step

Files (1)

• SKILL.md

SKILL.md (first lines)

# /pr-autopilot:step <PR_NUMBER>

One iteration of the PR review→fix loop. Invoked by `/loop` (Claude Code's dynamic-mode loop driver). Returns by either:

- Calling `ScheduleWakeup(delaySeconds=90, prompt="/loop /pr-autopilot:step <PR_NUMBER>", reason="polling for reviewer re-review")` to continue the loop
- Omitting the `ScheduleWakeup` call to terminate the loop

## Required tools on user machine

- `gh` (GitHub CLI, authenticated)
- `jq` (JSON processor)
- `git`

## Configuration

Read from `~/.claude/settings.json` under `prAutopilot`. Defaults if missing:

```json
{
  "prAutopilot": {
    "pollIntervalSeconds": 90,
    "fixIterationCap": 5,
    "pollTickCap": 10,
    "stallTickCap": 6,
    "reviewers": {
      "cursor":     { "enabled": true,  "login": "cursor[bot]", "scoreRegex": "(?i)score:\\s*([1-5])" },
      "copilot":    { "mode": "final-only", "login": "copilot-pull-request-reviewer[bot]" },
      "copilotSwe": { "mode": "off", "login": "copilot-swe-agent", "scoreRegex": "(?i)readiness:\\s*([1-5])" },
      "codex":      { "mode": "off", "postCommentsToPR": false },
      "claudeSelf": { "enabled": false, "rubricFile": "SELF-REVIEW-RUBRIC.md" }
    },
    "flowVerify": {
      "mode": "off",
      "catalog": "flows.md",
      "command": "npm run verify:flow",
      "resultsJson": "test-results-verify/results.json",
      "baseUrl": "http://localhost:3000"
    },
    "preCommitProfiles": {
      "default-pnpm": [{"cmd": "pnpm run typecheck"}, {"cmd": "pnpm run lint"}, {"cmd": "pnpm test"}],
      "default-npm":  [{"cmd": "npm run typecheck"},  {"cmd": "npm run lint"},  {"cmd": "npm test"}],
      "default-yarn": [{"cmd": "yarn typecheck"},     {"cmd": "yarn lint"},     {"cmd": "yarn test"}]
    },
    "autoMerge": {
      "allowedTargetBranches": ["dev"],
      "neverMergeToBranches": ["master", "main", "production"],
      "mergeMethod": "squash"
    }
  }
}
```

**Per-repo pre-commit profile (optional):** `preCommitProfiles` keys are repo basenames; without one, step 11 falls back to `default-<package-manager>` from the lockfile. Example with per-entry `if` conditions: `${CLAUDE_SKILL_DIR}/references/config.md`.

**Flow-verification config (v0.5.4):** `prAutopilot.flowVerify` gates a **behavioural** check on top of the review legs — reviewers read the diff, this one drives the app. `mode: "off"` (the default) skips it entirely, so repos without the harness are unaffected. `mode: "final-only"` runs it once, after every reviewer is green and before the auto-merge gate. It never starts a server and never merges on its own. See [`references/flow-verify.md`](references/flow-verify.md) and [`reviewers/FLOW-VERIFY-SETUP.md`](../../reviewers/FLOW-VERIFY-SETUP.md).

**Auto-merge config (v0.4):** `prAutopilot.autoMerge` tunes the *target* + *method* guards for `safeAutoMerge` (see `references/auto-merge.md`). Safe hardcoded defaults shown above — it works with zero config. The per-repo **opt-in itself is file-based** (`~/.pr-autopilot/automerge-repos`, separate from v0.3's `allowed-repos`); absent ⇒ auto-merge OFF everywhere; the notify-and-stop control flow is unchanged from v0.3 (the notification just adds a "merge manually" hint).

