Hypothesis under test by the owner: coherent extractions turn to mush
after merging because the auditor minted too many bridging edges of its
own — inference, not identity, and out of the auditor's scope. The graph
now connects across sources solely through identified equivalences;
relations between propositions come from the extractor alone.
- auditSchema drops 'edges' (submit_review = links only); the submit-seam
required-field doctrine reworded per tool
- AUDITOR_INSTRUCTIONS: the BRIDGING EDGES section dies; deliverable is
the identity verdicts only
- AuditVerdicts drops edges; ExtractedEdge endpoints are extractor-local
ids only; mergeFindings stops consuming verdict edges and its resolve()
drops the raw-node-id branch (dead — the extractor never sees graph ids)
- extractor untouched: submit_findings keeps propositions + follows/
supports edges, required as before
- tests: 17 submit fixtures + 9 typed verdict literals slimmed to links
- docs: architecture.md division-of-trust + pipeline diagram updated
Every notes chunk and chat coverage batch is its own job — registry row,
usage record, failure domain. Extracts fan out process-wide behind a
3-permit extract semaphore on globalThis (new 'queued' phase); audits and
merges stay serialized behind the existing FIFO gate ('waiting'), so each
auditor explores the graph after every earlier job's merge.
- ExtractionPhase gains 'queued'; RunningExtraction gains part {i,n};
the feed shows one row per job with a '· i/N' suffix (status route)
- abortExtraction is source-wide: every job of a fan-out dies together;
merged jobs keep their coverage checkpoints
- failure isolation at the fan-out (Promise.allSettled + aggregation):
a failed chunk no longer kills its siblings — the run returns what
actually merged; job-level aborts log, pre-fan-out errors still throw
- usage ledgers: one record per job (source-keyed), failed jobs bill
their partial spend
- runJob/aggregateJobs helpers; setPhase/recordStep/extract/audit/
notesExtract key the registry by jobKey (src#i) — notesExtract's
signal lookup was the one real bug the new tests caught (source-id
lookup meant fan-out aborts never reached the model call)
- semaphore is handoff-correct (fresh takes increment; handed-off
permits don't double-count — an inflated counter deadlocks at 3)
- tests: isolation pins rewritten (stage failure now resolves zeros),
single-flight/registry tails updated, per-job usage counts, and a new
describe pinning the fan-out: 3 extracting + 5 queued, parts 1..8,
source-wide abort, per-job usage on failure, extract peak == 3,
audit peak == 1
- docs/architecture.md: serialized-audit paragraph rewritten for
per-chunk jobs; two-front-doors wording aligned
Owner ruling 2026-09-28: the auditor has no control over the
extractor's output and never alters it, so re-emitting the full
findings per submit was pure payload cost — the retry collapse the
schema slim attacked, halved again. submit_review is now links
(local proposition id -> existing graph node; EQUIVALENCE BIAS
carried in the linkTo description) + bridging edges ONLY; the
merger joins extractor findings + verdicts.
mergeFindings(chatId, findings, verdicts, ...) — propositions come
from findings verbatim, links join by id, edges emit extractor's
then auditor's through one dedup-preserving emitter. The old echo
form hid an accidental drop channel: an auditor that omitted a
proposition from its echo silently dropped it from the merge.
Verdicts-only removes it structurally — extractor output always
fully merges (tests updated to full-merge expectations).
AuditResult/AuditedProposition -> IdentityLink/AuditVerdicts;
audit() logs links/bridges; batch view renders anchorless like
transcriptView; auditor prompt rewritten to the verdicts-only
deliverable; architecture.md division-of-trust updated.
OpenRouter log diagnosis (owner, 2026-09-28): extraction time was
dominated by tool-validation retries. The failing call had the whole
propositions array emitted as ONE JSON-STRING ('propositions.0: must
be object'), a truncated placeholder artifact inside it, and
word-salad statements late in the array — payload-size collapse.
Field audit: 'evidence' was never persisted anywhere and its
faithfulness-checking consumer died with the fully-trusted ruling;
'messages' fed messageId granularity that NOTHING downstream reads,
and the model sent [] anyway (chat-level provenance in practice).
'aliases' stays (optional now): the auditor's find tool globs them.
Schema: propositions = {id, name, statement, aliases?}; edges =
{from, to, type}. Prompts lose the evidence/messages doctrine and
[id]-anchor guidance, gain a compact faithfulness line;
transcriptView renders anchorless. The auditor echo instruction
drops the evidence/messages mention, so its payload halves too.
Heal: healStringifiedArrays on the submit tool's prepareArguments
(pi-agent-core's sanctioned pre-TypeBox shim) parses a stringified
propositions/edges back into arrays — the LaTeX-restore doctrine:
heal weak-model wire behavior at the seam. Test drives the real
agent-loop seam with a stringified submit.
Extraction may run in parallel across sources (extractors are
source-local and graph-blind), but the auditor explores the LIVE
graph through its tools and stamps identity verdicts against what
it finds — two concurrent audits would each link into a graph the
other is about to change, the structural mint of duplicate nodes.
withAuditLock: a FIFO promise-chain mutex on globalThis (the same
module-instance idiom as the run registries — a module-level lock
would be per-route-instance and useless in dev). The check runs
before any LLM is invoked or the graph read; the lock spans audit
through merge inclusive (the merge is the graph mutation) and is
reacquired per batch, so interleaved multi-batch runs each audit
against the latest graph instead of starving each other. Applied at
both front doors — chat batches and notes imports.
New 'waiting' phase (extracting → waiting → auditing → merging):
visible in the status feed so a queued run never looks stalled in
extracting. Abort-aware on both sides of the queue: a run aborted
while waiting leaves without auditing (its chain link still
releases — successors never hang); a run aborted mid-audit releases
in finally.
Tests pin: a second run queues at the gate and its auditor is
invoked only after the first run's merge is in the ledgers (the
post-merge-graph invariant itself); a run aborted while waiting
never audits, checkpoints nothing, and leaves a healthy chain for a
third run. Gates: typecheck, 1061/1061, lint, build.
Owner's first live test of the relinked pipeline surfaced three issues,
diagnosed against the ledgers:
1. Detail-card edge labels were backwards (UI only — data and canvas
arrows were correct). Semantics pinned 2026-09-28: an edge A→B
means A implies B (follows) or A supports B (supports). Outgoing
edges now list under 'What follows from this' / 'What this
supports', incoming under 'Follows from' / 'Supported by'.
Verified against the ledger: Invariance --supports--> Twin
paradox / Relativity of simultaneity; Argumentation ethics
--follows--> Homesteading — arrows correct, only the words were
swapped. docs/architecture.md wording fixed to match.
2. Session-local terminology leaked into node names ('Criterion C',
'Action axiom (F1)'). Both extractor prompts gain a naming
doctrine: names use established, discipline-canonical terminology
(glossary-headword test); never source-local bookkeeping (letters,
numbering, positional labels); local shorthand goes to aliases.
