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).
Owner report, 2026-10-04: starting a session with map picks showed ONE
empty user bubble, no agent turn; a follow-up woke the agent with the
right material but the marker never rendered. Two breaks, one root —
the record no longer starts empty and user-side data parts had no
render lane:
- The first-send gate checked initialMessages.length === 0; the
graph-context marker at message[0] made that false, so the ?prompt
turn NEVER fired. New pure helper initialPromptOwed (the escAbortTarget
pattern): the prompt is owed while no user/assistant message exists —
scaffolding is not conversation. Pinned in tests/graph-context.test.ts.
- The transcript's user branch rendered text parts only; a data-only
message (graph-context, practice-thinking) fell through as an EMPTY
bubble — the data-practice-thinking case I anchored to lived in the
ASSISTANT switch, dead on arrival. Data-only user messages now hoist
to a bubble-free card lane (the subagent-steer precedent); the two
dead assistant-switch cases are gone.
Gates: 1296 pass / 0 fail; typecheck, lint, build clean.
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.
The map picks no longer glue into the opening prompt as seed text
(owner, 2026-10-04 late). POST /api/chats accepts graphContext
{propositions}; the route resolves the picks against the live graph,
derives the reasoning chains among them (chainsForSelection →
formatChainLine), and writes a graph-context custom marker as
message[0] — before the learner's first user message, in either mode.
- Core: newGraphContextMessage (model-facing content = the
contextualization framing + proposition lines + chain lines;
details = the structured pick); toLlmMessages flattens it to a user
message like every marker, so the model reads it for free.
- View: projects to a user-side data-graph-context part.
- UI: GraphContextCard — minimal, muted, dashed, collapsed by default
(counts only); expand reveals name — statement rows and the chain
lines. Same compact-expandable pattern as the practice cards.
- Home: sends graphContext with the creation whenever picks exist;
the prompt param carries the learner's own words only.
Root cause of 'doesn't register': the slice-6 floating-bar edit never
landed (a silent no-match replace) — the Body accepted selectMode but
the bar ignored it, so the picker showed Clear + Delete and nothing
ever confirmed. Found by re-reading; the test suite never pinned the
select-mode bar.
The owner vision instead: the chain table's selection is PERSISTENT
and caller-owned.
- ChainTableOverlay gains controlledSelection/onSelectionChange (every
check, clear, and delete-prune reports up immediately; uncontrolled
fallback keeps the delete flow self-contained).
- Select mode's floating bar: Done (closes; the pick already
registered) — the delete fence never renders while picking.
- GraphView threads the lifted set through, resolving ids → nodes
against the live snapshot.
- Home: the pick bar above the composer shows 'N propositions
selected' + Clear in EITHER mode (dashed 'Pick propositions from the
map' hint only in practice mode with nothing picked); visible hover
(bg-accent) so the chip reads clickable. Practice sends still seed
from the pick set.
Drives handleChatRequest with the real practice agent + faux engine:
pose persists its ack and owes no continuation; a commit arrives as the
'Committed answer:' user message and the conclude turn continues to the
reveal; a nudge carries the 'Thinking trace' label and never concludes;
Esc's silent marker deactivates at once and the next input rides the
synthesized dismissed-conclude pair (shared CONCLUDE_ACK_TEXT).
- needsContinuation: a persisted problem-pose ack (terminate:true ended
the turn deliberately) owes no continuation — reloads after a pose no
longer wake a phantom turn; error results stay owed like any
interrupted tool.
- Chat header: practice sessions swap the phase chip for a quiet
Practice tag (no probe→teach arc applies) and hide the coverage ring
(practice never extracts — an ever-filling ring would lie).
- ChatSummary gains pendingProblem (activeProblemOf at projection); the
sidebar meta says 'Problem waiting' for a practice row with a posed
problem. Kind threads through Chat/[id] page.
- Home picker chip gains Clear (picks were otherwise irremovable).
- Sidebar: the three-way kind filter (Teaching/All/Practice, persisted
in localStorage) above the search; two new-session doors (Teaching →
/chat/new, Practice → /chat/new?mode=practice); practice rows carry
the dumbbell badge and 'Practicing…' busy meta.
- Draft: ?mode=practice swaps the copy (title, empty state, 'What do
you want to practice?') and startDraftSession posts kind:'practice'.
- Graph entry (home composer): teach|practice switch on the stats row's
left edge (StatsBar leadingSlot); practice mode shows the propositions
chip that opens the chain table as a picker (its floating bar becomes
'Practice these' — no delete fence); the picks seed the opening prompt
('Propositions from the map:…') while the title stays the typed text.
