Skill miner

Proposals come from your history, not from guesses. Each one cites how many times the pattern actually happened.

1. Mine (local, read-only, asked first)

Transcripts hold whole sessions, including anything ever pasted into them: keys, customer data, private code. {{cli}} mine reads the JSONL files under ~/.claude/projects on this machine and makes no network calls.

• ⛔ Ask the user before each run, naming the scope (this project or --all).
• Report aggregates only: skill names, counts, chains, session ids. Never quote, summarise or paste transcript content, in the report or in chat, and don't name what a session contained, not even to say it was left out: an example is its session id and the step order. To understand a manual step, describe its shape ("edits the changelog, then tags"), not its text.

{{cli}} mine                 # sessions of this project (and its worktrees)
{{cli}} mine --all           # every project under ~/.claude/projects
{{cli}} mine --json > mine.json

The output has top skills, bigram chains with confidence (how often A is immediately followed by B, out of all runs of A), trigrams, and recent session sequences. Installed bee skills are counted under their unprefixed names.

2. Find proposals

• Composite skill: a chain with count ≥ 5 and confidence ≥ 0.5. Package it as one skill that calls the existing skills in order with gates between them, like {{skill:ship-change}} does. Never copy their content into it.
• New skill: the same manual multi-step work repeated across sessions with no skill behind it (look at recent sequences and at what happens between skill calls).
• Retire signal: a skill nobody has run in 60+ days. Check it with {{skill:skill-eval}} before proposing removal.

3. Report

A table per proposal: the pattern, count, confidence, example session ids, and the proposed skill name and steps. Write it to SKILL_MINING_REPORT.md, or to the user if they prefer.

4. Scaffold (only when asked)

⛔ Ask before writing files. A new skill gets SKILL.md with a precise description (the description decides when an agent loads it) and at least one eval scenario ({{skill:skill-eval}}). That way its value is measured from day one.