# learning-harness A standalone learning app (web + Electron desktop): a **teacher agent** (Next.js; the Vercel AI SDK's `UIMessage` protocol end-to-end, LLM engine [pi-ai](docs/pi-engine-swap.md) via OpenRouter), probe→teach lessons with graded quiz/ask interactions, a formal **session plan tree** the teacher authors once (`plan_init`) and whose progress is derived from tagged quiz calls, and a persistent cross-session **knowledge graph** of what the learner knows — the graph is the app's home page. Not a chat app with a graph bolted on: sessions are probe→teach lessons, the plan tree is a structured, schema-validated object, quiz outcomes reach the graph through the extraction pipeline as ledger events, and the graph feeds back as the teacher's session context (semantic top-K under a token budget) so probing compounds across sessions. - [`docs/architecture.md`](docs/architecture.md) is the **canonical architecture spec** — layers, the import law, the five systems, the invariants, the seam directory. (The pre-audit spec is archived under `docs/archive/` — history only.) - Living design doc (all decisions): [`docs/vision.md`](docs/vision.md) - Build plan + wave/gate results: [`docs/plan.md`](docs/plan.md) ## Commands ```bash npm install # once npm run dev # http://localhost:3000 npm run build # production build (standalone output) npm start # serve the production build npm test # all 44 suites (node:test via tsx; deterministic, no LLM) npm run lint # eslint — includes the import law (architecture §2) npm run typecheck # tsc --noEmit npm run desktop:dev # next dev + Electron shell pointed at it npm run desktop:dist # build + package the desktop app (electron-builder) # one-shot data scripts (run from the master tree only — data never crosses trees) npx tsx --env-file=.env.local scripts/import-notes.ts # bulk-import notes → graph npx tsx --env-file=.env.local scripts/backfill-graph.ts # wipe + re-extract every chat npx tsx --env-file=.env.local scripts/replay-layout.ts # re-derive layout from the ledgers npx tsx scripts/sync-pricing.ts # regenerate core/prices.ts tables ``` ## Environment (`.env.local`) Secrets and machine-specific paths only. Model choice and thinking budgets are the user's, not the environment's — BYOK model slots live in **settings** (`models` section: strong/cheap/vision, provider + model id + optional thinking level; per-role and per-chat overrides ride on top). The extraction batch threshold is a settings knob too (`graph.extractionChunkTokens`, default 4096, floor 500). | Variable | Purpose | |---|---| | `OPENROUTER_API_KEY` | required — all LLM calls go through OpenRouter | | `GOOGLE_SEARCH_API_KEY` + `GOOGLE_CSE_ID` | researcher's `web_search` (Google Custom Search) | | `DEEPGRAM_API_KEY` | optional — dictation (STT token minting; audio streams browser→Deepgram, never through the server) | | `CHROME_PATH` | optional explicit Chrome path for headless verification (puppeteer-core uses system Chrome; nothing is downloaded) | The packaged desktop app takes keys from `userData/auth.json` (BYOK, mode 0600, injected into the server child as env only). ## Architecture Five layers, dependency arrows pointing inward only, machine-enforced as four ESLint zones (shell → barrels only; core and agent never import below themselves). Two composition roots, one writer per store. The full spec — layers, systems, runtime flows, invariants, seams: [`docs/architecture.md`](docs/architecture.md). ```mermaid flowchart TB LEARNER["the learner"] subgraph STAGE["STAGE — presentation shell (browser)"] UI["session UI · question dock · plan panel · graph canvas · activity dock · sidebar"] end subgraph CLASSROOM["CLASSROOM — session engine (owns data/chats)"] TURN["turn pipeline — handleChatRequest:
liveness · steer claiming · write-ahead persist · stream"] AGENTS["prober → teacher agents
(stateless per-turn, built from the record)"] CHATDATA[("chat records
data/chats/*.json")] end subgraph WORKBENCH["WORKBENCH — background labor (owns data/jobs)"] SUBS["researcher · demo-builder · visual-maker
(pi agents driven to completion)"] WAKE["wake pipeline — per-chat serialized steer delivery,
server-side wake turns, crash recovery"] end subgraph LIBRARY["LIBRARY — world model (owns data/graph ledgers)"] EXTRACT["extraction pipeline:
extractor → auditor → merger
(the SOLE graph writer)"] DERIVE["derivation — graphSnapshot · birth-session clusters
(pure reads)"] LEDG[("append-only ledgers
data/graph/*.jsonl")] end subgraph ROOM["MACHINE ROOM — host"] HOST["Electron main → standalone Next server ·
config.ts · LH_DATA_DIR · BYOK keys"] end LEARNER -->|"text · graded answers"| UI UI -->|"POST /api/chat + addToolOutput"| TURN TURN -->|"quiz / ask stream to browser
