Demos execute JS and need a browser; the app's own binary now provides
one: main.cjs boots a hidden window + CDP port under --browser-service
(before the single-instance lock - a helper is deliberately a second
instance of this binary), lifetime tied to the server's stdin (EOF =
quit, no orphans). headless.ts connects over CDP as PRIMARY (owner
ruling 2026-10-10), system Chrome stays the fallback, graceful refusal
when neither. Electron's CDP has no Target.createTarget, so the helper's
pre-created window is THE page, reused and reset per job behind a
serialized queue. PDFs and SVGs never reach a browser anymore (native
since stages 1-2); stale Chrome wording in model-facing tool copy made
engine-neutral. Live-proofed end-to-end from the dev tree (helper seen
by process watch, js-ran 42 probe, no orphan). Gates 1523/1523.
Visual-maker SVGs never execute script, so a browser added nothing but a
process. resvg rasterizes in-process now (white background and fitTo
width mirror the old Chrome-page semantics; svgViewport unchanged,
quirks included for parity). The old Chrome path stays verbatim as the
fallback for exotic-SVG parse failures - it was the only path before,
so nothing regresses.
The resvg binary (4MB per platform, one per build) rides the standalone
via the same tracing pattern. Gates 1519/1519; native path pinned
against real SVGs (intrinsic size, viewBox-only 900 default, opts
override, system-font text).
The browser was never the PDF engine - pdfjs always painted the pages;
Chrome only hosted the canvas tab. Now pdfjs-dist v6's own Node
integration (@napi-rs/canvas, its optionalDependency) hosts in-process:
same paint code, Skia canvas, zero browsers, zero temp files, no base64
hop. The whole pdf-reader runs browser-free - the Chrome gate on
pdf_render is gone.
The Skia binary (27MB per platform, exactly one per build - npm installs
only the building machine's) rides the standalone via the same
outputFileTracingIncludes pattern as pdfjs-dist. Verified: builds green
(1515), native render pinned against the real fixture (canonical
single-IDAT PNGs, dpi/72 arithmetic, page-range skipping), and the
canvas loads + paints from inside .next/standalone.
Stages 2-3 (resvg for SVG, the --browser-service demo helper) follow.
THE START BUG (owner report, reproduced live both kinds): new sessions
(?prompt flow — practice and teach alike) sat dead until a second
message, which then persisted WITHOUT the first. Root cause, proven
both directions: the submit POST rode the view's consumeAbort
controller; React StrictMode's simulated remount fires Diet A's
unmount cleanup (model.stop()) ~8ms in and killed the POST mid-flight;
the page's sentInitial guard had already flipped, so it never retried.
Teach was equally broken — only unnoticed, because every teach chat in
use predates the diet and practice starts are always fresh. Fix: the
submit law — POSTs are not abortable by view teardown (they are ~100ms
acks since Diet B and cannot starve the pool); after the ack, a
consumeAbort identity check guards the stream attach. Pinned with the
stop-mid-submit regression test. Live: StrictMode on, practice draft
POST 200/67ms, stream attached, message persisted, problem posed.
PRACTICE PROMPT (00-base): the opening rule — material arrives two
ways, both first-class (picked propositions primary; free-form —
topics, course material, files/PDFs — fully supported, never refused);
as soon as material exists the reply poses, at most one short line
before the problem call, no warm-up interview. Plus the hard law: no
problem is ever conversation — every problem through the tool (no
thinking field/no reveal/no verdict attaches to prose).
GRAPH-CONTEXT PREAMBLE: custom markers flatten to user-role on the
wire, so the old "Contextualization — the learner picked…" read like a
system broadcast in the learner's voice channel. Now the bracket
convention: "[Session context] — picked from the learner's knowledge
map…".
SIDEBAR: busy rows unify to "Working…" for both kinds (the spinner's
default label included); "Practicing…"/"Teacher is working…" retired.
Gates: 1490/1490, typecheck, lint. doctrine's stale doc pointer fixed.
The tools_fails_2 diagnosis (glm-5.3 dropping trailing required fields):
- probe-phase quiz/explain renamed to probe/recall — a shared name carried
the probe call shape into teaching (the first teach quiz dropped nodeId);
display seams accept the old names forever
- teach quiz+explain schemas lead with the required pair (type, nodeId);
explain gains the predict/check type and joins the reveal law (checks
collapse, predictions arrive open; no-type records render check-style)
- plan_init/probe_done results go context-slim (owner, context diet):
content = validated nodeIds + binding-id reminder / fixed ack; details
keeps the tree for the popup and {ok:true} — serializer-verified that
details never reaches the model; failure echoes dropped
- the cards carry the tool's own name (Probe/Recall) and the quiz type
reads on the label (Prediction/Check) in composer and transcript
The whitehead incident (2026-10-09): glm-5.3 emitted quiz type with
quotes baked in ('"predict"') — rejected ten times, byte-identical
retries, the lesson never started — and correctAnswer with the JSON
key-value colon glued on ([':parmenides']). Third GLM wobble incident,
and each prior heal was wired only where its incident landed; this one
landed precisely where the practice heal never reached.
Layer 1 — healEncodedArguments moves to core/argument-heal (the display
seam needs it too — the tool-parts law) with withArgumentHeal +
healToolRecord (agent/tools/heal) wired at EVERY registry: prober,
teacher, practice, bareQuestionTools (it had no prepareArguments at
all), demo-builder, and submit-agent (composing after
healStringifiedArrays; the auditor's extraTools wrapped).
