Documentation menu
On this page

ADR 0029 — Conversation-authored skills: observer + curator (v53-F1)

Date: 2026-07-17 · Status: accepted

Context

Skills were learned from ONE source — repeated worker runs. A workflow taught in chat evaporated when the chat ended; v51-F4's create_skill requires someone to notice the pattern and write it down. The operator's goal: "learn while performing things … so next time it saves time and eventually uses fewer tokens."

Decision

  1. The observer is opt-in, heuristic-only, and off the request path. A settings toggle (set_skill_observer, carded, default OFF — the v47-F7 posture for ambient behavior). v1 detects one honest signal: a completed turn with ≥3 tool steps. No model call rides the sweep. It runs on the existing ticker with a message-id cursor — not as a post-turn hook inside the chat engine (deviation from the plan draft, recorded: the ticker is genuinely non-blocking and needs no generator surgery; a proposal arriving one tick later costs nothing).

  2. Drafts, never skills (ADR 0016). The observer writes SkillCandidate rows, provenance="conversation", status draft. A candidate never self-promotes.

  3. The human gate without the worker test. skill approve admits a conversation draft directly — the v51-F4 reasoning extended: the test gate exists to prove a LEARNED WORKER recipe runs; a chat procedure has no runnable worker test, and the approve verdict IS the human gate. Learned drafts still must pass their test first (pinned). promote_to_template preserves provenance="conversation" so the registry records the generator.

  4. The curator surfaces, never acts. Stale drafts (unreviewed >30d) are named in the digest. It never archives, merges, or deletes — delete_skill is a carded verb, and a tick silently unloading an approved skill would be a shadow path around that card (the v53 review correction). Template IDLENESS surfacing is deferred with a named trigger: templates carry no usage tracking; when last-used tracking exists, the curator surfaces idle templates the same way.

Consequences

  • With the observer on, a repeated multi-step ask becomes a reviewable draft; approved, it appears in the chat skill index (ADR 0027) and the Queen follows the recipe instead of re-deriving it.
  • Correction-detection and cross-chat repetition analysis are recorded future work, not shipped heuristics.