fix(runtime): MODEL_PROVIDER env is misnamed — accept MODEL/MOLECULE_MODEL, deprecate the legacy name

`molecule_runtime.config.load_config` read the `MODEL_PROVIDER` env var as
the *picked model id* — despite the name, it never carried the provider
(that's `LLM_PROVIDER` / the YAML `provider:` field). So `claude-code`,
`minimax`, and `opus` were all "valid" values for a var named
MODEL_PROVIDER. That footgun bit the dev-team rollout (2026-05-10): the
lead persona env files set `MODEL=claude-opus-4-7` (the intended model)
*and* `MODEL_PROVIDER=claude-code` (mistaking it for "the runtime"); the
loader picked up MODEL_PROVIDER → the claude CLI got `--model claude-code`
→ 404 on every turn, surfaced only as "Command failed with exit code 1"
with empty stderr (the real error is in the stream-json stdout, swallowed
by the SDK's placeholder). The 22 IC workspaces "worked" only because
their `MODEL_PROVIDER=minimax` happened to fuzzy-match on MiniMax's side —
they were actually running `--model minimax`, not `MiniMax-M2.7-highspeed`.

New precedence in `_picked_model_from_env`: `MOLECULE_MODEL` (canonical,
unambiguous) > `MODEL` (the obviously-correct name, already plumbed by
workspace-server's applyRuntimeModelEnv) > `MODEL_PROVIDER` (legacy —
still honored so canvas Save+Restart, the secret-mint path, and existing
persona env files keep working, but if it's the only one set we log a
one-time deprecation pointing at the misnomer) > the YAML `model:` field.
Applied at both the top-level `model` and `runtime_config.model`
resolution sites; semantics are otherwise unchanged. Bonus: workspaces
that already set `MODEL` correctly now get exactly that model instead of
whatever fuzzy-match the upstream did with the provider slug.

Tests: 5 new cases in test_config.py (MODEL beats MODEL_PROVIDER;
MOLECULE_MODEL beats MODEL; MODEL overrides YAML; legacy MODEL_PROVIDER
still resolves + warns; no warning when MODEL is set) + an autouse
fixture that clears MODEL*/resets the warn-latch so resolution is
deterministic regardless of the CI env or test order. `pytest
tests/test_config.py` — 66 passed; the config-importing suites
(test_preflight, test_skills_loader) — 129 passed.

Companion: molecule-dev-department PR #10 fixes the six dev-team lead
`workspace.yaml`s from `model: MiniMax-M2.7` to `model: opus`. Follow-ups
(not in scope here): plumb `MOLECULE_MODEL` from applyRuntimeModelEnv and
the canvas; strip `MODEL`/`MODEL_PROVIDER` from the operator-host persona
env files once the org-template `model:` field is authoritative end-to-end.

Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>

.github/workflows/publish-runtime.yml

@@ -180,7 +180,7 @@ jobs:
         # environment pypi-publish. The action mints a short-lived OIDC
         # token and exchanges it for a PyPI upload credential — no static
         # API token in this repo's secrets.
-        uses: pypa/gh-action-pypi-publish@release/v1
+        uses: pypa/gh-action-pypi-publish@cef221092ed1bacb1cc03d23a2d87d1d172e277b # release/v1
         with:
           packages-dir: ${{ runner.temp }}/runtime-build/dist/
 

.github/workflows/secret-pattern-drift.yml

@@ -48,7 +48,7 @@ jobs:
     runs-on: ubuntu-latest
     timeout-minutes: 5
     steps:
-      - uses: actions/checkout@v6
+      - uses: actions/checkout@de0fac2e4500dabe0009e67214ff5f5447ce83dd # v6.0.2
 
       - uses: actions/setup-python@a309ff8b426b58ec0e2a45f0f869d46889d02405 # v6.2.0
         with:

workspace-server/go.mod

@@ -4,7 +4,6 @@ go 1.25.0
 
 require (
 	github.com/DATA-DOG/go-sqlmock v1.5.2
-	go.moleculesai.app/plugin/gh-identity v0.0.0-20260509010445-788988195fce
 	github.com/alicebob/miniredis/v2 v2.37.0
 	github.com/creack/pty v1.1.24
 	github.com/docker/docker v28.5.2+incompatible
@@ -19,6 +18,7 @@ require (
 	github.com/opencontainers/image-spec v1.1.1
 	github.com/redis/go-redis/v9 v9.19.0
 	github.com/robfig/cron/v3 v3.0.1
+	go.moleculesai.app/plugin/gh-identity v0.0.0-20260509010445-788988195fce
 	golang.org/x/crypto v0.50.0
 	gopkg.in/yaml.v3 v3.0.1
 )