Layer 2 — stripModelValueDebris (core/value-debris): leading/trailing
colons, quotes, backticks, whitespace strip at the MEMBERSHIP seams —
coerceCorrectAnswer (boundary check + display + grading in one),
normalizeQuizType (persisted wobbled records keep their role),
resolveTargetId (copy-syntax on targets).
Layer 3 — schema description hygiene: the quiz type and practice
mode/outcome/rung/missClass values now read as bare words — the quoted
JSON-key style invited quote-copying.
Display mirror — parseQuizInput heals the raw record on read, so a
boundary-healed quiz whose record keeps raw args renders instead of
becoming an invisible pending quiz (learner lockout).
Pins: tests/tool-heal.test.ts (the incident shapes end-to-end through
pi-ai validation, the wiring, the strips, the mirror). The stringy-
options pins flip to boundary-parity per tool-parts' own law; the
chat-flow garbage pin keeps its honest-reject intent with an unhealable
shape. 1471 tests, typecheck, lint green.
Limits out (owner rulings): the 'Typically 5-15 nodes' cap and the
'keep the map small' pressure are dead — a plan is as elaborate as the
subject needs; what keeps it readable is correct structure (unconditional
truths as the hanging points), not compression. Labels stay short; size
is free. The method section now carries the goal reiteration: thorough,
deep, real understanding — sessions run as long as the understanding
needs, plans as elaborate as the subject requires; no schedule, nothing
to rush. (Code was already limit-free: no maxTurns, no parser caps.)
The adaptation + detour law, merged into one section ('The map moves
with him'): when the ground shifts mid-lesson — a missing foundation, a
deeper confusion — never rush the repair; no wall of text that 'covers'
the ground to get back on plan. Update the map if the structure changed,
then build the missing ground the only way anything is built here: one
reasoning step at a time, verified, at the session's one and only pace.
Detours are the teaching, not interruptions to get through.
The probe enters at the top (owner-observed failure: eigenvectors, then
high-school chain rule, then partials — strands opened at textbook
floors). New entry-point rules: open near the hardest thing the goal
needs and let a miss walk you down; strands inform each other (a strand
crushed at height says neighbors start high); the verdict + confidence +
thinking field TOGETHER set the next question's difficulty — confident
correct jumps sharply, hesitant correct steps gently. Efficiency is
resolution: never descend into already-bracketed territory; the
questions you save buy the specific, nuanced ones that locate the edge.
The one emphasis device for load-bearing ideas: GitHub-style alert
callouts. The prompt teaches only > [!IMPORTANT] (one device, one
meaning — an unconditional truth laid down, a payoff landed, a hinge
definition) with the restraint law: everything else stays in prose, if
everything is boxed nothing is. The renderer accepts all five GitHub
markers and boxes them identically — a stray [!NOTE] still meant 'this
deserves emphasis', and a marker line must never leak as text.
Rendering (owner rulings): SHARP corners, white stroke, no background,
no label, no icon — just the outline — and the box hugs its text
(width: fit-content capped at the column) instead of stretching
full-width into empty space. Same grammar as the plan blocks: this is
structure, not chrome.
Seam: the Streamdown blockquote override in md-blocks. parseAlertMarker
is pure and exported — Streamdown wraps blocks in its own styled
components, so detection reads the hast node's tagName, and a standalone
marker drops its <br> hard-break artifact along with the empty
paragraph. Plain blockquotes pass through untouched.
Rider (owner, same round): tables hug their contents — width:
max-content on the data-streamdown table (width:auto would still
stretch, the classic table trap); wide ones keep the md-blocks
overflow-x scroller. Graceful in exports: a callout is still a
blockquote there.
6 render pins: the box mounts with white stroke + fit-content + no
rounding, all five markers box, standalone-marker drops the empty
paragraph, multi-paragraph structure kept, plain quotes untouched.
Accuracy (shared, one law across every agent): the target shifts from
surface facts to understanding itself — 'the deeper risk isn't a wrong
name or date, it's a wrong understanding: an interpretation of an idea
that seems solid and isn't.' The bar is the doubt itself: checks are
cheap, and weighing whether a doubt is worth checking is how errors slip
through. Thinking stays the agent's own.
Research (teacher 40-jobs): floor framing — verify your understanding
before it becomes his; dispatch before the fact crosses your lips, teach
around the gap until the brief lands. Planning (10-planning): planning is
the natural moment for verification; the map is a hypothesis and the
probe only its first sketch — rewriting is a normal, common event (the
session adapts to what teaching reveals); and plan_init IS the
presentation — never draw the plan again in prose (the double-plan bug).
New shared segment 16-mental-model.md: the thinking fields are the
session's primary data — read them in full, measure what they reveal,
adapt everything on the model they build. Composes into teacher, prober,
and practice (shared-fragments list updated).
Practice: the rarely-needed framing is dead — fact checks get their own
law (a wrong model solution grades the learner against a lie; pose once
the brief lands), files and map readers become plain side tools.
LaTeX: the question-tools craft list gains the point-of-use bullet
('LaTeX everywhere — plain-text math in a question is a craft failure'),
the formatting law names EVERY field, and the practice problem statement
line carries it too.
Prompt pins updated: felt-doubt bar + mental-model doctrine in both
teacher and practice compositions, no-double-plan rule, rewrite-is-
normal law, rarely-needed framing asserted gone.
web_fetch_pdf downloads (25MB cap, probe-validated) into a per-job temp
dir and returns the triage summary — page count, size class (the pdf
doctrine's <=10/<=40 thresholds), flagged math/image-heavy pages. The
researcher's toolset then gains pdf_info / pdf_extract / pdf_search over
that dir, scoped by a resolve-and-startswith wall: the researcher's
filesystem ends at the temp dir. Render is deliberately absent — the
researcher slot has no vision; visual reading stays the pdf job's job.
web_fetch: a PDF URL now returns a POINTER to web_fetch_pdf instead of
dead-ending, and every output carries 'Source: <final url after
redirects> (HTTP <status>)' — the 'ill-formed request' lesson: a
200-served error body must be distinguishable from a real page.
