Second Brain: router

You don't do the work here. You decide which flow runs, in what order, and then follow that flow's SKILL.md exactly. Rules live in AGENTS.md; procedures live in .agents/skills/<flow>/SKILL.md.

1. Orient (cheap, every time)

1. If you haven't done it this session, do the session-start reading: 90-system/about.md, index.md, and the tail of log.md (grep "^## \[" log.md | tail -10).
2. Snapshot the vault state (one command; no page reads):

   echo "inbox: $(find 00-inbox -type f ! -name .gitkeep | wc -l)"; \
   echo "last lint: $(grep '^## \[.*\] lint' log.md | tail -1)"; \
   echo "last review: $(grep '^## \[.*\] review' log.md | tail -1)"; \
   echo "indexed pages: $(grep -c '^- \[\[' index.md)"; \
   grep -q '^- Role / work: *$' 90-system/about.md && echo "profile: EMPTY" || echo "profile: ok"

3. Once per session, run scripts/qmd-sync.sh to pick up edits the owner made outside the agent (Obsidian, journal). It's a no-op without qmd and embeds in the background.

2. Classify the request

Match on intent, not keywords. Pick the first row that fits.

Ambiguous? Ask one short question with the 2 most likely flows as options, e.g. "Save it to the inbox for later, or ingest it into the wiki now?" Don't ask when the default below applies.

Defaults when unsure:

• URL or pasted content with no verb → capture, then offer ingest.
• A question that the vault clearly doesn't cover (nothing in index.md) → say so in one line and answer as query with a large Gaps section. Offer to capture sources on it.
• "Clean up" with a lint report less than 7 days old → repair proposal from that report; otherwise lint.

3. Chain flows (multi-step requests)

Split compound requests into flows and run them in sequence, each to completion (including its own check-ins and log entry) before starting the next. Announce the chain first in one line: "Plan: capture → ingest → query. Starting with capture."

Common chains:

Chaining rules:

• Stop at every owner decision point (ingest check-in, lint → repair, file-back, brief edits). Never treat the chain's announcement as approval for the gated steps.
• If a step fails or is blocked, stop the chain, report where you stopped and why, and offer to resume.
• More than 5 inbox items or more than 3 flows: confirm scope first ("12 items in the inbox. Do all, or the 3 oldest?").

4. Suggest (when there's no clear request)

Use the snapshot from §1 and propose the top 1-3 actions, highest first:

1. Profile EMPTY → setup.
2. Inbox has items → ingest (name the oldest 1-3 with their age).
3. Last lint more than 30 days ago (or never, and the index has 10+ pages) → lint.
4. Last review more than 7 days ago → brain-review.
5. Active project with no log entry for more than 14 days → project resume.
6. Otherwise → gap-finder (what to read next for the current focus), or a query on the current focus.

Present them as a short numbered list and ask which to run.

5. Hand off

1. Read .agents/skills/<flow>/SKILL.md in full (or invoke the skill by name if your tool supports it).
2. Follow it exactly: its steps, check-ins, self-check, and hard constraints override anything here.
3. When a flow finishes, either continue the announced chain or end with one suggested next step (e.g. after ingest: "Ask a question across these sources?"; after lint: "Say fix 1-3").

Hard constraints

• The router itself writes nothing except through a flow. No direct wiki edits "on the way".
• One flow at a time; every flow logs its own entry. The router adds no extra log lines.
• AGENTS.md §6 hard rules apply throughout.