Learning System + Knowledge Wiki

New here? This is a reusable local-first learning system, currently mid-migration. The canonical runtime is learning-pi (pi skills, subagents, prompts, verification gate). The Open WebUI layer in this repo is a frozen legacy snapshot — readable, but do not build new behavior on it. Start with the pi setup and the Discussions front door. Cloning this repo gives you the owner's live learning state as an example — to reuse it, build your own empty state with learning-pi/init-state.sh (Path B in Set up below) and fill it with your own material, rather than continuing the AIEFS schedule below.

This repo is the notebook: every session, concept, review, and wiki page lives here as Markdown + JSON. The tutor program that acts on it lives in learning-pi (self-hosted Open WebUI support is frozen, see OPENWEBUI.md).

How learning works here, in 30 seconds: you open the notebook (cd ~/learning-system && pi) and say /lesson (next curriculum lesson) or /teach <topic>. A Scout gathers live sources, the Tutor checks what you know, teaches in small steps with questions, and a Clerk files everything away — session notes, concept schedule, wiki pages. Later, /review quizzes you on whatever is due, mistakes first. A verification gate fact-checks the tutor, so unverified claims never reach you.

Two ways in: chat in your browser via Open WebUI + pi (deploy/, the face in front of the same tutor), or terminal-only pi (cd <notebook> && pi). Same tutor, same notebook — pick per session.

• Repo: https://github.com/delightaheebwa/learning-system
• Track: AIEFS (AI Engineering from Scratch — Rohit 20 phases; Mission 0 Catch-Up done 2026-09-05, P1 L07 done 2026-09-05, P1 L08 Optimization done 2026-09-10, P1 L09 Information Theory done 2026-09-24, P1 L10 Dimensionality Reduction in-progress — paused 5/6, CP5 mini 4 taught 2026-10-16, exit ticket unrun). SWE archived 2026-09-01 (43 concepts paused, strictly out of scope). Prior aie archived 2026-07-28.
• Scheduling: manual — /review reviews what's due, /ingest files new material away, /teach / /lesson / /continue learn. Each lesson pulls live sources (roadmap doc + Further Reading), in the lesson's language (Python/TS/Rust).

Layout

OPENWEBUI.md              setup + operating guide (read this first)
AGENTS.md                 agent context: repo conventions, change protocol, gotchas
scripts/
  setup_openwebui.py        one-shot installer: skills, gate Pipe filter, model presets, prompts, subagent system prompt
  ops.py                    runtime sidecar (state/bundle/apply/attempt)
  learner_history.py        regenerates Core/Learner History.md (tutor context)
  audit_openwebui.py        read-only repo↔live drift audit
  ops_test.py               unittest suite
  wiki_provenance.py        wiki trust-mix report + provenance lint/stamp
  wiki_provenance_test.py   tests for wiki_provenance.py
infra/                     portability: docker-compose.yml, backup/restore, .env.example, VERSIONS.md, ACCESS.md
Learning System/          live state + history
  Core/                     💡 Learning Profile, 📚 Active Concepts (the review schedule),
                            Attempts.json, 🧯 Mistakes.md, Learner History.md (generated),
                            📦 Concept Archive, templates
  Sessions/  Reviews/ (incl. Quality Gates/, Session Audits/)  Lessons/  Learning Records/
  Concept Notes/ (archived stub — do not add)  build-artifacts/  .tmp/ (ephemeral digests)
  Archive/                  ALL legacy material, era-folded (AIE, SWE, C-project, legacy tools, plans, frozen) — do not grep
  AGENTS.md                 behavioral conventions (consistency checks, git sync)
Knowledge Wiki/           curated wiki (raw/sources, raw/assets, wiki/, index.md, log.md, AGENTS.md)
Skills/                   FROZEN Open WebUI snapshot (canonical behavior lives in learning-pi/.pi/skills/)
  learning-system/SKILL.md        review + ingest flows (Clerk runs reviews AND ingests)
  learning-teach/SKILL.md         probe → plan → teach loop (+ fact-check & quiz-audit subagent gates, lesson-end review gate)
  learning-review/SKILL.md        ingest quality gate (review-gate subagent task — standalone ingests AND lesson-end ingests)
  llm-wiki/SKILL.md               wiki building/maintenance rules
  learning-review/openwebui/      gate_pipe.py + gate_schema.py (the deterministic enforcement)

Set up (one-shot per platform)

Two paths. Full prerequisites + verification live in learning-pi/README.md#set-up-one-shot-per-platform.

