agent-learning-mine

$ARGUMENTS

Purpose

Mine past session transcripts for friction and write it into learning/observations/ as raw observation files. This is retroactive capture — a sibling to the in-band capture a run does live, feeding the same observations/ folder under the same rules. It catches what live capture misses: friction in older sessions, and the quiet kind an agent never self-reported.

It is the proposal's "optional, fenced mine.py feeder" — best-effort, and degrades gracefully. The markdown observations are the durable layer; they survive even if transcript access ever vanishes.

What this is not

• Not the review/consolidate pass — it never clusters, proposes, or writes a lesson. That is learning-review, run separately afterward.
• Not a fix-proposer — it writes raw observations only; the no-fix rule is the whole point.
• Not live infrastructure — you run it by hand, it writes evidence, it stops.

Required context — defer to these, don't restate them

The observation contract lives in the learning folder; this skill carries only the mining procedure. Read these first and obey them for the card shape, the five kinds, and the no-fix / immutable / one-file-per-sighting rules:

• learning/AGENTS.md — the contract.
• learning/observations/_TEMPLATE.md — the exact file you write (frontmatter, Saw. / Not claiming., no fix slot).

Task

1. Ground in the contract above. The rules are theirs; this skill is only the procedure.
2. Distill the corpus. Run python3 .claude/skills/agent-learning-mine/scripts/extract.py (optionally --since YYYY-MM-DD). It turns the raw transcripts (1+ GB) into small per-session digests — human turns verbatim, /overwhelmed firings, and friction-flagged assistant lines — in a scratch dir outside the repo, and prints that path. It fails loud if the transcript format has drifted; if it does, stop and report — don't mine garbage.
3. Read existing observations in learning/observations/ so you don't re-nominate what's already captured.
4. Nominate candidates from the digests. For a large corpus, fan out reader subagents (Sonnet is plenty — this is extraction against a clear rubric) over digest batches. Each candidate is one sighting + a verbatim quote as its evidence + a proposed kind — and no fix (defer to the template). You hold synthesis: dedup identical sightings (keep genuine cross-session repeats — those are the signal), drop weak or fix-smuggling ones, and verify each quote against the digest before trusting it.
5. Write raw observation files per the template — one file per sighting, immutable shape. Stamp the run: field with the session id it came from, marked as mined (the template allows a session id when there is no packet).
6. Report and stop. Do not consolidate or propose. Run /learning-review agentic-build separately when you want the clustering pass.

Hard lines

• Raw observations only — no fix, no lesson, no routing decision (the template's three rules govern).
• Fenced + best-effort — the transcript store is an undocumented Claude Code internal; fail loud on drift, never silently emit empty or malformed cards.
• Digests are scratch — they hold transcript content; leave them in the temp dir, never commit them.
• Mining sees only noticed friction — say so in your report; a confident-but-wrong run no one caught leaves no trace, so absence from the log is not proof.

Final response

End with:

• the scratch digest dir + how many sessions it distilled
• how many observations you wrote, by kind, and from which sessions
• the honest blind-spot line (mining sees only what was noticed)
• No lesson was proposed and nothing was consolidated — run /learning-review agentic-build for that.