fix(a2a-mcp): use readline() not read(65536) for pipe-safe stdio

a2a_mcp_server.py main()'s stdio read loop used
`await loop.run_in_executor(None, stdin.read, 65536)`. On a PIPE,
read(n) blocks until n bytes accumulate OR EOF. A live MCP client
(openclaw bundle-mcp, Claude Code, Cursor) sends one ~150-byte
newline-delimited request and keeps stdin OPEN waiting for the reply,
so neither condition is met: the server never parses `initialize` and
the client times out (~30s; openclaw: "MCP error -32000: Connection
closed"). This silently broke peer visibility for every pipe-spawned
MCP host while passing all existing stdio tests, which only fed stdin
from a regular file or a heredoc-pipe that CLOSES (EOF returns
immediately). readline() returns as soon as one newline-delimited
line is available — exactly the JSON-RPC framing — and is
backward-compatible with the EOF/file cases.

Root cause of the 2026-05-15 openclaw peer-visibility outage
(workspace 95744c11): the molecule MCP server could not complete the
handshake over openclaw's stdio pipe, so the agent fell back to
native sessions_list. The openclaw template adapter fix
(template-openclaw#16) works around this via HTTP transport; this
patch fixes the stdio root cause so stdio works for all CLI MCP hosts.

Regression coverage:
- tests/test_a2a_mcp_server.py::TestStdioKeepOpenPipe — spawns the
  real a2a_mcp_server.py, writes one request over a pipe, and
  DELIBERATELY keeps stdin open. FAILS (15s timeout, empty response)
  on read(65536); PASSES on readline(). Verified both directions.
- ci-mcp-stdio-transport.yml: new "pipe held OPEN, no EOF" step that
  reproduces the literal openclaw failure (the prior steps only
  exercised EOF-closing stdin, which is why the outage shipped green).

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

.gitea/workflows/ci-mcp-stdio-transport.yml

@@ -158,8 +158,68 @@ jobs:
             echo "NOTE: No warning in output (may be suppressed by log level)"
           fi
 
+      - name: Reproduce openclaw failure — pipe held OPEN, no EOF
+        run: |
+          set -euo pipefail
+          echo "=== keep-stdin-open pipe (the real openclaw / Claude Code case) ==="
+          echo ""
+          echo "Before the readline() fix this HANGS: main() did"
+          echo "  stdin.read(65536)  -> on a pipe, blocks until 64KB OR EOF."
+          echo "An MCP client sends one ~150B initialize and keeps stdin"
+          echo "open waiting for the response, so the server never parsed"
+          echo "the request and the client timed out (openclaw: 'MCP error"
+          echo "-32000: Connection closed'). The earlier regular-file /"
+          echo "heredoc-pipe steps PASSED through this bug because a file"
+          echo "(or a closing heredoc) yields EOF immediately."
+          echo ""
+
+          # Drive the server through a real pipe that stays OPEN: write
+          # one initialize, do NOT close stdin, and require a response
+          # within a hard timeout. read(65536) -> no output -> timeout
+          # kills it -> FAIL. readline() -> immediate response -> PASS.
+          python - <<'PYEOF'
+          import json, subprocess, sys, time, select
+
+          proc = subprocess.Popen(
+              [sys.executable, "a2a_mcp_server.py"],
+              stdin=subprocess.PIPE, stdout=subprocess.PIPE,
+              stderr=subprocess.STDOUT,
+              env={**__import__("os").environ},
+          )
+          req = json.dumps({
+              "jsonrpc": "2.0", "id": 1, "method": "initialize",
+              "params": {"protocolVersion": "2024-11-05",
+                         "capabilities": {},
+                         "clientInfo": {"name": "keepopen", "version": "1"}},
+          }) + "\n"
+          proc.stdin.write(req.encode())
+          proc.stdin.flush()
+          # Deliberately DO NOT close proc.stdin — mirror a live MCP client.
+
+          deadline = time.time() + 15
+          line = b""
+          while time.time() < deadline:
+              r, _, _ = select.select([proc.stdout], [], [], 1)
+              if r:
+                  line = proc.stdout.readline()
+                  if line:
+                      break
+          proc.kill()
+
+          if not line:
+              print("FAIL: no response within 15s on an open pipe — "
+                    "stdin.read(65536) regression is back")
+              sys.exit(1)
+          resp = json.loads(line.decode())
+          assert resp.get("id") == 1 and "result" in resp, \
+              f"unexpected response: {line[:200]!r}"
+          assert resp["result"]["serverInfo"]["name"] == "molecule", \
+              f"wrong serverInfo: {line[:200]!r}"
+          print("PASS: server answered initialize on a still-open pipe")
+          PYEOF
+
       - name: Run unit tests for stdio transport
         run: |
           set -euo pipefail
           echo "=== Running stdio transport unit tests ==="
-          python -m pytest tests/test_a2a_mcp_server.py::TestStdioPipeAssertion -v --no-cov
+          python -m pytest tests/test_a2a_mcp_server.py::TestStdioPipeAssertion tests/test_a2a_mcp_server.py::TestStdioKeepOpenPipe -v --no-cov