3. Near-duplicate propositions and lonely nodes (~40% of fresh-run
nodes edge-less; 'Argumentation ethics' vs 'A priori of
argumentation' duplicated despite the auditor having linked that
phrase once). Root cause is upstream of the auditor: the extractor
had no self-merge or edge-completeness duty. Both prompts gain:
MERGE YOUR OWN OUTPUT (same claim at different precision = one
proposition, phrasings to aliases; nothing downstream dedupes) and
edge COMPLETENESS (a proposition that arose as evidence for,
consequence of, or prerequisite of another must carry the edge —
a graph of lonely nodes is a failed extraction).
Owner rulings on the open question: keep two edge types (no
'relates to' junk drawer); re-measure loneliness after these fixes.
Gates: typecheck, 1059/1059, lint, build.
The graph becomes raw reasoning chains only. Quiz outcomes and
understanding observations are no longer formalized; assessment
moves to the future practice module.
Events ledger removed end-to-end: LedgerEvent/LedgerEventKind/
UnderstandingLevel/ExtractionAttribution/AuditResult.dropped and
snapshot nodeObservations deleted from contracts; quizEvent builder
and the events step of mergeFindings gone; appendEvent/readEvents
dropped from the port and store; LEDGER_FILES loses events.jsonl
with a one-time purge of a stale file next to the nodes ledger;
graph-deletion and the delete dialog lose event purge + quiz-answer
impact counts; GraphSessionProposition becomes a pure birth list.
Whole-session markdown source view: new core/session-markdown.ts is
the shared formatting engine (id-anchored speaker blocks, quiz/
recall/question blocks, tool one-liners) — the extractor's prompt
view and the future user-facing export render from one engine.
Coverage fingerprints deliberately stay on the old line format
(content-hash semantics; re-formatting never re-opens covered
history). The whole session goes to the extractor — no phase cut.
Auditor refocused as the identity linker: never sees the source;
explores the graph via new auditor-specific tools in
agent/tools/graph.ts (find — glob over name/statement/aliases;
list_sessions; list_propositions with full statement text; no
timestamps) instead of an injected full-graph dump; equivalence
bias — paraphrases merge, only genuinely distinct propositions stay
apart; no drop channel, the extractor is fully trusted; may draw
bridging edges (typed follows/supports) between existing nodes and
the batch, resolved by the merger with chat-scoped message-less
provenance. Agentic loop, maxTurns 16.
UI: the node detail card gains an edges section above the sessions
list (follows-from / supported-by / what-follows / what-it-supports,
rendered only when present, click-to-select neighbors); the
understanding-note chips are gone.
Prober list_propositions reworded to what a session covered —
names only, no levels. docs/architecture.md + README updated to
the as-built pipeline. Gates: typecheck, 1059/1059, lint, build.
Two follow-up rulings on the explain tool:
1. No 'I don't know' row — not needed in free recall. If he doesn't
know, he says so in his own words; his text is the only answer
channel. Dock: text mode is rowless again (the composer IS the
surface); doctrine and the shared prompt chunk updated; the format
branch and card revert to quiz-only dontKnow.
2. explanation + correctAnswer collapse into ONE field — the separate
explanation was quiz remnant (MC needs the why outside its options;
a free-text commitment carries its own reasoning). Schema is now
{question, correctAnswer, details?}; the reveal shows the single
committed answer; no explanation snapshots anywhere on the explain
path.
1047/1047 tests, typecheck/lint/build green.
Sycophancy mitigation as a structural commitment: the model writes its
full correctAnswer (plus explanation, in that order) BEFORE the learner
sees the question; the commitment rides the persisted call args, is
revealed on submit, and grades his free text in the model's next reply —
the client never grades free recall.
- tools.explain setting (segmented 'Quizzing — Multiple choice | Free
recall' in Agent settings' Tools group): free arms explain and drops
quiz for BOTH agents (prober variant carries no nodeId, same pattern
as the probe quiz); mechanics keys never pass filterEnabledTools —
quiz stays unwritable (the 2026-09-19 wedge ruling stands)
- doctrine lives pi-natively in the tool descriptions (owner ruling —
no conditional system-prompt chunks): explain carries the commitment
rule; the MC option-building craft moved out of the prompt into the
quiz description; shared/20-question-tools.md slims to the
quiz-agnostic core
- display seams: parseExplainInput (LaTeX-debris restoration included),
dock text-mode with an 'I don't know' row (kbd row 0), AnsweredCard
reveal (The answer / Why, no verdict chips), [Explain answer] format
- phases/plan seams recognize both quizzing tools' nodeId tags; explain
answers deliberately feed no ledger evidence (no client grading to
key on — a future ruling can parse the model's grade)
1049/1049 tests, typecheck/lint/build green.
The owner's capture-vs-composition check confirmed the wire sends exactly
what composes (byte-identical save the nodeId paragraph ebfe058 had since
changed). This wave removes the implementation meta that leaked into the
prompts and reorders so the method carries weight:
- base: owner's new opener verbatim (install understanding; the
understanding definition — which absorbs the old 'Understanding is
structure' paragraph); Voice moved in from shared/10 (deleted); the
philosophy keeps edge-dependence + logistics ('standard learning'
dropped per owner); intro's two-agents/fragments sentence gone.
- teacher/05-method opens with the owner's methodology bridge ('No other
teaching methods come close') — paragraph 3 of the opener, placed here
so the shared base never points the prober at methodology it doesn't
carry.
- deleted: teacher/60-probing-in-brief (self-evident from shared history;
its mid-lesson mini-probe move salvaged into 20-teaching as one
sentence), the 'Session shape — the durable artifact' section
(extraction-pipeline meta), the fragments-meta lines in shared/20 and
05-method, 'the transcript's seam' phrasing, prober's 'the teacher
that takes over' / 'the handoff is automatic'.
- composition reorder (listSourceSegments): base -> the agent's own
fragments -> shared/ appendix. Teacher method now starts at 12% in
(was 62%); prober doctrine at 34%. shared/20 opens with the owner's
decision rule (quiz only for known-correct-answer, ask for everything
else).
- HANDOFF_TEXT: the owner's user-message version — first-person, paces
the planning (unconditional truths, motivated derivations), ends at
plan_init + accept-or-changes.
Teacher 31.4k -> 28.7k chars; prober 10.0k -> 9.6k. Gates: typecheck 0,
1018/1018, lint clean, build green. (Suite runs 16s; the earlier 10-min
timeout was stale node_modules after master's df1328e 0.87.1 sync —
npm install fixed it.)
