learnharness/docs/architecture.md

Learning Harness — Architecture

Status: the canonical architecture spec — describes the as-built system, verified against the code by a full audit on 2026-09-18 (every file under src/, app/, components/, electron/, scripts/, the lint config, and the test suite; lint and typecheck clean on the audited tree). The pre-audit spec is archived at archive/architecture-2026-09-08.md — history only. Product reasoning lives in docs/vision.md; file-by-file layout lives in the README.

What this is. A standalone learning app (web + Electron desktop): the learner's knowledge accumulates into one persistent, cross-session knowledge graph, built from tutoring sessions in which a teacher agent authors a formal, visible dependency tree (the plan tree) and verifies understanding with structured questions. The graph is the home page; the session is the classroom; everything else serves those two.

Design principles, in priority order:

1. Maximum control. Every dependency edge is explicit and machine-checked. Nothing reaches sideways for convenience.
2. Easy editability. A worker can trace any behavior to one file. Layers are named by responsibility, not by framework accident.
3. Zero technical debt. The shape you see is the shape that runs. Deprecated paths are deleted, not parked.
4. Crystal-clear architecture. Two barrels as the entire public API, two composition roots, one writer per store, named seams. When a rule can be a lint rule, it is.

The one-sentence model: a pure domain core at the center; every external effect (fs, LLMs, browser, OS) behind an adapter at the edge; two composition roots as the only places the edges are plugged together; and every persistent store written by exactly one system.

---

1. Two axes, one law

Hold the architecture along two orthogonal axes:

Static structure — the hexagon. One pure core (src/core); everything else is an adapter (implements a core port or drives the core) or a composition root (the only code that wires adapters to core). Dependency arrows point inward only — toward purity, toward stability. Two barrels are the entire public API: @/src/core (browser-safe) and @/src/server (the server-side surface).

Runtime ownership — six systems. Each persistent store has exactly one owning writer-system; systems talk only through named, idempotent protocols:

System	Owns (sole writer)	Role
Classroom — session engine	data/chats/*.json	the turn pipeline, prober→teacher, plan tree, questions, phases
Workbench — background labor	data/jobs/*.json	job runner, researcher/demo-builder/visual-maker, steer delivery
Library — world model	data/graph/* ledgers	extraction pipeline (sole graph writer), derivation, birth-session clusters
Practice — problem sets	data/practice/*.json	partitioner→generator→grader pipeline, self-contained sets (§7)
Stage — presentation shell	nothing server-side	components/, graph canvas, docks — pure consumer via /api/*
Machine room — host	process + config	Electron main, standalone Next server, config.ts, data-dir seam, BYOK keys

The unifying law (generates everything else): imports point inward, writes point to append-only evidence, and everything else is derived. Dependency flows toward purity; data flows toward evidence; no state is stored where it can be recomputed.

---

2. The layers and the import law

Dependency arrows point down only. Each layer may import the core barrel; nothing skips a boundary for convenience.

┌──────────────────────────────────────────────────────────────┐
│ SHELL      app/            Next.js route handlers (transport) │
│            components/     session + home UI (render only)    │
├──────────────────────────────────────────────────────────────┤
│ SERVER     src/server      composition roots (stores.ts,      │
│            chat-handler.ts) + named service seams + the      │
│            pi engine bridge. Zero business logic              │
├──────────────────────────────────────────────────────────────┤
│ AGENT      src/agent       prober + teacher loops, tool       │
│            surfaces, prompt fragments, sub-agents, the       │
│            extraction pipeline. Deps arrive via factories —  │
│            never imported from below                          │
├──────────────────────────────────────────────────────────────┤
│ CORE       src/core        pure domain: contracts (6 domains),│
│            ports, merger + derivation, coverage, question-   │
│            core, plan-tree, phases, layout, prices. No IO     │
│            (two files use node:crypto — pure                  │
│            but node-bound, hence kept out of the barrel)      │
├──────────────────────────────────────────────────────────────┤
│ STORE      src/store       fs adapters: chats · graph        │
│            ledgers · jobs (implement core ports) + two       │
│            portless settings adapters. The ONLY modules     │
│            that touch data/                                   │
└──────────────────────────────────────────────────────────────┘

The import law (eslint.config.mjs — the config's sole job; four import/no-restricted-paths zones, all "error"; each verified to fire; lint is a gate on every change):

Zone	May import	Rationale
components/**	@/src/core only (the barrel)	The UI renders domain data; it never runs domain logic or touches IO
app/**	@/src/core + @/src/server (the barrels)	Route handlers are transport; they delegate to server entry points
src/core/**	nothing below itself (never agent/server/store)	The core is pure; adapters implement its ports, never the reverse
src/agent/**	never server/store	The agent layer receives its dependencies (ports, engines, seams) via factories/closures

flowchart TB
    subgraph SHELL["SHELL — render + transport, no domain logic"]
        APP["app/ — pages + 28 API route handlers"]
        COMP["components/ — session + home UI, graph canvas"]
    end

    subgraph SERVER["SERVER — src/server — composition + seams"]
        SBAR["barrel index.ts — THE server surface"]
        ROOTS["composition roots: stores.ts (storage) · chat-handler.ts (turn)"]
        SEAMS["service seams: turn-service · jobs (wake) · turn-liveness · commit-positions · observability · model-catalog · settings-body"]
        PIB["pi/ — engine bridge: record-adapter (record → pi messages) · ui-stream (pi events → SSE)"]
    end

    subgraph AGENT["AGENT — src/agent — LLM loops, deps injected"]
        TEACH["prober + teacher agents · tools · prompt fragments"]
        SUBS["sub-agents: researcher · demo-builder · visual-maker"]
        PIPE["extraction pipeline"]
    end

    subgraph CORE["CORE — src/core — pure domain, zero IO"]
        CONTRACTS["contracts/ — 6 domains: plan · graph · chat · jobs · probe · models"]
        PORTS["ports.ts — ChatStore · GraphStore · JobStore (types only)"]
        BRAIN["merger + derivation · coverage · question-core · plan-tree · phases · layout · prices · tool-parts"]
        CBAR["barrel index.ts — browser-safe exports ONLY"]
    end

    subgraph STORE["STORE — src/store — fs adapters, only data/ writers"]
        ADP["chats.ts · graph.ts · jobs.ts (implement core ports)"]
        PORTLESS["render-settings.ts (portless)"]
    end