workspace-server/internal/handlers/a2a_poll_ingest_persist_test.go

@@ -0,0 +1,160 @@
+package handlers
+
+// Regression coverage for the POLL-mode arm of the canvas user-message
+// data-loss bug (internal#470 sibling — tracked on internal#471).
+//
+// Bug (reported 2026-05-16 by CTO Hongming): "in canvas i sometimes lose
+// my own message when i exit chat". The push-mode arm was fixed by
+// #1347 (persistUserMessageAtIngest — a SYNCHRONOUS, before-dispatch,
+// context.WithoutCancel INSERT). #1347's framing asserted "poll-mode
+// workspaces were never affected — logA2AReceiveQueued already persists
+// at ingest". That assertion is OVERSTATED.
+//
+// Hongming's tenant (slug `hongming`, org 2c940477-...) has 4 workspaces,
+// ALL runtime=external with empty URL → ALL delivery_mode=poll (proven
+// empirically: a benign A2A probe returns the synthetic
+// {"delivery_mode":"poll","status":"queued"} envelope for every one).
+// So his reported loss is the POLL path, NOT the push path #1347 fixes.
+//
+// Root cause (poll arm): the poll-mode short-circuit (a2a_proxy.go ~402)
+// calls logA2AReceiveQueued and then IMMEDIATELY returns the synthetic
+// 200 {status:"queued"} to the canvas. But logA2AReceiveQueued's durable
+// INSERT runs inside h.goAsync(...) — a DETACHED goroutine with NO
+// happens-before barrier against the HTTP response. The canvas sees 200
+// ("message accepted") while the activity_logs row may not yet be — and,
+// on a workspace-server restart / deploy / OOM / EC2 hibernation between
+// the 200 and the goroutine's commit, NEVER will be — durable. There is
+// also no fallback (unlike push-mode's legacy-INSERT fallback): a
+// swallowed LogActivity error loses the message with only a log line.
+// Chat-history reads activity_logs (postgres_store.go:165-187); a missing
+// row = message gone on reopen. That is exactly Hongming's symptom.
+//
+// Fix (parity with push-mode): the poll-mode ingest persist of the
+// canvas user message must be SYNCHRONOUS — committed before the queued
+// 200 is returned — on a context.WithoutCancel derived context, so a
+// client disconnect on chat-exit and a post-response restart cannot lose
+// it. Behavior is never worse than today (best-effort; a persist error
+// still returns queued).
+//
+// TEST DESIGN NOTE: sqlmock.ExpectationsWereMet() hangs indefinitely if
+// the expected query never fires. We use a select+default+time.After
+// pattern so the test FAILS fast (not hangs) when the production code
+// regresses to async (the INSERT never fires before handler returns),
+// while still returning promptly when all expectations are met. The
+// insertDelay is kept small (50ms) to minimise suite-level timing
+// impact under -race detection, where mock delays are amplified by
+// the instrumenter's goroutine overhead.
+
+import (
+	"bytes"
+	"encoding/json"
+	"net/http"
+	"net/http/httptest"
+	"testing"
+	"time"
+
+	"github.com/DATA-DOG/go-sqlmock"
+	"github.com/gin-gonic/gin"
+)
+
+// TestProxyA2A_PollMode_PersistsUserMessageSynchronouslyBeforeQueuedResponse
+// is the defining contract: for a poll-mode workspace, the canvas user
+// message MUST be durably INSERTed into activity_logs BEFORE the synthetic
+// queued 200 is returned to the client — with NO reliance on a detached
+// async goroutine completing later.
+//
+// The test proves the ordering by making the INSERT block briefly and
+// asserting the handler does NOT return until the INSERT has completed.
+// Pre-fix (INSERT in h.goAsync, response returned immediately) the
+// handler returns ~instantly while the INSERT is still pending in the
+// goroutine → the elapsed time is far below the injected INSERT delay and
+// ExpectationsWereMet() is racy/unmet at return. Post-fix (synchronous
+// persist before the queued response) the handler return is gated on the
+// INSERT, so elapsed >= the injected delay and the expectation is met