The runner owns the temp lifecycle (mkdtemp before, recursive cleanup in
the job's finally — never enters the record) and web_fetch_pdf urls ride
the sources ledger. Doctrine step 5 added: fetch, triage, extract only
the pages that matter, cite as (p. N).
Tests build a REAL minimal PDF programmatically (correct xref offsets —
pdfjs parses it like any download): zero network, real parse.
URL.protocol carries a trailing colon ("https:", WHATWG); the colon-less
comparison (a direct port from the py harness, where urlparse().scheme
has none) rejected EVERY valid URL ever fetched — the tool never fetched
a page in its life. Owner-reported via OpenRouter logs: 8 valid URLs in,
8 'invalid URL' results out.
The new pin stubs globalThis.fetch so the happy path runs hermetically —
a valid https URL must REACH the fetch stage (and ftp:// still rejected).
dist rebuilt (desktop:dist) — bundle verified: "http:"!==t.protocol
Owner ruling (2026-10-06): every concluded problem opens a pause. The
agent may speak its closing words about the learner's last thinking
after the conclude result, then ENDS ITS TURN — a system-prompt law
deliberately, per owner correction, NOT tool-level terminate. The next
problem is never posed until the learner explicitly moves on; the pause
is the session's teaching space (repairs, clarifications, follow-ups —
still surgical, never re-teaching).
The move-on signal is a first-class input: { type: "next" } appends a
practice-next marker (content = single-sourced NEXT_PROBLEM_TEXT — the
model's unambiguous go-ahead: pose per the selection rules or wrap if
the arc is complete). Honest 400s: practice sessions only, nothing
active (mid-problem input rides the dock), riding after any pending
dismissal synthesis exactly like text.
The fast path (owner approved: composer-band chip + Enter-on-empty):
- Composer.enterAction gains "next" — Enter on an EMPTY composer in
the pause fires the signal; typing anything makes Enter a send again;
busy (steer/stop) outranks the pause.
- The Next chip hovers at the composer band's top-right (pick-pill
chrome register) whenever the pause is live and a problem has ever
been posed; the placeholder becomes the hint ("Ask about this
problem — or press Enter for the next one…", exact);
- the transcript renders the signal as a compact centered NEXT PROBLEM
line (mono-caps chrome); session-core appends it optimistically.
CONCLUDE_ACK_TEXT rewritten to carry the law; doctrine §VIII gains the
pause law (owner-sanctioned edit #3); 00-base's conclude/dismissal
bullets rewritten + the between-problems section added.
- shared/10-voice.md (agent-neutral) + shared/15-accuracy.md (research-tool
truth contract) compose for teacher, prober, AND practice
- PRACTICE_SHARED_FRAGMENTS grows beyond 40-formatting.md
- practice 00-base.md drops its stale near-copy ("You have no research tool
here" was false since round 10); the solve-and-shortcut-hunt line folds
into the problem paragraph
- teaching system-prompt.md slims (Voice/Accuracy sections removed)
- pins updated: practice-agents composition, prompt-overrides segment mirror
A session whose material is not in the map is a SUPPORTED mode now:
target is optional (absent = untagged problem, poses freely), an
explicit [] heals away at the seam (the record's canonical form is
tagged-or-absent), and when targets ARE present the round-9 law is
unchanged — whole-graph validation, names normalized to ids. The
rejection message teaches the exit ('omit target entirely'), so a
model inventing targets for an off-graph topic converges to omitting
on the retry; omitting always works, no brick is possible. Prompt
carries the clause: material in the map → ids; not → omit, problems
stand on their own.
The practice registry widens beyond problem+conclude: read/ls/find/grep
(the learner-filesystem set, always), research + pdf jobs (research
honestly absent when no engine is live; demo/visual stay teach-only),
and the prober's map readers behind the SAME settings gate as the
prober's (graphToolsEnabled at the composition root). Base prompt
carries the conditional doctrine: reach for them only when the
situation genuinely requires — no propositions → look at the map;
learner references a file → read it; load-bearing uncertainty →
research it. Not toggle-filtered: problem/conclude are the session,
these are exceptions.
Four pins on the mastery-trace contract: a name target normalizes to
the canonical id (case-insensitive, whole-graph set), an id passes
through untouched, an unknown target rejects one-line AND does not arm
the gate (honest activation — the next valid pose goes through), and
the legacy node-source-less wiring stays unvalidated. The base prompt
now states the convention: target takes the proposition ids from the
session context — ids are how problems are traced.
problemTool takes the whole-graph node source (chat-handler passes
graphStore.readNodes, read fresh per pose): prepareArguments silently
normalizes name targets to canonical ids so the RECORD carries ids
(the mastery trace groups on them); execute rejects unresolvable
targets with a one-line invalidInputResult — no re-listing, the [id]
proposition list already sits in the agent's context. Validation set
is the whole graph, not the picks: pickless sessions and future
graph-tools poses stay valid. Absent node source (tests) = no
validation.
Owner round 2 (2026-10-05), Slice H: the Nudge/Commit two-lane buttons
are gone. One Send; Enter sends, Shift+Enter is a newline. Whether a
submission claims an answer is apparent from the thinking itself — the
agent reads it (hint ladder for traces, moves+verdict for claims).
- The thinking input loses the commit field end-to-end: ChatInput,
parseChatInput, newThinkingMessage (details and the model-facing label
both collapse to the single 'Your thinking'), the optimistic marker.
