").
+		WillReturnRows(sqlmock.NewRows([]string{"content"}).AddRow("orphan-fact"))
+
+	cfg := verifyConfig{
+		DB:       db,
+		Plugin:   &stubVerifyPlugin{},
+		Resolver: &stubVerifyResolver{namespaces: []ResolvedNamespace{}},
+	}
+	devnull, _ := os.Open(os.DevNull)
+	defer devnull.Close()
+	report, _ := verifyParity(context.Background(), cfg, devnull)
+	// Legacy has rows but plugin can't see any → mismatch.
+	if report.Mismatches != 1 {
+		t.Errorf("report = %+v, want 1 mismatch", report)
+	}
+}
+
+func TestVerifyParity_PickSampleError(t *testing.T) {
+	db, mock, _ := sqlmock.New()
+	defer db.Close()
+	mock.ExpectQuery("SELECT id::text FROM workspaces").
+		WillReturnError(errors.New("dead"))
+	cfg := verifyConfig{DB: db, Plugin: &stubVerifyPlugin{}, Resolver: &stubVerifyResolver{}}
+	devnull, _ := os.Open(os.DevNull)
+	defer devnull.Close()
+	_, err := verifyParity(context.Background(), cfg, devnull)
+	if err == nil || !strings.Contains(err.Error(), "pick sample") {
+		t.Errorf("err = %v", err)
+	}
+}
+
+// --- Truncate ---
+
+func TestVerifyTruncate(t *testing.T) {
+	if got := truncate("short", 10); got != "short" {
+		t.Errorf("got %q", got)
+	}
+	if got := truncate(strings.Repeat("a", 200), 10); !strings.HasSuffix(got, "…") {
+		t.Errorf("expected ellipsis: %q", got)
+	}
+}
+
+// --- CLI: -verify mode ---
+
+func TestRun_VerifyVsApplyMutuallyExclusive(t *testing.T) {
+	stderr, _ := os.OpenFile(os.DevNull, os.O_WRONLY, 0)
+	defer stderr.Close()
+	stdout, _ := os.OpenFile(os.DevNull, os.O_WRONLY, 0)
+	defer stdout.Close()
+	err := run([]string{"-verify", "-apply"}, stdout, stderr)
+	if err == nil || !strings.Contains(err.Error(), "exactly one") {
+		t.Errorf("err = %v", err)
+	}
+}
+
+func TestRun_VerifyAloneIsValid(t *testing.T) {
+	t.Setenv("DATABASE_URL", "")
+	t.Setenv("MEMORY_PLUGIN_URL", "http://x")
+	stderr, _ := os.OpenFile(os.DevNull, os.O_WRONLY, 0)
+	defer stderr.Close()
+	stdout, _ := os.OpenFile(os.DevNull, os.O_WRONLY, 0)
+	defer stdout.Close()
+	err := run([]string{"-verify"}, stdout, stderr)
+	// Will fail later on missing DATABASE_URL, NOT on the
+	// mutually-exclusive-modes check. Asserts that -verify is
+	// recognized as a valid mode.
+	if err == nil || !strings.Contains(err.Error(), "DATABASE_URL") {
+		t.Errorf("err = %v, want DATABASE_URL error (-verify alone is a valid mode)", err)
+	}
+}

workspace-server/cmd/memory-plugin-postgres/E2E.md

@@ -0,0 +1,68 @@
+# Real-subprocess E2E for memory-plugin-postgres
+
+The default `go test ./...` suite covers the plugin via in-process
+sqlmock tests (PR-3). This directory ALSO ships build-tag-gated tests
+that spawn the real binary against a live postgres — to catch
+classes of bug in-process tests can't see:
+
+- Boot-path regressions (env var typos, panic-on-startup)
+- Wire-format bugs sqlmock smooths over (the `pq.Array` issue we
+  hit during PR-3 development)
+- HTTP/socket encoding edge cases
+- C1 idempotency (real upsert against real postgres)
+
+## Running
+
+The tests skip silently unless an operator opts in with both:
+- The `memory_plugin_e2e` build tag
+- `MEMORY_PLUGIN_E2E_DB` env var pointing at a writable postgres
+
+### Quick local run (with docker)
+
+```bash
+docker run --rm -d --name memory-plugin-e2e-pg \
+  -e POSTGRES_PASSWORD=test -e POSTGRES_USER=test -e POSTGRES_DB=test \
+  -p 5432:5432 \
+  pgvector/pgvector:pg16
+
+# Wait a few seconds for postgres to accept connections
+until docker exec memory-plugin-e2e-pg pg_isready -U test >/dev/null 2>&1; do sleep 0.5; done
+
+MEMORY_PLUGIN_E2E_DB=postgres://test:test@localhost:5432/test?sslmode=disable \
+  go test -tags memory_plugin_e2e -v -count=1 ./cmd/memory-plugin-postgres/
+
+docker stop memory-plugin-e2e-pg
+```
+
+### CI integration
+
+These tests are NOT in the default required-checks set. Operators
+gating cutover on the suite should add a separate workflow step:
+
+```yaml
+- name: Memory plugin E2E
+  if: ${{ contains(github.event.pull_request.labels.*.name, 'memory-v2') }}
+  run: |
+    MEMORY_PLUGIN_E2E_DB=${{ secrets.MEMORY_PLUGIN_TEST_DSN }} \
+      go test -tags memory_plugin_e2e -v -count=1 ./cmd/memory-plugin-postgres/
+```
+
+## What each test pins
+
+| Test | Covers |
+|---|---|
+| `TestE2E_BootAndHealth` | Binary builds, starts, advertises all 5 capabilities |
+| `TestE2E_FullCommitSearchForgetRoundTrip` | Real wire encoding (no sqlmock), full agent flow |
+| `TestE2E_IdempotencyKey` | C1 fix end-to-end — upserts against real postgres |
+
+## What's still NOT covered
+
+- Migration drift (assumes the migrations dir is at the conventional
+  path; operator-customized layouts need their own test)
+- Plugin-internal recovery (kill backing store mid-request, etc.)
+- Concurrent commits with id collisions across processes
+- TTL eviction (would need to extend test runtime past `expires_at`)
+
+These gaps apply equally to forks of this binary; they're listed in
+[`testing-your-plugin.md`](../../../docs/memory-plugins/testing-your-plugin.md)
+under "what the harness does NOT cover".

workspace-server/cmd/memory-plugin-postgres/boot_e2e_test.go

@@ -0,0 +1,289 @@
+//go:build memory_plugin_e2e
+
+// Package main's real-subprocess boot test (#293 fixup, RFC #2728).
+//
+// Build-tag gated so it only runs when an operator explicitly opts in:
+//
+//   MEMORY_PLUGIN_E2E_DB=postgres://test:test@localhost:5432/test?sslmode=disable \
+//     go test -tags memory_plugin_e2e -v ./cmd/memory-plugin-postgres/
+//
+// Why a separate build tag:
+//   - The default `go test ./...` run shouldn't require docker or a
+//     live postgres
+//   - CI gates that DO want to run this can set the env var + tag
+//   - Operators verifying a custom plugin against the contract can
+//     copy this file as the template (replace the binary build step
+//     with their own)
+//
+// What this exercises that PR-11's swap test doesn't:
+//   - Real `go build` of cmd/memory-plugin-postgres/
+//   - Real binary boot via os/exec — catches mixed-key panics, missing
+//     env vars, crash-on-startup issues that in-process tests skip
+//   - Real postgres connection — catches wire-format bugs (e.g. the
+//     pq.Array regression we hit during PR-3)
+//   - Real HTTP round-trip with a TCP socket — catches encoding edge
+//     cases sqlmock + httptest can't see
+//
+// What this does NOT cover:
+//   - Schema migration drift (assumes the migrations dir is at the
+//     conventional path; operator-customized layouts need their own
+//     test)
+//   - Plugin-internal recovery (kill backing store mid-request, etc.)
+
+package main
+
+import (
+	"bytes"
+	"context"
+	"encoding/json"
+	"fmt"
+	"net/http"
+	"os"
+	"os/exec"
+	"path/filepath"
+	"runtime"
+	"testing"
+	"time"
+
+	mclient "github.com/Molecule-AI/molecule-monorepo/platform/internal/memory/client"
+	"github.com/Molecule-AI/molecule-monorepo/platform/internal/memory/contract"
+)
+
+const (
+	bootProbeTimeout = 30 * time.Second
+	bootProbeStep   = 500 * time.Millisecond
+)
+
+// requireE2EDB returns the test DSN. Skips the test (not fails) when
+// the env var is unset — keeps `-tags memory_plugin_e2e` runs from
+// crashing on dev machines without postgres.
+func requireE2EDB(t *testing.T) string {
+	t.Helper()
+	dsn := os.Getenv("MEMORY_PLUGIN_E2E_DB")
+	if dsn == "" {
+		t.Skip("MEMORY_PLUGIN_E2E_DB not set — skipping real-subprocess boot test")
+	}
+	return dsn
+}
+
+// buildBinary compiles cmd/memory-plugin-postgres/ to a temp dir.
+// Returns the path of the built binary. Test cleanup deletes it.
+func buildBinary(t *testing.T) string {
+	t.Helper()
+	dir := t.TempDir()
+	out := filepath.Join(dir, "memory-plugin-postgres")
+	if runtime.GOOS == "windows" {
+		out += ".exe"
+	}
+	// Find the cmd dir relative to this file.
+	_, thisFile, _, _ := runtime.Caller(0)
+	cmdDir := filepath.Dir(thisFile)
+	build := exec.Command("go", "build", "-o", out, ".")
+	build.Dir = cmdDir
+	build.Env = os.Environ()
+	if outErr, err := build.CombinedOutput(); err != nil {
+		t.Fatalf("go build failed: %v\n%s", err, outErr)
+	}
+	return out
+}
+
+// startBinary launches the built binary with the supplied env. Returns
+// the *exec.Cmd (test cleanup kills it) and the http URL it's listening
+// on. Polls /v1/health until ready or times out.
+func startBinary(t *testing.T, binary, dsn, listen string) (*exec.Cmd, string) {
+	t.Helper()
+	url := "http://" + listen
+	cmd := exec.Command(binary)
+	cmd.Env = append(os.Environ(),
+		"MEMORY_PLUGIN_DATABASE_URL="+dsn,
+		"MEMORY_PLUGIN_LISTEN_ADDR="+listen,
+		// Migrations dir lives next to the cmd source. The binary
+		// reads it relative to cwd by default; we set the env var
+		// override so the test doesn't depend on cwd.
+		"MEMORY_PLUGIN_MIGRATIONS_DIR="+migrationsDirForTest(t),
+	)
+	stdout := &bytes.Buffer{}
+	stderr := &bytes.Buffer{}
+	cmd.Stdout = stdout
+	cmd.Stderr = stderr
+	if err := cmd.Start(); err != nil {
+		t.Fatalf("start binary: %v", err)
+	}
+	t.Cleanup(func() {
+		if cmd.Process != nil {
+			_ = cmd.Process.Kill()
+			_ = cmd.Wait()
+		}
+		if t.Failed() {
+			t.Logf("binary stdout:\n%s", stdout.String())
+			t.Logf("binary stderr:\n%s", stderr.String())
+		}
+	})
+
+	deadline := time.Now().Add(bootProbeTimeout)
+	for time.Now().Before(deadline) {
+		resp, err := http.Get(url + "/v1/health")
+		if err == nil {
+			_ = resp.Body.Close()
+			if resp.StatusCode == 200 {
+				return cmd, url
+			}
+		}
+		// Bail early if the binary already exited.
+		if cmd.ProcessState != nil && cmd.ProcessState.Exited() {
+			t.Fatalf("binary exited during boot: stderr:\n%s", stderr.String())
+		}
+		time.Sleep(bootProbeStep)
+	}
+	t.Fatalf("binary did not become ready within %v", bootProbeTimeout)
+	return nil, ""
+}
+
+func migrationsDirForTest(t *testing.T) string {
+	t.Helper()
+	_, thisFile, _, _ := runtime.Caller(0)
+	return filepath.Join(filepath.Dir(thisFile), "migrations")
+}
+
+// TestE2E_BootAndHealth: build + start the real binary, hit /v1/health,
+// confirm capabilities match what the built-in plugin declares. Catches
+// "binary doesn't start" / "wrong env var name" / "panics on first
+// request" classes that in-process tests miss.
+func TestE2E_BootAndHealth(t *testing.T) {
+	dsn := requireE2EDB(t)
+	binary := buildBinary(t)
+	_, url := startBinary(t, binary, dsn, "127.0.0.1:19100")
+	cl := mclient.New(mclient.Config{BaseURL: url})
+
+	hr, err := cl.Boot(context.Background())
+	if err != nil {
+		t.Fatalf("Boot: %v", err)
+	}
+	if hr.Status != "ok" {
+		t.Errorf("status = %q", hr.Status)
+	}
+	wantCaps := map[string]bool{"fts": true, "embedding": true, "ttl": true, "pin": true, "propagation": true}
+	gotCaps := map[string]bool{}
+	for _, c := range hr.Capabilities {
+		gotCaps[c] = true
+	}
+	for c := range wantCaps {
+		if !gotCaps[c] {
+			t.Errorf("capability %q missing — built-in plugin should declare all 5", c)
+		}
+	}
+}
+
+// TestE2E_FullCommitSearchForgetRoundTrip: the full agent flow against
+// real postgres + real HTTP. Catches wire-format regressions (the
+// pq.Array bug we hit during PR-3 development) and contract-level
+// drift between Go bindings and the spec.
+func TestE2E_FullCommitSearchForgetRoundTrip(t *testing.T) {
+	dsn := requireE2EDB(t)
+	binary := buildBinary(t)
+	_, url := startBinary(t, binary, dsn, "127.0.0.1:19101")
+	cl := mclient.New(mclient.Config{BaseURL: url})
+
+	ctx := context.Background()
+	ns := fmt.Sprintf("workspace:e2e-%d", time.Now().UnixNano())
+
+	// 1. Upsert namespace.
+	if _, err := cl.UpsertNamespace(ctx, ns, contract.NamespaceUpsert{Kind: contract.NamespaceKindWorkspace}); err != nil {
+		t.Fatalf("UpsertNamespace: %v", err)
+	}
+	t.Cleanup(func() { _ = cl.DeleteNamespace(context.Background(), ns) })
+
+	// 2. Commit a memory.
+	resp, err := cl.CommitMemory(ctx, ns, contract.MemoryWrite{
+		Content: "user prefers tabs over spaces",
+		Kind:    contract.MemoryKindFact,
+		Source:  contract.MemorySourceAgent,
+	})
+	if err != nil {
+		t.Fatalf("CommitMemory: %v", err)
+	}
+	if resp.ID == "" {
+		t.Fatal("plugin returned empty memory id")
+	}
+
+	// 3. Search and find the memory we just wrote.
+	sresp, err := cl.Search(ctx, contract.SearchRequest{Namespaces: []string{ns}, Query: "tabs"})
+	if err != nil {
+		t.Fatalf("Search: %v", err)
+	}
+	if len(sresp.Memories) == 0 {
+		t.Errorf("Search returned 0 memories, want at least 1")
+	}
+	found := false
+	for _, m := range sresp.Memories {
+		if m.ID == resp.ID && m.Content == "user prefers tabs over spaces" {
+			found = true
+			break
+		}
+	}
+	if !found {
+		got, _ := json.Marshal(sresp.Memories)
+		t.Errorf("committed memory not found in search results: %s", got)
+	}
+
+	// 4. Forget the memory.
+	if err := cl.ForgetMemory(ctx, resp.ID, contract.ForgetRequest{RequestedByNamespace: ns}); err != nil {
+		t.Fatalf("ForgetMemory: %v", err)
+	}
+
+	// 5. Search again — gone.
+	sresp, err = cl.Search(ctx, contract.SearchRequest{Namespaces: []string{ns}, Query: "tabs"})
+	if err != nil {
+		t.Fatalf("Search after forget: %v", err)
+	}
+	for _, m := range sresp.Memories {
+		if m.ID == resp.ID {
+			t.Errorf("forgotten memory still in search results")
+		}
+	}
+}
+
+// TestE2E_IdempotencyKey covers the C1 fix end-to-end: same id passed
+// twice should upsert (one row, updated content), not duplicate.
+func TestE2E_IdempotencyKey(t *testing.T) {
+	dsn := requireE2EDB(t)
+	binary := buildBinary(t)
+	_, url := startBinary(t, binary, dsn, "127.0.0.1:19102")
+	cl := mclient.New(mclient.Config{BaseURL: url})
+
+	ctx := context.Background()
+	ns := fmt.Sprintf("workspace:e2e-idem-%d", time.Now().UnixNano())
+	if _, err := cl.UpsertNamespace(ctx, ns, contract.NamespaceUpsert{Kind: contract.NamespaceKindWorkspace}); err != nil {
+		t.Fatalf("UpsertNamespace: %v", err)
+	}
+	t.Cleanup(func() { _ = cl.DeleteNamespace(context.Background(), ns) })
+
+	fixedID := "11111111-2222-3333-4444-555555555555"
+	for i, content := range []string{"first version", "second version (updated)"} {
+		if _, err := cl.CommitMemory(ctx, ns, contract.MemoryWrite{
+			ID:      fixedID,
+			Content: content,
+			Kind:    contract.MemoryKindFact,
+			Source:  contract.MemorySourceAgent,
+		}); err != nil {
+			t.Fatalf("commit %d: %v", i, err)
+		}
+	}
+
+	sresp, err := cl.Search(ctx, contract.SearchRequest{Namespaces: []string{ns}})
+	if err != nil {
+		t.Fatalf("Search: %v", err)
+	}
+	matches := 0
+	for _, m := range sresp.Memories {
+		if m.ID == fixedID {
+			matches++
+			if m.Content != "second version (updated)" {
+				t.Errorf("upsert did not update content: got %q", m.Content)
+			}
+		}
+	}
+	if matches != 1 {
+		t.Errorf("upsert produced %d rows for id=%s, want 1", matches, fixedID)
+	}
+}

