Files
learnharness/components/chat.tsx
T
amosblomqvist 1d45d77303 feat: graph context injection — the session's prior knowledge, steered in at session start (vision-context-injection.md)
One shot per chat, fired at the first turn's onFinish: the session's
topic (its only typed message) is embedded as an EmbeddingGemma retrieval
query; the semantically closest graph propositions are selected under the
session's token budget (pure budget, no similarity threshold — the
decided doctrine) and delivered as a marker-only steer through the wake
pipeline, queuing behind pending questions like every sub-agent result.

- contracts: ChatSettings.graphContextTokens (default 1024, max 4096,
  step 128, 0 = off); GraphContextData + graphContextText; the
  graph-context JobKind + buildGraphContextSteer (id steer_ctx_<chatId>)
- core/graph-context.ts: brute-force cosine top-K, greedy chars/4 fill
  under budget, most-recent UnderstandingLevel fold, count cap 64,
  parse-don't-cast at the marker seam
- server: embedQuery (role-asymmetric query format, same warm singleton
  — vector-space parity by construction); the runner (startGraphContext-
  Injection one-shot gates; runGraphContextJob validates ledger parity);
  wake branches (chatIdOf/jobIdOf, finalizeJob done, recoverStaleJobs
  re-deliver or contextRespawn); abortJob kills context jobs SILENTLY
  (nothing was promised in the lesson); turn-service trigger seam
- adapter: data-graph-context flattens to the canonical block (malformed
  parts have no model face)
- UI: the budget slider (home dock + /chat/new draft, the session-data
  strip); the quiet SystemLine "~537 · 10 propositions"; dock kind
  "Knowledge context"; model-picker PUT now spreads settings (a model
  pick can no longer clear the sibling override)

Live-smoked end to end on the real embedder + real GLM: seeded ledger
(10 propositions), session on "modal logic: possible worlds and the
necessity box" — job in the dock, queued behind the prober's question,
delivered on the answer, SystemLine in the transcript, wake turn reads
the block and continues the probe. Gates: tsc, lint, 459/459, build.
2026-09-15 07:54:50 +03:00

1082 lines
46 KiB
TypeScript
Raw Blame History

This file contains ambiguous Unicode characters
This file contains Unicode characters that might be confused with other characters. If you think that this is intentional, you can safely ignore this warning. Use the Escape button to reveal them.
"use client";
import { createContext, Fragment, useCallback, useEffect, useLayoutEffect, useMemo, useRef, useState } from "react";
import { useChat } from "@ai-sdk/react";
import { isToolUIPart, lastAssistantMessageIsCompleteWithToolCalls } from "ai";
import { createChatTransport } from "@/components/chat-transport";
import { useStickToBottomContext } from "use-stick-to-bottom";
import {
Conversation,
ConversationContent,
ConversationEmptyState,
ConversationScrollButton,
} from "@/components/ai-elements/conversation";
import { Message, MessageContent, MessageResponse } from "@/components/ai-elements/message";
import { Reasoning, ReasoningContent, ReasoningTrigger } from "@/components/ai-elements/reasoning";
import { Shimmer } from "@/components/ai-elements/shimmer";
import { Separator } from "@/components/ui/separator";
import { AnsweredCard } from "@/components/answered-card";
import { artifactRenderers } from "@/components/artifacts";
import { SubagentSteerCard } from "@/components/artifacts/steer-card";
import { Composer } from "@/components/composer";
import { JobsContext } from "@/components/jobs-context";
import { StatsBar } from "@/components/stats-bar";
import { graphContextOfMessage, type ChatSettings } from "@/src/core";
import { formatTokens } from "@/components/usage-format";
import { PlanBlock, PlanPopup, latestTree } from "@/components/plan-block";
import { PlanMarginalia } from "@/components/plan-window";
import { WaypointsIcon } from "lucide-react";
import type { PlanBlockData } from "@/src/core";
import { QuestionDock } from "@/components/question-dock";
import { SidebarToggle } from "@/components/sidebar";
import type { ArtifactPartProps, JobRecord, Phase3, QuestionAnswerData, TeacherUIMessage, TurnUsage } from "@/src/core";
import { PhaseChip, PhaseDivider, planNodeCount } from "@/components/phase-divider";
import { findPendingToolCall, phaseAt3, phaseTimeline } from "@/src/core";
/** One width for everything: transcript, cards, and the bottom blob share this column. */
/**
* Width tokens (wave-5 reskin, owner decision): the teaching PROSE spans the
* full session width with a thin px-6 margin; the composer/question DOCK is
* deliberately narrower (~46rem centered) per the reference. Same ELEMENT,
* same place — width polish only.
*/
const COLUMN = "mx-auto w-full max-w-none px-6";
const DOCK = "mx-auto w-full max-w-[46rem] px-4";
// The DOCK container's fixed width (46rem) and the plan marginalia's
// geometry (owner, plan-placement round): the bare diagram sits in the
// dock band's RIGHT flank — MARGIN_GAP from the DOCK container's edge and
// the same from the column's right edge. The box width is pure CSS, so
// the diagram shrinks CONTINUOUSLY with the flank; MARGIN_MIN_W is only
// the fade floor (owner, resize round) — below it the marginalia fades
// out over the normal transition instead of popping away, and it stays
// mounted while the question owns the band so widening fades it back.
const DOCK_W = 46 * 16;
const MARGIN_GAP = 16;
const MARGIN_MIN_W = 220;
// The dock overlay's gradient band (h-24) — part of the reserved tail.
const GRADIENT_H = 96;
// Fold thresholds (owner, scroll round 1, tuned round 3): pixel distance
// from the bottom — position, not velocity. Leave the bottom → the box
// folds (visual only); reach it again → it re-opens. 96 (was 64, owner
// round 3) tolerates a small peek upward without folding; the dead band
// vs EXPAND_DIST absorbs trackpad drift.
const COLLAPSE_DIST = 96;
const EXPAND_DIST = 16;
/**
* Server-sync channel for artifact renderers: the CENTRAL jobs poll in Chat
* calls syncFromServer() when a background job settles — the wake turn's
* steer + response are already persisted by then, so fresh messages replace
* local state in one pull. sendText feeds the demo error-fix loop.