+// deterministically at return WITHOUT any waitAsyncForTest()/sleep.
+func TestProxyA2A_PollMode_PersistsUserMessageSynchronouslyBeforeQueuedResponse(t *testing.T) {
+	mock := setupTestDB(t)
+	setupTestRedis(t)
+	broadcaster := newTestBroadcaster()
+	handler := NewWorkspaceHandler(broadcaster, nil, "http://localhost:8080", t.TempDir())
+
+	const wsID = "ws-poll-sync-persist"
+	// Keep delay small: -race detection amplifies mock delays significantly.
+	// A 50ms delay is sufficient to prove synchronous blocking (~50× the
+	// normal INSERT latency) without bloating the full ./... suite runtime.
+	const insertDelay = 50 * time.Millisecond
+
+	expectBudgetCheck(mock, wsID)
+
+	// lookupDeliveryMode → poll, triggering the short-circuit.
+	mock.ExpectQuery("SELECT delivery_mode FROM workspaces WHERE id").
+		WithArgs(wsID).
+		WillReturnRows(sqlmock.NewRows([]string{"delivery_mode"}).AddRow("poll"))
+
+	// workspace-name lookup inside logA2AReceiveQueued.
+	mock.ExpectQuery(`SELECT name FROM workspaces WHERE id`).
+		WithArgs(wsID).
+		WillReturnRows(sqlmock.NewRows([]string{"name"}).AddRow("Poll WS"))
+
+	// The durable user-message write. We delay it so a synchronous
+	// persist visibly gates the handler return; a detached-goroutine
+	// persist (pre-fix) does not. The fix must keep using
+	// context.WithoutCancel so this write survives a chat-exit cancel.
+	mock.ExpectExec("INSERT INTO activity_logs").
+		WillDelayFor(insertDelay).
+		WillReturnResult(sqlmock.NewResult(0, 1))
+
+	w := httptest.NewRecorder()
+	c, _ := gin.CreateTestContext(w)
+	c.Params = gin.Params{{Key: "id", Value: wsID}}
+
+	// callerID == "" (no X-Workspace-ID) → this is a canvas_user message,
+	// exactly Hongming's case.
+	body := `{"jsonrpc":"2.0","id":"poll-canvas-1","method":"message/send","params":{"message":{"role":"user","parts":[{"text":"my own message"}]}}}`
+	c.Request = httptest.NewRequest("POST", "/workspaces/"+wsID+"/a2a", bytes.NewBufferString(body))
+	c.Request.Header.Set("Content-Type", "application/json")
+
+	start := time.Now()
+	handler.ProxyA2A(c)
+	elapsed := time.Since(start)
+
+	// Defining assertion #1: the handler must not have returned the
+	// queued response before the durable INSERT committed. Pre-fix this
+	// fails (elapsed ≈ 0, INSERT still racing in goAsync).
+	if elapsed < insertDelay {
+		t.Fatalf("poll-mode queued response returned in %v, before the %v user-message INSERT — "+
+			"the message is not durable when the client/process goes away (DATA LOSS). "+
+			"Persist must be synchronous before the queued 200.", elapsed, insertDelay)
+	}
+
+	// Defining assertion #2: the durable write actually happened by the
+	// time the handler returned. ExpectionsWereMet() hangs indefinitely if
+	// the mock never fires (e.g. production code regressed to async),
+	// so we check it in a goroutine with a hard 2s timeout — fails fast
+	// (no CI hang) on regression while returning promptly on success.
+	expectDone := make(chan error, 1)
+	go func() { expectDone <- mock.ExpectationsWereMet() }()
+	select {
+	case err := <-expectDone:
+		if err != nil {
+			t.Fatalf("user-message INSERT was not durable at handler return (unmet sqlmock expectations): %v", err)
+		}
+	case <-time.After(2 * time.Second):
+		t.Fatalf("ExpectationsWereMet() hung for >2s — INSERT mock never fired. " +
+			"Likely cause: production code regressed logA2AReceiveQueued to goAsync " +
+			"(INSERT fires after handler returns, not before).")
+	}
+
+	// Sanity: still the correct poll-mode envelope + status.
+	if w.Code != http.StatusOK {
+		t.Fatalf("expected 200 (queued), 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("response is not valid JSON: %v", err)
+	}
+	if resp["status"] != "queued" || resp["delivery_mode"] != "poll" {
+		t.Errorf("poll envelope changed: got status=%v delivery_mode=%v, want queued/poll",
+			resp["status"], resp["delivery_mode"])
+	}
+}