workspace-server/cmd/memory-plugin-postgres/main.go

@@ -0,0 +1,182 @@
+// memory-plugin-postgres is the built-in implementation of the memory
+// plugin contract (RFC #2728). Operators run it next to workspace-
+// server; workspace-server points MEMORY_PLUGIN_URL at it.
+//
+// Owns its own postgres tables (see migrations/). When an operator
+// swaps in a different plugin, this binary's tables become orphaned
+// — not auto-dropped. Document this in the plugin docs (PR-10).
+package main
+
+import (
+	"context"
+	"database/sql"
+	"errors"
+	"fmt"
+	"log"
+	"net"
+	"net/http"
+	"os"
+	"os/signal"
+	"strings"
+	"syscall"
+	"time"
+
+	_ "github.com/lib/pq"
+
+	"github.com/Molecule-AI/molecule-monorepo/platform/internal/memory/pgplugin"
+)
+
+const (
+	envDatabaseURL = "MEMORY_PLUGIN_DATABASE_URL"
+	envListenAddr  = "MEMORY_PLUGIN_LISTEN_ADDR"
+	envSkipMigrate = "MEMORY_PLUGIN_SKIP_MIGRATE"
+
+	defaultListenAddr = ":9100"
+)
+
+func main() {
+	if err := run(); err != nil {
+		log.Fatalf("memory-plugin-postgres: %v", err)
+	}
+}
+
+// run is the boot path. Extracted from main() so tests can drive it
+// with synthesized env. Returns nil on graceful shutdown, an error on
+// failure to bring up.
+func run() error {
+	cfg, err := loadConfig()
+	if err != nil {
+		return fmt.Errorf("config: %w", err)
+	}
+
+	db, err := openDB(cfg.DatabaseURL)
+	if err != nil {
+		return fmt.Errorf("open db: %w", err)
+	}
+	defer db.Close()
+
+	if !cfg.SkipMigrate {
+		if err := runMigrations(db); err != nil {
+			return fmt.Errorf("migrate: %w", err)
+		}
+	}
+
+	store := pgplugin.NewStore(db)
+	handler := pgplugin.NewHandler(store, func() error {
+		ctx, cancel := context.WithTimeout(context.Background(), 2*time.Second)
+		defer cancel()
+		return db.PingContext(ctx)
+	})
+
+	srv := &http.Server{
+		Addr:              cfg.ListenAddr,
+		Handler:           handler,
+		ReadHeaderTimeout: 5 * time.Second,
+	}
+
+	// Listen separately so we can log the bound port (handy when
+	// :0 is used in tests).
+	ln, err := net.Listen("tcp", cfg.ListenAddr)
+	if err != nil {
+		return fmt.Errorf("listen %s: %w", cfg.ListenAddr, err)
+	}
+	log.Printf("memory-plugin-postgres listening on %s", ln.Addr())
+
+	// Run server in a goroutine; main waits on signal.
+	errCh := make(chan error, 1)
+	go func() {
+		if err := srv.Serve(ln); err != nil && !errors.Is(err, http.ErrServerClosed) {
+			errCh <- err
+		}
+	}()
+
+	sigCh := make(chan os.Signal, 1)
+	signal.Notify(sigCh, syscall.SIGINT, syscall.SIGTERM)
+
+	select {
+	case <-sigCh:
+		log.Println("shutdown signal received")
+	case err := <-errCh:
+		return fmt.Errorf("serve: %w", err)
+	}
+
+	ctx, cancel := context.WithTimeout(context.Background(), 10*time.Second)
+	defer cancel()
+	return srv.Shutdown(ctx)
+}
+
+type config struct {
+	DatabaseURL string
+	ListenAddr  string
+	SkipMigrate bool
+}
+
+func loadConfig() (*config, error) {
+	dbURL := strings.TrimSpace(os.Getenv(envDatabaseURL))
+	if dbURL == "" {
+		return nil, fmt.Errorf("%s is required", envDatabaseURL)
+	}
+	addr := strings.TrimSpace(os.Getenv(envListenAddr))
+	if addr == "" {
+		addr = defaultListenAddr
+	}
+	return &config{
+		DatabaseURL: dbURL,
+		ListenAddr:  addr,
+		SkipMigrate: os.Getenv(envSkipMigrate) == "1",
+	}, nil
+}
+
+func openDB(databaseURL string) (*sql.DB, error) {
+	db, err := sql.Open("postgres", databaseURL)
+	if err != nil {
+		return nil, err
+	}
+	db.SetMaxOpenConns(25)
+	db.SetMaxIdleConns(5)
+	db.SetConnMaxLifetime(30 * time.Minute)
+	ctx, cancel := context.WithTimeout(context.Background(), 5*time.Second)
+	defer cancel()
+	if err := db.PingContext(ctx); err != nil {
+		return nil, fmt.Errorf("ping: %w", err)
+	}
+	return db, nil
+}
+
+// runMigrations applies the schema migrations bundled at
+// cmd/memory-plugin-postgres/migrations/. Idempotent on repeat boot.
+//
+// Implementation note: rather than embedding the full migrate engine,
+// we read the migration files at boot from a known relative path. The
+// down migrations are deliberately NOT applied here — that's a manual
+// operator action. This keeps the binary tiny and avoids dragging in
+// golang-migrate's drivers.
+func runMigrations(db *sql.DB) error {
+	// Find the migrations directory. In `go run` mode it's relative
+	// to the cmd dir; in the prebuilt binary case it's expected next
+	// to the binary OR via env var override.
+	dir := os.Getenv("MEMORY_PLUGIN_MIGRATIONS_DIR")
+	if dir == "" {
+		// Best-effort: try the cwd-relative path that works for `go test`.
+		dir = "cmd/memory-plugin-postgres/migrations"
+	}
+	entries, err := os.ReadDir(dir)
+	if err != nil {
+		return fmt.Errorf("read migrations dir %q: %w", dir, err)
+	}
+	for _, e := range entries {
+		if e.IsDir() || !strings.HasSuffix(e.Name(), ".up.sql") {
+			continue
+		}
+		path := dir + "/" + e.Name()
+		data, err := os.ReadFile(path)
+		if err != nil {
+			return fmt.Errorf("read %q: %w", path, err)
+		}
+		if _, err := db.Exec(string(data)); err != nil {
+			return fmt.Errorf("apply %q: %w", path, err)
+		}
+		log.Printf("applied migration %s", e.Name())
+	}
+	return nil
+}

workspace-server/cmd/memory-plugin-postgres/migrations/001_memory_v2.down.sql

@@ -0,0 +1,3 @@
+-- Down migration for memory_v2 plugin schema (RFC #2728).
+DROP TABLE IF EXISTS memory_records;
+DROP TABLE IF EXISTS memory_namespaces;

workspace-server/cmd/memory-plugin-postgres/migrations/001_memory_v2.up.sql

@@ -0,0 +1,47 @@
+-- Memory v2 plugin schema (RFC #2728).
+--
+-- These tables are owned by the built-in postgres memory plugin, NOT
+-- by workspace-server. When an operator swaps in a different memory
+-- plugin (Pinecone, Letta, custom), these tables become orphaned —
+-- not auto-dropped. Operator drops them when they're confident they
+-- don't want to switch back.
+--
+-- Lives under cmd/memory-plugin-postgres/migrations/ (NOT
+-- workspace-server/migrations/) to make the ownership boundary
+-- visible: workspace-server has zero knowledge of these tables.
+
+CREATE EXTENSION IF NOT EXISTS vector;
+
+CREATE TABLE IF NOT EXISTS memory_namespaces (
+    name        TEXT PRIMARY KEY,
+    kind        TEXT NOT NULL CHECK (kind IN ('workspace','team','org','custom')),
+    expires_at  TIMESTAMPTZ,
+    metadata    JSONB,
+    created_at  TIMESTAMPTZ NOT NULL DEFAULT now()
+);
+
+CREATE TABLE IF NOT EXISTS memory_records (
+    id           UUID PRIMARY KEY DEFAULT gen_random_uuid(),
+    namespace    TEXT NOT NULL REFERENCES memory_namespaces(name) ON DELETE CASCADE,
+    content      TEXT NOT NULL,
+    kind         TEXT NOT NULL CHECK (kind IN ('fact','summary','checkpoint')),
+    source       TEXT NOT NULL CHECK (source IN ('agent','runtime','user')),
+    expires_at   TIMESTAMPTZ,
+    propagation  JSONB,
+    pin          BOOLEAN NOT NULL DEFAULT false,
+    embedding    vector(1536),
+    content_tsv  tsvector GENERATED ALWAYS AS (to_tsvector('english', content)) STORED,
+    created_at   TIMESTAMPTZ NOT NULL DEFAULT now()
+);
+
+-- Indexes:
+--  - namespace: every search filters by namespace list
+--  - content_tsv: FTS path
+--  - embedding: semantic search (partial because most rows have no embedding)
+--  - expires_at: TTL janitor scans
+CREATE INDEX IF NOT EXISTS idx_memory_records_namespace ON memory_records(namespace);
+CREATE INDEX IF NOT EXISTS idx_memory_records_fts ON memory_records USING GIN (content_tsv);
+CREATE INDEX IF NOT EXISTS idx_memory_records_embedding ON memory_records
+    USING ivfflat (embedding) WHERE embedding IS NOT NULL;
+CREATE INDEX IF NOT EXISTS idx_memory_records_expires ON memory_records (expires_at)
+    WHERE expires_at IS NOT NULL;

workspace-server/cmd/server/main.go

@@ -18,6 +18,7 @@ import (
 	"github.com/Molecule-AI/molecule-monorepo/platform/internal/events"
 	"github.com/Molecule-AI/molecule-monorepo/platform/internal/handlers"
 	"github.com/Molecule-AI/molecule-monorepo/platform/internal/imagewatch"
+	memwiring "github.com/Molecule-AI/molecule-monorepo/platform/internal/memory/wiring"
 	"github.com/Molecule-AI/molecule-monorepo/platform/internal/provisioner"
 	"github.com/Molecule-AI/molecule-monorepo/platform/internal/registry"
 	"github.com/Molecule-AI/molecule-monorepo/platform/internal/router"
@@ -166,6 +167,16 @@ func main() {
 		wh.SetCPProvisioner(cpProv)
 	}
 
+	// Memory v2 plugin (RFC #2728): build the dependency bundle once
+	// here so all three handlers (MCPHandler, AdminMemoriesHandler,
+	// WorkspaceHandler) get the same plugin/resolver pair. memBundle
+	// is nil when MEMORY_PLUGIN_URL is unset — every consumer
+	// nil-checks before using.
+	memBundle := memwiring.Build(db.DB)
+	if memBundle != nil {
+		wh.WithNamespaceCleanup(memBundle.NamespaceCleanupFn())
+	}
+
 	// External-plugin env mutators — each plugin contributes 0+ mutators
 	// onto a shared registry. Order matters: gh-identity populates
 	// MOLECULE_AGENT_ROLE-derived attribution env vars that downstream
@@ -306,7 +317,7 @@ func main() {
 	cronSched.SetChannels(channelMgr)
 
 	// Router
-	r := router.Setup(hub, broadcaster, prov, platformURL, configsDir, wh, channelMgr)
+	r := router.Setup(hub, broadcaster, prov, platformURL, configsDir, wh, channelMgr, memBundle)
 