* (contracts.ts "THE ASYNC PROTOCOL" §4.)
*/
export const ChatActionsContext = createContext<{
chatId: string;
syncFromServer: () => Promise<void>;
/** Send a user message into the session (demo error-fix loop, W6). */
sendText: (text: string) => void;
} | null>(null);
/** useLayoutEffect that is silent during SSR (client components still render on the server in Next). */
const useIsoLayoutEffect = typeof window === "undefined" ? useEffect : useLayoutEffect;
/** A wake-pipeline steer message: user-role delivery whose metadata marks it
* as a sub-agent result — rendered as an artifact card, never a bubble. */
/** The knowledge-context card's one-liner: the numbers the observability
* doctrine wants visible at the seam — tokens and proposition count
* (parse-don't-cast; a malformed part still renders a quiet generic line). */
function graphContextLine(message: { parts: ReadonlyArray<{ type: string; data?: unknown }> }): string {
const data = graphContextOfMessage(message);
return data
? `Knowledge context — ~${formatTokens(data.tokens)} · ${data.propositions.length} propositions`
: "Knowledge context";
}
/** The quiet system line (custom-message scheme rendering): centered, muted,
* small — the transcript's voice for events that are not speech (wake
* scaffolding, the learner's plan verdict). */
function SystemLine({ children, testId }: { children: React.ReactNode; testId?: string }) {
return (
<div
data-testid={testId}
className="flex items-center justify-center gap-1.5 py-1 text-xs text-muted-foreground"
>
{children}
</div>
);
}
function isSubagentSteer(m: TeacherUIMessage): boolean {
return (m.metadata as { subagentResult?: unknown } | undefined)?.subagentResult != null;
}
/**
* The teaching session. All chat state lives in useChat; the server persists
* the canonical transcript after every turn.
*
* The bottom blob is ONE persistent shell whose contents swap (chat composer ↔
* question dock ↔ collapsed bar). The shell is sticky INSIDE the scroller —
* never overlapping the transcript at rest, never covering the scrollbar —
* and DockOverlay animates its height on variant changes.
*/
export function Chat({
chatId,
title,
initialMessages,
initialPrompt,
initiallyInFlight,
initialSettings,
}: {
chatId: string;
title: string;
initialMessages: TeacherUIMessage[];
initialPrompt?: string;
/** A teacher turn is running server-side but attached to no client
* (navigated away mid-turn, returned while in flight). */
initiallyInFlight?: boolean;
/** The record's per-chat model/thinking settings (the picker's seed). */
initialSettings?: ChatSettings;
}) {
const { messages, sendMessage, addToolOutput, setMessages, status, error, stop } =
useChat<TeacherUIMessage>({
id: chatId,
messages: initialMessages,
transport: createChatTransport(chatId),
// After the learner answers (addToolOutput), automatically send the
// continuation request so the teacher reacts to the answer.
sendAutomaticallyWhen: lastAssistantMessageIsCompleteWithToolCalls,
});
// New chat with a topic: send the first message exactly once.
const sentInitial = useRef(false);
useEffect(() => {
if (!sentInitial.current && initialPrompt && initialMessages.length === 0) {
sentInitial.current = true;
void sendMessage({ text: initialPrompt });
}
}, [initialPrompt, initialMessages.length, sendMessage]);
// A turn running server-side but attached to no client of ours (navigated
// away mid-turn, returned while in flight): the record is write-ahead
// persisted (answer included), but the teacher's reaction only lands when
// the orphaned turn finishes — poll for that, don't resurrect.
const [turnInFlight, setTurnInFlight] = useState(initiallyInFlight ?? false);
// Mount continuation: if the stored transcript ends with a fulfilled tool
// (a job result landed while no client was open), resume the lesson — the
// SDK's sendAutomaticallyWhen only fires inside addToolOutput, so reloads
// need this nudge. Once per mount; the turn that follows ends in text (or
// a pending question), so no loop. NEVER while a turn is already running
// server-side — that would double-run the teacher (the server would 409
// it; the poll below resumes instead).
const mountContinuation = useRef(false);
useEffect(() => {
if (mountContinuation.current) return;
mountContinuation.current = true;
if (
!turnInFlight &&
status === "ready" &&
messages.length > 0 &&
lastAssistantMessageIsCompleteWithToolCalls({ messages })
) {
void sendMessage();
}
// mount-only by design (turnInFlight read at initial value)
}, []);
const pending = findPendingToolCall(messages);
const localStream = status === "submitted" || status === "streaming";
const busy = localStream || turnInFlight;
// Teach-phase check (W5, vision §2): a pending question that lands once a
// plan node is OCCUPIED (phaseAt3 "teach") starts collapsed behind a
// "Check your understanding" reveal button; probe- and plan-phase
// questions auto-open as today. A pending question can never precede its
// phase's seam in the transcript (the teacher's turn ends at the question
// call), so the session's current phase IS the question's phase.
const requiresReveal = pending !== null && phaseAt3(messages) === "teach";
// The phase at every message (one linear fold): the transcript draws a
// color-coded dotted divider at each boundary; the header chip reads the
// session's current phase.
const phases = useMemo(() => phaseTimeline(messages), [messages]);
// A tool call still streaming its input: the teacher is writing a question.