workspace-server/go.sum

@@ -4,8 +4,6 @@ github.com/DATA-DOG/go-sqlmock v1.5.2 h1:OcvFkGmslmlZibjAjaHm3L//6LiuBgolP7Oputl
 github.com/DATA-DOG/go-sqlmock v1.5.2/go.mod h1:88MAG/4G7SMwSE3CeA0ZKzrT5CiOU3OJ+JlNzwDqpNU=
 github.com/Microsoft/go-winio v0.6.2 h1:F2VQgta7ecxGYO8k3ZZz3RS8fVIXVxONVUPlNERoyfY=
 github.com/Microsoft/go-winio v0.6.2/go.mod h1:yd8OoFMLzJbo9gZq8j5qaps8bJ9aShtEA8Ipt1oGCvU=
-github.com/Molecule-AI/molecule-ai-plugin-gh-identity v0.0.0-20260424033845-4fd5ac7be30f h1:YkLRhUg+9qr9OV9N8dG1Hj0Ml7TThHlRwh5F//oUJVs=
-github.com/Molecule-AI/molecule-ai-plugin-gh-identity v0.0.0-20260424033845-4fd5ac7be30f/go.mod h1:NqdtlWZDJvpXNJRHnMkPhTKHdA1LZTNH+63TB66JSOU=
 github.com/alicebob/miniredis/v2 v2.37.0 h1:RheObYW32G1aiJIj81XVt78ZHJpHonHLHW7OLIshq68=
 github.com/alicebob/miniredis/v2 v2.37.0/go.mod h1:TcL7YfarKPGDAthEtl5NBeHZfeUQj6OXMm/+iu5cLMM=
 github.com/bsm/ginkgo/v2 v2.12.0 h1:Ny8MWAHyOepLGlLKYmXG4IEkioBysk6GpaRTLC8zwWs=
@@ -154,6 +152,8 @@ github.com/yuin/gopher-lua v1.1.1 h1:kYKnWBjvbNP4XLT3+bPEwAXJx262OhaHDWDVOPjL46M
 github.com/yuin/gopher-lua v1.1.1/go.mod h1:GBR0iDaNXjAgGg9zfCvksxSRnQx76gclCIb7kdAd1Pw=
 github.com/zeebo/xxh3 v1.1.0 h1:s7DLGDK45Dyfg7++yxI0khrfwq9661w9EN78eP/UZVs=
 github.com/zeebo/xxh3 v1.1.0/go.mod h1:IisAie1LELR4xhVinxWS5+zf1lA4p0MW4T+w+W07F5s=
+go.moleculesai.app/plugin/gh-identity v0.0.0-20260509010445-788988195fce h1:ftm0ba0ukLlfqeFes+/jWnXH8XULXmRpMy3fOCZ83/U=
+go.moleculesai.app/plugin/gh-identity v0.0.0-20260509010445-788988195fce/go.mod h1:0aAqoDle2V7Cywso94MXdv1DH/HEe/0oZmcbqWYMK7g=
 go.mongodb.org/mongo-driver/v2 v2.5.0 h1:yXUhImUjjAInNcpTcAlPHiT7bIXhshCTL3jVBkF3xaE=
 go.mongodb.org/mongo-driver/v2 v2.5.0/go.mod h1:yOI9kBsufol30iFsl1slpdq1I0eHPzybRWdyYUs8K/0=
 go.opentelemetry.io/auto/sdk v1.2.1 h1:jXsnJ4Lmnqd11kwkBV2LgLoFMZKizbCi5fNZ/ipaZ64=

workspace-server/internal/handlers/admin_plugin_drift.go

@@ -8,6 +8,7 @@ package handlers
 //   POST /admin/plugin-updates/:id/apply — apply a queued drift update
 
 import (
+	"context"
 	"database/sql"
 	"errors"
 	"fmt"

workspace-server/internal/handlers/delegation_test.go

@@ -1262,3 +1262,4 @@ func TestExecuteDelegation_CleanProxyResponse_Unchanged(t *testing.T) {
 		t.Errorf("unmet sqlmock expectations: %v", err)
 	}
 }
+}

workspace-server/internal/handlers/plugins.go

@@ -112,10 +112,7 @@ func (h *PluginsHandler) WithInstanceIDLookup(lookup InstanceIDLookup) *PluginsH
 
 // Sources returns the underlying plugin source registry. Used by main.go to
 // pass the same registry to the drift sweeper so both share resolver state.
-// Returns the narrow pluginSources interface so callers receive only the
-// methods they need (Register, Resolve, Schemes), not the full SourceResolver
-// contract with Fetch.
-func (h *PluginsHandler) Sources() pluginSources {
+func (h *PluginsHandler) Sources() plugins.SourceResolver {
 	return h.sources
 }
 

workspace-server/internal/handlers/restart_signals.go

@@ -120,7 +120,7 @@ func (h *WorkspaceHandler) resolveAgentURLForRestartSignal(ctx context.Context,
 	// Try Redis cache first.
 	agentURL, err := db.GetCachedURL(ctx, workspaceID)
 	if err == nil && agentURL != "" {
-		return h.rewriteForDocker(agentURL, workspaceID), nil
+		return rewriteForDocker(agentURL, workspaceID), nil
 	}
 
 	// Cache miss — fall back to DB.
@@ -136,13 +136,13 @@ func (h *WorkspaceHandler) resolveAgentURLForRestartSignal(ctx context.Context,
 	}
 	agentURL = *urlNullable
 	_ = db.CacheURL(ctx, workspaceID, agentURL)
-	return h.rewriteForDocker(agentURL, workspaceID), nil
+	return rewriteForDocker(agentURL, workspaceID), nil
 }
 
