This file is the repo-level entry point for agents working on shared agent protocols.
When asked to read AGENTS.md and prepare to work:
- Identify the repo root, branch, and worktree status.
- Read this file,
README.md,docs/README.md, anddocs/CURRENT.md. - For a protocol change, read the affected
SKILL.md, required references, scripts, tests, and the active plan underdocs/agent_plans/when one exists. - Report the task boundary and whether the worktree is clean before editing.
Do not load every historical plan by default. Dated plans under
docs/agent_plans/ are provenance records unless docs/CURRENT.md or the human
names one as active.
Protocol directories are executable workflow packages. Keep them focused on:
SKILL.md;scripts/;tests/;references/;- reusable templates or guard documents.
Put dated implementation plans, review threads, closeout reports, and other
worktree/PR artifacts under docs/agent_plans/, not inside protocol
directories.
For structured review gates in this repo, use
structured-review/scripts/claude_structured_review.py from this checkout. Do
not substitute user-home absolute paths, sibling checkouts, self-review, or
chat-only review when the shared protocol says the runner is available.
Keep the main checkout clean and on main and do feature work in linked
worktrees. Consumption of these protocols happens through pinned submodules
and committed skill mounts, not through this checkout's path; see
README.md (Canonical Source).
- Use
docs/README.mdfor document placement and lifecycle rules. - Use
docs/CURRENT.mdfor the current source-of-truth map. - Use
docs/backlog.ymlfor this repo's protocol-development backlog. - Use
backlog-maintenance/SKILL.mdbefore adding, updating, closing, or querying backlog items in this repo. - Do not commit local user-home absolute paths, secrets, credentials, private account details, or generated private run outputs.
When adding, renaming, moving, or retiring a protocol, update README.md,
docs/CURRENT.md, and any affected index in the same change.