 	// HTTP server with graceful shutdown
 	srv := &http.Server{

workspace-server/internal/handlers/admin_memories.go

@@ -1,23 +1,83 @@
 package handlers
 
 import (
+	"context"
+	"database/sql"
 	"log"
 	"net/http"
+	"os"
+	"strings"
 	"time"
 
 	"github.com/Molecule-AI/molecule-monorepo/platform/internal/db"
+	mclient "github.com/Molecule-AI/molecule-monorepo/platform/internal/memory/client"
+	"github.com/Molecule-AI/molecule-monorepo/platform/internal/memory/contract"
+	"github.com/Molecule-AI/molecule-monorepo/platform/internal/memory/namespace"
 	"github.com/gin-gonic/gin"
 )
 
+// envMemoryV2Cutover gates whether admin export/import routes through
+// the v2 plugin (PR-8 / RFC #2728). When unset, the legacy direct-DB
+// path runs unchanged so operators who haven't enabled the plugin
+// keep working.
+const envMemoryV2Cutover = "MEMORY_V2_CUTOVER"
+
 // AdminMemoriesHandler provides bulk export/import of agent memories for
 // backup and restore across Docker rebuilds (issue #1051).
-type AdminMemoriesHandler struct{}
+//
+// PR-8 (RFC #2728): when wired with the v2 plugin via WithMemoryV2 AND
+// MEMORY_V2_CUTOVER is true, export reads from the plugin's namespaces
+// and import writes through the plugin. Both paths preserve the
+// SAFE-T1201 redaction shipped in F1084 + F1085.
+type AdminMemoriesHandler struct {
+	plugin   adminMemoriesPlugin
+	resolver adminMemoriesResolver
+}
+
+// adminMemoriesPlugin is the slice of the memory plugin client we
+// call from this handler.
+type adminMemoriesPlugin interface {
+	CommitMemory(ctx context.Context, namespace string, body contract.MemoryWrite) (*contract.MemoryWriteResponse, error)
+	Search(ctx context.Context, body contract.SearchRequest) (*contract.SearchResponse, error)
+	UpsertNamespace(ctx context.Context, name string, body contract.NamespaceUpsert) (*contract.Namespace, error)
+}
+
+// adminMemoriesResolver mirrors the namespace resolver methods this
+// handler calls.
+type adminMemoriesResolver interface {
+	WritableNamespaces(ctx context.Context, workspaceID string) ([]namespace.Namespace, error)
+	ReadableNamespaces(ctx context.Context, workspaceID string) ([]namespace.Namespace, error)
+}
 
 // NewAdminMemoriesHandler constructs the handler.
 func NewAdminMemoriesHandler() *AdminMemoriesHandler {
 	return &AdminMemoriesHandler{}
 }
 
+// WithMemoryV2 attaches the v2 plugin + resolver. Production wiring
+// path; main.go calls this after Boot()-ing the plugin client.
+func (h *AdminMemoriesHandler) WithMemoryV2(plugin *mclient.Client, resolver *namespace.Resolver) *AdminMemoriesHandler {
+	h.plugin = plugin
+	h.resolver = resolver
+	return h
+}
+
+// withMemoryV2APIs is the test-only wiring that takes interfaces.
+func (h *AdminMemoriesHandler) withMemoryV2APIs(plugin adminMemoriesPlugin, resolver adminMemoriesResolver) *AdminMemoriesHandler {
+	h.plugin = plugin
+	h.resolver = resolver
+	return h
+}
+
+// cutoverActive reports whether the export/import path should route
+// through the v2 plugin.
+func (h *AdminMemoriesHandler) cutoverActive() bool {
+	if os.Getenv(envMemoryV2Cutover) != "true" {
+		return false
+	}
+	return h.plugin != nil && h.resolver != nil
+}
+
 // memoryExportEntry is the JSON shape for a single exported memory.
 type memoryExportEntry struct {
 	ID            string    `json:"id"`
@@ -36,9 +96,17 @@ type memoryExportEntry struct {
 // SECURITY (F1084 / #1131): applies redactSecrets to each content field
 // before returning so that any credentials stored before SAFE-T1201 (#838)
 // was applied do not leak out via the admin export endpoint.
+//
+// CUTOVER (PR-8 / RFC #2728): when MEMORY_V2_CUTOVER=true and the v2
+// plugin is wired, reads from the plugin instead of agent_memories.
 func (h *AdminMemoriesHandler) Export(c *gin.Context) {
 	ctx := c.Request.Context()
 
+	if h.cutoverActive() {
+		h.exportViaPlugin(c, ctx)
+		return
+	}
+
 	rows, err := db.DB.QueryContext(ctx, `
 		SELECT am.id, am.content, am.scope, am.namespace, am.created_at,
 		       w.name AS workspace_name
@@ -91,6 +159,9 @@ type memoryImportEntry struct {
 // before both the deduplication check and the INSERT so that imported memories
 // with embedded credentials cannot land unredacted in agent_memories (SAFE-T1201
 // parity with the commit_memory MCP bridge path).
+//
+// CUTOVER (PR-8 / RFC #2728): when MEMORY_V2_CUTOVER=true and the v2
+// plugin is wired, writes through the plugin instead of agent_memories.
 func (h *AdminMemoriesHandler) Import(c *gin.Context) {
 	ctx := c.Request.Context()
 
@@ -100,6 +171,11 @@ func (h *AdminMemoriesHandler) Import(c *gin.Context) {
 		return
 	}
 
+	if h.cutoverActive() {
+		h.importViaPlugin(c, ctx, entries)
+		return
+	}
+
 	imported := 0
 	skipped := 0
 	errors := 0
@@ -175,3 +251,310 @@ func (h *AdminMemoriesHandler) Import(c *gin.Context) {
 		"total":    len(entries),
 	})
 }
+
+// exportViaPlugin reads memories from the v2 plugin and emits them in
+// the legacy memoryExportEntry shape so existing tooling that consumes
+// the export keeps working.
+//
+// Optimization (#289 fix): the previous implementation was O(workspaces)
+// in BOTH resolver CTE walks AND plugin search calls. For a 1000-tenant
+// org, that's 1000 × resolver + 1000 × HTTP, where most are redundant
+// because workspaces sharing a team/org root see identical namespaces.
+//
+// New strategy:
+//   1. Single SQL pass walks parent_id chains, returning each
+//      workspace's root_id alongside its name.
+//   2. Group workspaces by root → unique tree count is typically <<
+//      workspace count.
+//   3. Resolve namespaces ONCE per root (any workspace under that
+//      root produces the same readable list).
+//   4. Build a UNION of namespaces across all roots; single plugin
+//      search call.
+//   5. Map each memory back to a workspace_name via a namespace→ws
+//      lookup table built up from step 3.
+//
+// Net cost: 1 SQL + N_roots resolver calls + 1 plugin call (vs
+// N_workspaces resolver + N_workspaces plugin in the old code).
+func (h *AdminMemoriesHandler) exportViaPlugin(c *gin.Context, ctx context.Context) {
+	// 1. One SQL pass: every workspace + its root id.
+	wsRows, err := loadWorkspacesWithRoots(ctx, db.DB)
+	if err != nil {
+		log.Printf("admin/memories/export (cutover): workspaces query: %v", err)
+		c.JSON(http.StatusInternalServerError, gin.H{"error": "export query failed"})
+		return
+	}
+
+	// 2. Group by root → list of workspaces.
+	rootToWorkspaces := make(map[string][]workspaceRow, len(wsRows))
+	for _, w := range wsRows {
+		rootToWorkspaces[w.RootID] = append(rootToWorkspaces[w.RootID], w)
+	}
+
+	// 3. Resolve team/org namespaces once per root, then add each
+	// member's private workspace:<id> namespace explicitly.
+	//
+	// IMPORTANT: ReadableNamespaces(rootID) returns
+	// {workspace:rootID, team:rootID, org:rootID}. Calling it once
+	// per root is enough for team:/org:/custom: (those are shared by
+	// every member of the root group), but the workspace: namespace
+	// it returns is rootID's only — child members' private
+	// workspace:<childID> namespaces would be silently dropped from
+	// the export. Inject each member's workspace:<id> below to keep
+	// coverage parity with the legacy per-workspace iteration.
+	nsToOwner := make(map[string]string)       // namespace → workspace_name (first matching wins)
+	allNamespaces := make(map[string]struct{}) // union for plugin search
+	for rootID, members := range rootToWorkspaces {
+		readable, err := h.resolver.ReadableNamespaces(ctx, rootID)
+		if err != nil {
+			log.Printf("admin/memories/export (cutover) root=%s: resolve: %v", rootID, err)
+			continue
+		}
+		// Collect non-workspace namespaces (team:/org:/custom:/...) from
+		// the root view; these are identical across every member.
+		for _, ns := range readable {
+			if strings.HasPrefix(ns.Name, "workspace:") {
+				continue
+			}
+			allNamespaces[ns.Name] = struct{}{}
+			if _, alreadyMapped := nsToOwner[ns.Name]; alreadyMapped {
+				continue
+			}
+			if owner := pickOwnerForNamespace(ns.Name, members); owner != "" {
+				nsToOwner[ns.Name] = owner
+			}
+		}
+		// Inject each member's private workspace:<id> namespace + its
+		// owner. Children's private memories live in workspace:<childID>
+		// which the root-only resolve doesn't surface.
+		for _, m := range members {
+			ns := "workspace:" + m.ID
+			allNamespaces[ns] = struct{}{}
+			nsToOwner[ns] = m.Name
+		}
+	}
+
+	if len(allNamespaces) == 0 {
+		c.JSON(http.StatusOK, []memoryExportEntry{})
+		return
+	}
+
+	// 4. Single plugin search across the union.
+	nsList := make([]string, 0, len(allNamespaces))
+	for ns := range allNamespaces {
+		nsList = append(nsList, ns)
+	}
+	resp, err := h.plugin.Search(ctx, contract.SearchRequest{Namespaces: nsList, Limit: 100})
+	if err != nil {
+		log.Printf("admin/memories/export (cutover): plugin search: %v", err)
+		c.JSON(http.StatusOK, []memoryExportEntry{})
+		return
+	}
+
+	// 5. Map each memory to a workspace_name, redact, emit.
+	seen := make(map[string]struct{})
+	memories := make([]memoryExportEntry, 0, len(resp.Memories))
+	for _, m := range resp.Memories {
+		if _, dup := seen[m.ID]; dup {
+			continue
+		}
+		seen[m.ID] = struct{}{}
+		owner := nsToOwner[m.Namespace]
+		redacted, _ := redactSecrets(owner, m.Content)
+		memories = append(memories, memoryExportEntry{
+			ID:            m.ID,
+			Content:       redacted,
+			Scope:         legacyScopeFromNamespace(m.Namespace),
+			Namespace:     m.Namespace,
+			CreatedAt:     m.CreatedAt,
+			WorkspaceName: owner,
+		})
+	}
+	c.JSON(http.StatusOK, memories)
+}
+
+// workspaceRow bundles the per-workspace fields the optimized export
+// needs (id + name + root for grouping).
+type workspaceRow struct {
+	ID     string
+	Name   string
+	RootID string
+}
+
+// loadWorkspacesWithRoots returns one row per workspace with its root
+// id computed via a recursive CTE. Single SQL pass — replaces the
+// previous N×ReadableNamespaces pattern that walked each tree
+// independently.
+func loadWorkspacesWithRoots(ctx context.Context, conn *sql.DB) ([]workspaceRow, error) {
+	rows, err := conn.QueryContext(ctx, `
+		WITH RECURSIVE chain AS (
+			SELECT id, parent_id, name, id AS root_id, 0 AS depth
+			FROM workspaces
+			WHERE parent_id IS NULL
+			UNION ALL
+			SELECT w.id, w.parent_id, w.name, c.root_id, c.depth + 1
+			FROM workspaces w
+			JOIN chain c ON w.parent_id = c.id
+			WHERE c.depth < 50
+		)
+		SELECT id::text, name, root_id::text FROM chain ORDER BY name
+	`)
+	if err != nil {
+		return nil, err
+	}
+	defer rows.Close()
+	out := make([]workspaceRow, 0)
+	for rows.Next() {
+		var w workspaceRow
+		if err := rows.Scan(&w.ID, &w.Name, &w.RootID); err != nil {
+			return nil, err
+		}
+		out = append(out, w)
+	}
+	return out, rows.Err()
+}
+
+// pickOwnerForNamespace returns the workspace_name to attribute a
+// namespace to in the export. workspace:<id> namespaces map to the
+// matching member; team:* / org:* / custom:* fall back to the first
+// member of the root group (canonical owner).
+func pickOwnerForNamespace(ns string, members []workspaceRow) string {
+	if strings.HasPrefix(ns, "workspace:") {
+		wantID := strings.TrimPrefix(ns, "workspace:")
+		for _, m := range members {
+			if m.ID == wantID {
+				return m.Name
+			}
+		}
+	}
+	// Non-workspace namespaces: attribute to first member of the root
+	// group. Stable because loadWorkspacesWithRoots returns ORDER BY
+	// name, so the same root group always picks the same owner.
+	if len(members) > 0 {
+		return members[0].Name
+	}
+	return ""
+}
+
+// importViaPlugin writes the entries through the plugin instead of
+// directly to agent_memories. Workspaces are resolved by name like
+// the legacy path. Scope→namespace mapping mirrors the PR-6 shim.
+func (h *AdminMemoriesHandler) importViaPlugin(c *gin.Context, ctx context.Context, entries []memoryImportEntry) {
+	imported := 0
+	skipped := 0
+	errs := 0
+
+	for _, entry := range entries {
+		var workspaceID string
+		if err := db.DB.QueryRowContext(ctx,
+			`SELECT id::text FROM workspaces WHERE name = $1 LIMIT 1`,
+			entry.WorkspaceName,
+		).Scan(&workspaceID); err != nil {
+			log.Printf("admin/memories/import (cutover): workspace %q not found, skipping", entry.WorkspaceName)
+			skipped++
+			continue
+		}
+
+		// Redact BEFORE the plugin sees it (SAFE-T1201 parity).
+		content, _ := redactSecrets(workspaceID, entry.Content)
+
+		ns, err := h.scopeToWritableNamespaceForImport(ctx, workspaceID, entry.Scope)
+		if err != nil {
+			log.Printf("admin/memories/import (cutover): %v", err)
+			skipped++
+			continue
+		}
+
+		// Idempotent namespace upsert before commit.
+		if _, err := h.plugin.UpsertNamespace(ctx, ns, contract.NamespaceUpsert{
+			Kind: namespaceKindFromLegacyScope(entry.Scope),
+		}); err != nil {
+			log.Printf("admin/memories/import (cutover): upsert ns %s: %v", ns, err)
+			errs++
+			continue
+		}
+
+		if _, err := h.plugin.CommitMemory(ctx, ns, contract.MemoryWrite{
+			Content: content,
+			Kind:    contract.MemoryKindFact,
+			Source:  contract.MemorySourceAgent,
+		}); err != nil {
+			log.Printf("admin/memories/import (cutover): commit %s: %v", ns, err)
+			errs++
+			continue
+		}
+		imported++
+	}
+
+	c.JSON(http.StatusOK, gin.H{
+		"imported": imported,
+		"skipped":  skipped,
+		"errors":   errs,
+		"total":    len(entries),
+	})
+}
+
+// scopeToWritableNamespaceForImport mirrors the PR-6 shim translation.
+// Returns the namespace string the resolver picks for the requested
+// scope; errors out cleanly on GLOBAL or unmapped values so importing
+// a malformed entry doesn't crash the run.
+func (h *AdminMemoriesHandler) scopeToWritableNamespaceForImport(ctx context.Context, workspaceID, scope string) (string, error) {
+	writable, err := h.resolver.WritableNamespaces(ctx, workspaceID)
+	if err != nil {
+		return "", err
+	}
+	wantKind := contract.NamespaceKindWorkspace
+	switch strings.ToUpper(scope) {
+	case "", "LOCAL":
+		wantKind = contract.NamespaceKindWorkspace
+	case "TEAM":
+		wantKind = contract.NamespaceKindTeam
+	case "GLOBAL":
+		wantKind = contract.NamespaceKindOrg
+	default:
+		return "", &skipImport{reason: "unknown scope: " + scope}
+	}
+	for _, ns := range writable {
+		if ns.Kind == wantKind {
+			return ns.Name, nil
+		}
+	}
+	return "", &skipImport{reason: "no writable namespace of kind " + string(wantKind)}
+}
+
+// skipImport is a typed error so the caller can distinguish "skip
+// this entry" from a hard failure.
+type skipImport struct{ reason string }
+
+func (e *skipImport) Error() string { return "skip: " + e.reason }
+
+// legacyScopeFromNamespace reverses the namespace→scope mapping for
+// the export shape. Mirrors namespaceKindToLegacyScope from the PR-6
+// shim but is lifted out so admin_memories doesn't depend on the MCP
+// handler's helpers.
+func legacyScopeFromNamespace(ns string) string {
+	switch {
+	case strings.HasPrefix(ns, "workspace:"):
+		return "LOCAL"
+	case strings.HasPrefix(ns, "team:"):
+		return "TEAM"
+	case strings.HasPrefix(ns, "org:"):
+		return "GLOBAL"
+	default:
+		return ""
+	}
+}
+
+// namespaceKindFromLegacyScope returns the contract.NamespaceKind for
+// a legacy scope value. Unknown defaults to workspace so importing
+// an unexpected row still produces a typed namespace.
+func namespaceKindFromLegacyScope(scope string) contract.NamespaceKind {
+	switch strings.ToUpper(scope) {
+	case "TEAM":
+		return contract.NamespaceKindTeam
+	case "GLOBAL":
+		return contract.NamespaceKindOrg
+	default:
+		return contract.NamespaceKindWorkspace
+	}
+}
+

