v2.4.1 — 🏁 The factory checks its own claims

The factory now checks its own claims. A frontier-model audit of every component (35 agents · 42 playbooks · 42 skills · 137 practices) produced a four-phase upgrade roadmap — and all four phases shipped, each one enforced by machinery rather than documentation, then sealed by the factory's own learning loop.

Highlights

🧾 Eval coverage is a CI-enforced ratchet. A new agent or skill cannot ship without a verified eval case; a covered component that loses its cases is a blocked regression. Everything pre-existing is grandfathered by name in a committed manifest — visible debt that can only shrink. First agent-level behavioral evals (problem-sharpener, devils-advocate) verified live and baselined.

🤝 Agent handoffs are typed contracts. The six core chains (idea→discovery→synthesis, architect→builder, keyword→SEO, pricing launch→scale, brand→landing copy) declare Produces / Shape / Preconditions / On-gap. The interview synthesizer refuses batches under 5; the builder refuses to build without written architecture; keyword briefs cannot invent search volumes.

🎛️ The model policy is wired, not declared. Every spawnable agent carries its model tier as frontmatter (15 opus / 15 sonnet / 4 haiku); a tested resolver reads it at the spawn boundary; stakes escalation is active and up-only — anything touching auth, payments, migrations, RLS, or data deletion runs on the top tier regardless of role.

🧠 Recall is push, not pull. /work-on and /full-cycle auto-inject brain recall at session start — anti-patterns first, because defenses beat context. Product-level learnings are indexed and feed the promotion pipeline.

🚁 The heavy verdicts go fleet-native. Validation, security checks, code review, devil's advocacy, and portfolio reviews can fan out as blind parallel workers, pass every blocking finding through an adversarial refuter ("default to refuted"), and synthesize with a judge that reports disagreement instead of averaging it away. Serial execution stays first-class.

🔁 The loop closed on itself. /learn-loop's first dual-scored run: five candidates scored independently twice, the fresh-eyes scorer rejected two of them, the operator ratified the survivors. The commit that introduced the retro rule had to pass the retro rule to merge.

Also inside

• Site-audit capability set (v1.33–v1.34): the "looks static, but it's live" trap — an env-gated backend is still a backend; the backend-reality check now runs first in /security-check, the 70-check security checklist, and the launch workbook.
• /pr stowaway guardrail: PR branches cut from origin/main explicitly, publish-set review before push — an unpushed parallel-session commit can never again ride into a squash-merge unreviewed.
• Retro + decision disciplines with teeth: every changelog entry needs a retro or a recorded skip (CI-enforced); decision capture is a forced end-of-session question.
• Practices ledger: 134 → 137 (35 ✅ proven with dated incidents); playbooks 41 → 42 (fleet-patterns).

Get started

curl -fsSL https://raw.githubusercontent.com/hamza-ali-shahjahan/hamzaish/main/install.sh | sh

Then open Claude Code and type /builder-mode <your idea>. Full beginner path: docs/start-here.md.

License

AGPL-3.0-or-later. Built on a thousand generosities — see ACKNOWLEDGMENTS.