// Shown as a transcript shimmer, exactly like the thinking indicator.
const questionStreaming =
!pending &&
messages.at(-1)?.role === "assistant" &&
(messages.at(-1)?.parts.some((p) => isToolUIPart(p) && p.state === "input-streaming") ?? false);
const [text, setText] = useState("");
// Jobs never gate the composer (contracts.ts "THE ASYNC PROTOCOL" §4): job
// tools resolve with an ack mid-turn, so the learner can interject while
// sub-agents run — a feature, not a race: the wake pipeline queues steers
// until the current turn ends. Only a QUESTION pending or a turn in
// flight (client or server-side wake turn) gates sending.
const canSend = text.trim().length > 0 && !pending && !turnInFlight;
const syncFromServer = useCallback(async () => {
const res = await fetch(`/api/chats/${chatId}`);
if (!res.ok) return;
const record = (await res.json()) as { messages?: TeacherUIMessage[]; turns?: TurnUsage[] };
if (Array.isArray(record.turns)) setTurns(record.turns);
if (!Array.isArray(record.messages)) return;
setMessages(record.messages);
// setMessages does NOT fire sendAutomaticallyWhen (the SDK only checks
// it inside addToolOutput) — without this the teacher would stay silent
// after a job result lands. sendMessage() with no args re-requests with
// the current list: the continuation. (contracts.ts protocol §4.)
if (
status === "ready" &&
lastAssistantMessageIsCompleteWithToolCalls({ messages: record.messages })
) {
void sendMessage();
}
}, [chatId, setMessages, sendMessage, status]);
// The live plan popup (wave 4): opened by any transcript plan block, the
// top-bar plan chip, and the dock band's marginalia.
const [planOpen, setPlanOpen] = useState(false);
// The session's live plan (drives the top-bar chip; the marginalia
// derives its own from the same transcript).
const planTree = useMemo(() => latestTree(messages), [messages]);
// ── The plan verdict gate (docs/probe-teach-split.md §4) ──
// The plan has been presented (a tree exists) and the session still sits in
// the planning phase → the composer's send square becomes the stacked
// Accept / Decline buttons. Acceptance is the explicit plan→teach seam
// (a data-plan-accepted part on the message); decline asks for revision.
const planVerdict = useMemo(
() => phaseAt3(messages) === "plan" && planTree !== null,
[messages, planTree],
);
const sendVerdict = useCallback(
(accepted: boolean) => {
const note = text.trim();
setText("");
// TWO REAL MESSAGES, ONE TURN (the custom-message scheme, user-decided
// 2026-09-12): a typed note is the learner's own user message, and the
// verdict is its own custom message — marker part ONLY, no human-
// authored text (the model-facing wording lives in the adapter:
// record-adapter flattens it; nothing fake is persisted as speech).
// setMessages + sendMessage() (no args) fires ONE request carrying
// both — the same single-turn mechanism syncFromServer already uses,
// so no 409 dance and no double teacher turn.
const stamp = Date.now().toString(36);
const rnd = crypto.randomUUID().slice(0, 8);
const appended: TeacherUIMessage[] = [];
if (note) {
appended.push({
id: `msg_${stamp}_${rnd}`,
role: "user",
parts: [{ type: "text", text: note }],
});
}
appended.push({
id: `verdict_${stamp}_${rnd}`,
role: "user",
parts: [
{
type: "data-verdict",
id: `vd_${stamp}_${rnd}`,
data: { accepted, at: new Date().toISOString() },
},
],
});
setMessages([...messages, ...appended]);
void sendMessage();
},
[text, messages, setMessages, sendMessage],
);
// ── Session usage (wave-4 stats bar) ──
// The money channel: the record's turns log, folded client-side by
// core/prices.ts. Seeded once on mount (the page doesn't carry turns),
// refreshed when the local stream settles (each turn appends to the log
// server-side), and kept fresh by syncFromServer whenever it already
// fetches the record. No turns yet → the bar renders "—" everywhere.
const [turns, setTurns] = useState<TurnUsage[]>([]);
// The per-chat model control state (model-picker writes via PUT; the
// record is the source of truth, lifted here so the stats bar re-renders).
const [settings, setSettings] = useState<ChatSettings | undefined>(initialSettings);
// The dock's reserved tail height (DockOverlay reports it; the spacer at
// the end of the transcript renders it). Lives here because the spacer is
// inside the scroller while the dock overlays OUTSIDE it — two parents.
// 208 ≈ gradient(96) + composer block(~112): the SSR/pre-hydration
// guess so the static paint has a sane tail. The first layout effect
// replaces it with the measured truth before paint.
const [dockReserve, setDockReserve] = useState(208);
const syncTurns = useCallback(async () => {
try {
const res = await fetch(`/api/chats/${chatId}`);
if (!res.ok) return;
const record = (await res.json()) as { turns?: TurnUsage[] };
if (Array.isArray(record.turns)) setTurns(record.turns);
} catch {
// transient blip: stats stay on the last known figures
}
}, [chatId]);
useEffect(() => {
void syncTurns();
}, [syncTurns]);
// After the learner's own turn ends the log has a fresh entry — and the
// Post-turn adoption + the HANDOFF WATCH. Two things can trail our own
// stream's end: (1) the server's persist/endTurn — milliseconds, and the
// record is only complete once not-in-flight (plan blocks attach at
// persist); (2) a WAKE TURN — the probe→teacher handoff fires
// ~0.75–1.2s later (the pump's grace), and NO jobs poll is running to
// notice it: probe-phase sessions have zero job traffic, so the jobs
// poll is switched off and the open client never learns the teacher's
// turn exists (the freeze the user hit 2026-09-11 — the sidebar saw
// busy→idle while the session sat frozen). So: adopt at the first idle
// read as before, then keep watching briefly — if turnInFlight flips
// true, hand off to the wake-adoption poll below (busy shimmer, adopt
// at finish).
