Core-survival labelings (vision section 1, DECIDED 2026-09-13):
ClusterLabeling records in a new append-only ledger
(data/graph/cluster-labels.jsonl), immutable, id = clu_ + sha256 of the
sorted core so re-naming races and re-runs dedupe at the store seam.
matchClusterLabelings (core/clusters.ts, pure): a community keeps a
record's label while at least half its core survives in it; both sides
single-successor (a split keeps one name; greedy by overlap, ties to
the newer record).
The namer (src/agent/subagents/cluster-namer.ts): ONE singular cheap LLM
call (new 'labeler' role, glm-flash/low) names every new/changed
community at once, given member names + statements (capped at 12
members per cluster). refreshClusterLabels runs POST-EXTRACTION
(user-decided) from both front doors, fire-and-forget — naming failure
never fails the run; refreshes serialize behind a chain so parallel
imports never mint duplicate labels. import-notes.ts awaits
clusterNamingSettled before exit.
Payload: GraphSnapshot.clusters (>= MIN_CLUSTER_SIZE 3) joins detection
with labelings at read time — claimed clusters carry the record's label
and id; unlabeled ones carry a member-hash id and null label. Detection
is deterministic so no memo in v1 (ms at current scale; revisit at
vault scale).
Tests (372/372): matcher half-rule/split/merge/greedy/tie cases; naming
end-to-end through importNotes (one call, one record, claim on
re-refresh), naming failure non-fatal; snapshot derivation with
labeling join; 11 test pipelines now pin an explicit labeler seam.
detectCommunities(nodes, edges) in src/core/clusters.ts (new dep
graphology-communities-louvain): Louvain modularity over the undirected,
unweighted follows+supports structure (user-decided 2026-09-13). The
random-walk traversal runs on an inlined seeded mulberry32 so the same
graph always yields the same partition; community indices are still
unstable across edits, which is exactly why labelings (phase B) key to
sorted MEMBER SETS. follows+supports between the same pair collapse to
one edge; self-loops and dangling edges are skipped defensively; no
min-size filter here (the >=3 render threshold is a presentation rule).
Tests: two-cliques-plus-bridge split, determinism, singleton/empty
graphs, pair collapse, defensive skips, member sorting. 363/363.
Real-data smoke over the current notes graph: 338 nodes / 211 edges ->
138 communities, modularity 0.95, and the big ones are exactly the
learn-session topics (databases, differential forms, sales, servers,
vim, copulas...). 32 communities clear the >=3 render threshold.
The most suitable default is the stricter end — labels only once
zoomed in past 2x fit (user-decided 2026-09-13). The slider window
extends below it (0.25x) so the default is no longer the floor.
The label-visibility zoom threshold (was the LABELS_CAMERA_RATIO=1
constant) becomes the fourth persisted setting: labels appear when the
camera ratio is at/below it (1x = fit zoom; slider 0.5x-4x, shown with
an x suffix). Deliberately THE single zoom knob per the cluster design
(vision section 1): cluster labels will fade in over the same span as
node labels fade out, gated inversely on this same setting.
Live-verified: at a zoomed-out camera (labels previously hidden),
raising the threshold to 4x rebuilds and re-shows labels without any
camera movement; resetting restores the fade. 355/355 tests,
typecheck, lint clean.
Three sliders behind a gear in the search panel (user-decided
2026-09-13), persisted per browser in localStorage with clamped
sanitizing on load. Default node size dropped 8 → 6 for dense graphs.
Mechanics: the mount effect is split — the data effect builds the
graphology graph + one frozen forceAtlas2 pass per snapshot; the render
effect mounts sigma and re-runs on settings change only. Sizes are
stamped onto the same graph instance (layout never recomputed), and the
arrowhead program is constructed per rebuild with ratios solved from
the settings (head px = thickness × lengthToThicknessRatio × 2; wideness
keeps the stock 0.8 shape). The camera position carries across rebuilds
via a ref, so a slider tweak never re-fits or re-lays-out the view;
useDeferredValue smooths slider drags.
Live-verified in the browser: sliders persist to localStorage, the
renderer rebuilds with zero console errors, hover/picking survives the
rebuild, and a zoomed-out camera stays zoomed out across a settings
change (labels-gate opacity as the observable). 355/355 tests,
typecheck, lint, build green.
Diagnosis from the real 25-file notes import (user-pasted session logs):
almost every run hit 'Validation failed for tool submit_findings/review:
edges must have required properties edges' — the model (glm flash, low
thinking) omits top-level fields when they're empty, and pi-ai's
structural JSON-Schema validation runs BEFORE the zod layer, so each
omission cost one of the extractor's maxTurns 4 (full payload
regeneration) and occasionally burned out the run entirely (the two
FAILED imports).
The model is technically wrong, but the schema provokes it:
required-but-usually-empty arrays are a known weak-model anti-pattern.
- lhTool's model-facing declaration now converts with io: "input"
(the declaration describes what the MODEL sends): .default() fields
drop out of required. No-op for teacher tools (none use defaults).
- findingsSchema.edges / auditSchema.dropped: .default([]) — omission
means "none"; attribution: .default({quiz:{}}) for quiz-free sessions.
- Prompts hardened in all three agents: include every top-level field,
never omit because empty.
Structural probe: validateToolArguments now passes the exact pasted
failure shape; zod re-parse in execute applies the defaults. 355/355
tests, typecheck, lint, build green. (lh-tool.ts is the shared engine
seam — flagged to master: io:"input" is the correct mode for every
model-facing declaration, teacher path verified unaffected.)