- Legacy records' commit flags are deliberately unread.
- Doctrine: the two labeled channels become one read instruction; 'answer
commits' section retitled 'Claimed answers' — a claim is not a
progress signal.
- ThinkingCard: one label, collapsed arrival (the unified row).
- The dock placeholder no longer explains lanes that no longer exist.
Owner round (2026-10-05): the problem pose only really needs statement,
mode, target, modelSolution; conclude needs no reveal field.
- problemInputObject drops decisionPoints, expectedWrongMoves, strategy,
rationale. Halves the pose payload (the GLM incident's failing args
were 3.3KB — a direct weak-model hardening).
- concludeInputObject drops solutionReveal: the reveal is the
modelSolution committed at pose time, shown verbatim at conclude.
- Doctrine pass (00-base mechanics, 10-doctrine V.4/V.5): the commitment
is the solution alone; the reveal instruction points at it.
Fixtures slimmed to the new shapes (flow, agents, heal). ConcludeCard's
render of legacy solutionReveal input stays until the card redesign
lands (Slice C).
The bkt-practice record showed ONE turn in which the model concluded
(clean ack) and then every pose died on 'a problem is already active'
until it gave up ('the problem slot is stuck — the conclude calls report
success'). Two flaws, both fixed:
1. GATES READ A FROZEN RECORD. problem/conclude gates re-read the stored
record per execute — but the record freezes at the turn's write-ahead
(persistTurn lands only at turn END), so a mid-turn conclude → pose
pair could never see each other. practiceTools now builds ONE
turn-scoped gate (createPracticeGate) seeded once from the record;
the executes flip it. Fresh closure per turn (the agent is built per
request) — no cross-turn leakage.
2. PHANTOM ACTIVATION. activeProblemOf ignored tool RESULTS, so the
composer mounted from a pose the gate itself had refused (owner
ruling: 'either it goes through or it doesn't'). Honest activation:
a problem activates only on its clean ack; a conclude frees only on
its clean ack; errored and dangling calls are skipped. Retroactively
heals the BKT record (all its poses errored, all concludes clean →
no active problem). pendingProblemDismissalOf follows the same law
via the shared cleanToolCallIds helper.
Pinned: the mid-turn conclude→pose pair works in one toolset; the gate
re-arms after a successful pose; lone errored/dangling poses activate
nothing; a clean pose under a rejection stays active; an errored
conclude frees nothing.
The GLM incident (2026-10-04): glm-5.3 sent mode as a JSON-string-encoded
enum ('"recall"') on 19 consecutive problem calls; TypeBox's error never
named the expected values, so the model re-emitted the identical payload
until the dev server was killed. And mid-turn the learner was trapped:
Esc's question/problem rungs only dismissed — the dismiss route 409'd
against the turn-owned record and nothing could stop the running turn.
- healEncodedArguments (practice.ts, the prepareArguments seam): a STRING
where the schema expects a literal union / boolean / null / number /
array is JSON.parse-d and kept when it satisfies the schema. Valid args
pass through untouched (same reference); unhealable values stay for
validation to reject honestly. Wired on BOTH problem and conclude.
- dismissal-park.ts: the dismiss route parks a mid-turn dismissal (202);
the turn's settle path (onFinish, after persistTurn, before endTurn —
the same ordering law as the record write itself) applies it against
the persisted record, only if the dismissed thing really is pending.
- confirmAbort: question/problem discard while busy now ALSO aborts the
in-flight turn; the dialog line says so (busy prop).
No error-result cap (owner ruling). 16 new pins (heal incl. pi-ai
validateToolArguments end-to-end; park registry, apply paths, the 202
route contract).
The practice agent wrote wrong LaTeX delimiters — 40-formatting.md (the
\(...\)/\[...\) delimiter law, markdown, mermaid) lived in shared/,
which practice deliberately excluded. Owner ruling: formatting applies
in every setting. Practice now composes its own directory plus a
CURATED shared allowlist (40-formatting.md only — the question/
confidence/file mechanics stay out, practice has no such tools); the
composition test pins the appendix and the order.
- tools/practice.ts: problem (terminate pose; commitment hidden in args)
and conclude (non-terminating; outcome/verdict/reveal/rung/missClass/
retest), gated by activeProblemOf — pose-while-active and conclude-
with-nothing-active both return invalidInputResult.
- prompts/practice/00-base.md (voice + accuracy + tool mechanics) and
10-doctrine.md (the owner's doctrine verbatim; one approved edit —
§I material arrival generalized). Practice composes its own directory
ALONE: no teaching base, no shared fragments.
- createPracticeAgent (teacher-agent.ts); chat-handler branch — kind
'practice' builds the practitioner from turn 1 and skips extraction.
- problemResult/concludeResult acks in tool-execute (conclude text
single-sourced: CONCLUDE_ACK_TEXT).
Owner ruling 2026-10-04: the module is a sunk cost, retrievable from
history if ever revisited. Removed: core practice contracts + pipeline
helpers, the PracticeStore port + adapter + data dir path, server
service + API routes (sets, clusters picker, practice aborts), the
practice pages + dialog/exam/list UI, chain-table's Generate handoff,
sidebar nav, activity-dock practice rows, and the partitioner/
problemGenerator/grader model roles (BACKGROUND_MODEL_ROLES 8 → 5).
Survivors, moved to core/chains.ts (pure + browser-safe): ReasoningChain,
chainsForSelection, formatChainLine — the chain table and graph
emission still consume them.
pi-ai 1.0's onProviderStreamEvent, at the layer our incident class lives
(GLM arg truncation, DeepSeek id resets) — env-gated (owner choice:
no app surface, off by default; raw events carry full prompts), one
JSONL per chat under data/debug/provider-events/ with a 20MB cap and a
run-start marker for record correlation.