workspace-server/internal/handlers/admin_memories_cutover_test.go

@@ -0,0 +1,800 @@
+package handlers
+
+import (
+	"bytes"
+	"context"
+	"encoding/json"
+	"errors"
+	"net/http"
+	"net/http/httptest"
+	"strings"
+	"testing"
+	"time"
+
+	"github.com/DATA-DOG/go-sqlmock"
+	"github.com/gin-gonic/gin"
+
+	platformdb "github.com/Molecule-AI/molecule-monorepo/platform/internal/db"
+	"github.com/Molecule-AI/molecule-monorepo/platform/internal/memory/contract"
+	"github.com/Molecule-AI/molecule-monorepo/platform/internal/memory/namespace"
+)
+
+// --- stubs ---
+
+type stubAdminPlugin struct {
+	upserts  []string
+	commits  []commitRecord
+	searches []contract.SearchRequest
+	commitFn func(ctx context.Context, ns string, body contract.MemoryWrite) (*contract.MemoryWriteResponse, error)
+	searchFn func(ctx context.Context, body contract.SearchRequest) (*contract.SearchResponse, error)
+	upsertFn func(ctx context.Context, name string, body contract.NamespaceUpsert) (*contract.Namespace, error)
+}
+
+type commitRecord struct {
+	NS      string
+	Content string
+}
+
+func (s *stubAdminPlugin) UpsertNamespace(ctx context.Context, name string, body contract.NamespaceUpsert) (*contract.Namespace, error) {
+	s.upserts = append(s.upserts, name)
+	if s.upsertFn != nil {
+		return s.upsertFn(ctx, name, body)
+	}
+	return &contract.Namespace{Name: name, Kind: body.Kind, CreatedAt: time.Now().UTC()}, nil
+}
+func (s *stubAdminPlugin) CommitMemory(ctx context.Context, ns string, body contract.MemoryWrite) (*contract.MemoryWriteResponse, error) {
+	s.commits = append(s.commits, commitRecord{NS: ns, Content: body.Content})
+	if s.commitFn != nil {
+		return s.commitFn(ctx, ns, body)
+	}
+	return &contract.MemoryWriteResponse{ID: "out-1", Namespace: ns}, nil
+}
+func (s *stubAdminPlugin) Search(ctx context.Context, body contract.SearchRequest) (*contract.SearchResponse, error) {
+	s.searches = append(s.searches, body)
+	if s.searchFn != nil {
+		return s.searchFn(ctx, body)
+	}
+	return &contract.SearchResponse{}, nil
+}
+
+type stubAdminResolver struct {
+	readable []namespace.Namespace
+	writable []namespace.Namespace
+	err      error
+}
+
+func (s *stubAdminResolver) ReadableNamespaces(_ context.Context, _ string) ([]namespace.Namespace, error) {
+	return s.readable, s.err
+}
+func (s *stubAdminResolver) WritableNamespaces(_ context.Context, _ string) ([]namespace.Namespace, error) {
+	return s.writable, s.err
+}
+
+func adminRootResolver() *stubAdminResolver {
+	return &stubAdminResolver{
+		readable: []namespace.Namespace{
+			{Name: "workspace:root-1", Kind: contract.NamespaceKindWorkspace, Writable: true},
+			{Name: "team:root-1", Kind: contract.NamespaceKindTeam, Writable: true},
+			{Name: "org:root-1", Kind: contract.NamespaceKindOrg, Writable: true},
+		},
+		writable: []namespace.Namespace{
+			{Name: "workspace:root-1", Kind: contract.NamespaceKindWorkspace, Writable: true},
+			{Name: "team:root-1", Kind: contract.NamespaceKindTeam, Writable: true},
+			{Name: "org:root-1", Kind: contract.NamespaceKindOrg, Writable: true},
+		},
+	}
+}
+
+// installMockDB swaps platformdb.DB with a sqlmock for a test.
+func installMockDB(t *testing.T) sqlmock.Sqlmock {
+	t.Helper()
+	mockDB, mock, err := sqlmock.New()
+	if err != nil {
+		t.Fatalf("sqlmock new: %v", err)
+	}
+	prev := platformdb.DB
+	platformdb.DB = mockDB
+	t.Cleanup(func() {
+		_ = mockDB.Close()
+		platformdb.DB = prev
+	})
+	return mock
+}
+
+// --- cutoverActive ---
+
+func TestCutoverActive(t *testing.T) {
+	cases := []struct {
+		name     string
+		envVal   string
+		plugin   adminMemoriesPlugin
+		resolver adminMemoriesResolver
+		want     bool
+	}{
+		{"env unset", "", &stubAdminPlugin{}, adminRootResolver(), false},
+		{"env true but unwired", "true", nil, nil, false},
+		{"env false", "false", &stubAdminPlugin{}, adminRootResolver(), false},
+		{"env true wired", "true", &stubAdminPlugin{}, adminRootResolver(), true},
+	}
+	for _, tc := range cases {
+		t.Run(tc.name, func(t *testing.T) {
+			t.Setenv(envMemoryV2Cutover, tc.envVal)
+			h := &AdminMemoriesHandler{plugin: tc.plugin, resolver: tc.resolver}
+			if got := h.cutoverActive(); got != tc.want {
+				t.Errorf("got %v, want %v", got, tc.want)
+			}
+		})
+	}
+}
+
+// --- WithMemoryV2 wiring ---
+
+func TestWithMemoryV2_AttachesDeps(t *testing.T) {
+	h := NewAdminMemoriesHandler().WithMemoryV2(nil, nil)
+	// Both nil pointers — wiring still attaches them; cutoverActive
+	// reports false because the interface values are nil.
+	if h.plugin == nil && h.resolver == nil {
+		// expected
+	}
+}
+
+func TestWithMemoryV2APIs_AttachesDeps(t *testing.T) {
+	h := NewAdminMemoriesHandler().withMemoryV2APIs(&stubAdminPlugin{}, adminRootResolver())
+	if h.plugin == nil || h.resolver == nil {
+		t.Error("withMemoryV2APIs must attach both interfaces")
+	}
+}
+
+// --- Export via plugin ---
+
+func TestExport_RoutesThroughPluginWhenCutoverActive(t *testing.T) {
+	t.Setenv(envMemoryV2Cutover, "true")
+	mock := installMockDB(t)
+
+	mock.ExpectQuery("WITH RECURSIVE chain").
+		WillReturnRows(sqlmock.NewRows([]string{"id", "name", "root_id"}).
+			AddRow("ws-1", "alpha", "ws-1"))
+
+	plugin := &stubAdminPlugin{
+		searchFn: func(_ context.Context, body contract.SearchRequest) (*contract.SearchResponse, error) {
+			return &contract.SearchResponse{Memories: []contract.Memory{
+				{ID: "mem-1", Namespace: "workspace:root-1", Content: "fact x", Kind: contract.MemoryKindFact, Source: contract.MemorySourceAgent, CreatedAt: time.Now().UTC()},
+				{ID: "mem-2", Namespace: "team:root-1", Content: "team y", Kind: contract.MemoryKindFact, Source: contract.MemorySourceAgent, CreatedAt: time.Now().UTC()},
+			}}, nil
+		},
+	}
+	h := NewAdminMemoriesHandler().withMemoryV2APIs(plugin, adminRootResolver())
+
+	gin.SetMode(gin.TestMode)
+	w := httptest.NewRecorder()
+	c, _ := gin.CreateTestContext(w)
+	c.Request = httptest.NewRequest("GET", "/admin/memories/export", nil)
+	h.Export(c)
+
+	if w.Code != http.StatusOK {
+		t.Fatalf("code = %d body=%s", w.Code, w.Body.String())
+	}
+	var entries []memoryExportEntry
+	if err := json.Unmarshal(w.Body.Bytes(), &entries); err != nil {
+		t.Fatalf("unmarshal: %v", err)
+	}
+	if len(entries) != 2 {
+		t.Errorf("entries = %d", len(entries))
+	}
+	// Legacy scope label must be in the export
+	scopes := map[string]bool{}
+	for _, e := range entries {
+		scopes[e.Scope] = true
+	}
+	if !scopes["LOCAL"] || !scopes["TEAM"] {
+		t.Errorf("expected LOCAL+TEAM scopes, got %v", scopes)
+	}
+}
+
+func TestExport_DeduplicatesByMemoryID(t *testing.T) {
+	t.Setenv(envMemoryV2Cutover, "true")
+	mock := installMockDB(t)
+
+	// Two workspaces, both will see the same team-shared memory.
+	mock.ExpectQuery("WITH RECURSIVE chain").
+		WillReturnRows(sqlmock.NewRows([]string{"id", "name", "root_id"}).
+			AddRow("ws-1", "alpha", "ws-1").
+			AddRow("ws-2", "beta", "ws-2"))
+
+	plugin := &stubAdminPlugin{
+		searchFn: func(_ context.Context, body contract.SearchRequest) (*contract.SearchResponse, error) {
+			return &contract.SearchResponse{Memories: []contract.Memory{
+				{ID: "mem-shared", Namespace: "team:root-1", Content: "team-fact", Kind: contract.MemoryKindFact, Source: contract.MemorySourceAgent, CreatedAt: time.Now().UTC()},
+			}}, nil
+		},
+	}
+	h := NewAdminMemoriesHandler().withMemoryV2APIs(plugin, adminRootResolver())
+	gin.SetMode(gin.TestMode)
+	w := httptest.NewRecorder()
+	c, _ := gin.CreateTestContext(w)
+	c.Request = httptest.NewRequest("GET", "/admin/memories/export", nil)
+	h.Export(c)
+
+	var entries []memoryExportEntry
+	_ = json.Unmarshal(w.Body.Bytes(), &entries)
+	if len(entries) != 1 {
+		t.Errorf("dedup failed; got %d entries, want 1", len(entries))
+	}
+}
+
+func TestExport_SkipsWorkspaceWhenResolverFails(t *testing.T) {
+	t.Setenv(envMemoryV2Cutover, "true")
+	mock := installMockDB(t)
+	mock.ExpectQuery("WITH RECURSIVE chain").
+		WillReturnRows(sqlmock.NewRows([]string{"id", "name", "root_id"}).
+			AddRow("ws-1", "alpha", "ws-1"))
+
+	plugin := &stubAdminPlugin{}
+	resolver := &stubAdminResolver{err: errors.New("resolver dead")}
+	h := NewAdminMemoriesHandler().withMemoryV2APIs(plugin, resolver)
+	gin.SetMode(gin.TestMode)
+	w := httptest.NewRecorder()
+	c, _ := gin.CreateTestContext(w)
+	c.Request = httptest.NewRequest("GET", "/admin/memories/export", nil)
+	h.Export(c)
+
+	// Should still 200 with empty memories — failure is per-workspace.
+	if w.Code != http.StatusOK {
+		t.Errorf("code = %d body=%s", w.Code, w.Body.String())
+	}
+}
+
+func TestExport_SkipsWorkspaceWhenPluginSearchFails(t *testing.T) {
+	t.Setenv(envMemoryV2Cutover, "true")
+	mock := installMockDB(t)
+	mock.ExpectQuery("WITH RECURSIVE chain").
+		WillReturnRows(sqlmock.NewRows([]string{"id", "name", "root_id"}).
+			AddRow("ws-1", "alpha", "ws-1"))
+
+	plugin := &stubAdminPlugin{
+		searchFn: func(_ context.Context, _ contract.SearchRequest) (*contract.SearchResponse, error) {
+			return nil, errors.New("plugin dead")
+		},
+	}
+	h := NewAdminMemoriesHandler().withMemoryV2APIs(plugin, adminRootResolver())
+	gin.SetMode(gin.TestMode)
+	w := httptest.NewRecorder()
+	c, _ := gin.CreateTestContext(w)
+	c.Request = httptest.NewRequest("GET", "/admin/memories/export", nil)
+	h.Export(c)
+
+	if w.Code != http.StatusOK {
+		t.Errorf("code = %d", w.Code)
+	}
+}
+
+func TestExport_WorkspacesQueryFails(t *testing.T) {
+	t.Setenv(envMemoryV2Cutover, "true")
+	mock := installMockDB(t)
+	mock.ExpectQuery("WITH RECURSIVE chain").
+		WillReturnError(errors.New("db dead"))
+
+	plugin := &stubAdminPlugin{}
+	h := NewAdminMemoriesHandler().withMemoryV2APIs(plugin, adminRootResolver())
+	gin.SetMode(gin.TestMode)
+	w := httptest.NewRecorder()
+	c, _ := gin.CreateTestContext(w)
+	c.Request = httptest.NewRequest("GET", "/admin/memories/export", nil)
+	h.Export(c)
+
+	if w.Code != http.StatusInternalServerError {
+		t.Errorf("code = %d, want 500", w.Code)
+	}
+}
+
+func TestExport_EmptyReadable(t *testing.T) {
+	t.Setenv(envMemoryV2Cutover, "true")
+	mock := installMockDB(t)
+	mock.ExpectQuery("WITH RECURSIVE chain").
+		WillReturnRows(sqlmock.NewRows([]string{"id", "name", "root_id"}).
+			AddRow("ws-1", "alpha", "ws-1"))
+
+	resolver := &stubAdminResolver{readable: []namespace.Namespace{}}
+	h := NewAdminMemoriesHandler().withMemoryV2APIs(&stubAdminPlugin{}, resolver)
+	gin.SetMode(gin.TestMode)
+	w := httptest.NewRecorder()
+	c, _ := gin.CreateTestContext(w)
+	c.Request = httptest.NewRequest("GET", "/admin/memories/export", nil)
+	h.Export(c)
+	if w.Code != http.StatusOK {
+		t.Errorf("code = %d", w.Code)
+	}
+	if !strings.Contains(w.Body.String(), "[]") {
+		t.Errorf("expected empty array, got %s", w.Body.String())
+	}
+}
+
+func TestExport_RedactsSecretsInPluginPath(t *testing.T) {
+	t.Setenv(envMemoryV2Cutover, "true")
+	mock := installMockDB(t)
+	mock.ExpectQuery("WITH RECURSIVE chain").
+		WillReturnRows(sqlmock.NewRows([]string{"id", "name", "root_id"}).
+			AddRow("ws-1", "alpha", "ws-1"))
+
+	plugin := &stubAdminPlugin{
+		searchFn: func(_ context.Context, _ contract.SearchRequest) (*contract.SearchResponse, error) {
+			return &contract.SearchResponse{Memories: []contract.Memory{
+				{ID: "mem-1", Namespace: "workspace:root-1", Content: "API_KEY=sk-1234567890abcdefghijk0123456789", Kind: contract.MemoryKindFact, Source: contract.MemorySourceAgent, CreatedAt: time.Now().UTC()},
+			}}, nil
+		},
+	}
+	h := NewAdminMemoriesHandler().withMemoryV2APIs(plugin, adminRootResolver())
+	gin.SetMode(gin.TestMode)
+	w := httptest.NewRecorder()
+	c, _ := gin.CreateTestContext(w)
+	c.Request = httptest.NewRequest("GET", "/admin/memories/export", nil)
+	h.Export(c)
+
+	if strings.Contains(w.Body.String(), "sk-1234567890abcdef") {
+		t.Errorf("export leaked unredacted secret: %s", w.Body.String())
+	}
+}
+
+// --- Import via plugin ---
+
+func TestImport_RoutesThroughPluginWhenCutoverActive(t *testing.T) {
+	t.Setenv(envMemoryV2Cutover, "true")
+	mock := installMockDB(t)
+	mock.ExpectQuery("SELECT id::text FROM workspaces").
+		WithArgs("alpha").
+		WillReturnRows(sqlmock.NewRows([]string{"id"}).AddRow("root-1"))
+
+	plugin := &stubAdminPlugin{}
+	h := NewAdminMemoriesHandler().withMemoryV2APIs(plugin, adminRootResolver())
+
+	body, _ := json.Marshal([]memoryImportEntry{
+		{Content: "fact x", Scope: "LOCAL", WorkspaceName: "alpha"},
+	})
+	gin.SetMode(gin.TestMode)
+	w := httptest.NewRecorder()
+	c, _ := gin.CreateTestContext(w)
+	c.Request = httptest.NewRequest("POST", "/admin/memories/import", bytes.NewReader(body))
+	c.Request.Header.Set("Content-Type", "application/json")
+	h.Import(c)
+
+	if w.Code != http.StatusOK {
+		t.Fatalf("code = %d body=%s", w.Code, w.Body.String())
+	}
+	if len(plugin.commits) != 1 {
+		t.Errorf("commits = %d, want 1", len(plugin.commits))
+	}
+	if plugin.commits[0].NS != "workspace:root-1" {
+		t.Errorf("ns = %q", plugin.commits[0].NS)
+	}
+}
+
+func TestImport_SkipsUnknownWorkspace(t *testing.T) {
+	t.Setenv(envMemoryV2Cutover, "true")
+	mock := installMockDB(t)
+	mock.ExpectQuery("SELECT id::text FROM workspaces").
+		WithArgs("ghost").
+		WillReturnError(errors.New("no rows"))
+
+	plugin := &stubAdminPlugin{}
+	h := NewAdminMemoriesHandler().withMemoryV2APIs(plugin, adminRootResolver())
+
+	body, _ := json.Marshal([]memoryImportEntry{
+		{Content: "x", Scope: "LOCAL", WorkspaceName: "ghost"},
+	})
+	gin.SetMode(gin.TestMode)
+	w := httptest.NewRecorder()
+	c, _ := gin.CreateTestContext(w)
+	c.Request = httptest.NewRequest("POST", "/admin/memories/import", bytes.NewReader(body))
+	c.Request.Header.Set("Content-Type", "application/json")
+	h.Import(c)
+
+	var resp map[string]int
+	_ = json.Unmarshal(w.Body.Bytes(), &resp)
+	if resp["skipped"] != 1 || resp["imported"] != 0 {
+		t.Errorf("resp = %v", resp)
+	}
+}
+
+func TestImport_PluginUpsertNamespaceError(t *testing.T) {
+	t.Setenv(envMemoryV2Cutover, "true")
+	mock := installMockDB(t)
+	mock.ExpectQuery("SELECT id::text FROM workspaces").
+		WillReturnRows(sqlmock.NewRows([]string{"id"}).AddRow("root-1"))
+
+	plugin := &stubAdminPlugin{
+		upsertFn: func(_ context.Context, _ string, _ contract.NamespaceUpsert) (*contract.Namespace, error) {
+			return nil, errors.New("upsert dead")
+		},
+	}
+	h := NewAdminMemoriesHandler().withMemoryV2APIs(plugin, adminRootResolver())
+
+	body, _ := json.Marshal([]memoryImportEntry{
+		{Content: "x", Scope: "LOCAL", WorkspaceName: "alpha"},
+	})
+	gin.SetMode(gin.TestMode)
+	w := httptest.NewRecorder()