    CFG["config.ts — leaf facts: PATHS · LEDGER_FILES · engine-construction facts (no models — BYOK slots own those)"]

    COMP -->|"lint zone: @/src/core only"| CBAR
    APP -->|"lint zone: @/src/core + @/src/server"| CBAR
    APP -->|"lint zone: @/src/core + @/src/server"| SBAR
    SBAR --> ROOTS
    SBAR --> SEAMS
    ROOTS -->|"instantiates + passes ports"| ADP
    ROOTS -->|"factory calls, ports injected"| TEACH
    ROOTS -->|"factory calls, ports injected"| PIPE
    SEAMS --> PIB
    TEACH --> CONTRACTS
    TEACH --> BRAIN
    SUBS --> CONTRACTS
    PIPE --> BRAIN
    ADP -->|"implement"| PORTS
    ADP --> CFG
    CBAR --> CONTRACTS
    CBAR --> PORTS
    CBAR --> BRAIN
    SERVER --> CFG
    AGENT --> CFG

    CORE -.->|"✗ FORBIDDEN (lint)"| SERVER
    AGENT -.->|"✗ FORBIDDEN (lint)"| SERVER
    COMP -.->|"✗ FORBIDDEN (lint)"| SERVER

Two rules are documentation-only by design. The ESLint plugin's except can't express "all but one", so (a) src/server/stores.ts is intended to be the only server module importing adapters — but the server barrel also re-exports the two portless settings adapters and observability.ts imports chatStore directly; and (b) the browser-safety of the core barrel is a doc-comment convention (see §2.1). tests/, scripts/, and electron/ are deliberately outside the law (lint-excluded); tests import freely across layers, scripts act as extra composition roots.

2.1 The core barrel and browser-safety

src/core/index.ts exports only browser-safe modules — pure logic + types, no node builtins, no fs, no network. Deliberately not barrel-exported (server code imports them directly): graph.ts and coverage.ts (need node:crypto), clusters.ts, session-clusters.ts and layout-settle.ts (server-only roles). Port types are types-only, hence safe to re-export. This split is what lets the same core serve browser (grading, layout, plan statuses, cost folding) and server (merger, derivation, coverage).

2.2 Contracts — one file per domain

src/core/contracts.ts is a barrel over seven domain files, each owned outright by its stream (schema changes in a domain you don't own go through the master session):

• contracts/plan.ts — plan tree types + derived statuses (derivePlanStatuses)
• contracts/graph.ts — graph domain: ledgers, findings/audit shapes, snapshot, birth-session clusters
• contracts/chat.ts — session records, usage/money channel, settings, handoff text
• contracts/jobs.ts — the ack-and-steer protocol: job records, steer message builders, deterministic steer ids, usage accumulation
• contracts/probe.ts — probe assessment + the shared zod schema
• contracts/models.ts — client-safe model-catalog shapes
• contracts/practice.ts — problem sets, reasoning chains, params, grade verdicts

---

3. Ports, adapters, and the composition roots

Storage is per-entity semantic ports, never one generic interface: ChatStore, GraphStore, JobStore, PracticeStore (src/core/ports.ts — types only). The fs adapters (src/store/{chats,graph,jobs,practice}.ts) implement them and contain persistence and nothing else. One more adapter (src/store/render-settings.ts) is a portless free function re-exported through the server barrel — a deliberate pragmatic bypass, documented here so nobody mistakes it for the pattern.

Idioms. Graph evidence lives in append-only JSONL ledgers (corrupt-line-tolerant reads; strict shape guards drop invalid lines, never guess). Chats, jobs, positions, sources, and the settings blobs are whole-file read-modify-write JSON. All writes are synchronous, single-process (atomicity rides the Node event loop). Merge/dedup/derivation semantics are reader-side, in core — the store never decides.

• src/server/stores.ts is THE composition root for storage: it instantiates the adapters and hands the port objects to everything else (chat-handler, job runner, extraction pipeline, practice pipeline, graph/positions/chats routes, turn services). It is the single place a storage decision (JSON vs SQLite, local vs remote) gets re-made.
• src/server/chat-handler.ts is THE composition root for a turn: it wires the stores, the job runner, the extraction pipeline, the practice pipeline, and the pi engine into the stream response (setTurnRunner/setJobRespawn register its seams for the wake pipeline). All logic lives behind the seams it calls.
• The agent layer is dependency-inverted throughout: createJobRunner({ chatStore, jobStore, deliver }), createExtractionPipeline({ chatStore, graphStore, extractor?, auditor?, embedder }), createTeacherTools({ chatId, chatStore, jobStore, spawn }), createProberAgent/createTeacherAgent(engine?, tools?, …). Tests inject the same ports production wires — or swap an in-memory port.

