by XRenSiu
Done_When Pipeline v1.0 — turn fuzzy natural-language requirements into machine-verifiable completion contracts, then run a multi-agent acceptance loop against an implementation. Nine skills in a two-layer topology (per HTML v2 architecture). Layer 1: TWO contract producers + SIX independently-invocable review skills — `/acceptance-spec` (NL → EARS spec + done_when.yaml with existence/behavior/rules schema + spec-robustness.md anti-gaming companion), `/test-suite-generator` (EARS → 5-layer test pyramid: existence/unit/integration/e2e/mutation; the v0.x fitness rubric layer was retired per HTML v2 §3.5 fitness-check dissolution), and six review skills each user-invocable on their own: `/code-reviewer` (diff → findings, focus-driven: security/logic/perf/style/all; Detective Loop not flowchart; 5-finding cap; cross-vendor adversarial mode), `/qa-reviewer` (actually runs tests, classifies maintenance-vs-genuine failures, emits go/no-go), `/pm-reviewer` (Agent-as-Judge: LOCATE/READ/RETRIEVE atoms; requirements normalized from EARS/Jira/Linear/PRD/issue; 4-state TicketCompliance verdict where requires_human_verification is the formal home for genuinely-unautomatable evaluation), `/spec-drift-detector` (code archaeologist: detects spec/code factual divergence without judging which is correct; git_blame traces commit_introducing_drift; 3 divergence types: timing/behavior/contract), `/spec-gaming-detector` (assumes author is gaming; 6 RHD patterns absolute + diff mode; outputs spec_robustness_gaps for contract hardening), `/meta-judge` (synthesizes findings via 4 actions: dedupe/weight/arbitrate/classify; HARD WALL — does NOT re-review code; pluggable rules source). Layer 2: `/acceptance-fleet` is pure orchestrator — dispatches the 6 review skills in parallel against an impl, hands findings to /meta-judge, decodes verdict into four-state ratchet (DONE/FIX/SPEC_DRIFT/GAMING_RISK), persists every iteration to ratchet-log/iteration-NNN/. Anti-gaming structural guarantee: implementation agents MUST NOT see evaluator prompts; minimum medium isolation (mixed Claude sizes) enforced; cross-vendor (Codex/Gemini) preferred for adversarial-reviewer and spec-gaming-detector to break the Claude-reviewing-Claude sycophancy loop per Milvus benchmark. Borrows: PR-Agent diff schema + TicketCompliance 4-state; Greptile v3 Detective Loop; Anthropic Code Review verification step + fleet-by-focus; DevAI Agent-as-Judge (LOCATE/READ/RETRIEVE); Playwright Healer maintenance-vs-genuine; Dartmouth/Yale Meta-Judge (replaces multi-agent debate); 4-Eyes Principle from finance; Weaver framework for weighted weak-verifier ensemble; Komorebi AI Specification Self-Correction. Design philosophy: verifiable beats judgeable (even for things that feel subjective); debate amplifies bias (use meta-judge synthesis, not debate); independence-by-default (each review skill works standalone — done_when.yaml is just one of its consumers).
Claude Code1 Skill