The probe intro was redundant with 60-probing-in-brief.md (and its
no-probe fallback described a state the code can't produce — the teacher
only exists post-probe). 'How the map tracks itself' folded into one
tagging paragraph; grain and rewriting compressed; the live-map context
kept to one sentence. The discovery-chain thinking block stays verbatim,
and the stress-test stays. Presenting: ask him to confirm the plan
before teaching begins — medium unprescribed (owner ruling 2026-09-27;
dock-question confirmation leaves the teach divider floated next to the
planning divider when no text reply splits the group — accepted as
cosmetic).
The inferred teach seam sat as a ternary branch in the messages.map
chain (alongside the no-bubble marker branches), so the ENTIRE first
teaching group — thinking, prose, the tagged quiz call — rendered AS
the divider and vanished below it (the Lacan session, 2026-09-27).
The marker branches replace contentless scaffolding; the teach group is
a real message. The divider now prepends inside the group's Fragment.
Also tightens 10-planning.md: the presenting turn asks NOTHING — no
quiz, no ask_user_question, no 'shall we begin?' (the teacher was
substituting a confirmation question for the dead verdict buttons).
The Accept/Decline verdict gate is gone (owner ruling 2026-09-26):
presenting the plan ends the teacher's turn; the learner's next message
is a natural reply. Teaching begins at the first quiz call tagged with a
nodeId (probe quizzes carry none) — both folds (view phaseAt3, native
phaseAt3Lh) advance on the call itself, and the teach divider renders
before that whole assistant block. No backwards compat: verdict markers,
the verdict input event, optimistic appends, composer buttons, and
projection all deleted; prompts and docs rewritten to the natural flow.
Three owner rulings from the gradient-descent session
(chat_muhskq6p_e8c3600b — 7 failed quiz calls, answers leaked in
option descriptions, planning divider floating at the transcript top):
1. RETRYING LINES (general): any tool call that errors back to the
agent (validation failure, malformed args — the retry loop) renders
one quiet "Retrying…" SystemLine instead of silence. Question tools
previously rendered null, so a flailing quiz loop read as a stall.
Aborts are terminal ("Operation aborted") and stay silent. Verified
live: exactly 7 lines render at the 7 failure sites of the session.
2. QUIZ OPTION DESCRIPTIONS: NUKED. The model used them to telegraph
correctness in every quiz of the session ("that moves uphill — it
climbs f instead of descending it") — answerable by elimination.
The field is gone from BOTH quiz schemas (probe + teach); ask keeps
it (ungraded — nothing to leak). The dock additionally suppresses
descriptions for quiz kind, so pre-fix records stop leaking too.
Verified live: the session's pending quiz renders bare labels.
Plus the diagnosed failure-mode fixes: GLM sometimes emits `options`
as a JSON-encoded string (6 of 7 failures). prepareArguments hooks
on all three question tools coerce it BEFORE pi's validation (the
call lands instead of bouncing); the client coerces again in
normalizeOptions (pi persists raw call args). Schema descriptions
hardened ("REQUIRED ... never a JSON-encoded string").
3. DIVIDERS AT THE MARKERS: the probe-handoff SystemLine
("Preparing your plan…") IS NOW the Planning divider; the accepted
verdict IS NOW the Teaching divider ("✓ Plan accepted" line gone) —
one signal per seam, not two. The fold-based divider block is
DELETED: Wave D grouping merged the whole probe into one giant
assistant group, so the inclusive-fold comparison floated the
divider to the transcript top (reproduced against the real record:
6 view messages, phase flipped at the giant group). The markers
render in the right spot by record construction. A DECLINED verdict
is not a seam — it keeps the "↻ Plan revision requested" line.
Verified live: plan divider after the probe concluded, teach
divider after the plan presentation, nothing at the top.
Tests: 1004/1004 — string-coercion pins (normalizeOptions + the
prepareArguments hooks on all three tools, identity fast path),
quiz-schema-has-no-description pin (ask keeps it).
Owner correction (2026-09-26, later): a real lesson is ONE continuous
teach->quiz->teach run - there are no run-would-stop moments mid-lesson,
and the run ends that DO exist are quiz stopper exits where delivery
must not fire. Subagent results must land BETWEEN messages INSIDE the
run, never splitting a toolCall from its toolResult.
Vehicle 1 now hands markers to agent.steer() - pi's STEERING queue:
the loop drains it at TURN BOUNDARIES (after a completed assistant
message AND its tool results - the bonded pair survives injection) and
again at RUN START (a continuation drains before its first LLM call),
then CONTINUES the run with the marker in context.
steeringMode "all" replaces followUpMode.
registerLiveAgent moved BEFORE runPiTurn and now hands pending markers
itself: the start poll drains the queue SYNCHRONOUSLY inside the
runPiTurn call, so a later registration would miss it. A marker
stranded by a quiz-terminated run therefore rides the ANSWER run's
first LLM call - the native replacement for the old continuation
coalescing.
Settlement reconciliation, the idle wake, probe-handoff exclusion,
and the registry-as-truth design are UNCHANGED; stranding is now
exactly the stopper-exit case.
Tests: jobs 45/45 (fakeAgent.steer; new pin - registration steers
pending markers into the new run, parked wake never fires); chat-flow
24/24 (new e2e through the real composition - a gated mid-run turn, a
marker delivered while the turn streams, record shape
[call, result, MARKER, response] with the marker adjacent to the
completed unit, the model SEES the marker in its next prompt, zero
wakes); full suite 999/999, tsc/lint/build green.
Live smoke (PORT=3004): research dispatched in the probe turn parked
behind the dangling quiz; on the quiz answer the marker injected at
the answer run's START poll - record [answer result, marker, teacher
response], job flipped done/delivered at settlement, brief card
rendered with its sources toggle.
Plan doc: docs/PI-NATIVE-MIGRATION-PLAN.md gains the pivot section.
Deferred (owner list): thinking blocks rendering after messages;
vanishing tool-call output tokens.
Delivery now rides pi's OWN follow-up queue when a run is live, and a
carry-everything idle wake otherwise. One registry, two vehicles;
claimSteers and continuation coalescing are deleted (owner rulings,
2026-09-26: follow-ups only; one queue; the wake is the only appender).
THE LIVE RUN (vehicle 1): the chat-handler registers the streaming
turn's agent (registerLiveAgent, next to registerTurn); the pump hands
every eligible queued marker to agent.followUp() - pi polls that queue
when the run would otherwise stop, injects the markers
(message_start/message_end on the live wire), and CONTINUES the same
run. followUpMode "all" joins every handed marker in one continuation
turn. The question stopper runs before the follow-up poll, so an open
quiz is never overridden - a handed marker strands at settlement and
waits (today's pacing). probe-handoffs never take this path (the
handoff must START the teacher's run - a record-level branch).
SETTLEMENT RECONCILIATION: settleSteers(chatId, persisted) at onFinish
(after unregisterLiveAgent) resolves+flips entries the run delivered
and returns stranded ones to pending - the registry stays the truth;
the dead agent's queued copy is abandoned. This is what makes the
handoff safe: pi owns no retry for a stranded follow-up.