The engine side is adapted too: src/agent/models.ts (piEngineFor) resolves each role's engine from the user's model slots (BYOK: settings.json models section — strong/cheap/vision; a chain that ends empty refuses honestly with 503 models_not_configured) — and the teacher/prober conversation runs the SESSION'S OWN record: POST /api/chats stamps every newborn verbatim from the strong slot (the new-session default, chosen in setup), in-session picks write the session's record (PUT /api/chats/:id/settings), and the strong slot never reaches an existing session again (legacy unstamped records fall back to it at resolve time) — per-session models, owner 2026-09-28. Beyond the conversation, src/server/pi/ bridges the two SDK worlds: record-adapter.ts (persisted UIMessage record → stateless pi Message[], dangling tool calls get synthesized INTERRUPTED results) and ui-stream.ts (pi agent events → AI-SDK UIMessage SSE).

flowchart LR
    subgraph DRIVING["DRIVING ADAPTERS — inbound"]
        BROWSER["Browser UI — components/: question dock (addToolOutput) · graph canvas · docks"]
        HTTP["HTTP transport — app/api/*: 28 route handlers"]
        NOTES["Notes front door — POST /api/graph/import"]
        ELEC["Electron main — thin launcher, spawns server, LH_DATA_DIR + BYOK env"]
    end

    subgraph HEXCORE["THE CORE — src/core (pure domain)"]
        direction TB
        SESS["session brain: phases · plan-tree · question-core · pending · steer builders"]
        GBRAIN["graph brain: merger + derivation · coverage · clusters · layout"]
        PORTSBOX["PORTS (types only): ChatStore · GraphStore · JobStore · PracticeStore"]
    end

    subgraph DRIVEN["DRIVEN ADAPTERS — outbound"]
        FS["fs stores — src/store (data/)"]
        LLM["LLM engines — pi-ai/OpenRouter via piEngineFor"]
        HEAD["headless Chrome — demo verify · SVG render"]
        DEEP["Deepgram STT — token-minting route, browser-side WebSocket"]
        WEB["web tools — Google CSE + Readability/Jina (researcher)"]
    end

    BROWSER --> HTTP
    NOTES --> HTTP
    ELEC --> HTTP
    HTTP --> SESS
    HTTP --> GBRAIN
    SESS --> PORTSBOX
    GBRAIN --> PORTSBOX
    PORTSBOX --> FS
    SESS --> LLM
    GBRAIN --> LLM
    SESS --> HEAD
    SESS --> WEB
    BROWSER --> DEEP

---

4. The six systems (runtime ownership and protocols)

The systems view: each box owns its store; edges are the protocols, each with the adapter/seam that implements it.

flowchart TB
    LEARNER["the learner"]

    subgraph STAGE["STAGE — presentation shell (owns nothing server-side)"]
        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"]
        AGENTS["prober → teacher (stateless per-turn, from the record)"]
        CHATDATA["chat records — data/chats/*.json"]
    end

    subgraph WORKBENCH["WORKBENCH — background labor (owns data/jobs)"]
        JOBRUN["job runner"]
        SUBAGENTS["researcher · demo-builder · visual-maker"]
        WAKE["wake pipeline — per-chat serialized queues on globalThis"]
    end

    subgraph PRAC["PRACTICE — problem sets (owns data/practice)"]
        PPART["practice pipeline — partitioner → one generator per problem · one grader per answer<br/>(feedback-only; never steers the teacher)"]
        PDATA["problem sets — data/practice/*.json"]
    end

    subgraph LIBRARY["LIBRARY — world model (owns data/graph ledgers)"]
        EXTRACT["extraction pipeline — extractor → auditor → merger (SOLE graph writer)"]
        DERIVE["derivation — graphSnapshot · birth-session clusters (reads, never write)"]
        GRAPHDATA["append-only ledgers — data/graph/*.jsonl"]
    end

    subgraph ROOM["MACHINE ROOM — host"]
        HOST["Electron main → standalone Next server · config.ts · LH_DATA_DIR · BYOK"]
    end

    LEARNER -->|"text · answers [transport adapter]"| UI
    UI -->|"POST /api/chat [question answers ride addToolOutput]"| TURN
    TURN -->|"quiz / ask_user_question stream to browser [client-question tools, no execute]"| UI
    TURN -->|"saveMessages · persistTurn [ChatStore port]"| CHATDATA
    TURN -->|"spawn(job) — fire-and-forget ack [job tools + JobStore port]"| JOBRUN
    JOBRUN --> SUBAGENTS
    SUBAGENTS -->|"result → steer message [steer builders in core/jobs.ts]"| WAKE
    WAKE -->|"exactly-once wake turn [turnRunner seam → same handleChatRequest]"| TURN
    TURN -->|"triggerExtraction when uncoveredTokens ≥ chunkTokens [trigger seam, single-flight]"| EXTRACT
    EXTRACT -->|"reads transcript as evidence [ChatStore port]"| CHATDATA
    EXTRACT -->|"append-only ledger writes [GraphStore port — sole writer]"| GRAPHDATA
    DERIVE -->|"pure read: snapshot for home [graphSnapshot]"| GRAPHDATA
    UI -->|"POST /api/practice/sets · answers [HTTP transport]"| PPART
    PPART -->|"reads the graph ONCE at creation — propositions+chains frozen into the set [GraphStore port, pure read]"| GRAPHDATA
    PPART --> PDATA
    UI -->|"GET /api/graph (read) · PUT positions (sanctioned exception) · POST import"| DERIVE
    HOST -.->|"hosts every server-side system"| TURN

Four server-side systems, each the sole writer of its own store, connected by named protocols: the turn request (Stage→Classroom), the ack-and-steer loop (Classroom→Workbench→Classroom), the extraction trigger + transcript read (Classroom→Library), and the practice request (Stage→Practice over HTTP, plus Practice's one-time graph snapshot read at creation). The Stage touches systems only through the thin HTTP transport; the teacher reaches the Library directly through the gated prober tools (list_sessions / list_propositions) rather than any injected context.

