菜单
Orchestrate task execution via fresh subagents with dispatch, monitoring, and result collection
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
| name | subagent-execution-loop |
| description | Orchestrate task execution via fresh subagents with dispatch, monitoring, and result collection |
| version | 1.0.0 |
| category | experiment-execution |
| type | tactic |
| orchestrates | ["implementer-dispatch","execution-monitoring","result-collection"] |
| dependencies | {"sops":["execution-monitoring","implementer-dispatch","result-collection","ponytail:ponytail","ponytail:ponytail-review","superpowers:receiving-code-review","superpowers:requesting-code-review","superpowers:subagent-driven-development","superpowers:test-driven-development"]} |
执行循环直接 load superpowers:subagent-driven-development 当引擎，不再自写循环伪码。 每个任务在引擎内部按以下贯穿规则执行：
Skill load ponytail:ponytail —— 任务开始即开启精简反射。Skill load superpowers:test-driven-development —— 每任务 RED → GREEN → REFACTOR。Skill load superpowers:requesting-code-review —— 任务 diff 出来后派 reviewer 子代理。Skill load ponytail:ponytail-review —— code-review 之后，对 diff 再过一道
过度工程审（delete/stdlib/native/yagni/shrink）。Skill load superpowers:receiving-code-review —— 核验 review 反馈再落地，反馈错就反驳。DARE 原生的 implementer-dispatch / execution-monitoring / result-collection 作为 引擎内每任务的派单、盯状态、收结果三步保留。
| Condition | Action |
|---|---|
| model 选择 / retry / 超时 / 死锁 | 交给 superpowers:subagent-driven-development 引擎处理 |
| code-review 后 | 先 ponytail:ponytail-review 查过度工程，再 receiving-code-review 落地 |
| Budget < 10% remaining | 停，报告 partial（DARE 预算治理） |
| >50% 关键路径 BLOCKED | 中止执行（DARE 排程判据） |
Optional, no fixed order; the final leaf is always a sop.
| SOP | When to use |
|---|---|
| execution-monitoring | Monitor execution progress, detect anomalies, and report status |
| implementer-dispatch | Dispatch execution subagent — select model by complexity, construct prompt with full task context |
| ponytail:ponytail | Lazy-senior reflex: simplest thing that holds; mark every deliberate shortcut |
| ponytail:ponytail-review | Audit the diff for over-engineering (delete/stdlib/native/yagni/shrink) |
| result-collection | Collect experiment outputs — metrics, logs, artifacts — into structured result set |
| superpowers:receiving-code-review | Verify review feedback before applying; push back when wrong |
| superpowers:requesting-code-review | Dispatch a code-reviewer subagent after each task |
| superpowers:subagent-driven-development | Execute the plan via a fresh subagent per task with two-stage review |
| superpowers:test-driven-development | RED -> GREEN -> REFACTOR per task |
Strategy: Attack an isomorphism claim by demanding an explicit structure-preserving map and trying to break it. Targets any multi-language claim of the form 'X ≅ Y ≅ … across N mathematical languages'. Forces the claim to either earn the word 'isomorphism' or be demoted to 'analogy'. Methods: category theory (functor/natural-iso criteria), model theory, Lakatos monster-barring.
Strategy: Dialectic engine retuned for truth-seeking, not survival. A defender steelmans a claim into its MOST falsifiable form, a critic attacks to refute it, a judge classifies the exchange into BROKEN/CORROBORATED/UNFALSIFIABLE — the judge does NOT pick a winner or score persuasiveness. Methods: Irving debate (repurposed), Toulmin argumentation, Mayo severe testing.
Strategy: Run BEFORE building any validator (sandbox/simulation/benchmark). Builds a non-circularity matrix of theory-claim × validator-assumption to detect when a validator would 'confirm' a theory only because it was built on the theory's own premises. A circular validator's PASS carries zero evidential weight. Methods: Cartwright nomological machines, Winsberg sanctioning-of-simulations, tautology detection.
Strategy: Attack a beautiful unified result on the suspicion that its beauty is the bug. Distinguishes EARNED simplicity (forbids/predicts/subsumes) from DECORATIVE simplicity (re-describes/relabels/accommodates). Directly serves the Occam aesthetic by making it a falsifiable bar, not a vibe. Methods: Sober parsimony-as-evidence, MDL, Meehl risky prediction, accommodation-vs-prediction.
Campaign: Truth-seeking adversarial validation for scientific research artifacts (NOT publication defense). Core question: Where have we fooled ourselves, and is each load-bearing claim even falsifiable? Win-condition is INVERTED from survival/resilience to active refutation. Methods: Popper falsificationism, Lakatos Proofs and Refutations, Mayo severe testing, Platt strong inference.
Strategy: Attack the evidential weight of an 'independent convergence' claim. When N reasoning paths all reach the same conclusion, the confidence boost is real only if the paths were actually independent. Measures shared-prior / shared-blindspot contamination and corrects the over-counted confidence. Methods: Bayesian agreement-as-evidence, correlated-error analysis, jury theorem assumptions.