agent-qa-authoring — скилл для Claude Code: тесты и установка

---
name: agent-qa-authoring
description: Use when creating, editing, validating, or running agent-qa tests, suites, or hooks. Prefer agent-qa MCP tools, enforce canonical agent-qa IDs, and use the bundled schema reference to avoid hallucinated config keys or YAML fields.
metadata:
  short-description: Author agent-qa tests safely
---

# agent-qa Authoring

## Workflow

1. Discover the surface from the protocol itself: `tools/list` names every tool, and every caller
   sees and may call the same set. For the full picture — the capability catalog, the v1 to v2
   tool-name migration map, and this workspace's posture — run `npx agent-qa init --agent`, which
   prints one JSON document and writes nothing.
2. Read the workspace posture from that same manifest: whether this workspace has a configuration
   that reads. No MCP tool reads configuration.
3. Load `references/agent-qa-contracts.json` when you need exact schema fields or ID contracts.
4. Never mint an ID and never hand-write one. IDs are minted host-side where the entity is written:
   name what you want, then read the id back off the entity card `agent_qa_get_entity` returns.
5. Never validate an ID by hand either. The one URI codec validates every `agentqa://` reference as
   it is parsed, including the one `agent_qa_get_entity` resolves — a reference that comes back
   resolved is a reference that was well-formed.
6. Validate before you write. `agent_qa_refine_test`'s verify arm returns the same verdict that
   gates the write; there is no standalone linter over pasted text, on purpose, because a second
   validator can disagree with the one that gates.
7. Authoring changes apply in the call, and each answer names what it wrote:
   - A new test or suite file: `agent_qa_write_test`, which names the suite file the draft joins and
     writes it unless a keep-gate rejected it. Its `written` receipt names the file it wrote, or why
     none was written, with the draft's disposition and the four keep-gate results: that is how well
     the written test was verified.
     - If it answers `refused` with `statedValueRefusal`, the draft lost a value the request stated
       (a quoted string, number, price or date): show the user its `stage` and each of its
       `missingValues`, and stop. Never reword the request to get past the check; asking again in
       their own words is the user's call.
   - An edit to a test: `agent_qa_refine_test`.
   - A new hook or another configuration change: `agent_qa_change_config`, which writes the change
     and answers the files and fields it changed.
   - Each change is on disk when the call answers; nothing waits for a second call. `git status`
     and `git diff` show what changed, and `git restore` reverts a file.
8. There is no delete verb on any surface, by rule rather than by omission. Removing a test, a suite
   or a hook is a repository edit a person makes.
9. Editing an existing hook has no tool yet: it is an edit to `workspace.hooksFile` in the workspace
   until the owning milestone rules on the verb.
10. Keep file paths matched by `workspace.testMatch` or `workspace.suiteMatch`.

## Required ID Contracts

- Test IDs: `t_` + 10 id-agent words.
- Suite IDs: `s_` + 10 id-agent words.
- Hook IDs: `h_` + 10 id-agent words.
- Run IDs: `r_` + 10 id-agent words.
- Observation IDs: `obs_` + 10 id-agent words.

## Before Running

- Verify the draft first (step 6).
- Start runs with `agent_qa_start_run` rather than shelling out, and name the test files to run
  with `patterns` (paths or globs). Its input also lists `scope`, which the local agent-qa server
  refuses, so never pass it.
- A run starts in the call, with no prompt. A run that drives a target declared
  `environment: production` is refused, naming the target, unless `agent-qa.config.yaml` has
  `allowProduction: true` at the top level. No argument or flag allows it.
- `agent_qa_list_runs` enumerates recent runs; `agent_qa_get_run_digest` reports one.

## Secrets

- Only the values in the file `workspace.secretsFile` names are secrets. agent-qa replaces each one
  with `[secret:<name>]` in the text it sends to a model and in the records, logs and output it
  writes, and never restores it.
- To type a secret in a step, write `{{secret:<name>}}`. Never write `[secret:<name>]` in a step: it
  would type the placeholder, so the action is refused.
- Nothing else is masked: not `workspace.envFile` values, environment variables, cookies, saved auth
  state or emails, and not screenshots or recordings. The one exception: the dashboard's model
  connection check keeps the key or header value it is testing out of the provider's error message.
  Use temporary, limited test credentials.

## Do Not

- Do not invent config keys.
- Do not use legacy root config buckets.
- Do not hand-write IDs.
- Do not mutate files outside the configured workspace patterns.
- Do not call a v1 tool name. Every retired name and what answers its need instead is carried by the
  migration map in `npx agent-qa init --agent`; calling one returns that ruling rather than a
  result.