---

5. The session (one chat turn)

handleChatRequest is framework-free (plain Request/Response) with a buildAgent test seam so tests drive it with scripted engines. The turn, in exact order:

sequenceDiagram
    participant B as Browser (useChat)
    participant R as POST /api/chat
    participant H as handleChatRequest
    participant L as turn-liveness (globalThis)
    participant CS as chatStore adapter
    participant A as prober/teacher agent (pi engine)
    participant F as onFinish seams

    B->>R: messages (+ graded question output via addToolOutput)
    R->>H: handleChatRequest(req)
    H->>L: beginTurn(chatId)
    Note over H,L: already running → 409; retryingFetch waits, re-fetches<br/>the record, merges unsent client messages by id, re-POSTs
    H->>H: claimSteers — coalesce queued steers into outgoing
    H->>CS: saveMessages (write-ahead: persist even before the stream)
    H->>A: build agent from record (probe_done part in transcript ⇒ teacher)
    Note over H,A: recordToPiMessages: stateless rebuild; dangling<br/>tool calls get synthesized INTERRUPTED results
    A-->>B: SSE stream (ui-stream bridge)
    Note over B: quiz / ask_user_question have NO server execute:<br/>rendered in the dock; graded client-side via question-core
    B->>B: QuestionDock: parse → shuffle (seeded) → buildAnswer (grade)
    B->>R: continuation (sendAutomaticallyWhen fires)
    A-->>F: onFinish
    F->>CS: persistTurn (canonical, idempotent) + appendTurnUsage
    F->>F: triggerExtraction — if uncovered tokens ≥ chunkTokens, fire async
    F->>L: endTurn (ALWAYS last — the registry never locks)

• Client-side question tools (quiz, ask_user_question) have no server execute: the 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 continuation. A pending quiz persisted in the record is exactly what makes resume mid-question free. Deterministic seeded shuffling (keyed by toolCallId) keeps server and client renders identical.
• The prober→teacher handoff: the first turn(s) run the prober (floor/ceiling via quiz, goal via ask); probe_done (server-executed) files the assessment and queues a handoff steer; from the first tool-probe_done part in the transcript, the handler builds the teacher agent for every later turn. Phases (core/phases.ts) fold strictly: probe → plan → teach.
• The formal plan tree is author-once: the teacher authors the session's dependency map with a single schema-validated plan_init tool (a mermaid-subset graph, parsed deterministically with line-numbered errors, validated as a DAG). Progress is never mutated — statuses (current/visited/upcoming) are derived from tagged quiz calls in the transcript; the tree panel and the plan blocks render that derivation.
• Job tools are server-executed, ack-and-steer: research/demo/visual resolve immediately mid-turn with a fire-and-forget ack (the teacher keeps teaching); spawn(job, input) fires the background runner.
• Each turn service is a named, error-contained seam taking explicit deps; failures log one [chat <id>] line and never touch the response. Swapping any of them touches nothing else.

5.1 The wake pipeline (Workbench → Classroom)

On sub-agent 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 (provider conversion drops data parts) — and hands it to the wake pipeline:

flowchart TB
    R["job runner finishes sub-agent"] -->|"deliver(steer) — injectable seam"| ENQ["enqueue steer in per-chat queue (globalThis)"]
    ENQ --> PUMP["pump(chatId) — one loop per chat (pumping set serializes)"]
    PUMP --> GATE{"turn-free AND no pending question? (400ms poll)"}
    GATE -->|no| PUMP
    GATE -->|yes| REFRESH["re-fetch fresh record, re-check steerDelivered (claim race)"]
    REFRESH --> RUN["turnRunner(chatId, record + steer) — the SAME handleChatRequest path"]
    RUN -->|"200"| FIN["finalizeSteer → finalizeJob (done/error)<br/>— status flips only after the wake turn persists"]
    RUN -->|"409"| BACKOFF["backoff 500→1000→2000→5000ms → retry"]
    CRASH["crash recovery: listJobs → recoverStaleJobs (stale >15min)<br/>re-deliver results / respawn via setJobRespawn"] --> PUMP

Job status flips done/error only after the wake turn persists. Delivery is exactly-once: deterministic steer ids (steer_<jobId>, steer_probe_<chatId>), steerDelivered checks at every claim point, and claimSteers finalizing in-transcript steers instead of re-appending. The client adopts the wake turn through the same turnInFlight machinery as returning mid-stream (busy shimmer, composer gate, append-only mergeAdopted adoption) 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 (needs a vision-capable model — the vision slot in settings).

---

6. The knowledge graph — two channels, one writer

The extraction pipeline is the sole graph writer. Nothing else appends to data/graph/*.jsonl — not the teacher, not the server, not a migration script. (The one sanctioned exception: PUT /api/graph/positions writes layout positions only, sanitized, with an incremental layout settle when the graph grew. Backfill scripts bypass this by design, as one-time operations.)