- PracticeDock (components/practice-dock.tsx): the active problem
replaces the composer — statement (markdown+LaTeX) above, the 'Your
thinking' composer-twin below with the two submit lanes (Nudge =
thinking without a commit; Commit = answer commit; ⌘/Ctrl+Enter
commits). Dictation claims the thinking field.
- Transcript cards (practice-cards.tsx + the artifact registry): the
problem card (compact-expandable statement — the commitment NEVER
renders), the conclude card (outcome + verdict + reveal; the
dismissal synthesis is one muted line), and the user-side thinking
register cards (Your thinking / Your solution).
- view.ts: the problem pose joins the placeholder fork (the ack never
flips the part); practice-thinking markers project as data parts;
problem-dismissed markers project nothing.
- chat.tsx: activeProblemOf drives the dock branch (question > problem
> composer), the esc layering gains the problem rung (below question,
above turn) with its discard confirm + the silent marker route; the
thinking submissions POST through the session's optimistic appends.
- 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).
The native shapes beneath the practice tools (owner design 2026-10-04):
- activeProblemOf — the conclude-before-next-problem gate's fold (a
conclude call OR a dismissal marker deactivates; the pose-time ack
never does — nudges keep the card up).
- The thinking input: {problemCallId, commit, text} → practice-thinking
marker; the label (THINKING_HINT/COMMIT_LABEL, single-sourced in
contracts) is the only register difference the model sees.
- Ask-style problem dismissal: silent route appends the UI-only marker;
the user's NEXT input prepends the synthesized conclude pair (call
outcome:'dismissed' + shared ack) so the agent reads the discard in
its own vocabulary and the slot frees.
- toLlmMessages: dismissal marker filtered; pair passes natively.
Practice sessions (owner, 2026-10-04): the marker rides the record and
the normalized summary (absent field reads as teach — pre-practice
records never migrate); POST /api/chats accepts {kind: 'practice'}.
The agent branch, tools, and UI land in the following slices.
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).
Esc over a mermaid focus, the plan popup, or the setup dialog opened the
question-dismiss confirm underneath — KeyboardShortcuts' window listener
registers first, so the overlays' own self-closing Esc listeners could
never outrun the lh:kbd-abort-turn dispatch. The state gate in ChatView
is the defense (the architecture's own pattern from the abort dialog):
escAbortTarget, exported pure, rules priority by stacking order — abort
confirm (claimed by its dialog) > mermaid (paints above plan) > plan >
setup (self-closing) > question > turn. The gate closes the overlay; the
overlays' own listeners stay (idempotent). Pinned in esc-layering.test.
DeepSeek via OpenRouter returns tool-call ids that reset every response —
a one-call-per-turn session names every call call_0. Everything keys on
toolCallId (the fold's call↔result join, the pending gate, answer
routing, the option-shuffle seed), so the Lacan session corrupted: the
quiz answer overwrote the ask card, the ask question vanished, and the
order scrambled. Record itself was always intact.
- src/server/pi/call-ids.ts (new): ToolCallIdUniquifier at the wire
boundary — colliding ids remint once (orig~2, ~3…), the same mints at
settlement; tables key by message OBJECT IDENTITY (streaming messages
carry no ids; the message_end object is reference-identical to the
state object). History never rewrites; providers see paired ids, which
is protocol-legal (pi synthesizes ids for providers that send none).
- ui-stream: runPiTurn remaps events + settles through the uniquifier
(spec.usedToolCallIds, seeded by chat-handler from the record).
- view.ts: findToolPart scans newest-first — pre-fix duplicate-id
records (the existing Lacan session) heal on next render.
- tests: 9 unit pins (call-ids.test.ts) + the two-call_0 e2e (chat-flow):
ask answered as ask, quiz as quiz, ids unique in the record.
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
Root cause: doSteerOrSend branched on turnInFlight, but the local dedup
(chat.tsx ~370) clears that flag the moment the own stream takes over —
during our own turn busy was true and turnInFlight false, so Enter fell to
the plain-send path whose busy guard returned silently. Branch on busy:
own stream AND foreign/wake turns both have a live agent registered by the
time the send's POST resolved, so the steer lands (a raced settle 409s and
falls back to the plain send).
The button now follows the text: busy + typed text = Send (click steers —
steering must be clickable, not Enter-only); busy + empty = Stop. ESC keeps
its confirm-popup abort in every state, so stop is never unreachable.
STEERING, NOT FOLLOW-UPS — pi-agent-core's two rails, used deliberately:
agent.steer() injects at the next response boundary INSIDE the live run
(one logical turn, the model changes course); followUp() would only extend
an ended run — not built. The same rail job markers already ride
(steeringMode 'all', the 2026-09-26 doctrine).
