Assurance qualité pour aventures ZIL

---

name: quality-assurance description: Coordinate independent technical, blind functional, artistic, and audience-accessibility release passes for a packaged ZIL adventure

Run four independent release perspectives and synthesize them without collapsing different kinds of evidence into one generic bug list.

Inputs

• Completed Stages 1-8 of the adventure workflow
• Packaged playable build and current tests
• Declared genre, artistic intent, and target audience

Pass Order

1. Technical release gate — @technical-tester

Invoke @technical-tester first. It may inspect source and design materials and must follow skill testing. Block later passes for crashes, startup failure, corrupted persistence, or an unreachable golden path. Lesser technical findings can proceed to the shared remediation ledger while review continues.

Expected output: <game-name>-technical-report.md plus focused RED regressions. The report must contain a complete room-exit matrix proving opposite-direction returns, including documented intentional one-way or non-Euclidean exceptions.

2. Blind functional playtest — @game-tester

Use a new agent context. The tester must follow the blindness boundary in .opencode/agents/game-tester.md; do not pass it technical findings, design rationale, source discoveries, or walkthrough knowledge before its organic findings are frozen.

Expected output: <game-name>-bugs.md plus focused RED regressions for reproducible functional bugs.

3. Artistic review — @artistic-tester

Use another fresh context and follow skill artistic-review. Preserve its first-experience pass before it reads authored intent.

Expected output: <game-name>-artistic-review.md.

4. Audience-fit and accessibility — @accessibility-tester

Use a fresh context and follow skill accessibility-testing. Supply the declared target audience, not technical or artistic verdicts. Require independent fresh saves for its personas.

Expected output: <game-name>-accessibility-review.md.

Synthesis

Create or update test/QUALITY.md with separate sections for:

• technical invariants and regressions;
• functional bugs and regressions;
• artistic findings and revision decisions;
• audience/accessibility barriers and persona evidence;
• cross-cutting findings observed independently by more than one pass;
• prioritized remediation owners and verification method.

Use these classes:

| Class | Typical evidence | Verification | |-------|------------------|--------------| | Technical defect | Failed invariant, state, test, or walkthrough | Automated regression | | Functional play defect | Exact organic command/output and expected behavior | Parser-level regression | | Artistic finding | Transcript evidence compared with stated intent | Targeted replay and editorial review | | Accessibility barrier | Persona-specific stall or loss of operability/perceivability | Repeat persona scenario; regression only for stable mechanics |

Deduplicate shared root causes but retain each perspective's evidence. For example, a missing synonym may be one source fix with both a functional regression and an accessibility impact note.

Release Decision

The QA gate is not complete until:

• critical/high technical and functional defects are fixed and GREEN;
• the golden path passes from a fresh save;
• every artistic REVISE finding has a documented revision or an explicit design decision;
• every material accessibility barrier has a mitigation, accepted limitation, or revised audience claim;
• affected specialist scenarios have been rerun after remediation.

After synthesis, invoke skill bug-fixing for code defects and the relevant authoring skill for artistic or accessibility revisions. Then run a short confirmation pass from every affected perspective.

Anti-Patterns

• One context plays organically after already reading the walkthrough or design solution.
• A single severity table mixes crashes, pacing opinions, and access barriers without type or evidence.
• Trope counting substitutes for artistic judgment.
• “The walkthrough passes” substitutes for novice usability.
• Subjective preferences become brittle text regressions.
• Accessibility is reduced to easier puzzles or more hints.