pipeline-doctorlisted

# pipeline-doctor — symptom → diagnosis → safe repair (CLI only) Recurring pipeline breakages used to be fixed by hand-editing the SQLite state db. This skill replaces that with a fixed recipe table: every repair goes through a whitelisted `scripts/pipeline.py` command, so every state change stays auditable. **Never hand-written SQL, never direct edits to the db file.** Project truth: `CLAUDE.md`. ## 1. Triggers / Non-triggers - **Triggers:** "the pipeline is stuck", "状态机卡住了", "the vault lock won't release", "window-done reports a JSON error", "it won't let me re-run profile / re-profile the source", "diagnose the pipeline / what state is the pipeline in". - **Non-triggers (never fire):** - content quality / coverage / audit → `kb-qa`; semantic health → `wiki-lint-semantic`. - adding a book / writing pages → `ingest`; post-publish retrospective → `kb-postmortem`. - "distill this failure into the skill" → `skill-evolve`. - anything that would require raw SQL / editing the SQLite file → refuse and explain the CLI path. ## 2. Inputs - The symptom as described by the user + any command stderr. - `status` / `next` output (per-source stage/status + next human action). - If windows are involved: `ingest-stats --source <src>` for a quick read of window states. ## 3. Outputs - A diagnosis (which recipe row matched) and the executed safe repair, or an explicit hand-back when no recipe matches. Destructive steps (`reset-source --apply`) always show their dry-run plan fi