THE IDLE WAKE (vehicle 2): waitForIdle (unchanged gates) then ONE turn
carrying EVERY pending marker as its input appends - dequeue-at-fire
(the Wave B regression lesson: the wake run's own follow-up poll must
never see its own cargo; a 409 pushes the cargo back and retries).
CLIENT: applyWireEvent now upserts non-assistant messages by id, so
in-run injected markers render live instead of dropping until settle.
Tests: jobs 44/44 (live handoff, settlement stranding, handoff-waits,
one-wake-carries-all); chat-flow 23/23 (the double-append pin drives
the real pump through the new input); session-core 28/28 (custom-role
wire appends). Full suite 997/997, tsc/lint/build green.
Live smoke (PORT=3004): research job delivered mid-probe - parked
behind the dangling questions exactly per semantics, landed once on
the open gates, teacher folded the brief in and re-dispatched; the
second (good) result rendered the full 10-source brief card and the
lesson continued from it. No duplicate ids anywhere in the record.
Known gap flagged to owner (NOT from this change): the server accepts
malformed answer events via raw API (answers without the answers[]
shape) and AnsweredCard then crashes the page render - surfaced by the
smoke's scripted answers, repaired by hand in the scratch record.
Server: ui-stream.ts is verbatim pi AgentEvent passthrough (SSE data
lines + [DONE]); only app-level events are the loop-level error and
turn-settled terminal; onError carries raw provider text (owner ruling).
Client: useChat + the AI-SDK transport replaced by components/session-core.ts
(SessionModel: 409 identity-retry, optimistic c_ appends, absorbed()
adoption, needsContinuation, shared consume for POST + reattach) and the
session.ts hook; the golden-master grouping canon moved client-side into
core/view.ts (projectView); scanners read the native record (phaseAt3Lh,
danglingQuestionOf, graph toolParts, plan-blocks, runner, extraction).
Coverage: the native message is the unit (fork 3) with the call+result
inseparability bond rule; view-line-less messages (markers, error
results) are invisible to coverage and never checkpointed.
Deletes: chat-transport, projection, ui-message, their tests (ported to
tests/session-core.test.ts), and the ai + @ai-sdk/react dependencies —
the Vercel SDK is fully out of the tree. Wire pins in chat-flow rewritten
onto the AgentEvent dialect (eventsOf parser).
Gates: tsc clean, 994/994 tests, eslint clean, build green; live smoke
on PORT=3004 (verbatim wire → native record).
The chat record is now pi-native agent history (LhMessage[]), the AI SDK's
UIMessage transcript contract is gone from the server, and every client turn
arrives as ONE explicit input event.
Core (new):
- src/core/native-messages.ts — the LhMessage taxonomy (user/assistant/
toolResult + lhMarker customs: verdict, probe-handoff, subagent-steer,
subagent-abort, plan-block), constructors, and toLlmMessages — the single
model-facing seam (convertToLlm): strips ids, flattens markers to user
text, drops UI-only plan-blocks, synthesizes INTERRUPTED results for
dangling aborted calls, never throws.
- src/core/projection.ts — projectNative: the one-way native → UIMessage
projection (consecutive assistant+toolResult runs group into one
assistant view message; toolResults attach backward by toolCallId;
markers project to data parts; plan-blocks splice after the last
successful plan_init). Every existing consumer reads through it.
- src/core/input-events.ts — the ChatInput union (text/answer/verdict/
continue), parseChatInput validation, and inputAppends (answer targets
the dangling question call or 400s; verdict = optional note user
message + native verdict marker).
Server:
- chat-handler.ts — the turn root takes {chatId, input}; write-ahead
persists input appends + claimed steers before the turn; claimSteers
coalesces the whole queue with the handler's steerDelivered filter as
the dedupe seam; teacher-vs-prober branch on hasProbeDone.
- turn-service.ts — persistTurn stamps ids, drops clean question
placeholder toolResults (the record ends at the unanswered call),
emits plan-block markers idempotently (plan-blocks.ts rewrite), and
persists abort partials.
- ui-stream.ts — PiTurnSpec takes the native base; the wire dialect is
unchanged. jobs.ts steers are native markers (steer_<jobId> dedup ids);
waitForIdle reads danglingQuestionOf. teacher-agent.ts wires
convertToLlm: toLlmMessages. record-adapter.ts DELETED wholesale.
- GET /api/chats/[id] serves messages: projectNative(record.messages).
Client:
- chat-transport.ts — the 409 retry is IDENTITY (re-POST the same event
body after the turn settles — the event was never persisted
pre-acceptance; the merge logic is gone). The transport consumes the
stashed input event via prepareSendMessagesRequest and throws loudly on
a missing stash.
- chat.tsx — every send call site stashes exactly one ChatInput event
(text, answer, verdict, continue) before invoking the SDK send.
Tests: full native migration (chat-flow rewritten to input events and
native record pins; jobs/extraction/demos/graph/usage/plan-blocks/
probe-done/reattach-transport converted; record-adapter.test.ts deleted;
chat-transport tests pin the identity retry). 982/982, typecheck, lint,
build all green.
- src/agent/lh-tool.ts deleted: the zod-wrapped LhTool + toPiTool bridge
(z.toJSONSchema per call, clientQuestion flag, hand-rolled result mapping)
is gone. Tools are pi-agent-core native AgentTools now.
- New src/agent/tool-execute.ts owns what the wrapper added: the abort guard
(runWithAbortGuard: run signal + 30s per-call timeout + 250ms grace,
'Operation aborted' / timeout wording preserved), ToolInputError (the
retryable '- problem at path' format), result mapping (string verbatim /
object JSON / image parts, details rides the record's tool part output),
and questionResult() — placeholder + terminate:true (owner Option A:
belt of terminate + slim makeQuestionStopper with the clean-question
rule; maxTurns deleted from the teaching turn entirely per owner).
- All 11 tool files + tools.ts registries converted to TypeBox parameters
(typebox@1.3.7, pinned to pi-agent-core's copy). zod refinements
(quizInputProblems, askOptionProblems, live-nodeId) moved from schemas
into execute — the model sees the same retryable error shape.
- question-core + contracts/probe TypeBox-authored (single source shared
with the UI); tool-parts parse seam runs the boundary's full chain
(Convert → Check → Clean → refinement), fixing the live-smoke lockout:
pi's boundary Value.Converts correctAnswer: 17 → '17' but the record
keeps raw args, so a parse without Convert rejected a posed quiz
(invisible pending question, turn ended). Pinned with the real smoke
emission shape.
- Subagents converted: demo-builder/researcher AgentTool-native,
visual-maker/pdf-reader declaration-only pi-ai Tools with local execute
fns, extraction submit schemas re-authored in TypeBox (owner option 2:
zod defaults → Type.Optional + fill at the submit seam).