flowchart TB
    TRIG["trigger seam (turn-service): coverageOf(record, coverageLedger)<br/>— uncovered tokens ≥ graph.extractionChunkTokens (settings)?"] -->|"yes, single-flight per source"| RUN["extraction.runExtraction(chatId)"]
    IMPORT["POST /api/graph/import (notes front door)<br/>deterministic src id = sha256(path); content-fp guard skips unchanged re-imports"] -->|"importNotes: shared pipeline, second extractor prompt"| RUN
    RUN --> BATCH["batchRuns — greedy token-budgeted batches of oldest uncovered TEACHER-PROSE messages"]
    BATCH --> EXT["EXTRACTOR (LLM) — sees ONLY the session markdown view<br/>(tree-blind, graph-blind; shared engine: core/session-markdown.ts)"]
    EXT --> AUD["AUDITOR (LLM, agentic loop) — sees findings ONLY, never the source;<br/>explores the graph via find / list_sessions / list_propositions tools,<br/>judges identity ONLY (link-or-create, equivalence bias) — never mints edges"]
    AUD --> MERGE["MERGER (deterministic, core/graph.ts) — the ONLY ledger writer:<br/>identity via validated linkTo · dedupe edges · coverage checkpoint"]
    MERGE --> SETTLE["settleGraphLayout — resting positions ledger"]
    MERGE -->|"failure at ANY stage"| NOTHING["checkpoints nothing — the next trigger re-runs"]

Raw reasoning chains only (2026-09-27 cleanup; teacher-prose view 2026-09-28): the extractor reads ONLY the session — and since the 2026-09-28 ruling, ONLY the teacher agent's prose: plain paragraphs under the session title, no learner messages, no thinking, no tool calls or results, no quiz blocks. Coverage fingerprints, slot tokens, and batch budgets all derive from that same selection (messageViewLines in core/coverage.ts), so "covered" is exactly "what the extractor saw" — a quiz answer or learner edit can never open a run, and only a teacher-prose edit re-opens its message. The old call+result inseparability bond is deleted (nothing left to keep together); batching is pure greedy fill. The shared markdown engine (core/session-markdown.ts) keeps the full session format for the future user-facing export and is no longer used by extraction. The extractor proposes free-form propositions (true statements; a misconception is stated as its true correction) and follows/supports edges. The submit schemas are slim (owner, 2026-09-28): propositions carry id/name/statement/optional aliases and edges carry from/to/type — the messages and evidence fields are gone (provenance is chat-level; quote fields were never persisted and weak models burned retries on them), and a prepareArguments heal at the pre-validation seam rescues stringified-payload submits. Every node is an objective, world-facing fact. No quiz outcomes, no understanding assessments — the events ledger and UnderstandingLevel are deleted; the learner's relationship to a proposition is practice territory — which landed 2026-09-29 as feedback-only strengthening (§7): grades live on problem sets, and nothing ever writes learner states back to the graph.

Serialized audit, parallel extracts — per-chunk jobs (owner ruling 2026-09-28): every chunk (notes file) and every batch (chat coverage) is its own JOB — registry row, usage record, failure domain. Extracts fan out in parallel across all jobs and sources, throttled by a process-wide 3-permit extract semaphore on globalThis (the provider-facing rate stance; a job waiting for a permit shows queued). The audit→merge section remains a process-wide FIFO critical section — a mutex on globalThis (the same module-instance idiom as the run registries): before any auditor LLM call or graph read, a job waits (waiting phase) until every earlier job's merge has landed, so each auditor explores the graph as it actually stands — two concurrent audits would each link into a graph the other is about to change, the structural mint of duplicate nodes. Failure isolation: one job failing (bad chunk, degraded model output) is absorbed at the fan-out — its messages stay uncovered and honestly re-runnable while its siblings still merge; the run returns what actually merged. Abort is source-wide: every job of a source's fan-out dies together, and merged jobs keep their checkpoints. The status feed shows one row per job (file.md · 2/5 suffixes for chunked sources), and each job persists its own usage record (failed jobs bill their partial spend too).

Division of trust (the pipeline's core rule): the extractor is fully trusted and is the only stage that ever sees the source. The auditor never sees the transcript — it receives the findings JSON alone and explores the existing graph through read-only tools (find, list_sessions, list_propositions — statement text included, no timestamps) to reach its verdicts. VERDICTS-ONLY (owner ruling 2026-09-28): the auditor's submit is { links } — identity links only (which findings propositions ARE existing nodes). The edges argument is GONE (owner, same evening): the auditor never mints edges — cross-source connection happens solely through identified equivalences, and relations between propositions come from the extractor alone. It never restates the findings; the merger joins findings + verdicts, so an order of magnitude less is emitted and there is no structural way for the auditor to alter or drop extractor output. Equivalence bias (owner ruling 2026-09-27): paraphrases, rewordings, and compressions of the same claim MERGE; only genuinely distinct propositions stay apart. The deterministic merger validates and appends: link resolution, edge dedupe by (from, to, type), coverage checkpoints. There is no drop channel — nothing rejects the extractor's output. Per-job in-memory resolution (owner, 2026-09-29): the mappings ledger is deleted — each job resolves its own proposition ids against the graph in memory and merges independently, so no job can see or collide with another's ids (the persisted table once silently dropped same-id propositions across chunks and welded wrong edges). Cross-job and cross-run identity is the auditor's links alone, plus the coverage ledger (nothing re-extracts) and the notes content-fingerprint guard.

Coverage, not quiz-counting (supersedes the v1 every-N-quiz-answers trigger): each message's view (the compact line format core/coverage.ts has always hashed) is hashed into a per-chat fingerprint ledger (coverage.jsonl); "covered" means exactly "a run committed on this content." The prompt view is the richer markdown document — a deliberate 2026-09-27 split: re-formatting presentation never re-opens covered history. Uncovered = fingerprint diff, never dual-tracked; runs are greedy token-budgeted batches (graph.extractionChunkTokens — a settings knob, default 4096, floor 500, read fresh each turn) of the oldest uncovered messages — quiz-less sessions extract too. A coarse sessions.jsonl checkpoint still marks each merged run. The trigger seam (turn-service.ts) is trigger-agnostic from the pipeline's side: when a better mechanism lands (session end, idle, explicit flush), only the seam changes. Single-flight per source; the run registry powers the coverage UI and true abort.

