cubetribe/claudecode_godmode-on/quality-gates

Quality Gates (Verification Matches the Evidence)

After @builder completes, "checks" means, per Core Rule 5:

1. Deterministic checks — always. Typecheck, lint, tests, build, run by hook. A

compiler result is a fact, not a second opinion, and it costs no context when it passes.

1. UX gate — when the sprint declared `ux_gate: auto`. @tester.
2. Security gate — on security surfaces. @security (auth code, secrets handling,

.github/workflows/).

1. Review pass — on risk or doubt. Pull the /code-review skill. No standing

agent re-reads the same diff.

Whichever of 2–4 apply run in PARALLEL:

• All APPROVED -> continue to @scribe
• Any BLOCKED -> back to @builder with merged feedback

Execution Pattern

                    @builder completes
                           │
                 ┌─────────┴──────────┐
                 ▼                    ▼
        ┌─────────────────┐   (parallel, only if applicable)
        │ Deterministic    │   ┌──────────┐ ┌───────────┐ ┌──────────────┐
        │ hook (always)    │   │ @tester  │ │ @security │ │ /code-review │
        │ typecheck/lint/  │   │ if       │ │ if        │ │ if risk or   │
        │ tests/build      │   │ ux_gate: │ │ security  │ │ doubt        │
        │ 0 bytes on pass  │   │ auto     │ │ surface   │ │              │
        └────────┬─────────┘   └────┬─────┘ └─────┬─────┘ └──────┬───────┘
                 │                  │              │               │
                 └──────────────────┴──────────────┴───────────────┘
                                     SYNC POINT
                                         │
                              Apply Decision Matrix

Decision Matrix

Every gate that ran for this sprint MUST report before the matrix applies:

Rules:

• You MUST wait for every gate that applies to this sprint before deciding.
• @scribe can ONLY be called when every gate that ran is APPROVED (or the hook passed

and no other gate applied).

• If ANY gate is BLOCKED → back to @builder with specific feedback.
• Escalation follows docs/orchestrator/META-DECISIONS.md's Tier 1 self-resolution:

max 2 attempts, same agent, same scope, before Tier 2 (orchestrator resolution).

Deterministic Hook (Always)

Runs via hook, not a subagent — a compiler or test runner result is a fact, not a second opinion. Emits 0 bytes of context on success; output only on failure. Project-detection is required: the hook must recognize what kind of project it is running in (Node/TS, Swift, Flutter, other) and pass through cleanly on unknown project types instead of blocking.

@tester Gate (when ux_gate: auto)

Runs only when the sprint frontmatter declares ux_gate: auto (see docs/templates/SPRINT_TEMPLATE.md). Default is human; skip when the write scope has no UI paths. If the playwright MCP is unreachable, the sprint falls back to human and logs it — it does not block.

Screenshot viewports: 375×667 (mobile), 768×1024 (tablet), 1920×1080 (desktop).

Minimum output: 800 characters Required sections: Summary, Screenshots Table, Console Errors, Performance Metrics, Verdict

@security Gate (on security surfaces)

Runs on auth code, secrets handling, .github/workflows/, and other security-sensitive surfaces flagged by the sprint's risk signals. Opens evidence @builder did not have (threat-modeling read of the diff), so it stays a model pass, not a hook.

Review Pass (/code-review, on risk or doubt)

Pulled on demand when risk or doubt about code judgment warrants a second read of the diff — never as a standing agent. This is the replacement for the judgment portion of the former @validator: a compiler is a fact, but naming, structure, and maintainability calls need a model pass, invoked only when it is worth the cost.

Judgment-class decisions (architecture choice between valid alternatives, design taste, malformed-request suspicion) remain mandatory human escalations regardless of how many of the gates above pass — see docs/orchestrator/QUALITY-GATES.md for the full decision matrix. Unanimous agreement among same-tier agents on a judgment question is correlated evidence, not independent evidence, and does not waive the human escalation.

Fail-Safe (Graceful Degradation)

If a gate agent crashes (MCP failure, timeout):

1. Full Report — Normal operation, all checks pass
2. Partial Report — Some checks completed, others failed/timed out
3. Failure Report — Agent crashed, structured error output

Failure report includes: error type, suggested action (retry/escalate/skip), completion percentage.

Agent Return Contract

Each agent that ran writes a full report to reports/vX.Y.Z/sprint-NN/<NN>-<agent>-report.md (canonical numbering: docs/templates/REPORT_TEMPLATES.md) if it holds Write; read-only agents return their verdict and whoever dispatched them persists it. The return message to the Orchestrator is the structured verdict only — separate from the on-disk report.

STATUS: APPROVED | BLOCKED | DONE
- finding 1 (one line max)
- finding 2
- finding 3
report: <absolute path to full report>

Important: scripts/validate-agent-output.js enforces min-length on the file, not the return message, for agents that still hold Write. This contract does not change validation behavior.

CC_GodMode — © 2025–2026 Dennis Westermann ([dennis-westermann.de](https://www.dennis-westermann.de)). Proprietary — not open source. Free for private, non-commercial use; redistribution or re-hosting outside GitHub is prohibited; attribution required. Official source: [github.com/cubetribe/ClaudeCode_GodMode-On](https://github.com/cubetribe/ClaudeCode_GodMode-On). See LICENSE.