- Accepted delta, flagged: TypeBox-native schemas skip pi's JSON-Schema
primitive coercion fast path... (boundary still Value.Converts), and
prober-quiz unknown keys ride through (Value.Clean strips at display).
- Tests: 988 passing (was 967 at the recon baseline); tool-abort and
read-tool rewritten against the native seams; question-core/question-
tools/tool-parts/demos/web-search/graph-tools/plan-tree/probe-done/
tool-toggles converted to Value.Check/execute pins.
Live smoke (PORT=3004, real model): native ls turn streams end-to-end; a
real quiz pose (first attempt missing correctAnswer → tool-input-error →
model retry → posed) terminates the turn cleanly and persists pending.
baseAgentOptions (teacher-agent.ts) and runSubagent (subagents/pi-run.ts)
passed getApiKey: () => openRouterApiKey(). pi-agent-core resolves
(config.getApiKey ? await config.getApiKey(provider) : undefined) || config.apiKey
(agent-loop.js:191) — a provided hook OVERRIDES pi-ai's native per-provider
env resolution, and this one ignored its provider arg: every non-OpenRouter
provider was sent the OpenRouter key and 401'd (owner repro: openai and
anthropic direct). Deletion restores the architecture's stated intent —
the keys seam reflects file-saved keys into env and pi-ai reads them per
request. openRouterApiKey() export kept (now consumer-less; visual-maker/
pdf-reader read process.env.OPENROUTER_API_KEY directly for their own
OpenRouter-specific fetch paths — correct, untouched).
Pinned by tests/api-key-hook.test.ts: built teacher + prober expose NO
getApiKey hook. Gates: typecheck 0, 987/987, lint clean, build green.
The prober was planning perfect 7-question batches instead of probing. Three
surfaces now say the same thing — one question, adapt, next question:
- prober/10 rewritten to ~2.6k (from 5.6k): serial discipline leads (ask,
wait, choose the next question from the answer); 'How your questions
arrive' deleted (owner: does nothing); 'You cannot get there by thinking'
deleted (owner: didn't reduce thinking); 'never plan a batch' and the
first-message-length line dropped per owner — no overcorrection, no
strong-instruction inflation. Binary search compressed to four moves.
- tool descriptions: the async-era fossil 'Ask exactly one question per
tool call; after calling, end your turn and wait' — which permitted
multi-call batches — swapped for 'ONE question per turn: ... his answer
determines the next' in all four places (prober quiz + ask descriptions,
teacher quiz description, both zod question-field describes).
- shared/40-formatting: markdown/LaTeX coverage now explicitly names quiz
questions, options, and explanations ('in prose and in quiz questions
alike') — the prober wasn't using LaTeX in questions.
Prober composed prompt ~13.2k -> ~10.0k chars. Pins preserved verbatim
('measurement should be efficient', 'the read tool can inspect it directly').
Gates: typecheck 0, 986/986, lint clean, build green.
Owner-diagnosed: the turn gate checked only model coordinates, so a
pinned slot whose provider key was deleted passed the gate and died
mid-stream masked to 'An error occurred.' The key is now gated where
the model is — by what would RUN (env-first keyValue truth, so an
env-shadowed key passes).
- core/settings.ts: the provider_key_missing envelope family (code,
provider, slots, human line naming the provider from AUTH_PROVIDER_KEYS
labels) + ProviderKeyMissingError + isProviderKeyRefusal matcher.
- chat-handler turn gate: refuses 503 before beginTurn when the resolved
conversation's provider key is gone (slots: [strong]).
- Job runner: same check via an injected dep (agent layer cannot read
the key store) — a keyless engine refuses the spawn through the
existing failure-shape convention, no phantom run.
- Extraction pipeline: refuseIfStagesUnready refuses BEFORE any batch
(first offender: extractor before auditor); the trigger skips info-
level, /api/graph/import surfaces the 503 envelope per file.
- Tests: 8 new pins (turn gate both ways, runner refusal, pipeline
first-offender + pass-through, envelope factory, matcher); the three
turn-driving suites seed OPENROUTER_API_KEY like any real setup.
- Doc fix: stale EXTRACTION.chunkTokens comment in the manual-extract
route (left over from eed4b4e).
Known residual (pre-existing, flagged for master): the manual extract
route logs BOTH config refusals as console.error (fire-and-forget path);
the coverage ledger keeps everything uncovered either way.
Owner-locked extraction-UX redesign (dispatch 2026-09-24), part 1 (data
layer; the UI chunk follows).
THRESHOLD -> SETTING
- config.ts's EXTRACTION.chunkTokens dies; the default moves into
FACTORY_SETTINGS.graph.extractionChunkTokens (4096, the old constant's
value). GraphSettings gains the optional field, Required at resolution
via the standard ?? FACTORY pattern — pre-redesign records resolve the
default cleanly.
- FLOOR 500 (the implementer's call, documented on the field): every run
pays the graph-carrying prompt overhead out of the write channel's
money — a smaller budget fires penny-batches that cost more per
extracted token than the content is worth. The sanitizer CLAMPS up
(bd68486 clamp-never-reject discipline: fractions floor, junk drops);
the settings PUT REJECTS below-floor with the rationale in the message.
No ceiling: a large budget just delays extraction — the user's stated
cost/thoroughness tradeoff, not a misconfiguration to police.
- The trigger (turn-service.triggerExtraction) takes the budget as a
parameter — chat-handler passes effective.graph.extractionChunkTokens,
the fresh per-turn read, same as every other gate.
- The pipeline's BATCH side threads through the same number:
ExtractionPipelineDeps gains an OPTIONAL getChunkTokens getter with the
factory default (Teacher's 24 test fixtures stay untouched; the two
call-site lines inside extraction.ts are the only cross-band edits,
flagged). chat-handler + backfill/import-notes wire the live
effectiveSettings — trigger and batches share one budget, always.
PER-SESSION COVERAGE READ
- sessionCoverage(chatId) (server/chats.ts): the chat view's ring-meter
block — totalMessages/coveredCount/uncoveredCount/coveredTokens/
uncoveredTokens (the same coverageOf derivation /api/graph/coverage
serves, scoped to one chat), running (chat-kind extraction registered
for this id), and lastFoldedAt (the chat's newest coverage row's fold
— a cheap single pass; null before any fold). GET /api/chats/[id]
carries it as coverage: {...}; unknown chat -> null (the route 404s
first).
- /api/graph/coverage (the all-sessions list) compiles untouched — the
per-session read is a scoping of it, not a supersession; the UI's
sidebar/ring chunk decides which surface it consumes.
Tests: factory/sanitizer pins (clamp-up, fraction-floor, junk-drop),
resolution pins (absent -> 4096, stored 2000 wins), settings-body pins
(valid/absent/fraction/below-floor 400/junk 400), sessionCoverage pins
(derived split, newest-fold scoping, cross-session isolation, running via
the registry, null on unknown chat). 893/893.