(client tools, no server execute)"| UI TURN --> CHATDATA TURN -->|"spawn(job) — fire-and-forget ack"| SUBS SUBS -->|"result → steer message"| WAKE WAKE -->|"exactly-once wake turn"| TURN TURN -->|"trigger when uncovered tokens ≥ chunkTokens"| EXTRACT EXTRACT -->|"reads transcript as evidence"| CHATDATA EXTRACT --> LEDG EXTRACT -->|"semantic top-K context injection (one-shot steer)"| TURN LEDG --> DERIVE UI -->|"GET /api/graph · PUT positions · POST import"| DERIVE HOST -.->|"hosts the server side"| TURN ``` **Core invariant:** the extraction pipeline (extractor → auditor → merger) is the sole graph writer — tree-blind and post-hoc, triggered by token coverage, not quiz counts. The LLM stages propose, verify, and judge identity (link-or-create against the current graph; each finding carries verbatim evidence and message-granular provenance); the deterministic merger owns every ledger append and rebuilds anything the model could silently mis-copy (event kinds, timestamps, idempotency refs, provenance) from the transcript. Every node is an objective, world-facing fact; the learner's relationship to it lives entirely in the event ledger. The teacher stays a teacher; everything below the ledgers is derived. ## Layout ``` app/ Next.js App Router — pages + 16 thin API route handlers components/ session + home UI, question dock, plan panel, graph canvas + view, artifact renderers (registry: artifacts/), dictation, activity dock, keyboard machine src/ agent/ system-prompt.md base teaching behavior (the methodology) prompts/{prober,teacher}/ composed numbered fragments per role teacher-agent.ts prober/teacher factories (engine injectable for tests) tools.ts tool composition; quiz / ask_user_question (client tools) tools/plan.ts plan_init (author-once mermaid plan, deterministic parser) tools/probe.ts probe_done (probe→teach handoff seam) tools/jobs.ts research / demo / visual tool surfaces tools/web.ts web_search (Google CSE) + web_fetch (researcher only) subagents/ researcher · demo-builder · visual-maker · pdf-reader, extraction pipeline (extractor/auditor + merger handoff), runner (async protocol), headless Chrome helper, abort registry server/chat-handler.ts THE turn composition root (liveness · steers · persist · stream) server/turn-service.ts turn seams: persist · extraction trigger server/jobs.ts the wake pipeline: serialized per-chat steer delivery, server-side wake turns, crash recovery server/stores.ts THE storage composition root — the one place adapters wire in server/turn-liveness.ts per-chat turn registry (one turn per chat, ever) server/observability.ts the activity feed (running jobs + extraction, abort) server/pi/ the engine bridge: record-adapter (record → pi messages) · ui-stream (pi events → UIMessage SSE) server/index.ts the server barrel (the only server import app/ may use) core/contracts/{plan,graph,chat,jobs,probe,models}.ts the contracts — one file per domain core/ports.ts storage ports — ChatStore · GraphStore · JobStore (core code depends on these types only) core/ question-core, plan-tree, phases, pending, graph (merger/derivation), coverage, layout, clusters, prices, tool-parts — all pure: storage arrives as parameters store/ fs adapters implementing the ports: chats · graph ledgers · jobs (+ render-settings, default-settings) electron/ desktop launcher + security guard (spawns the standalone server; 1 IPC channel; LH_DATA_DIR + BYOK env) scripts/ build-server.mjs (packaging guards), import-notes, backfill-graph, replay-layout, sync-pricing data/ chats/ · graph/ (six ledgers + positions/sources/ render-settings maps) · jobs/ · settings.json — runtime state, git-ignored ``` ## Key mechanisms **Client-side question tools.** `quiz` and `ask_user_question` have no server `execute`: the tool call streams to the browser, the learner answers in the dock, the client grades via `question-core` and returns the verdict with `addToolOutput` — which fires the SDK continuation so the teacher resumes mid-turn. Options are shuffled deterministically (seeded by toolCallId) so server and client renders match. A pending quiz persisted in the chat record is exactly what makes **resume mid-question free**. **The formal plan tree (author-once).** The teacher authors the session's dependency tree with one schema-validated tool — `plan_init`, whose mermaid-subset input is parsed deterministically (line-numbered errors) and validated as a DAG. The tree persists in the chat record; progress is never mutated — `current`/`visited`/upcoming are **derived** from tagged quiz calls in the transcript, and the panel renders that derivation. **The async sub-agent protocol (ack-and-steer).** `research` / `demo` / `visual` are server-executed tools: the call resolves immediately mid-turn with a fire-and-forget ack and the teacher keeps teaching; a background runner works in parallel (the client polls `/api/activity` and `/api/jobs`). On completion the runner builds a **steer message** — a user-role `UIMessage` whose text part is the model-facing result and whose `data-subagent-artifact` part carries the full result for the client — and the **wake pipeline** delivers it: a per-chat serialized queue waits for the turn to end and no question to be pending, then fires a server-side teacher turn that weaves the result into the lesson. Delivery is exactly-once (deterministic steer ids, delivered-flag checks at every claim point); stalled jobs are recovered on read; aborted jobs deliver an abort steer. The client adopts the wake turn through the same `turnInFlight` machinery as returning mid-stream (busy shimmer, composer gate) and renders steers as artifacts — never a bubble. Sub-agents **verify, never one-shot**: the demo-builder headlessly executes its HTML in system Chrome before delivery; the visual-maker renders each SVG and looks at it (vision-capable model — the `vision` slot in settings). **The knowledge graph (two channels, one writer).