Wired at every agent construction: teacher/prober turns (AgentBuildOptions.
chatId), jobs via LoopContext (researcher, demo-builder through
runSubagent; visual-maker + pdf-reader through their streamOnce), and
the extraction pipeline (extractor + auditor keyed by source id).
practice deliberately unwired (sunk-cost module).
Authoring refinements (quizInputProblems, nodeId checks, ask/explain
refinements, submit-agent pipeline laws) now return invalidInputResult —
same model-facing failure and retry semantics pi minted from the throw,
but details keeps {tool, problems} as data instead of the throw path's
empty {}. ToolInputError deleted; wire shape unchanged (verified against
pi's own thrown-error conversion in agent-loop).
The teaching cycle gains its motivation question. The teach quiz schema
now carries a REQUIRED type field — predict (posed before the teaching,
arrives open, the learner predicts the next step) or check (posed after,
arrives collapsed behind the reveal). The prober's quiz stays role-less.
- core: quizInputObject.type union + normalizeQuizType + the reveal law
as pure quizNeedsReveal; QuestionData.questionType; note → thinking
rename (legacy reads fold back); readingThinking channel on answers;
compactAnswerForModel emits thinking / thinkingWhileReading
- tools: descriptions stripped to mechanics — QUIZ_OPTION_CRAFT and
EXPLAIN_DOCTRINE deleted, doctrine moved to the prompt layer
(reverses the 2026-09-27 placement, owner 2026-10-03); shared
quizDescription/explainDescription builders differ only in deltas
- prompts: shared/20-question-tools.md rewritten (option craft,
transcribed-speech leniency, the two thinking channels); new
teacher/28-predictions.md (cycle, roles, skip rule, resolution law);
20-teaching + 30-question-principles scoped to the new taxonomy
- UI: predictions render open with a Prediction eyebrow, verdict shows
but the explanation tucks behind a toggle; the reading-thinking field
mounts while a check waits (dictation claims it until reveal);
answered card separates thinking vs thinking-while-reading blocks
- requiresReveal gates on pending.tool === 'quiz' (owner: only graded quizzes
arrive collapsed; ask/explain always open)
- ESC-abort chain opens AbortTurnDialog everywhere (owner ruling); from a
pending question abort IS dismissal — dismissedAnswer rides the answer
path, the teacher acknowledges the skip in its next turn; popup-Esc is
claimed via the state gate (two window listeners fire per Esc)
- 25-confidence.md: stripped the slider/JSON/calibration mechanics — pure
reading doctrine remains
- toolQuestionKind named once in core (dock parity), dock switched to it
Owner design round 2026-10-01: every multiple-choice quiz answer now
carries the learner's stated probability that his selected answer is
correct — calibration data for the prober and teacher, the flow itself
unchanged.
Definition (owner ruling, after weighing chance-normalization): RAW
probability. Sub-chance reports are diagnostic gold, not dead space —
'picked an option, believes it wrong' is a real state. The agents get
the sophistication: new shared/25-confidence.md composes for both
agents (chance level, sub-chance = distrusts the pick, correct+low =
fragile, wrong+high = committed misconception, IDK = genuine gap).
- question-core: AnswerRequest/QuestionAnswerData.confidence (quiz
only, optional — old records valid), buildAnswer clamps and drops it
on dontKnow, formatAnswerMessage states the line, CONFIDENCE_STOPS +
DEFAULT_CONFIDENCE_INDEX exported.
- question-dock: snapped slider below 'I don't know', above 'Your
thinking' — native range over the stop INDEX so ←/→ step the uneven
grid cleanly, Tab chain rows → slider → thinking, Enter submits from
the slider too. Rests at 50 and always sends (owner: 50 must mean a
real coin flip, never 'untouched'). 'I don't know' GRAYS it out —
layout holds, no value sent.
Deliberately untouched: explain free-recall card (quiz-only scope),
answered-card rendering of confidence, ask flow.
The partitioner was told to reference chain ids, but no rendered line
ever showed one — chainLine emitted bare 'premises → conclusion', so
the model re-typed whole rendered lines into submit_partition and every
call failed 'unknown chain ids' validation (owner live report,
2026-09-30 evening). The generator's YOUR TARGET block had the same
blindness.
- chainLine now leads with the chain's id: [arg_x] n_a, n_b → n_c;
a bare proposition repeats its own id ([n_y] n_y (no premises)) —
mechanically teaching that its chain id IS its proposition id.
- PARTITIONER_INSTRUCTIONS rewritten to chains-primary vocabulary
(was still 'follows-edges, bijective re-encoding' — pre-chains
framing) with the id-only doctrine and singular propositions
first-class.
- partitionSchema/problemSchema gained field descriptions: ids
VERBATIM, never re-type, never invent.
- Generator OUTPUT bullet + target-block header say the same.
Design ruling en route: partitioner stays an LLM partitioning chains
(deterministic partitioning and propositions-as-targets both
considered and declined by the owner).
Owner live-test verdict on the spine (2026-09-30): no perceptible gain —
the real problems (restatement nodes, shallow re-derivations, same
conclusion from many angles) are mint-time issues a post-hoc chain
minter cannot fix. This arc moves assembly INTO the extractor and adds
the deterministic dedupe the owner specified:
- Digest-aware extraction: each stretch's extractor prompt carries THE
GRAPH SO FAR (its own source's merged propositions + chains, ids
verbatim); build on existing ids instead of restating, lean on
established conclusions as premises. Mixed-namespace chain endpoints
(fresh p-ids OR existing n_-ids) validated at the submit seam.