Path A — clone this repo (owner's live data, as an example):

[ -d ~/learning-pi ] || git clone https://github.com/delightaheebwa/learning-pi.git ~/learning-pi
git clone https://github.com/delightaheebwa/learning-system.git ~/learning-system
~/learning-pi/install.sh ~/learning-system
cd ~/learning-system && pi  # approve/trust once

Start a fresh track rather than continuing the schedule below (you have no write access here anyway).

Path B — your own fresh state (recommended for actual use):

[ -d ~/learning-pi ] || git clone https://github.com/delightaheebwa/learning-pi.git ~/learning-pi
~/learning-pi/init-state.sh ~/my-learning  # empty scaffold, zero owner data, fresh git history
~/learning-pi/install.sh ~/my-learning
cd ~/my-learning && pi  # approve/trust once
# then: git remote add origin <your-url> && git push -u origin main;
# define your track in Learning System/MISSION.md + CURRICULUM.md; run /teach <topic> or /lesson

Dependency one-liners (once git, python3, and mise are present the minimal path above works):

• Omarchy/Arch: sudo pacman -S --needed git openssh python base-devel curl unzip mise deno nodejs npm
• Ubuntu/Debian/WSL: sudo apt-get update && sudo apt-get install -y git python3 curl unzip ca-certificates nodejs npm + curl -fsSL https://mise.run | sh
• macOS: brew install git python mise deno node (after xcode-select --install)
• Windows: not natively supported — use WSL2 (see learning-pi setup section), keep repos on the Linux filesystem.

Daily use (once set up)

cd ~/learning-system && pi  # or ~/my-learning for your own state

Then /review, /solo, /ingest, /teach, /lesson, /continue, /pause, /audit (/solo, /audit, /show are pi-only; /swe below is legacy Open WebUI-only). See learning-pi/README.md for the full usage and AUDIT-ROADMAP.md for what is done (P0–P2, P4 foundation-beats, scout-audit gate, teach rewrite, bridge + deploy) and deferred (P3).

Setup in Open WebUI (one time — frozen legacy)

Run the installer, then start chatting:

OPENWEBUI_API_KEY=sk-... python3 scripts/setup_openwebui.py

That creates/updates: the 4 Skills, the Scout / Learning Tutor / Clerk Model presets, the Gate Pipe Filter, the global subagent system prompt (keyed GATE:), and the 7 legacy Prompts (/swe, /review, /ingest, /teach, /lesson, /continue, /pause — frozen; canonical pi has 9: /review /solo /ingest /teach /lesson /continue /pause /audit /show, no /swe). Verification gates run as foreground subagent tasks (delegate_task, background:false) with Pydantic envelope validation via the Gate Pipe; verifiers run on the calling preset's own base model (Open WebUI has no separate verifier-model setting — verified 2026-09-08 against utils/subagents.py::delegate) — enforcement rests on the deterministic Pipe checks, not model independence. (The legacy review_gate + fact_check + quiz_gate Tools were retired; do not bind them.) After install, models are changed in one place each in the Open WebUI UI — see OPENWEBUI.md. The installer only applies model values on a first install (or when empty); re-running it preserves your UI-chosen models. See OPENWEBUI.md for the full walkthrough, the system prompt, and the prompt texts.

The repo must also be present in the Open Terminal workspace (the model's file workspace) with git push wired — see OPENWEBUI.md and Learning System/AGENTS.md. (Legacy Open WebUI path only — the canonical pi path uses the current checkout + dispatch and needs no Open Terminal workspace.)

Principles

• Clarity is not comprehension
• Prerequisites come first
• Retrieval beats recognition
• Open questions stay open until resolved
• Active Concepts is the persistent memory

Community

This system is built in public, and Discussions is the front door:

• 💬 Discussions — Q&A, ideas, show-and-tell, and the build log.
• 🐛 Issues — concrete bugs and tracked work. Route by layer: pi runtime / gate / skills → learning-pi/issues; state / wiki / Open WebUI snapshot → learning-system issues. If unsure, open in learning-system and say which layer you think it is.
• 🔒 Security reports — see SECURITY.md (do not post publicly).

The pi control layer lives in delightaheebwa/learning-pi, but its conversations happen here too. See CONTRIBUTING.md and the Code of Conduct.

License

MIT © 2026 Delight Aheebwa