** The write channel: the extraction pipeline — extractor (LLM: sees only the session, rendered as readable markdown by the shared `core/session-markdown.ts` engine — propositions as true statements, `follows`/`supports` edges with verbatim evidence; fully trusted, graph-blind) → auditor (LLM, an agentic loop that never sees the source: explores the existing graph through `find` / `list_sessions` / `list_propositions` tools, stamps each proposition with an identity verdict — link to an existing node (equivalence bias: paraphrases merge) or create — and may draw bridging edges between existing nodes and the new batch) → merger (`mergeFindings`, deterministic: applies the verdicts, link-appends phrasings/provenance, resolves and dedupes edges, checkpoints coverage) → layout settle. No quiz outcomes or understanding assessments live in the graph (2026-09-27 cleanup); practice (`/practice`) consumes the graph as feedback-only problem sets and writes nothing back. Triggered when a chat's **uncovered tokens** (message-view fingerprints, `coverage.jsonl`) cross `graph.extractionChunkTokens` (a settings knob, default 4096, floor 500 — read fresh per turn); failure checkpoints nothing, the next trigger re-runs. The same pipeline serves the **notes front door** (`POST /api/graph/import`: any `*.md`/`*.txt` becomes a source; re-imports are idempotent by path hash). The read channel: `GET /api/graph` derives the snapshot from the ledgers — a pure read; the home view is a hand-rolled Canvas 2D renderer (one rAF loop, dirty-flag idle, d3-zoom camera, zoom-gated labels, cluster hulls) over a d3-force layout, with server-persisted render settings and drag-committed positions. Clusters are **birth-session groupings**: a node joins the cluster of the session that birthed it (a chat, or an imported notes file as a file-kind session), derived at read — no embedding pass, no naming pass. The teacher reaches the graph on its own terms through the gated prober tools (`list_sessions` / `list_propositions`, the `agent.graphTools` setting) — there is no automatic context injection. **Practice (problem sets).** The graph's `follows`-edges become reasoning chains; the `/practice` creation dialog picks a birth-session cluster (the graph view's own groups) and freezes its propositions + chains into a self-contained `ProblemSet`. Three named-level axes steer generation (ladder rung 0–4, cue distance 0–2, Theory|Application — never a number to interpolate; defaults Direct exercise · Adjacent context · Theory), plus an optional free-form note the agents weigh but need not follow. Creation has two entries: the cluster picker on `/practice`, or the graph view's chain table — every reasoning chain in one selectable overlay whose pinned preview shows the exact `N propositions → M chains` the agents will receive — handing a free-form node selection to the same dialog. A partitioner agent covers every chain exactly once, one generator per partition writes problems against the owner's problem-engineering framework (commit-before-pose: statement + solution + rubric together), and grading is **strengthening, not assessment** — one attempt per problem with its own submit button, instant solution/rubric reveal, then a score 0–10 with brief reasoning landing live; nothing ever written back to the graph. Runs are registry-backed background jobs on the activity dock (no steer, never wakes the teacher); sets live in `data/practice`, crash-swept on boot. **Dictation.** Optional speech-to-text: the server mints short-lived Deepgram tokens; audio streams browser→Deepgram over WebSocket, never through the server. ## Testing All deterministic, no LLM, no network (engines are injected — pi-ai's faux provider scripted per test; `globalThis.fetch` stubbed where transport matters): 1. **Pure core units** — merger (via the real fs adapters, backup-and-restored), derivation, coverage, question grading, plan tree, phases, clusters, layout, tool-parts, prices. 2. **Server flows** — `handleChatRequest` driven end-to-end with scripted engines: persistence, liveness, steer claiming, the wake pipeline (409 retry, quiz deferral, idempotency), crash recovery, the data-dir seam (including a fresh child process). 3. **Sub-agent loops** — researcher/demo-builder/visual-maker with fake headless APIs; the extraction pipeline end-to-end (audit drops, failure-checkpoints-nothing, single-flight, abort). 4. **Render tests** — transcript, cards, docks, composer, sidebar, coverage via React server-side rendering. ## Lineage The teaching methodology comes from the Obsidian-based POC (`~/Desktop/Obsidian/vault/π/learn`; its `knowledge graph design.md` was the original epistemic-model source, since amended — see the amendments table in `docs/vision.md`). The harness was originally built on the pi agent SDK, pivoted to standalone Next.js + the Vercel AI SDK, and later swapped its LLM engine back to pi-ai while keeping the AI SDK's `UIMessage` as the client protocol ([`docs/pi-engine-swap.md`](docs/pi-engine-swap.md)). ## Data Nothing under `data/` is ever tracked by git. Chats, graph ledgers, jobs, positions, and the settings records are plain files (JSON/JSONL) — inspect them freely; they are the canonical records. The desktop app relocates the whole directory via `LH_DATA_DIR`.