transcript-mininglisted

# Transcript Mining Session transcripts are the raw evidence for workflow audits and skill distillation, but they are huge (single sessions reach several MB; projects reach 100+ MB). This skill encodes the discipline that keeps mining cheap and the findings sound. ## Hard rules (violations are exactly the observed waste modes) 1. **Never Read a transcript file whole.** A single `Read` of a session JSONL can overflow the context. Always extract with `python3`/`jq` and bring back only aggregates or matched lines. 2. **Scan ONLY the named session files.** If the task names sessions, touch nothing else — do not "also check related sessions for completeness" (observed failure: an executor re-scanned an entire project instead of 7 named files). If targets are unknown, select by *metadata only* (`ls -lat`, sizes, mtimes) — never by reading contents to decide. 3. **Delegate per-project fan-out to subagents**, one per project directory, each returning a quantitative report — never stream raw transcript content into the main context. 4. **Content vs mechanics separation.** When distilling knowledge (e.g. debugging playbooks): mine only the product/environment content (symptom → probe → resolution). Never mine workflow mechanics (command behavior, gate handling) — transcripts predate current workflow versions and that layer is stale by default. 5. **Provenance, not memoir.** Distilled entries are stated at the failure-class / pattern level, stack-a