workspace-server/internal/handlers/a2a_proxy_helpers.go

@@ -504,25 +504,49 @@ func lookupDeliveryMode(ctx context.Context, workspaceID string) string {
 // reads in PR 3 — that's how a poll-mode workspace receives inbound A2A
 // without a public URL.
 func (h *WorkspaceHandler) logA2AReceiveQueued(ctx context.Context, workspaceID, callerID string, body []byte, a2aMethod string) {
+	// DATA-LOSS FIX (internal#471 — poll-mode sibling of #1347/internal#470):
+	// this is the ONLY durable write of a poll-mode inbound message,
+	// including a canvas_user message (callerID == "") typed in the canvas
+	// chat. It MUST be SYNCHRONOUS and complete BEFORE the caller returns
+	// the synthetic {status:"queued"} 200 — otherwise the canvas sees the
+	// send acknowledged while the activity_logs row is still racing in a
+	// detached goroutine, and a workspace-server restart / deploy / OOM /
+	// EC2 hibernation between the 200 and the goroutine's commit loses the
+	// user's message permanently (chat-history reads activity_logs, so a
+	// missing row = message gone on reopen). Hongming's tenant is entirely
+	// poll-mode (4 external workspaces, no URL — verified empirically), so
+	// his reported loss is THIS path; #1347 (push-mode, persists AFTER the
+	// poll short-circuit) structurally cannot cover it.
+	//
+	// Mirrors persistUserMessageAtIngest's discipline:
+	//   - context.WithoutCancel: a client disconnect on chat-exit (which
+	//     cancels the inbound request ctx) MUST NOT abort this write.
+	//   - SYNCHRONOUS (no goAsync): the row must be durable before the
+	//     queued 200 is returned to the caller.
+	//   - Best-effort: LogActivity already logs+swallows INSERT errors, so
+	//     a hiccup never blocks or fails the user's send (behavior for
+	//     that one request is never worse than the pre-fix async path).
+	// The post-commit broadcast still fires inside LogActivity; a missed
+	// WebSocket event is not data loss (the durable row is the truth the
+	// canvas re-reads on reopen).
+	insCtx, cancel := context.WithTimeout(context.WithoutCancel(ctx), 30*time.Second)
+	defer cancel()
+
 	var wsName string
-	db.DB.QueryRowContext(ctx, `SELECT name FROM workspaces WHERE id = $1`, workspaceID).Scan(&wsName)
+	db.DB.QueryRowContext(insCtx, `SELECT name FROM workspaces WHERE id = $1`, workspaceID).Scan(&wsName)
 	if wsName == "" {
 		wsName = workspaceID
 	}
 	summary := a2aMethod + " → " + wsName + " (queued for poll)"
-	h.goAsync(func() {
-		logCtx, cancel := context.WithTimeout(context.WithoutCancel(ctx), 30*time.Second)
-		defer cancel()
-		LogActivity(logCtx, h.broadcaster, ActivityParams{
-			WorkspaceID:  workspaceID,
-			ActivityType: "a2a_receive",
-			SourceID:     nilIfEmpty(callerID),
-			TargetID:     &workspaceID,
-			Method:       &a2aMethod,
-			Summary:      &summary,
-			RequestBody:  json.RawMessage(body),
-			Status:       "ok",
-		})
+	LogActivity(insCtx, h.broadcaster, ActivityParams{
+		WorkspaceID:  workspaceID,
+		ActivityType: "a2a_receive",
+		SourceID:     nilIfEmpty(callerID),
+		TargetID:     &workspaceID,
+		Method:       &a2aMethod,
+		Summary:      &summary,
+		RequestBody:  json.RawMessage(body),
+		Status:       "ok",
 	})
 }
 

workspace-server/internal/handlers/org_helpers.go

@@ -177,7 +177,7 @@ func isEnvIdentPart(c byte) bool {
 	return isEnvIdentStart(c) || (c >= '0' && c <= '9')
 }
 
-// loadWorkspaceEnv reads the org root .env and the workspace-specific .env .env and the workspace-specific .env
+// loadWorkspaceEnv reads the org root .env and the workspace-specific .env
 // (workspace overrides org root). Used by both secret injection and channel
 // config expansion.
 //

workspace-server/internal/provisioner/provisioner.go

@@ -189,6 +189,24 @@ const containerNamePrefix = "ws-"
 // (the wiped-DB case after `docker compose down -v`).
 const LabelManaged = "molecule.platform.managed"
 
