| name | neuro-link |
| description | Main orchestrator for the neuro-link-recursive knowledge/memory control plane. Use this whenever the user says /neuro-link, asks about brain status, wants to trigger a scan, ingest a source, curate a wiki topic, check or dispatch the task queue, edit config, or get a system overview. Also trigger when the user says "what's in the brain right now", "run neuro", "what's pending", or refers to the vault's overall health. This is the primary entry point — it does not do heavy work itself; it dispatches to the right specialized skill (neuro-scan, neuro-surgery, hyper-sleep, crawl-ingest-update, job-scanner, wiki-curate, reasoning-ontology) and composes their outputs. Prefer this over invoking specialized skills directly when the user's intent is ambiguous. |
/neuro-link
Top-level orchestrator. Think of this skill as a dispatcher with taste: it parses the user's intent, picks the right downstream skill(s), assembles their outputs, and reports back with a brief executive summary.
Subcommand map
Match the user's message against this table. If ambiguous, ask a single disambiguating question rather than guessing.
| Pattern | Dispatch to |
|---|
/neuro-link status or "how's the brain" or "is X running" | scripts/status.sh (no subagent needed) |
/neuro-link scan or "scan for issues" or "find gaps" | /neuro-scan |
/neuro-link surgery or "fix the failures" or "repair" | /neuro-surgery |
/neuro-link sleep or "do nightly maintenance" | /hyper-sleep |
/neuro-link ingest <source> or "ingest this paper" | /crawl-ingest-update |
/neuro-link tasks or "what's pending" or "run the queue" | list 00-neuro-link/tasks/*.md + dispatch /job-scanner |
/neuro-link config or "edit neuro-link config" | open config/neuro-link.md frontmatter only |
/neuro-link health or "vault health" | tv_quick_health_check + tv_get_broken_links via MCP |
/neuro-link curate <topic> or "write wiki about" | /wiki-curate |
/neuro-link ontology <domain> or "refresh ontology" | /reasoning-ontology |
Tool preferences
For reads, prefer these MCP tools in this order:
- TurboVault (
tv_*) — for vault-wide lexical search, link graph, frontmatter SQL, centrality. Use tv_search over tv_read_note + grep.
- neuro-link-recursive (
nlr_*) — for schema-aware wiki reads (nlr_wiki_read, nlr_wiki_list) and full RAG (nlr_rag_query, nlr_rag_query_verified).
- Direct file reads (
Read, Glob) — only for infrastructure files (config/, .claude/, server/, scripts/). Never for 02-KB-main/ — use nlr_wiki_read so access is logged.
For writes to knowledge artifacts, there is one rule: always go through the schema-aware tool, never tv_write_note or Write directly.
02-KB-main/*.md → nlr_wiki_create / nlr_wiki_update (frontmatter schema enforced)03-Ontology-main/**/*.md → /reasoning-ontology (maintains both tiers + InfraNodus sync)04-Agent-Memory/logs.md → append via scripts/log.sh only (preserves JSONL ordering)
The why: direct writes bypass schema validation and break the auto-RAG index, which keyword-matches against frontmatter. Silent breakage is worse than a loud tool error.
Auto-RAG delegation
/neuro-link never fetches context itself. The UserPromptSubmit hook at hooks/auto-rag-inject.sh runs before every prompt and injects relevant vault content. This skill reads that injection from the hook output and extends it only if the subcommand needs something the hook didn't grab (e.g., ontology detail for /ontology subcommand).
Output contract
Always end with a 3-line executive summary the user can act on:
Did: <what happened>
State: <current system status after the action>
Next: <one recommended follow-up or "nothing">
This matters because the user typically uses /neuro-link as a status probe while they're doing other work. A 3-line summary fits in their peripheral attention; a long prose reply pulls them away from what they were doing.
When to refuse / escalate
- Destructive requests without
--force: if the user asks to delete a collection, wipe memory, or reset state, refuse and require explicit --force with a description of intent. Log the denial to 04-Agent-Memory/logs.md.
- Unknown subcommands: don't guess — ask "did you mean X, Y, or Z?" referencing the subcommand table above.
- Conflicts between MCP servers: if
nlr_* and tv_* return contradictory results for the same read, report the conflict rather than picking one. This usually signals an index drift (auto-rag hook log in state/llm_logs/ will show when it last indexed).
References
references/tool-namespaces.md — full nlr_* vs tv_* tool map with rationalereferences/subcommand-dispatch.md — deeper dispatch logic including edge cases
Scripts
scripts/status.sh — fast health probe (heartbeat + Qdrant + Neo4j + MCP)scripts/log.sh — append-only log helper for 04-Agent-Memory/logs.md
