reflect: detect Codex no-verify thrash

[EdbertChan]

Summary

Codex rollout audits now recover patch and command history from tool calls.

That lets reflect-on-thrash flag edit streaks with no verification and repeated failed command outputs.

The fixtures prove a Codex no-verify session fires, while clean Codex input stays silent.

Review Claim

Codex rollout audits expose same-problem thrash from tool-call history, and reflect-on-thrash fires on a no-verify fixture while staying silent on clean Codex input.

Review Lane

behavior

Review Unit

engine-runtime

Safety Invariant

Synthetic fixtures only; no live transcript content is committed, no Invoker repair workflow is modified, and the PR is not merged. Assumption: upstream non-interactive task supplied this invariant, so chat confirmation is unconfirmed.

Slice Rationale

This slice adds the catstack-owned catch for the reflected Codex session. Broader Invoker worker changes, log-bounding, and SQLite fixture linting are separate review units.

Non-goals

• No Invoker repair workflow changes.
• No live transcript fixtures or paths.
• No long-gate monitor, CI-log hook, or SQLite fixture lint.

Test Plan

Test Plan

• python3 -m unittest discover -s engine/skills/reflect/scripts/tests -v
• python3 -m unittest discover -s engine/hooks/reflect-on-thrash/tests -v
• python3 engine/skills/reflect/scripts/token_audit.py codex <live Codex rollout> --out /tmp/reflect-01a082c3-audit-after.json
• python3 - <<'PY' import sys sys.path.insert(0, 'engine/hooks/reflect-on-thrash') import detect print(detect.thrash_hits('<live Codex rollout>')) PY
• python3 engine/skills/reflect/scripts/top_sessions.py 10
• git diff --check
• python3 engine/skills/make-pr/scripts/preflight.py --base origin/main
• node engine/skills/draft-pr/scripts/lint-diff-atomicity.mjs --base origin/main

Revert Plan

Revert Plan

• Safe to revert? Yes
• Revert command: git revert 43732b4
• Post-revert steps: None
• Data migration? No

---

Note

Low Risk
Offline audit/heuristic changes and synthetic test fixtures only; no auth, data migration, or user-facing runtime paths beyond richer reflect flags on Codex sessions.

Overview
Codex rollout audits now reconstruct Bash and patch/Edit steps from custom_tool_call / exec payloads (parsed cmd and apply_patch paths), then apply the same same-problem thrash heuristics Claude mode already used: no-verify-edit-streak (≥3 edits without test/build-style verification or a direct run of the edited file) and recurring-failure-signatures (repeated non-zero exit tool output with numbers normalized). Those flags land in audit_codex results and --out JSON (n_recurring_failures, longest_edit_streak_no_verify).

Tests and docs add synthetic Codex JSONL helpers, unit cases for streak vs verified patch vs recurring failures, a reflect-on-thrash hook test that a calm three-patch Codex session hits no-verify-edit-streak without intervention-must-automate, and an update to cost-audit.md describing Codex --out thrash coverage.

^{Reviewed by Cursor Bugbot for commit 43732b4. Bugbot is set up for automated code reviews on this repo. Configure here.}

[cursor]

Bugbot couldn't run - usage limit reached

Bugbot is counted against Cursor usage for this user or team, and this run hit a usage or spend limit.

A user or team admin can review and increase usage limits in the Cursor dashboard.

(requestId: serverGenReqId_86897a73-945e-4a17-90ff-9363a0901236)

[mergify]

Tick the box to add this pull request to the merge queue (same as @mergifyio queue).

• Queue this pull request