The same extractor→auditor→merger pipeline now runs over raw personal
notes in arbitrary formats (vision §1, DECIDED 2026-09-13):
- NotesSource records (data/graph/sources.json, deterministic id from
path hash → idempotent re-imports) play the chatId role in mappings,
checkpoints, and observation events; usage lands on the source record
(appendSourceUsage, idempotent per startedAt, kept across upserts).
- mergeFindings is chat-store-free: callers pass MergeSource
{messages, foldedAt} — sessions pass the chat record, notes pass
messages:null (no quiz-event rebuild; a deterministic rebuild needs
structured quiz tool parts).
- Second extractor prompt (NOTES_EXTRACTOR_INSTRUCTIONS) for general-
purpose extraction; understanding judges the AUTHOR's grasp. Auditor
wording input-agnostic. extract/audit share one generic runSubmitAgent.
- graphSnapshot.fileSources (srcId→filePath); notes sources render as
static path entries in the detail card (FileTextIcon, never a link).
- One shared pipeline instance (exported from chat-handler): same
single-flight registry, live status chips (label = file path), usage
accounting, and true abort for both front doors.
- Transport: POST /api/graph/import (multipart, 3-wide worker pool) +
Import notes… button in the graph view; scripts/import-notes.ts for
batch runs (concurrency 4).
- real-data-extraction/ gitignored (raw import sources never committed).
355/355 tests; typecheck, lint, build green.
Community detection as the graph's at-a-glance structure: Louvain on
graphology, seeded; structure-based (never layout-based); entirely
inferred from the graph, entirely separate from extraction; size-
gated (>= ~3 nodes); named by one cheap cached LLM call per new/changed
cluster.
Labeling persistence — the core-survival rule (user-decided): each
append-only labeling record stores the label + the member set it was
produced with (the core); a recomputed community keeps its label
while >= half its members are from that core, relabels under half
(new record appended, old records stay as history); single successor on
splits/merges; color keys to the labeling record.
Rendering: dashed hull + blurred low-opacity fill behind everything,
per-cluster stable hue, cluster labels inverse-gated against the node
label span (domain view zoomed out / proposition view zoomed in);
boundary geometry computed at draw time (pure function of member
positions) so future physics never invalidates it; inter-cluster edges
stay visible (meaningful bridges).
Multi-scale clusters (hierarchical Louvain dendrogram) deliberately
parked. Open questions: boundary aesthetic/label placement (visual
testing), threshold tuning (synthetic benchmark — now the primary
clustering-quality validation, sparsity being the honest risk).
Live-driving the real browser caught it (the user's report: "they don't
work as buttons"): the stats row sits OUTSIDE the composer shell's
pointer-events-auto div, inside the sticky scroll wrapper's
pointer-events-none — every control on the row was unclickable, including
the pre-existing stats-detail popover button. The row now re-enables
pointer-events itself.
Also, per the user's sizing note: bar text 10px → 11px with a touch more
breathing room, and the model + thinking controls gain quiet pill
affordance (rounded-full border, card bg, ▾ chevron) — still muted by
design, but visibly buttons.
Verified live end-to-end in the browser: click opens the model menu ABOVE
(366 rows, search autofocused), "qwen flash" narrows to 5, picking
qwen3.8-flash persists + flips the label, the thinking menu shows exactly
that model's supported levels (Off/Minimal/Low/Medium/High), Medium
persists on the record via PUT, labels update. Test session settings
restored after the drive. Gates 350/350.
The model-control wave (user-decided design, 2026-09-12): ONE controller
for the conversation agents (prober/teacher are functionally one agent);
background sub-agents stay config-pinned. pi-ai owns the registry — the
shipped 366-model OpenRouter catalog (cost, context, modality,
thinkingLevelMap), getSupportedThinkingLevels/clampThinkingLevel for the
per-model thinking truth, and streamSimple's automatic clamp. No second
registry to drift out of sync.
Seams:
- config.ts: ReasoningEffort retired for pi's ThinkingLevel
(off|minimal|low|medium|high|xhigh|max). GLM roles go honest at "low" —
pi's catalog says GLM-5.3-flash accepts exactly low|high|max, and the old
"minimal" was silently clamped to low by streamSimple since the swap;
nothing changes behaviorally. visualMaker → qwen/qwen3.8-flash (fixes
the dead qwen/qwen3.8-max slug — the catalog only ships -0902; also the
user's cost pick, 0.15/0.47 per-1M).
- src/agent/models.ts: piEngineFor(role, settings?) — provider+id resolved
from the catalog, thinking level from settings ?? config, CLAMPED to the
model's supported list (a settings change between models can never send
an invalid level). PiEngine gains `level` (observability/tests).
- ChatRecord.settings {model?: {provider, id}, thinkingLevel?} — written
by PUT /api/chats/[id]/settings (replace-whole, no updatedAt churn),
read by chat-handler at turn composition: agents are stateless per turn,
so the change is next-turn-effective by construction. The turns log
still records the ACTUAL billed model (the honest witness).
- GET /api/models: the configured-provider catalog as client-safe DTOs
(core/contracts/models.ts) + config defaults; a second provider joins
the picker by registering it in piModels() — nothing else changes.
- PUT settings validation gate (src/server/model-catalog.ts): unknown
provider/model or unsupported level → 400 carrying the supported list,
never a silent clamp.
- scripts/sync-pricing.ts: regenerated PRICES/CONTEXT_LIMITS from pi-ai's
shipped openrouter.json (offline; same source as the engine + picker).
All 366 models priced; glm's stale 0.15/0.5 → pi's 0.075/0.25, limit
1310720 → 1048576 — cost is inferred at read time, so history re-prices.
