lidge-jun/codexclaw / plugins/codexclaw/skills/qa
cxc-qa
VerifiedUse for manual web, TUI, CLI, API and desktop QA. Triggers: smoke test, visual QA, screenshot check, CJK clipping, 수동 QA, 실제로 되는지 확인, 동작 확인, 직접 돌려봐.
| File | Kind | Size |
|---|---|---|
| SKILL.md | skill md | 3.1 KB |
| agents/openai.yaml | config | 155 B |
| references/adversarial-oracles.md | markdown | 2.5 KB |
| references/cli-tui-qa.md | markdown | 3.2 KB |
| references/evidence-contract.md | markdown | 3.7 KB |
| references/http-api-qa.md | markdown | 2.9 KB |
| references/scope-and-binding.md | markdown | 1.3 KB |
| references/teardown-receipts.md | markdown | 392 B |
| references/visual-qa.md | markdown | 7.1 KB |
| scripts/validate-evidence.mjs | script | 21.1 KB |
references/evidence-contract.md
View on GitHub## 3. Evidence contract
Artifacts live under `<evidence root>/<sessionId>/qa/<scenario-id>/`, where the evidence
root is `.codexclaw/evidence/` in repo mode and `<durable>/evidence/` from `cxc state path` otherwise:
- `invocation.txt` — the exact command(s)/steps, copy-pasteable.
- the artifact(s) — capture, screenshot, response, transcript.
- `verdict.json` — `{ "scenario": "<id>", "criterion": "<what this proves>",
"surface": "http|cli|tui|web|gui", "verdict": "PASS|FAIL|NA",
"artifactRefs": ["<relative paths>"], "note": "<one line>",
"capturedAt": "<RFC3339>", "sourceSnapshotAt": { "kind": "resolved|unavailable",
"commitSha": "<sha>", "dirty": <bool>, "capturedAt": "<RFC3339>",
"treeHash": "<sha256, dirty only>" } }`.
On `web` and `gui` only, also `"captureChecks": { "signature": <bool>,
"nonEmpty": <bool>, "dimensionsMatch": <bool>, "composited": <bool> }` —
all four keys (QA-CAPTURE-INTEGRITY-01 in `visual-qa.md`).
A desktop row that depends on a built artifact also sets
`"desktopArtifact": true` and `"criterionIds": ["c-3"]` (its goalplan
criteria) and lists exactly one `artifact-identity.json` in `artifactRefs`.
The identity schema is DESKTOP-ARTIFACT-01 in cxc-dev-devops
[native-desktop-acceptance.md](../../dev-devops/references/native-desktop-acceptance.md).
Rules:
- Every PASS names at least one non-empty artifact in `artifactRefs`.
- `capturedAt` and `sourceSnapshotAt` are required on every surface: knowing
when evidence was captured, and against which tree, does not depend on what
kind of surface produced it (QA-EVIDENCE-FRESHNESS-01).
- `inferred` and `partial` verdicts DO NOT EXIST — a scenario either ran
against the real surface or it did not.
- A scenario that cannot run is a FAIL carrying the blocker and the missing
prerequisite, not a skip.
- `NA` is legal only when the class structurally cannot apply to the surface
(e.g. viewport class on a headless API), always with a recorded reason.
- This directory is shared with the SubagentStop receipt gate's root
(`<durable>/evidence/`, using the durable directory shown by `cxc state path`); main-session QA artifacts do not interact with
worker receipts (the gate validates only the worker's own
`EVIDENCE_RECORDED:` marker path).
- A worker that CANNOT write there was dispatched wrong, not gated wrong: read-only
lanes belong on `agent_type:"explorer"`, which the gate never touches. The retry
budget is terminal, so a mis-routed worker is released rather than trapped — but its
verdict stays unresolved and blocks goal completion until a receipt settles it.
After every scenario is done, emit the aggregate receipt. Replace `<evidence-root>` with
the resolved durable directory plus `/evidence`:
```
node plugins/codexclaw/skills/qa/scripts/validate-evidence.mjs \
"<evidence-root>/<sessionId>/qa/" --emit-receipt
```
It validates every `verdict.json`, confirms they all describe the same tree,
and writes `<evidence-root>/<sessionId>/qa-receipt.json`. Validated verdict
and identity files are hashed into typed `artifactManifest` entries
(`path`, `sha256`, `kind`, optional `criterionIds`), rechecked whenever the
receipt is read. A schemaVersion 2+ plan with a recorded final gate needs an
identity entry for each non-native desktop criterion; default v1 plans rely on
this ingress validation only. Any failure
leaves no receipt behind, including deleting one an earlier run produced —
a receipt that outlives the QA it attests to is worse than none.
Current limitation: the final gate does not pick this file up on its own. Until
the `final-gate` lifecycle CLI exists, point `finalGate.qaReceiptPath` at that
path by editing the goalplan directly. Remove this paragraph when that CLI lands.
Related skills
| Skill | Stars | Updated | Works with | Status | Install |
|---|---|---|---|---|---|
| finishing-a-development-branchobra/superpowersUse when implementation is complete, all tests pass, and you need to decide how to integrate the work | Stars: 322k | Updated | Claude CodeCodex | Verified | Install |
| systematic-debuggingobra/superpowersUse when encountering any bug, test failure, or unexpected behavior, before proposing fixes | Stars: 322k | Updated | Claude CodeCodex | Verified | Install |
| test-driven-developmentobra/superpowersUse when implementing any feature or bugfix, before writing implementation code | Stars: 322k | Updated | Claude CodeCodex | Verified | Install |
| skill-creatoranthropics/skillsCreate new skills, modify and improve existing skills, and measure skill performance. Use when users want to create a skill from scratch, edit, or optimize an existing skill, run evals to test a skill, benchmark skill performance with variance analysis, or optimize a skill's description for better triggering accuracy. | Stars: 198k | Updated | Claude CodeCodex | Verified | Install |