const prevStatusRef = useRef(status);
useEffect(() => {
if (prevStatusRef.current !== "ready" && status === "ready") {
// immediate: turns-only refetch — the stats bar updates even while
// the server's persist/endTurn lags the client by milliseconds
void syncTurns();
let cancelled = false;
let timer: ReturnType<typeof setTimeout> | undefined;
let adopted = false;
const WATCH_ATTEMPTS = 8; // 8 × 500ms: persist lag + the ~1.2s handoff
const adopt = async (attempt: number) => {
if (cancelled) return;
try {
const res = await fetch(`/api/chats/${chatId}`);
if (res.ok) {
const record = (await res.json()) as { turnInFlight?: boolean };
if (!adopted) {
// phase 1: wait out our own turn's persist lag, then adopt once
if (record.turnInFlight !== true) {
adopted = true;
void syncFromServer();
}
} else if (record.turnInFlight === true) {
// phase 2 (the watch): a wake turn started — the handoff.
setTurnInFlight(true); // → the wake-adoption poll takes over
return;
}
}
} catch {
/* transient — keep polling */
}
if (attempt < WATCH_ATTEMPTS && !cancelled) timer = setTimeout(() => void adopt(attempt + 1), 500);
};
void adopt(0);
return () => {
cancelled = true;
if (timer) clearTimeout(timer);
};
}
prevStatusRef.current = status;
}, [status, chatId, syncFromServer, syncTurns]);
// ── Central job observability (contracts.ts "THE ASYNC PROTOCOL" §4) ──
// ONE poll of /api/jobs drives every job surface (transcript dispatch
// cards, the sub-agents panel) and re-syncs the transcript exactly once
// when a job settles. A settle no longer implies the record is complete:
// CONTINUATION COALESCING flips the job done at claim time — while the
// learner's own answer turn is still streaming (and a settle can race a
// second job's wake turn). Syncing mid-turn would swap in a stale snapshot
// mid-stream, which breaks the SDK's in-place continuation merge (duplicate
// message ids). So: a settle during a busy turn defers the sync; the next
// idle poll tick adopts the finished record. Poll fast while jobs run,
// slow otherwise (new-dispatch discovery), and not at all when this
// session has never dispatched a job.
const [jobs, setJobs] = useState<JobRecord[]>([]);
const jobPartCount = useMemo(
() =>
messages.reduce(
(n, m) =>
n +
m.parts.filter(
(p) => p.type === "tool-research" || p.type === "tool-demo" || p.type === "tool-visual",
).length,
0,
),
[messages],
);
const anyRunning = jobs.some((j) => j.status === "running");
const pollMs = jobPartCount === 0 && jobs.length === 0 ? 0 : anyRunning ? 1000 : 4000;
const localStreamRef = useRef(false);
localStreamRef.current = localStream;
const jobStatusRef = useRef<Map<string, string>>(new Map());
// A settle seen during a busy turn — the sync is owed, not dropped.
const pendingSyncRef = useRef(false);
useEffect(() => {
if (pollMs === 0) return;
let cancelled = false;
const poll = async () => {
if (cancelled) return;
try {
const res = await fetch(`/api/jobs?chatId=${encodeURIComponent(chatId)}`);
if (!res.ok) return;
const data = (await res.json()) as { jobs?: JobRecord[]; turnInFlight?: boolean };
if (cancelled) return;
const list = data.jobs ?? [];
for (const j of list) {
const prev = jobStatusRef.current.get(j.id);
if (prev === "running" && j.status !== "running") {
// The record is only guaranteed complete when no turn is in
// flight — locally or server-side. A coalesced settle fires
// mid-turn by design; adopt the record on the first idle tick.
if (localStreamRef.current || data.turnInFlight === true) {
pendingSyncRef.current = true;
} else {
void syncFromServer();
}
}
jobStatusRef.current.set(j.id, j.status);
}
if (
pendingSyncRef.current &&
!localStreamRef.current &&
data.turnInFlight !== true
) {
pendingSyncRef.current = false;
void syncFromServer();
}
setJobs(list);
// Wake-turn adoption: a server-side teacher turn with no client request
// of ours (sub-agent delivery) — the same machinery as returning
// mid-turn: busy state + record poll to completion.
if (data.turnInFlight === true && !localStreamRef.current) setTurnInFlight(true);
} catch {
/* transient fetch failure — keep polling */
}
};
void poll();
const timer = setInterval(poll, pollMs);
return () => {
cancelled = true;
clearInterval(timer);
};
}, [pollMs, jobPartCount, chatId, syncFromServer]);
// Re-attach after mid-turn navigation (the key fix): poll the record until
// the orphaned turn finishes (turnInFlight flips false server-side in
// onFinish, AFTER persist), then adopt the completed transcript — the
// syncFromServer above already runs the job-continuation check. Until then
// the session reads as busy: thinking shimmer, composer gated.