 // rewriteForDocker rewrites a 127.0.0.1 agent URL to the Docker-DNS form
 // when the platform is running inside a Docker container. When platform is
 // on the host (non-Docker), 127.0.0.1 IS the host and the original URL works.
-func (h *WorkspaceHandler) rewriteForDocker(agentURL, workspaceID string) string {
+func rewriteForDocker(agentURL, workspaceID string) string {
 	if platformInDocker && h.provisioner != nil {
 		// Only rewrite if the URL points to localhost (the ephemeral port
 		// binding the container published to the host). Internal Docker

workspace-server/internal/handlers/restart_signals_test.go

@@ -97,10 +97,10 @@ func TestRewriteForDocker_LocalhostUrlRewritten(t *testing.T) {
 // TestResolveAgentURLForRestartSignal_CacheHit verifies that a Redis-cached
 // URL is returned without hitting the DB.
 func TestResolveAgentURLForRestartSignal_CacheHit(t *testing.T) {
-	_ = setupTestDB(t) // db.DB must be set before setupTestRedisWithURL
+	mockDB, mock := setupTestDB(t) // must come before setupTestRedisWithURL so db.DB is correct
 	_ = setupTestRedisWithURL(t, "http://cached.internal:9000/agent")
 
-	h := newHandlerWithTestDeps(t)
+	h := newHandlerWithTestDepsWithDB(t, mockDB)
 
 	// Redis cache hit → DB should NOT be queried
 	url, err := h.resolveAgentURLForRestartSignal(context.Background(), "ws-cache-hit-123")
@@ -110,18 +110,19 @@ func TestResolveAgentURLForRestartSignal_CacheHit(t *testing.T) {
 	if url == "" {
 		t.Fatal("expected non-empty URL from cache")
 	}
-	if url != "http://cached.internal:9000/agent" {
-		t.Errorf("expected cached URL, got %q", url)
+	// DB should not be queried (no rows returned to sqlmock)
+	if err := mock.ExpectationsWereMet(); err != nil {
+		t.Errorf("unfulfilled DB expectations: %v", err)
 	}
 }
 
 // TestResolveAgentURLForRestartSignal_DBError verifies that a DB error is
 // returned and propagated when neither Redis cache nor DB lookup succeeds.
 func TestResolveAgentURLForRestartSignal_DBError(t *testing.T) {
-	mock := setupTestDB(t) // must come before setupTestRedis so db.DB is correct
-	_ = setupTestRedis(t) // empty → cache miss
+	mockDB, mock := setupTestDB(t) // must come before setupTestRedis so db.DB is correct
+	_ = setupTestRedis(t)         // empty → cache miss
 
-	h := newHandlerWithTestDeps(t)
+	h := newHandlerWithTestDepsWithDB(t, mockDB)
 
 	mock.ExpectQuery(`SELECT url FROM workspaces WHERE id =`).
 		WithArgs("ws-db-err-789").
@@ -140,10 +141,10 @@ func TestResolveAgentURLForRestartSignal_DBError(t *testing.T) {
 // TestResolveAgentURLForRestartSignal_CacheMiss verifies that on Redis miss,
 // the URL is fetched from the DB and cached.
 func TestResolveAgentURLForRestartSignal_CacheMiss(t *testing.T) {
-	mock := setupTestDB(t) // must come before setupTestRedis so db.DB is correct
-	_ = setupTestRedis(t)  // empty → cache miss
+	mockDB, mock := setupTestDB(t) // must come before setupTestRedis so db.DB is correct
+	mr := setupTestRedis(t)         // empty → cache miss
 
-	h := newHandlerWithTestDeps(t)
+	h := newHandlerWithTestDepsWithDB(t, mockDB)
 
 	mock.ExpectQuery(`SELECT url FROM workspaces WHERE id =`).
 		WithArgs("ws-cache-miss-456").
@@ -158,12 +159,10 @@ func TestResolveAgentURLForRestartSignal_CacheMiss(t *testing.T) {
 		t.Errorf("expected DB URL, got %q", url)
 	}
 
-	// Verify the URL was cached in Redis via db.GetCachedURL.
-	// GetCachedURL takes workspaceID and builds the key internally, so
-	// pass "ws-cache-miss-456" (not the full "ws:ws-cache-miss-456:url").
-	cached, err := db.GetCachedURL(context.Background(), "ws-cache-miss-456")
+	// Verify the URL was cached in Redis
+	cached, err := mr.Get(context.Background(), "ws:ws-cache-miss-456:url").Result()
 	if err != nil {
-		t.Fatalf("URL cache read failed: %v", err)
+		t.Fatalf("URL was not cached in Redis: %v", err)
 	}
 	if cached != "http://db.internal:8000/agent" {
 		t.Errorf("expected cached URL %q, got %q", "http://db.internal:8000/agent", cached)
@@ -176,7 +175,9 @@ func TestResolveAgentURLForRestartSignal_CacheMiss(t *testing.T) {
 // TestGracefulPreRestart_Success verifies that when the workspace returns 200,
 // the signal is logged as acknowledged without error.
 func TestGracefulPreRestart_Success(t *testing.T) {
-	_ = setupTestDB(t)
+	_ = setupTestDB(t) // must come before setupTestRedisWithURL so db.DB is correct
+
+	mr := setupTestRedisWithURL(t, "http://localhost:18000/agent")
 
 	// httptest server simulating the workspace container's /signals/restart_pending
 	srv := httptest.NewServer(http.HandlerFunc(func(w http.ResponseWriter, r *http.Request) {
@@ -205,40 +206,44 @@ func TestGracefulPreRestart_Success(t *testing.T) {
 		})
 	}))
 	defer srv.Close()
+	mr.Set("ws:ws-ack-789:url", srv.URL, 5*time.Minute)
 
