hermit-agent: Command for Claude Code

/code-polish-hermit — state-machine polish loop

Unlike /code-polish, this variant is a distributed state machine. Hermit runs its own review-and-apply loop until it is self-clean, and Claude provides two independent review passes on top as a quality gate. Claude's token spend is capped at ~2 review rounds.

Arguments

$ARGUMENTS — PR number (required).

• --strategy {aggressive|moderate|conservative} — delegation depth. If omitted, read from the model-strategy map in the checkpoint config (same mechanism as /feature-develop-hermit).

Prerequisites

• Hermit gateway + MCP server running.
• Valid PR branch / worktree resolved.
• Project's test runner works (Hermit will invoke it).

State machine

HERMIT_LOOP          ← Hermit self-reviews and fixes until clean
     ↓
CC_REVIEW_1          ← Claude review pass 1
     │               ↓ findings? ←─────┐
     │                                 │
     └── clean ──→ CC_REVIEW_2         ↑
                        │              │
                        ↓              │
                     clean             │
                        ↓              │
                      DONE             │
                                       │
              findings? ──→ HERMIT_LOOP (iteration+1)

The loop terminates when two Claude review passes back to back are clean, or the global iteration limit (5) is hit, or Hermit fails three times in a row.

Workflow

Step 1 — lock

mkdir -p .hermit
cat > .hermit/active-task.lock <<EOF
{
  "task_id": "<filled-after-run_task>",
  "started_at": "$(date -Iseconds)",
  "skill": "code-polish-hermit",
  "state": "HERMIT_LOOP",
  "iteration": 0,
  "pr": "$ARGUMENTS"
}
EOF

Step 2 — state HERMIT_LOOP (initial entry)

mcp__hermit__run_task(cwd=<worktree>, background=true, model=<--model or "">)

You are Hermit. Polish PR #<PR-number> to your own "first clean" state.

Loop (max 5 internal iterations):
1. Self-review the PR diff against the project's review criteria.
   Output Pn findings.
2. If findings are empty → STOP and report `self_clean: true`.
3. Otherwise, apply every finding.
4. Re-run the review.
5. Repeat.

Rules:
- TDD: every code change updates or adds tests; confirm the test
  runner passes before closing an iteration.
- Preserve existing behaviour — never silently drop validation,
  error handling, or cleanup.
- Do NOT commit or push.
- Stay in scope: touch only files in this PR's diff (or the
  refactoring targets listed in `.omc/plans/<branch>-patterns.md`
  if it exists).
- Ambiguous fix direction → call ask_user_question.

Completion:
- self_clean: true, OR internal iteration limit reached.
- Return: final state, iterations used, files modified.