useEffect(() => {
if (!turnInFlight) return;
let cancelled = false;
let timer: ReturnType<typeof setTimeout> | undefined;
const poll = async () => {
if (cancelled) return;
try {
const res = await fetch(`/api/chats/${chatId}`);
if (res.ok && !((await res.json()) as { turnInFlight?: boolean }).turnInFlight) {
setTurnInFlight(false);
void syncFromServer();
return;
}
} catch {
/* transient network hiccup — keep polling */
}
if (!cancelled) timer = setTimeout(poll, 1500);
};
timer = setTimeout(poll, 1000);
return () => {
cancelled = true;
if (timer) clearTimeout(timer);
};
}, [turnInFlight, chatId, syncFromServer]);
const onSend = () => {
const t = text.trim();
if (!t || busy || pending) return;
setText("");
void sendMessage({ text: t });
};
return (
<ChatActionsContext.Provider value={{ chatId, syncFromServer, sendText: (t: string) => void sendMessage({ text: t }) }}>
<JobsContext.Provider value={jobs}>
{/* The session surface: the transcript column (header + scroller) and
the plan-tree panel as full-height SIBLINGS — the panel spans top to
bottom exactly like the left sidebar, and the header (name + phase
chip) spans from the left sidebar's edge to the panel's edge. The
dock overlays the column bottom (outside the scroller); a spacer in
the transcript reserves its height. */}
<div className="flex h-dvh">
<div className="flex min-w-0 flex-1 flex-col">
{/* The top bar is its own block (owner, top-fade round 4): in flow
above the scroller on the plain page background — no gradient. Its
bottom outline (border-b border-border) matches the one that
separates the left sidebar — same width, same color (owner, r5). h-10, not py-2: the toggle (a size-9 icon button)
renders only while the sidebar is CLOSED — an auto-height row
would collapse when it disappears and slide the title + chip up. */}
<header className="flex h-10 items-center gap-3 border-b border-border px-3">
<SidebarToggle />
<div className="min-w-0 flex-1 truncate font-title text-2xl tracking-tight">{title}</div>
{/* The plan chip (owner, plan-placement round): session-level
structure beside the phase chip — access to the full-DAG popup
at every width; the marginalia is the wide-screen glance view. */}
{planTree && (
<button
type="button"
data-testid="plan-chip"
onClick={() => setPlanOpen(true)}
title="Open the full plan"
className="flex shrink-0 cursor-pointer items-center gap-1.5 rounded-full border border-border/80 px-2 py-0.5 text-xs font-medium uppercase tracking-wide text-muted-foreground transition-colors hover:border-border hover:text-foreground"
>
<WaypointsIcon className="size-3.5 text-primary" aria-hidden />
Plan
</button>
)}
<PhaseChip phase={phases.at(-1) ?? "probe"} />
</header>
<div className="relative flex min-h-0 flex-1">
<Conversation className="h-full min-w-0 flex-1" initial="instant">
<ConversationContent className="min-h-full gap-0 p-0">
<div className={`${COLUMN} flex flex-1 flex-col gap-4 pt-4`}>
{messages.length === 0 && !busy && (
<ConversationEmptyState
title="New session"
description="The teacher will open by probing what you already know."
/>
)}
{messages.map((m, mi) =>
m.role === "user" && isSubagentSteer(m) ? (
// Sub-agent delivery (wake-pipeline steer): NO message bubble — the
// artifact card sits in the transcript column directly, full width
// and centered like the teacher's artifacts (contracts.ts "THE ASYNC
// PROTOCOL" §4). Key fallback: legacy transcripts (pre id-hygiene)
// can still hold empty ids — the index keeps keys unique for them.
<SubagentSteerCard key={m.id || `m-${mi}`} message={m} />
) : m.role === "user" && m.parts.some((p) => p.type === "data-graph-context") ? (
// The prior-knowledge injection's delivery (vision-context-injection.md):
// quiet system line, the numbers only — the block itself was for the
// model (adapter flattening); the part data stays the audit trail.
<SystemLine key={m.id || `m-${mi}`} testId="graph-context">
{graphContextLine(m)}
</SystemLine>
) : m.role === "user" && m.parts.some((p) => p.type === "data-probe-handoff") ? (
// The probe→teach handoff wake (custom-message scheme): quiet
// system line, never a user bubble — the wake is scaffolding,
// the "Planning" divider marks the seam, and the model-facing
// instruction lives in the adapter.
<SystemLine key={m.id || `m-${mi}`} testId="probe-handoff">
Preparing your plan…
</SystemLine>
) : m.role === "user" && m.parts.some((p) => p.type === "data-verdict") ? (
// The learner's plan verdict (custom-message scheme): its own
// message, the quiet marker — the accept is the phase seam, a
// decline asks for revision. Never a user bubble: the model-
// facing wording is the adapter's, and this is the learner's
// ACT, not their words (a typed note, if any, rendered above
// as its own real message).
<SystemLine
key={m.id || `m-${mi}`}
testId={
(m.parts.find((p) => p.type === "data-verdict") as
| { data?: { accepted?: boolean } }
| undefined)?.data?.accepted
? "verdict-accepted"
: "verdict-declined"
}
>
{(m.parts.find((p) => p.type === "data-verdict") as
| { data?: { accepted?: boolean } }
| undefined)?.data?.accepted
? "✓ Plan accepted"
: "↻ Plan revision requested"}
</SystemLine>
) : (
<Fragment key={m.id || `m-${mi}`}>
{mi > 0 && phases[mi] !== phases[mi - 1] && (
<PhaseDivider phase={phases[mi]!} nodeCount={planNodeCount(m)} />
)}
<Message from={m.role} className={m.role === "assistant" ? "max-w-full" : undefined}>
<MessageContent className={m.role === "assistant" ? "w-full" : undefined}>
{m.role === "user" ? (
// User text renders plain: the model's LaTeX pipeline is not
// applied to learner-authored text ($ stays $, no surprises).
<div className="lh-prose whitespace-pre-wrap">
{m.parts.filter((p) => p.type === "text").map((p) => p.text).join("\n")}
</div>
) : (
m.parts.map((part, pi) => {
const isLastPart = mi === messages.length - 1 && pi === m.parts.length - 1;
switch (part.type) {
case "text":
return <MessageResponse key={pi}>{part.text}</MessageResponse>;
case "reasoning":
return (
<Reasoning key={pi} isStreaming={status === "streaming" && isLastPart} className="mb-2">
<ReasoningTrigger />
<ReasoningContent>{part.text}</ReasoningContent>
</Reasoning>
);
case "step-start":
return pi > 0 ? <Separator key={pi} className="my-3" /> : null;
case "data-plan-block":
// The plan in the chat (wave 4): a frozen snapshot
// block — rendered as carried, never re-derived.
return (
<PlanBlock
key={pi}
data={part.data as PlanBlockData}
onOpen={() => setPlanOpen(true)}
/>
);
case "tool-quiz":
case "tool-ask_user_question": {
// Pending/streaming questions live in the dock, not the
// transcript; answered ones render as graded records.