Two front doors, one pipeline: chat transcripts (quiz-triggered) and raw personal notes (POST /api/graph/import — each file becomes a NotesSource playing the chatId role, deterministic id from the path hash; a content-fingerprint stamp makes an unchanged re-import a TRUE no-op — no model calls, no usage — while changed text re-runs the whole file, owner 2026-09-29). Notes chunk like chats (owner, 2026-09-28): large files split at line boundaries under the same token budget, and each chunk/batch fans out as its own job exactly like a chat run — parallel extracts under the semaphore, serialized audits under the gate. No single model call carries a whole file. Both submit tools require their top-level fields (edges/links present even when empty): a payload missing them fails validation rather than silently merging structure-less nodes.

The read channel is pure derivation. GET /api/graph reads the ledgers (graphSnapshot: merged nodes/edges, birth-session clusters, positions) — the home view never triggers a write. Node detail cards show a node's edges — outgoing under "What follows from this" / "What this supports", incoming under "Follows from" / "Supported by" (an edge A→B means A implies or supports B, owner semantics 2026-09-28) — and its birth sessions.

The teacher probes, nothing is injected (supersedes the semantic-injection experiment, killed 2026-09-23): the teacher reaches the graph on its own terms through two read-only prober tools — list_sessions (birth sessions, the learner's files) and list_propositions (the names of what a session covered) — gated by the agent.graphTools setting (composing with graph.enabled). No topic embedding, no top-K selection, no session-start steer; the agent decides what to look at, and the graph never writes into a lesson unbidden.

Rendering: a hand-rolled Canvas 2D renderer (components/graph-canvas.ts — one rAF loop, dirty-flag idle, d3-zoom camera, zoom-gated label fade, cluster hulls computed at draw time) over a d3-force layout in core (layout.ts, warm-started incrementalLayout + settleGraphLayout). Render settings and drag-committed positions persist server-side; the camera persists per-device in localStorage.

---

7. Practice — strengthening, not assessment (owner ruling 2026-09-29)

Practice turns the graph into problem sets the learner reasons through out loud. The purpose ruling first: this is strengthening, not assessment. Grades are feedback for the learner — a verdict, an explanation, a look at the model solution — and nothing else. Grading writes nothing to the graph, no understanding levels exist anywhere, and the vision doc's exam machinery (scores, streaks, spaced repetition over failure rates) stays parked. A set exercises what the learner already met; it certifies nothing.

The chain model. Practice sees a node selection as reasoning chains: the selection's follows-edges grouped by conclusion (chainsForSelection, src/core/practice.ts). A chain is premises → conclusion — one per conclusion node, a singleton for an edge-less node; premise-only nodes belong to no chain. The grouping is a bijection with the follows-edge set (chainsToGraph is its exact inverse, round-trip test-pinned) — no chain exists that isn't the edges, no edge that isn't in a chain. Supports edges are invisible to practice (owner ruling: supports are grounding, not reasoning steps). The partition must cover every chain exactly once — a true partition; duplicates fail validation exactly like gaps (owner amendment 2026-09-29: variety accumulates ACROSS sets, never within one — the at-least-once doctrine is dead). chainCoverage is the soundness check: only missing fails a partition, duplicated is informational.

Self-contained sets. A ProblemSet freezes its world at creation: the selected node ids, a propositions snapshot (Record<nodeId, {name, statement}>), and the derived chains all persist into the set (data/practice/*.json, one file per set, PracticeStore port). Generation, grading, coverage, and the UI read ONLY the set — later extractions can never shift the ground under a live or historical set.

Two-stage generation. One partitioner agent reads the chains plus the three creator-set dials (searchSolutionRatio, cueDistance, presentation — each 0..1, shown as anchor text in the UI, never raw numbers) and submits the true partition; each partition becomes one problem's target. Then one generator agent per problem: its prompt carries the problem-engineering framework verbatim (the owner's doctrine — the three distances, the search-to-solution axis, the rung ladder, reverse/discriminate/probe-the-premises, the standing prohibitions), YOUR TARGET (the assigned chains echoed verbatim — the echo law: the submit's chain set must equal the assignment exactly), and THE FULL SET (the whole selection: outside propositions are supporting material the generator may draw on with minimal use and maximal focus; the never-announce law covers ALL propositions). The submit is commit-before-pose — statement, solution, and rubric arrive together (the same sycophancy mitigation the quiz/explain mechanic uses); a problem never exists without its pre-written answer. Generators run in parallel (a process-wide 3-permit semaphore shared with grading) with per-problem failure isolation: one bad problem errors alone; the set still reaches ready. Dedup across sets: each generator also sees the problems of the ≤5 most recent sets sharing a node, so situations don't repeat between sets; within a set, parallel generators can't see each other — accepted for V1 (owner ruling).

Grading is feedback-only. Submitting answers fires one grader job per answered problem: the grader sees the problem, the propositions its chains touch, the pre-written solution and rubric, and the learner's answer — never the full set (minimal, by ruling). It returns a verdict (solid/partial/missed) and feedback that names the failed recall and teaches the boundary. The verdict lands on the problem and the UI reveals it with the solution and rubric; the answer locks. Nothing reaches the graph, no level, no count, anywhere. A grader failure leaves the answer gradeless and re-submittable.