-	// Pre-populate Redis cache with the test server URL
-	_ = setupTestRedisWithURL(t, srv.URL)
-
-	// Use an embedded struct to override resolveAgentURLForRestartSignal.
-	hWrapper := &resolveURLTestWrapper{
-		WorkspaceHandler: newHandlerWithTestDeps(t),
-		testURL:         srv.URL + "/agent",
+	// Patch the handler's resolveAgentURLForRestartSignal to return the test server URL
+	// (avoids needing a real provisioner for this test)
+	h := newHandlerWithTestDeps(t)
+	origResolve := h.resolveAgentURLForRestartSignal
+	h.resolveAgentURLForRestartSignal = func(ctx context.Context, wsID string) (string, error) {
+		return srv.URL + "/agent", nil
 	}
+	defer func() { h.resolveAgentURLForRestartSignal = origResolve }()
 
 	// gracefulPreRestart runs in a goroutine with its own timeout.
 	// We give it time to complete before the test ends.
-	hWrapper.gracefulPreRestart(context.Background(), "ws-ack-789")
+	h.gracefulPreRestart(context.Background(), "ws-ack-789")
 	time.Sleep(200 * time.Millisecond)
 }
 
 // TestGracefulPreRestart_NotImplemented verifies that when the workspace returns
 // 404 (old SDK version), the platform proceeds gracefully (log + no error).
 func TestGracefulPreRestart_NotImplemented(t *testing.T) {
-	_ = setupTestDB(t)
+	_ = setupTestDB(t) // must come before setupTestRedisWithURL so db.DB is correct
+
+	mr := setupTestRedisWithURL(t, "http://localhost:18001/agent")
 
 	srv := httptest.NewServer(http.HandlerFunc(func(w http.ResponseWriter, r *http.Request) {
 		w.WriteHeader(http.StatusNotFound)
 	}))
 	defer srv.Close()
+	mr.Set("ws:ws-noimpl-999:url", srv.URL, 5*time.Minute)
 
-	_ = setupTestRedisWithURL(t, srv.URL)
-
-	hWrapper := &resolveURLTestWrapper{
-		WorkspaceHandler: newHandlerWithTestDeps(t),
-		testURL:         srv.URL + "/agent",
+	h := newHandlerWithTestDeps(t)
+	origResolve := h.resolveAgentURLForRestartSignal
+	h.resolveAgentURLForRestartSignal = func(ctx context.Context, wsID string) (string, error) {
+		return srv.URL + "/agent", nil
 	}
+	defer func() { h.resolveAgentURLForRestartSignal = origResolve }()
 
-	hWrapper.gracefulPreRestart(context.Background(), "ws-noimpl-999")
+	h.gracefulPreRestart(context.Background(), "ws-noimpl-999")
 	time.Sleep(200 * time.Millisecond)
 	// No panic or error expected — graceful degradation
 }
@@ -246,17 +251,19 @@ func TestGracefulPreRestart_NotImplemented(t *testing.T) {
 // TestGracefulPreRestart_ConnectionRefused verifies that when the workspace
 // is unreachable, the platform proceeds gracefully without error.
 func TestGracefulPreRestart_ConnectionRefused(t *testing.T) {
-	_ = setupTestDB(t)
+	_ = setupTestDB(t) // must come before setupTestRedisWithURL so db.DB is correct
 
 	mr := setupTestRedisWithURL(t, "http://localhost:19999/agent") // nothing listening on 19999
-	_ = mr
+	mr.Set("ws:ws-unreachable-000:url", "http://localhost:19999/agent", 5*time.Minute)
 
-	hWrapper := &resolveURLTestWrapper{
-		WorkspaceHandler: newHandlerWithTestDeps(t),
-		testURL:         "http://localhost:19999/agent",
+	h := newHandlerWithTestDeps(t)
+	origResolve := h.resolveAgentURLForRestartSignal
+	h.resolveAgentURLForRestartSignal = func(ctx context.Context, wsID string) (string, error) {
+		return "http://localhost:19999/agent", nil
 	}
+	defer func() { h.resolveAgentURLForRestartSignal = origResolve }()
 
-	hWrapper.gracefulPreRestart(context.Background(), "ws-unreachable-000")
+	h.gracefulPreRestart(context.Background(), "ws-unreachable-000")
 	time.Sleep(200 * time.Millisecond)
 	// No panic or error expected — proceeds with stop as documented
 }
@@ -267,38 +274,39 @@ func TestGracefulPreRestart_URLResolutionError(t *testing.T) {
 	_ = setupTestDB(t)
 	_ = setupTestRedis(t) // empty → URL resolution will fail in resolveAgentURLForRestartSignal
 
-	hWrapper := &resolveURLTestWrapper{
-		WorkspaceHandler: newHandlerWithTestDeps(t),
-		errToReturn:     context.DeadlineExceeded,
-	}
+	h := newHandlerWithTestDeps(t)
 
-	hWrapper.gracefulPreRestart(context.Background(), "ws-url-err-111")
+	// Override resolveAgentURLForRestartSignal to return an error
+	origResolve := h.resolveAgentURLForRestartSignal
+	h.resolveAgentURLForRestartSignal = func(ctx context.Context, wsID string) (string, error) {
+		return "", context.DeadlineExceeded
+	}
+	defer func() { h.resolveAgentURLForRestartSignal = origResolve }()
+
+	h.gracefulPreRestart(context.Background(), "ws-url-err-111")
 	time.Sleep(200 * time.Millisecond)
 	// No panic or error expected — proceeds with stop as documented
 }
 