if (part.state === "output-available") {
return <AnsweredCard key={pi} answer={part.output as QuestionAnswerData} />;
}
return null;
}
default: {
// Artifact extension point (contracts.ts): job
// parts and future artifacts render via the
// registry — add renderers there, not here.
if (part.type.startsWith("tool-")) {
const Renderer = artifactRenderers[part.type.slice("tool-".length)];
if (Renderer) return <Renderer key={pi} part={part as ArtifactPartProps["part"]} />;
}
return null;
}
}
})
)}
</MessageContent>
</Message>
</Fragment>
))}
{(status === "submitted" || turnInFlight) && <Shimmer className="text-sm">Thinking…</Shimmer>}
{questionStreaming && <Shimmer className="text-sm">Writing a question…</Shimmer>}
{error && (
<div className="rounded-lg border border-destructive/50 bg-destructive/10 px-3 py-2 text-sm">
Something went wrong: {error.message}
</div>
)}
</div>
{/* The dock's reserved tail (DockOverlay reports its height):
keeps at-rest transcript content scrolling fully clear of
the overlay — the flow-space replacement for the old
sticky dock. Zero only pre-measurement (the first layout
effect reports before paint). */}
<div aria-hidden data-testid="dock-spacer" style={{ height: dockReserve }} />
</ConversationContent>
{/* The dock: an overlay sibling of the scroller (the
use-stick-to-bottom pattern — like the scroll button). Before
the button in DOM order so the button keeps painting above
the dock, as it always has. */}
<DockOverlay
pending={pending}
chatId={chatId}
settings={settings}
onSettingsSaved={setSettings}
plan={<PlanMarginalia messages={messages} />}
onOpenPlan={() => setPlanOpen(true)}
requiresReveal={requiresReveal}
phase={phases.at(-1) ?? "probe"}
turns={turns}
onReserve={setDockReserve}
composer={
<Composer
// Stop applies only to a LOCAL stream: a server-side turn has no
// client request to abort (stop() would be a no-op lie).
busy={localStream}
value={text}
onChange={setText}
onSend={onSend}
onStop={() => stop()}
canSend={canSend}
verdict={
planVerdict
? {
onAccept: () => sendVerdict(true),
onDecline: () => sendVerdict(false),
}
: undefined
}
/>
}
onSubmitAnswer={(output) => {
if (!pending) return;
void addToolOutput({
tool: pending.tool as "quiz" | "ask_user_question",
toolCallId: pending.toolCallId,
output,
});
}}
/>
<ConversationScrollButton />
</Conversation>
</div>
</div>
</div>
{planOpen && <PlanPopup messages={messages} onClose={() => setPlanOpen(false)} />}
</JobsContext.Provider>
</ChatActionsContext.Provider>
);
}
/**
* The floating blob: ONE persistent shell, OVERLAID at the bottom of the
* column (absolute — a sibling of the scroller, the use-stick-to-bottom
* pattern; it replaced the old in-flow sticky dock, which died of the
* fold-latency problem below). Contents swap between the chat composer and
* the question dock; shell height morphs between variants (WAAPI, quick). A
* spacer at the end of the transcript (rendered by ChatView from onReserve
* reports) holds the dock's flow-space, so at-rest content scrolls fully
* clear of the overlay.
*
* Why an overlay (owner, scroll round 1): an in-flow dock IS scroll geometry
* — folding it shrinks the scroll content, so the fold had to wait until the
* learner was a full dock-height past the bottom (any closer and the browser
* clamp pinned them back, re-expanding the box: the oscillation). Decoupled
* from the flow, the fold is VISUAL ONLY: it fires as soon as the learner
* leaves the bottom, and the reserved tail only ever GROWS while a question
* is open (a shrinking tail would reintroduce the clamp).
*
* Collapse: scroll up past COLLAPSE_DIST and the dock folds into a slim bar;
* click it (or scroll back to the bottom) to re-expand. The question form
* stays MOUNTED while collapsed (hidden, not unmounted) so a half-made
* selection survives.
*/
function DockOverlay({
pending,
requiresReveal,
phase,
turns,
composer,
onSubmitAnswer,
chatId,
settings,
onSettingsSaved,
onReserve,
plan,
onOpenPlan,
}: {
pending: ReturnType<typeof findPendingToolCall>;
/** Teach-phase pending question: arrives collapsed behind the reveal button (W5). */
requiresReveal: boolean;
/** Session phase at the pending question — scales the thinking field. */
phase: Phase3;
/** The session's usage log — the stats bar beneath the composer folds it. */
turns: TurnUsage[];
composer: React.ReactNode;
onSubmitAnswer: (output: QuestionAnswerData) => void;
/** The model control's pass-through (the stats bar hosts it). */
chatId: string;
settings?: ChatSettings;
onSettingsSaved: (settings: ChatSettings) => void;
/** Reports the dock's settled height (block + gradient) — ChatView's
* dockReserve state feeds the transcript's spacer from it. */
onReserve: (h: number) => void;
/** The plan marginalia's content (the bare diagram) — rendered only
* while a question owns the band. Null until a plan exists. */
plan?: React.ReactNode;
/** Opens the full-DAG popup (the marginalia + chip share it). */
onOpenPlan: () => void;
}) {
const { isAtBottom, scrollToBottom, scrollRef } = useStickToBottomContext();
// Teach-phase mount (session reload with a pending check): the dock starts
// as the collapsed invitation too — the check must not pop open on reload.
const [expanded, setExpanded] = useState(!requiresReveal);
// Reveal state (W5): false until the learner clicks the invitation. While
// false, the scroll machine must NOT auto-expand at the bottom — the
// learner reads the teaching first and opens the check when ready. Once
// revealed, the normal collapse machine rules apply untouched.
const [revealed, setRevealed] = useState(false);
const revealPending = requiresReveal && !revealed;
const pendingId = pending?.toolCallId ?? null;
const initialPendingRef = useRef(pendingId);
// New question MID-SESSION: open expanded if the learner is following at
// the bottom (scroll to reveal it); reading history → arrive as the bar.
