sorokin.oleg.edu/learnharness
1
0
Fork 0
Code Issues Pull Requests Actions Packages Projects Releases 6 Wiki Activity

Merge current main into PR61 finishing branch

Browse Source
This commit is contained in:
amosblomqvist
2026-10-28 07:08:36 +02:00
parent 25a4687611 3dd52110bb
commit 812ea23419
170 changed files with 10528 additions and 2018 deletions
Download Patch File Download Diff File Expand all files Collapse all files
.github/ISSUE_TEMPLATE/bug_report.yml
+67
View File
@@ -0,0 +1,67 @@
name: Bug report
description: Report something that is not working as expected.
labels: [bug]
body:
- type: markdown
attributes:
value: |
Thanks for helping improve LearnHarness. Describe what you observed; you do not need to know the cause.
**This report is public.** Remove API keys, tokens, passwords and private information before posting. Never attach credential files or your whole data folder. Share only the learning content you are comfortable making public.
- type: textarea
id: description
attributes:
label: What happened?
description: Describe the problem and what you expected to happen instead.
validations:
required: true
- type: input
id: version
attributes:
label: LearnHarness version or build
description: Include the app version, installer filename or source commit. If you cannot find it, write "unknown".
validations:
required: true
- type: input
id: operating_system
attributes:
label: Operating system
description: For example, Windows, macOS or your Linux distribution. Include the OS version if known.
validations:
required: true
- type: textarea
id: reproduction
attributes:
label: Steps or context (optional)
description: How can we reproduce it? If it is intermittent or you are unsure, describe what you were doing when it happened.
validations:
required: false
- type: input
id: installation
attributes:
label: Installation method (optional)
description: For example, desktop installer, AppImage, system package, self-hosted web or source build.
validations:
required: false
- type: input
id: provider_model
attributes:
label: Provider and model (if relevant)
description: Include the provider, model and API-key or subscription connection type if relevant. Do not include any keys or tokens.
validations:
required: false
- type: textarea
id: transcript_excerpt
attributes:
label: Sanitized raw JSON or chat-log excerpt (optional)
description: A few relevant entries around the affected turn or tool call are especially helpful. Paste the excerpt directly, preserving JSON escaping where possible. Remove credentials, personal information and private learning content first. Leave this blank if the records are difficult to find.
render: json
validations:
required: false
- type: textarea
id: screenshots
attributes:
label: Screenshots or other context (optional)
description: Attach screenshots or add other relevant details. Check for private information before posting.
validations:
required: false
.github/ISSUE_TEMPLATE/config.yml
+1
View File
@@ -0,0 +1 @@
blank_issues_enabled: true
.github/ISSUE_TEMPLATE/feature_request.yml
+17
View File
@@ -0,0 +1,17 @@
name: Feature request
description: Suggest a new feature or an improvement.
labels: [enhancement]
body:
- type: markdown
attributes:
value: |
Ideas are welcome; you do not need a finished design or implementation plan. A request opens a discussion, not a commitment to build it.
**This report is public.** Do not include credentials or private information.
- type: textarea
id: idea
attributes:
label: What would you like to see?
description: Describe your idea and the problem it would help you solve.
validations:
required: true
.github/ISSUE_TEMPLATE/question.yml
+15
View File
@@ -0,0 +1,15 @@
name: Question
description: Ask about using LearnHarness.
labels: [question]
body:
- type: markdown
attributes:
value: |
**This report is public.** Do not include credentials or private information.
- type: textarea
id: question
attributes:
label: What is your question?
description: Include any relevant context. Version, OS, provider or model details can help when they relate to your question.
validations:
required: true
.gitignore
+3
View File
@@ -13,6 +13,9 @@ dist/
# dev auth/keys store (auth-keys wave): repo-root auth.json — secrets never committed
auth.json
chatgpt-credentials.json*
github-copilot-credentials.json*
xai-credentials.json*
# dev-fresh launcher scratch (fresh-user sim): disposable per launch
.fresh-data/
README.md
+4 -1
View File
@@ -12,7 +12,8 @@
<a href="https://learnharness.ai">Download</a> ·
<a href="#how-it-works">How it works</a> ·
<a href="docs/development.md">Documentation</a> ·
<a href="https://github.com/amosblomqvist/learnharness/issues">Issues</a>
<a href="https://github.com/amosblomqvist/learnharness/issues">Issues</a> ·
<a href="https://discord.gg/5tCHHQEB2e">Discord</a>
</p>
LearnHarness uses AI agents to plan and teach lessons from your current understanding to your goal. It builds a detailed picture of what you understand and how you reason. That picture is updated as you learn.
@@ -52,6 +53,8 @@ Add your provider API key and choose your models, then start a session with what
All three platforms support in-app updates.
**Dictation:** Create a [Deepgram API key](https://console.deepgram.com/) with the **Member role or higher** ([required for token minting](https://developers.deepgram.com/reference/auth/tokens/grant)) and add it under **Settings → Providers → Dictation**. Choose the language below the key; **Multilingual** supports automatic recognition and mixed-language speech. A `403 Insufficient permissions` error during token minting can indicate an insufficient key role.
