[jgoux]

After a stack is reopened, its E2E log subscriptions can connect before the detached owner starts and silently stop, leaving failure reports with only logs from the previous run. Attach them after the owner starts and include service health, wake status, and the last error in failure diagnostics.

Test fixtures only; runtime behavior is unchanged. Follow-up to #6684.

[github-actions]

🤖 AI Review

Codex reported no findings. Claude reported three low-severity diagnostic gaps, all confirmed by the checked-out code. They affect whole-stack test failure observability rather than production behavior: log-stream failures are discarded, stop diagnostics omit available status details, and status-query failures omit their causes.

Findings

Severity	Location	Category	Sources	Claim
🟡 MINOR	packages/stack/tests/whole-stack/fixture.ts:74	observability	claude	Log-watch fibers discard subscription failures, so a failed log subscription can leave an empty diagnostic tail without explaining why.
⚪ NIT	packages/stack/tests/whole-stack/fixture.ts:221	observability	claude	When a status RPC fails, the fallback emits only service=unknown and discards the failure cause.
⚪ NIT	packages/stack/tests/whole-stack/scenarios.ts:159	observability	claude	Stop-path diagnostics report only lifecycle from each service status, omitting health, wakeEnabled, and error details available in the same observation.

Findings outside the diff

• 🟡 MINOR packages/stack/tests/whole-stack/fixture.ts:74 — Log-watch fibers discard subscription failures, so a failed log subscription can leave an empty diagnostic tail without explaining why.
• ⚪ NIT packages/stack/tests/whole-stack/scenarios.ts:159 — Stop-path diagnostics report only lifecycle from each service status, omitting health, wakeEnabled, and error details available in the same observation.

Stats

Claude findings: 3 · Codex findings: 0 · Confirmed: 3 · Refuted: 0 · Uncertain: 0

---

Models: claude-opus-5 + gpt-5.6-sol · Trigger: auto · Workflow run

This review runs once per PR. A maintainer can request another with a /ai-review comment.

[jgoux]

All three findings are now addressed in 813b821, superseding the earlier deferrals:

• Log-stream failures report a service-specific message and their cause. Interruption-only teardown remains quiet.
• Failed status queries include their cause while retaining observations for other services.
• Stop diagnostics include lifecycle, health, wake status, and the last service error, and still propagate the original stop failure.

Formatting, generic and Effect lint, and stack type checks passed. The native default lifecycle/reopen E2E passed with no log-stream error noise during normal teardown. The earlier parallel run was interrupted before its configured timeout and is not counted as a passing validation.