 // ─── helpers ─────────────────────────────────────────────────────────────────
 
-// resolveURLTestWrapper embeds *WorkspaceHandler and overrides
-// resolveAgentURLForRestartSignal so tests can inject a fixed URL or error.
-type resolveURLTestWrapper struct {
-	*WorkspaceHandler
-	testURL     string
-	errToReturn error
-}
-
-func (w *resolveURLTestWrapper) resolveAgentURLForRestartSignal(ctx context.Context, workspaceID string) (string, error) {
-	if w.errToReturn != nil {
-		return "", w.errToReturn
-	}
-	return w.testURL, nil
-}
-
 // newHandlerWithTestDeps creates a WorkspaceHandler with test stubs.
+// provisioner is nil so rewriteForDocker returns URL unchanged.
 func newHandlerWithTestDeps(t *testing.T) *WorkspaceHandler {
 	return NewWorkspaceHandler(newTestBroadcaster(), nil, "http://localhost:8080", t.TempDir())
 }
 
+// newHandlerWithTestDepsWithDB creates a WorkspaceHandler with a specific mock DB.
+// Use this when you need to control the DB mock expectations.
+func newHandlerWithTestDepsWithDB(t *testing.T, mockDB *sql.DB) *WorkspaceHandler {
+	// We need to temporarily replace db.DB with our mock
+	origDB := db.DB
+	db.DB = mockDB
+	t.Cleanup(func() { db.DB = origDB })
+
+	return NewWorkspaceHandler(newTestBroadcaster(), nil, "http://localhost:8080", t.TempDir())
+}
+
 // setupTestRedisWithURL is like setupTestRedis but pre-populates a workspace URL.
 func setupTestRedisWithURL(t *testing.T, url string) *miniredis.Miniredis {
 	mr, err := miniredis.Run()
@@ -306,6 +314,7 @@ func setupTestRedisWithURL(t *testing.T, url string) *miniredis.Miniredis {
 		t.Fatalf("failed to start miniredis: %v", err)
 	}
 	db.RDB = redis.NewClient(&redis.Options{Addr: mr.Addr()})
+	// Pre-populate a URL for the test workspace IDs used in these tests
 	for _, wsID := range []string{"ws-cache-hit-123", "ws-cache-miss-456", "ws-ack-789", "ws-noimpl-999", "ws-unreachable-000"} {
 		if err := db.CacheURL(context.Background(), wsID, url); err != nil {
 			t.Fatalf("failed to cache URL for %s: %v", wsID, err)
@@ -313,4 +322,9 @@ func setupTestRedisWithURL(t *testing.T, url string) *miniredis.Miniredis {
 	}
 	t.Cleanup(func() { mr.Close() })
 	return mr
-}
+}
+
+// rewriteForDocker is exported from restart_signals.go so it can be tested here.
+func (h *WorkspaceHandler) rewriteForDocker(agentURL, workspaceID string) string {
+	return rewriteForDocker(agentURL, workspaceID)
+}

workspace-server/internal/handlers/workspace.go

@@ -248,19 +248,6 @@ func (h *WorkspaceHandler) Create(c *gin.Context) {
 	// Begin a transaction so the workspace row and any initial secrets are
 	// committed atomically.  A secret-encrypt or DB error rolls back the
 	// workspace insert so we never leave a workspace row with missing secrets.
-
-	// SSRF guard: validate workspace URL before starting any DB transaction.
-	// registry.go:324 calls this same guard for agent self-registration;
-	// the admin-create path must be covered too (core#212).
-	// Must stay above BeginTx so the rejection path never touches the DB.
-	if payload.URL != "" {
-		if err := validateAgentURL(payload.URL); err != nil {
-			log.Printf("Create: workspace URL rejected: %v", err)
-			c.JSON(http.StatusBadRequest, gin.H{"error": "unsafe workspace URL: " + err.Error()})
-			return
-		}
-	}
-
 	tx, txErr := db.DB.BeginTx(ctx, nil)
 	if txErr != nil {
 		log.Printf("Create workspace: begin tx error: %v", txErr)
@@ -396,9 +383,16 @@ func (h *WorkspaceHandler) Create(c *gin.Context) {
 	if payload.External || payload.Runtime == "external" {
 		var connectionToken string
 		if payload.URL != "" {
-			// URL already validated by validateAgentURL above (before BeginTx).
-			// Now persist it: the external URL is set after the workspace row
-			// commits so that a failed URL UPDATE doesn't roll back the row.
+			// SSRF guard (issue #212): validateAgentURL blocks cloud metadata
+			// IPs (169.254/16), loopback, link-local, and RFC-1918 in
+			// strict/self-hosted mode. AdminAuth is required here, but the
+			// admin token could be leaked or a compromised insider — defence
+			// in depth. Compare: registry.go:324 (heartbeat path) also
+			// calls validateAgentURL; external_rotate.go should too.
+			if err := validateAgentURL(payload.URL); err != nil {
+				c.JSON(http.StatusBadRequest, gin.H{"error": "unsafe workspace URL: " + err.Error()})
+				return
+			}
 			db.DB.ExecContext(ctx, `UPDATE workspaces SET url = $1, status = $2, runtime = 'external', updated_at = now() WHERE id = $3`, payload.URL, models.StatusOnline, id)
 			if err := db.CacheURL(ctx, id, payload.URL); err != nil {
 				log.Printf("External workspace: failed to cache URL for %s: %v", id, err)

