code-review-loop

Code Review Loop (up to 3 iterations)

Input: a git base SHA (or other explicit range) to review against (for example: <sha> or HEAD~1).

Preconditions

• Confirm you are in the intended repo: git rev-parse --show-toplevel
• Ensure you have a base commit SHA (prefer the commit from before the change set).
  • If none was provided and you cannot choose confidently, ask the user for the correct base SHA.
• Start from a clean working tree (recommended):
  • git status --porcelain should be empty.
  • If it’s not empty, either commit/stash first or ask the user what to do (don’t mix unrelated changes into review fixes).
• Ensure both CLIs are available (codex and claude) because reviewer/responder must be different vendors.

Expected runtime

Typically runs 15–45 minutes, but allow at least 120 minutes.

Do not interrupt/restart the subagent if it looks stuck. The loop script owns stuck/timeout handling and will exit on its own when it completes or when --timeout is reached. Prefer watching the periodic heartbeat output (--heartbeat-seconds) instead of tailing logs.

Why Claude sometimes looked “stuck”

Claude Code print mode can be silent in --output-format text until it finishes (including during tool work). This repo defaults Claude subprocesses to --output-format stream-json --include-partial-messages so the wrapper sees measurable progress and inactivity timeouts mainly trigger on true hangs.

Loop

Run the review loop script:

resolve_skill_dir() {
  local name="$1"
  local repo_root=""
  repo_root="$(git rev-parse --show-toplevel 2>/dev/null || pwd)"

  local candidates=(
    "$repo_root/.agents/skills/$name"
    "$repo_root/.claude/skills/$name"
    "$HOME/.agents/skills/$name"
    "$HOME/.codex/skills/$name"
    "$HOME/.claude/skills/$name"
  )

  for d in "${candidates[@]}"; do
    if [[ -d "$d" ]]; then
      echo "$d"
      return 0
    fi
  done

  echo "Error: skill '$name' not found in repo-scoped or user-scoped skill dirs." >&2
  return 1
}

CODE_REVIEW_LOOP_SKILL_DIR="$(resolve_skill_dir code-review-loop)"
python3 "$CODE_REVIEW_LOOP_SKILL_DIR/scripts/run_review_loop.py" <base-sha>

Options:

• --max-iterations N (default: 3)
• --cli codex|claude (pin responder provider; reviewer is selected from provider pool and always differs from responder)
• --provider-pool codex,claude (provider pool for both subprocesses; reviewer/responder vendors are always distinct)
• --codex-model-pool ... and --claude-model-pool ... (random model selection; supports model@effort)
• --progress-log <path> and --heartbeat-seconds N
• --artifacts-dir <path> (optional; stores reviewer outputs for responders to read)
• --timeout N (per-subprocess timeout in seconds; default: 7200)
• --review-only (responder does not apply fixes/commit)

Claude automation knobs (env vars; defaults shown):

• PROMPT_TEMPLATES_CLAUDE_OUTPUT_FORMAT=stream-json (text|json|stream-json)
• PROMPT_TEMPLATES_CLAUDE_MIN_VERSION=2.1.33 (fail fast if installed Claude Code is older)
• PROMPT_TEMPLATES_CLAUDE_STREAM_LOG_MAX_BYTES=10485760 (per invocation; set 0 for unlimited)
• PROMPT_TEMPLATES_CLAUDE_INACTIVITY_TIMEOUT_SECONDS=180 (set 0 to disable; in text/json mode it is disabled unless explicitly set)
• PROMPT_TEMPLATES_CLAUDE_INACTIVITY_MIN_RUNTIME_SECONDS=30
• PROMPT_TEMPLATES_CLAUDE_PROMPT_BUDGET_BYTES=0 (disabled by default; set >0 to enforce)

Behavior:

• Each iteration runs in two fresh subprocesses:
  • Reviewer subprocess: selected from the provider pool and generates review findings from BASE_SHA..HEAD (it fetches the diff locally via git diff).
  • Responder subprocess: agrees/rejects findings and applies only agreed fixes (it fetches the diff locally via git diff and reads reviewer output from the artifacts dir).
• The script enforces that reviewer and responder always run on different vendors (codex vs claude).
• Claude subprocesses run in a native PTY (when available) to avoid non-termination/hang modes that require a TTY.
• Claude subprocesses have an inactivity timeout; on classifiable Claude automation failures the script retries once.
• If reviewer=claude fails after retry and responder is not pinned to codex, the script performs a role-swap fallback and re-runs the iteration with reviewer=codex,responder=claude (it prints a CLAUDE_FALLBACK: ... line when this happens).
• Reviewer subprocesses stop early once a valid reviewer JSON object has been emitted (Codex) or once Claude emits a terminal type=result event (stream-json).
• When the responder subprocess runs on Codex, it is invoked with --sandbox danger-full-access and -a never so it can run git commands (including committing fixes) without getting stuck on approvals or .git lock writes.
• In non---review-only mode, the script enforces commit hygiene per iteration: if responder leaves working-tree fix changes, it auto-commits them before the next iteration; if responder accepts findings but produces no committed changes, the loop stops.
• The script stops early when reviewer reports no findings, or when the responder agrees with none of the feedback.
• After completion, it emits a required REVIEW_PROGRESS log summary including:
  • total iterations completed;
  • which iteration (if any) terminated early due to no feedback or no agreement.
• Default model pools:
  • Codex: gpt-5.2@high
  • Claude: claude-opus-4-6[1m]@high

Troubleshooting:

• Look for CLAUDE_RETRY: / CLAUDE_FALLBACK: lines on stderr to confirm retry/fallback behavior.
• Heartbeats include idle_seconds=... and byte counters; PTY mode uses pty_bytes=....
• --progress-log may include full tool outputs (diffs, file contents). Treat it as sensitive; stream-json logs are truncated by default via PROMPT_TEMPLATES_CLAUDE_STREAM_LOG_MAX_BYTES.
• If you see CLAUDE_FALLBACK_UNAVAILABLE: ... responder is pinned to codex, rerun with --cli claude to allow the role-swap fallback.