- UI (components/model-picker.tsx): stats-bar right side becomes model +
thinking buttons; menus open ABOVE; fuzzy search over "provider id
name" (core/fuzzy.ts — pi's /model pattern); model rows show name/id/
price/context; the thinking menu lists exactly the chosen model's
supported levels (GLM: low|high|max — no dead options, no "off").
Model switch drops an incompatible pinned level (falls back to config,
which clamps honestly) rather than carrying a 400.
Tests (350/350, +15): catalog projection + defaults, the settings gate,
pi thinking truth pinned against upgrades (glm low|high|max; minimal→low;
off→low — GLM cannot disable thinking), the engine factory's
override/fallback/clamp/throw matrix, fuzzy ranking, store round-trip,
and the re-priced folds (usage, stats bar, activity dock).
Live smoke (disposable driver, deleted): catalog route 366 + defaults;
glm+minimal → 400 + supported; turn 1 billed qwen/qwen3.8-flash under
the override; mid-session switch flipped turn 2 to glm; clear-overrides
restored the config pin. ALL CHECKS PASS on real OpenRouter.
The custom-message scheme (user-decided 2026-09-12): system-authored
user-role messages are NOT human messages — they carry marker data parts
only, the adapter owns their model-facing wording, and the UI renders
them under control. Nothing fake is ever persisted as human speech.
The inventory (all migrated):
- probe→teacher handoff wake: was a text-bearing user message rendering
as a fake user bubble; now data-probe-handoff marker-only — the
adapter flattens it to HANDOFF_TEXT (core/contracts/chat.ts), the UI
renders a quiet 'Preparing your plan…' system line.
- plan accept/decline: was boilerplate text ('The plan is accepted…',
'I'd like the plan revised…') masquerading as learner speech — the
decline wasn't even marked; now data-verdict {accepted, at} on its own
message, adapter flattens to verdictText(), UI renders the quiet
verdict marker ('✓ Plan accepted' / '↻ Plan revision requested').
sendVerdict sends TWO REAL MESSAGES when the learner typed a note
(the note as a genuine user message, then the verdict) in ONE turn
via setMessages + sendMessage() — no 409 dance, no double teacher turn.
- sub-agent steers: the abort steer gains its data-subagent-aborted
marker part; the UI keys steer rendering off PARTS (artifact/aborted),
not metadata — metadata stays as pump mechanics only.
- record-adapter: systemTextOf() flattens marker-only messages; steers
keep their text (payload — briefs are content, not boilerplate).
phases.ts: plan→teach advances on data-verdict{accepted:true} only;
learner text alone can never sanction teaching (pinned). data-plan-
accepted retired, no back-compat (user-decided: wiped-data doctrine).
Pinned: record-adapter.test.ts (verdict/handoff flattening, note+verdict
order, steer payload kept, UI-only parts skipped), phases.test.ts
(verdict marker semantics, out-of-order, text-never-fakes), jobs.test.ts
(abort marker part), demos.test.tsx (badge renders off the part, not
metadata).
Live GLM smoke (disposable driver, deleted): probe→conclude→wake→plan→
decline-with-note (revised, still plan)→accept-with-note (teach) — ALL
CHECKS PASS, including marker-only handoff/verdict on the wire and zero
fake human speech in the transcript.
Gates: tsc, lint, build, 335/335.
The architectural answer to the validation-crash question: validation IS
first at the model-facing boundary (pi validates arguments before
execute; failures return to the model same-turn — all three live
incidents retried successfully). But the same bytes also exist UNVALIDATED
by design in two other places: the live wire streams raw tool-input
deltas DURING generation, and the transcript keeps the raw input of
REJECTED calls (honesty + the replay invariant). The consumers were
supposed to trust part STATES — but that trust was enforced by
TypeScript casts ('part.input as X'), which are compile-time lies with
no runtime check behind them. That is the hole both crashes came
through, and it should not have been possible.
The structural rule now: model-authored INPUT is consumed only through
parsers running the SAME zod schema the model faced — boundary-valid ⇔
consumer-parseable, mechanically. Server/learner-authored OUTPUT stays
trusted (our code authors it).
- src/core/tool-parts.ts (new): parseQuizInput (display parse over the
shared quizInputBase — strips the teacher-only nodeId, accepts both
prober and teacher boundary shapes), parseAskInput, parseProbeDoneInput.
- contracts/probe.ts: probeDoneInputSchema is now core-owned zod — the
probe_done tool declares it (single source of truth), the seam runs
the literal same schema.
- question-dock.tsx: the dock parses its input at entry; unparsable
renders the quiet malformed-question fallback, never a crash.
- probe-done-card.tsx: the input preview goes through the parse; the
old state reasoning and shape guards collapse into it.
Pinned in tests/tool-parts.test.ts: both exact live crash shapes
(stringified quiz options, stringified assessment), boundary-parity
cases for all three schemas, nodeId stripping, mid-stream partials;
render-cards pins the dock fallback markup.
Gates: tsc, lint, build, 328/328.
Two live bugs from the modal-realism stress runs:
1. Multi-select ask: picking the custom 'Other…' row cleared every
already-picked option, and picking any option cleared a typed custom
answer — single-select displacement logic applied to both modes. The
dock's selection is now a pure state machine in question-core
(toggleOption/toggleOther over DockSelection, unit-tested): single
displaces (unchanged), multi coexists — neither channel clears the
other; the typed text always survives.
2. ProbeDoneCard crashed on a REJECTED probe_done call: the model
stringified 'assessment' (the same double-encoding habit as the
options bug), zod rejected the call, the model retried — but the
card's input-preview fallback trusted the errored part's unvalidated
input shape and died on assessment.strands.map. The preview now reads
input only while the call is pending or streaming (guarded with
Array.isArray); an output-error part renders a quiet
malformed-and-retried note instead of the garbage.
