| name | dartai-adversarial-quality-loop |
| description | Adversarial cooperation loop for code quality verification with plan adjustment at each phase. 對抗協作循環:實施者與驗證者互相制衡,逐階調整計劃,確保代碼品質。 Use when: execute dart task, run quality pipeline, adversarial review, implement with verification |
| disable-model-invocation | true |
| context | fork |
Adversarial Quality Loop (Ralph Wiggum Pattern)
實施者與驗證者對抗協作之持續執行環,確保代碼品質。驗證者主動尋缺,實施者守護與修復。
This skill body holds trigger conditions, the phase index, dispatch prerequisites, and core principles. Per-phase DO/DO-NOT detail, eagle-eye violation patterns, reviewer dispatch prompts, and cite-verify rules live in references/*.md and are loaded on demand at the matching decision branch.
Trigger
Use when the task-executor agent runs the full adversarial pipeline on a Dart task. Caller is task-executor under dartai:start, dartai:task, or workflow:start-loop.
Agent Dispatch Prerequisites
This loop dispatches reviewer subagents in parallel via Agent (alias Task). Two prerequisites apply at the dispatch site:
- Top-level driver only. Subagents cannot spawn subagents — the harness scopes the deferred-tool list per-agent and does not surface
Agent/Task to nested runners. If this skill executes inside a subagent, stop and report to the parent. Do not fall back to inline reviewer logic — that would collapse the adversarial separation the loop depends on.
- 驗
Agent schema 可呼。 三態分類(前 Agent 缺非「我在 subagent」之證):
- Preloaded —
Agent(alias Task)見於頂層 <functions> 塊。直用。
- Deferred — 列名於
<system-reminder> deferred-tools 段而 schema 缺。直呼必 InputValidationError。先 ToolSearch query="select:Agent" max_results=1 載 schema,本回合續可用。
- Neither — 真無。告 user 分類結果,毋默降級為 inline reviewer——該降級會塌陷 adversarial 分離。
Core Principles
Context-Sized Tasks
每環中任務須:
- Scoped: 單次專注會話可完成
- Isolated: 與其他並發工作獨立
- Measurable: 完成定義明確
- Bounded: 每任務最多3-5個文件
Plan Adjustment Protocol
每階段結束後,自動(勿停待確認):審視所發現內容;依發現更新餘下任務;若發現阻塞問題則重新排序;在任務評論中記錄調整;立即繼續下一階段,除非BLOCKED。
plan_adjustment_rules:
automatic_continuation:
description: "Phase transitions are automatic, not approval gates"
behavior: "Adjust plan silently and continue"
when_to_stop:
- "Task scope exceeds limits (split required)"
- "Critical blocker found (cannot proceed)"
- "Security vulnerability discovered (must fix first)"
- "All tests failing with no clear fix"
when_to_continue:
- "Minor issues found (add fix tasks, continue)"
- "Scope clarification needed (note and continue)"
- "New edge cases discovered (add to plan, continue)"
- "Pattern conflicts found (note for later, continue)"
never_ask:
- "Should I continue?"
- "Do you want me to proceed?"
- "Is this plan okay?"
- "Ready for next phase?"
Phase Index & Reference Pointers
Phase 0 — Git Hygiene & TDD Setup
Phase 1 — Read Grilled Task Spec
Phase 2 — Adversarial Implementation
Phase 3 — Concurrent Adversarial Review (verdict-file channel)
Phase 4 — Quality Gate Verification (lint, test, coverage, static analysis)
Phase 4.5 — Review for Plan Updates (C-class refactor discovery)
Phase 5 — Post-Task Deep Review (security, performance, PM, replan)
Phase 5.5 — Lightweight Cite Verify (post-commit, pre-Done)
Phase 6 — Final Validation
Each phase entry is a fetch instruction the executor MUST follow before acting on that phase. Skipping the reference and improvising from this body alone is a protocol violation.
| Entering | Load reference before proceeding |
|---|
| Phase 0 (git + TDD setup) | references/phases.md § "Phase 0" |
| Phase 1 (spec confirm) | references/phases.md § "Phase 1" |
| Phase 2 (implementation + self-review) | references/phases.md § "Phase 2" + references/eagle-eye.md (verifier role) |
| Phase 3 (parallel reviewers) | references/dispatch.md (full prompts, verdict-file channel, result handling) |
| Phase 4 (quality gates) | references/phases.md § "Phase 4" |
| Phase 4.5 (refactor discovery) | references/phases.md § "Phase 4.5" |
| Phase 5 (post-task deep review) | references/phases.md § "Phase 5" |
| Phase 5.5 (cite verify) | references/cite-verify.md |
| Phase 6 (final validation) | references/phases.md § "Phase 6" + "Loop Continuation Protocol" |
| Any verifier moment (self-review, Phase 3, Phase 5) | references/eagle-eye.md for the seven violation classes |
Eagle-Eye Violations (rejection grounds)
When the verifier role activates (Phase 2 self-review, Phase 3 parallel reviewers, Phase 5 deep review), load references/eagle-eye.md before approving anything. Seven violation classes — any single occurrence is grounds for immediate rejection:
- Scope creep & gold plating
- Over-engineering & complexity
- Incomplete work markers (TODO/FIXME/XXX/HACK/STUB/...)
- Giving up / "too hard" cop-outs
- Codebase integration failure (code doesn't blend)
- Test ownership violations (blame, "not my test")
- Eagle-eye verification checklist (automated + manual)
Detection rules, pattern lists, and verdict semantics for each class live in references/eagle-eye.md. The body of this skill names the classes; do not approve from class names alone — load the reference for detection rules.
Verdict-File Channel (Phase 3 + Phase 5)
Reviewers write verdicts to .dartai/reports/<task-id>/<role>.md (line-oriented per verdict-schema.md). The driver gates on file content, not stdout body. Always clear the reports dir at Phase 3 entry, even on retry-after-fix — stale verdict files would otherwise be parsed as the current decision.
Full protocol (clear/recreate, parallel dispatch prompts for all five always-on + two conditional reviewers, Monitor fallback, retry rules, conditional-skip handling): references/dispatch.md.
Loop Continuation Protocol (Phase 6 close)
- On Success: Update task status to Done; add completion comment with summary; log metrics; RETURN to main loop (it spawns the next task subagent).
- On Failure: Document the specific failure point; add failure comment with details; recommend whether a fix task is needed; identify blocked downstream tasks; RETURN to main loop (do NOT stop the loop — the loop owns continuation decisions). You are one iteration.
- Plan Adjustment Summary is logged into
.dartai/loop-state.json under the task entry — see references/phases.md § "Loop Continuation Protocol" for the full schema.
Metrics
每任務追蹤:每階段耗時、各點所作調整、各類別發現問題、修復有效率。
跨任務追蹤:常見失敗模式、階段瓶頸、調整頻率、品質趨勢。
Companion References
references/phases.md — full Phase 0/1/2/4/4.5/5/6 spec (DO/DO-NOT, verification criteria, plan adjustment checkpoints, loop continuation schema).
references/dispatch.md — Phase 3 reviewer dispatch playbook: prepare reports dir, parallel reviewer prompts (5 always-on + 2 conditional), verdict-file consumption rules, Monitor fallback, retry/skip handling.
references/eagle-eye.md — seven violation classes with full detection rules, pattern lists, examples, and verdict semantics. Load whenever the verifier role is active.
references/cite-verify.md — Phase 5.5 lightweight citation verification: form-specific check rules (file:line, symbol, git_sha, memory_id), retry budget, Tier separation.
Related
- 檔案後端孿生:
workflow:adversarial-quality(file-backed 同一循環)。