qestack
Explains the qestack gated workflow that proves AI-written tests catch bugs, and routes to the right qe-* skill. Use when the user names qestack or a qe-* skill, or asks which qestack step comes next in a repository with .qestack/.
- License 许可
- Apache-2.0
- 最近更新
- 2026-11-14
- Stars
- ★ 1
- Forks
- 0
01 前言 Front Matter
- name
- qestack
- license
- Apache-2.0
- description
- Explains the qestack gated workflow that proves AI-written tests catch bugs, and routes to the right qe-* skill. Use when the user names qestack or a qe-* skill, or asks which qestack step comes next in a repository with .qestack/.
- compatibility
- Needs the qestack engine (Python 3.11+), git and a shell. Full guard only in Claude Code.
03 SKILL.md 正文
qestack
qestack proves that AI-written tests catch bugs. It does not trust "all green" or coverage. It uses evidence: traceability to requirements, red-before-green, test locks, mutation score, hidden faults and an independent judge.
This skill and the qe-* skills are the gated workflow. They need the qestack engine. The other skills of the plugin are the QA library. They need no engine. For QA work outside the gated workflow, use qa-navigator.
Preflight (every qestack skill)
- Run
qestack --version. If the command is not found, tell the user to install the engine and stop:pipx install "git+https://github.com/quark-sri/qestack#subdirectory=engine" - The engine version must have the same major and minor number as this plugin (0.1.x). If not, tell the user to install the matching engine and stop.
- Check that
.qestack/project.tomlexists. If not, use theqe-setupskill first. - Find the feature: the folder name under
specs/, for example003-order-discount. If the user did not give it and the branch name does not show it, ask the user. Pass it as--featureto everyqestackcommand and to every agent.
How to start a qestack agent
Use the Agent tool. The plugin agents have the prefix qestack:, for example subagent_type: qestack:qe-test-writer. Put these lines at the start of the prompt. Leave out the lines that do not apply.
Feature: <feature folder name>
Mode: <mode of the agent>
Phase: prove | harden (qe-test-writer in improve mode only)
Iteration: <n> (qe-test-judge in test-review mode only)
Feedback: <path> (qe-test-writer in improve mode only)
If the host cannot start qestack: agents
Some hosts cannot start the plugin agents, for example an agent other than Claude Code, or Claude Code without the plugin. Then do not do the role yourself. A new session has no memory of this session, so the role stays independent.
- Stop at the step that starts the agent.
- Find the agent text:
references/agents/<agent>.mdof this skill, for examplereferences/agents/qe-test-judge.md. If that folder does not exist, use the agent file of the plugin,plugins/qestack/agents/<agent>.md. - Tell the user to start a new session of their agent in the same repository. Give the user the text to paste into it: "Act as the agent in
<path of the agent text>. Follow it exactly." and the start lines above. - Wait until the user says that the role is done. Then continue with the next step. Read only the files that the role writes, for example
plan-review.json. Do not read the fault details. - Use one new session for each role and each start. Never use the session of one role for another role.
Never do the writer, the judge or the fault seeder role in this session. This session saw the plan, the tests, the code or the faults. The designer reads only the spec. You can do the designer role in this session only if this session did not read the implementation. Then say "Independent: no" in the report.
Run the long commands in the background
qestack gate, qestack flaky, qestack mutation and qestack seed can run longer than the time limit of one shell command. Run them with run_in_background: true and wait for the completion notice. Do not poll. If a run stops early, run it again: the engine reverts a fault that a stopped seed run left in the code.
Choose the skill
| The user wants to | Skill | When |
|---|---|---|
| Set up qestack in a repository | qe-setup |
Once per repository, or when the stack changes |
| Check that a spec is testable | qe-grill-spec |
Before test design |
| Design tests from a spec | qe-plan |
Before code exists |
| Write tests from the plan | qe-write |
After qe-plan, before implementation |
| Prove the tests fail first, then lock them | qe-prove |
After qe-write, before implementation |
| Improve the tests until they meet the bar | qe-harden |
After implementation |
| Get an independent verdict on one feature | qe-judge |
Any time |
| Measure an existing test suite | qe-audit |
Any time, read-only |
| Check the judge against human verdicts | qe-calibrate |
Each sprint |
With the plugin, the full skill names have the prefix qestack:, for example /qestack:qe-plan.
Standard flow for a new feature
qe-grill-spec -> qe-plan -> qe-write -> qe-prove -> (implement the feature) -> qe-harden
With speckit: qestack init --speckit adds /speckit.testplan, /speckit.testwrite and /speckit.testquality. They run the same skills.
Rules for all qestack work
- Read the guardrails:
qestack show guardrails. - Report numbers from the engine. Do not give opinions without evidence.
- Never weaken a test to pass a gate. Never change thresholds or protected files.
- Do not edit qestack result files (gate JSON files,
feedback-<n>.md,quality-report.md,faults/). Run the command again. The hook blocks these edits. - Without the hook (any host other than Claude Code with the plugin or
qestack init --hook), nothing blocks these edits in the session. Then the CI guard is the authority: runqestack guard --feature <feature>before each commit, and in CI. - A correct test that fails shows a product bug. Report it as a finding.
04 同仓库更多 Skills
unit-testing
Writes, fixes or improves unit tests for rules, calculations, validation and state logic: arrange-act-assert, exact oracles, boundary sets, doubles only at boundaries, controlled clock and randomness. Use when the user says 'write unit tests', 'TDD this' or 'mock this dependency'. Not for HTTP endpoints (use api-testing).
ui-automation
Writes, fixes or structures browser end-to-end tests for critical journeys: role, label or test ID locators, auto-waiting assertions, API data setup, page objects, auth reuse and traces. Use when the user says 'UI test', 'end-to-end test' or 'our UI tests are brittle'. Not for business rules (use unit-testing).
testcase-reviewer
Reviews automated tests or manual test cases like a senior test architect: scores fidelity, oracles, assertion depth, edge cases and determinism, and finds tests that pass on wrong code. Use when the user says 'review my tests', 'test smells' or 'review AI-written tests'. Not for fault injection (use mutation-probe).
testability-review
Scores how testable a feature, module or design is before tests exist (seams for clock, randomness and network, hidden state, observability) and gives the smallest production change per blocker. Use when the user asks 'is this testable' or 'why is this hard to test'. Not for writing the tests (use unit-testing).
test-strategy
Writes a test strategy for a product, system or team: quality goals, test levels and their share, environments, data, automation scope, non-functional testing, metrics and a roadmap. Use when the user says 'test strategy' or 'QA strategy'. Not for one sprint or release plan (use test-planning).
test-runner
Finds how a project runs its tests, runs the right subset, and triages each failure as product bug, test defect, environment problem or flaky test, with evidence. Use when the user says 'run the tests' or asks 'which tests fail and why'. Not for pipeline design (use ci-cd-integration).