The jobs pattern — registry, not JobRecords. Practice runs are NOT Workbench jobs: no JobStore records, no steer messages, nothing ever wakes the teacher — a practice run isn't chat-scoped (a set outlives its birth session and has no lesson to re-enter; its result is itself, persisted in its own store and read by its own page). What it shares with the Workbench is the observability surface: a globalThis run registry (the module-instance idiom — __lhPracticeRuns), one row per set-phase (setId:generation, setId:grading), feeding the activity dock (GraduationCap rows; abort is set-scoped and idempotent; a boot sweep marks orphaned partitioning/generating sets as error "Interrupted (server restarted)" — the job runner's crash-honesty at set grain). Store writes serialize behind a per-set mutex; the pipeline is the sole writer of data/practice.

API + UI. Creation is GRAPH-DOMAIN, not chat-domain (owner ruling 2026-09-29: "problem sets always take in nodes — graph objects"): POST /api/practice/sets takes {nodeIds, params, title?} ONLY — no chatId, no clusterId; the UI resolves a picked cluster to its member node ids and passes the cluster's label as the title. GET /api/graph/clusters serves the picker: the birth-session partition itself (birthSessionPartition, same ordering — size desc, ties by id) — EVERY group regardless of size (owner ruling, same evening: MIN_CLUSTER_SIZE is a hull-drawing rule; practice availability must not inherit a visual threshold), each as {key, kind: "chat"|"notes", label, nodeIds, nodeCount}; a cluster whose naming record is gone keeps label: null — honest, never guessed. 202 {id} with generation fire-and-forget; 400 empty selection (empty intersection with live nodes — the vanished-nodes edge) / bad params; 503 model/key refusal. GET /api/practice/sets · GET/DELETE /api/practice/sets/[id] (delete aborts live runs first) · POST /api/practice/sets/[id]/answers (202; 409 when not ready). Surfaces: /practice (the list — status/counts rows, delete confirm, the new-set dialog with the three dials and the cluster picker — the ONLY creation entry for now; the graph-view cluster entry comes later with the free-form selection UI) and /practice/[id] (the exam page — numbered problems, markdown statements with math healing at the display seam, a textarea + focus-gated dictation bar per problem, one Submit per set, the post-grade reveal). Entry: the sidebar's Practice nav row.

Model wiring. Three roles — partitioner, problemGenerator, grader — all pinned to the strong slot (quality-critical, one settings edit from cheap) and in BACKGROUND_MODEL_ROLES (per-role overrides). The stage agents are single-submit loops on the shared runSubmitAgent seam (submit-agent.ts, extracted from extraction for exactly this): a validate hook turns a bad submit into a tool-input error the model retries against, within maxTurns 4; exhaustion fails the stage honestly (the set says so, the chains re-run cleanly on a new set).

---

8. Process boundaries (the machine room)

flowchart LR
    subgraph DESKTOP["Electron app (production)"]
        MAIN["Electron main (electron/main.cjs)<br/>thin launcher + security guard<br/>1 IPC channel: lh:expand-window"] -->|"spawn: ELECTRON_RUN_AS_NODE,<br/>loopback port, LH_DATA_DIR=userData/data"| NEXT["standalone Next server (child)<br/>ALL app logic: turns · extraction · jobs · graph"]
        MAIN -->|"loadURL, sandboxed renderer"| REND["Browser renderer (components/)<br/>fetch/WebSocket to own origin only"]
    end
    subgraph OUTSIDE["outside the process"]
        OR["OpenRouter (BYOK key from auth.json, env-injected — never argv)"]
        DG["Deepgram STT (browser-direct WebSocket)"]
        GCSE["Google CSE / Jina (researcher tools)"]
        CHROME["system Chrome (headless: demo verify · SVG render)"]
        DISK["data/ ledgers (via LH_DATA_DIR seam)"]
    end
    NEXT --> OR
    NEXT --> GCSE
    NEXT --> CHROME
    NEXT --> DISK
    REND --> DG
    REND -->|"fetch /api/*"| NEXT

The Electron main is a launcher + security guard, never a host for app logic: single-instance lock, deny-by-default permissions, origin-pinned navigation, sandboxed renderer. All Next.js code runs unchanged in a child process (output: "standalone"), relocated by the LH_DATA_DIR seam (also the test seam — data-dir-seam.test.ts covers it, including a fresh-child-process case). The build hard-fails if repo data/ would leak into the bundle or if prompt assets fall out of the trace (scripts/build-server.mjs).

---

9. The foundational invariants — what ALWAYS holds

Tier A — machine-enforced (eslint.config.mjs; lint fails the build)

Law	Zone
Imports point inward. src/core never imports agent/server/store	zone 3
The shell sees barrels only. components/** → @/src/core; app/** → @/src/core + @/src/server	zones 1–2
The agent layer never imports server or store — every dependency arrives as a factory/closure argument	zone 4

Tier B — gate-enforced elsewhere

Law	Gate
Strict types, no silent indexing (strict, noUncheckedIndexedAccess, verbatimModuleSyntax)	tsc --noEmit
No session data in the bundle; prompt assets traced	build-server.mjs hard-fails
The core barrel stays browser-safe (node-bound/heavy modules deliberately excluded)	barrel doc-comment; chunk-verified when the exclusions were set

Tier C — conventional, verified to hold in the code today

Law
One composition root per concern — stores.ts (storage), chat-handler.ts (a turn)
One writer per ledger — the merger is the sole graph writer; chat records written only by the turn pipeline; job records only by the runner; positions PUT is the only sanctioned graph-adjacent exception
The store never decides — adapters persist and shape-guard; merge/dedup/derivation semantics are reader-side, in core
Derived, never dual-tracked — append-only evidence; every state recomputed at read (snapshot, plan statuses, phases, costs, coverage)
Idempotency via deterministic keys — steer ids, ledger refs, coverage fingerprints, source content fingerprints, cluster content hashes; re-running anything converges
LLMs judge; code mechanizes — grading, merging, fingerprints, layout are deterministic; a failed LLM stage checkpoints nothing
The persisted record is the only source of truth — agents are stateless per-turn rebuilds; no unanswered tool call reaches a provider
Failure containment — side channels (context, extraction, usage, naming) are fire-and-forget behind try/catch seams; a side-channel failure never touches the response or the persist; the turn registry clears on every path
Absence is not zero — unreported tokens, absent prices, absent levels render as absent, never fabricated
Schema-valid ⇔ parseable — boundary zod schemas are shared verbatim by producer and consumer

