Hurricane AIHurricaneAI
← Back to document history
Document version

Hurricane Learner v1

This is the stored snapshot for the approved document version. The diff below shows what changed from the previous version.

Building
Source path
learner/README.md
Source commit
No commit recorded
Created at
Jul 6, 2026, 3:40 PM UTC
Source digest
9cc15770c7c9be6e8bcecc279ce6da42dc82d41b4e649f04009857b7a1966f1c

Document snapshot

Hurricane Learner

Summary

Learner turns sources you trust into knowledge an AI can actually use. Point it at articles, YouTube videos, files, and PDFs on a topic; it digests them into a persistent, per-topic knowledge base (a searchable memory of everything you fed it) and, on demand, synthesizes that knowledge into a portable agent skill — a single SKILL.md playbook that imports straight into Hurricane Sentinel.

Status

Building. Internal only, built from source on Linux with python3.12; there is no public download. It needs a local embeddings/chat backend (Ollama by default) to do real work.

What works today

Learner is a learner CLI over a SQLite knowledge base:

  • Ingest sources on a topic — article_url, video_url (YouTube transcript), file_path (text or PDF), or inline text, one at a time or from a JSON manifest. Each source is loaded, optionally LLM-refined, chunked, embedded, and stored, with distilled knowledge items (rules, workflows, definitions, warnings, examples) extracted alongside the raw chunks.
  • A persistent per-topic knowledge base in ~/.learner (SQLite) — sources, chunks, embeddings, and knowledge items survive across runs. list shows what's stored; refresh re-indexes a topic's sources; export writes a migration bundle.
  • RAG over the KB — query does semantic search; rag-answer retrieves context and answers a question with grounded citations.
  • build-skill — synthesizes a topic's stored knowledge into a SKILL.md playbook (a .skill.md file or a .skill.zip) whose YAML frontmatter is Sentinel-import-compatible, so a learned topic drops straight into an agent.
  • Configurable model routing — talk to a provider direct, or route model calls through a Muzzle listener (muzzle://host:port) so traffic is inspected and keys stay centralized.
  • Configurable embeddings — ollama (default), hf, or openai. The base install stays slim; hf and openai are opt-in install extras (learner[hf] / learner[openai]).

How it fits

Learner is the fourth Hurricane product, and it feeds the other three:

  • Muzzle guards model input/output.
  • Leash guards what an agent can do on the OS.
  • Sentinel runs agents inside both guards, with a human in the loop.
  • Learner builds the knowledge and skills those agents use.

A skill Learner writes is imported by Sentinel and attached to an agent — so the research you do once becomes durable capability an agent carries into every run.

Roadmap

  • Phase 1 — standalone CLI. ✅ This build: ingest → per-topic KB → RAG → build-skill → Sentinel-ready SKILL.md.
  • Phase 2 — Sentinel add-on. A live learner_query tool so an agent can consult a Learner knowledge base mid-run, not just import a static skill.
  • Phase 3 — web console. Drive ingest, review, and skill-building from the Sentinel browser console.

More

Diff from previous

This is the first approved version, so there is no previous diff.