Mean QA

MeanQA

Find defects a happy-path pass misses. Preserve what was actually tested, including passing, failed, blocked, and interrupted work. Never confuse a test plan with a run or a screenshot with working behavior.

First action: create the proof

Every campaign gets <project-root>/.proof/<timestamp-slug>/PROOF.md, including planning-only and blocked campaigns. Before HTTP/browser/database probes or test execution, run:

sh <skill-dir>/scripts/start-proof.sh <project-root> <run-slug> <environment> <build>

Derive the project root from the app/service under test; honor a user-specified artifact location. Use unknown for metadata not yet established. Read back the returned document before any probe. The helper never overwrites an earlier run and publishes a nonempty initial document. If you are only designing, immediately mark it DESIGNED — NOT RUN; if blocked, record BLOCKED, the failed prerequisite and next action. No empty "successful" report.

If creation fails, stop probes, report the refused path and use another user-permitted persistent location if available. Never pretend the file exists. This is deterministic when invoked, not a runtime hook: skill instructions cannot guarantee capture if the agent bypasses them or is killed. An abandoned IN PROGRESS report means incomplete, never pass.

Update the report and case artifacts after each meaningful case, before the next one. Save the pre-change state before changing it. On normal completion record the actual end time, verdict, coverage and remaining gaps; on interruption, record INCOMPLETE if execution still permits a write. Do not overwrite earlier runs or silently remove evidence. Keep .proof local/private; check ignore rules before staging and never automatically commit its contents.

Load only what the task needs

• Always read findings.md: oracle, false-positive control and honest verdicts.
• Planning: read attack.md and cases.md. Mark every proposed case not run.
• Execution: read environment.md and evidence.md before probing. Add attack/cases when designing new cases, not when simply re-running existing ones.
• Report formatting: read report.md when writing the completed report.
• External publishing only: read publishing.md once publishing is requested or already authorized. A PR existing, or a request to verify a PR, does not authorize uploads, comments or body edits.

Four questions before the happy path

1. When a dependency cannot answer, what value does it return and how does the caller distinguish failure from "nothing found"? Trace error channels; do not assume the absence of an error return means success.
2. What is the harm in both directions — wrongful action and wrongful denial?
3. What invariant must always hold, and what sequence breaks it? Check conserved value, one intent/one effect, legal transitions and ownership.
4. What is irreversible — charges, disclosures, notifications, deletion? Prioritize by consequence and recoverability, not convenient tests.

Then run applicable cheap probes against in-scope, authorized test targets: malformed input, wrong types, missing/extra fields, untrusted client values, boundaries, duplicate/concurrent requests and unauthorized identities. Record skipped or unavailable probes with the coverage consequence. A mandate to test never authorizes a production mutation or another tenant.

Predict the oracle and fails-if condition before each case. Use stable case IDs, exact sanitized inputs, observed outcome and a repeatable reproduction. A surprising intentional policy is a policy question, not automatically a bug. Rank findings by demonstrated consequence; name unknown causes as unknown. Run the relevant baseline before an authorized fix and the same proof plus focused regressions afterward. QA alone does not authorize fixing.

Evidence volume without token waste

Capture as much relevant, safe, nonduplicative evidence as available: commands and exit codes, requests/responses with timings, direct persistence reads, logs/traces, browser steps, before/after screenshots, console/network, head/build identity and environment scope. Mark unavailable channels explicitly; missing telemetry is not a clean result. Synthetic fixtures first.

Keep complete sanitized captures in cases/<id>/; give every results row a matching evidence block with observation, decisive excerpt and links to complete files. Embed screenshots in the report. Do not paste large bodies repeatedly into the report or conversation. Preserve every executed case, not merely a sample. Cap oversized/continuous captures with an explicit bound and truncation notice; never claim an uncaptured remainder was preserved.

Keep tool output bounded: inspect indexed files and targeted excerpts; don't re-read all logs or full reports after each append. Chat gets findings, coverage, gaps and a proof link. More evidence on disk need not mean more model tokens. Report measured usage only if exposed by the runtime; otherwise label token estimates and do not invent a dollar cost.

Safety and finish

Never print, extract, paste or persist credentials, session cookies, auth headers or personal data. Redact before writing or displaying artifacts; screenshots and network bodies can leak secrets too. If safe capture is impossible, retain sanitized observations and mark the resulting proof gap.

An unknown environment is production. Production stays bounded/read-only under the user's applicable confirmation rules. Prove a mutation target is synthetic and the action authorized before using it. Do not switch accounts, log the user out, change global browser configuration, or publish evidence just to finish QA.

Before finishing, check each row has real, readable evidence files, all image links resolve, and the report matches what ran. Leave user tabs/sessions intact; clean up only resources created for this run. State what was not tested. A stakeholder's risk acceptance is separate from the evidence-based verdict.