Session ingest

Ingest this session into the wiki

Capture the conversation you just had — from any repo — and fold it into the vault like a scoped ingest. A deterministic script scrapes the transcript, a Haiku subagent summarizes it (keeping the bulky raw text out of this context), then you do the wiki integration after a confirmation step.

Vault root

Resolve the vault root first. Run this in a Bash tool call (not via !-inline execution — the static permission check rejects shell expansions, so the inline form fails with "Contains expansion"):

d="${CLAUDE_SKILL_DIR:-$(pwd)}"; if [ -n "${SECOND_BRAIN_DIR:-}" ]; then echo "$SECOND_BRAIN_DIR"; else while [ "$d" != "/" ] && [ -n "$d" ]; do if [ -f "$d/CLAUDE.md" ] && [ -d "$d/wiki" ]; then echo "$d"; break; fi; d=$(dirname "$d"); done; fi

Steps

0. Resolve the vault root (precedence):

   1. If $ARGUMENTS is non-empty, use it as the vault path.
   2. Else use the auto-resolved path injected above ($SECOND_BRAIN_DIR, or derived from this skill's location).
   3. If neither yields a directory containing CLAUDE.md + wiki/, stop and tell the user to set SECOND_BRAIN_DIR (in ~/.claude/settings.json env, or the repo's settings) or pass the path as an argument. All vault writes below use this absolute root.

1. Scrape the transcript: python3 ${CLAUDE_SKILL_DIR}/scripts/scrape_session.py ${CLAUDE_SESSION_ID} It writes a clean digest to /tmp/claude-session-<id>.md and prints STAGING_FILE=, TITLE=, DATE=, CWD=, SLUG=, TURNS=. Capture the staging path. If it exits with an error (no transcript), stop and report.

2. Summarize via a Haiku subagent. Spawn an Agent with model: haiku whose prompt gives it the absolute staging-file path and this output contract — it reads the file and returns ONLY:

   • topic (short, for the page slug) and a 1–2 line summary
   • what was done / decided
   • entities touched: projects, people, organizations, goals (names)
   • action items (owner + absolute due date where stated)
   • open threads / next steps and key takeaways The raw transcript stays in the subagent's context, not this one.

3. Confirm before writing. Read <vault>/index.md to dedup against existing pages. Then show the user the proposed sessions/ page (title + summary) and the list of wiki pages you'll create/update. Wait for approval. Convert any relative dates to absolute (today's date is in session context).

4. Full ingest (mirror the ingest discipline; all paths absolute under the vault root):

   • Write <vault>/wiki/sessions/YYYY-MM-DD-<topic-slug>.md. If <vault>/wiki/_templates/session.md exists, use it; otherwise use the embedded structure below.
   • Fan out: create or update each touched projects/, people/, organizations/, goals/, devlog/ page. Add new facts; cite back with [[sessions/<slug>]]; prefer updating an existing page over a near-duplicate; on conflicts add a > [!warning] Contradiction callout linking both and flag it; bump each touched page's updated: to today.
   • Update <vault>/index.md with the new session and any new pages (one-line summaries).
   • Append to <vault>/log.md: ## [YYYY-MM-DD] session-ingest | <topic> with 1–3 bullets.
   • Update <vault>/home.md if active projects, open goals, or recent decisions changed.

5. Clean up. Remove the /tmp staging file.

Embedded sessions/ page structure (used when no template is present)

---
date: YYYY-MM-DD
repo: <cwd the session ran in>
projects: ["[[projects/project-name]]"]
tags: [session]
---

# <Topic> — YYYY-MM-DD

## Summary
1–2 lines.

## What was done
- ...

## Decisions
- ...

## Action items
- [ ] <action> — owner: [[people/firstname-lastname]] — due: YYYY-MM-DD

## Open threads
- ...

## Links
- [[projects/project-name]] · [[people/firstname-lastname]]

Done when

• A sessions/ page exists for this session, relevant entity/project pages are updated with cite-backs, index.md + log.md reflect it, and the /tmp staging file is removed.