Add an event-driven agent layer to the repo: GitHub Actions workflows that run Claude Code or Codex on a self-hosted runner to review PRs, run health audits, regenerate files, etc.
Recipes (prompts + config) live in .agents/recipes/ so the team can iterate on them through normal PRs.
Why
We have interactive agent tooling (skills, AGENTS.md) but nothing that runs autonomously on repo events or schedules. This closes that gap.
Rough scope
- Recipe format and shared runner context in
.agents/recipes/
- PR review workflow (on
pull_request events)
- Daily maintenance workflow with rotating suites: docs audit (docstrings, broken links), dependency audit (pinning, CVEs, upgrade safety), smoke tests (regressions)
- Runner memory (
.agents/memory/) so the agent doesn't start fresh every run - tracks known issues, deduplicates findings, detects deltas
- Tool-agnostic - start with Claude Code, add Codex later
Agent actions are additive (comments, PRs) - never autonomous merges or approvals.
Add an event-driven agent layer to the repo: GitHub Actions workflows that run Claude Code or Codex on a self-hosted runner to review PRs, run health audits, regenerate files, etc.
Recipes (prompts + config) live in
.agents/recipes/so the team can iterate on them through normal PRs.Why
We have interactive agent tooling (skills, AGENTS.md) but nothing that runs autonomously on repo events or schedules. This closes that gap.
Rough scope
.agents/recipes/pull_requestevents).agents/memory/) so the agent doesn't start fresh every run - tracks known issues, deduplicates findings, detects deltasAgent actions are additive (comments, PRs) - never autonomous merges or approvals.