Forge Scribe — never lose what the owner asked

Self-improvement substrate: before applying this skill, read learnings.md in this skill's own folder and honor its corrections. After a run that produced a genuine correction, append it there with a date and real evidence — never invent a lesson.

Every owner prompt is kept word-for-word (secrets scrubbed) and becomes a tracked item R-n that stays visible until it is planned, done with evidence and acked by the owner. The only way to change scribe state is node .claude/forge-bin/forge-scribe.cjs. Never edit the files under .claude/forge-scribe/ by hand.

When to use

• The owner sends a new wish, change, cancel or preference mid-project ("onthoud dit", "ik heb nog een wijziging", "note my change", "scrap that", "voortaan…").
• The owner asks what they asked before ("wat wilde ik ook alweer", "open wensen", "did you miss anything I asked").
• End-of-run reconciliation (before finalize) and on resume of a long project.

Not for: storing a standing rule directly (that is /forge remember), writing a PRD (forge-prd), or taking meeting minutes of a document.

Context rule

Only DIGEST.md (<= 1200 chars) goes back into context. Never paste inbox.jsonl or ledger.jsonl.

Wiring (hooks, rule, finalize)

The hook prints [forge-scribe] Genoteerd: R-n … per captured prompt (the digest only when it changed); act on it. A Stop reminder blocks once per unplanned item; before completion log owner_requests_reconciled (the writer refuses it unless gate is OK) — see config/orchestration/HOOKS_OPT_IN.md section 7.

Only the owner's words (origin filter, WP 3.8)

The hook also fires for text that is not the owner: a subagent hand-back (<agent-message from=...>, [Subagent hand-back]), a cross-session or teammate message, a system notification, a task notification. Such text is judged on its START (or a wrapper that fills almost all of it, never a mere mention) and is not recorded: one line in skipped.jsonl (reason system-origin, char count, hashed session, never the text), no item, no digest line, no REPEAT or split hint. Owner prose that quotes such a tag is recorded as usual. When unsure the scribe records and marks the item origin_uncertain (the receipt says "origin uncertain"): check those words are really the owner's. A hand-back recorded before WP 3.8 is set aside once with reclassify --system-origin --apply (backup first, text kept, folds as superseded with reason system-origin; the gate re-checks it, so it can never hide owner text; reopen brings one back).

Repeats (OWNER REPEATS v2)

REPEAT xN of R-a,R-b — why: R-a 75% (shared words) means the owner had to ask again: make the HARD RULE now, then link it with transition R-n repeat-rule --ref standing_rule:<id>. Only owner items are compared, on the owner's own words: pasted command or skill bodies (# /forge (global)), fenced code, block quotes, long quotations and a re-pasted earlier prompt are cut first, so two different /forge tasks never match on the shared boilerplate. A match needs at least two shared meaningful words and an overlap (dice) of 0.40 or more, or 60% of the new prompt already said before (config repeat.* in FORGE_SCRIBE.json). At most the 3 closest earlier asks are named. Stored links from before WP 3.8 are re-judged on display, so an old false "REPEAT x37" no longer shows. The split hint counts only the owner's own asks.

Toggle corrections ([tc], WP I.8c)

An owner message that says Forge should have turned something on, used it or done it unasked ("zet de lanes aan", "ik moet steeds vragen om ...", "je bent de watchdog vergeten", "why didn't you run the review", "stop asking me") gets the TAG toggle-correction (not a class, not a status). It is judged on the owner's own words only (hand-backs, pasted blocks, quotes and re-pasted prompts never count) and needs a corrective cue; a plain task ("zet de lamp aan in de app") is never tagged. The receipt ends with [tc] toggle-correction (make it a default/rule/check): do that in this run. [tc] marks the item in the digest and list; node .claude/forge-bin/forge-anticipate.cjs metric counts them per mission. Older records: retag --toggle-correction (dry run), then --apply (backup first; --skip R-n for a doubtful hit).

Lifecycle

unclassified -> open -> planned -> in-progress -> done (evidence) -> acked (owner); side states: defer-proposed (an agent's defer, with reason; NOT final until the owner acks it, then deferred = a recorded carry-over), cancelled (owner), superseded, answered.

Evidence types: wp_completed:<wp_id> (linked wp, real event in the run), check_passed:<event_id>, commit:<sha>, standing_rule:<rule id> (preference items only).

Preferences

A preference item ("voortaan…", "always…") is NOT stored as a rule by the scribe. Run forge-scribe.cjs route R-n, ask the owner to confirm, then forge-standing.cjs remember "<text>" (the existing /forge remember path) and close with done --evidence standing_rule:<id>.

Mid-run change protocol

1. impact R-n --run <id> → proposed wp_ids. 2. Confirm with link. 3. Amend a not-started package, or re-create a completed one as a new agent_work_package_created with supersedes and req_ids:["R-n"].
2. Head Chef triages only that delta.

Acks are batched

At the end of a run present ONE list of done-unacked items to the owner (batch-the-asks rule). Never pause mid-run per item. Ack/cancel are refused unless the prompt is a real owner message in the session transcript — a classifier, not a proof.

When to dispatch the Scribe agent (sonnet/low, GOAT quick route)

The scribe specialist (.claude/agents/scribe.md, exec-reviewer class, no Write/Edit) is dispatched by the Lead only for a prompt flagged multi_ask (3+ asks), end-of-run reconciliation, or a resume with

= 20 open items. The normal path costs no LLM call: the Lead runs the CLI directly.

Honesty

Never report an item as planned, done or acked unless forge-scribe.cjs output says so; quote the literal gate exit code.