Sync Claude Sessions

Pipeline skill: turn recent Claude Code session transcripts into indexed knowledge. Runs standalone or as step 1 of /data-ingestion.

Process

1. Find the last sync point. Read raw/sessions/.last-sync (a single ISO timestamp). If missing, default to 14 days ago.
2. Locate session data. Claude Code stores transcripts under ~/.claude/projects/<project-slug>/*.jsonl. List files modified after the last sync point.
   • Cloud fallback: in a fresh remote container (scheduled routine runs), that directory is missing or holds only this session — the owner's desktop CLI history is unreachable there. Do NOT report "nothing new"; instead digest the current session so far (its task, outcome, mistakes, corrections) as the sole entry, and state in the report that full local history syncs only when this skill runs on the desktop.
   • If there is genuinely nothing to digest (no history AND no meaningful current-session activity), record that and skip to step 6.
3. Digest, don't dump. For each new/updated session, extract only:
   • Date, project, one-line task description
   • Outcome (succeeded / abandoned / partial)
   • Mistakes or repeated corrections (user re-explaining, reverted edits, failed commands retried)
   • Skill opportunities (workflows done manually 2+ times that could become a skill or command) Use jq/grep to pull user messages and tool errors rather than reading whole transcripts into context.
4. Write the digest. Append one file: raw/sessions/YYYY-MM-DD_session-digest.md with a section per session. Never copy secrets, tokens, or credentials out of transcripts.
5. Index. Add a row to the Session Digests table in wiki/index.md summarizing the recurring mistakes and skill opportunities found.
6. Update the sync point. Write the current ISO timestamp to raw/sessions/.last-sync.
7. Report. Sessions processed, top recurring mistake, top skill opportunity. If run from /data-ingestion, keep this to 3 lines.