+	c, _ := gin.CreateTestContext(w)
+	c.Request = httptest.NewRequest("POST", "/admin/memories/import", bytes.NewReader(body))
+	c.Request.Header.Set("Content-Type", "application/json")
+	h.Import(c)
+
+	var resp map[string]int
+	_ = json.Unmarshal(w.Body.Bytes(), &resp)
+	if resp["errors"] != 1 || resp["imported"] != 0 {
+		t.Errorf("resp = %v", resp)
+	}
+}
+
+func TestImport_PluginCommitError(t *testing.T) {
+	t.Setenv(envMemoryV2Cutover, "true")
+	mock := installMockDB(t)
+	mock.ExpectQuery("SELECT id::text FROM workspaces").
+		WillReturnRows(sqlmock.NewRows([]string{"id"}).AddRow("root-1"))
+
+	plugin := &stubAdminPlugin{
+		commitFn: func(_ context.Context, _ string, _ contract.MemoryWrite) (*contract.MemoryWriteResponse, error) {
+			return nil, errors.New("commit dead")
+		},
+	}
+	h := NewAdminMemoriesHandler().withMemoryV2APIs(plugin, adminRootResolver())
+
+	body, _ := json.Marshal([]memoryImportEntry{
+		{Content: "x", Scope: "LOCAL", WorkspaceName: "alpha"},
+	})
+	gin.SetMode(gin.TestMode)
+	w := httptest.NewRecorder()
+	c, _ := gin.CreateTestContext(w)
+	c.Request = httptest.NewRequest("POST", "/admin/memories/import", bytes.NewReader(body))
+	c.Request.Header.Set("Content-Type", "application/json")
+	h.Import(c)
+
+	var resp map[string]int
+	_ = json.Unmarshal(w.Body.Bytes(), &resp)
+	if resp["errors"] != 1 {
+		t.Errorf("resp = %v", resp)
+	}
+}
+
+func TestImport_RedactsBeforePluginSeesContent(t *testing.T) {
+	t.Setenv(envMemoryV2Cutover, "true")
+	mock := installMockDB(t)
+	mock.ExpectQuery("SELECT id::text FROM workspaces").
+		WillReturnRows(sqlmock.NewRows([]string{"id"}).AddRow("root-1"))
+
+	plugin := &stubAdminPlugin{}
+	h := NewAdminMemoriesHandler().withMemoryV2APIs(plugin, adminRootResolver())
+
+	body, _ := json.Marshal([]memoryImportEntry{
+		{Content: "API_KEY=sk-1234567890abcdefghijk0123456789", Scope: "LOCAL", WorkspaceName: "alpha"},
+	})
+	gin.SetMode(gin.TestMode)
+	w := httptest.NewRecorder()
+	c, _ := gin.CreateTestContext(w)
+	c.Request = httptest.NewRequest("POST", "/admin/memories/import", bytes.NewReader(body))
+	c.Request.Header.Set("Content-Type", "application/json")
+	h.Import(c)
+
+	if len(plugin.commits) != 1 {
+		t.Fatalf("commits = %d", len(plugin.commits))
+	}
+	if strings.Contains(plugin.commits[0].Content, "sk-1234567890") {
+		t.Errorf("plugin received unredacted content: %q", plugin.commits[0].Content)
+	}
+}
+
+func TestImport_SkipsUnknownScope(t *testing.T) {
+	t.Setenv(envMemoryV2Cutover, "true")
+	mock := installMockDB(t)
+	mock.ExpectQuery("SELECT id::text FROM workspaces").
+		WillReturnRows(sqlmock.NewRows([]string{"id"}).AddRow("root-1"))
+
+	plugin := &stubAdminPlugin{}
+	h := NewAdminMemoriesHandler().withMemoryV2APIs(plugin, adminRootResolver())
+
+	body, _ := json.Marshal([]memoryImportEntry{
+		{Content: "x", Scope: "WEIRD", WorkspaceName: "alpha"},
+	})
+	gin.SetMode(gin.TestMode)
+	w := httptest.NewRecorder()
+	c, _ := gin.CreateTestContext(w)
+	c.Request = httptest.NewRequest("POST", "/admin/memories/import", bytes.NewReader(body))
+	c.Request.Header.Set("Content-Type", "application/json")
+	h.Import(c)
+
+	var resp map[string]int
+	_ = json.Unmarshal(w.Body.Bytes(), &resp)
+	if resp["skipped"] != 1 {
+		t.Errorf("resp = %v", resp)
+	}
+}
+
+func TestImport_SkipsWhenResolverErrors(t *testing.T) {
+	t.Setenv(envMemoryV2Cutover, "true")
+	mock := installMockDB(t)
+	mock.ExpectQuery("SELECT id::text FROM workspaces").
+		WillReturnRows(sqlmock.NewRows([]string{"id"}).AddRow("root-1"))
+
+	plugin := &stubAdminPlugin{}
+	resolver := &stubAdminResolver{err: errors.New("dead")}
+	h := NewAdminMemoriesHandler().withMemoryV2APIs(plugin, resolver)
+
+	body, _ := json.Marshal([]memoryImportEntry{
+		{Content: "x", Scope: "LOCAL", WorkspaceName: "alpha"},
+	})
+	gin.SetMode(gin.TestMode)
+	w := httptest.NewRecorder()
+	c, _ := gin.CreateTestContext(w)
+	c.Request = httptest.NewRequest("POST", "/admin/memories/import", bytes.NewReader(body))
+	c.Request.Header.Set("Content-Type", "application/json")
+	h.Import(c)
+
+	var resp map[string]int
+	_ = json.Unmarshal(w.Body.Bytes(), &resp)
+	if resp["skipped"] != 1 {
+		t.Errorf("resp = %v", resp)
+	}
+}
+
+// TestExport_BatchesPluginCallsByRoot pins the I3 fix: previously the
+// export ran one resolver + one plugin search per workspace (N+1 in
+// both); now it groups by root and runs one resolver + one plugin
+// search per UNIQUE root.
+//
+// Setup: 3 workspaces under 1 root → 1 resolver call + 1 plugin call
+// (was: 3 resolver + 3 plugin in the old code). The plugin search
+// receives 5 namespaces: each member's workspace:<id> + team:root-1
+// + org:root-1. (Children's workspace:<id> namespaces must be
+// included or admin export silently drops their private memories.)
+func TestExport_BatchesPluginCallsByRoot(t *testing.T) {
+	t.Setenv(envMemoryV2Cutover, "true")
+	mock := installMockDB(t)
+
+	mock.ExpectQuery("WITH RECURSIVE chain").
+		WillReturnRows(sqlmock.NewRows([]string{"id", "name", "root_id"}).
+			AddRow("root-1", "alpha", "root-1").
+			AddRow("child-1", "alpha-child", "root-1").
+			AddRow("child-2", "alpha-grandchild", "root-1"))
+
+	pluginSearchCount := 0
+	plugin := &stubAdminPlugin{
+		searchFn: func(_ context.Context, body contract.SearchRequest) (*contract.SearchResponse, error) {
+			pluginSearchCount++
+			if len(body.Namespaces) != 5 {
+				t.Errorf("plugin search call %d: namespaces len = %d, want 5 (3 workspace + team + org); got %v", pluginSearchCount, len(body.Namespaces), body.Namespaces)
+			}
+			return &contract.SearchResponse{}, nil
+		},
+	}
+	h := NewAdminMemoriesHandler().withMemoryV2APIs(plugin, adminRootResolver())
+
+	gin.SetMode(gin.TestMode)
+	w := httptest.NewRecorder()
+	c, _ := gin.CreateTestContext(w)
+	c.Request = httptest.NewRequest("GET", "/admin/memories/export", nil)
+	h.Export(c)
+
+	if w.Code != http.StatusOK {
+		t.Errorf("code = %d body=%s", w.Code, w.Body.String())
+	}
+	if pluginSearchCount != 1 {
+		t.Errorf("plugin search called %d times, want 1 (was 3 with the old N+1 code)", pluginSearchCount)
+	}
+}
+
+// perWorkspaceResolver mimics the real resolver: ReadableNamespaces
+// returns the SPECIFIC workspace's view (workspace:<that ID> +
+// team:<root> + org:<root>), not a constant set. The legacy
+// stubAdminResolver hides the I3 silent-drop bug by ignoring its
+// workspace-id argument.
+type perWorkspaceResolver map[string][]namespace.Namespace
+
+func (r perWorkspaceResolver) ReadableNamespaces(_ context.Context, ws string) ([]namespace.Namespace, error) {
+	v, ok := r[ws]
+	if !ok {
+		return nil, errors.New("perWorkspaceResolver: unknown ws " + ws)
+	}
+	return v, nil
+}
+func (r perWorkspaceResolver) WritableNamespaces(_ context.Context, ws string) ([]namespace.Namespace, error) {
+	return r.ReadableNamespaces(nil, ws)
+}
+
+// TestExport_IncludesEveryMembersPrivateNamespace pins the I3 follow-up
+// fix: when a root group has multiple members, the export must surface
+// each member's workspace:<id> namespace, not just the root's. Before
+// the fix, calling ReadableNamespaces(rootID) returned only
+// workspace:rootID + team:rootID + org:rootID — every child workspace's
+// private memories were silently dropped from admin export.
+func TestExport_IncludesEveryMembersPrivateNamespace(t *testing.T) {
+	t.Setenv(envMemoryV2Cutover, "true")
+	mock := installMockDB(t)
+
+	mock.ExpectQuery("WITH RECURSIVE chain").
+		WillReturnRows(sqlmock.NewRows([]string{"id", "name", "root_id"}).
+			AddRow("root-1", "alpha", "root-1").
+			AddRow("child-1", "alpha-child", "root-1").
+			AddRow("child-2", "alpha-grandchild", "root-1"))
+
+	resolver := perWorkspaceResolver{
+		"root-1": {
+			{Name: "workspace:root-1", Kind: contract.NamespaceKindWorkspace, Writable: true},
+			{Name: "team:root-1", Kind: contract.NamespaceKindTeam, Writable: true},
+			{Name: "org:root-1", Kind: contract.NamespaceKindOrg, Writable: true},
+		},
+		"child-1": {
+			{Name: "workspace:child-1", Kind: contract.NamespaceKindWorkspace, Writable: true},
+			{Name: "team:root-1", Kind: contract.NamespaceKindTeam, Writable: true},
+			{Name: "org:root-1", Kind: contract.NamespaceKindOrg, Writable: true},
+		},
+		"child-2": {
+			{Name: "workspace:child-2", Kind: contract.NamespaceKindWorkspace, Writable: true},
+			{Name: "team:root-1", Kind: contract.NamespaceKindTeam, Writable: true},
+			{Name: "org:root-1", Kind: contract.NamespaceKindOrg, Writable: true},
+		},
+	}
+
+	var passedNamespaces []string
+	plugin := &stubAdminPlugin{
+		searchFn: func(_ context.Context, body contract.SearchRequest) (*contract.SearchResponse, error) {
+			passedNamespaces = append(passedNamespaces, body.Namespaces...)
+			return &contract.SearchResponse{Memories: []contract.Memory{
+				{ID: "m-root", Namespace: "workspace:root-1", Content: "root private", Kind: contract.MemoryKindFact, Source: contract.MemorySourceAgent, CreatedAt: time.Now().UTC()},
+				{ID: "m-child1", Namespace: "workspace:child-1", Content: "child-1 private", Kind: contract.MemoryKindFact, Source: contract.MemorySourceAgent, CreatedAt: time.Now().UTC()},
+				{ID: "m-child2", Namespace: "workspace:child-2", Content: "child-2 private", Kind: contract.MemoryKindFact, Source: contract.MemorySourceAgent, CreatedAt: time.Now().UTC()},
+				{ID: "m-team", Namespace: "team:root-1", Content: "shared team", Kind: contract.MemoryKindFact, Source: contract.MemorySourceAgent, CreatedAt: time.Now().UTC()},
+			}}, nil
+		},
+	}
+	h := NewAdminMemoriesHandler().withMemoryV2APIs(plugin, resolver)
+
+	gin.SetMode(gin.TestMode)
+	w := httptest.NewRecorder()
+	c, _ := gin.CreateTestContext(w)
+	c.Request = httptest.NewRequest("GET", "/admin/memories/export", nil)
+	h.Export(c)
+
+	if w.Code != http.StatusOK {
+		t.Fatalf("code = %d body=%s", w.Code, w.Body.String())
+	}
+
+	// Every member's private namespace must reach the plugin search.
+	want := []string{"workspace:root-1", "workspace:child-1", "workspace:child-2", "team:root-1", "org:root-1"}
+	got := make(map[string]bool, len(passedNamespaces))
+	for _, ns := range passedNamespaces {
+		got[ns] = true
+	}
+	for _, w := range want {
+		if !got[w] {
+			t.Errorf("plugin search missing namespace %q (got %v)", w, passedNamespaces)
+		}
+	}
+	if len(passedNamespaces) != 5 {
+		t.Errorf("plugin search namespace count = %d, want 5 (3 workspace + team + org)", len(passedNamespaces))
+	}
+
+	// Children's private memories must appear in the export, attributed
+	// to the right workspace_name.
+	var entries []memoryExportEntry
+	if err := json.Unmarshal(w.Body.Bytes(), &entries); err != nil {
+		t.Fatalf("unmarshal: %v", err)
+	}
+	byID := map[string]memoryExportEntry{}
+	for _, e := range entries {
+		byID[e.ID] = e
+	}
+	for _, exp := range []struct{ id, ns, owner string }{
+		{"m-root", "workspace:root-1", "alpha"},
+		{"m-child1", "workspace:child-1", "alpha-child"},
+		{"m-child2", "workspace:child-2", "alpha-grandchild"},
+	} {
+		e, ok := byID[exp.id]
+		if !ok {
+			t.Errorf("export missing memory %s — children's private memories silently dropped", exp.id)
+			continue
+		}
+		if e.Namespace != exp.ns {
+			t.Errorf("memory %s namespace = %q, want %q", exp.id, e.Namespace, exp.ns)
+		}
+		if e.WorkspaceName != exp.owner {
+			t.Errorf("memory %s owner = %q, want %q", exp.id, e.WorkspaceName, exp.owner)
+		}
+	}
+}
+
+// TestPickOwnerForNamespace covers the namespace→workspace_name
+// attribution helper introduced in I3.
+func TestPickOwnerForNamespace(t *testing.T) {
+	members := []workspaceRow{
+		{ID: "root-1", Name: "alpha", RootID: "root-1"},
+		{ID: "child-1", Name: "alpha-child", RootID: "root-1"},
+	}
+	cases := []struct {
+		name string
+		ns   string
+		want string
+	}{
+		{"workspace ns matches member id", "workspace:child-1", "alpha-child"},
+		{"workspace ns no match → first", "workspace:foreign", "alpha"},
+		{"team ns → first member of root group", "team:root-1", "alpha"},
+		{"org ns → first member", "org:root-1", "alpha"},
+		{"custom ns → first member", "custom:foo", "alpha"},
+	}
+	for _, tc := range cases {
+		t.Run(tc.name, func(t *testing.T) {
+			if got := pickOwnerForNamespace(tc.ns, members); got != tc.want {
+				t.Errorf("pickOwnerForNamespace(%q) = %q, want %q", tc.ns, got, tc.want)
+			}
+		})
+	}
+	if got := pickOwnerForNamespace("workspace:abc", nil); got != "" {
+		t.Errorf("empty members must return \"\", got %q", got)
+	}
+}
+
+// --- Helper functions ---
+
+func TestLegacyScopeFromNamespace(t *testing.T) {
+	cases := []struct {
+		in   string
+		want string
+	}{
+		{"workspace:abc", "LOCAL"},
+		{"team:abc", "TEAM"},
+		{"org:abc", "GLOBAL"},
+		{"custom:abc", ""},
+		{"", ""},
+	}
+	for _, tc := range cases {
+		if got := legacyScopeFromNamespace(tc.in); got != tc.want {
+			t.Errorf("legacyScopeFromNamespace(%q) = %q, want %q", tc.in, got, tc.want)
+		}
+	}
+}
+
+func TestNamespaceKindFromLegacyScope(t *testing.T) {
+	cases := []struct {
+		in   string
+		want contract.NamespaceKind
+	}{
+		{"LOCAL", contract.NamespaceKindWorkspace},
+		{"local", contract.NamespaceKindWorkspace},
+		{"TEAM", contract.NamespaceKindTeam},
+		{"GLOBAL", contract.NamespaceKindOrg},
+		{"weird", contract.NamespaceKindWorkspace},
+	}
+	for _, tc := range cases {
+		if got := namespaceKindFromLegacyScope(tc.in); got != tc.want {
+			t.Errorf("namespaceKindFromLegacyScope(%q) = %q, want %q", tc.in, got, tc.want)
+		}
+	}
+}
+
+func TestSkipImport_ErrorMessage(t *testing.T) {
+	e := &skipImport{reason: "unknown scope: WEIRD"}
+	if !strings.Contains(e.Error(), "unknown scope: WEIRD") {
+		t.Errorf("Error() = %q", e.Error())
+	}
+}
+
+// --- Confirm legacy paths still work when env is unset ---
+
+func TestExport_LegacyPathWhenCutoverInactive(t *testing.T) {
+	t.Setenv(envMemoryV2Cutover, "")
+	mock := installMockDB(t)
+	mock.ExpectQuery("SELECT am.id, am.content, am.scope, am.namespace").
+		WillReturnRows(sqlmock.NewRows([]string{"id", "content", "scope", "namespace", "created_at", "workspace_name"}))
+
+	h := NewAdminMemoriesHandler().withMemoryV2APIs(&stubAdminPlugin{}, adminRootResolver())
+	gin.SetMode(gin.TestMode)
+	w := httptest.NewRecorder()
+	c, _ := gin.CreateTestContext(w)
+	c.Request = httptest.NewRequest("GET", "/admin/memories/export", nil)
+	h.Export(c)
+
+	if w.Code != http.StatusOK {
+		t.Errorf("code = %d body=%s", w.Code, w.Body.String())
+	}
+	if err := mock.ExpectationsWereMet(); err != nil {
+		t.Errorf("legacy SQL path not exercised: %v", err)
+	}
+}