## Your keys, your data
- **Your keys and models.** No LearnHarness subscription or markup. You pay your providers for usage; a live cost meter shows model costs.
app/api/chats/[id]/route.ts
+3 -1
View File
@@ -1,4 +1,4 @@
import { compactionInFlight, deleteChatSession, getChat, isTurnRunning, sessionCoverage, sessionCost, stores, type ChatDeleteMode } from "@/src/server";
import { compactionInFlight, deleteChatSession, getChat, isTurnRunning, sessionCoverage, sessionCost, stores, turnFailure, type ChatDeleteMode } from "@/src/server";
export async function GET(_req: Request, { params }: { params: Promise<{ id: string }> }) {
const { id } = await params;
@@ -19,6 +19,8 @@ export async function GET(_req: Request, { params }: { params: Promise<{ id: str
messages: chat.messages,
pricing: sessionCost(chat),
turnInFlight: isTurnRunning(id),
// Also reaches idle views that missed the failing turn's stream.
turnError: turnFailure(id),
compacting: compactionInFlight(id),
coverage: sessionCoverage(id),
});
app/api/chats/[id]/stream/route.ts
+4 -3
View File
@@ -5,8 +5,9 @@
* submitter too attaches here. One SSE stream per chat: a running turn →
* subscribe to the live turn object: FULL replay of the event log, then
* live chunks until settle ([DONE]). No running turn → the cheap terminal
* ({"type":"turn-settled"} + [DONE]): the record is current, the client
* refetches it.
* ({"type":"turn-settled"} + [DONE]): the client refetches the record.
* A retained settlement-failure notice precedes that terminal when a
* reply could not be saved before the submitter attached.
*
* 404 mirrors the POST handler's unknown-chat. Framework route is a thin
* wrapper over the server barrel (the import law); tests drive the seams
@@ -18,6 +19,6 @@ export async function GET(_req: Request, ctx: { params: Promise<{ id: string }>
const { id } = await ctx.params;
if (!getChat(id)) return Response.json({ error: "Unknown chat" }, { status: 404 });
const handle = runningTurn(id);
if (!handle) return settledTurnResponse();
if (!handle) return settledTurnResponse(id);
return piTurnResponse(handle);
}
app/api/settings/[section]/route.ts
+9 -7
View File
@@ -5,12 +5,13 @@
* discipline, same as the per-chat and defaults routes); sections are
* independent, so partial writes can never clobber each other.
*
* section ∈ general | chat | agent | appearance | graph | compaction
* section ∈ general | chat | agent | appearance | graph | compaction | models | dictation
* Body: the complete desired section object. Parse + validate strictly
* (server/settings-body.ts): a bad shape/type is a 400 carrying the
* supported list where one exists. The chat and models sections
* additionally pass the catalog gate (validateChatSettings /
* validateModelsSettings — unknown provider/model = 400). Thinking
* validateModelsSettings — unknown provider/model = 400, except unchanged
* model-slot selections may remain saved while unavailable). Thinking
* levels never reject a write (owner, interlude 2026-09-22): a pinned
* level CLAMPS to the named model's nearest supported level and the
* clamped value is stored — model swaps just work, same as the
@@ -63,6 +64,8 @@ export async function PUT(request: Request, ctx: { params: Promise<{ section: st
if (!parsed.ok) {
return Response.json({ error: parsed.error }, { status: 400 });
}
// Compare against the same record that this whole-section write replaces.
const record = readSettingsRecord() ?? { version: 1 };
let agentHealed: AgentSettings | undefined;
// Section-specific cross-validation (against the catalog / the prompt
@@ -73,11 +76,11 @@ export async function PUT(request: Request, ctx: { params: Promise<{ section: st
return Response.json({ error: verdict.error, supported: verdict.supported }, { status: 400 });
}
}
// Model slots wave (2026-09-20): the models section gets the same
// catalog gate the chat section gets — a slot must name a real
// provider+model and a thinking level it supports.
// New/replaced slot models pass the catalog gate. Unchanged saved model
// references may survive disconnect, including thinking-preference edits,
// so each unavailable slot can be repaired independently of the others.
if (section === "models") {
const verdict = validateModelsSettings(parsed.settings as ModelsSettings);
const verdict = validateModelsSettings(parsed.settings as ModelsSettings, record.models);
if (!verdict.ok) {
return Response.json({ error: verdict.error, supported: verdict.supported }, { status: 400 });
}
@@ -124,7 +127,6 @@ export async function PUT(request: Request, ctx: { params: Promise<{ section: st
// Read-modify-write ONE section of the record; everything else rides
// along untouched (other sections, unknown keys — sanitize preserves).
const record = readSettingsRecord() ?? { version: 1 };
const next = { ...record, [section]: sectionOut };
writeSettingsRecord(next);
return Response.json({ settings: next });
app/api/subscriptions/browser/route.ts
+6
View File
@@ -0,0 +1,6 @@
import { subscriptionBrowserResponse } from "@/src/server";
export const runtime = "nodejs";
export const dynamic = "force-dynamic";
export const GET = subscriptionBrowserResponse;
export const POST = subscriptionBrowserResponse;
app/api/subscriptions/route.ts
+6
View File
@@ -0,0 +1,6 @@
import { subscriptionControlsResponse } from "@/src/server";
export const runtime = "nodejs";
export const dynamic = "force-dynamic";
export const GET = subscriptionControlsResponse;
export const POST = subscriptionControlsResponse;
app/chat/[id]/page.tsx
+2 -1
View File
@@ -1,7 +1,7 @@
import { notFound } from "next/navigation";
import { Chat } from "@/components/chat";
import { slotsConfigured } from "@/src/core";
import { compactionInFlight, getChat, isTurnRunning, readSettingsRecord } from "@/src/server";
import { compactionInFlight, getChat, isTurnRunning, readSettingsRecord, turnFailure } from "@/src/server";
export const dynamic = "force-dynamic";
@@ -29,6 +29,7 @@ export default async function ChatPage({
// derived app-side (core/view.ts projectView; GET serves the same
// native messages).
