fix(agents): narrow session lock scope

CHANGELOG.md

@@ -13,6 +13,7 @@ Docs: https://docs.openclaw.ai
 
 ### Fixes
 
+- Agents/sessions: acquire the session write lock only after cold bootstrap, plugin, and tool setup so fallback runs are not blocked by stalled pre-model startup work. Thanks @codex.
 - Gateway/models: skip external OpenRouter and LiteLLM pricing refreshes for local/self-hosted model endpoints so startup does not wait on remote pricing catalogs for local-only Ollama, vLLM, and compatible providers. Thanks @codex.
 - CLI/plugins: stop security-blocked plugin installs from retrying as hook packs, so normal plugin packages report the scanner failure without a misleading "not a valid hook pack" follow-up. Fixes #61175; supersedes #64102. Thanks @KonsultDigital and @ziyincody.
 - Control UI/Dreaming: require explicit confirmation before applying restart-impacting Dreaming mode changes, with restart warning copy and loading feedback. Fixes #63804. (#63807) Thanks @bbddbb1.

src/agents/pi-embedded-runner/run/attempt.spawn-workspace.context-engine.test.ts

@@ -253,6 +253,27 @@ describe("runEmbeddedAttempt context engine sessionKey forwarding", () => {
     expectCalledWithSessionKey(afterTurn, sessionKey);
   });
 
+  it("resolves bootstrap context before acquiring the session write lock", async () => {
+    const events: string[] = [];
+    hoisted.resolveBootstrapContextForRunMock.mockImplementation(async () => {
+      events.push("bootstrap");
+      return { bootstrapFiles: [], contextFiles: [] };
+    });
+    hoisted.acquireSessionWriteLockMock.mockImplementation(async () => {
+      events.push("lock");
+      return { release: async () => {} };
+    });
+
+    await createContextEngineAttemptRunner({
+      contextEngine: createContextEngineBootstrapAndAssemble(),
+      sessionKey,
+      tempPaths,
+    });
+
+    expect(events).toEqual(expect.arrayContaining(["bootstrap", "lock"]));
+    expect(events.indexOf("bootstrap")).toBeLessThan(events.indexOf("lock"));
+  });
+
   it("forwards modelId to assemble", async () => {
     const { bootstrap, assemble } = createContextEngineBootstrapAndAssemble();
     const contextEngine = createTestContextEngine({ bootstrap, assemble });

src/agents/pi-embedded-runner/run/attempt.ts

@@ -639,16 +639,6 @@ export async function runEmbeddedAttempt(
       agentId: sessionAgentId,
     });
 
-    const sessionLock = await acquireSessionWriteLock({
-      sessionFile: params.sessionFile,
-      maxHoldMs: resolveSessionLockMaxHoldFromTimeout({
-        timeoutMs: resolveRunTimeoutWithCompactionGraceMs({
-          runTimeoutMs: params.timeoutMs,
-          compactionTimeoutMs: resolveCompactionTimeoutMs(params.config),
-        }),
-      }),
-    });
-
     const sessionLabel = params.sessionKey ?? params.sessionId;
     const contextInjectionMode = resolveContextInjectionMode(params.config);
     const agentDir = params.agentDir ?? resolveOpenClawAgentDir();
@@ -1205,6 +1195,19 @@ export async function runEmbeddedAttempt(
     let systemPromptText = systemPromptOverride();
     const userPromptPrefixText = bootstrapRouting.userPromptPrefixText;
 
+    // Keep the session lock scoped to transcript/session mutations. Cold plugin
+    // and tool setup can be slow, and holding the lock there blocks CLI fallback
+    // from taking over the same session when a gateway run stalls before model I/O.
+    const sessionLock = await acquireSessionWriteLock({
+      sessionFile: params.sessionFile,
+      maxHoldMs: resolveSessionLockMaxHoldFromTimeout({
+        timeoutMs: resolveRunTimeoutWithCompactionGraceMs({
+          runTimeoutMs: params.timeoutMs,
+          compactionTimeoutMs: resolveCompactionTimeoutMs(params.config),
+        }),
+      }),
+    });
+
     let sessionManager: ReturnType<typeof guardSessionManager> | undefined;
     let session: Awaited<ReturnType<typeof createAgentSession>>["session"] | undefined;
     let removeToolResultContextGuard: (() => void) | undefined;