v0.2.0 — Self-Improving SDD: Adversarial Review, Feedback Promotion & Quality Gates
Prospec v0.2.0
First major release of the benwu95/prospec fork,
building on upstream ci-yang/prospec v0.1.5.
v0.1.x gave you a linear SDD pipeline (story → plan → tasks → implement → verify → archive).
v0.2.0 closes two feedback loops on top of it so the workflow doesn't just repeat —
it compounds. Every change now passes through an independent adversarial reviewer, and
recurring lessons promote (human-gated) into versioned team rules. The whole pipeline is
wrapped in objective quality gates.
122 files changed · +5,119 / −8,851 · 10 merged PRs (#2–#11)
✨ Headline features
/prospec-review — Adversarial review → fix loop (new)
An independent, fresh-context reviewer audits the entire change diff between implement
and verify. Only verifier-confirmed, drop-in critical findings are auto-fixed; major
findings are proposed (downgraded to WARN), the rest escalate to you. It's spec-aware —
it checks against delta-spec REQs, dependency direction, and module conventions, which a
generic reviewer can't. The commit boundary stays after verify hits grade S/A, so
implement + review + verify fixes land in one atomic commit (Prospec prompts; never
auto-commits).
/prospec-learn — Feedback promotion pipeline (new)
Collects recurring lessons (session corrections, repeated verify FAILs, recurring review
criticals from the cross-stage quality_log), scores them with an explicit, reproducible
rule (frequency + impact-module count — not a black box), and — only with explicit human
approval — promotes them across three tiers: personal → team _playbook.md → Constitution
rule. Everything is auditable and version-controlled. This is what makes Prospec get
smarter with use, not just bigger.
🛡️ Quality machinery (now across every workflow Skill)
- Output Contract — each Skill self-reports
Met N/M | Overall: PASS|WARN|FAILagainst
objective criteria, so you stop hand-checking artifacts. - Entry / Exit dual gates — preconditions checked before a Skill runs, Constitution
compliance after; WARN/FAIL records persist via a cross-stagequality_logso an
early concern resurfaces downstream. - Executable Constitution —
initseeds rules with RFC-2119 severity
(MUST → FAIL / SHOULD → WARN / MAY → advisory);/prospec-verifygrades against them.
🔌 Agent support & toolchain
- Antigravity CLI replaces Gemini CLI. Antigravity / Codex / GitHub Copilot now unify on
the agents.md open standard — a singleAGENTS.md+.agents/skills/
(Claude Code keepsCLAUDE.md+.claude/skills/). - pnpm-only toolchain on Node ≥ 22.13 (was Node ≥ 20). Installs directly from the GitHub
fork —npm i -g github:benwu95/prospec— with aprepare-script build (npm publish path
removed).
📚 Knowledge & specs
module-map.yamlis now generated duringinit(no longer a manual step)._index.mdgains an aliases column + a shared glossary for better cross-module recall.- Oversized module READMEs are split into sub-modules automatically.
/prospec-verifydimension 4/5 re-scoped to knowledge ↔ code consistency, decoupled
from the feature-spec gate.
🐞 Fixes
- BUG-001 — tech-stack detection is now config-first (respects
.prospec.yamlinstead
of mis-detecting).
📊 Compared to v0.1.5
| v0.1.5 | v0.2.0 | |
|---|---|---|
| Skills | 11 | 13 (+/prospec-review, +/prospec-learn) |
| Tests | 350 | 549 |
| Node | ≥ 20.0.0 | ≥ 22.13.0 |
| AI agents | Claude · Gemini · Copilot · Codex | Claude · Antigravity · Copilot · Codex |
| Install | npm publish | GitHub fork (pnpm-only) |
| Quality gates | — | Output Contract + Entry/Exit + executable Constitution |
| Feedback loops | linear pipeline | + adversarial review + lesson promotion |
⬆️ Upgrade notes
After prospec agent sync, delete the now-unused legacy agent files:
GEMINI.md, .gemini/skills/, .codex/skills/, .github/copilot-instructions.md,
.github/instructions/.
Bump your environment to Node ≥ 22.13 and use pnpm.