initialMessages={chat.messages}
initialError={turnFailure(chat.id)}
initialConfigured={slotsConfigured(record?.models)}
initialPrompt={prompt}
// A teacher turn that outlived its client: the remounted view must
components/ai-elements/message.tsx
+11 -1
View File
@@ -34,7 +34,7 @@ import {
type PluginConfig,
} from "streamdown";
import { mdComponents } from "@/components/md-blocks";
import { transformMathDelimiters } from "@/src/core";
import { repairMathForDisplay, transformMathDelimiters } from "@/src/core";
export type MessageProps = HTMLAttributes<HTMLDivElement> & {
from: LhViewMessage["role"];
@@ -388,6 +388,16 @@ export const MessageResponse = memo(
MessageResponse.displayName = "MessageResponse";
/** Recovery is a tool-text presentation concern, not a shape/answer-data
* transform. Ordinary assistant/user prose keeps MessageResponse unchanged. */
export function ToolContentResponse({ children, ...props }: MessageResponseProps) {
return (
<MessageResponse {...props}>
{typeof children === "string" ? repairMathForDisplay(children) : children}
</MessageResponse>
);
}
export type MessageToolbarProps = ComponentProps<"div">;
export const MessageToolbar = ({
components/answered-card.tsx
+13 -12
View File
@@ -5,9 +5,9 @@ import { ConfidenceRail } from "@/components/confidence-rail";
import { GlyphMark } from "@/components/ui/glyph-mark";
import { ThinkingBlock } from "@/components/thinking-block";
import { cn } from "@/lib/utils";
import { MessageResponse } from "@/components/ai-elements/message";
import { ToolContentResponse } from "@/components/ai-elements/message";
import type { QuestionAnswerData } from "@/src/core";
import { parseAskInput, parseQuizInput } from "@/src/core";
import { answerForPresentation, parseAskInput, parseQuizInput } from "@/src/core";
/**
* Static rendering of an answered question (tool part in `output-available`).
@@ -19,7 +19,8 @@ import { parseAskInput, parseQuizInput } from "@/src/core";
* client — the card reveals his answer, the committed correct answer, and
* the explanation, with NO verdict chips (owner ruling 2026-09-27).
*/
export function AnsweredCard({ answer, input }: { answer: QuestionAnswerData; input?: unknown }) {
export function AnsweredCard({ answer: savedAnswer, input }: { answer: QuestionAnswerData; input?: unknown }) {
const answer = answerForPresentation(savedAnswer, input);
const [whyOpen, setWhyOpen] = useState(false); // the tucked explanation (predictions, owner 2026-10-03)
if (answer.dismissed) {
return (
@@ -86,13 +87,13 @@ export function AnsweredCard({ answer, input }: { answer: QuestionAnswerData; in
))}
</div>
<div className="mb-3 font-medium">
<MessageResponse>{answer.questionText}</MessageResponse>
<ToolContentResponse>{answer.questionText}</ToolContentResponse>
</div>
{parsedDetails && (
// The details line, same paint as the pending dock (owner,
// 2026-10-12): full markdown, LaTeX included.
<div className="mb-3 text-sm text-muted-foreground">
<MessageResponse>{parsedDetails}</MessageResponse>
<ToolContentResponse>{parsedDetails}</ToolContentResponse>
</div>
)}
@@ -131,14 +132,14 @@ export function AnsweredCard({ answer, input }: { answer: QuestionAnswerData; in
</span>
)}
<span className={cn(!sel && "text-muted-foreground")}>
<MessageResponse>{opt.label}</MessageResponse>
<ToolContentResponse>{opt.label}</ToolContentResponse>
{descByIndex.get(opt.index) && (
// Ask-option subtext (owner, 2026-10-12): same markdown path
// as the pending dock — LaTeX included. Quiz options carry
// none (owner ruling 2026-09-26); the map stays empty for them.
<MessageResponse className="lh-prose block text-xs text-muted-foreground">
<ToolContentResponse className="lh-prose block text-xs text-muted-foreground">
{descByIndex.get(opt.index)}
</MessageResponse>
</ToolContentResponse>
)}
</span>
</div>
@@ -211,9 +212,9 @@ export function AnsweredCard({ answer, input }: { answer: QuestionAnswerData; in
{whyOpen && (
<div className="mt-1.5 rounded-lg bg-muted py-2 text-sm text-foreground/50">
<span className="font-medium text-muted-foreground">Why: </span>
{/* The markdown wrapper must go truly inline — a block <p> after the
"Why:" label created the oversized leading gap (owner's fix list). */}
<MessageResponse className="inline [&_p]:inline [&_p]:m-0">{answer.explanation}</MessageResponse>
{/* Keep the first paragraph beside the label, not every paragraph:
later paragraphs must retain their recovered document structure. */}
<ToolContentResponse className="inline [&>p:first-child]:inline [&>p:first-child]:m-0">{answer.explanation}</ToolContentResponse>
</div>
)}
</div>
@@ -225,7 +226,7 @@ export function AnsweredCard({ answer, input }: { answer: QuestionAnswerData; in
{isExplain && answer.committed && (
<div className="mt-3 rounded-lg bg-muted py-2 text-sm text-muted-foreground">
<span className="font-medium text-foreground">The answer: </span>
<MessageResponse className="inline [&_p]:inline [&_p]:m-0">{answer.committed}</MessageResponse>
<ToolContentResponse className="inline [&>p:first-child]:inline [&>p:first-child]:m-0">{answer.committed}</ToolContentResponse>
</div>
)}
</div>
components/artifacts/dispatch-card.tsx
+2 -2
View File
@@ -15,7 +15,7 @@ import { BrailleSpinner } from "@/components/braille-spinner";
*
* job unknown → "dispatched" pulse (the jobs feed hasn't seen it yet)
* running → live activity card (activity line + elapsed)
* done/error → quiet one-liner ("delivered" / "failed") — the artifact
* done/error → quiet one-liner ("finished" / "failed") — the artifact
* itself arrives in the steer message right below, so the
* box must not duplicate it.