+// AgentUID / AgentGID are the uid/gid of the unprivileged `agent` user that
+// every workspace template creates and drops to via `gosu agent` before
+// exec'ing the runtime (the a2a_mcp_server runs under this uid). The value is
+// fixed at 1000:1000 across all templates — see:
+//   - workspace-configs-templates/claude-code-default/Dockerfile (`useradd -u 1000 ... agent`)
+//   - workspace-configs-templates/hermes/Dockerfile               (`useradd -u 1000 ... agent`)
+//   - workspace/entrypoint.sh                                     (`exec gosu agent` — "uid 1000")
+//
+// Files the platform injects into /configs AFTER the entrypoint's
+// `chown -R agent:agent /configs` (the post-start #418 re-injection and the
+// pre-start #1877 volume write) must be owned by this uid/gid, otherwise the
+// agent-uid MCP server hits EACCES reading /configs/.auth_token, sends an
+// empty bearer, and the platform 401s on /registry/{id}/peers (list_peers).
+const (
+	AgentUID = 1000
+	AgentGID = 1000
+)
+
 // managedLabels is the canonical label map applied to every workspace
 // container + volume. Pulled out so a future addition (e.g. instance
 // UUID for multi-platform-shared-daemon disambiguation) is one edit.
@@ -862,8 +880,18 @@ func buildTemplateTar(templatePath string) (*bytes.Buffer, error) {
 	return &buf, nil
 }
 
-// WriteFilesToContainer writes in-memory files into /configs in the container.
-func (p *Provisioner) WriteFilesToContainer(ctx context.Context, containerID string, files map[string][]byte) error {
+// buildConfigFilesTar builds the tar stream that WriteFilesToContainer streams
+// into /configs via CopyToContainer. Every entry is stamped Uid/Gid = agent
+// (AgentUID/AgentGID) so the files land agent-owned after extraction. This is
+// the issue #418 post-start re-injection path: it runs AFTER the template
+// entrypoint's `chown -R agent:agent /configs`, so without explicit ownership
+// in the tar header the files extract as root:root (tar Uid/Gid default 0) and
+// the agent-uid MCP server can no longer read /configs/.auth_token (and
+// /configs/.platform_inbound_secret) → empty bearer → list_peers 401.
+//
+// Pulled out as a pure function so the ownership contract is unit-testable
+// without a live Docker daemon (mirrors buildTemplateTar).
+func buildConfigFilesTar(files map[string][]byte) (*bytes.Buffer, error) {
 	var buf bytes.Buffer
 	tw := tar.NewWriter(&buf)
 
@@ -876,8 +904,10 @@ func (p *Provisioner) WriteFilesToContainer(ctx context.Context, containerID str
 				Typeflag: tar.TypeDir,
 				Name:     dir + "/",
 				Mode:     0755,
+				Uid:      AgentUID,
+				Gid:      AgentGID,
 			}); err != nil {
-				return fmt.Errorf("failed to write tar dir header for %s: %w", dir, err)
+				return nil, fmt.Errorf("failed to write tar dir header for %s: %w", dir, err)
 			}
 			createdDirs[dir] = true
 		}
@@ -886,19 +916,30 @@ func (p *Provisioner) WriteFilesToContainer(ctx context.Context, containerID str
 			Name: name,
 			Mode: 0644,
 			Size: int64(len(data)),
+			Uid:  AgentUID,
+			Gid:  AgentGID,
 		}
 		if err := tw.WriteHeader(header); err != nil {
-			return fmt.Errorf("failed to write tar header for %s: %w", name, err)
+			return nil, fmt.Errorf("failed to write tar header for %s: %w", name, err)
 		}
 		if _, err := tw.Write(data); err != nil {
-			return fmt.Errorf("failed to write tar data for %s: %w", name, err)
+			return nil, fmt.Errorf("failed to write tar data for %s: %w", name, err)
 		}
 	}
 	if err := tw.Close(); err != nil {
-		return fmt.Errorf("failed to close tar writer: %w", err)
+		return nil, fmt.Errorf("failed to close tar writer: %w", err)
 	}
+	return &buf, nil
+}
 
-	return p.cli.CopyToContainer(ctx, containerID, "/configs", &buf, container.CopyToContainerOptions{})
+// WriteFilesToContainer writes in-memory files into /configs in the container,
+// agent-owned (see buildConfigFilesTar).
+func (p *Provisioner) WriteFilesToContainer(ctx context.Context, containerID string, files map[string][]byte) error {
+	buf, err := buildConfigFilesTar(files)
+	if err != nil {
+		return err
+	}
+	return p.cli.CopyToContainer(ctx, containerID, "/configs", buf, container.CopyToContainerOptions{})
 }
 
 // CopyToContainer exposes CopyToContainer from the Docker client for use by other packages.
@@ -988,13 +1029,28 @@ func (p *Provisioner) ReadFromVolume(ctx context.Context, volumeName, filePath s
 	return clean, nil
 }
 
