Amadeus
2026wipnothing activates itself
Built for OpenAI Build Week · submitted jul 2026The problem
A local-first procedural-memory compiler for coding harnesses. It captures a Codex run a human explicitly opted into, helps that human turn a successful run into a compact skill or a deterministic script, and saves the result as a reviewable preview. Ships as a CLI, a project-scoped MCP server and a local dashboard.
The constraint
Nothing may activate itself. A skill library that auto-promotes what looks successful is a library that quietly poisons every future run — so the preview/active boundary is enforced structurally: a `reviewed_preview` candidate is excluded from retrieval entirely, and only a passing matched evaluation plus a named human promotion installs the frozen package.
What the git log says
- $ Capture starts only on an explicit --consent, scoped to that one session; model reasoning is excluded and API-key variables are stripped from the child environment.
- $ Redaction before persistence — API keys, tokens, passwords, auth headers and private-key blocks replaced with a marker; user patterns rejected if they use lookarounds, backreferences or nested repetition.
- $ Authoring removes known no-ops and duplicate lines, validates safe regex, and enforces a 500-line hard limit with a 120-line compactness warning.
- $ Retrieval abstains unless the top match clears both a score threshold and a margin over the runner-up; only the winning skill's body enters the managed prompt, never the candidate store.
- $ Capture sits behind a HarnessAdapter interface — Codex is the only live adapter today, and the README says so rather than claiming portability it hasn't tested.
What I learnt
Governance is the product. The interesting part wasn't generating skills — it was the boundary that stops a plausible one from being used: consent to capture, a named reviewer with a written rationale, and retrieval that abstains rather than guesses.