Pinned: question-core.test.ts (3 dock-selection machine cases),
render-cards.test.tsx (4 probe-done card cases incl. the exact live
stringified shape, plus the multi-select ask markup pin).
Gates: tsc, lint, build, 315/315.
Live observation (modal-realism run): the teacher batched two quizzes in
one turn; the dock showed the SECOND question first, so the answering
order was reversed vs the saved record and the transcript cards.
Cause: findPending scanned parts BACKWARD within a message, so with two
input-available quizzes it picked the last one first. No overwriting was
involved — parts are created in call order on both wire and record; the
scan direction alone reversed the interaction order. (pi executes a
parallel batch with starts + result messages in call order; the
execution-end order is completion-racy, which also allowed a millisecond
dock-subject flip mid-turn.)
Fix A — src/core/pending.ts: parts now scan FORWARD within a message
(messages stay newest-first): the FIRST unanswered question is the dock
subject; the second appears only after the first is answered.
Fix B — src/agent/lh-tool.ts: question tools opt into sequential
execution (pi runs the batch in call order when any called tool opts
in), so tool-input-available chunks always arrive in call order and the
dock subject can never flip mid-turn. Question executes are instant.
Pinned in chat-flow.test.ts: a parallel two-quiz batch asserts wire
order (A's available chunk before B's), part persistence order, and the
dock selection sequence (A first, then B after A is answered).
Gates: tsc, lint, build, 307/307.
Live crash (modal-realism session): GLM double-encoded the quiz options
array as a JSON STRING inside the tool-call arguments. The bridge emitted
tool-input-available at toolcall_end — BEFORE validation runs — so the
client briefly saw the malformed quiz as a pending question and the dock
crashed in normalizeOptions ((string).map). The later tool-input-error
chunk could not revive the tree.
The old SDK gated this with InvalidToolInputError: validation ran before
the input-available chunk. The bridge now holds the part in
input-streaming until tool_execution_end and only then emits
tool-input-available (clean) or tool-input-error (invalid, structural
JSON Schema failure or the zod re-run inside execute).
Server-side behavior was already correct (the record shows the error
part + a successful model retry) — this is purely the wire-ordering hole.
Pinned in chat-flow.test.ts: the refinement-layer retry (correctAnswer
mismatch) and the exact live structural shape (string options) both
assert the invalid call never becomes input-available on the wire.
Gates: tsc, lint, build, 306/306.
The Vercel AI SDK survives only as Band D (useChat client + UIMessage
record types); @openrouter/ai-sdk-provider is uninstalled.
- src/agent/models.ts: PiEngine {model, streamFn} via createModels +
openrouterProvider subpath; thinkingLevelFor maps config effort none->off
- src/agent/lh-tool.ts: neutral LhTool (zod authoring, optional execute,
clientQuestion flag); toPiTool converts via z.toJSONSchema and re-runs
zod inside execute so refinements ride the error-result retry path
- src/server/pi/record-adapter.ts: UIMessage records -> pi Messages
(dangling tool calls get synthesized interrupted results)
- src/server/pi/ui-stream.ts: SSE bridge emitting AI-SDK-shaped chunks so
the useChat client parses unchanged (continuation id rule, question
tools never emit output, tool-input-error for bad authoring)
- teacher-agent.ts: stateless per-turn pi Agents; shouldStopAfterTurn
stops only when every question-tool call resolved cleanly (errors retry)
- chat-handler.ts: buildAgent(outgoing) factory seam (stateless doctrine);
assertValidUiMessages owns the retired SDK validation gate
- sub-agents via shared runSubagent (pi-run.ts); extraction stages stop
after a clean submit; visual-maker keeps its hand-rolled loop over
direct streamSimple calls; aborted model calls bill no usage
- tests: all model seams ported to pi-ai faux provider (scripted
FauxResponseSteps, agent factories); handleChatRequest second param
is now an agent factory
Gates: tsc, lint, build, 305/305. Live GLM smoke (disposable driver,
deleted): prober streaming + persistence, question seam, turn-2
continuation same message id, per-turn real token usage - ALL PASS.
Data wiped at the pivot (chats/graph/jobs, .gitkeeps kept).
Two user-reported problems from the muscle-protein-synthesis session,
both diagnosed from the live data (chat_mtxeml9e_6118e59f).
The freeze: after probe_done, the open session never showed the
teacher's wake turn — only navigating away and back loaded it. Not an
agent-switch problem: the server-side handoff worked perfectly
(write-ahead persist, liveness registry, the sidebar's busy dot saw the
wake the whole time). The gap was client attach: the only mid-session
wake detector the open page has is the jobs poll, and probe-phase
sessions have zero job traffic, so that poll is OFF; the post-turn
adoption poll read turnInFlight=false at ~0ms (the wake fires at
~0.75-1.2s, after the pump's grace), adopted, and stopped. Fix: the
post-turn adoption effect now keeps a short WATCH WINDOW after adopting
(~4s, 8x500ms) — if turnInFlight flips true, it hands off to the
existing wake-adoption poll (busy shimmer, adopt at finish).
The mermaid failures: four failed plan_init calls, all caused by ONE
hidden rule — underscores are forbidden in node ids (mps_translation,
net_balance, per_meal). The id regex rejected '_' BEFORE the slug check,
so the specific, actionable message never fired; the generic 'can't read
the edge — expected a[title] --> b[title]' lied (the syntax was exactly
that), and the teacher spent three attempts hunting punctuation it
suspected (colons, commas, multi-word labels) — visible in its own
reasoning trail. Per the user's principle — enforce only the slugs,
free everything else: underscores are first-class slug chars; the id
CANDIDATE class is broad so every bad id reaches the specific slug
message; chained edges (a --> b --> c) expand to pairs; ==>, -.-> are
edge synonyms; any direction header accepted (we layout ourselves);
id["title"] quote pairs are stripped. The tool description, zod
describe, and the teacher's planning prompt now state the one strict
rule instead of the old subset liturgy.
