Plan Critique Loop

Plan Critique Loop (up to 3 iterations)

Input: an existing plan doc path (for example: plan/my-feature.md).

Preconditions

• Confirm you are in the intended repo: git rev-parse --show-toplevel
• Confirm the plan doc exists: test -f <plan-doc-path>
• Ensure git can create commits before running the loop (user.name/user.email configured).
• Do not start implementation in this skill; only critique and revise the plan.
• Ensure both CLIs are available (codex and claude) because critic/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.

If you use --progress-log, do not tail -f it into the main context unless you must debug; prefer checking progress via log line counts (for example: wc -l <progress-log>) and/or the heartbeat counters.

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 critique 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
}

PLAN_CRITIQUE_LOOP_SKILL_DIR="$(resolve_skill_dir plan-critique-loop)"
python3 "$PLAN_CRITIQUE_LOOP_SKILL_DIR/scripts/run_critique_loop.py" <plan-doc-path>

Options:

• --max-iterations N (default: 3)
• --cli codex|claude (pin responder provider; critic is selected from provider pool and always differs from responder)
• --provider-pool codex,claude (provider pool for both subprocesses; critic/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 critic outputs for responders to read)
• --timeout N (per-subprocess timeout in seconds; default: 7200)

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:
  • Critic subprocess: selected from the provider pool and generates critique feedback (it reads the plan doc from disk).
  • Responder subprocess: triages critic feedback (agree/reject) and applies only agreed edits using the selected responder provider/model (it reads the plan doc from disk and reads critic output from the artifacts dir).
• The script enforces that critic and responder always run on different vendors (codex vs claude).
• Any subprocess that runs on Codex is invoked with --sandbox danger-full-access, -a never, and --search.
• Claude subprocesses run in non-interactive mode with --dangerously-skip-permissions, in a native PTY (when available), and have an inactivity timeout.
• On classifiable Claude automation failures the script retries once. If critic=claude fails after retry and responder is not pinned to codex, it performs a role-swap fallback and re-runs the iteration with critic=codex,responder=claude.
• After each iteration, the script stages and commits plan-file edits (if any) before starting the next iteration.
• The script stops early when responder reports no actionable feedback.
• No manual accept/reject input is required for the loop to revise the plan.
• After completion, it appends a deterministic changelog block to the plan doc recording:
  • total iterations executed;
  • which iteration (if any) stopped due to no feedback.
• Default model pools:
  • Codex: gpt-5.2@high
  • Claude: claude-opus-4-6[1m]@high