| name | dartai-post-task-reviewer |
| description | Forked-context deep post-task reviewer — preloads verdict schema + OWASP security + deep code analysis + PM/docs accuracy + replan lens. 對抗深度後任務審查(fork上下文)。 Use when: dispatch post-task-reviewer subagent, deep review after fast gates, OWASP audit, PM/docs accuracy check, replan recommendation |
| context | fork |
| agent | dartai:post-task-reviewer |
Post-Task Reviewer Skill (companion)
對抗性深度後任務審查的薄入口技能。當主循環在快速門通過、品質門綠色後fork出深度審查子代理時,此技能在fork上下文內預加載審查視角,使OWASP遍歷、架構分析、文檔交叉檢查、重新規劃等大量中間步驟不污染主線程,僅最終verdict YAML塊回流(深度細節通過evidence_path離線)。
Why this skill exists
- Context isolation — Post-task review is the heaviest reviewer (security walkthrough + architecture pass + doc accuracy + replan). Without fork, intermediate findings would dwarf the main loop's window. Fork is the difference between a sustainable loop and one that compacts every iteration.
- Single source of truth for output shape — The agent file (
dartai:post-task-reviewer) describes how to think. This skill points at what to emit (verdict-schema) and what playbooks to load (adversarial-quality-loop + code-quality).
- Evidence_path is non-optional in practice — Deep review typically generates findings beyond the ≤30-line verdict budget. Reviewers SHOULD write to
.dartai/reports/<task-id>/post-task-reviewer.md and reference it via evidence_path.
- Feature detection — Fallback below.
Loaded context
When this skill activates inside the forked subagent:
- Verdict schema —
plugins/dartai/skills/verdict-schema.md
- Adversarial quality loop —
plugins/dartai/skills/adversarial-quality-loop.md (architecture lens, completeness checklist)
- Code quality —
plugins/dartai/skills/code-quality.md (refactor-first hook, review-for-plan-updates pipeline for replan output)
- Agent playbook —
plugins/dartai/agents/post-task-reviewer.md (security audit, deep code analysis, PM/docs review, replan)
Review lens (one-liner)
代碼通過快速門。以更深分析破解之。發現並行審查者遺漏之處,然後依所有發現更新計劃。
Output contract
Emit a single fenced YAML block as the final message body, ≤30 lines. Schema in plugins/dartai/skills/verdict-schema.md. Strongly prefer writing depth to .dartai/reports/<task-id>/post-task-reviewer.md and setting evidence_path — deep review almost always exceeds the inline budget.
verdict: pass | fail | warn
confidence: high | med | low
blockers:
- "<file:line> — <one-line description>"
advisories:
- "<one-line nit or follow-up>"
evidence_path: ".dartai/reports/<task-id>/post-task-reviewer.md"
Replan recommendations also belong in the evidence file (under a ## Replan heading); the orchestrator surfaces them to the operator on demand.
Fallback (pre-2.1 harness)
If the runtime does not honor context: fork:
- The skill still loads via the agent's
skills: array; same playbooks in scope.
- Reviewer still emits verdict-only YAML; gate semantics unchanged.
- Only isolation benefit is lost. Because this reviewer is the heaviest, the cost is most visible here — but behavior is preserving and the gate still works.
Related
dartai:verdict-schema — wire format
dartai:adversarial-quality-loop — the loop this reviewer closes
dartai:code-quality — refactor proposal pipeline used for replan
dartai:agents/post-task-reviewer — the reviewer agent that binds this skill