Define Auto Review proof loop MVP #35

Objective

Define and implement the Auto Review MVP as a Codex-native proof loop, preserving useful Every Code behavior without porting the old background reviewer product wholesale.

Finish Line

Auto Review MVP is scoped as a Codex-native proof loop with bounded, fresh, target-applicable findings.

Current Status

State: PR #44 is open for the Auto Review proof loop MVP.

Implemented in PR #44:

• Exec-harness helpers classify Auto Review runs as current, stale, or detached from branch/head metadata.
• Summaries surface only current findings, stay quiet for clean/stale/detached runs, and enforce hard count/byte caps.
• Detail payloads are recovered by stable finding id and bounded by caller-provided byte cap.
• Run sidecars save/load by safe run id and reject path traversal.
• Tests cover target applicability, bounded summaries, bounded detail lookup, quiet clean output, restart recovery, unsafe run ids, missing run ids, numeric finding ids, and missing targets.
• README/AGENTS guidance documents the proof-loop contract and the future ContextualUserFragment requirement for model-visible context.

Validation run locally:

• python3 -m unittest discover -s tools/codex-exec-harness -p 'test_harness.py'
• python3 -m unittest discover -s tools/codex-exec-harness -p 'test_*.py'
• ruff check --diff tools/codex-exec-harness/harness.py tools/codex-exec-harness/test_harness.py
• ruff check tools/codex-exec-harness/harness.py tools/codex-exec-harness/test_harness.py
• git diff --check
• just fmt from codex-rs completed with known existing Ruff/TOML parser warnings for exclude-newer = "7 days".

Review status:

• Scout agents recommended the narrow exec-harness contract instead of premature runtime review plumbing.
• Review agents found path traversal, unbounded summary, byte-count, and edge-case test gaps; PR Define Auto Review proof loop harness contract #44 includes fixes and tests.

Next action:

• Watch PR Define Auto Review proof loop harness contract #44 CI and merge when green, then close Define Auto Review proof loop MVP #35 if the merged result satisfies the finish line.

Scope

In:

• Fresh versus stale finding classification.
• Target-applicable versus detached history handling.
• Bounded detail retrieval.
• Restart/recovery behavior.
• Quietness gate for clean builds or no-finding cases.

Out:

• No wholesale old Auto Review lifecycle/store/UI port.
• No expansion into Auto Drive.
• No unbounded model-context injection from review artifacts.

Acceptance Criteria

• Review findings are tied to current branch/head or explicitly marked stale/detached.
• Detail payloads are bounded and recoverable by stable id.
• Clean builds do not produce noisy review output.
• Integration or harness coverage proves restart recovery.
• Any context injected from review artifacts follows the model-visible context rules in AGENTS.md.

Relationships

Child of #28. Depends on classification ledger for old proof-loop evidence. Related to future agent/local LLM work if model selection becomes configurable.

Validation

Run focused Auto Review proof tests or harness scenarios once implemented, plus any changed crate tests.

Decisions

Auto Review is Rewrite unless the classification ledger proves a small old component can be safely ported.

Open Questions

Which Codex-native review/guardian surfaces should own the first proof loop?