- Serialized per source: a source's batches/chunks run strictly in
order (each extract reads every earlier merge); cross-source runs
stay parallel; source-wide abort stops the loop. The spine stage,
role, phase, usage stage, cost row, and setup roster are gone.
- Core-ideas prompt: goal reframed as uncovering core ideas and showing
how everything derives from the foundations; direction upward, prefer
lean derivations, never re-derive what the graph already holds.
- Chain simplification (merger maintenance): simplifyChains drops
provably redundant chains at every merge — prefer-fewest-premises +
ancestry-cover ([A,B]⟹X with [X,C]⟹Y covers [A,B,C]⟹Y, transitive);
same-key provenance lines drop together; rewriteLedgers under the
audit lock; droppedChains counted separately in stats.
Chunk budget stays 8192 (owner: keep for now, maybe lower later).
Per-batch pipeline is now extract → spine → audit → merge (owner order,
2026-09-30), the spine inside the serialized audit section so it reads
the graph exactly as the following audit will. Measured motive: 0 of 133
chains in the live corpus crossed a stretch boundary — no stage minted
cross-stretch chains.
- src/agent/subagents/spine.ts: SESSION/NOTES spine preambles + shared
SPINE_CORE (span doctrine; extraction doctrines carried over), the
submit_spine TypeBox schema, validateSpine (mechanical span law: every
chain touches >=1 fresh endpoint AND >=1 existing digest endpoint;
unknown ids, self-reference, and id collisions are loud), the digest
prompt builder, and the runner (strong tier, retry-only).
- extraction.ts: 'assembling' phase, spine stage in the per-need gate,
usage/billing stage, assemble() inside withAuditLock for both front
doors; first stretch of a source skips the call (empty digest, no
spend); the audit sees COMBINED propositions (chunk + connectives),
never chains; merger receives the combined findings.
- core/graph.ts mergeFindings: chain endpoints resolve through the
job-local map OR pass through verbatim as existing n_ ids (the spine's
mixed namespace).
- EXTRACTION_CORE: steps-of-reasoning selection bar (examples, notation
glosses, and restatements that no argument needs are not steps).
- Factory chunk budget 4096 -> 8192 (fewer seams, more local context).
- healStringifiedArrays moved to the shared submit seam (submit-agent.ts).
- Cost panel: Assembly row; usage records carry a spine stage (optional
on legacy records).
- Tests +5: prompt pins, span-law matrix, digest rendering, seeded
end-to-end flow (spanning chain + connective, verbatim n_ endpoint,
spine billed as its own stage), non-spanning refused + retried.
Data analysis of the five-source corpus (490 nodes, 244 chains) found
three pathologies the goal-first core let through:
1. Narrative-adjacency chains — premises that occasioned rather than
entailed conclusions (concentrated in the polemical sources).
2. Session-scaffolding minted as knowledge — the source's own axiom
numbering (A1/F2/UT2), arc labels ('the crux') persisted as nodes.
3. Full-discourse volume — glossary and argument-arc material at equal
rank with load-bearing ideas (~200 nodes per source).
Three doctrine lines, same terse register as the core:
- entailment is the only link that earns a chain; historical/narrative
connections keep their pieces as bare propositions or drop them
- load-bearing ideas only (what a reference on the subject would teach)
- propositions must stand alone for a reader who never saw the source;
no leaning on the source's own scaffolding
Substring pins for all three; shared by session + notes extractors by
construction.
- architecture.md section 6 rewritten: stored truth = propositions + chains
(chains.jsonl); deriveView as the lossy image (implication nodes +
untyped arrows, rebuilt per read); the bijection framing (independent
arguments for one conclusion stay distinct — the corruption the edge
store invited); new deletion-semantics paragraph (proposition death
cascades to every referencing chain; arg_ death = one argument);
supports retirement stated plainly; auditor's propositions-only scope;
clusters stated as birth-session partition truth (no Louvain claim)
- section 7 practice chain model: stored-chains inclusion law replaces
follows-edge grouping; chainsToGraph bijection claim deleted
- README: graph clauses (stored chains, derived view, chain dedupe),
practice clause, ledger inventory, and two long-stale line-5 claims
(ledger events, semantic top-K) corrected
- contracts/graph.ts GraphSnapshot comment: Louvain claim replaced with
birth-session truth; extraction.ts comments no longer narrate the
dead expandFindings/chainsToEdges seam
Owner rulings 2026-09-30: reasoning chains are the stored truth; the graph
canvas renders a lossy derived image. No migration, no backcompat — old
graph data was wiped.
- contracts: GraphChain (arg_-minted id, premises, conclusion) replaces
typed edges; GraphEdge survives only as the untyped derived arrow;
GraphSnapshot carries stored chains; supports edges retired entirely
- store: chains.jsonl ledger (edges.jsonl purged at load alongside legacy
events.jsonl); readChains/appendChain seam; merge dedupes by
(sorted premises, conclusion), re-assertions append under the same arg_ id
- core: deriveView mints implication nodes (name ⟹, statement via
formatChainLine) plus premise→imp→conclusion arrows; clusters stay
proposition-only (birthSessionPartition); chainsForSelection now groups
STORED chains (chain rides iff every member selected) — independent
arguments for one conclusion never union; chainsToGraph/chainsToEdges die
- extraction: extractor submits chains natively (expandFindings deleted);
auditor receives propositions only; supports fan-out gone
- deletion: full death of a proposition kills every referencing chain;
arg_-prefixed node deletes kill that chain; chainsRemoved stats
- UI: every arrow renders the ⟹ double shaft; supports toggle + thickness
slider and detail-card edge sections deleted (card keeps name +
proposition); chain table + practice pipeline read stored chains
- tests: 1188 green across the new seams
- dialog: optional 'Anything else?' textarea (typing-only, 2000-char mirror)
- API: body gains note? — trimmed, whitespace-only omitted, hard 400 over 2000
- pipeline: note persisted on the set and threaded to both agent stages
- agents: THE LEARNER'S NOTE prompt block (present only when set) + the
weighed-never-owed stance pinned in both doctrines
- fix en route: doubled THE PARAMETERS heading in both prompts since slice 3
(renderParams owns the heading; the literal blocks removed)
Rulings: each problem generation is its own visible dock job (the
extraction per-chunk pattern), and practice shares extraction's
semaphore — one process-wide 3-permit pool for ALL background LLM
jobs; total background LLM concurrency bounded at 3.
