Skip to content

SKILL.md

Latest commit

 

History

History
48 lines (36 loc) · 2.76 KB

SKILL.md

File metadata and controls

48 lines (36 loc) · 2.76 KB
Raw
name brain
description One-glance health of the "second brain" so failures aren't SILENT — is the local search server alive, is the RAG index fresh, is the session-memory ledger filling, did the nightly distillation run. Trigger on "/brain", "/memory", "brain health", "is the reindex alive". READ-ONLY diagnostic with a green/red verdict per component.
license MIT

/brain — second-brain health at one glance

🧒 When reporting to a non-technical operator, end with a short plain-words recap (rule eli5-always).

Catches SILENT breakage of the memory/RAG stack (the things that kept breaking month after month: the reindex, the local search server, the memory pilot). Read-only, 0 tokens.

What it does (one script)

python $IMPORTS_ROOT/brain_health.py

Checks 5 things and colors them 🟢/🟡/🔴:

  1. Search server on its local port — alive or not (if it's down, auto-recall silently disappears); also shows whether the graph-assisted mode is on.
  2. Index _brain_e5.npy — how fresh (🟡 if the reindex lags >48h).
  3. TurnState ledger — how many turns recorded (memory Phase 1), when the last one was.
  4. Nightly distillation — candidates in quarantine + when the last run happened.
  5. A/B direct-vs-graph recall — how many runs + the operator's verdicts (👍/👎).

Writes a dashboard $OBSIDIAN_VAULT/_Dashboards/Brain-Health.html (the operator reads with their eyes, [[prefer-visual-dashboards]]). Exit code 0/1/2 = ok/warn/red (for scripts).

When to fix (on 🔴/🟡)

  • 🔴 search server down → run the restart script (as admin, see [[always-on-memory-pilot]]); or reboot (the at-logon task brings it up).
  • 🟡 index lagging → gpu_check.py [--kill], then brain_embed_update.py [--wait-gpu 10] ([[reindex-routine]]).
  • 🟡 ledger empty → fine if memory Phase 1 was just enabled (it fills from the next sessions on).

What it does NOT do

It doesn't fix anything and doesn't write to the vault. It's a diagnostic. Fixing is a separate explicit step (the "read before you fix" rule, [[verify-existing-before-proposing]]).

About & contact

Built and battle-tested at Palo Alto AI Research Lab — a fleet of Claude Code machines running 24/7 as a second brain and synthetic cofounder. Every skill here survived real production use before publication.