// Teach-phase checks ALWAYS arrive collapsed behind the reveal button,
// even at the bottom. On session mount the dock simply starts in its
// initial state (expanded for probe, invitation for teach) — no scroll, no
// morph.
useEffect(() => {
if (!pendingId || pendingId === initialPendingRef.current) return;
if (requiresReveal) {
setRevealed(false);
setExpanded(false);
} else {
setExpanded(isAtBottom);
if (isAtBottom) void scrollToBottom();
}
// fire only when a NEW question lands
}, [pendingId]);
// ── Fold machine (position, not velocity — owner, scroll round 1) ──
// The dock is an overlay: folding is visual-only and scroll geometry never
// moves, so the thresholds can hug the bottom. Leave it → fold; reach it
// again → re-open. The dead band keeps trackpad drift from chattering.
const expandedRef = useRef(expanded);
expandedRef.current = expanded;
const pendingRef = useRef(!!pending);
pendingRef.current = !!pending;
const requiresRevealRef = useRef(requiresReveal);
requiresRevealRef.current = requiresReveal;
const revealedRef = useRef(revealed);
revealedRef.current = revealed;
// The plan marginalia's flank width (owner, plan-placement round): from
// the DOCK container's right edge to the block's right edge — the bare
// diagram's canvas. Tracked LIVE by a ResizeObserver on the block
// (owner, resize round): every geometry change — window resize, sidebar
// toggle, scrollbar appearance, the box's own rewrap — reports the real
// width instantly, so the fade floor never lags behind reality (the old
// render-time measurement went stale during drag-resizes and the mount
// decision popped late).
const [flankW, setFlankW] = useState(0);
const flankWRef = useRef(0);
// Part A of the same round (the resize mute): any block size change
// stamps this — the fold machine ignores scroll events for a beat after,
// because a resize/rewrap re-clamps scrollTop and emits PHANTOM scroll
// events mid-transient (the box flickered to the bar and back).
const geometryAtRef = useRef(0);
useIsoLayoutEffect(() => {
const block = blockRef.current;
if (!block) return;
const measure = () => {
const next = Math.round(block.offsetWidth / 2 - DOCK_W / 2 - 2 * MARGIN_GAP);
// Stamp the mute on WIDTH changes ONLY: column resizes are the
// phantom-scroll source. Height changes are the dock's own doing
// (morphs, question rewraps) — stamping those would self-mute the
// fold machine and swallow the at-bottom auto re-expand right
// after a fold (verified live: the box got stuck collapsed).
if (Math.abs(next - flankWRef.current) > 1) {
flankWRef.current = next;
setFlankW(next);
geometryAtRef.current = Date.now();
}
};
measure();
const ro = new ResizeObserver(measure);
ro.observe(block);
return () => ro.disconnect();
}, []);
// The overlay must not paint over the scroller's scrollbar (owner, scroll
// round 2): the scroller reserves symmetric gutters (scrollbar-gutter:
// stable both-edges), so inset the overlay's right edge by the gutter —
// zero on macOS overlay scrollbars, the real width on classic platforms.
const [sbInset, setSbInset] = useState(0);
useEffect(() => {
const el = scrollRef.current;
if (!el) return;
const measure = () => setSbInset(Math.round((el.offsetWidth - el.clientWidth) / 2));
measure();
window.addEventListener("resize", measure);
return () => window.removeEventListener("resize", measure);
}, [scrollRef]);
// Wheel over the dock must still scroll the transcript (owner, scroll
// round 2): the overlay is a SIBLING of the scroller, so the browser's
// native scroll chaining stops at the overlay (the old in-flow dock
// chained for free). Internal scrollers — the question area, textarea
// overflow, the model menus — keep native scrolling, including their
// own edge chaining; only when nothing under the cursor can take the
// delta do we forward it to the transcript.
useEffect(() => {
const root = blockRef.current;
const el = scrollRef.current;
if (!root || !el) return;
const canScroll = (n: HTMLElement, dir: number) => {
if (n.scrollHeight - n.clientHeight < 2) return false; // not scrollable
return dir < 0 ? n.scrollTop > 0 : n.scrollTop < n.scrollHeight - n.clientHeight - 0.5;
};
const onWheel = (e: WheelEvent) => {
if (e.deltaY === 0) return;
for (let n = e.target as HTMLElement | null; n && n !== root; n = n.parentElement) {
if (n instanceof HTMLElement && canScroll(n, e.deltaY)) return; // native
}
e.preventDefault();
el.scrollTop += e.deltaY * (e.deltaMode === 1 ? 16 : e.deltaMode === 2 ? el.clientHeight : 1);
};
root.addEventListener("wheel", onWheel, { passive: false });
return () => root.removeEventListener("wheel", onWheel);
}, [scrollRef]);
// ── The reserve: the spacer's height, reported upward ──
// Grows immediately (typing, new variants). NEVER drops while a question
// is open — a shrinking reserved tail is scroll geometry and the browser
// clamp would pin a nearby learner to the bottom (the old oscillation).