Part 1 — the ledger + remnants interlude (master-dispatched, owner ruling
2026-09-23: nothing configures the graph beyond chat models):
- The embedding ledger surface is gone: readEmbeddings/writeEmbeddings
ports, the EmbeddingsLedger contract, EMBEDDINGS_FILE (config, store
impls, graphRevision source list), and graph-deletion's vector-pruning
step. On-disk data files stay untouched (inert history).
- The extraction embedder seam dies wholesale: ExtractionPipelineDeps
drops the embedder dep (chat-handler + backfill/import-notes stubs go
with it), the 'embedding' phase leaves ExtractionPhase, the usage
capture/persist helpers die, and stages.embedding leaves
ExtractionRunUsage (historical records keep their shape — reads are
lenient). Extraction is two stages again, like the local-embedder era.
- Consumer-less remnants: latestQuizOutcomes + QUIZ_EVENT_KINDS,
clusterContentId, matchClusterLabelings + Community + ClusterLabeling,
the readClusterLabels/appendClusterLabel port seams + store impls +
LEDGER_FILES.clusterLabels. NamingRunUsage/readNamingRuns STAY
(standing ruling — the cost pill renders historical naming spend);
clusterHue + MIN_CLUSTER_SIZE stay (live).
- The 'labeler' role dies: BACKGROUND_MODEL_ROLES, ROLE_SLOT, ModelRole,
the catalog pin projections, and the UI roster strings. RolePins keys
shrink with the union.
The notes birth fallback (owner-approved 2026-09-23, surfaced by the
interlude's live smoke: production notes imports create provenance-EMPTY
nodes (importNotes passes messages: null — the file's anchor is the
mapping line), so the provenance-only birth rule left every imported
file unclustered and invisible to the prober's session list, against the
locked file-birth ruling. A provenance-empty node now births under its
FIRST mapping line's chatId (ledger order); only notes-born nodes ever
hit the fallback (chat nodes always carry provenance; a re-encountered
node's link-append re-opens the provenance rule). Deletion-safe: mapping
lines of deleted chats are stripped alongside their provenance.
Live-verified on port 3004: a fresh notes import clusters under its
src_ id with the file path as the label, end to end with no embedder.
845/845 tests, typecheck/lint/build green.
EOF
)
The owner's zero-configuration ruling (2026-09-23): nothing configures
the graph beyond chat models. The prior-knowledge injection — its budget
dial, its runner, its marker part, its dock/dispatch entries — is
removed across all bands; the prober now reads the learner's persistent
knowledge map on demand through two gated tools instead:
- list_sessions: every birth session, names + kinds + counts + dates
(graphSessionIndex, Platform's core seam)
- list_propositions(session): what was born there + the session's own
latest understanding level + that evaluation's date
(graphSessionPropositions) — names/levels/dating ONLY, no statement
text (owner ruling: the map's organization is the prober's business,
its prose is not)
Registered via createProberTools when chat-handler passes graphStore
(the graphToolsEnabled gate decided server-side, per-turn fresh); the
prober doctrine gains the map-review nudge at session start.
Also: runner's JobKind special-casing collapses (KIND_ROLE/KIND_RUNNERS
are plain records again), semantic-clusters.ts dies with its test, the
embedder dep's doc tells the wave-7 stub story, and ports.ts's
readClusterLabels/readNamingRuns surface stays untouched for Platform's
ledger-port interlude.
848/848 tests, typecheck/lint/build green.
Wave 7 chunk A (950ddfd) replaced semantic clustering with birth-session
clusters labeled by session names — the LLM namer became dead code, but
extraction still called it, minting naming runs nobody reads on every
merge. Real model spend wasted per extraction.
Removed (agent band):
- src/agent/subagents/cluster-namer.ts — the naming sub-agent, deleted
whole (refreshClusterLabels, clusterNamingSettled, the naming loop).
- extraction.ts: both post-merge garnish blocks (chat + notes paths),
the labeler test seam on ExtractionModels, and the import. The labeler
role's only consumer is gone; the 'labeler' entry in
BACKGROUND_MODEL_ROLES/ROLE_SLOT (core band) is now consumer-less —
flagged to Platform, not crossed.
- scripts/import-notes.ts: the settle-before-exit line + the named-
clusters count in the summary.
- tests: the cluster-naming describe (naming + failure-non-fatal tests)
and the dead CLUSTER_* fixtures; the labeler-less BYOK test repurposed
as the kill pin — a merge mints NO naming runs and NO label records.
NOT touched (flagged, outside the agent band): matchClusterLabelings
(core/clusters.ts), clusterContentId (core/graph.ts — now consumer-
less), the appendClusterLabel/readClusterLabels port seams + store
impls + ClusterLabeling type (store/core band), tests/clusters.test.ts's
matchClusterLabelings describe. Cost-summary's naming stage untouched
per Platform's ruling (historical spend display reads readNamingRuns,
which stays live).
Gates: typecheck, 921/921, lint, build.
Platform's partial-spend interlude made embedNewNodes fail BY RESULT:
landed vectors + partial usage ride the return value instead of a
throw. The extraction-side doc comments still claimed 'failed embeds
contribute nothing' — now they state the pay-once truth: skipped
(unconfigured) embeds carry no usage, failed ones carry the landed
batches' steps, and a mid-run failure leaves a three-stage run's
partial embedding spend on the ledger. The call-site comment says
clustering runs with what LANDED (not 'without the new nodes'), and
recordEmbedFailure's doc marks the catch belt-and-suspenders for
injected embedders that still throw (production no longer does).
Wave-6 review (Teacher stream): the injection runner's null-embedder
branch failed with 'graph.embeddingModel is null', but the resolver
collapses BOTH an unset model AND a set model whose provider key
vanished into null — a key race mid-session would have reported a
false 'null' setting. The message now names both causes honestly.
Pinned by a new failure-mode test (null resolver → honest error, never
a phantom query).
Review verdicts (no changes needed): the session-deletion guard sits at
every chat-merge boundary (per batch, fresh lookup, rides the existing
AbortError path — nothing merged, nothing checkpointed, registry
cleaned in the finally); persistRunUsage's deleted-chat skip composes
with appendSourceUsage's vanished-source no-op so no usage record
dangles; notes sources have no deletion surface so the notes path needs
no guard. The embedding tail's non-fatal contract holds (failure →
recordEmbedFailure → no stage, extractor/auditor usage still persists
through the catch).
