agent-ci-optimizer

An agent skill that makes CI faster without making it lie. It audits a repo's push → trustworthy green → merge loop, measures where the time goes, and proposes one reversible, benchmarked change at a time across seven modules: fast runners, native lint/typecheck, bigger machines and test-worker reuse, caching and blobless clones, parallel jobs, affected tasks, and pre-push hooks. It starts read-only and treats a false green as worse than a slow build.

Contents

• SKILL.md: the procedure agents follow.
• references/: discovery matrix, playbook, safety rules, measurement, examples, dated sources, config recipes.
• templates/: audit report, optimization PR, paste-ready brief for agents without the skill.
• scripts/: Python 3.10+ stdlib tools.
  • discover_repo.py inventories the stack and never opens secret files.
  • collect_gh.py and analyze_gh.py pull read-only GitHub Actions data and produce p50/p95 baselines.
  • benchmark.py runs paired timing.
  • select_affected.py is a fail-closed affected-test selector.

Install

The skill directory name must be agent-ci-optimizer. Check each target first with ls -ld <target>. If an old entry exists, rename or remove it, then link with ln -sn:

SKILL_DIR="$PWD"   # this clone
ln -sn "$SKILL_DIR" ~/.agents/skills/agent-ci-optimizer
ln -sn ../../.agents/skills/agent-ci-optimizer ~/.claude/skills/agent-ci-optimizer
ln -sn ~/.agents/skills/agent-ci-optimizer ~/.codex/skills/agent-ci-optimizer
ln -sn ~/.agents/skills/agent-ci-optimizer ~/.pi/agent/skills/agent-ci-optimizer

A symlink install serves this working tree as-is, so uncommitted edits reach every agent. To ship a copy without repo tooling, run git archive HEAD | tar -x -C <dest>.

Test

python3 -m unittest discover -s tests -v   # scripts, links, anchors, frontmatter
actionlint                                  # workflow lint

© 2026 PyModel LLC. MIT license.