+// writeAuthTokenVolumeCmd is the shell command the throwaway alpine container
+// runs to seed /vol/.auth_token. alpine runs it as root, so without the
+// explicit `chown 1000:1000` the file stays root:root after the template
+// entrypoint's `chown -R agent:agent /configs` has already run — the agent-uid
+// (AgentUID) MCP server then gets EACCES reading it → empty bearer →
+// list_peers 401. Pulled out as a pure function so the ownership contract is
+// unit-testable without a live Docker daemon. Issue #1877.
+func writeAuthTokenVolumeCmd() string {
+	return fmt.Sprintf(
+		"mkdir -p /vol && printf '%%s' $TOKEN > /vol/.auth_token && chmod 0600 /vol/.auth_token && chown %d:%d /vol/.auth_token",
+		AgentUID, AgentGID,
+	)
+}
+
 // WriteAuthTokenToVolume writes the workspace auth token into the config volume
 // BEFORE the container starts, eliminating the token-injection race window where
 // a restarted container could read a stale token from /configs/.auth_token before
 // WriteFilesToContainer writes the new one. Issue #1877.
 //
 // Uses a throwaway alpine container to write directly to the named volume,
-// bypassing the container lifecycle entirely.
+// bypassing the container lifecycle entirely. The written file is chowned to
+// the agent uid/gid (see writeAuthTokenVolumeCmd).
 func (p *Provisioner) WriteAuthTokenToVolume(ctx context.Context, workspaceID, token string) error {
 	if p == nil || p.cli == nil {
 		return ErrNoBackend
@@ -1002,7 +1058,7 @@ func (p *Provisioner) WriteAuthTokenToVolume(ctx context.Context, workspaceID, t
 	volName := ConfigVolumeName(workspaceID)
 	resp, err := p.cli.ContainerCreate(ctx, &container.Config{
 		Image: "alpine",
-		Cmd:   []string{"sh", "-c", "mkdir -p /vol && printf '%s' $TOKEN > /vol/.auth_token && chmod 0600 /vol/.auth_token"},
+		Cmd:   []string{"sh", "-c", writeAuthTokenVolumeCmd()},
 		Env:   []string{"TOKEN=" + token},
 	}, &container.HostConfig{
 		Binds: []string{volName + ":/vol"},