*/
@@ -54,7 +54,7 @@ export function DispatchCard({ toolCallId, kind }: { toolCallId: string; kind: J
>
{failed ? <XIcon className="size-3.5 text-red-500" /> : <CheckIcon className="size-3.5" />}
<span className="font-medium uppercase tracking-wide">{label}</span>
<span>{failed ? "failed" : "delivered"}</span>
<span>{failed ? "failed" : "finished"}</span>
</div>
);
}
components/artifacts/svg-image.tsx
+37
View File
@@ -0,0 +1,37 @@
/** Two display-only repairs for inputs the old HTML embedding accepted.
* Do not rewrite stored artifacts/downloads or attempt general XML repair.
* Comments, CDATA and processing instructions are copied verbatim. */
export function svgImageSource(svg: string): string {
let foundRoot = false;
return svg.replace(
/<!--[\s\S]*?(?:-->|$)|<!\[CDATA\[[\s\S]*?(?:\]\]>|$)|<\?[\s\S]*?(?:\?>|$)|<\/?[a-zA-Z](?:[^<>"']|"[^"]*"|'[^']*')*>|&nbsp;/g,
(token) => {
if (token.startsWith("<!--") || token.startsWith("<![CDATA[") || token.startsWith("<?")) return token;
if (!foundRoot && /^<svg(?=[\s/>])/i.test(token)) {
foundRoot = true;
// Consume complete attribute values, so text mentioning xmlns
// inside a different attribute cannot masquerade as a declaration.
const attrs = token.matchAll(/\s([^\s=/>]+)(?:\s*=\s*(?:"[^"]*"|'[^']*'|[^\s>]+))?/g);
const hasNamespace = Array.from(attrs).some((attr) => attr[1] === "xmlns");
if (!hasNamespace) token = token.replace(/^<svg/i, '$& xmlns="http://www.w3.org/2000/svg"');
}
return token.replace(/&nbsp;/g, "&#160;");
},
);
}
/** Keep generated SVG in an image document, outside the app's CSS and ID scope.
* This is a display boundary, not a sanitizer for the original SVG download. */
export function SvgImage({
svg,
className,
width,
}: {
svg: string;
className?: string;
width?: number;
}) {
// The source is already an in-memory vector image; no Next image optimization.
// eslint-disable-next-line @next/next/no-img-element
return <img src={`data:image/svg+xml;charset=utf-8,${encodeURIComponent(svgImageSource(svg))}`} alt="Figure" className={className} width={width} draggable={false} />;
}
components/artifacts/visual-card.tsx
+12 -18
View File
@@ -8,6 +8,7 @@ import { Toggle } from "@/components/settings/controls";
import { useSessionMermaid } from "@/components/session-overlay";
import { downloadFile } from "@/components/artifacts/download";
import { DispatchCard } from "@/components/artifacts/dispatch-card";
import { SvgImage } from "@/components/artifacts/svg-image";
import type { ArtifactPartProps, VisualResult } from "@/src/core";
/**
@@ -16,22 +17,10 @@ import type { ArtifactPartProps, VisualResult } from "@/src/core";
* the chat view's jobs feed). The SVG itself renders from the steer message
* (sub-agent result delivery) via VisualOutputCard.
*
* The SVG is model-authored for the learner's own machine (their own teacher
* wrote it, verified by looking) — dangerouslySetInnerHTML is acceptable
* here, with a cheap hygiene pass (sanitizeSvg: strip <script> blocks, on*
* handlers, and javascript:/data:text/html URLs) since it lands in the
* app's DOM.
* Both display modes use an SVG image document: generated markup and its
* styles never join the app's DOM. Downloads still contain the original SVG.
*/
/** Cheap SVG hygiene — not a full sanitizer, just removing the obvious exec vectors. */
export function sanitizeSvg(svg: string): string {
return svg
.replace(/<script[\s\S]*?<\/script\s*>/gi, "")
.replace(/<script[^>]*\/?>/gi, "")
.replace(/\son[a-z]+\s*=\s*(?:"[^"]*"|'[^']*'|[^\s>]+)/gi, "")
.replace(/(href|xlink:href)\s*=\s*(?:"\s*(?:javascript|data:text\/html)[^"]*"|'\s*(?:javascript|data:text\/html)[^']*'|(?:javascript|data:text\/html)[^\s>]*)/gi, "");
}
/** The svg root's DECLARED width in px, when there is one — the natural
* size the card displays at instead of stretching to full width (owner,
* 2026-10-07). %, em, "auto", or a missing attribute → null (the caller
@@ -122,13 +111,18 @@ export function VisualOutputCard({
<div
className={
natural !== null
? "w-fit max-w-full overflow-hidden rounded-lg bg-white [&_svg]:h-auto [&_svg]:max-w-full"
: "w-full overflow-hidden rounded-lg bg-white [&_svg]:h-auto [&_svg]:w-full"
? "w-fit max-w-full overflow-hidden rounded-lg bg-white"
: "w-full overflow-hidden rounded-lg bg-white"
}
style={{ filter: inverted ? "invert(1) hue-rotate(180deg)" : undefined }}
dangerouslySetInnerHTML={{ __html: sanitizeSvg(result.svg) }}
data-testid="visual-svg"
/>
>
<SvgImage
svg={result.svg}
width={natural ?? undefined}
className={natural !== null ? "block h-auto max-w-full" : "block h-auto w-full"}
/>
</div>
<div className="mt-2 flex items-center justify-between gap-2">
<div className="text-xs font-medium uppercase tracking-wide text-muted-foreground">Figure</div>
<div className="flex items-center gap-1.5">
components/chat.tsx
+198 -127
View File
@@ -5,7 +5,7 @@ import { useChatSession } from "@/components/session";
import { visibleQueuedSteers } from "@/components/session-core";
import { SetupDialog } from "@/components/setup-dialog";
import { Hint } from "@/components/ui/hint";
