Best possible solution:

Preserve normal timeout-triggered compaction for internal timeouts while adding a targeted user-abort guard, likely based on externalAbort or abort reason, with regression tests that keep real timeout attempts marked aborted.

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

Yes for the PR-introduced regression by source inspection: ordinary timeouts call abortRun(true), which sets both aborted and timedOut, so the proposed guard skips timeout compaction. No live proof was provided for the contributor's claimed Control UI stop scenario.

Is this the best way to solve the issue?

No. The change should distinguish user-initiated aborts from internal timeout aborts instead of blocking all aborted timeout attempts from timeout recovery.

Full review comments:

• [P1] Preserve timeout compaction for internal timeouts — src/agents/pi-embedded-runner/run.ts:1560
  The runner's own timeout path calls abortRun(true), and that sets both aborted = true and timedOut = true, so this guard excludes ordinary internal LLM timeouts from timeout-compaction recovery. The test changes avoid that by removing aborted from normal timeout fixtures, but real timeout attempts keep it, so high-context timeouts will fall through instead of compacting.
  Confidence: 0.94

Overall correctness: patch is incorrect
Overall confidence: 0.94

What I checked:

• PR diff changes the compaction gate: The PR changes the timeout-compaction condition to timedOut && !aborted && ..., and removes aborted: true from the normal retry-cap timeout fixtures. (src/agents/pi-embedded-runner/run.ts:1560, 9c503bb6c01a)
• Current main compacts on timed-out high-context attempts: Current main's timeout-compaction block runs when timedOut is true and the attempt was not already in compaction or tool execution timeout. (src/agents/pi-embedded-runner/run.ts:1560, 9924fff1afc7)
• Real timeout attempts also set aborted: The internal timeout path calls abortRun(true), and abortRun(true) sets both aborted = true and timedOut = true, so !aborted excludes normal runner timeouts. (src/agents/pi-embedded-runner/run/attempt.ts:2613, 9924fff1afc7)
• Existing regression coverage codifies aborted timeout compaction: Current main has a focused test named still attempts compaction for timed-out attempts that set aborted, plus retry-cap tests where timed-out attempts include aborted: true. (src/agents/pi-embedded-runner/run.timeout-triggered-compaction.test.ts:368, 9924fff1afc7)
• History points to the timeout-compaction block and tests: Blame for the timeout-compaction condition, abort timeout flagging, and aborted-timeout regression tests all points to commit cc84606 in the available local history. (src/agents/pi-embedded-runner/run.ts:1557, cc84606946eb)
• Real behavior proof is absent: The live PR body is empty and comments contain only a ClawSweeper status placeholder, with no screenshot, recording, logs, terminal output, or linked artifact showing the after-fix Control UI stop behavior. (9c503bb6c01a)

Likely related people:

• Shakker: Available blame for the timeout-compaction block, abort timeout flagging, and the aborted-timeout tests points to the same commit authored by Shakker; local history for these paths is otherwise sparse. (role: available-history area contributor; confidence: medium; commits: cc84606946eb; files: src/agents/pi-embedded-runner/run.ts, src/agents/pi-embedded-runner/run/attempt.ts, src/agents/pi-embedded-runner/run.timeout-triggered-compaction.test.ts)

Remaining risk / open question:

• The PR has no real after-fix proof for the claimed Control UI stop flow.
• The correct guard likely needs to distinguish caller/user aborts from internal timeout aborts; using the generic aborted flag conflates both paths.

Codex review notes: model gpt-5.5, reasoning high; reviewed against 9924fff1afc7.