Best possible solution:

Land this PR after ordinary maintainer merge handling so Active Memory uses one setup-grace-aware timeout budget across the hook, watchdog, and embedded runner.

Do we have a high-confidence way to reproduce the issue?

Yes. Source inspection on current main shows the raw inner timeout handoff while the outer hook and watchdog use timeoutMs + setupGraceTimeoutMs; the focused unit harness has a setup-grace scenario that the PR updates to assert the combined embedded timeout.

Is this the best way to solve the issue?

Yes. Passing the resolved timeoutMs + setupGraceTimeoutMs into runEmbeddedPiAgent is the narrowest maintainable fix because it preserves existing clamps and aligns all Active Memory timeout layers without changing recall behavior.

What I checked:

• Current main raw embedded timeout: Current main still calls runEmbeddedPiAgent with timeoutMs: params.config.timeoutMs, so the embedded runner can self-timeout before the outer setup-grace-aware watchdog. (extensions/active-memory/index.ts:2289, 54a81e00807b)
• Outer budgets already include setup grace: The Active Memory watchdog uses params.config.timeoutMs + params.config.setupGraceTimeoutMs, and hook registration also uses the combined budget. (extensions/active-memory/index.ts:2436, 54a81e00807b)
• Embedded runner consumes caller timeout: The embedded runner configures its HTTP runtime and schedules the run abort timer from params.timeoutMs, making the Active Memory handoff value behaviorally significant. (src/agents/pi-embedded-runner/run/attempt.ts:709, 54a81e00807b)
• PR head fixes the handoff: The PR computes embeddedTimeoutMs from the resolved config timeout plus setup grace and passes that value to runEmbeddedPiAgent. (extensions/active-memory/index.ts:2275, 5305f9c9f7fc)
• Regression, docs, and changelog included: The PR updates the setup-grace assertion to expect the combined budget, removes the now-stale docs caveat, and adds an Unreleased Fixes changelog entry crediting the contributor. (extensions/active-memory/index.test.ts:2259, 5305f9c9f7fc)
• Head checks: GitHub check-runs for the PR head include successful check, check-prod-types, check-docs, check-lint, build-artifacts, and security fast checks; only expected skipped/cancelled non-gating automation appears in the list. (5305f9c9f7fc)

Likely related people:

• vincentkoc: Local blame attributes the current Active Memory embedded-runner call, setup-grace watchdog, and focused test block in this checkout to commit ea45950a9d8e, which introduced the current files in available history. (role: current-line history owner; confidence: medium; commits: ea45950a9d8e; files: extensions/active-memory/index.ts, extensions/active-memory/index.test.ts)
• hyspacex: Related merged PR fix(active-memory): preserve setup time outside recall timeout #72620 added the plugin-level setup grace behavior and regression coverage that this PR finishes applying across the embedded-runner boundary. (role: setup-grace feature introducer; confidence: medium; commits: 4203f288fcba; files: extensions/active-memory/index.ts, extensions/active-memory/index.test.ts)
• steipete: The latest discussion says steipete rewrote the branch as a narrow maintainer patch, and local history shows recent adjacent Active Memory maintenance in 5e9135f2e2. (role: recent maintainer and PR rewriter; confidence: high; commits: 5305f9c9f7fc, 5e9135f2e23d; files: extensions/active-memory/index.ts, extensions/active-memory/index.test.ts, CHANGELOG.md)
• RayWoo: The merged setup-grace documentation PR introduced the docs caveat that this PR removes once the embedded-runner budget is fixed. (role: recent docs maintainer; confidence: medium; commits: 153e4c59cfd5; files: docs/concepts/active-memory.md)

Codex review notes: model gpt-5.5, reasoning high; reviewed against 54a81e00807b.