Verified: tsc, 305/305 (the exact muscle-protein mermaid is now a
passing test), lint, build; real-browser puppeteer smoke on live GLM —
the teacher's wake turn appeared in the open page ~50s after the probe
concluded with ZERO navigation, verdict gate + first teach check +
plan window all live in-page (the single console 404 is the app's
missing favicon, environmental noise).
The accept-plan UI lands (docs/probe-teach-split.md §4, formerly the
interim rule): presenting the plan no longer moves the phase — the
learner's ACCEPT does. While the presented plan awaits its verdict, the
composer's send square becomes two stacked buttons (green Accept plan /
red Decline); typed text rides the verdict as his note, Enter keeps
sending plain side questions (phase unmoved), decline asks for a
revision and the buttons stay.
The seam: accept sends a user message carrying a data-plan-accepted
part (PlanAcceptedData in contracts/chat.ts, TeacherUIDataParts
'plan-accepted'); phaseAt3 reads it as the plan→teach event (plan_init
demoted to presentation-only); convertToModelMessages drops data parts,
so the teacher sees only the text. Mid-lesson rewrites need no fresh
verdict — acceptance happens once, at the first presentation.
The teacher's planning prompt now ends the turn at presentation and
waits for the verdict (never teaches into an unaccepted plan).
Gates: tsc clean, 304/304 (phases tests rewritten to the acceptance
seam incl. decline/side-question/failed-init cases; composer verdict
render tests), lint, build. Live GLM smoke ALL PASS: phase pinned at
'plan' after plan_init, presentation turn carried NO check, accept
click flipped the summary to 'teach', then tagged checks moved the map
(current=rates, visited=composite); zero console errors.
The successor of the injected plan blocks (docs/probe-teach-split.md §6):
always on once a tree exists, a zoomed-in window (2:1, ~2¼ node widths)
centered on the CURRENT node, re-centering by pure derivation as tagged
checks move the lesson — statuses derive live from the transcript on
every render (derivePlanStatuses), glow on current, fallback focus on
the lesson's first node, click opens the existing full-DAG popup.
Bottom-left, mirroring the activity dock's bottom-right; hidden on
narrow viewports; the learner's visual polish pass comes later.
PlanDiagram gains a viewBoxOverride (exported); latestTree exported.
Tests: viewBox centering/clamping (pure) + render (hidden pre-plan,
windowed viewBox, glowing current, starting fallback).
Emission runs at persistTurn — after the whole turn is composed — so
appending the block buried it under the same turn's teaching prose and
quiz calls (user-observed in chat_mtwgb9lj: the DAG appeared only after
the first check). Splice it in immediately after the (last) successful
plan_init part instead. Two positional tests added.
Diagnosis (2026-09-11 user testing, minimal-effort GLM): sessions stuck on
'plan' forever. Two causes: (1) the three-phase view read the stored tree's
currentNodeId, which nothing ever writes post author-once (2026-09-03) — a
structural dead seam; (2) the teacher stochastically left quiz.nodeId
untagged, and done-by-quiz-result semantics would have falsely completed
nodes anyway.
The redesign (docs/probe-teach-split.md) abolishes derived occupancy in
favor of explicit events:
- Two agents over one transcript: createProberAgent (quiz without a
nodeId field, ask, probe_done) and createTeacherAgent (full toolset,
plan_init onward). The chat-handler construction branch keys on a
probe_done part in the transcript. MODELS.prober added.
- probe_done: server-side seam tool — carries the learner assessment
(strands + free-form edge read), refuses while a question is open or
the probe is already concluded, and queues a handoff steer through the
generalized wake pipeline (steers now dedupe by id, not jobId; job
finalization only for job steers). The teacher auto-wakes and authors
the plan; the learner sees the assessment via the probe-done card.
- phases: probe ends at a concluded probe_done, plan ends at a
successful plan_init — strictly sequential, no skipping, no legacy
compatibility (data wiped at the pivot per doctrine).
- Plan map = position, not completion: derivePlanStatuses reads only
tagged quiz CALLS (state/outcome feed nothing); current = latest
tagged node with consecutive duplicates collapsed, visited = earlier
distinct stays (a revisit stays in both — current wins at render).
PlanTree is structure-only; currentNodeId/doneNodeIds and the dead
mutation helpers are gone. No fallback current.
- Prompts split per agent (prompts/prober/, prompts/teacher/ over the
trimmed shared base); graph-context injection removed entirely
(shelved, docs/probe-teach-split.md §5); knowledge-graph context no
longer reaches either agent.
- Plan blocks are full-only snapshots at plan_init (neighborhood
machinery retired; the persistent DAG window is the future UI wave);
usage capture unchanged.
Gates: tsc clean, 294/294 tests (phases/plan-tree/plan-blocks/question-
tools rewritten to the new semantics; probe-done + handoff-wake tests
added), lint, build. Live GLM smoke PASS: prober concluded with an honest
assessment, teacher auto-woke, authored a 10-node plan with the block
attached, then tagged checks on slope-instinct → secant derived current/
visited exactly as designed; grading honest both ways (driver off-by-one
proved a wrong answer grades false).
Live smoke exposed two bugs in the turn usage log:
- onFinish messages carry no ids (withMessageIds fills copies inside
persistTurn), so fresh turns keyed messageId "" — provenance lost.
- The SDK coalesces continuation turns into ONE response message (a whole
probe conversation lands in a single assistant message), so per-messageId
idempotency swallowed every continuation turn's usage.
