v2.22.0 — 📊 The measured factory

The factory stops deciding from memory and starts deciding from measurement — cost-to-outcome model routing, a model-independence bench, three telemetry instruments (what happened · how it felt · what got caught), and a scaffold that's born green. Four cycles since v2.18.0.

Highlights

• Cost-to-outcome routing + model-independence. A cross-model bench (bun run bench) scores each skill's eval across model tiers into a capability-per-dollar leaderboard; the runtime router picks the cheapest model that clears the bar and escalates only the failures (FAIL_BUILDABLE → climb a tier). A check-model-independence ratchet flags any skill that only passes on one model. Honest boundary: the factory optimizes the harness / router / skills against outcomes — never model weights.
• Three telemetry instruments. Traces (what happened) + friction (how it felt) + the new defect registry (which check caught what, at catch time) ground every /learn-loop retro in data, not recollection.
• A reward-wired hill-climb. A reward ledger records the outcomes customers actually care about (executed eval · e2e · activation), and /goal prefers an executed eval signal alongside its fresh-eyes rubric.
• A scaffold born green. The starter typechecks and builds out of the box, check-starter asserts install + typecheck + build, and every external call goes through a conscious retry door.

What's inside

v2.22.0 — the guard-fire defect registry · bun run defect log|report records every real catch (CI red, eval regression, live assertion, hook block, review) as one structured row — catcher · stage · target · severity · fixed-by — seeded with the factory's historical catches so the first report is already useful. /hamzaish now consults the control plane interactively: the weekly mandate outranks the menu.

v2.21.0 — cost-to-outcome loops · model-independence bench + capability-per-dollar leaderboard, an evidence-based router (measured beats guessed) with a cheap→frontier cascade, and a reward ledger + executed-eval signal feeding /goal. New doctrine: cost-to-outcome & model-independence.

v2.20.0 — the flywheel tightens · the friction/delight instrument, a capture doctrine with an anti-rot do-NOT-capture list, a MECE gate on new skills with a quarterly archive-never-delete curator, and the eval debt paid down (5 → 8 covered entities).

v2.19.0 — a scaffold born green · starter typecheck/build truth, check-starter that runs the real commands, real /privacy + /terms pages, and retry doors on every external call.

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>. New to terminals? Start at docs/start-here.md.

License

AGPL-3.0 — clean, no added clauses. Free for builders; commercial license on request.