DELETE /api/chats/[id]?mode=preserve-graph|delete-nodes (mode required):
- preserve-graph: only extraction bookkeeping dies (coverage rows,
legacy checkpoints) — graph ledgers byte-untouched
- delete-nodes: the chat's ledger traces die. Node survival = surviving
provenance point / mapping line (notes imports carry empty
provenance) / surviving event. The creator's aliases + provenance are
stripped from a surviving node's create line (owner ruling
2026-09-23: name/statement/createdAt are the shared identity);
orphan nodes, their edges, events, mappings, positions, embedding
vectors and cluster members die; cluster labels stay (immutable)
Gates: 404 unknown, 409 turn_in_flight / jobs_running (with jobIds),
503 extraction_winding_down after a 10s abort-and-poll; the purge is
one sync block (ledgers → jobs → chat record last). rewriteLedgers is
atomic (tmp + rename). graphRevision (max mtime over graph data files)
rides the activity feed + the response as the change stamp.
Anti-resurrection: the chat pipeline re-checks chat liveness at each
merge — a run that slips past the wind-down aborts with nothing merged
or checkpointed; its partial spend dies with the chat.
Fix 1 — clamp, never reject (owner ruling, interlude 2026-09-22): a
model swap with a pinned thinking level must JUST WORK. The owner hit
the write-side 400 live (Claude Fable 5.1/xhigh → GLM 5.3 flash:
"z-ai/glm-5.3-flash does not support thinking level xhigh"). The
supported-level verdicts in validateChatSettings and
validateModelsSettings are DEAD — the WRITE clamps the pin to the new
model's nearest supported level (pi's clampThinkingLevel, the same
translation piEngineForSpec applies at runtime) and stores the clamped
result; the server is the truth — the record shows the clamped level
after write.
New pure helpers in model-catalog.ts, applied by ALL THREE write
routes (section, defaults, per-chat — one verdict seam, so every
surface that used the old 400 gets the clamp uniformly):
- clampModelsLevels: every slot's level clamps against the SLOT'S OWN
model (the parser already requires a model on each slot write).
- clampChatLevel: the body's own model, else the record's resolved
strong slot (the chain's fallback); no model anywhere → rides
unchecked (runtime clamps at engine build; the turn gate refuses
the model-less chat anyway).
Fix 2 — DEFAULT_THINKING_ORDER moves from config.ts to core/settings
(core barrel, AUTH_KEYS precedent — one source of truth, no drift).
The UI imports it to auto-populate a thinking level on FRESH model
picks; the ordering has no other client role. Legacy unset records
keep the runtime defaultThinkingLevel fallback at engine build
untouched. Import sites updated: agent/models.ts, tests.
Verified live on 3004: the owner's exact case stores max (not 400);
a bare chat-level write clamps to low against the strong slot's glm;
unknown provider/model still 400s with the provider list.
Tests: the level-verdict suites became clamp suites (the owner's case
pinned verbatim: xhigh→max, minimal→low, off→low, supported
pass-through, level-less slots ride); the machine-aware
expectBareLevelVerdict chain is retired — clampChatLevel takes the
strong slot as an explicit param, so the tests are pure again.
Gates: typecheck, lint, 839/839, build.
The fix wave (4e1f545) put research + pdf in the prober's registry per
the owner ruling, but the doctrine never mentioned them. Owner chose
(2026-09-22) a brief grounded mention over schema-only discovery: the
paragraph extending 'read his actual material' now covers pdf (same
grounding for his PDF material, no copy-pasting) and research
(sparingly, only when pinning a fact would genuinely sharpen a probe
question — probing is measurement, not exploration, and a research
brief is never a reason to stall the questions).
The BYOK refusal path (engineFor came up empty) only flipped the job
record to error and returned: no steer. But the teacher had already
dispatched — the ack promised a result as a new message — so the lesson
stalled at the dock card with the turn over and the teacher waiting
forever on a brief that would never arrive. Every other failed job
steers (the research/demo/visual/pdf catch paths all deliver the kind's
failure shape through finishJob); the refusal now follows the SAME
convention:
- record the honest error (envelope wording + missing tier) WITHOUT
flipping status — finalizeJob owns the terminal flip after delivery;
- build the kind's failure result (Research failed:/FAILED:/PDF failed:
shapes the isSubagentFailure heuristics read) and steer it through
finishJob, so the wake pipeline resumes the lesson and the teacher
reads the refusal and tells the learner.
Also cleaned the stale engineFor doc (the model-slots-era block still
said 'factory-pin chain ... absent → config pins' — two comments had
accumulated on the field; merged into one BYOK-accurate block).
Platform's refusal test pinned the old no-steer semantics; updated to
assert the new convention (steer delivered, record flipped by
finalizeJob after). The other piEngineFor callers (teacher-agent,
pi-run, visual-maker, pdf-reader) are deliberate honest-thrower
fallbacks (opts.engine ?? piEngineFor(role)) — missed-injection
detectors, kept as-is.
Owner decision 2026-09-22: 'it doesn't make sense to have any default
model configured anywhere; the application is bring your own keys.'
config.ts: MODELS deleted; ModelRole becomes an explicit 9-role union;
DEFAULT_THINKING_ORDER (medium, low, high, minimal, xhigh, max) bakes
the unset-level default — first entry the model supports; explicitly
set levels keep pi's native nearest-clamp.
core/settings-resolution: the factory layer dissolves — the chain is
role override → slot → NOTHING. resolveSettings(record) takes no
factory. Empty chains resolve to null/undefined; a thinkingLevel-only
slot is inert (stored, never run). resolveConversationSpec: chat.model
override over the strong slot, null when both empty. resolveRoleSpec
spans all nine roles for the job/extraction engines.
Refusal envelope (core/settings.ts): modelsNotConfigured(missing) —
{error, code: 'models_not_configured', missing: [slot tiers]} +
ModelsNotConfiguredError with .envelope. Per-need gates, each honest:
- chat-handler: universal turn gate (new + continuing turns) returns
503 before beginTurn — no liveness, no write-ahead persist, no graph
trigger before the gate fires.
- runner.runJob: unresolvable engine lands the job as error with the
envelope wording; no phantom loop, no result steer.
- extraction: refuseIfStagesUnconfigured throws before any stage
(turn-service info-skips so coverage retries next turn; the import
route surfaces per-file code+missing); labeler-less runs skip
cluster naming with a console line and still merge.
Write side: slot entries must carry a model — level-only slot writes
are 400s (absent slot key = clear). piEngineForSpec resolves unset
thinking levels via DEFAULT_THINKING_ORDER; piEngineFor now throws
honestly; thinkingLevelFor/modelIdFor deleted; teacher-agent, pi-run,
visual-maker, pdf-reader all read engine.level.
model-catalog: defaults/slots/rolePins null when the chain is empty
(contracts nullable) — no phantom ids in payloads. Stores drop
MODEL_PINS entirely.
Tests: pin-leaners converted to explicit injection (per-chat model
pins on scripted chats — exactly what the setup flow writes; faux
engines for agent-construction tests). New coverage: the 503 gate
envelope, the runner refusal, the extraction refusal + labeler skip.
Gates: typecheck, lint, 815/815, build.