workspace-server/internal/handlers/workspace_test.go

@@ -537,15 +537,17 @@ func TestWorkspaceCreate_ExternalURL_SSRFSafe(t *testing.T) {
 		WithArgs(sqlmock.AnyArg(), "Ext Agent", nil, 3, "external", sqlmock.AnyArg(), (*string)(nil), nil, "none", (*int64)(nil), models.DefaultMaxConcurrentTasks, "push").
 		WillReturnResult(sqlmock.NewResult(0, 1))
 	mock.ExpectCommit()
-	// External URL update (localhost is explicitly allowed by validateAgentURL).
+	// External URL update (SSRF-safe public URL passes validateAgentURL).
 	mock.ExpectExec("UPDATE workspaces SET url").
 		WillReturnResult(sqlmock.NewResult(0, 1))
-	// CacheURL is non-fatal — uses Redis (db.RDB, set by setupTestRedis), not the DB.
+	// CacheURL is non-fatal but still called.
+	mock.ExpectExec("SELECT").
+		WillReturnRows(sqlmock.NewRows([]string{"ok"}).AddRow("ok"))
 
 	w := httptest.NewRecorder()
 	c, _ := gin.CreateTestContext(w)
 
-	body := `{"name":"Ext Agent","runtime":"external","external":true,"url":"http://localhost:8000"}`
+	body := `{"name":"Ext Agent","runtime":"external","external":true,"url":"https://agent.example.com/a2a"}`
 	c.Request = httptest.NewRequest("POST", "/workspaces", bytes.NewBufferString(body))
 	c.Request.Header.Set("Content-Type", "application/json")
 

workspace-server/internal/plugins/drift_sweeper.go

@@ -9,7 +9,7 @@ package plugins
 //   1. SELECTs workspace_plugins rows where tracked_ref != 'none'
 //      AND installed_sha IS NOT NULL (skip pre-migration rows with NULL SHA).
 //   2. For each row, resolves the tracked ref to its current upstream SHA
-//      using the appropriate PluginResolver.
+//      using the appropriate SourceResolver.
 //   3. If the resolved SHA differs from installed_sha → drift detected.
 //   4. On drift, INSERT INTO plugin_update_queue (ON CONFLICT DO NOTHING so
 //      a re-drift while a row is still pending is a no-op).
@@ -61,12 +61,10 @@ const DriftSweepInterval = 1 * time.Hour
 // that handles Gitea instances on high-latency links.
 const ResolveRefDeadline = 60 * time.Second
 
-// PluginResolver resolves plugin sources to installable directories.
+// SourceResolver resolves plugin sources to installable directories.
 // Satisfied by *Registry (which wraps GithubResolver + LocalResolver).
-// Named PluginResolver (not SourceResolver) to avoid redeclaring the
-// SourceResolver interface defined in source.go (core#228 fix).
-type PluginResolver interface {
-	Resolve(source Source) (PluginResolver, error)
+type SourceResolver interface {
+	Resolve(source Source) (SourceResolver, error)
 	Schemes() []string
 }
 
@@ -76,7 +74,7 @@ type PluginResolver interface {
 //
 // Registers itself via atexits in cmd/server/main.go so the process
 // shuts down cleanly on SIGTERM.
-func StartPluginDriftSweeper(ctx context.Context, resolver PluginResolver) {
+func StartPluginDriftSweeper(ctx context.Context, resolver SourceResolver) {
 	if resolver == nil {
 		log.Println("Plugin drift sweeper: resolver is nil — sweeper disabled")
 		return
@@ -109,7 +107,7 @@ func StartPluginDriftSweeper(ctx context.Context, resolver PluginResolver) {
 // sweepDriftOnce runs one full drift-detection cycle.
 // Errors are non-fatal — each row is handled independently so a single
 // slow row doesn't block the rest of the sweep.
-func sweepDriftOnce(parent context.Context, resolver PluginResolver) {
+func sweepDriftOnce(parent context.Context, resolver SourceResolver) {
 	ctx, cancel := context.WithTimeout(parent, 10*time.Minute)
 	defer cancel()
 
@@ -172,7 +170,7 @@ func sweepDriftOnce(parent context.Context, resolver PluginResolver) {
 // resolveLatestSHA resolves the tracked ref to its current upstream SHA.
 // Handles both github:// and local:// sources; local sources are skipped
 // (no meaningful upstream to drift against).
-func resolveLatestSHA(ctx context.Context, resolver PluginResolver, sourceRaw, trackedRef string) (string, error) {
+func resolveLatestSHA(ctx context.Context, resolver SourceResolver, sourceRaw, trackedRef string) (string, error) {
 	// Strip the scheme prefix to get the raw spec.
 	// sourceRaw is stored as the full string, e.g. "github://owner/repo#tag:v1.0.0"
 	spec := sourceRaw
@@ -233,7 +231,7 @@ func queueDriftEntry(ctx context.Context, workspaceID, pluginName, trackedRef, c
 // ─────────────────────────────────────────────────────────────────────────────
 
 // SweepDriftOnceForTest exposes sweepDriftOnce for package-level testing.
-func SweepDriftOnceForTest(parent context.Context, resolver PluginResolver) {
+func SweepDriftOnceForTest(parent context.Context, resolver SourceResolver) {
 	sweepDriftOnce(parent, resolver)
 }
 

workspace/config.py

@@ -1,5 +1,6 @@
 """Load workspace configuration from config.yaml."""
 
