verifier

Verifier Skill

The verifier is the default verification path for tiny, small, and medium complexity tasks. It does NOT design test matrices, does NOT spawn investigators, does NOT write new tests beyond what unity-dev already specified. Its job is to mechanically run the verification bundle and report PASS/FAIL.

Mandatory verification contract. Even at the lightweight verifier level, the run MUST satisfy @.claude/skills/qa-validation/verification-contract.md:

• Static layer — confirm compilation is clean and no banned patterns were introduced (asmdef breakage, Burst-incompatible code in hot paths, structural changes inside query iteration, editor-only API in runtime builds). Use debug-check-compilation + console-get-logs as evidence.
• Runtime layer — execute the verification bundle from impl_result.json and capture concrete evidence (tests-run, logs, screenshots).

If a layer is impossible per contract §7, record the reason in verification_result.notes and continue. Otherwise: missing layer → status="BLOCKED". Reasoning alone is NEVER sufficient.

When You Are Spawned

• Triage classified the task as small or medium.
• unity-dev has finished and emitted workspace/impl_result.json with a filled-in verification_bundle.
• The orchestrator validated the bundle is non-empty before spawning you.

Inputs

• workspace/impl_result.json — the verification bundle, repro steps, expected state, invariants, and edge cases unity-dev wants checked.
• workspace/triage.json — for risk context.

Procedure (Fixed Six Steps)

1. Static layer (compile + architecture). Call debug-check-compilation (or REST equivalent). Read console-get-logs for editor errors/warnings. Scan the changed files for banned patterns listed in verification-contract.md §1. Record findings under static_verification on the result artifact.
2. Read the bundle. verification_bundle.repro_steps is your script.
3. Run each repro step via MCP (preferred) or by reading the changed files to confirm the expected state.
4. Check every invariant. Each unmet invariant is a regression.
5. Sample edge cases. Cover at least half of edge_cases[]. Note any not checked under notes.
6. Emit workspace/verification_result.json with both static_verification and runtime_verification populated (or notes recording the §7 exemption that lets you omit one).

If at any point a repro step cannot run (compilation broken, MCP unreachable, required asset missing): set status="BLOCKED", write the reason, and stop. Do NOT improvise tests. Compilation broken in the static layer is an automatic BLOCKED — the runtime layer cannot be trusted on a broken build.

Output

workspace/verification_result.json matching .claude/schemas/verification_result.schema.json:

{
  "status": "PASS|FAIL|BLOCKED",
  "method": "verifier",
  "regressions": ["…"],
  "edge_cases_checked": ["…"],
  "stress_results": [],
  "evidence": ["static: debug-check-compilation clean", "runtime: tests-run PASS 12/12"],
  "notes": "edge cases skipped: X, Y (reason)",
  "risk_level": "LOW|MEDIUM|HIGH",
  "fail_reason": null,
  "static_verification": {
    "compile_clean": true,
    "asmdef_clean": true,
    "banned_patterns_found": [],
    "notes": "no editor-only API leaked into runtime asmdef"
  },
  "runtime_verification": {
    "method": "EditMode",
    "tests_executed": 12,
    "tests_passed": 12,
    "edge_cases_covered": 4,
    "notes": ""
  }
}

Risk level rubric:

• LOW — all invariants pass, all edge cases sampled, no regressions
• MEDIUM — invariants pass, some edge cases skipped, no regressions
• HIGH — any regression OR fewer than half edge cases sampled

What You DO NOT Do

• Design new test matrices (that is the tester agent's job, spawned only at large/critical complexity)
• Edit code (Phase 3 forbids writes; see .claude/docs/rules/mcp-phase-gates.md)
• Spawn bug-investigation (escalate via fail_reason instead)
• Wait indefinitely for a fix — write BLOCKED and return

Escalation

If verification FAILs:

1. Write verification_result.json with status="FAIL" and a precise fail_reason (which invariant + how it failed + which file).
2. Return. The orchestrator routes back to unity-dev for one retry.
3. After two FAILs on the same task: orchestrator escalates to tester (or human, per the retry policy in team.md).

Validation

Before declaring complete:

python .claude/scripts/orchestrate.py validate workspace/verification_result.json verification_result

If validation fails, fix the artifact and re-validate. The orchestrator's finalize gate will block any artifact that fails schema validation.