TurnUsage gains id: string (unique per request, the append idempotency
key); messageId stays as transcript provenance, taken from the id-filled
list. appendTurnUsage dedupes by id; continuation turns with the same
messageId each keep their usage.
- src/core/plan-blocks.ts: pure emission builder — full block on successful
plan_init, neighborhood block when the derived plan state moved since the
newest existing block; the as-of state reads from the block's frozen
snapshot (transcript quiz parts mutate in place on answer, so
re-derivation is post-mutation blind). Idempotent by construction.
- turn-service persistTurn enriches with withPlanBlocks at the canonical
write; chat.tsx adopts the record once the server agrees the turn is
done (endTurn flips only after persist).
- chat-handler: per-request usage collector via agent-level onStepFinish
(StepResult.usage), appended at onFinish to ChatRecord.turns via
appendTurnUsage, idempotent per response message id.
Two-part root cause (found by the lh-ui stream + master together): (1)
src/store/jobs.ts updateJob unconditionally overwrote updatedAt with now(),
silently clobbering any explicit patch timestamp — JobPatch types it as
patchable but the adapter ignored it, so order-sensitive tests rode on
same-millisecond tie-breaking (failed ~4/5 in isolation). The store now
honors patch.updatedAt ?? now() — semantics-preserving for every existing
caller, which never passed updatedAt explicitly. (2) the activity-feed
test's finally wiped the ENTIRE shared data/jobs dir — with node:test
running test files as concurrent processes, that deletes other processes'
live fixtures; cleanup is now scoped to the test's own created ids.
POST /api/activity/abort landed on main (fde46c2). 404 not_found /
409 not_running mean the process reached the desired end state on its
own — refresh and clear the abort state, no error surfaced. Genuine
failures (network/400) show 'Abort failed — try again' on the blob.
The plan now lives IN the chat (plan blocks, wave-4 item 3) and background
processes live in the activity dock (item 1), so the right sidebar's
tree panel + sub-agent fleet column is retired from the chat view:
- delete components/tree-panel.tsx (three-mode rail/list/graph panel) and
components/sub-agents-panel.tsx; their graph interaction machinery lives
on in plan-block.tsx and fleet observability in activity-dock.tsx.
- app/layout.tsx drops TreePanelModeProvider (the mode no longer exists).
- chat.tsx drops the xl:flex right column; the middle column is the view.
- tests: render-tree-panel.test.tsx deleted; the sub-agents panel suite
removed from demos.test.tsx (dispatch cards' jobs-context stays — the
central poll drives transcript dispatch cards).
- components/plan-block.tsx: hand-built SVG (user decision) — deterministic
layered layout (Kahn topo + longest-path depth, rows centered, authoring
tie-break), bezier edges, the retired panel's status color language.
- PlanBlock renders data-plan-block parts AS CARRIED (frozen snapshots,
never re-derived): kind full = borderless, viewBox = layout extent so
nodes hug the column edges, natural height; kind neighborhood = own
small labeled 'Plan fragment' container with the current node glowing
(pulsing primary halo).
- PlanPopup: full-screen overlay with the LIVE plan — latest tree from the
transcript, statuses derived at click time via derivePlanStatuses
(contracts/plan.ts). Esc / backdrop / X close; drag pan, wheel
zoom-to-cursor, zoom buttons; fit-floor semantics harvested from the
retired tree panel (zoom floor = the fit, pan locks at it).
- chat.tsx: data-plan-block case in the part switch + popup mount.
- src/core/ui-message.ts: TeacherUIDataParts gains "plan-block": PlanBlockData
(additive, type-only — lane touch flagged to master).
- tests/render-plan-block.test.tsx: pure layout (layering, centering,
dangling-edge tolerance), wrap, frozen-snapshot render, live derivation.
Extraction is real spend and background work; the dock needs to see
and stop it. User decisions: persist all cost data (chat record),
per-stage usage granularity, and TRUE cancellation for abort.
USAGE CAPTURE (live + persisted):
- extraction.ts: per-call onStepFinish captures each billed step's
usage (SDK LanguageModelUsage -> CallUsage: input/output/cache
read+write/reasoning) into the registry's live per-stage usage
(extractor/auditor; the merger is deterministic — no model call).
- /api/graph/status runs now carry usage (model + steps per stage);
cost/context folds are the consumer's (core/prices.ts — nothing
stores dollars).
- ChatRecord.extractions: append-only per-run usage log (mirrors the
turns log): startedAt key, at, durationMs, merge stats when the
merger ran, stages. Persisted on success AND failure — a failed run
still spent money. Idempotent per startedAt; the list payload
(ChatSummary) never leaks it.
- EmissionStats moved to contracts/graph.ts (its contract home);
core/graph.ts re-exports for existing importers.
- Verified end-to-end against the real pipeline (backfill one chat):
per-step token counts, reasoning tokens, and the auditor's growing
prompt-cache reads captured and persisted.
TRUE ABORT (the dock's abort button, user decision):
- RunningExtraction owns one AbortController per run, shared across
phases; abortExtraction(chatId) fires it and returns whether a run
was aborted (idempotent no-op otherwise). The pending model call
receives the signal — the provider cancels the request, spend
stops. throwIfAborted() at phase boundaries ends runs aborted in
sync gaps.
- Abort is a deliberate stop, not a failure: same ledger discipline
(nothing merged, nothing checkpointed, partial usage persisted) but
the run reports zero emissions instead of throwing. Exported via
src/server/observability.ts for master's /api/activity/abort route.
- Tests: mid-run abort (hanging mock that rejects on the signal),
live-accumulation + persisted-record assertions, store-seam
idempotency, and the summary-payload leak guard.
