Agent-Skills.md

Agent Skills: PR Guardian

>-

UncategorizedID: s-hiraoku/synapse-a2a/pr-guardian

Repository

s-hiraoku/synapse-a2a
s-hiraoku

Install this agent skill to your local

pnpm dlx add-skill https://github.com/s-hiraoku/synapse-a2a/tree/HEAD/.agents/skills/pr-guardian

Skill Files

Browse the full folder contents for pr-guardian.

Download Skill

Loading file tree…

.agents/skills/pr-guardian/SKILL.md

Skill Metadata

Name
pr-guardian
Description
>-

PR Guardian

Watch a PR, detect problems, fix them, repeat — until everything is green.

Why this exists

After pushing a PR you typically wait for CI, check for merge conflicts, and respond to CodeRabbit review comments. Each issue requires a separate manual step. PR Guardian automates the entire feedback loop: it polls the PR status at a fixed interval, dispatches the appropriate fix skill when a problem is found, and exits only when all checks pass (or the iteration limit is reached).

Quick start

/pr-guardian                     # Default: 5 min interval, 20 max iterations
/pr-guardian --interval 3        # Poll every 3 minutes
/pr-guardian --max-iterations 10 # Give up after 10 cycles

Workflow

Step 0: Validate preconditions

Before starting the loop, verify:

  1. gh CLI is installed and authenticated
  2. Current branch has an open PR (gh pr view)
  3. Display the PR number and URL

If any check fails, report the error and stop.

Step 1: Poll PR status

Run the bundled status script:

bash <skill-path>/scripts/poll_pr_status.sh

This returns a JSON object with these fields:

| Field | Type | Meaning | |---------------------|---------|-----------------------------------------------| | branch | string | Current branch name | | pr_number | int | PR number | | checks_total | int | Total CI checks (excludes CodeRabbit) | | checks_passed | int | CI checks that passed | | checks_failed | int | CI checks that failed | | checks_running | int | CI checks still in progress | | has_conflict | bool | Merge conflicts detected | | coderabbit_check | string | CodeRabbit check status: none/pending/pass/fail | | coderabbit_comments | int | Number of CodeRabbit inline comments | | all_green | bool | True when everything passes (including CodeRabbit) |

Important: all_green is only true when ALL of these hold:

  • No merge conflicts
  • No CI failures
  • No CI checks still running
  • CodeRabbit check is NOT pending AND NOT none (review must have completed at least once)
  • No unresolved CodeRabbit inline comments (comments marked ✅ Addressed are excluded)
  • At least one CI check exists

Step 2: Evaluate and report

Print a concise status line each cycle:

[PR Guardian] Cycle 3/20 — PR #42
  CI: 4 passed, 1 failed, 0 running
  Conflicts: none
  CodeRabbit: review completed, 2 inline comments
  → Action: fixing CodeRabbit comments

Step 3: Dispatch fixes (priority order)

When issues are found, fix them one category at a time in this order. Only fix one category per cycle — after fixing, go back to Step 1 to re-poll, because a fix may have changed the state (e.g., a push triggers new CI runs).

Priority 1 — Merge conflicts

Merge conflicts block everything else. When has_conflict is true:

  1. Invoke /fix-conflict
  2. After resolution, the push triggers new CI runs — go back to Step 1

Priority 2 — CI failures

When checks_failed > 0 and checks_running == 0 (wait for running checks to finish before diagnosing failures):

  1. Invoke /fix-ci
  2. After the fix commit is pushed, go back to Step 1

Priority 3 — CodeRabbit review comments

When coderabbit_check == "pass" and coderabbit_comments > 0 and CI is passing:

  1. Invoke /fix-review
  2. /fix-review will verify each finding against current code before fixing
  3. After the fix commit is pushed, go back to Step 1

IMPORTANT: Only dispatch /fix-review when coderabbit_check is "pass" (review completed). If it is "pending", wait — the review is still in progress and comments may not be final yet.

Step 4: Handle pending states

The correct action is to wait (not fix) when:

  • checks_running > 0: CI is still running — do not diagnose yet
  • coderabbit_check == "pending": CodeRabbit review is in progress — wait for it to finish before looking at comments
  • coderabbit_check == "none": CodeRabbit hasn't started yet (common right after a push) — wait at least one cycle for it to appear
  • checks_total == 0 and it's the first cycle: checks haven't started yet — wait for GitHub Actions to pick up the push

Sleep for the poll interval and re-check. Do not invoke any fix skill while checks are still running or reviews are pending — you would be fixing based on incomplete information.

Step 5: Check exit conditions

After each cycle, check:

  1. All green (all_green == true): Print success message and exit
  2. Max iterations reached: Print summary of remaining issues and exit
  3. Same failure repeated 3 times: The automatic fix is not working. Print what failed and suggest manual intervention, then exit

If none of these conditions are met, sleep for the configured interval and go back to Step 1.

Step 6: Sleep between cycles

sleep <interval_minutes * 60>

The default interval is 5 minutes (300 seconds). This gives CI enough time to run after a fix push before the next poll.

Configuration

| Flag | Default | Description | |--------------------|---------|------------------------------------| | --interval | 5 | Minutes between polls | | --max-iterations | 20 | Max cycles before giving up |

Exit messages

Success:

[PR Guardian] All checks green! PR #42 is ready.
  CI: 6/6 passed | Conflicts: none | CodeRabbit: pass, 0 comments
  Completed in 3 cycles over 15 minutes.

Max iterations:

[PR Guardian] Reached 20 cycles without all-green.
  Remaining issues:
    - CI: tests failing (pytest assertion error in test_auth.py)
    - CodeRabbit: 2 unresolved inline comments
  Manual intervention needed.

Repeated failure:

[PR Guardian] Same CI failure persisted after 3 fix attempts.
  Failing check: tests
  Last error: AssertionError in test_auth.py:42
  Automatic repair is not resolving this — please fix manually.

Important notes

  • This skill composes /fix-conflict, /fix-ci, and /fix-review. It does not reimplement their logic — it orchestrates them.
  • Only one fix category is attempted per cycle to avoid cascading changes that confuse the state.
  • The skill respects both CI timing AND CodeRabbit timing: it waits for running checks AND pending reviews to finish before attempting fixes.
  • CodeRabbit check status (pending/pass/fail) is tracked separately from inline comment count. A pending check means the review is still running — do NOT dispatch /fix-review yet.
  • Each fix skill handles its own commit and push. PR Guardian only handles the polling loop and dispatch.
  • /fix-review verifies each CodeRabbit finding against the current code before applying fixes — it does not blindly apply suggestions.