Fix wave (2026-09-20), two owner-approved fixes:
Fix 1 — the searchAvailable gate was dead code on the live wiring path:
createTeacherTools RECEIVED ctx.searchAvailable but dropped it when
building the inner JobToolContext, so jobTools' research gate always took
its ?? true default (jobTools-direct tests passed; the production path
never saw the signal). Threaded through — the gate now reads the live
server signal on the teacher registry too.
Fix 2 — the prober registry gains the research + pdf dispatch tools
(owner ruling, 2026-09-20): these two are phase-omnipresent — the prober
can verify facts or read learner PDFs while assessing. demo/visual stay
teach-phase-only (they need the lesson's plan context). Same gating as
the teacher side: toggle-filtered, research subject to the search gate
(honestly absent when no engine is live — the prober's dispatch closes
over the SAME jobTools builder, one gate, one semantics). probe_done
stays always-on-by-construction; the learner-fs set untouched.
- ProberToolContext (extends ProbeDoneContext with the job wiring);
createProberAgent takes it; chat-handler injects jobStore + spawn +
searchAvailable into the prober ctx.
- TeacherToolContext.spawn's stale input union fixed (PdfInput was
missing from the TYPE — runtime already passed it).
- Tests: the threaded-gate describe in web-search.test.ts pins the LIVE
factory path both ways (teacher + prober); question-tools' registry
key-sets gain research/pdf; the prober toggle test pins the
research:false path.
Gates: typecheck, 807/807 tests, lint, build — all green.
Chunk B, item 3 of the search port + slots wave (2026-09-20): the flagged
gap closed — ExtractionModels deps were engine VALUES, keeping
extractor/auditor/labeler config-pin-bound while the job runner went
slot-native. The pipeline now takes the runner's exact treatment:
- ExtractionPipelineDeps.engineFor: (role) => PiEngine — resolved FRESH
per stage call, so a settings PUT between stages lands. Direct test
values (extractor/auditor/labeler) remain the priority seam, checked
FIRST; absent values fall to the chain; absent chain → config pins
(unchanged).
- All six stage call sites (chat + notes extract/audit, the cluster
namer x2) resolve through stageEngine(role, direct).
- chat-handler's extraction composition injects the full chain —
piEngineForSpec(resolveRoleSpec(role, readSettings(), MODEL_PINS)) —
same wiring as the job runner.
Chunk B, item 4 (session-level slot-anchored model picks): VERIFIED
already complete from Platform's model-slots wave — the conversation
engine chain (chat-handler) anchors on effective.models.strong, the
picker's default surface (catalogDefaultsFor) falls back to strong,
both pinned in tests. No phase split (user decision, 2026-09-20).
Capability-filtered pickers live in the settings UI (main 343be10).
Tests: engineFor fills the stages end-to-end (both propositions merged
through the chain, consulted in pipeline order) + the direct-value
priority pin (engineFor that throws when consulted).
Gates: typecheck, 795/795 tests, lint, build — all green.
Chunk A of the search port wave (2026-09-20), per master's assignment:
- web.ts: the Google Custom Search backend is gone — web_search now runs
on the core search port (SearchProvider), injected as a dependency
(import law: the tool never imports server code). createWebTools(search)
resolves the LIVE provider fresh per call; null = the honest
unavailability string (no silent fallback, owner decision). web_fetch
unchanged (needs no engine).
- runner: JobRunnerDeps.searchProvider — the server composition threaded
into the researcher; jobs already in flight when a key disappears still
fail honestly through the same seam, never crash.
- jobTools: the research dispatch tool DROPS from the teacher's registry
when ctx.searchAvailable() is false (owner decision: registry-level
drop — the model never sees the tool, no doomed dispatch, no error
card). Absent seam = available (tests). chat-handler injects
server/searchAvailable + searchProvider.
- tests: stubFetch (the Google wire stub) retired; jobs.test.ts threads
a recording stub provider through all four research runners and asserts
the scripted query hit the port; new web-search.test.ts covers the
factory contract (null → honest string, hits, error strings, fresh
resolution, web_fetch independence) + the registry drop matrix.
Gates: typecheck, 793/793 tests, lint, build — all green.
Model slots wave (2026-09-20), chunk B:
- settings: new `models` section {strong, cheap, vision} slots (each
field independent, same shape as a role override); ROLE_SLOT maps all
nine roles to tiers (strong=prober+teacher, cheap=researcher/
demoBuilder/extractor/auditor/labeler, vision=visualMaker+pdfReader);
resolution chain per role: override → slot → factory pin.
- config: ModelSpec gains provider (all pins openrouter, the BYOK
default); MODEL_PINS in stores.ts now spans all nine roles (the strong
slot's factory rep IS the prober pin).
- engine: piModels() registers EVERY builtin provider catalog;
piEngineForSpec(resolved spec) is the new engine factory (pi-ai
resolves per-provider keys from env itself); piEngineFor stays as the
config-pin fallback. chat-handler resolves the conversation chain
(chat override → strong slot → pin) and injects engineFor into the job
runner, closing the old gap where per-role overrides never reached
job engines.
- /api/models: models federate ONLY keyed providers (providers surface
with keyPresent, env-wins, fresh per call); slots (resolved tier pins)
join rolePins; catalogDefaultsFor inherits the strong slot; the models
settings section gets the same catalog gate as chat.
- pricing: PRICES/CONTEXT_LIMITS uniformly keyed provider:id
(regenerated from ALL pi-ai catalogs, 1354 keys); usage stamps
(TurnUsage, JobUsage, extraction stages) switch to provider:id —
owner decision: no legacy compat, old stamps render unpriced.
Search port + model slots wave, chunk A (owner design, 2026-09-20):
- core/search.ts: SearchHit + SearchProvider port + SearchStatus —
pure, browser-safe; adapters arrive as injected dependencies.
- server/search.ts: Tavily + Brave adapters (wire shapes VERIFIED
against docs.tavily.com's OpenAPI + Brave's dashboard docs — Tavily
auth is the Bearer header, not the legacy body api_key; Brave is
X-Subscription-Token) + engine resolution per owner decisions: a key
ALONE activates its engine (no engine choice needed), explicit
agent.search.engine pins the preference (both keys: Tavily first),
a pinned engine with a missing key is honestly unavailable — no
silent fallback, no default engine.
- Curated keys catalog: GOOGLE_SEARCH_API_KEY/GOOGLE_CSE_ID retire
(CSE closed to new customers, dies Jan 2027 — stored values become
preserved unknown vars); TAVILY_API_KEY (tvly- shape) + BRAVE_API_KEY
(length guard) join. UI keys-section TOOL_IDS now derives from the
catalog — cannot drift.
- GET /api/search: {available, engine, reason} — the honest status for
the settings gray-out; researcher-disable semantics ride the same
signal (Teacher stream rewires the tool onto the port).
- agent.search settings field (schema + sanitizer + strict parser);
web.ts's WebSearchResult is now the core SearchHit (re-exported).
amosblomqvist