Best possible solution:

Keep the store cleanup for every resumed-session FailoverError, but move recoverable fresh-session retry under the CLI runner hook lifecycle or an equivalent hook-aware seam before merge.

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

Yes. Source inspection gives a high-confidence path: current main only clears stale CLI bindings for session_expired, while the CLI execution layer can raise timeout and unknown FailoverError reasons from the same resumed-session path.

Is this the best way to solve the issue?

No. The cleanup boundary is right, but retrying timeout/unknown from attempt-execution.ts is not the narrowest maintainable fix because it runs after the CLI runner can emit a failed final hook event.

Full review comments:

• [P2] Keep recoverable retry inside the CLI runner lifecycle — src/agents/command/attempt-execution.ts:535-539
  By treating timeout and unknown as recoverable here, the outer catch retries after runCliAgent has already emitted agent_end with success: false. If the fresh retry succeeds, hook consumers observe a failed final state while the caller returns success, so the recoverable retry needs to move under the CLI runner lifecycle or an equivalent hook-aware seam.
  Confidence: 0.88

Overall correctness: patch is incorrect
Overall confidence: 0.88

What I checked:

• Current main cleanup gate: Current main only enters the stale CLI binding cleanup/retry path when err.reason === "session_expired". (src/agents/command/attempt-execution.ts:512, c3a0fb9325c9)
• Current main emits timeout/unknown FailoverError reasons: The CLI execution layer throws FailoverError with reason timeout for watchdog/overall timeouts and classifies other nonzero exits to a reason or unknown. (src/agents/cli-runner/execute.ts:569, c3a0fb9325c9)
• CLI runner hook boundary: For unhandled FailoverError reasons, runPreparedCliAgent emits agent_end with success: false before rethrowing; only session_expired has an in-runner retry path before the final success hook. (src/agents/cli-runner.ts:328, c3a0fb9325c9)
• Hook contract: The hook docs define agent_end as observing final messages, success state, and run duration, so a failed final event followed by a successful returned attempt is inconsistent. Public docs: docs/plugins/hooks.md. (docs/plugins/hooks.md:109, c3a0fb9325c9)
• PR retry location: PR head adds timeout/unknown as recoverable in attempt-execution.ts, then retries with runCliWithSession(undefined) after runCliAgent has already thrown from the runner boundary. (src/agents/command/attempt-execution.ts:535, 3d5856561e6a)
• PR proof is mocked regression coverage: PR head adds mocked runCliAgent tests for auth, timeout, and unknown failover paths; the PR body's real-behavior section reports Vitest output rather than a real CLI/backend run or redacted runtime logs. (src/agents/command/attempt-execution.cli.test.ts:226, 3d5856561e6a)

Likely related people:

• steipete: Current-main blame for the affected attempt-execution and cli-runner failover blocks points to Peter Steinberger, and GitHub commit history shows repeated recent maintenance across the affected agent/session files. (role: recent maintainer and likely follow-up owner; confidence: high; commits: a34d4ef9d9, 82c11deaa2cb, fa689295c649; files: src/agents/command/attempt-execution.ts, src/agents/cli-runner.ts, src/agents/command/session-store.ts)
• vincentkoc: Recent merged work hardened CLI runner hooks and transcript loading around the same agent_end/history surface that this PR can regress. (role: adjacent hook behavior owner; confidence: medium; commits: f40f8a60cce5, 51f9f94cc361; files: src/agents/cli-runner.reliability.test.ts)
• HFConsultant: Introduced the atomic CLI session clearing path and regression coverage that this PR relies on when clearing stale persisted bindings. (role: session-store cleanup history; confidence: medium; commits: 647f4ee8ce4f; files: src/agents/command/session-store.ts, src/agents/command/session-store.test.ts)
• wenxu007: Recently worked on FailoverError propagation through the CLI/failover path, which is adjacent to the reasons and metadata this PR changes. (role: failover error attribution history; confidence: medium; commits: 9df0ae676723; files: src/agents/cli-runner.ts, src/agents/cli-runner.reliability.test.ts)

Remaining risk / open question:

• No real CLI/backend run or redacted runtime logs currently prove stale-session recovery after the patch; the supplied proof is mocked Vitest output.

Codex review notes: model gpt-5.5, reasoning high; reviewed against c3a0fb9325c9.