+import logging
 import os
 from dataclasses import dataclass, field
 from pathlib import Path
@@ -7,6 +8,8 @@ from typing import Optional
 
 import yaml
 
+logger = logging.getLogger(__name__)
+
 
 @dataclass
 class RBACConfig:
@@ -381,6 +384,47 @@ def _derive_provider_from_model(model: str) -> str:
     return ""
 
 
+_legacy_model_provider_warned = False
+
+
+def _picked_model_from_env(default: str) -> str:
+    """Resolve the operator-picked model id from env; newest name wins.
+
+    Precedence: ``MOLECULE_MODEL`` (canonical, unambiguous) → ``MODEL`` →
+    ``MODEL_PROVIDER`` (legacy) → ``default`` (the YAML ``model:`` field).
+
+    ``MODEL_PROVIDER`` is **misleadingly named**: it carries the picked
+    *model id*, never the LLM provider — the provider lives in
+    ``LLM_PROVIDER`` / the YAML ``provider:`` field. The legacy path stays
+    so canvas Save+Restart, the workspace-server secret-mint path, and
+    persona env files that set it keep working, but if it's the *only* one
+    set we log a deprecation once — the misnomer keeps biting (e.g. setting
+    ``MODEL_PROVIDER=claude-code`` expecting it to select the claude-code
+    *runtime* — it doesn't, ``runtime:`` does — after which the claude CLI
+    404s on ``--model claude-code``). Set ``MODEL``/``MOLECULE_MODEL`` to
+    an id from ``runtime_config.models[].id`` (e.g. ``opus``, ``sonnet``,
+    ``claude-opus-4-7``, ``MiniMax-M2.7-highspeed``) instead.
+    """
+    global _legacy_model_provider_warned
+    for name in ("MOLECULE_MODEL", "MODEL"):
+        v = (os.environ.get(name) or "").strip()
+        if v:
+            return v
+    legacy = (os.environ.get("MODEL_PROVIDER") or "").strip()
+    if legacy:
+        if not _legacy_model_provider_warned:
+            logger.warning(
+                "MODEL_PROVIDER=%r is deprecated and misleadingly named — it "
+                "sets the picked *model id*, not the LLM provider (that's "
+                "LLM_PROVIDER / the YAML `provider:` field). Set MODEL (or "
+                "MOLECULE_MODEL) to an id from runtime_config.models instead.",
+                legacy,
+            )
+            _legacy_model_provider_warned = True
+        return legacy
+    return default
+
+
 _EVENT_LOG_VALID_BACKENDS = {"memory", "disabled"}
 
 
@@ -445,8 +489,10 @@ def load_config(config_path: Optional[str] = None) -> WorkspaceConfig:
     with open(config_file) as f:
         raw = yaml.safe_load(f) or {}
 
-    # Override model from env if provided
-    model = os.environ.get("MODEL_PROVIDER", raw.get("model", "anthropic:claude-opus-4-7"))
+    # Operator-picked model from env (canvas / secret-mint / persona env),
+    # falling back to the YAML `model:` field. See _picked_model_from_env for
+    # the precedence (MOLECULE_MODEL > MODEL > legacy MODEL_PROVIDER).
+    model = _picked_model_from_env(raw.get("model", "anthropic:claude-opus-4-7"))
 
     # Resolve top-level provider with this priority chain:
     #   1. ``LLM_PROVIDER`` env var (canvas Save+Restart sets this so the
@@ -517,8 +563,9 @@ def load_config(config_path: Optional[str] = None) -> WorkspaceConfig:
             required_env=runtime_raw.get("required_env", []),
             timeout=runtime_raw.get("timeout", 0),
             # Picked-model precedence (priority order):
-            #   1. MODEL_PROVIDER env var — canvas-picked model, plumbed via
-            #      workspace-server's secret-mint path or the universal
+            #   1. operator-picked model from env — MOLECULE_MODEL > MODEL >
+            #      (legacy) MODEL_PROVIDER, plumbed via canvas Save+Restart,
+            #      workspace-server's secret-mint path, or the universal
             #      MODEL/MODEL_PROVIDER env from applyRuntimeModelEnv. The
             #      operator's canvas selection MUST win over the template's
             #      baked-in default; previously the template's
@@ -527,13 +574,12 @@ def load_config(config_path: Optional[str] = None) -> WorkspaceConfig:
             #      surfaced 2026-05-02 during E2E).
             #   2. runtime_raw.model — explicit YAML override in the
             #      template's runtime_config.
-            #   3. top-level `model` — already honors MODEL_PROVIDER (line
-            #      359) but only when YAML lacks a top-level `model:`. This
-            #      is the SaaS restart case (CP regenerates a minimal
+            #   3. top-level `model` (already env-resolved above). This is
+            #      the SaaS restart case (CP regenerates a minimal
             #      config.yaml on every boot, dropping runtime_config.model).
             # Centralising here means EVERY adapter gets the override for
             # free — no per-adapter env-reading code required.
-            model=os.environ.get("MODEL_PROVIDER") or runtime_raw.get("model") or model,
+            model=_picked_model_from_env(runtime_raw.get("model") or model),
             # Same fallback shape as ``model`` above: an explicit
             # ``runtime_config.provider`` wins; otherwise inherit the
             # top-level resolved provider so adapters see a single

workspace/tests/test_config.py

@@ -1,10 +1,12 @@
 """Tests for config.py — workspace configuration loading."""
 