import { useZeroKeys, ZeroKeysPill } from "@/components/setup-cta";
import { useModelSetup, selectedModelUnavailable, openModelPicker, ZeroKeysPill } from "@/components/setup-cta";
import { useStickToBottomContext, type StickToBottomContext } from "use-stick-to-bottom";
import {
Conversation,
@@ -26,6 +26,8 @@ import { Composer } from "@/components/composer";
import { useEnterSubmits } from "@/components/enter-submits";
import { blurAfterPointerClick } from "@/components/focus-hygiene";
import { dockStateKey } from "@/components/dock-state-key";
import { clearOpeningPrompt, readOpeningPromptEntry, stashOpeningPrompt, OPENING_PROMPT_FAILURE } from "@/components/opening-prompt";
import { forwardDockWheel, useDockHeight } from "@/components/dock-height";
import { QueuedSteersBody, type QueuedSteerLine } from "@/components/queued-steers";
import { DictationBar, finishDictation, joinDictation, useDictation } from "@/components/dictation";
import { useFieldCycle } from "@/components/field-cycle";
@@ -38,7 +40,6 @@ import {
messageAnchorId,
textPartMessageId,
kbdEventForAction as kbdEventForChat,
needsContinuation,
type ChatInput,
type ChatSettings,
type LhMessage,
@@ -57,7 +58,7 @@ import { PracticeDock } from "@/components/practice-dock";
import { ThinkingCard, type PracticeThinkingData } from "@/components/practice-cards";
import { GraphContextCard, type GraphContextData } from "@/components/graph-context-card";
import { TopBar } from "@/components/top-bar";
import type { ArtifactPartProps, JobRecord, Phase3, QuestionAnswerData, TurnUsage, SessionPricing } from "@/src/core";
import type { ArtifactPartProps, JobRecord, Phase3, QuestionAnswerData, QuestionFormDraft, TurnUsage, SessionPricing } from "@/src/core";
import type { LhViewMessage } from "@/src/core";
import { PhaseChip, PhaseDivider } from "@/components/phase-divider";
import {
@@ -66,9 +67,11 @@ import {
type CompactionDividerData,
} from "@/components/compaction-divider";
import { CoverageRing } from "@/components/extraction-coverage";
import { Button } from "@/components/ui/button";
import {
activeProblemOf,
dismissedAnswer,
endsInFailedTurn,
findPendingToolCall,
phaseAt3,
phaseTimeline,
@@ -78,8 +81,6 @@ import {
} from "@/src/core";
import { ReadingThinkingField } from "@/components/reading-thinking";
import { currentUserThinkingOpen, setAllUserThinkingOpen } from "@/components/thinking-block";
import { slotsConfigured, type ModelsSettings } from "@/src/core";
import { useSettingsSaved } from "@/components/settings-signal";
/** One width for everything: transcript, cards, and the bottom blob share this column. */
/**
@@ -162,6 +163,18 @@ export function parseProviderKeyRefusal(message: string | undefined): boolean {
}
}
/** Retry sends no composer text. Its refusal must not resurrect the draft
* from an earlier, already-durable text send over today's unsent composer.
* Retire that fallback synchronously, never when the async Retry settles:
* a later ordinary send may own a new recovery draft by then. */
export function retryFailedResponse(
send: (input: ChatInput) => void,
submittedDraft: { current: string | null },
): void {
submittedDraft.current = null;
void send({ type: "retry" });
}
/** useLayoutEffect that is silent during SSR (client components still render on the server in Next). */
const useIsoLayoutEffect = typeof window === "undefined" ? useEffect : useLayoutEffect;
@@ -198,6 +211,7 @@ export function Chat({
chatId,
title,
initialMessages,
initialError,
initialPrompt,
initiallyInFlight,
initiallyCompacting = false,
@@ -207,6 +221,8 @@ export function Chat({
chatId: string;
title: string;
initialMessages: LhMessage[];
/** Memory-only settlement notice supplied on reopening. */
initialError?: string;
/** The session kind (practice sessions, owner 2026-10-04): practice
* renders no phase chip (no probe→teach arc applies) and a Practice
* tag rides the header instead. */
@@ -236,8 +252,8 @@ export function Chat({
// (core/view.ts projectView); sends are ONE explicit input event each
// ({chatId, input} — the stash indirection died with the SDK), and the
// turn stream is the VERBATIM AgentEvent passthrough the server emits.
const { messages, nativeMessages, contextTokens, send, steer, steerQueue, submitAnswer, dismissQuestion, dismissProblem, setMessages, adoptRecord, status, error, stop, clearError, resumeStream, setReattachSettled } =
useChatSession({ chatId, initialMessages, initiallyInFlight });
const { messages, nativeMessages, hasConfirmedConversation, contextTokens, send, continueIfOwed, steer, steerQueue, submitAnswer, dismissQuestion, dismissProblem, questionDraft, saveQuestionDraft, setMessages, adoptRecord, status, error, stop, clearError, resumeStream, setReattachSettled } =
useChatSession({ chatId, initialMessages, initialError, initiallyInFlight });
const sessionFindScopeRef = useRef<HTMLDivElement | null>(null);
const scrollContextRef = useRef<StickToBottomContext | null>(null);
@@ -246,16 +262,6 @@ export function Chat({
// The reattach settled terminal → the sync-and-continue seam (below).