Wave 4 (agent-side observability + control), reconciled with the dock
contract (master relay 2026-09):
- abortJob in src/server/jobs.ts, exported through the server barrel for
POST /api/activity/abort (master's thin route): fires the job's
AbortController (real kill — the in-flight model call aborts, not just
the next loop check), drops any queued result steer, flips the record
to error "Aborted by learner" + activity "aborted" synchronously, and
queues a one-line abort steer (buildAbortSteer) through the same wake
pipeline so the teacher's promise doesn't hang in the lesson. Shares
the result steer's deterministic id space, so abort + racing result
can never both deliver (steerDelivered dedupe).
- Abort registry (src/agent/subagents/aborts.ts) on globalThis, mirroring
the wake pipeline's module-state pattern; runner registers/unregisters
per run, so the route's module instance sees the loop's controller.
- Runner usage capture: LoopContext (abortSignal + onUsage) threads into
all three kind loops; per-step usage folds via accumulateJobUsage into
JobRecord.usage (model from config MODELS, absent-never-zero), flushed
in runJob's finally so an aborted job's money record survives the kill.
- Contracts: JobUsage token fields optional + lastStep (most recent
call's accounting, for the dock's per-blob context line vs
CONTEXT_LIMITS); SubagentResultMeta.aborted; AbortSteerMessage;
buildAbortSteer; SteerMessage union generalized.
- demo-builder/visual-maker accept abortSignal + onUsage (visual folds
generateText usage per iteration).
- stillRunning/finishJob/jobContext guards: an aborted job is never
overwritten by the runner (terminal no matter what the wake does);
failJob and the pump's finalizeJob keep the flip.
- UI: abort steer renders a quiet aborted chip (job-aborted-badge).
- Prompt 20-jobs.md: one line about aborted jobs.
Tests: accumulateJobUsage folds (undefined-preservation, lastStep),
runner usage capture, real-kill abort (sync flip, silent unwind,
abort steer wakes), usage-so-far survives the kill, aborting a
delivering job drops the queued result steer, outcome reasons,
abort-chip render. 330/330, typecheck/lint/build green.
One quiet row beneath the composer/question shell (not flush to the
screen bottom; static height, outside the morphing shell):
- LEFT: context window + session cost, clickable → per-turn detail
popover (context, total, unpriced count, one row per turn with
model/steps/cost). All folds from core/prices.ts — sessionCost,
contextWindow, turnCost — no cost math in components.
- RIGHT: model + "Thinking —" — a placeholder slot for the coming
model/thinking-level CONTROL (later work with its contract changes);
display-only today, model read from the last turn's usage,
pretty-printed (Glm 5.3 Flash).
- Data: the record's turns log, seeded on mount, refreshed when the
local stream settles, kept fresh by the existing record re-syncs.
No turns yet → "—" everywhere; unpriceable turns never fold into a
fake $0 (the doctrine).
- formatUsd/formatTokens moved to components/usage-format.ts (shared
with the activity dock); prettyModel joins them, pure + tested.
One reserved corner of the shell, one blob per running BACKGROUND
process — session sub-agents (research/demo/visual) and global
extraction runs. Teaching turns are deliberately NOT here: the dock
exists to surface what the learner can't otherwise see; a running
teaching turn is already fully visible where it happens (user decision).
- Data: ONE poll of GET /api/activity (jobs across ALL chats,
title-joined, oldest-first + extraction runs), folded client-side
with core/prices.ts (jobCost). Adaptive interval — 2.5s while
anything runs, 10s idle, 4× slower on hidden tabs; a failed fetch
keeps the last feed (no flicker to empty on a blip).
- Per blob: kind icon, elapsed (ticking), cost, activity line, and an
explicit END on the blob's face. Click → detail popover (context
line, model, steps, open-session link). The context line is the LAST
model call's tokens vs the model's limit (JobUsage.lastStep — pending
agents-stream contract; "—" until it populates), NEVER a cumulative
sum. Absent data renders "—", never zero (the doctrine).
- Header: total running cost (sum of priced blobs, "—" when nothing
is priced). Narrow viewports collapse to a count badge → drawer.
- END posts to /api/activity/abort {kind, id} — the backend is being
composed by master over the streams' mechanisms; until it exists a
click locks the button and the popover says "Abort not available
yet" (never a dead control).
- Idle renders NOTHING. Mounted once in the root layout; every view
gets it for free.
toBlobs / formatElapsed / formatUsd / formatTokens are pure + tested;
SSR render tests pin blobs, the "—" doctrine, the face END, and the
priced total.
The dock is visible in every view, so it needs every running process —
sub-agent jobs across ALL chats (not just the open one) plus the global
extraction runs — in one poll. /api/jobs stays chat-scoped and untouched:
it is the transcript machinery's poll and carries turnInFlight for the
re-attach seam (chat-transport), the seam we spent three live rounds
stabilizing; making it polymorphic would fork its response shape by mode
and put regression risk on exactly the wrong path. The dock is a different
consumer with a different shape (cross-chat, running-only, cost-bearing),
served by a thin route over server/observability.ts:
- activityFeed(): running jobs from every chat (title-joined, oldest-first
on last activity) + extraction runs + busyChats (chats with a teacher
turn in flight — the biggest spender; data made available, rendering is
a dock decision).
- GET /api/activity — the dock's single poll everywhere.
- Cost math stays client-side (core/prices.ts folds); raw usage rides the
records and the transport stays thin.
Decided with the user over the polymorphic-/api/jobs and
separate-all-jobs-endpoint alternatives.