- POST /api/chats/[id]/steer: 404/400 honest guards; no live turn → 409
and the client silently falls back to a plain text send (the settle
race); live → steerLiveChat hands the message to the live agent's
steering queue AND parks it in the userSteers registry — the stranding
safety net. A steer stranded behind a question termination or abort
stays parked and rides the chat's NEXT run (handOffEligible), the same
deliver-or-requeue lifecycle as job markers. Reconciliation at settle:
delivered ids drop (reconcileUserSteers).
- Client: steer(text) queues + POSTs without flipping any turn state;
the drain's verbatim message events upsert the real user card by the
client-minted id (steer_<localid>) and the indicator fold drops the
line. canSend no longer gates on turnInFlight — Enter mid-turn steers;
Enter with an empty composer still stops; the square button stays Stop.
- The queued-steer indicator (owner spec): quiet one-liners above the
composer shell — right-aligned to where its corner rounding starts,
truncated with an ellipsis, fixed max width, ~half opacity, no
background, no container. Tunable later.
- Verified from pi-agent-core source: steering drains per completed turn
inside the loop; injected messages emit message_start/end verbatim;
finishTurn {action:'end'} (our question stopper) exits BEFORE the
steering poll — hence the registry.
Live test finding: answer A then dismiss B woke the agent — the continuation
fold treated a mixed [answered, dismissed] trailing batch as 'reaction owed'
(all-dismissed rule only). Dismiss A then answer B and dismiss-both were
already correct.
The law now: the trailing batch's LAST resolution decides. Ending on a
dismissal = the user signaled stop — no wake, the answers' reactions ride the
next plain message. Ending on a real answer = the reaction is owed. One law
with the pump's pause (endsInDismissal reads the same last result), so
continuation, delivery-holding, and the pause can never disagree. The wrong
pin ('a real answer in the batch still demands the reaction') flipped.
The incident: deepseek batched quiz + ask_user_question in one assistant
message; the quiz executed first (sequential) and terminated the turn; the
answer to the quiz IMMEDIATELY woke a reaction turn (fork-1 design assumed
one dangling question max), the model complained about the unanswered ask,
and the ask became permanently unanswerable — its toolResult could only
append AFTER the intervening assistant message, a shape DeepSeek rejects.
The gate: for answer inputs, if a question call still dangles after the
append, persist write-ahead + emit record-changed + return the settled
terminal — no turn. The dock advances to the sibling; the LAST answer fires
the one reaction turn with every result adjacent to its calling message.
Nothing else generates while a question dangles (job wakes wait, continuation
folds refuse, the composer gates), so adjacency holds by construction.
Also repaired the stranded session data (ask call stripped from the batch;
backup at /tmp/chat_murkbeq2_backup.json).
- {type:"dismiss"} input event: parse + pure appends with the answer-event
orphan discipline (dangling-call 400, kind match, answer must be dismissed)
- POST /api/chats/[id]/dismiss: the no-turn composition — write-ahead append,
409 while a turn runs, emits record-changed (new bus word; idle views sync)
- needsContinuation: an all-dismissed trailing batch demands nothing;
endsInDismissal = the derived pause predicate (tail-based, no latch)
- wake pipeline: while the record ends in a dismissal the pump holds every
delivery in the steer queue — the user's next send is the wake (until the
text-turn steer claim lands, settlement re-pump delivers right after it)
- session-core.dismissQuestion: optimistic skipped card + silent POST;
chat.tsx confirmAbort question-mode swaps submitAnswer for it; ESC-abort
popup everywhere (AbortTurnDialog) with popup-Esc claimed via state gate
- parallel question calls dismiss call-by-call (results bind by id)
- 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
Consistency ruling (owner, 2026-10-01): the finish-first send and the
run-end focus are properties of the COMPOSER, not of one page.
- home (graph view, raw textarea): topicRef mirror; start() finishes
dictation first (before the setup/key guards, so an intercepted draft
is the complete one) then reads the mirror; the start square stays
live while a run holds deferred speech; run end focuses the topic
input.
- new-session: the shell owns the dictation instance — handleSend wraps
the parent's send with finish-first, the composer stays sendable
during a run, and run end focuses it via the Composer's inputRef. The
parent's send reads a textRef mirror (applyText at every write): the
flush lands via setState after the shell's await, and the parent's
closure must not read a stale draft.