workspace-server/internal/handlers/chat_files.go

@@ -30,6 +30,7 @@ package handlers
 
 import (
 	"context"
+	"database/sql"
 	"fmt"
 	"io"
 	"log"
@@ -102,14 +103,45 @@ const chatUploadDir = "/workspace/.molecule/chat-uploads"
 // of bug as the original SaaS provision drift fixed in #2366; this
 // extraction prevents that class on the consumer side.
 func resolveWorkspaceForwardCreds(c *gin.Context, ctx context.Context, workspaceID, op string) (wsURL, secret string, ok bool) {
+	var deliveryMode sql.NullString
 	if err := db.DB.QueryRowContext(ctx,
-		`SELECT COALESCE(url, '') FROM workspaces WHERE id = $1`, workspaceID,
-	).Scan(&wsURL); err != nil {
+		`SELECT COALESCE(url, ''), delivery_mode FROM workspaces WHERE id = $1`, workspaceID,
+	).Scan(&wsURL, &deliveryMode); err != nil {
 		log.Printf("chat_files %s: workspace lookup failed for %s: %v", op, workspaceID, err)
 		c.JSON(http.StatusNotFound, gin.H{"error": "workspace not found"})
 		return "", "", false
 	}
 	if wsURL == "" {
+		// Distinguish the two empty-URL classes so the user sees an
+		// actionable error rather than a misleading "not registered yet"
+		// (which implies waiting will help):
+		//
+		//  push-mode → URL just isn't on the row yet (workspace
+		//    restart in progress, or first /registry/register hasn't
+		//    landed). 503 + "not registered yet" is correct — retry
+		//    after the next heartbeat (~30s) will likely succeed.
+		//
+		//  anything else (poll-mode, NULL, empty string) → URL is
+		//    structurally absent. The platform never dispatches to a
+		//    non-push workspace, so chat upload (which is HTTP-forward
+		//    by design) cannot proceed by waiting. Returning 503 here
+		//    would loop the canvas client forever. 422 signals "this
+		//    request can't succeed against THIS workspace's
+		//    configuration" — the only fix is to re-register the
+		//    workspace with a publicly-reachable URL.
+		//
+		// Live-observed 2026-05-04: external runtime workspaces (e.g.
+		// molecule-sdk-python on a mac laptop) register with
+		// delivery_mode=NULL. The narrow "poll" check missed them; the
+		// invariant we actually want is "URL empty + not-push = no
+		// dispatch path, ever".
+		if !deliveryMode.Valid || deliveryMode.String != "push" {
+			c.JSON(http.StatusUnprocessableEntity, gin.H{
+				"error":  "workspace has no callback URL — chat " + op + " requires push-mode + public URL",
+				"detail": "This workspace registered without a publicly-reachable URL (delivery_mode is not 'push'). The platform cannot dispatch chat uploads to it. Re-register the workspace with a public URL in push mode (e.g. via ngrok / Cloudflare tunnel) to enable chat file " + op + ".",
+			})
+			return "", "", false
+		}
 		c.JSON(http.StatusServiceUnavailable, gin.H{"error": "workspace url not registered yet"})
 		return "", "", false
 	}

workspace-server/internal/handlers/chat_files_test.go

@@ -58,16 +58,38 @@ func uploadFixture(t *testing.T) (*bytes.Buffer, string) {
 	return &buf, mw.FormDataContentType()
 }
 
