AGENTS.md · git:20260905.c5eb68d · 2026-09-05 · sha256 4a9ba43d1a59d147

AGENTS.md git:20260905.c5eb68dA

Immutable. This exact content is served forever at /api/v1/blob/4a9ba43d1a59d147.

# AGENTS.md

## Repo Purpose

This repository stores reusable agent skills. Most work should be limited to
root documentation, `.agents/skills/`, and the validation scripts
under `scripts/`.

## Structure

- Skills live in `.agents/skills/<skill-name>/`
- Each skill must include `SKILL.md`
- `SKILL.md` frontmatter must contain exactly `name` and `description`; the
  declared name must match the skill folder
- Supporting files such as `references/`, `scripts/`, and `assets/` are allowed
  when they directly serve the skill
- `README.md` is the published skill inventory and routing guide; keep it
  aligned whenever a skill is added, removed, or retitled

## Skill Taxonomy

`README.md` is the source of truth for the current skill inventory, family
routing, and composition examples. Do not duplicate that catalog here.

- Start with one principle skill for language- or discipline-specific work,
  add matching domain or project overlays, and add an orthogonal workflow skill
  only when the task needs that mode
- A baseline overlay is the routine default; a canonical overlay is the
  stronger option for harder work in the same domain and should preserve the
  baseline's core expectations
- A companion overlay adds an optional workflow without replacing its baseline
- A process overlay governs workflow or enforcement; a template overlay is a
  pattern for repo-local adaptation
- A system skill builds or installs a repo-owned system instead of guiding
  ordinary execution
- Keep these relationships explicit when adding or revising skills so adjacent
  skills do not drift into overlapping jobs

## Working Rules

- Preserve the existing folder-based layout unless the task explicitly requires a structural change
- Keep instructions concise, specific, and grounded in files that actually exist
- Use the repo's role vocabulary consistently when naming or describing skills
- Update only the skill folder and root docs relevant to the task
- When adding or changing a skill, keep its trigger boundaries distinct from
  neighboring skills and update the relevant `README.md` family and defaults
- Treat `README.md` and `AGENTS.md` as durable repo docs; do not add process or automation claims that the repository does not implement
- Prefer small, targeted edits over broad repo rewrites
- Author skills, examples, and templates independently. Keep external references
  when primary documentation, specifications, research, or tool evidence supports
  a concrete procedure, constraint, or maintenance decision
- Obtain explicit user agreement before incorporating third-party material that
  requires additional licenses or notices; retain required notices for any such
  material that remains

## Validation

There is no application build or lint workflow.

Keep reusable validation procedures, fixtures, expected outcomes, and checker
code in the repository. Never commit individual skill-maintenance run reports,
transcripts, logs, run-specific output snapshots, per-run hashes, or result tables, including
under skill `references/` or `assets/` directories.

When retaining run artifacts, use the host's task-artifact storage outside the
checkout, or an OS temporary directory outside it when no task storage is available.
Report their location in the task, not in committed files. Retain concise
source/version facts that support current guidance, and turn useful failures
into minimal reusable cases instead of shipping their run history.

For changes to skills, supporting files, root documentation, or validation
behavior, run:

```console
python scripts/check_skills.py
```

When `scripts/check_skills.py` or `scripts/test_check_skills.py` changes, also
run:

```console
python scripts/test_check_skills.py
```

When the metadata parser, executable skill resources, or their tests change,
also run:

```console
python scripts/test_skill_resources.py
```

Report unavailable Bash, real CMake/CTest, or PyYAML checks as skipped, not
passed. Tool discovery and optional overrides are documented in `README.md`.

When `scripts/check-skills.sh` changes, run its wrapper on a Bash-capable
system:

```console
bash scripts/check-skills.sh
```

Before finishing, verify:

- new or edited skills still live under `.agents/skills/`
- each skill folder still has `SKILL.md`
- the applicable validation commands above pass
- root docs do not claim nonexistent commands or automation
- root docs still match the current root files and published skill inventory
- examples and references point to real files
- likely skill-name references in Markdown point to published local skills

## Safety

- Do not rename or move published skill folders without explicit instruction
- Do not add generated content unless the task requires it
- Do not invent setup, CI, packaging, or release workflows that are not present in the repo

## Scope

Use this root file for the whole repository. Nested `AGENTS.md` files are not needed unless a subtree later gets a materially different workflow.