Known deviations from the ideal (documented, deliberate or accepted)

• Two portless settings adapters re-exported through the server barrel; observability.ts imports chatStore directly. The "only stores.ts imports adapters" rule is documentation-only.
• scripts/ and electron/ are lint-exempt; scripts act as extra composition roots (and backfills are, by design, extra graph writers); tests/ imports freely across layers.
• Small conventional-only sync points: SIDEBAR_AUTOCLOSE mirrored in electron/main.cjs ↔ components/sidebar.tsx; IPC channel names mirrored by comment; the jobs dir path hardcoded in the adapter rather than config.ts.

---

10. Observability (minimal, by decision)

Structured-logging frameworks are still deferred; the baseline is one correlation-tagged log line per event at the seams ([chat <id>] on turn services, [extraction <id>] on the pipeline, [job <id>] on runners, [graph] on reads). What has grown past log lines is the activity feed (server/observability.ts → GET /api/activity → the activity dock): running jobs, extraction runs, and practice runs (per-set rows: generation/grading phases), per-run usage (money channel), and true abort (POST /api/activity/abort — kind job/extraction/practice). When structured logging lands, it goes behind these seams — no call sites change.

---

11. Testing

All deterministic, no network: engines are injected (pi-ai's faux provider — createFauxCore/setResponses — wrapped as the PiEngine production uses), globalThis.fetch is stubbed where transport matters, and pure logic gets in-memory port implementations. Suites either mirror production wiring with the real fs adapters (backup-and-clear of the ledgers they touch) or swap in-memory ports deliberately ("this suite never touches data/graph"). Four strata:

1. Pure core units — merger via real adapters, derivation, coverage, question grading, plan tree, phases, clusters, layout, tool-parts, prices, fuzzy search.
2. Server flows — handleChatRequest driven end-to-end with scripted faux 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); the practice pipeline end-to-end (partition echo/coverage retries, failure isolation, semaphore width, abort mid-generation, grading, boot sweep).
4. Render tests — transcript, cards, docks, composer, sidebar, coverage, stats, practice surfaces (creation dialog, list, exam states) via React server-side rendering (renderToStaticMarkup), no DOM library.

Gates on every change: tsc clean, all tests, eslint clean, next build clean.

---

12. The seam directory

Every place the architecture is designed to be cut — the "one file" a future change should touch:

Seam	File	What it swaps
Storage decision (JSON→SQLite, remote)	src/server/stores.ts	the adapters, once
LLM engine / model per role	settings models slots + src/agent/models.ts (piEngineFor)	provider wiring, one place
Turn composition	src/server/chat-handler.ts	what a turn consists of
Extraction trigger	src/server/turn-service.ts (triggerExtraction)	when the Library runs
Practice orchestration	src/agent/subagents/practice-pipeline.ts	generation/grading flow without touching contracts
Graph access for teacher	src/agent/tools/graph.ts (list_sessions / list_propositions)	what the teacher can probe, one gate
Steer delivery (wake)	src/server/jobs.ts + injected deliver	how background results re-enter lessons
Agent construction (tests)	handleChatRequest(req, buildAgent?)	scripted engines, no refactor
Sub-agent verification	sub-agents/headless.ts (headlessOverride)	demo/SVG verification, no Chrome
Data location (Electron)	LH_DATA_DIR env (config.ts PATHS)	where data/ lives, nothing else
Desktop host	electron/main.cjs	launcher/security shell only

---

13. Decision history — superseded decisions (do not re-litigate without new evidence)

Each of these was implemented, lived in real data, and was replaced; the replacement reasoning is in docs/vision.md's amendments table.

Superseded	Replaced by	When
Live-emission graph construction (per-session, plan-tree-derived)	Post-hoc, tree-blind extraction pipeline	2026-09-02 (vision §1)
taught events, misconception nodes, node descriptions/kinds	True propositions only; wrongness lives in the ledger; name+statement+aliases	2026-09-02/08
Editable plan tree (plan_add_node…plan_mark_done)	Author-once plan_init + derived statuses	2026-09-08 era
Hotspot + 1-hop graph context	Semantic top-K under token budget, delivered as a steer	2026-09-14 (vision-context-injection)
every-N-quiz-answers extraction trigger	Token-coverage batches (fingerprint ledger)	2026-09-15 (vision-extraction-coverage)
sigma.js + graphology + forceAtlas2 + Louvain clustering	Hand-rolled Canvas 2D renderer, d3-force layout, embedding-based agglomerative clusters	2026-09-13 era
Vercel AI SDK LLM loop (MockLanguageModelV4 tests)	pi-ai engine (faux-provider tests), AI SDK kept as the UIMessage protocol	pi-engine-swap

14. Deliberately deferred

Structured logging (§10); SQLite/remote storage (drop-in behind the ports, §3); multi-scale cluster hierarchy (Louvain's dendrogram idea, parked); the anchored calibration rubric; alternative extraction triggers beyond the coverage seam (session end, idle); retroactive bootstrap of the POC sessions (vision §1 consolidation). When any of these land, this doc changes with the commit that lands them — it describes the system, not an intention.