-// expectURL stubs the SELECT that resolves the workspace's url.
+// expectURL stubs the SELECT that resolves the workspace's url +
+// delivery_mode. Defaults delivery_mode to "push" — most tests don't
+// care about the mode and just want a URL to forward to. Use
+// expectURLAndMode when the test needs a specific mode (e.g. the
+// poll-mode 422 path).
 func expectURL(mock sqlmock.Sqlmock, workspaceID, url string) {
-	mock.ExpectQuery(`SELECT COALESCE\(url, ''\) FROM workspaces WHERE id = \$1`).
+	expectURLAndMode(mock, workspaceID, url, "push")
+}
+
+// expectURLAndMode is the explicit form for tests that need to
+// exercise the delivery_mode branch (e.g. poll-mode workspaces get
+// a 422 instead of a 503 when URL is empty — the platform can't
+// dispatch to a non-push workspace at all).
+func expectURLAndMode(mock sqlmock.Sqlmock, workspaceID, url, mode string) {
+	mock.ExpectQuery(`SELECT COALESCE\(url, ''\), delivery_mode FROM workspaces WHERE id = \$1`).
 		WithArgs(workspaceID).
-		WillReturnRows(sqlmock.NewRows([]string{"url"}).AddRow(url))
+		WillReturnRows(sqlmock.NewRows([]string{"url", "delivery_mode"}).AddRow(url, mode))
+}
+
+// expectURLNullMode is the production-observed shape: external runtime
+// workspaces (molecule-sdk-python on user infra) register with
+// delivery_mode = NULL, not "poll". Caught 2026-05-04 — the narrow
+// "poll" check missed three of three real workspaces in user reports.
+func expectURLNullMode(mock sqlmock.Sqlmock, workspaceID, url string) {
+	mock.ExpectQuery(`SELECT COALESCE\(url, ''\), delivery_mode FROM workspaces WHERE id = \$1`).
+		WithArgs(workspaceID).
+		WillReturnRows(sqlmock.NewRows([]string{"url", "delivery_mode"}).AddRow(url, nil))
 }
 
 // expectURLMissing stubs the SELECT to return sql.ErrNoRows.
 func expectURLMissing(mock sqlmock.Sqlmock, workspaceID string) {
-	mock.ExpectQuery(`SELECT COALESCE\(url, ''\) FROM workspaces WHERE id = \$1`).
+	mock.ExpectQuery(`SELECT COALESCE\(url, ''\), delivery_mode FROM workspaces WHERE id = \$1`).
 		WithArgs(workspaceID).
 		WillReturnError(sql.ErrNoRows)
 }
@@ -201,9 +223,13 @@ func TestChatUpload_NoURL(t *testing.T) {
 	mock := setupTestDB(t)
 	setupTestRedis(t)
 
-	// Workspace registered but URL hasn't been reported yet (mid-boot).
+	// Workspace registered (push-mode) but URL hasn't been reported
+	// yet (mid-boot). 503 + "not registered yet" is the right surface — the
+	// canvas client can retry after the next heartbeat picks up the URL.
+	// Push mode is the only branch that produces 503; everything else
+	// (poll, NULL, empty) gets 422 because no amount of waiting helps.
 	wsID := "00000000-0000-0000-0000-000000000042"
-	expectURL(mock, wsID, "")
+	expectURLAndMode(mock, wsID, "", "push")
 
 	h := NewChatFilesHandler(NewTemplatesHandler(t.TempDir(), nil))
 	body, ct := uploadFixture(t)
@@ -211,7 +237,65 @@ func TestChatUpload_NoURL(t *testing.T) {
 	h.Upload(c)
 
 	if w.Code != http.StatusServiceUnavailable {
-		t.Errorf("expected 503 when workspace url empty, got %d: %s", w.Code, w.Body.String())
+		t.Errorf("expected 503 when workspace url empty (push mode), got %d: %s", w.Code, w.Body.String())
+	}
+	if !strings.Contains(w.Body.String(), "not registered yet") {
+		t.Errorf("expected transient-state error message, got: %s", w.Body.String())
+	}
+}
+
+// TestChatUpload_PollModeEmptyURL pins the 422 distinguisher: a
+// poll-mode workspace has no URL by design, so chat upload (which is
+// HTTP-forward to the workspace) cannot succeed by retrying. Returning
+// 503 here would loop the canvas client forever; 422 + an actionable
+// message tells the user what to do.
+func TestChatUpload_PollModeEmptyURL(t *testing.T) {
+	mock := setupTestDB(t)
+	setupTestRedis(t)
+
+	wsID := "00000000-0000-0000-0000-000000000099"
+	expectURLAndMode(mock, wsID, "", "poll")
+
+	h := NewChatFilesHandler(NewTemplatesHandler(t.TempDir(), nil))
+	body, ct := uploadFixture(t)
+	c, w := makeUploadRequest(t, wsID, body, ct)
+	h.Upload(c)
+
+	if w.Code != http.StatusUnprocessableEntity {
+		t.Fatalf("expected 422 for poll-mode upload, got %d: %s", w.Code, w.Body.String())
+	}
+	if !strings.Contains(w.Body.String(), "push") {
+		t.Errorf("expected error to suggest push mode, got: %s", w.Body.String())
+	}
+}
+
+// TestChatUpload_NullModeEmptyURL — production-observed 2026-05-04:
+// external-runtime workspaces (molecule-sdk-python on user infra)
+// register with delivery_mode = NULL, not "poll". The earlier narrow
+// poll-only check fell through to the misleading 503. The fix is the
+// inverse-of-push test: anything not exactly "push" with empty URL
+// can't dispatch and gets the actionable 422.
+//
+// Three of three external workspaces in the user's tenant had this
+// shape (home hermes / runner mac mini / mac laptop, all
+// runtime=external + url='' + delivery_mode=NULL).
+func TestChatUpload_NullModeEmptyURL(t *testing.T) {
+	mock := setupTestDB(t)
+	setupTestRedis(t)
+
+	wsID := "30ba7f0b-b303-4a20-aefe-3a4a675b8aa4" // user's "mac laptop"
+	expectURLNullMode(mock, wsID, "")
+
+	h := NewChatFilesHandler(NewTemplatesHandler(t.TempDir(), nil))
+	body, ct := uploadFixture(t)
+	c, w := makeUploadRequest(t, wsID, body, ct)
+	h.Upload(c)
+
+	if w.Code != http.StatusUnprocessableEntity {
+		t.Fatalf("expected 422 for null-delivery-mode upload, got %d: %s", w.Code, w.Body.String())
+	}
+	if !strings.Contains(w.Body.String(), "callback URL") {
+		t.Errorf("expected error to mention callback URL, got: %s", w.Body.String())
 	}
 }
 

workspace-server/internal/handlers/external_connection.go

@@ -83,7 +83,20 @@ curl -fsS -X POST "{{PLATFORM_URL}}/registry/register" \
 const externalChannelTemplate = `# Claude Code channel — bridges this workspace's A2A traffic into your
 # Claude Code session. No tunnel/public URL needed (polling-based).
 #
-# 1. Save this token + workspace_id, then create ~/.claude/channels/molecule/.env:
+# Prereq: Bun installed (channel plugins are Bun scripts).
+#   bun --version    # must print a version number
+#
+# 1. Inside Claude Code, install the channel plugin from its GitHub repo.
+#    The plugin is NOT on Anthropic's default allowlist, so a one-time
+#    marketplace-add is needed before install:
+#
+#      /plugin marketplace add Molecule-AI/molecule-mcp-claude-channel
+#      /plugin install molecule@molecule-mcp-claude-channel
+#
+#    Then either run /reload-plugins or restart Claude Code so the
+#    plugin is registered.
+#
+# 2. Create the per-watched-workspace config file:
 mkdir -p ~/.claude/channels/molecule
 cat > ~/.claude/channels/molecule/.env <<'EOF'
 MOLECULE_PLATFORM_URL={{PLATFORM_URL}}
@@ -92,13 +105,32 @@ MOLECULE_WORKSPACE_TOKENS=<paste auth_token from create response>
 EOF
 chmod 600 ~/.claude/channels/molecule/.env
 
-# 2. Launch Claude Code with the channel enabled:
-claude --channels plugin:molecule@Molecule-AI/molecule-mcp-claude-channel
+# 3. Launch Claude Code with the channel enabled. Custom (non-Anthropic-
+#    allowlisted) channels need the --dangerously-load-development-channels
+#    flag to opt in — without it, you'll see "not on the approved channels
+#    allowlist" on startup.
+claude --dangerously-load-development-channels \
+  --channels plugin:molecule@molecule-mcp-claude-channel
 
+# You should see on stderr:
+#   molecule channel: connected — watching 1 workspace(s) at {{PLATFORM_URL}}
+#
 # Inbound A2A messages now surface as conversation turns. Claude's
 # replies route back via the reply_to_workspace MCP tool — no extra
 # wiring on your side.
 #
+# Common errors:
+#   "plugin not installed"            → Step 1 didn't run; run /plugin install
+#                                       inside Claude Code, then /reload-plugins.
+#   "not on approved channels allowlist" → Add --dangerously-load-development-channels
+#                                       to the launch command (Step 3).
+#   "config-missing"                  → ~/.claude/channels/molecule/.env not
+#                                       readable; re-run Step 2 and check chmod.
+#
+# Team/Enterprise orgs: the --dangerously-load-development-channels flag is
+# blocked by managed settings. Your admin must set channelsEnabled=true and
+# add the plugin to allowedChannelPlugins in claude.ai admin settings.
+#
 # Multi-workspace: comma-separate IDs and tokens (same order). See
 # https://github.com/Molecule-AI/molecule-mcp-claude-channel for
 # pairing flow, push-mode upgrade, and v0.2 roadmap.
@@ -186,3 +218,191 @@ async def main():
 if __name__ == "__main__":
     asyncio.run(main())
 `
+
+// externalHermesChannelTemplate — install snippet for operators whose
+// external agent IS a hermes-agent session. Routes the workspace's
+// A2A traffic into the running hermes gateway as platform messages
+// via the molecule-channel plugin.
+//
+// The plugin (Molecule-AI/hermes-channel-molecule) is a hermes
+// platform adapter that:
+//   1. Spawns ``python -m molecule_runtime.a2a_mcp_server`` as a
+//      stdio MCP subprocess (separate from any hermes-side MCP
+//      client connection).
+//   2. Long-polls ``wait_for_message`` on the platform's inbox.
+//   3. Dispatches each inbound activity into the hermes gateway as a
+//      MessageEvent — same code path Telegram/Discord use.
+//   4. Outbound replies route via ``send_message_to_user`` (canvas
+//      user) or ``delegate_task`` (peer agent) MCP tool calls.
+//
+// Result: hermes gets push parity with Claude Code / codex / openclaw —
+// canvas messages and peer A2A arrive as conversation turns mid-session,
+// not just at the start of a new ``hermes`` invocation.
+//
+// Plugin uses the upstream ``register_platform`` API shipped by
+// NousResearch/hermes-agent#17751 (merged 2026-04-30) and falls back
+// to the legacy ``register_platform_adapter`` shape on older forks —
+// same wheel installs cleanly on stock or patched hermes-agent.
+const externalHermesChannelTemplate = `# Hermes channel — bridges this workspace's A2A traffic into your
+# hermes-agent session. No tunnel/public URL needed (long-poll based,
+# same shape as the Claude Code channel).
+#
+# Prereq: a hermes-agent install on the target machine. Latest builds
+# (post #17751) ship the platform-plugin API natively; older ones are
+# also supported via the plugin's dual-mode fallback.
+#
+# 1. Install the runtime + plugin:
+pip install molecule-ai-workspace-runtime
+pip install 'git+https://github.com/Molecule-AI/hermes-channel-molecule.git'
+
+# 2. Export the workspace credentials:
+export MOLECULE_WORKSPACE_ID={{WORKSPACE_ID}}
+export MOLECULE_PLATFORM_URL={{PLATFORM_URL}}
+export MOLECULE_WORKSPACE_TOKEN="<paste from create response>"
+export MOLECULE_ORG_ID="<your org id>"
+
+# 3. Edit ~/.hermes/config.yaml — under your existing top-level
+#    gateway: block, add a plugin_platforms entry:
+#
+#      gateway:
+#        # ...your existing gateway settings...
+#        plugin_platforms:
+#          molecule:
+#            enabled: true
+#
+#    If you don't yet have a gateway: block, create one with just
+#    that plugin_platforms entry. Don't append blindly — YAML
+#    rejects duplicate top-level keys, so a second gateway: block
+#    will silently break hermes config loading.
+
+# 4. Restart the hermes gateway:
+hermes gateway --replace
+
+# Inbound canvas messages + peer A2A now arrive as MessageEvents —
+# same dispatch path Telegram/Discord/Slack use. The agent replies via
+# send_message_to_user / delegate_task MCP tool calls (already wired
+# by the plugin's molecule_runtime MCP subprocess).
+#
+# Source + issue tracker:
+# https://github.com/Molecule-AI/hermes-channel-molecule
+`
+
+// externalCodexTemplate — for operators whose external agent is a
+// codex CLI (@openai/codex) session. Wires the molecule_runtime A2A
+// MCP server into codex's config.toml so the agent can call
+// list_peers / delegate_task / send_message_to_user / commit_memory.
+//
+// Push parity caveat: codex's MCP client doesn't forward arbitrary
+// notifications/* from configured MCP servers (verified by reading
+// codex-rs/codex-mcp/src/connection_manager.rs in openai/codex). So
+// this snippet gives outbound tools but NOT mid-turn push from
+// inbound A2A. For full push parity on a codex external, the
+// equivalent of hermes-channel-molecule would be needed — a bridge
+// daemon that long-polls the platform inbox and calls codex's
+// turn/steer RPC. Tracked separately; this snippet is the
+// outbound-tool-only first cut.
+const externalCodexTemplate = `# Codex MCP config — outbound tool path. For operators whose external
+# agent is a codex CLI (@openai/codex) session.
+#
+# This wires the molecule platform's A2A MCP server into codex so
+# the agent can call list_peers / delegate_task / send_message_to_user
+# / commit_memory. Inbound A2A (canvas messages, peer-initiated tasks)
+# does NOT push into the running codex turn yet — codex's MCP runtime
+# doesn't route arbitrary notifications/* from configured MCP servers.
+# For inbound delivery into a codex session, pair with the Python SDK
+# tab for now.
+
+# 1. Install codex CLI + the workspace runtime wheel:
+npm install -g @openai/codex@^0.57
+pip install molecule-ai-workspace-runtime
+
+# 2. Edit ~/.codex/config.toml and add the block below. {{PLATFORM_URL}}
+#    and {{WORKSPACE_ID}} are stamped server-side; paste your auth
+#    token for MOLECULE_WORKSPACE_TOKEN before saving.
+#
+#    Don't append blindly — TOML rejects duplicate
+#    [mcp_servers.molecule] tables, so re-running on an existing
+#    config will break codex parsing. If [mcp_servers.molecule]
+#    already exists (e.g. you set this up before), replace the
+#    existing block instead of appending.
+
+mkdir -p ~/.codex
+# (then open ~/.codex/config.toml in your editor and paste:)
+#
+# [mcp_servers.molecule]
+# command = "python3"
+# args = ["-m", "molecule_runtime.a2a_mcp_server"]
+# startup_timeout_sec = 30
+#
+# [mcp_servers.molecule.env]
+# WORKSPACE_ID = "{{WORKSPACE_ID}}"
+# PLATFORM_URL = "{{PLATFORM_URL}}"
+# MOLECULE_WORKSPACE_TOKEN = "<paste from create response>"
+# MOLECULE_ORG_ID = "<your org id>"
+
+# 3. Run codex — the molecule tools are now available to the agent:
+codex
+`
+
+// externalOpenClawTemplate — for operators whose external agent is an
+// openclaw session. Wires the molecule MCP server via openclaw's
+// `mcp set` config + starts the openclaw gateway on loopback.
+//
+// Like the codex tab, this is outbound-only. Full push parity on an
+// external openclaw would need a sessions.steer bridge daemon (the
+// equivalent of hermes-channel-molecule for openclaw). Tracked
+// separately; outbound tools is the first cut.
+const externalOpenClawTemplate = `# OpenClaw MCP config — outbound tool path. For operators whose
+# external agent is an openclaw session.
+#
+# This wires the molecule platform's A2A MCP server into openclaw's
+# gateway so the agent can call list_peers / delegate_task /
+# send_message_to_user / commit_memory. Inbound A2A push into a
+# running openclaw run is not wired here yet — the platform-side
+# openclaw template (template-openclaw) implements the full
+# sessions.steer push path; an external setup would need the same
+# bridge daemon the template uses. For inbound delivery on an
+# external machine today, pair with the Python SDK tab.
+
+# 1. Install openclaw CLI + the workspace runtime wheel:
+npm install -g openclaw@latest
+pip install molecule-ai-workspace-runtime
+
+# 2. Onboard openclaw against your model provider (one-time setup).
+#    --non-interactive needs an explicit --provider + --model so it
+#    doesn't prompt; pick what matches your API key. Skip step 2 if
+#    you've already onboarded on this host.
+#
+#    openclaw onboard --non-interactive \
+#      --provider openai \
+#      --model gpt-5
+
+# 3. Wire the molecule MCP server. {{WORKSPACE_ID}} + {{PLATFORM_URL}}
+# are stamped server-side; paste the auth token before running.
+WORKSPACE_TOKEN="<paste from create response>"
+MOLECULE_ORG_ID="<your org id>"
+openclaw mcp set molecule "$(cat <<EOF
+{
+  "command": "python3",
+  "args": ["-m", "molecule_runtime.a2a_mcp_server"],
+  "env": {
+    "WORKSPACE_ID": "{{WORKSPACE_ID}}",
+    "PLATFORM_URL": "{{PLATFORM_URL}}",
+    "MOLECULE_WORKSPACE_TOKEN": "$WORKSPACE_TOKEN",
+    "MOLECULE_ORG_ID": "$MOLECULE_ORG_ID"
+  }
+}
+EOF
+)"
+
+# 4. Start the openclaw gateway as a durable background process.
+#    A bare '&' dies when the terminal closes; nohup + log file keeps
+#    the gateway alive across logout. For systemd-managed hosts,
+#    register a unit instead.
+nohup openclaw gateway --dev --port 18789 --bind loopback \
+  > ~/.openclaw/gateway.log 2>&1 &
+disown
+
+# 5. Run an agent turn — molecule tools are now available:
+openclaw agent --message "list my peers"
+`