A. SHARED POOL: new src/agent/subagents/bg-llm-semaphore.ts —
globalThis __lhLlmSlots, PERMIT_COUNT 3 (the ruling), the
fresh-take/handed-off-permit deadlock discipline preserved for
BOTH callers. extraction.ts's withExtractSlot is now a thin
wrapper (job-row signal → withBgLlmSlot); __lhExtractSlots and
__lhPracticeSlots both die. The partitioner now acquires a permit
too (it's a background LLM job — it didn't before).
B. REGISTRY ROWS: the aggregate setId:generation / setId:grading rows
are replaced by setId:partition (phase partitioning, label
'Partitioning — title'), setId:gen:problemId (queued→generating,
'Problem i/N — title'), setId:grade:problemId (queued→grading,
'Grading i/N — title'). Each row carries its OWN abort controller
and its OWN usage (one PracticeStepLog per row — the aggregate
partitioner/generator/grader folding dies). Row key = feed id =
abort key (the dock dup-key law).
C. ABORT SEMANTICS: abortPracticeRow(key) routes by shape — gen row:
THAT problem only (error 'Aborted', siblings continue, set settles
ready with one errored problem); partition row: the whole run;
grade row: that grade drops (answer kept, re-submittable).
abortPractice(setId) composes from row aborts + the synchronous
persisted marks (a set-abort's verdict stands before the settles
re-read it — the settle step keeps an existing error status).
/api/activity/abort kind practice now takes the ROW key (charset
admits ':').
D. FEED + DOCK: ActivityPractice/FeedPractice carry per-row identity
(id = row key, problemIndex/problemTotal); dock blobs render per
row, labels carry the i/N; the aggregate row is gone.
E. Boot sweep unchanged at set grain; the live-row check is now a
key-prefix scan (any live row of the set protects it). The
updatedAt heartbeat covers the new row scheme as-is.
Tests (20 in the pipeline suite): per-problem row shape/uniqueness/
i-N labels (partition row frozen behind a gated partitioner; per-row
usage ISOLATION pin — b's log never shows on c's row); abort-one-gen
(sibling completes, set ready with one errored problem); partition-
row abort; queued grade-row abort (dead-ticket path: answer kept,
re-submittable, sibling lands); shared-pool identity + contention
(practice queues behind extraction-style holders); registry/dock
feed mapping. Test isolation hardened: unique per-store set ids,
afterEach registry/lock/pool reset (a timed-out test can no longer
poison the rest).
architecture.md: jobs-pattern paragraph rewritten (per-problem rows,
shared pool, abort routing).
1149/1149, tsc 0, lint 0, build 0.
Prior state: zero formatting guidance anywhere in the agent module —
models wrote prose-only by default.
GENERATOR_DOCTRINE Part III gains the FORMATTING bullet (after the
THE PARAMETERS bullet, before PREVIOUS PROBLEMS): statement, solution,
and rubric render as full Markdown — math always LaTeX (inline $…$,
display $$…$$), code/pseudocode/formal systems in fenced blocks with
a language tag, identifiers inline as `code`, tables for parallel
structure, lists for enumerations; prose-only renderings of
mathematical or structured content are a defect.
GRADER_INSTRUCTIONS gains the feedback-formatting paragraph (before
the OUTPUT section): full Markdown, LaTeX math, fenced code, lists/
tables where they carry structure — teach the boundary in properly
formatted prose, not plain text.
PARTITIONER gets nothing (machine-shaped output).
Tests: substring pins on the exported doctrine constants — the exact
FORMATTING sentence + ordering (after presentation, before previous
problems), the exact grader paragraph + before-OUTPUT ordering.
Architecture.md §7 one clause (the formatting doctrine in the
generator's prompt inventory).
1144/1144, tsc 0, lint 0, build 0.
Owner report: created a set, 'Interrupted (server restarted)' showed
on list + set views, generation proceeded and overwrote it. The
sweep's 'generating + no registry row' test had a false-positive
mode: it ran at EVERY pipeline instantiation, and an instance could
evaluate against a registry whose fresh row it couldn't observe.
Three guards, per ruling:
1. HEARTBEAT: ProblemSet.updatedAt? (contract + doc) — bumped on
every store write; createSet stamps it = createdAt; saveSet
re-stamps (validator tolerates absence for legacy rows).
2. ONCE PER PROCESS: the sweep guards on globalThis
.__lhPracticeSweepDone — no instance re-evaluates after the
process's first sweep.
3. GRACE WINDOW: only sets whose heartbeat (updatedAt ?? createdAt)
is older than 120s are stamped. A live run saves at every phase/
problem transition, so it is structurally unreachable; a true
orphan is minutes stale at next boot. One [practice <id>] log
line per stamped set (the extraction/jobs console.error idiom).