Two owner rulings, 2026-10-01 ('nothing gets lost' round):
1. SUBMIT FINISHES DICTATION FIRST. Every submit path — the dock's
answer (Enter/Cmd+Enter/button) and the composer's send — stops any
live run and lets Deepgram's last finals land in the target field
BEFORE building the payload. Dictation.finish() resolves at cleanup,
post-flush; a module finisher registry (same pattern as the question
target) hands it to the dock without a prop bridge. Field values are
read from REF MIRRORS (thinkingTextRef/answerTextRef, the composer's
textRef): the flush lands via setState after an await, invisible to
the closure. The quiz flow the owner described now holds: dictate
reasoning → pick answer → pick confidence → Enter → dictation ends,
reasoning saved into the thinking field, THEN the answer submits.
The composer stays sendable while a run is live (its spoken text is
deferred in the run's buffer) so Enter can trigger the finish-first
send at all; nothing said → the send no-ops.
2. RUN END FOCUSES THE BARE COMPOSER. The falling edge of
dictation.active focuses the composer's textarea — only when the
composer is mounted: a pending question unmounts it (ref null, no
focus), so the thinking field is deliberately left alone. Composer
gains an optional inputRef for the page's focus handle.
The rail paint is extracted to components/confidence-rail.tsx (ramp,
tick stack, geometry laws) and shared by both surfaces so they cannot
drift: the dock's live slider paints the rail under its disc, the
answered card renders it disc-less — the missing handle itself says
'record, not control'. Same look otherwise: thin rail, stop color,
solid ticks, % read-out on the right, dock-mirrored placement (below
the options, above the thinking).
Renders only where a confidence exists: quiz answers with one sent;
never on 'I don't know' (none was sent), never on explain/ask, never
on pre-feature records (optional field, absent).
Round-4 regression, root-caused:
- The rail overlay (position:absolute) painted above the in-flow input
no matter the DOM order — CSS paints positioned elements after
in-flow ones — so the bar sliced across the disc ('disc behind the
bar').
- The wrapper's height was unpinned: the rail's top-1/2 centered
against a stretched box while the disc rode the input at the top —
vertical drift, which also exaggerated the disc's overhang and made
the original-size 16px disc read oversized.
Fix: the input is absolute inset-0 too (DOM order now governs — the
disc rides above the rail), and the wrapper is pinned to h-4 so rail
center and disc track share one box. Geometry otherwise unchanged from
round 4: rail spans the thumb-center range, 0/100 flush at the very
ends, solid ticks, no rim.
- 0/100 at the very edges: the paint moved off the input onto a rail
overlay spanning exactly the thumb-CENTER range (inset 8px = half
the 16px disc per side) — the disc rests flush against the rail's
ends, no bare stub poking past it. The input is now pure interaction:
full width, transparent, 16px-tall hit area, native drag.
- Ticks solid, not dashed; plain p% positions (the overlay IS the
center range), still background layers under the fill.
- Typing guard (owner): a range input is NOT a typing field — after a
click or drag on the slider the ↑/↓ option walk and ←/→ confidence
cycling keep working. Only real text fields (thinking, answer,
search, renames) own their keys. The slider's own keydown is gone —
the global machine owns arrows, and its preventDefault cancels the
native 1% step.
- Disc rim reverted: plain bg-current disc again.
Owner refinements 2026-10-01:
- Bar back to the original h-1.5 rail (colors unchanged).
- Disc bleeds into its own fill no more: 2px #141414 rim on the thumb
reads as a gap wherever it sits.
- Dashed stop ticks rebuilt as BACKGROUND LAYERS on the input itself:
#141414 base color, four dashed tick gradients (30/50/70/90 only —
the rail's ends already say 0 and 100) sized 1px x 100% at the
thumb-center inset, and the fill gradient now TRANSPARENT past the
thumb and painted on top — ticks live inside the bar, top to bottom,
visible only on the empty side, never over the filled portion.
Root-cause find: the sidebar 'highlighting' on ←/→ was :focus-visible —
pointer-clicked elements match it only after a keypress, so any key in
the quiz lit the still-focused session row. The arrows never moved
focus. Fix: the no-ring law (owner) — *:focus/:focus-visible outlines
removed app-wide in globals.css; surfaces keep their own affordances
and the dock's keyboard walk keeps its ring utilities.
Slider repairs:
- mousedown-preventDefault had killed click + drag in Chromium —
removed (with tabIndex -1); the slider is natively interactive again,
and the ring law makes its focus invisible anyway.
- Focused arrows step the STOP index on the input itself (native would
step 1% and snap straight back — dead keys); the global handler
skips INPUT targets, so no double-step. Enter still submits.
- Empty side: var(--muted) matched the card and vanished — now #141414
(the owner's main-background pick), and the pill carries a border
again (border-border).
- Dashed tick per snap stop below the bar, inset by the thumb's
half-width so each tick sits under the thumb's center at that stop.
amosblomqvist