setReattachSettled(() => reattachSettledRef.current());
// New chat with a topic: send the first message exactly once
// (initialPromptOwed — the pure gate, pinned in the test suite).
const sentInitial = useRef(false);
useEffect(() => {
if (!sentInitial.current && initialPromptOwed(initialMessages, initialPrompt)) {
sentInitial.current = true;
void send({ type: "text", text: initialPrompt! });
}
}, [initialPrompt, initialMessages, send]);
// A turn running server-side, attached to no live view of ours (navigate
// away mid-turn, or a wake turn discovered mid-session): the reattach
// stream is the primary channel (resumeStream — the same rendering path
@@ -434,9 +440,7 @@ export function Chat({
useEffect(() => {
if (mountContinuation.current) return;
mountContinuation.current = true;
if (!turnInFlight && status === "ready" && needsContinuation(initialMessages)) {
void send({ type: "continue" });
}
if (!turnInFlight) continueIfOwed(initialMessages);
// mount-only by design (turnInFlight read at initial value)
}, []);
@@ -458,7 +462,6 @@ export function Chat({
const busy = localStream || turnInFlight;
const busyRef = useRef(busy);
busyRef.current = busy;
// The attach handoff (stream-reattach wave): the moment a resumed
// stream takes over the rendering — the SDK left "ready", so
// submitted/streaming means chunks are flowing — localStream covers
@@ -532,10 +535,12 @@ export function Chat({
}, []);
const pendingQuestionId = pending?.toolCallId ?? null;
useEffect(() => {
// A new pending question never inherits the previous one's draft.
readingThinkingRef.current = "";
setReadingThinking("");
}, [pendingQuestionId]);
// Different calls never share a draft; a failed submission of THIS
// call restores the reading channel alongside the form's snapshot.
const reading = pendingQuestionId ? questionDraft(pendingQuestionId)?.readingThinking ?? "" : "";
readingThinkingRef.current = reading;
setReadingThinking(reading);
}, [pendingQuestionId, questionDraft]);
// The phase at every message (one linear fold): the transcript draws a
// color-coded dotted divider at each boundary; the header chip reads the
@@ -552,31 +557,79 @@ export function Chat({
textRef.current = v;
setText(v);
}, []);
// The BYOK setup flow (owner, 2026-09-22): the persistent CTA and the
// refusal interception share one SetupDialog mount. `configured` is
// seeded server-side (slotsConfigured on the record) and flips on Save.
const [configured, setConfigured] = useState(initialConfigured);
// The zero-keys truth (zero-keys wave): no provider key = the pill
// takes precedence over the models CTA (keys come first in the chain).
// Reads once on mount, re-reads on the "keys" signal.
const zeroKeys = useZeroKeys();
// Cross-surface freshness (fix wave): another surface's models write —
// setup save, composer pick, a settings write — re-reads the record so
// the setup CTA appears/disappears live (it used to lag until reload).
useSettingsSaved((section) => {
if (section !== "models") return;
fetch("/api/settings")
.then((r) => r.json() as Promise<{ settings?: { models?: ModelsSettings } | null } | null>)
.then((body) => setConfigured(slotsConfigured(body?.settings?.models)))
.catch(() => {});
});
// Saved slot completeness is live, but the session's selected route
// comes from its own ordinary model control, never today's Strong pin.
const { catalog, configured, zeroKeys } = useModelSetup(initialConfigured);
const catalogRef = useRef(catalog);
catalogRef.current = catalog;
const [selectedModel, setSelectedModel] = useState<string | undefined>();
const [setupOpen, setSetupOpen] = useState(false);
/** The last explicitly-sent draft (composer send): the
* refusal flow restores it into the composer, where it waits for the
* learner to re-send (Save never auto-sends, owner 2026-10-09).
* Auto-continuations (probe answers) never set it — their refusal
* opens the dialog with no replay. */
* opens recovery with no replay. */
const draftRef = useRef<string | null>(null);
const openingDraftRef = useRef<string | null>(null);
const [openingError, setOpeningError] = useState(false);
const openingMountedRef = useRef(false);
useEffect(() => {
openingMountedRef.current = true;
return () => { openingMountedRef.current = false; };
}, []);
const deliverOpeningPrompt = useCallback((prompt: string) => {
openingDraftRef.current = prompt;
// Persist uncertainty BEFORE the POST: another mount must restore,
// not race the still-arriving request. No storage means no auto-send.
if (!stashOpeningPrompt(chatId, prompt, true)) {
applyText(prompt);
setOpeningError(true);
return;
}
setOpeningError(false);
draftRef.current = prompt; // The existing 503 flow must restore it too.
applyText("");
void send({ type: "text", text: prompt }, () => {
clearOpeningPrompt(chatId);
openingDraftRef.current = null;
if (draftRef.current === prompt) draftRef.current = null;
}).then(accepted => {
if (!accepted && openingMountedRef.current && openingDraftRef.current === prompt && !textRef.current) {
applyText(prompt);
}
});
}, [chatId, send, applyText]);
const sentInitial = useRef(false);
useEffect(() => {
if (sentInitial.current) return;
sentInitial.current = true;
const entry = readOpeningPromptEntry(chatId);
const prompt = initialPrompt || entry?.text; // Legacy links still work.
if (!initialPromptOwed(initialMessages, prompt)) {
if (initialMessages.some(m => m.role === "user" || m.role === "assistant")) clearOpeningPrompt(chatId);
return;
}
openingDraftRef.current = prompt!;
if (entry?.attempted || initiallyInFlight) {
applyText(prompt!); // Interrupted start: the learner explicitly resends.
