nexus-recall

name	nexus-recall
description	The estate's wiki-first recall gate. Before answering, researching, or planning ANY task, deterministically pull grounding from the 2nd-Brain wiki via brain.js — score every estate index (vault OKF + MEMORY.md + skills index) without opening files, open the single best-matching section, and ground the answer in it before reaching outward. Global and reusable; the read-gate counterpart to nexus-copywriter. Fable-5 calibrated — cheap deterministic recall first (low effort), evidence-tagged, context-lean (one section, never a whole-vault dump). Triggers on any question, research request, "what do we know about", "have we decided", "is there prior context", or the start of any plan/spec/decision. Nexus G2 runs it as the deterministic wiki-mine.
disable-model-invocation	false
allowed-tools	Bash, Read, Grep, Glob

nexus-recall

Check the wiki before you think. The estate has accumulated hundreds of decisions, outcomes, and lessons in the 2nd-Brain vault. Answering from the model's memory when the vault already holds the answer is how drift and contradiction get in. This skill makes deterministic wiki recall a standing first step — the read-gate counterpart to nexus-copywriter's content gate.

When it fires (global)

Any task that could turn on prior estate context: a question, a research request, the start of a plan / spec / decision, "what do we know about X", "have we decided Y", "is there prior context". Nexus's G2 (Frame & mine) runs it as the deterministic wiki-mine before any external research.

The recall ladder (deterministic first, cheap first)

Run top to bottom; stop at the first hit. This is the Fable-5 effort discipline — the cheapest step that answers wins, and recall stays context-lean (pull one answering section, never dump the vault).

1. Deterministic index scan (brain.js). One command, ~10-80ms, no files opened until the winner:

   node "$HOME/2nd Brain/2nd Brain/_system/brain.js" find "<the question>"
   

   It scores every estate index line (vault OKF indexes + MEMORY.md + skills/index.md), opens the single best file, and returns its answering section + runners-up. Add --json for the structured {file, section, evidence} when a skill consumes it programmatically.
2. On a miss (weak/empty score) → Grep the vault, then add the missing index line so next time is deterministic (self-healing — the estate standard: a miss must leave the index better). Vault root: ~/2nd Brain/2nd Brain/.
3. Only then reach outward (WebSearch / deep-research) — and when the external answer is durable, it should be ingested back (wiki-ingest) so recall covers it next time.

Fable-5 calibration

• Effort — recall is the low-effort first pass; it must run before any high-effort research so the expensive tier only fires on genuine gaps.
• Evidence tag — recalled context is [VERIFIED] grounded in the returned file path + section (cite it: per Wiki/<file> §<section>). Never restate recalled content as the model's own claim.
• Context-lean (context-editing / compaction aware) — pull the one answering section, not the whole vault; brain.js opens exactly one file. Keeps the working context small by construction.
• Memory-tool aware — MEMORY.md is one of the scored indexes, so durable feedback/decisions are recalled alongside wiki pages in the same pass.

Output contract (for orchestration)

Emit, don't just prose:

recall{
  question,
  hit: true|false,
  source: "Wiki/<file> §<section>"   // the grounding citation, when hit
  evidence: "<the answering section, verbatim-trimmed>",
  tag: VERIFIED,                     // grounded in a real vault file
  go_external: true|false,           // true only on a miss or stale-but-time-sensitive
  index_repair?: "<line added to which index>"  // when a miss was self-healed
}

go_external: false when the wiki answers and the topic isn't time-sensitive — that is the whole point (avoid redundant external calls). go_external: true on a miss or when the wiki answer exists but the data is stale and the question is time-sensitive (then supplement, don't replace).

Hard rules

1. Recall before research. No external call on an estate-context question until brain.js has run.
2. A miss must repair the index. Grep-found answer → add the index line; never leave the miss.
3. Cite the grounding. Recalled facts carry per Wiki/<file> §<section>, never ungrounded.
4. One section, not the vault. Never dump whole files/folders into context — brain.js returns the one answering section by design; honour it.
5. Don't rebuild retrieval. brain.js is the engine; this skill orchestrates it. wiki-query (Margot-scoped) and wiki-ingest (write-back) are the siblings — this is the global read-gate.

Verification (this session, 2026-07-08)

brain.js find "fable free model routing" → scored 1812 index entries, opened 1 file, returned Wiki/opus-fable-mode-skill-model-routing-herk-… §"Model routing table" in 78ms. --json returns {file, section, evidence}. Engine confirmed working.

Provenance

Built 2026-07-08 to make the estate's existing deterministic-retrieval engine (brain.js, see memory [[second_brain_deterministic_retrieval]]) a global, always-on recall gate usable inside Nexus — rather than a CLI only some flows remember to run. Standard: vault Wiki/second-brain-standard.md.