workspace-server/internal/provisioner/token_ownership_test.go

@@ -0,0 +1,95 @@
+package provisioner
+
+import (
+	"archive/tar"
+	"errors"
+	"io"
+	"strings"
+	"testing"
+)
+
+// These tests pin the P0 fix for the fleet-wide list_peers 401 (Hermes and
+// every other template): the workspace-server token-injection paths wrote
+// /configs/.auth_token (and /configs/.platform_inbound_secret) as root:root
+// AFTER the template entrypoint's `chown -R agent:agent /configs` ran, so the
+// agent-uid (1000) MCP server (a2a_mcp_server, running via `gosu agent`) hit
+// `[Errno 13] Permission denied` reading the bearer → empty bearer → platform
+// 401 on /registry/{id}/peers (the literal tool_list_peers path).
+//
+// The agent uid is 1000:1000, verified from the templates:
+//   - workspace-configs-templates/claude-code-default/Dockerfile: `useradd -u 1000 ... agent`
+//   - workspace-configs-templates/hermes/Dockerfile:               `useradd -u 1000 ... agent`
+//   - workspace/entrypoint.sh / claude-code-default/entrypoint.sh:  `exec gosu agent` ("uid 1000")
+//
+// Both tests assert the real artifact (the tar headers Docker's CopyToContainer
+// honours for ownership, and the literal shell command the throwaway alpine
+// container runs), not a mock that bypasses ownership. They FAIL on pre-fix
+// code (no Uid/Gid in tar headers; no chown in the alpine command → root:root)
+// and PASS post-fix (agent-owned).
+
+// TestWriteFilesToContainerTar_FilesAreAgentOwned covers the issue #418
+// post-start re-injection path (WriteFilesToContainer): the tar it streams
+// into /configs via CopyToContainer must carry Uid/Gid = agent (1000) so the
+// extracted files land agent-readable, not root:root. This is the path that
+// (re)writes BOTH .auth_token and .platform_inbound_secret on a cadence.
+func TestWriteFilesToContainerTar_FilesAreAgentOwned(t *testing.T) {
+	files := map[string][]byte{
+		".auth_token":              []byte("tok-abc123"),
+		".platform_inbound_secret": []byte("inbound-secret-xyz"),
+		"nested/dir/file.txt":      []byte("data"),
+	}
+
+	buf, err := buildConfigFilesTar(files)
+	if err != nil {
+		t.Fatalf("buildConfigFilesTar: %v", err)
+	}
+
+	tr := tar.NewReader(buf)
+	seen := map[string]bool{}
+	for {
+		hdr, err := tr.Next()
+		if errors.Is(err, io.EOF) {
+			break
+		}
+		if err != nil {
+			t.Fatalf("read tar: %v", err)
+		}
+		if _, err := io.Copy(io.Discard, tr); err != nil {
+			t.Fatalf("drain %s: %v", hdr.Name, err)
+		}
+		seen[hdr.Name] = true
+		if hdr.Uid != AgentUID {
+			t.Fatalf("tar entry %q Uid = %d, want %d (agent) — root-owned injection causes the list_peers 401",
+				hdr.Name, hdr.Uid, AgentUID)
+		}
+		if hdr.Gid != AgentGID {
+			t.Fatalf("tar entry %q Gid = %d, want %d (agent)", hdr.Name, hdr.Gid, AgentGID)
+		}
+	}
+
+	for _, want := range []string{".auth_token", ".platform_inbound_secret"} {
+		if !seen[want] {
+			t.Fatalf("tar missing %q (seen: %v)", want, seen)
+		}
+	}
+}
+
+// TestWriteAuthTokenVolumeCmd_ChownsToAgent covers the issue #1877 pre-start
+// volume-write path (WriteAuthTokenToVolume): the throwaway alpine container
+// writes /vol/.auth_token then chmod 0600 but, pre-fix, never chowns it, so it
+// stays root:root (alpine runs the command as root). The literal command must
+// chown the file to the agent uid:gid so the agent-uid MCP server can read it.
+func TestWriteAuthTokenVolumeCmd_ChownsToAgent(t *testing.T) {
+	cmd := writeAuthTokenVolumeCmd()
+
+	if !strings.Contains(cmd, "chmod 0600 /vol/.auth_token") {
+		t.Fatalf("alpine cmd lost the 0600 chmod (regression): %q", cmd)
+	}
+
+	wantChown := "chown 1000:1000 /vol/.auth_token"
+	if !strings.Contains(cmd, wantChown) {
+		t.Fatalf("alpine cmd = %q, missing %q — without it .auth_token stays root:root "+
+			"and the agent-uid MCP server gets EACCES → empty bearer → list_peers 401",
+			cmd, wantChown)
+	}
+}

workspace/a2a_mcp_server.py

@@ -776,7 +776,23 @@ async def main():  # pragma: no cover
     buffer = b""
     while True:
         try:
-            chunk = await loop.run_in_executor(None, stdin.read, 65536)
+            # MUST be readline(), NOT read(65536). MCP is a line-delimited
+            # JSON-RPC stream where the client (openclaw bundle-mcp,
+            # Claude Code, Cursor, ...) sends one small (~150B) request
+            # and keeps stdin OPEN waiting for the response. A fixed-size
+            # `stdin.read(65536)` on a PIPE blocks until either 64KB
+            # accumulate OR EOF — neither happens during a normal MCP
+            # handshake — so the server never parses `initialize` and the
+            # client times out (~30s; openclaw: "MCP error -32000:
+            # Connection closed"). This made the stdio transport unusable
+            # for every pipe-spawned MCP host while passing tests/manual
+            # checks that fed stdin from a regular FILE (where read()
+            # returns immediately at the short file's end). readline()
+            # returns as soon as one newline-terminated line is available,
+            # which is exactly the JSON-RPC framing. Diagnosed 2026-05-15
+            # against a live openclaw workspace; see
+            # molecule-ai-workspace-runtime#61 (same fd-compat lineage).
+            chunk = await loop.run_in_executor(None, stdin.readline)
             if not chunk:
                 break
             buffer += chunk