return;
}
deliverOpeningPrompt(prompt!);
}, [chatId, initialPrompt, initialMessages, initiallyInFlight, deliverOpeningPrompt, applyText]);
useEffect(() => {
if (!hasConfirmedConversation) return;
clearOpeningPrompt(chatId);
// A later canonical read may confirm a lost acknowledgment. Clear
// only the unchanged recovered draft, never the learner's newer work.
const opening = openingDraftRef.current;
if (opening !== null && textRef.current === opening) applyText("");
if (draftRef.current === opening) draftRef.current = null;
openingDraftRef.current = null;
setOpeningError(false);
}, [chatId, hasConfirmedConversation, applyText]);
// THE BYOK REFUSAL (owner, 2026-09-22): a send that lands the server's
// 503 models_not_configured envelope renders as the SETUP FLOW, never
@@ -591,8 +644,9 @@ export function Chat({
if (!error) return;
const missing = parseModelsRefusal(error.message);
// The key-refusal twin (zero-keys wave): a pinned model whose
// provider key is gone gets the SAME treatment — revert-draft and
// open setup (the dialog derives to its first-key stage on its own).
// provider key is gone also restores the draft. With another provider
// connected, repair belongs to the ordinary current-session picker;
// otherwise setup derives its exclusive provider/loading stage.
const keyRefused = parseProviderKeyRefusal(error.message);
if (!missing && !keyRefused) return; // a real error keeps the honest banner
// ORDER MATTERS: clearError() flips this effect's own dep (`error`)
@@ -612,7 +666,8 @@ export function Chat({
draftRef.current = null;
}
clearError();
setSetupOpen(true);
if (keyRefused && catalogRef.current?.providers.some((p) => p.keyPresent)) openModelPicker();
else setSetupOpen(true);
})();
return () => {
cancelled = true;
@@ -679,6 +734,12 @@ export function Chat({
// server's registry mutex is the structural gate; this is the
// affordance).
const [compacting, setCompacting] = useState(initiallyCompacting);
// A failed model response (issue #110): Retry keeps the saved input
// and completed work, without another learner message. Offered from
// "error" too: a dropped stream leaves the model there, and send()
// recovers from it. Later compaction requires new input instead.
const retryable = (status === "ready" || status === "error") && !busy && !compacting && endsInFailedTurn(nativeMessages);
const retryFailedTurn = () => retryFailedResponse(send, draftRef);
const canSend =
(text.trim().length > 0 || dictation.active) && !pending && !activeProblem && !compacting;
@@ -717,20 +778,20 @@ export function Chat({
const syncFromServer = useCallback(async () => {
const res = await fetch(`/api/chats/${chatId}`);
if (!res.ok) return;
const record = (await res.json()) as { messages?: LhMessage[]; turns?: TurnUsage[]; pricing?: SessionPricing };
const record = (await res.json()) as { messages?: LhMessage[]; turns?: TurnUsage[]; pricing?: SessionPricing; turnError?: string };
if (Array.isArray(record.turns)) setTurns(record.turns);
setPricing(record.pricing);
if (!Array.isArray(record.messages)) return;
setMessages(record.messages);
setMessages(record.messages, record.turnError);
// The continuation seam (contracts.ts protocol §4): a job result
// landing in the record leaves the agent owing a continuation — the
// record ENDS on un-consumed results, the dock owns no question, and
// nothing runs. The native fold (needsContinuation) is the old SDK
// lastAssistantMessageIsCompleteWithToolCalls check.
if (status === "ready" && needsContinuation(record.messages)) {
void send({ type: "continue" });
}
}, [chatId, setMessages, send, status]);
// Current model truth also guards retained/live failures: canonical
// sync must never automatically replace a reply that failed to save.
continueIfOwed(record.messages);
}, [chatId, setMessages, continueIfOwed]);
// The settled terminal's other half (stream-reattach wave): the
// transport resolved turn-settled — no turn running, the record is
@@ -905,8 +966,8 @@ export function Chat({
// (endTurn emits from onFinish before the response terminates, so the
// SSE frame reliably beats [DONE]) — the handler below cannot sync
// there, it flags instead, and THIS flip consumes the flag: the record
// is guaranteed durable (endTurn fires only after persistTurn, the S2
// ordering law), so replacing the view onto it absorbs the tail's
// reflects durable truth (a failed settlement retains its notice,
// not its unsaved reply), so replacing the view absorbs the tail's
// streamed c_-id copies into base under their server ids. Without
// this the tail kept those copies forever and the next send's
// steer-adoption window rendered them a second time. The wake handoff
@@ -1143,6 +1204,18 @@ export function Chat({
if (busy || pending || compacting) return;
const t = textRef.current.trim();
if (!t) return;
if (zeroKeys) {
setSetupOpen(true);
return;
}
if (selectedModelUnavailable(catalog, selectedModel)) {
openModelPicker();
return; // No optimistic send and no change to the learner's draft.
