Skip to content

ci-sweeper.md

Latest commit

 

History

History
103 lines (81 loc) · 3.98 KB

ci-sweeper.md

File metadata and controls

103 lines (81 loc) · 3.98 KB
Raw

CI Sweeper — Cursor (Automations / Agent)

This example maps the CI Sweeper loop to Cursor while keeping week one report-only. Cursor Automations can run on schedules or GitHub events in a managed cloud sandbox. Cursor also has a local /loop skill for long-running agents, but it does not create a repository-owned, reviewable schedule manifest. Use a GitHub Action or another external scheduler when the cadence must live with the repository.

Setup

Copy the shared skills and state into Cursor's project-local paths:

mkdir -p .cursor/skills/ci-triage \
  .cursor/skills/minimal-fix \
  .cursor/skills/loop-verifier
cp starters/ci-sweeper/.claude/skills/ci-triage/SKILL.md \
  .cursor/skills/ci-triage/SKILL.md
cp templates/SKILL.md.minimal-fix \
  .cursor/skills/minimal-fix/SKILL.md
cp templates/SKILL.md.verifier \
  .cursor/skills/loop-verifier/SKILL.md
cp starters/ci-sweeper/ci-sweeper-state.md.example \
  ci-sweeper-state.md

Put always-on path denylists, attempt limits, and the no-auto-merge rule in .cursor/rules/ as well as in the Automation prompt.

Automation prompt (week one — report only)

Create a scheduled or GitHub-triggered Automation with a prompt like:

Run the ci-triage skill against the latest failing checks on the default branch.
Read ci-sweeper-state.md before classifying each failure.
Record the check URL, failing job or test, first failing commit, confidence,
and whether the failure is a regression, flake, infrastructure issue, or unknown.
Update ci-sweeper-state.md with the evidence and proposed next action.

Week one is report-only:
- do not edit source files or workflow files;
- do not create a worktree, branch, commit, pull request, or CI retry;
- do not post comments or change check status;
- do not merge or enable auto-merge.

Escalate security, release, permissions, secrets, and infrastructure failures.
Stop after reporting when the root cause is ambiguous.

Review the state file after every run. A prompt and .cursor/rules are not permission boundaries: remove write-capable MCP and PR-opening tools where the configuration allows, then use GitHub branch protection and least-privilege repository access as backstops. Cursor Automations can open PRs by default. If that capability cannot be removed, do not grant credentials that can push branches or open or merge PRs; export the state change as an artifact for a human to apply instead.

L2 fix attempts (only after tuning)

After the report-only output is consistently accurate, allow one narrowly scoped fix only for a confirmed repository regression:

  1. A human approves the failing check, allowed paths, and validation command.
  2. Create an isolated attempt with npx @cobusgreyling/loop-worktree create --run-id <run-id> --pattern ci-sweeper.
  3. Run minimal-fix in that worktree; never edit the main checkout.
  4. Run loop-verifier in a separate Agent session over the diff and test output.
  5. On rejection, mark and discard the attempt. On approval, present the diff to a human for the final decision.

Never enable auto-merge. Stop after three attempts on one root cause and escalate with the collected evidence.

Example ci-sweeper-state.md

# CI Sweeper State
Last run: 2026-07-23 06:00 UTC
Mode: report-only

## Failing checks

### validate / unit-tests
- URL: https://github.com/example/project/actions/runs/123
- First failing commit: abc1234
- Classification: likely regression (medium confidence)
- Evidence: test `parses empty input` fails consistently on two runs
- Proposed action: human review before an isolated fix attempt
- Attempts: 0 / 3

## Escalated
- deploy-production: permissions failure; no automated retry or fix

References