neuro-link

Non-HITL nightly brain maintenance. Use this whenever the user says /hyper-sleep, schedules overnight maintenance, or when triggered by cron during the configured sleep window. Also trigger when the user says "run maintenance", "do background cleanup", "catch up while I sleep", or hands off work expecting it done by morning. Handles the subset of findings from /neuro-scan that don't need human judgment: auto-ingests proposed sources, synthesizes with a hard confidence cap at 0.6, refreshes ontologies for topics with ≥3 new pages, executes folder-delegated low-risk improvements, warms the qmd embedding cache, and compacts agent memory logs older than 7 days. Hard 4-hour timeout, no deletions, no network access outside the configured source list, no edits above confidence=0.6 without HITL.

HITL-driven repair of the knowledge base, operating on the latest /neuro-scan report. Use this whenever the user says /neuro-surgery, asks to fix failures, resolve ontology inconsistencies, re-synthesize stale wiki pages, coordinate recursive-self-improvement execution, or review deep-reasoning proposals. Also trigger when the user says "fix the scan report", "clean up the failures", "audit contradictions", or wants to work through the backlog interactively. Every single write to 02-KB-main/ or 03-Ontology-main/ requires explicit user approval per-change — never batch approvals, never apply silently. Skilled at diffs: shows before/after side-by-side, waits for thumb up, then applies via schema-aware tools (nlr_wiki_update / reasoning-ontology — never bare Write or tv_edit_note on schema-bound paths).

Interactive guided first-run bootstrap for the neuro-link-recursive system. Use this whenever the user says /neuro-link-setup, asks to "set up neuro-link", "configure the brain", "bootstrap the vault", "install neuro-link", or when state/heartbeat.json shows status=initialized. Also use when the user says the system isn't working on first install, when MCP servers aren't reachable on fresh clone, or when they need to reconfigure secrets/models/hooks. Walks conversationally through prereqs, vault structure verification, secrets population, Octen/Qwen3 model downloads, MCP server registration, hook installation, skill generation, and first-run verification. Never auto-installs missing binaries — always prints the exact command so the user can review before running.

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.

Maintain a comprehensive toolbox wiki for third-party repos the user or their agents use often but do not own (vectorbt.pro, nautilus_trader, qmd, turbovault, optuna, ray, pufferlib, polars, etc). Use this whenever the user says /adjacent-tools-code-docs, asks to document a third-party tool, mentions a new dependency they will rely on regularly, or wants a local searchable wiki for an external library. Also trigger when the user says "add X to the toolbox", "document this library", "keep track of this tool", or the user asks questions about a library that isn't yet in 08-code-docs/toolbox/. Follows the Karpathy LLM-Wiki pattern with YAML frontmatter, citations, open questions, and contradictions. Keeps synced with upstream via doc-sync-embed-verify on releases. Outputs to 08-code-docs/toolbox/<tool-name>/.

Source ingestion pipeline that deposits raw content into 01-raw/ (SHA256-deduped), classifies into 01-sorted/, and optionally synthesizes wiki pages to 02-KB-main/. Use this whenever the user says /crawl-ingest-update, /crawl-ingest, asks to ingest a paper/book/repo/video/URL/vault, or drops a link expecting it to be added to the brain. Also trigger when the user says "grab this for the brain", "add this source", "read this paper into the KB", or forwards a PDF/URL expecting it indexed. Supports web pages, web crawls, GitHub repos, YouTube transcripts, standard PDFs, math-dense PDFs (via Marker v1.10+ + MinerU 2.5 MLX cross-check), arXiv papers (prefers ar5iv HTML → falls back to arXiv S3 LaTeX → last resort Marker on the PDF), math books, Stacks Project chapters (JSON API), nLab dumps, Wikipedia math bulk dumps, and Obsidian vault imports. MathWorld scraping is blocked (Wolfram TOS). Always preserves $$...$$ blocks byte-for-byte; canonicalizes equations via pylatexenc + latex2sympy2_extended + SymPy srepr.

Track repos the user has forked under the HyperFrequency org or personal account, maintaining semantic understanding of the functional diff between upstream and fork, and generating enriched docs for fork-specific functionality. Use this whenever the user says /forked-repos-with-changes, asks to track a fork, add a fork, review functional differences between upstream and fork, or hands off a forked repo expecting it documented. Also trigger when the user says "track this fork", "what did we change upstream", "generate fork docs", or mentions a fork repository in the context of adding it to the brain. Outputs to 08-code-docs/forked-up/<repo>/ and, for HyperFrequency org forks, feeds the shared Devin DeepWiki via doc-sync-embed-verify on gh push. Uses LLM-based semantic diff rather than git diff to produce readable functional differences, and pins the fork_point + last_merge_base in frontmatter so resync catches upstream changes that haven't been rebased.

Register the user's main codebases for context-aware retrieval and auto-RAG injection. Use this whenever the user says /main-codebase-tools, /main-codebase-tools add <repo>, asks to register a repo for context injection, or adds a new project they'll be working on regularly. Also trigger when the user says "track this repo", "add this to auto-rag", "index my codebase for Claude", or mentions a new project they expect Claude to know about. Performs full indexing via Context7 (API surface / doc snippets) and Augment Code / Auggie (semantic understanding), registers the repo's primary topics with the /auto-rag keyword router, subscribes fork repos to doc-sync-embed-verify for upstream change tracking, and generates a short CLAUDE.md addendum with architecture and gotchas. Each registered repo gets one spec file in this skill's directory with frontmatter pinning the indexed version.

Brain scanner that produces a structured health report and queues remediation tasks. Use this whenever the user says /neuro-scan, asks to check pending jobs, find knowledge gaps, scan for stale wikis, audit failures in hooks/skills/cron/harness-to-harness comms, or when scheduled hourly by cron. Also trigger when the user says "what's broken", "audit the brain", "what should I work on next", or is about to start a work session and wants situational awareness. Runs six independent scan passes — pending jobs, self-improvement proposals, failure logs, knowledge-graph gaps, stale wikis, and upstream doc drift — then writes a dated report to 06-Recursive/daily.md and auto-queues remediation tasks with source=neuro-scan to 00-neuro-link/tasks/.

Automated self-improvement loop with consortium grading and HITL gates. Use this whenever the user says /recursive-self-improvement, asks to run the improvement cycle, review consortium grades, check performance delta, or when this is triggered by scheduled cron. Also trigger when the user says "grade the agents", "see how we're doing", "what can be improved", or is about to make significant changes to prompts/skills/ontologies and wants a data-backed recommendation. Runs a 3+ agent consortium (never single-grader — that's overfit to one reviewer's biases), majority-votes on proposals, surfaces the proposal batch to the user for HITL approval (default manual), and only then executes approved adjustments to agent topology, ontologies, skills, or prompts. Never auto-applies changes to 02-KB-main or 03-Ontology-main without explicit user approval.