keyoku 2.12.1 — stable

keyoku is stable. The harness with muscle memory — it watches how you work in Claude Code, Cursor, or Codex, turns the goals you actually finish into reusable workflows, and brings them back the next time your agent faces something similar.

Why this is the stable baseline

• Muscle memory works end-to-end — capture (live goal_focus, explicit goal_record, or activity backfill), convergence on machine-checkable criteria, and reuse.
• Reuse needs no API key — the coding agent itself judges which learned workflow applies (matches by meaning, not just shared words). An SLM is an optional accelerator for headless runs.
• Verified, not asserted — 263 tests + a CI muscle-memory eval (precision@1 100%, 0% false positives); behavioral lift measured at +67–100pp across models.
• Bounded + local-first — every persistent log is size-capped, secrets redacted at write time, ~/.keyoku at 0700/0600, keyoku inspect shows exactly what's stored. No cloud, no telemetry.

Highlights since 2.0

• Muscle memory: build-then-verify capture, pitfalls (negative memory), live goal_focus, activity backfill + keyoku backfill repair, keyoku refine.
• Zero-dependency reuse: agent-as-judge recall; optional SLM re-rank; self-pruning suggestions.
• Trust & ops: keyoku inspect, release preflight, honest single-sourced version, all logs bounded.

This release (2.12.1)

• Bounded audit.jsonl (~1 MB / ~2000 entries) — the last unbounded log. Regression test added.

Install

npm install -g keyoku && keyoku init

Full history: CHANGELOG.md