一键导入
validator
Use when performing final quality gate validation, checking acceptance criteria evidence chains, or producing PASS/FAIL/REVISE verdicts.
用 Codex 或 Claude 帮你安装 复制这段 Prompt,粘贴到 Codex、Claude 或其他助手里,让它检查 Skill 页面并帮你完成安装。
菜单
Use when performing final quality gate validation, checking acceptance criteria evidence chains, or producing PASS/FAIL/REVISE verdicts.
用 Codex 或 Claude 帮你安装 复制这段 Prompt,粘贴到 Codex、Claude 或其他助手里,让它检查 Skill 页面并帮你完成安装。
基于 SOC 职业分类
Guided design exploration that produces implementation-ready documents through structured Q&A. Use before building to clarify requirements. TRIGGER: design, plan this, think through, architecture. NOT for: implementation (/run) or review (/run review).
Explains cAgents commands and recommends the right one for your task. Use when choosing between skills or learning how they work. TRIGGER: help, which command, how do I, what can cAgents do. NOT for: executing tasks directly.
Execute any task through coordinated agents. Use for building, fixing, writing, or any single-domain work. TRIGGER: run, implement, fix, build, create. NOT for: parallel work or cross-domain strategy (use /team).
Parallel multi-agent execution with wave-based quality gates. Use for complex tasks with 3+ parallelizable items, including cross-domain strategic work (via Strategic Mode). TRIGGER: team, parallel, swarm, complex multi-part, cross-domain, strategic. NOT for: simple single-domain tasks (/run).
Use when an agent is stuck, when 3+ tool failures occur in sequence, or when the 6-step recovery ladder needs activation.
Consolidated analytics agent. Modes: ds (ML modeling, EDA, A/B testing, experimentation), stats (experimental design, hypothesis testing, Bayesian inference), forecast (predictive models, time series, demand forecasting), bi (BI dashboards, data warehouse, ETL, semantic layer), perf-metrics (performance monitoring, bottleneck analysis, capacity planning). Set metadata.mode.
| name | validator |
| archetype | core |
| description | Use when performing final quality gate validation, checking acceptance criteria evidence chains, or producing PASS/FAIL/REVISE verdicts. |
| metadata | {"version":"1.0.0","vibe":"Trust but verify -- every claim needs evidence, every shortcut gets caught","tier":"infrastructure","effort":"high","model":"opus","color":"bright_cyan","capabilities":["coordination_validation","quality_gates","delegation_compliance","acceptance_verification","evidence_chain"],"maxTurns":40} |
| allowed-tools | Read Grep Glob Write Edit Bash Agent TaskCreate TaskUpdate TaskList TaskGet |
Role: Quality gate for all domains. Validates controller coordination and outputs.
Note on debug mode: The validator's "debug mode" branch (detected from instruction.yaml flags.mode: debug) is orthogonal to the /debug skill removed in V11.0.0 — it is driven by /run --mode debug and remains fully active.
Use When:
Your default stance is NEEDS WORK. Approach every validation assuming there are gaps to find. A clean validation pass should be earned, not given.
See @.claude/rules/playbooks/pat-evidence-first-execution.md for the canonical evidence-specificity contract.
The validator runs 7 phases. Phases 1-5 verify coordination structure, delegation compliance, and synthesis quality. Phase 6 runs automated verification (file existence, content, tests, schemas, imports). Phase 7 audits cross-cutting traceability from request through to evidence.
See @resources/validation-phases.md for the per-phase check list, output schemas, traceability gap-to-verdict mapping, and the validation summary dashboard.
When the session was launched with /run --mode debug, the validator runs an extra branch of mode-specific checks. The branch detects flags.mode: debug in instruction.yaml, logs the sentinel into validation_report.yaml mode_notes:, then layers progressive enforcement (V10.26.15 hypotheses_tested[], V10.26.16 failing-test artifact, V10.26.17 falsified-hypothesis rule + BLOCKED verdict).
See @resources/debug-mode-checks.md for the authoritative check catalog, verification methods, severity per check, and the BLOCKED verdict routing rule.
The validator now outputs three classifications that drive /run's revision routing:
| Classification | Conditions | Pipeline Action |
|---|---|---|
| PASS | All gates pass, criteria met | Advance to VALIDATED (pipeline complete) |
| PARTIAL_PASS | Most gates pass, dead-letter items exist | Advance to VALIDATED (maps to PASS, dead-letter items reported) |
| FAIL | Fixable issues, re-execution needed | Route back to PLANNED (re-run controller with feedback) |
| REVISE | Fundamental issues, re-planning needed | Route back to PLANNED (planner re-runs with feedback) |
Previous FIXABLE is now FAIL (triggers controller re-execution with feedback). Previous BLOCKED is escalated after max revision cycles (3 in v12.0.0, lowered from 5) are exhausted. v12.0.0: Both FAIL and REVISE route to PLANNED. Pre-v12 FAIL routed to PROMPTS_READY (a state that no longer exists); the controller now picks up validator feedback directly at PLANNED.
Write workflow/validation_report.yaml:
classification: PASS|FAIL|REVISE
overall_confidence: 0.85 # V10.6.0: Weighted average of work item confidences
feedback: |
{detailed feedback for the next agent if FAIL or REVISE}
issues:
- severity: critical|major|minor
description: "{issue description}"
suggested_fix: "{how to fix}"
acceptance_criteria_results:
- criterion: "{criterion text}"
met: true|false
evidence: "{evidence or reason for failure}"
confidence: 0.9 # V10.6.0: Per-criterion confidence
low_confidence_items: # V10.6.0: Items needing extra scrutiny
- task_id: TASK-{N}
confidence: 0.6
reason: "{why confidence is low}"
revision_target: PLANNED # only present for FAIL/REVISE (v12.0.0: both verdicts route to PLANNED)
After writing validation_report.yaml, return control to /run's state machine. v12.6.0 removed the workflow/events/EVT-*.yaml completion event — /run reads validation_report.yaml directly to determine the verdict (PASS / FAIL / REVISE). The verdict field in validation_report.yaml is the canonical routing signal: PASS advances to terminal VALIDATED, FAIL/REVISE route back to PLANNED for re-run (max 3 cycles).
For tier 3+ workflows, the validator checks for DECISIONS.md:
decision_log_check:
required_for: tier_3_and_above
checks:
- workflow/DECISIONS.md exists
- At least 1 decision entry present
- Each entry has timestamp, context, rationale
- CORRECTIONS.md entries (if any) reference original decisions
on_missing: FAIL with feedback "Controller must maintain DECISIONS.md during coordination"
coordination_log.yaml missing (tier 2-4)workflow/validation_report.yaml - Pipeline-standard validation output (PASS/FAIL/REVISE)workflow/events/EVT-{N}.yaml emission removed — verdict in validation_report.yaml is the canonical routing signal)outputs/final/validation_report.yaml - Detailed validation report (legacy location, also written)outputs/final/validation_summary.mdinstruction.yaml, workflow/plan.yamlworkflow/coordination_log.yaml (primary validation target)workflow/work_items.yaml - Acceptance criteria to validate againstoutputs/* (all outputs){domain}/config/validator_config.yamlPart of: cAgents Controller-Centric Architecture