Tests: stale orphan stamped; FRESH generating set untouched (grace);
once-guard (second factory instantiation never re-sweeps); live-row
protection re-armed via resetSweep() test seam; store pins for
updatedAt birth-stamp + re-stamp + legacy tolerance. Architecture.md
sweep sentence updated (once-per-process + grace semantics).
1142/1142, tsc 0, lint 0, build 0.
Contracts: LadderRung 0-4, CueDistance 0-2, Presentation theory|
application; PracticeParams carries the three named levels. Doc
comments hold the three definition tables verbatim (rung names +
definitions from the ladder doctrine; cue levels; presentation
levels).
Core: LADDER_RUNGS/CUE_LEVELS/PRESENTATIONS tables in src/core/
practice.ts (definitions verbatim) — ONE source spoken by the agents'
prompts, the dialog's stepped controls, and the exam header; core
barrel re-exports practice.ts (pure domain, browser-safe).
Agents: Part I's ladder intro now 'five rungs, numbered; the request
arrives as a rung' with bullets numbered 0-4 (definitions untouched);
Part III's parameters block replaced with the named-level doctrine
(rung = exact class, no continuum; cue counts lesson examples AND
previous problems' situations; presentation dresses, never changes);
partitioner's RESPECT THE PARAMETERS replaced (low rungs = many small
assignments, high = larger; far cue makes confusables-together MORE
important; presentation does not shape the partition). renderParams
outputs the exact named-level block. Anchor/0..1 language swept dead.
Service + store: strict union validation — floats between levels,
out-of-range, wrong strings, legacy dial shapes all 400 at the API
and fail isProblemSet (legacy sets skipped from lists; no migration —
data/practice is disposable scratch, owner wipes before testing).
UI: dialog = 5-detent ladder slider + 3-detent cue slider (each shows
the level NAME + one-line definition) + Theory|Application segmented
toggle; defaults rung 2 / cue 1 / theory; Create posts the new shape.
Exam header gains the levels line ('Direct exercise · Adjacent
context · Theory').
Tests: renderParams exact-block pin; strict-validation pins (new
tests/practice-service.test.ts — bad unions 400, legacy shape 400);
store pins (legacy floats + 2.5 rung fail); render pins (detents +
names + definitions, defaults, detent-swap, toggle active segment,
levels line); every params fixture moved off floats. 1138/1138, tsc
0, lint 0, build 0.
Docs: architecture.md — discretization paragraph (compact tables,
named-level invocation, defaults, one-table source), UI mentions;
README axes sentence.
Owner ruling: 'problem sets always take in nodes — graph objects' —
practice is graph-domain, not chat-domain. No session-created sets.
API: POST /api/practice/sets takes {nodeIds, params, title?} ONLY —
the chatId branch, UnknownChatError, and the any-provenance filter are
deleted; the selector is nodeIds (readNodes filtered to the selection,
unknown ids drop, empty live intersection → 400 empty_selection);
title optional from the UI (the picker's cluster label), dated default
stays. chatStore leaves the pipeline deps entirely (labels resolve
store-side in the clusters service).
GET /api/graph/clusters (new): the birth-session groups the graph view
renders — graphSnapshot's clusters verbatim (same ordering, same
MIN_CLUSTER_SIZE presentation filter), mapped to {key, kind
chat|notes, label (null when the naming record is gone — honest,
never guessed), nodeIds, nodeCount}; corrupt/empty graph → empty
list. GraphClusterGroup contract in core/contracts/graph.ts (client-
safe — the dialog imports it).
practice-dialog: session picker → cluster picker (rows show label +
proposition count, notes clusters marked); Create posts the cluster's
nodeIds + label as title + params. 400 empty-selection copy reworded
for the vanished-nodes edge ('Nothing practicable is left in that
cluster — the graph changed.'); 503 deep link unchanged. Preselected
mode gone — the dialog is the only creation entry for now (the graph-
view cluster entry comes later with free-form selection).
Sidebar: session right-click 'Practice this session's propositions'
removed (item + state + dialog wiring); the /practice nav row stays.
Tests: pipeline — chatId/any-provenance tests die, nodeIds selection
pins added (unknown-id filtering, dated-default title, title hint
rides, empty/vanished → EmptySelectionError); render — cluster picker
pins (label+count rows, notes kind, unlabeled cluster muted, empty
graph line), preselected pins removed; 16 render + 13 pipeline.
Docs: architecture.md §7 API+UI paragraph rewritten (graph-domain
ruling, /api/graph/clusters, cluster picker as the only entry, no
404-unknown-chat); README creation-flow line updated.
1130/1130, tsc 0, lint 0, build 0.
src/agent/subagents/practice.ts: createPracticeAgents(deps) with the
ExtractionModels mirror (per-stage engine overrides + engineFor +
providerKeyMissing + refuseIfStagesUnready, all three roles strong-tier).
Generator doctrine = the owner's problem-engineering framework verbatim
(minus Sources) + Part III wiring section per the green-lit amendments
(FULL SET input with minimal-use doctrine; degradation never in
meta-commentary). Partitioner: true-partition doctrine, validate via
chainCoverage + exactly-once + unknown-ids, error-back retries through
runSubmitAgent's new validate hook (ToolInputError → error tool result).
Generator: echo law (chains set === target). Grader: verdict union +
nonempty feedback (schema-enforced). renderChains/renderPropositions/
renderParams/renderPreviousProblems. BACKGROUND_MODEL_ROLES extended with
the practice trio (settings override symmetry). runSubmitAgent gains the
validate seam (null = accept). 16-test faux-engine suite incl. retry
paths, refusal gates, neverSubmitted.
amosblomqvist