AGENTS.md · diff
git:20260725.7dfa442 to git:20260911.c965a40
5 added, 21 removed. Audit A to A.
# AGENTS.md
Guidance for AI coding agents working in the **wise-claude** marketplace. This
is the repo's [AGENTS.md](https://agents.md) — a free-form project-instructions
file, not a loadable agent registry.
## Working in this repo
- This is a Claude Code plugin marketplace; the plugin lives in
`plugins/wise/`. Contributor procedures, conventions, and the workflow-engine
reference are in [`CONTRIBUTING.md`](./CONTRIBUTING.md) and
[`docs/wise/`](./docs/wise/). Read those before changing the plugin.
- The plugin is hand-edited directly — there is no build step and nothing in
the repo is generated.
- Validate before committing: `just check` (runs
- `python3 scripts/validate_repo.py` and
- `python3 -m pytest plugins/wise/tests -q`); `python3 -m py_compile
- plugins/wise/scripts/workflows.py`, `python3 -m json.tool` the JSON manifests,
+ repository validation, pytest, Ruff and mypy in the pinned Python development
+ environment); Python compilation, `python3 -m json.tool` on JSON manifests,
and `bash -n` the shell scripts also catch syntax slips.
## The wise SDLC agent roster
`wise` ships 13 SDLC role agents under
[`plugins/wise/agents/`](./plugins/wise/agents/), one markdown file per role,
catalogued in [`plugins/wise/AGENTS.md`](./plugins/wise/AGENTS.md). They are
real Claude Code plugin subagents — auto-discovered when the plugin is
installed; invoke a role as `subagent_type: wise:<name>` (e.g.
`wise:architect`). Frontmatter: `name`, `description`, `tools`, `model: inherit`,
`effort`, `color`. Plugin subagents ignore `hooks` / `mcpServers` /
`permissionMode`.
- The workflow engine dispatches `prompt` steps to them (per-step `agent:` field +
- workflow `agents:` policy — see
- [`docs/wise/workflows.md`](./docs/wise/workflows.md#agents-model-and-effort)).
-
- | Role | Default effort | Role | Default effort |
- |---|---|---|---|
- | `ceo` | high | `software-engineer` | medium |
- | `cto` | high | `qa-engineer` | medium |
- | `product-manager` | medium | `security-engineer` | high |
- | `engineering-manager` | medium | `devops-engineer` | medium |
- | `architect` | high | `sre` | high |
- | `ux-designer` | medium | `technical-writer` | low |
- | `code-reviewer` | high | | |
-
- Steps run in-conversation (subscription-covered): a step's `model:` is a real
- per-step override and `effort:` is a best-effort prompt nudge. A pinned model
- that has retired auto-falls-back to its alias with a notice. See
- [`docs/wise/workflows.md`](./docs/wise/workflows.md#agents-model-and-effort).
+ V2 workflows run `agent` steps through headless provider CLIs. Claude children
+ can adopt or delegate to these role cards. Model and effort resolution belongs
+ to the canonical Python engine; see [the workflow reference](docs/wise/workflows.md).
## Adding or editing a role
Edit the card at `plugins/wise/agents/<name>.md` directly (frontmatter +
persona prose), then update the table in
[`plugins/wise/AGENTS.md`](./plugins/wise/AGENTS.md). Full
procedure: [`CONTRIBUTING.md` §9.10](./CONTRIBUTING.md#910-the-agent-roster).