workspace/tests/test_a2a_mcp_server.py

@@ -2097,3 +2097,124 @@ def test_peer_metadata_set_replaces_existing_entry_in_place(_reset_peer_metadata
     )
     cached = a2a_client._peer_metadata[peer]
     assert cached[1]["name"] == "v2", "re-write must update the value in place"
+
+
+class TestStdioKeepOpenPipe:
+    """Regression for the openclaw peer-visibility outage (2026-05-15).
+
+    main()'s read loop used `await loop.run_in_executor(None,
+    stdin.read, 65536)`. On a PIPE, `read(n)` blocks until n bytes
+    accumulate OR EOF. A real MCP client (openclaw bundle-mcp, Claude
+    Code, Cursor) sends ONE ~150-byte newline-delimited request and
+    keeps stdin OPEN waiting for the reply — so neither condition is
+    met, the server never parses `initialize`, and the client times
+    out (~30s; openclaw surfaced "MCP error -32000: Connection
+    closed"). Every prior stdio test fed stdin from a regular file or
+    a heredoc-pipe that CLOSES (EOF), masking the bug.
+
+    These spawn the real a2a_mcp_server.py process, write one request
+    over a pipe, and DELIBERATELY keep stdin open. With the buggy
+    read(65536) the assertion times out and fails; with readline() it
+    passes promptly. This is the literal user-facing path, not a
+    mock — see feedback_smoke_test_vendor_truth_not_shape_match.
+    """
+
+    def _spawn(self):
+        import subprocess
+        env = dict(os.environ)
+        env.setdefault("WORKSPACE_ID", "00000000-0000-0000-0000-000000000001")
+        server = os.path.join(
+            os.path.dirname(os.path.dirname(os.path.abspath(__file__))),
+            "a2a_mcp_server.py",
+        )
+        return subprocess.Popen(
+            ["python3", server],
+            stdin=subprocess.PIPE,
+            stdout=subprocess.PIPE,
+            stderr=subprocess.STDOUT,
+            env=env,
+        )
+
+    def _read_line_with_deadline(self, proc, deadline_s=15):
+        import select
+        import time
+        end = time.time() + deadline_s
+        while time.time() < end:
+            r, _, _ = select.select([proc.stdout], [], [], 1)
+            if r:
+                line = proc.stdout.readline()
+                if line:
+                    return line
+        return b""
+
+    def test_initialize_answered_on_still_open_pipe(self):
+        """One initialize, stdin kept OPEN, response required <15s.
+
+        FAILS (times out -> empty line) on stdin.read(65536).
+        PASSES on stdin.readline().
+        """
+        proc = self._spawn()
+        try:
+            req = json.dumps({
+                "jsonrpc": "2.0", "id": 1, "method": "initialize",
+                "params": {
+                    "protocolVersion": "2024-11-05",
+                    "capabilities": {},
+                    "clientInfo": {"name": "keepopen", "version": "1"},
+                },
+            }) + "\n"
+            proc.stdin.write(req.encode())
+            proc.stdin.flush()
+            # NOTE: stdin is intentionally NOT closed — mirrors a live
+            # MCP client. Closing it here would yield EOF and let the
+            # buggy read(65536) return, hiding the regression.
+
+            line = self._read_line_with_deadline(proc, 15)
+        finally:
+            proc.kill()
+            proc.wait(timeout=5)
+
+        assert line, (
+            "no response within 15s on a still-open pipe — the "
+            "stdin.read(65536) pipe-blocking regression is back "
+            "(this is the exact openclaw peer-visibility outage)"
+        )
+        resp = json.loads(line.decode())
+        assert resp.get("id") == 1, f"unexpected id: {line[:200]!r}"
+        assert "result" in resp, f"no result envelope: {line[:200]!r}"
+        assert resp["result"]["serverInfo"]["name"] == "molecule", (
+            f"wrong serverInfo: {line[:200]!r}"
+        )
+
+    def test_two_sequential_requests_on_open_pipe(self):
+        """initialize THEN tools/list on the same open pipe — proves
+        the loop keeps reading line-by-line, not just the first 64KB
+        chunk. tools/list must include list_peers (the peer-visibility
+        tool the outage was about)."""
+        proc = self._spawn()
+        try:
+            proc.stdin.write((json.dumps({
+                "jsonrpc": "2.0", "id": 1, "method": "initialize",
+                "params": {"protocolVersion": "2024-11-05",
+                           "capabilities": {},
+                           "clientInfo": {"name": "x", "version": "1"}},
+            }) + "\n").encode())
+            proc.stdin.flush()
+            init = self._read_line_with_deadline(proc, 15)
+            assert init, "initialize unanswered on open pipe"
+
+            proc.stdin.write((json.dumps({
+                "jsonrpc": "2.0", "id": 2, "method": "tools/list",
+            }) + "\n").encode())
+            proc.stdin.flush()
+            tl = self._read_line_with_deadline(proc, 15)
+        finally:
+            proc.kill()
+            proc.wait(timeout=5)
+
+        assert tl, "tools/list unanswered — loop stopped after one read"
+        resp = json.loads(tl.decode())
+        names = {t["name"] for t in resp["result"]["tools"]}
+        assert "list_peers" in names, (
+            f"list_peers missing from tools/list: {sorted(names)}"
+        )