// In chat mode drops apply (the composer's typing shrink is today's
// harmless in-flow behavior; nothing auto-expands there).
const reserveRef = useRef(0);
const blockRef = useRef<HTMLDivElement>(null);
const offer = (blockH: number) => {
const next = blockH + GRADIENT_H;
if (next <= reserveRef.current + 0.5 && pendingRef.current) return; // fold immunity
reserveRef.current = next;
onReserve(next);
};
// Explicit expand (bar click): the smooth scroll to the bottom takes
// hundreds of ms, during which the position is still legitimately inside
// the collapse zone — the machine would instantly re-collapse the dock
// (open → snap shut → reopen at the bottom, the flicker). Suppress the
// collapse branch until the reveal scroll settles or is interrupted.
const expandIntentRef = useRef(false);
const intentTimerRef = useRef<ReturnType<typeof setTimeout>>(undefined);
const expandAndReveal = () => {
expandIntentRef.current = true;
setRevealed(true); // the teach-check invitation was accepted (W5)
setExpanded(true);
clearTimeout(intentTimerRef.current);
intentTimerRef.current = setTimeout(() => {
expandIntentRef.current = false;
}, 2000);
void Promise.resolve(scrollToBottom()).finally(() => {
clearTimeout(intentTimerRef.current);
expandIntentRef.current = false;
});
};
useEffect(() => {
const el = scrollRef.current;
if (!el) return;
const onScroll = () => {
if (!pendingRef.current) return;
// Resize mute (owner, resize round): a geometry change re-clamps
// scrollTop and the browser emits the clamp as a scroll event —
// reading the transient position folded the box to the bar for a
// flicker. Real scrolling resumes after the brief mute window.
if (Date.now() - geometryAtRef.current < 400) return;
const d = el.scrollHeight - el.clientHeight - el.scrollTop;
if (expandedRef.current) {
if (d > COLLAPSE_DIST && !expandIntentRef.current) setExpanded(false);
} else if (d < EXPAND_DIST) {
// Teach checks stay collapsed until the learner opens them (W5): the
// at-bottom auto-expand would pop the form over the teaching.
if (requiresRevealRef.current && !revealedRef.current) return;
setExpanded(true);
void scrollToBottom();
}
};
el.addEventListener("scroll", onScroll, { passive: true });
return () => el.removeEventListener("scroll", onScroll);
}, [scrollRef, scrollToBottom]);
// The settled block height (gradient excluded), offered to the spacer.
// Only ever called with NATURAL heights — never the animated mid-morph
// value (the iso effect's getAnimations guard wraps every call).
const measureBlock = () => {
const b = blockRef.current;
if (b) offer(b.getBoundingClientRect().height);
};
const mode = pending ? `question:${pending.toolCallId}` : "chat";
// The opening variant of a session appears instantly — the fade/rise is
// reserved for mid-session transitions.
const initialModeRef = useRef(mode);
// Height morph: animate the shell from its previous height to its new one
// whenever the variant (chat ↔ question ↔ collapsed bar) changes. Typing
// growth keeps the same stateKey and is NOT animated.
const shellRef = useRef<HTMLDivElement>(null);
const stateKey = `${mode}:${pending ? expanded : "-"}`;
const prevRef = useRef<{ key: string; height: number } | null>(null);
useIsoLayoutEffect(() => {
const el = shellRef.current;
if (!el) return;
const p = prevRef.current;
if (p && p.key === stateKey) {
// Same variant: keep the settled height fresh, but never measure while
// a morph is in flight — getBoundingClientRect would read the ANIMATED
// value and corrupt the start height of the next morph.
if (el.getAnimations().length === 0) {
const h = el.getBoundingClientRect().height;
p.height = h;
measureBlock(); // typing growth → the spacer grows with the box
}
return;
}
el.getAnimations().forEach((a) => a.cancel()); // flush to natural height before measuring
const h = el.getBoundingClientRect().height;
// Variant change → re-reserve. Growth applies instantly; a drop (the
// fold, or question → composer) is filtered by the offer policy.
measureBlock();
if (p && Math.abs(p.height - h) > 1) {
el.animate([{ height: `${p.height}px` }, { height: `${h}px` }], {
duration: 180,
easing: "cubic-bezier(0.2, 0.8, 0.2, 1)",
});
}
prevRef.current = { key: stateKey, height: h };
});
return (
<div className="pointer-events-none absolute inset-x-0 bottom-0" style={{ right: sbInset }}>
<div className="h-24 bg-gradient-to-t from-background via-background/85 to-transparent" />
<div ref={blockRef} className="relative bg-background pb-5">
<div className={DOCK}>
<div ref={shellRef} className="pointer-events-auto overflow-hidden rounded-2xl border bg-card shadow-lg">
<div key={mode} className={mode === initialModeRef.current ? undefined : "dock-variant"}>
{pending ? (
<QuestionDock
toolName={pending.tool as "quiz" | "ask_user_question"}
toolCallId={pending.toolCallId}
input={pending.input as never}
expanded={expanded}
reveal={revealPending}
phase={phase}
onExpand={expandAndReveal}
onSubmit={onSubmitAnswer}
/>
) : (
composer
)}
</div>
</div>
{/* The stats row: beneath the composer/question shell, not
flush to the screen bottom — static height, outside the
morphing shell so it never joins a height animation. */}
<StatsBar
turns={turns}
chatId={chatId}
settings={settings}
onSettingsSaved={onSettingsSaved}
/>
</div>
{/* The plan marginalia (owner, plan-placement round): the bare
diagram in the band's right flank — boundaless, meet-fitted, part
of the lesson's page rather than a floating widget. Mounted only
while a question owns the band (its reserved height is the canvas);
folding the dock fades + blurs it out of existence, so reading
the transcript in full wipes the margin clean. Click → popup.
Width (owner, resize round): the box shrinks continuously with
the flank; below MARGIN_MIN_W it FADES (never pops away — no
mid-question unmount), and widening past the floor fades it
back in. */}
{pending && plan && (
<button
type="button"
data-testid="plan-marginalia"
onClick={onOpenPlan}
title="Open the full plan"
style={{ left: `calc(50% + ${DOCK_W / 2 + MARGIN_GAP}px)` }}
className={`absolute bottom-5 right-4 top-0 cursor-pointer transition-[filter,opacity] duration-300 ease-out ${
expanded && flankW >= MARGIN_MIN_W
? "pointer-events-auto opacity-100 blur-0"
: "pointer-events-none opacity-0 blur-md"
}`}
>
{plan}
</button>
)}
</div>
</div>
);
}