workspace-server/internal/handlers/handlers_test.go

@@ -8,8 +8,6 @@ import (
 	"fmt"
 	"net/http"
 	"net/http/httptest"
-	"os"
-	"path/filepath"
 	"testing"
 	"time"
 
@@ -569,67 +567,6 @@ func TestProxyA2A_WorkspaceOffline(t *testing.T) {
 	}
 }
 
-// ---------- TestSharedContext ----------
-
-func TestSharedContext(t *testing.T) {
-	mock := setupTestDB(t)
-
-	// Create a temp configs directory with a workspace config
-	tmpDir := t.TempDir()
-	wsDir := filepath.Join(tmpDir, "test-workspace")
-	if err := os.MkdirAll(wsDir, 0755); err != nil {
-		t.Fatalf("failed to create config dir: %v", err)
-	}
-
-	// Write config.yaml with shared_context
-	configYAML := "name: Test Workspace\nshared_context:\n  - test.md\n"
-	if err := os.WriteFile(filepath.Join(wsDir, "config.yaml"), []byte(configYAML), 0644); err != nil {
-		t.Fatalf("failed to write config.yaml: %v", err)
-	}
-
-	// Write the shared context file
-	testContent := "# Shared Context\nThis is shared context content."
-	if err := os.WriteFile(filepath.Join(wsDir, "test.md"), []byte(testContent), 0644); err != nil {
-		t.Fatalf("failed to write test.md: %v", err)
-	}
-
-	handler := NewTemplatesHandler(tmpDir, nil)
-
-	// Mock DB returning workspace name that normalizes to "test-workspace"
-	mock.ExpectQuery("SELECT name FROM workspaces WHERE id =").
-		WithArgs("ws-ctx").
-		WillReturnRows(sqlmock.NewRows([]string{"name"}).AddRow("Test Workspace"))
-
-	w := httptest.NewRecorder()
-	c, _ := gin.CreateTestContext(w)
-	c.Params = gin.Params{{Key: "id", Value: "ws-ctx"}}
-	c.Request = httptest.NewRequest("GET", "/workspaces/ws-ctx/shared-context", nil)
-
-	handler.SharedContext(c)
-
-	if w.Code != http.StatusOK {
-		t.Errorf("expected status 200, got %d: %s", w.Code, w.Body.String())
-	}
-
-	var resp []map[string]interface{}
-	if err := json.Unmarshal(w.Body.Bytes(), &resp); err != nil {
-		t.Fatalf("failed to parse response: %v", err)
-	}
-	if len(resp) != 1 {
-		t.Fatalf("expected 1 file, got %d", len(resp))
-	}
-	if resp[0]["path"] != "test.md" {
-		t.Errorf("expected path 'test.md', got %v", resp[0]["path"])
-	}
-	if resp[0]["content"] != testContent {
-		t.Errorf("expected content %q, got %v", testContent, resp[0]["content"])
-	}
-
-	if err := mock.ExpectationsWereMet(); err != nil {
-		t.Errorf("unmet sqlmock expectations: %v", err)
-	}
-}
-
 // ---------- TestHeartbeatHandler_TaskChanged ----------
 
 func TestHeartbeatHandler_TaskChanged(t *testing.T) {
@@ -1218,53 +1155,6 @@ func TestWorkspaceGet_CurrentTask(t *testing.T) {
 	}
 }
 
-func TestSharedContext_NoSharedFiles(t *testing.T) {
-	mock := setupTestDB(t)
-
-	// Create a temp configs directory with a workspace config that has no shared_context
-	tmpDir := t.TempDir()
-	wsDir := filepath.Join(tmpDir, "empty-workspace")
-	if err := os.MkdirAll(wsDir, 0755); err != nil {
-		t.Fatalf("failed to create config dir: %v", err)
-	}
-
-	// Write config.yaml without shared_context
-	configYAML := "name: Empty Workspace\ndescription: No shared context\n"
-	if err := os.WriteFile(filepath.Join(wsDir, "config.yaml"), []byte(configYAML), 0644); err != nil {
-		t.Fatalf("failed to write config.yaml: %v", err)
-	}
-
-	handler := NewTemplatesHandler(tmpDir, nil)
-
-	// Mock DB returning workspace name that normalizes to "empty-workspace"
-	mock.ExpectQuery("SELECT name FROM workspaces WHERE id =").
-		WithArgs("ws-empty").
-		WillReturnRows(sqlmock.NewRows([]string{"name"}).AddRow("Empty Workspace"))
-
-	w := httptest.NewRecorder()
-	c, _ := gin.CreateTestContext(w)
-	c.Params = gin.Params{{Key: "id", Value: "ws-empty"}}
-	c.Request = httptest.NewRequest("GET", "/workspaces/ws-empty/shared-context", nil)
-
-	handler.SharedContext(c)
-
-	if w.Code != http.StatusOK {
-		t.Errorf("expected status 200, got %d: %s", w.Code, w.Body.String())
-	}
-
-	var resp []interface{}
-	if err := json.Unmarshal(w.Body.Bytes(), &resp); err != nil {
-		t.Fatalf("failed to parse response: %v", err)
-	}
-	if len(resp) != 0 {
-		t.Errorf("expected empty array, got %d items", len(resp))
-	}
-
-	if err := mock.ExpectationsWereMet(); err != nil {
-		t.Errorf("unmet sqlmock expectations: %v", err)
-	}
-}
-
 // TestActivityHandler_Report_SourceIDSpoofRejected verifies the #209 spoof
 // guard: a workspace authenticated for :id cannot inject activity rows with
 // source_id pointing at a different workspace. Bearer-auth middleware would

workspace-server/internal/handlers/mcp.go

@@ -83,6 +83,12 @@ type mcpTool struct {
 type MCPHandler struct {
 	database    *sql.DB
 	broadcaster *events.Broadcaster
+
+	// memv2 is the v2 memory plugin wiring (RFC #2728). nil-safe:
+	// every v2 tool calls memoryV2Available() first and returns a
+	// clear error rather than crashing when the operator hasn't set
+	// MEMORY_PLUGIN_URL.
+	memv2 *memoryV2Deps
 }
 
 // NewMCPHandler wires the handler to db and broadcaster.
@@ -217,6 +223,76 @@ var mcpAllTools = []mcpTool{
 			},
 		},
 	},
+
+	// ─────────────────────────────────────────────────────────────────
+	// v2 memory tools (RFC #2728). Coexist with legacy commit_memory /
+	// recall_memory; PR-6 aliases the legacy names. Surface here so
+	// agents calling tools/list see them when MEMORY_PLUGIN_URL is
+	// configured (handlers no-op cleanly when it isn't).
+	// ─────────────────────────────────────────────────────────────────
+	{
+		Name:        "commit_memory_v2",
+		Description: "Save a memory to a namespace. Defaults to your own workspace. Use list_writable_namespaces to discover what else you can write to. Server applies SAFE-T1201 redaction before storage.",
+		InputSchema: map[string]interface{}{
+			"type": "object",
+			"properties": map[string]interface{}{
+				"content":    map[string]interface{}{"type": "string"},
+				"namespace":  map[string]interface{}{"type": "string"},
+				"kind":       map[string]interface{}{"type": "string", "enum": []string{"fact", "summary", "checkpoint"}},
+				"expires_at": map[string]interface{}{"type": "string", "description": "RFC3339"},
+				"pin":        map[string]interface{}{"type": "boolean"},
+			},
+			"required": []string{"content"},
+		},
+	},
+	{
+		Name:        "search_memory",
+		Description: "Search memories across one or more namespaces. Empty namespaces = search everything readable. Server applies ACL intersection before querying.",
+		InputSchema: map[string]interface{}{
+			"type": "object",
+			"properties": map[string]interface{}{
+				"query":      map[string]interface{}{"type": "string"},
+				"namespaces": map[string]interface{}{"type": "array", "items": map[string]interface{}{"type": "string"}},
+				"kinds":      map[string]interface{}{"type": "array", "items": map[string]interface{}{"type": "string", "enum": []string{"fact", "summary", "checkpoint"}}},
+				"limit":      map[string]interface{}{"type": "integer"},
+			},
+		},
+	},
+	{
+		Name:        "commit_summary",
+		Description: "Save an end-of-session summary. Same shape as commit_memory_v2 but kind=summary and a 30-day default TTL.",
+		InputSchema: map[string]interface{}{
+			"type": "object",
+			"properties": map[string]interface{}{
+				"content":    map[string]interface{}{"type": "string"},
+				"namespace":  map[string]interface{}{"type": "string"},
+				"expires_at": map[string]interface{}{"type": "string"},
+			},
+			"required": []string{"content"},
+		},
+	},
+	{
+		Name:        "list_writable_namespaces",
+		Description: "List the namespaces this workspace can write to.",
+		InputSchema: map[string]interface{}{"type": "object", "properties": map[string]interface{}{}},
+	},
+	{
+		Name:        "list_readable_namespaces",
+		Description: "List the namespaces this workspace can read from.",
+		InputSchema: map[string]interface{}{"type": "object", "properties": map[string]interface{}{}},
+	},
+	{
+		Name:        "forget_memory",
+		Description: "Delete a memory by id. Only memories in namespaces you can write to can be forgotten.",
+		InputSchema: map[string]interface{}{
+			"type": "object",
+			"properties": map[string]interface{}{
+				"memory_id": map[string]interface{}{"type": "string"},
+				"namespace": map[string]interface{}{"type": "string"},
+			},
+			"required": []string{"memory_id"},
+		},
+	},
 }
 
 // mcpToolList returns the filtered tool list for this MCP bridge.
@@ -363,6 +439,14 @@ func (h *MCPHandler) dispatchRPC(ctx context.Context, workspaceID string, req mc
 // Tool dispatch
 // ─────────────────────────────────────────────────────────────────────────────
 
+// Dispatch is the public entry point external code (tests, future
+// out-of-package callers) uses to invoke a tool by name. Forwards
+// to the unexported dispatch so existing in-package call sites
+// stay unchanged.
+func (h *MCPHandler) Dispatch(ctx context.Context, workspaceID, toolName string, args map[string]interface{}) (string, error) {
+	return h.dispatch(ctx, workspaceID, toolName, args)
+}
+
 func (h *MCPHandler) dispatch(ctx context.Context, workspaceID, toolName string, args map[string]interface{}) (string, error) {
 	switch toolName {
 	case "list_peers":
@@ -381,6 +465,22 @@ func (h *MCPHandler) dispatch(ctx context.Context, workspaceID, toolName string,
 		return h.toolCommitMemory(ctx, workspaceID, args)
 	case "recall_memory":
 		return h.toolRecallMemory(ctx, workspaceID, args)
+
+	// v2 memory tools (RFC #2728). PR-6 will alias the legacy names to
+	// these; until then they are independent surfaces.
+	case "commit_memory_v2":
+		return h.toolCommitMemoryV2(ctx, workspaceID, args)
+	case "search_memory":
+		return h.toolSearchMemory(ctx, workspaceID, args)
+	case "commit_summary":
+		return h.toolCommitSummary(ctx, workspaceID, args)
+	case "list_writable_namespaces":
+		return h.toolListWritableNamespaces(ctx, workspaceID, args)
+	case "list_readable_namespaces":
+		return h.toolListReadableNamespaces(ctx, workspaceID, args)
+	case "forget_memory":
+		return h.toolForgetMemory(ctx, workspaceID, args)
+
 	default:
 		return "", fmt.Errorf("unknown tool: %s", toolName)
 	}