## Config → algorithm derivation

| Config field | `enabledForEachIter` | `enabledForFinal` | `requiresTrigger` | `postsThreads` | Score signal |
|---|---|---|---|---|---|
| `cursor.enabled=true` | yes | no | no (auto on push) | yes | `Score: N/5` regex |
| `copilot.mode=each-iter` | yes | no | yes (`requested_reviewers` API with `Copilot` — NOT `@copilot` mention which triggers SWE Agent) | yes | "0 unresolved threads" |
| `copilot.mode=final-only` | no | yes | yes (same API, one-shot at end) | yes | "0 unresolved threads" |
| `copilotSwe.mode=each-iter` | yes | no | yes (`gh pr comment "@copilot please review"`) | no — top-level conversational comments | Claude judgment on comment body |
| `copilotSwe.mode=review-score` | yes | no | yes (`@copilot please review` mention, **review-only**) | no — top-level conversational comment | `Readiness: N/5` regex (`scoreRegex`) |
| `copilotSwe.mode=final-only` | no | yes | yes (same mention, one-shot at end) | no | Claude judgment on comment body |
| `codex.mode=each-iter` | yes | no | yes (`codex exec review --base`) | no | pass/fail |
| `codex.mode=final-only` | no | yes | yes (one-shot at end) | no | pass/fail |
| `claudeSelf.enabled=true` | no (final-only by design) | yes | no (internal) | no | 1-5 vs rubric |
| `flowVerify.mode=final-only` | no (too slow for every iter) | yes | no (internal) | no | `results.json` counters, NOT the exit code |

Copilot has two products (Code Review = `copilot`, SWE Agent = `copilotSwe`) and SWE Agent has two roles (`each-iter` fixer vs `review-score` reviewer). Read `${CLAUDE_SKILL_DIR}/references/config.md` when choosing or explaining reviewer config. If Copilot Code Review does not respond within 10 poll ticks, suggest `copilotSwe` instead.

## Mode derivation

`prAutopilot.primaryFixer` (new setting, default `"auto"`):

| `primaryFixer` | Resolved mode | Notes |
|---|---|---|
| `"claude"` | X | ABORT if `copilotSwe.mode == "each-iter"` (would conflict — see `references/config.md`). Otherwise standard Mode X. |
| `"copilotSwe"` | Y | Force Mode Y; derive_mode returns "Y", pre-flight validates copilotSwe.mode == each-iter (ABORTs there if not). |
| `"auto"` (default) | derived | Inspect reviewer config — `derive_mode()` below. |

The prose form of these rules (including the `primaryFixer="claude"` + `each-iter` conflict) is in `${CLAUDE_SKILL_DIR}/references/config.md`; `derive_mode()` below is authoritative.

```python
function derive_mode(config):
  pf = config.primaryFixer  # default "auto"
  swe_each = config.reviewers.copilotSwe.mode == "each-iter"  # Mode Y fixer ONLY
  any_xreviewer = config.reviewers.cursor.enabled
                  or config.reviewers.copilot.mode == "each-iter"
                  or config.reviewers.codex.mode == "each-iter"
                  or config.reviewers.copilotSwe.mode == "review-score"  # SWE Agent as scoring reviewer (Mode X)

  if pf == "claude":
    if swe_each: return ABORT_CONFIG  # conflict — see message in pre-flight
    return "X"
  if pf == "copilotSwe":
    return "Y"  # pre-flight validates copilotSwe.mode == each-iter
  # pf == "auto"
  if swe_each and not any_xreviewer: return "Y"
  if any_xreviewer and not swe_each: return "X"
  if any_xreviewer and swe_each:     return ABORT_CONFIG  # ambiguous
  return ABORT_NO_DRIVER
```

## Pre-flight: config validation

```python
mode = derive_mode(config)
if mode == ABORT_NO_DRIVER:
  push_notification("config error", "no per-iter reviewer or fixer enabled; nothing would drive the loop")
  return  # terminate (no ScheduleWakeup)
if mode == ABORT_CONFIG
   or (mode == "X" and no reviewer in {cursor, copilot, codex, copilotSwe(review-score)} has enabledForEachIter)
   or (mode == "Y" and config.reviewers.copilotSwe.mode != "each-iter"):
  # exact config-error messages for each case: ${CLAUDE_SKILL_DIR}/references/config.md §"Pre-flight config errors"

From MarcinSufa/claude-pr-autopilot, MIT.

More from MarcinSufa/claude-pr-autopilot