greptile-gate - SKILL.md Agent Skill

name: greptile-gate description: After a PR is opened (or on demand), read Greptile's 0–5 confidence score and converge the PR toward 5/5 — grill on intent, fix review findings, re-trigger Greptile, loop up to 3 rounds, then run a final independent review. Use during /workflows:ship right after the PR is created, or when the user asks to check the Greptile score, run the Greptile gate, or read/address the Greptile review on a PR. Skips gracefully when Greptile isn't on the repo. user-invocable: true

Greptile Gate

Read the Greptile AI reviewer's verdict on an open PR, report its 0–5 confidence score, and converge the PR toward 5/5 — grilling on intent and fixing findings across up to 3 rounds, then a final independent review. Invoked by /workflows:ship right after the PR is created, and re-runnable standalone against any open PR.

The gate never merges — the developer merges manually. It converges and hands back.

Quick start

PR="$(gh pr view --json number -q .number)"
bash "${CLAUDE_PLUGIN_ROOT}/scripts/greptile-verdict.sh" --pr "$PR"

The verdict helper emits one JSON line: {"present":true,"score":3,…}, {"present":true,"score":null,…}, or {"present":false}. Always read the score from the helper — never parse the Greptile comment yourself.

Workflow

Resolve the PR. Use the PR number/url the user gave, else gh pr view --json number,url for the current branch. No PR → report and stop.
Read the verdict via greptile-verdict.sh --pr <PR>.
- present:false → "No Greptile verdict — Greptile isn't installed or hasn't reviewed yet. Skipping the gate." Exit 0. (Never block a ship.)
- present:true → report "Greptile scored this PR N/5" (or, score:null, "reviewed but posted no parseable score").
Convergence loop — repeat up to a maximum of 3 rounds, until the score is 5/5. If the current score is already 5/5, skip straight to the Final review. Otherwise run a round (human-in-the-loop):
1. Grill on intent — every round. Run the grill-with-docs skill to align on requirements and sharpen terminology; capture decisions in CONTEXT.md / ADRs. Where Greptile flagged a deliberate choice, reply to its comment explaining the intent so it learns.
2. Fix the code. Run /workflows:review; fix each finding with sequential-thinking and ultrathink; re-run /workflows:review until it returns nothing.
3. Confirm before pushing. Show the developer the changes and get approval — every round is human-in-the-loop.
4. Push the fixes to the PR branch.
5. Re-trigger Greptile. Record the trigger time, then post the re-review request:
```
TRIGGER_ISO="$(python3 -c 'import datetime; print(datetime.datetime.now(datetime.timezone.utc).isoformat().replace("+00:00","Z"))')"
gh pr comment "$PR" --body "@greptile-apps please re-review"
```
6. Wait for the fresh re-review (bounded — never hangs):
```
bash "${CLAUDE_PLUGIN_ROOT}/scripts/greptile-await.sh" --pr "$PR" --trigger "$TRIGGER_ISO"
```
  Final line is the terminal state:
  - FRESH_PASS → now 5/5 → go to Final review.
  - FRESH_FAIL → re-report the new N/5; if rounds remain, loop; else escalate (below).
  - TIMED_OUT → Greptile didn't respond within the bound → stop and hand back with context.
After 3 rounds without 5/5 → escalate: stop and hand the developer the remaining Greptile findings plus the full context of what each round tried. Do not merge.
Final review (only on 5/5). Dispatch one independent review agent to read the open PR diff (gh pr diff "$PR") with fresh eyes and report its verdict. Then stop — the gate never merges; the developer merges manually. /workflows:ship then resumes its terminal steps (Linear stays In Review, compound-learnings, audit, handbook-drift) on the converged code.

Notes

Skip-gracefully is mandatory. Greptile absent → report + exit 0; never hard-fail a ship.
@-handle. @greptile-apps is the confirmed trigger handle for this org. If it ever changes, the await fails safe (TIMED_OUT rather than hang) — but update it here, since a wrong handle posts the comment and triggers nothing.
The verdict reader keys off the Greptile author and the latest comment by timestamp; the freshness classifier treats a pre-trigger comment as stale, so an old summary never reads as a fresh re-review.
Requires authenticated gh, plus jq and python3 (the helpers check and error clearly).
grill-with-docs is a user-global skill — invoke its behavior; don't vendor a copy.