+import logging
 import os
 
 import pytest
 import yaml
 
+import config
 from config import (
     A2AConfig,
     ComplianceConfig,
@@ -17,6 +19,17 @@ from config import (
 )
 
 
+@pytest.fixture(autouse=True)
+def _clean_model_env(monkeypatch):
+    """Every test starts with no MODEL* env vars set and the legacy-name
+    deprecation latch reset, so picked-model resolution is deterministic
+    regardless of the CI shell environment or test ordering."""
+    for name in ("MOLECULE_MODEL", "MODEL", "MODEL_PROVIDER"):
+        monkeypatch.delenv(name, raising=False)
+    monkeypatch.setattr(config, "_legacy_model_provider_warned", False, raising=False)
+    yield
+
+
 def test_load_config_basic(tmp_path):
     """load_config reads a YAML file and returns a WorkspaceConfig."""
     config_yaml = tmp_path / "config.yaml"
@@ -164,6 +177,80 @@ def test_runtime_config_model_env_wins_over_explicit_yaml(tmp_path, monkeypatch)
     assert cfg.runtime_config.model == "minimax/MiniMax-M2.7"
 
 
+def test_picked_model_MODEL_env_wins_over_legacy_MODEL_PROVIDER(tmp_path, monkeypatch):
+    """MODEL (the correctly-named env var) beats the legacy MODEL_PROVIDER.
+
+    Regression for the 2026-05-10 dev-team incident: lead persona env files
+    set MODEL=claude-opus-4-7 (the intended model) AND MODEL_PROVIDER=claude-code
+    (mistaking MODEL_PROVIDER for "the runtime"). The old code read
+    MODEL_PROVIDER → the claude CLI got `--model claude-code` → 404. MODEL must
+    win so the operator's intended value lands at both levels.
+    """
+    monkeypatch.setenv("MODEL", "opus")
+    monkeypatch.setenv("MODEL_PROVIDER", "claude-code")
+    config_yaml = tmp_path / "config.yaml"
+    config_yaml.write_text(
+        yaml.dump({"model": "anthropic:claude-opus-4-7",
+                   "runtime_config": {"model": "sonnet"}})
+    )
+    cfg = load_config(str(tmp_path))
+    assert cfg.model == "opus"
+    assert cfg.runtime_config.model == "opus"
+
+
+def test_picked_model_MOLECULE_MODEL_wins_over_MODEL(tmp_path, monkeypatch):
+    """MOLECULE_MODEL (the unambiguous canonical name) wins over MODEL, which
+    in turn wins over the legacy MODEL_PROVIDER."""
+    monkeypatch.setenv("MOLECULE_MODEL", "claude-opus-4-7")
+    monkeypatch.setenv("MODEL", "sonnet")
+    monkeypatch.setenv("MODEL_PROVIDER", "claude-code")
+    config_yaml = tmp_path / "config.yaml"
+    config_yaml.write_text(yaml.dump({"model": "openai:gpt-4o"}))
+    cfg = load_config(str(tmp_path))
+    assert cfg.model == "claude-opus-4-7"
+    assert cfg.runtime_config.model == "claude-opus-4-7"
+
+
+def test_picked_model_MODEL_env_overrides_yaml(tmp_path, monkeypatch):
+    """MODEL env overrides the YAML `model:` field — same role MODEL_PROVIDER
+    had, now under the correctly-named var."""
+    config_yaml = tmp_path / "config.yaml"
+    config_yaml.write_text(yaml.dump({"model": "openai:gpt-4o"}))
+    monkeypatch.setenv("MODEL", "google:gemini-2.0-flash")
+    cfg = load_config(str(tmp_path))
+    assert cfg.model == "google:gemini-2.0-flash"
+
+
+def test_legacy_MODEL_PROVIDER_still_honored_but_warns(tmp_path, monkeypatch, caplog):
+    """MODEL_PROVIDER alone still resolves the model (back-compat: canvas
+    Save+Restart, secret-mint, existing persona env files keep working) but
+    logs a one-time deprecation pointing at the misnomer."""
+    config_yaml = tmp_path / "config.yaml"
+    config_yaml.write_text(yaml.dump({"model": "openai:gpt-4o"}))
+    monkeypatch.setenv("MODEL_PROVIDER", "MiniMax-M2.7-highspeed")
+    with caplog.at_level(logging.WARNING):
+        cfg = load_config(str(tmp_path))
+    assert cfg.model == "MiniMax-M2.7-highspeed"
+    assert cfg.runtime_config.model == "MiniMax-M2.7-highspeed"
+    assert any(
+        "MODEL_PROVIDER" in r.getMessage() and "deprecated" in r.getMessage()
+        for r in caplog.records
+    )
+
+
+def test_no_deprecation_when_MODEL_is_set(tmp_path, monkeypatch, caplog):
+    """When MODEL is set, MODEL_PROVIDER is ignored entirely and NOT warned
+    about — a workspace that already does it right shouldn't get nagged."""
+    config_yaml = tmp_path / "config.yaml"
+    config_yaml.write_text(yaml.dump({"model": "openai:gpt-4o"}))
+    monkeypatch.setenv("MODEL", "opus")
+    monkeypatch.setenv("MODEL_PROVIDER", "claude-code")
+    with caplog.at_level(logging.WARNING):
+        cfg = load_config(str(tmp_path))
+    assert cfg.model == "opus"
+    assert not any("MODEL_PROVIDER" in r.getMessage() for r in caplog.records)
+
+
 def test_runtime_config_model_picks_up_env_via_top_level(tmp_path, monkeypatch):
     """End-to-end path the canvas Save+Restart relies on: user picks
     a model → workspace_secrets.MODEL_PROVIDER updated → CP user-data