The 2026-09 observability wave's type-level foundation, landed before the
streams so nothing renders against guessed shapes:
- contracts/chat.ts: CallUsage + TurnUsage (per-model-call token accounting,
one entry per teacher/wake turn keyed by messageId) and ChatRecord.turns —
a DELIBERATELY parallel append-only usage log, not per-message metadata:
the client re-POSTs its message list every turn and its copies never
carry server-stamped fields, so anything inside messages gets clobbered
by the next persist (the wave-3 merge-rule lesson).
- contracts/chat.ts: PlanBlockData — the persisted data-plan-block part
(full DAG on plan_init, one-level neighborhood on plan updates, snapshot
semantics so history renders what was true then).
- contracts/jobs.ts: JobUsage + JobRecord.usage — a sub-agent run's
accumulated accounting, written by the runner.
- core/ports.ts + store/chats.ts: ChatStore.appendTurnUsage — idempotent
per messageId (abort replays / wake retries append once); listChats no
longer leaks turns into the list payload.
- core/prices.ts: PRICES + CONTEXT_LIMITS (generated from OpenRouter's
published pricing via scripts/sync-pricing.ts) + pure folds: callCost,
turnCost, sessionCost, contextWindow, jobCost. Unknown model or absent
fields render as absent, never zero.
- server/observability.ts: barrel-approved surface for the transport layer
(fixes the import-law violation in the graph status route that slipped
through the wave-3 merge verification).
Capture wiring (agent onStepFinish -> persist seam, plan-block emission)
lands next as master seam work; the streams start against these shapes.
Wave 3: the plan protocol collapses to one re-callable mermaid plan_init
(seven mutation tools deleted by design; statuses derive from quiz checks;
live quiz-nodeId validation with mid-turn retry). The steer-starvation arc:
steers deliver coalesced with the learner's answer (claimSteers), persist
dedupes same-id continuation copies, jobs settle-sync defers while a turn
is busy. Research steers gain the collapsed full-brief affordance. Model
reasoning effort set to minimal across roles.
With continuation coalescing, a job flips done at CLAIM time — while the
learner's own answer turn is still streaming. The poller's settle re-sync
fired on the flip, swapping in a mid-turn record snapshot mid-stream; the
steer after the assistant message broke the SDK's in-place continuation
merge, and the streaming response landed as a SECOND message with the same
id — duplicate React keys in the live transcript (the user's live test hit
it; the record itself stayed clean).
A settle during a busy turn (local stream OR server turnInFlight) now marks
a pending sync; the first idle poll tick adopts the finished record. This
also fixes the latent multi-job race: steer B settling while steer A's wake
turn is in flight would have grabbed the same mid-turn snapshot.
Out-of-lane surgical touch (flagged for master review): the wave-1 poller
block in components/chat.tsx.
With a steer coalesced into a quiz-answer continuation, the steer sits AFTER
the assistant message being continued, so the SDK's onFinish list carries
the input copy AND the extended response copy as two entries with the SAME
id — persistTurn saved both, the jobs poller's syncFromServer imported the
dup, and React keyed a single message twice (the user's live test hit it).
withMessageIds now dedupes by id keeping the LAST occurrence — the SDK's
final merge state (a superset of the earlier copy) — which also yields the
right semantic shape: [user, steer, assistant(extended)].
Out-of-lane surgical touch (flagged for master review): turn-service.ts.
Squashed the live session's already-corrupted record by hand (same dedupe).
The bug (seen live): a teacher that ends every turn with a new pending quiz
keeps the wake pipeline's idle gates permanently closed — the steer sits in
'delivering result' forever while the learner answers check after check.
- claimSteers (server/jobs.ts): the chat-handler claims every queued,
undelivered steer the instant a client turn wins the turn registry, and
appends it after the learner's answer — conversion-safe, the answered
quiz carries its output. Every answer flushes pending steers immediately.
- The pump keeps its post-idle duplicate-guard (re-check steerDelivered
before firing), so a claim racing the pump converges to exactly one
delivery either way.
- The stale-job sweep now skips chats with a turn in flight (a healthy but
slow wake no longer gets falsely re-queued as 'recovering').
- Out-of-lane surgical touch (flagged for master review): the 3-line
handleChatRequest seam (claim after beginTurn, persist+stream 'outgoing').
- Verified live end-to-end on the stuck session: answering the pending quiz
fired the steer + wake turn within a second; job flipped done/delivered.
- plan_init is now the ONLY plan tool: a mermaid 'graph TD' block parsed by a
constrained-subset parser (mermaid-plan.ts) with line-precise errors; the
seven mutation/tracking tools are gone by design (2026-09-03 design pass).
- Re-call plan_init = sanctioned wholesale rewrite; slugs are permanent.
- derivePlanStatuses (contracts/plan.ts): done = passed check tagged with a
quiz nodeId; current = open check's node else first topo-ready node — the
tracking the teacher no longer declares by hand.
- questionToolsFor(ctx) (tools.ts): the quiz schema chains a dynamic plan-
nodeId validation against the LIVE record — an invalid nodeId is rejected
at the call boundary and the teacher retries mid-turn, before any UI.
Static questionTools stay validator-less for the default path.
- Prompts rewritten for mermaid authoring + derived position.
- Out-of-lane surgical touches (flagged for master review):
src/core/question-core.ts (quizInputObject export + nodeId description),
src/core/contracts/plan.ts (derivePlanStatuses + doc).
Key-convention policy, settled: every field accepts ⌘/Ctrl+Enter
(metaKey || ctrlKey — Mac gets ⌘↵, Windows/Linux Ctrl↵, the Gmail
convention); plain Enter stays matched to the field's species —
sends in message fields (chat composer, free-text answers — the
universal chat convention, Shift+Enter still a newline), newline in
the thinking field (composition, for dumping).
amosblomqvist