}
if (openingDraftRef.current !== null) {
deliverOpeningPrompt(t);
return;
}
draftRef.current = t; // the refusal flow's preserve-and-replay seam
applyText("");
void send({ type: "text", text: t });
@@ -1574,9 +1647,15 @@ export function Chat({
</Shimmer>
</span>
)}
{openingError && <p role="alert" className="text-sm text-destructive">{OPENING_PROMPT_FAILURE}</p>}
{error && (
<div className="rounded-lg border border-destructive/50 bg-destructive/10 px-3 py-2 text-sm">
Something went wrong: {error.message}
<div className="flex items-center justify-between gap-3 rounded-lg border border-destructive/50 bg-destructive/10 px-3 py-2 text-sm">
<span>Something went wrong: {error.message}</span>
{retryable && (
<Button variant="outline" size="xs" data-testid="retry-failed-turn" onClick={retryFailedTurn}>
Retry
</Button>
)}
</div>
)}
</div>
@@ -1596,10 +1675,13 @@ export function Chat({
fold) and NEVER overlaps the composer. */}
<DockOverlay
pending={pending}
initialQuestionDraft={pending ? questionDraft(pending.toolCallId)?.form : undefined}
onSaveQuestionDraft={(form) => { if (pending) saveQuestionDraft(pending.toolCallId, { form }); }}
practice={activeProblem}
compactionRunning={compacting}
onSubmitThinking={submitThinking}
chatId={chatId}
onModelChange={setSelectedModel}
plan={hasPlan ? <PlanMarginalia messages={messages} /> : undefined}
onOpenPlan={() => setPlanOpen(true)}
requiresReveal={requiresReveal}
@@ -1663,8 +1745,10 @@ export function Chat({
// ride the same callback (QuestionAnswerData.dismissed). The
// reading-thinking draft rides along (owner 2026-10-03) — the
// check answer is its only carrier.
saveQuestionDraft(pending.toolCallId, { readingThinking: readingThinkingRef.current });
const reading = readingThinkingRef.current.trim() || undefined;
clearReadingThinking();
// The pending-id effect clears this only when the question
// actually leaves. An offline preflight keeps the draft visible.
submitAnswer(pending.tool, pending.toolCallId, reading ? { ...output, readingThinking: reading } : output);
}}
/>
@@ -1701,7 +1785,6 @@ export function Chat({
onClose={() => setSetupOpen(false)}
onDone={() => {
setSetupOpen(false);
setConfigured(true);
// Save only saves (owner ruling 2026-10-09): the draft went
// home into the composer at refusal time and STAYS there —
// the learner sends it themselves.
@@ -1752,7 +1835,8 @@ export function Chat({
* 2026-10-02 abort chain proper (question before turn). */
export type EscAbortTarget = "mermaid" | "plan" | "setup" | "question" | "problem" | "turn" | null;
/**
* Does the ?prompt topic still owe its first send? The gate is "no real
* Does the opening topic still owe its first send? The topic is the
* server's ?prompt= or the sessionStorage stash. The gate is "no real
* conversation yet" (no user/assistant message) — NOT emptiness: a
* session started with map picks carries a graph-context MARKER as
* message[0], and that scaffolding must never swallow the first send
@@ -1821,6 +1905,8 @@ export function shouldStampGeometryMute(nextH: number, prevH: number, pending: b
function DockOverlay({
pending,
initialQuestionDraft,
onSaveQuestionDraft,
practice,
compactionRunning,
onSubmitThinking,
@@ -1834,6 +1920,7 @@ function DockOverlay({
composer,
onSubmitAnswer,
chatId,
onModelChange,
onReserve,
plan,
onOpenPlan,
@@ -1843,6 +1930,8 @@ function DockOverlay({
queuedSteers,
}: {
pending: ReturnType<typeof findPendingToolCall>;
initialQuestionDraft?: QuestionFormDraft;
onSaveQuestionDraft: (draft: QuestionFormDraft) => void;
/** Compaction running for this chat (round 14) — the docks' submit
* affordance gate (the server's registry is the real one). */
compactionRunning: boolean;
@@ -1868,6 +1957,7 @@ function DockOverlay({
onSubmitAnswer: (output: QuestionAnswerData) => void;
/** The model control's pass-through (the stats bar hosts it). */
chatId: string;
onModelChange: (model: string | undefined) => void;
/** Reports the dock's settled height (block + gradient) — ChatView's
* dockReserve state feeds the transcript's spacer from it. */
onReserve: (h: number) => void;
@@ -1962,6 +2052,17 @@ function DockOverlay({
const revealedRef = useRef(revealed);
revealedRef.current = revealed;
const blockRef = useRef<HTMLDivElement>(null);
const viewportRef = useRef<HTMLDivElement>(null);
const bodyRef = useRef<HTMLDivElement>(null);
const shellRef = useRef<HTMLDivElement>(null);
const contentRef = useRef<HTMLDivElement>(null);
const mode = pending
? `question:${pending.toolCallId}`
: practice
? `practice:${practice.toolCallId}`
: "chat";
// 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
@@ -2000,28 +2101,33 @@ function DockOverlay({
return () => ro.disconnect();
}, []);
// The reserve grows immediately, but never shrinks while a question or
// practice problem owns the dock: shrinking flow space could pin a nearby
// learner back to the bottom. Composer mode may give space back.
const reserveRef = useRef(0);
const offer = useCallback((blockH: number) => {
const next = blockH + GRADIENT_H;
if (next <= reserveRef.current + 0.5 && pendingRef.current) return;
// Stamp growth at the source, including child-only field resizes, so
// stick-to-bottom's catch-up is not mistaken for the learner leaving.
if (shouldStampGeometryMute(next, reserveRef.current, pendingRef.current)) geometryAtRef.current = Date.now();
reserveRef.current = next;
onReserve(next);
}, [onReserve]);
// 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.
// The same measure owns the QUESTION/PROBLEM HEIGHT CAP (owner round,
// 2026-10-05): the old fixed fractions (55dvh / 40vh) cut long problems
// off absurdly short — the cap is now the AVAILABLE COLUMN (the
// scroller already starts below the top bar) minus a 2rem breathing
// gap, carried as --dock-max-h for the docks' scroll areas.
// Budget the whole solid dock, leaving only its question area flexible.
// The hook measures natural, visible content before the morph/reserve
// effect below. Short windows scroll the reading/question/answer body;
// the stats/model toolbar remains outside that overflow viewport.
const [sbInset, setSbInset] = useState(0);
const [dockMaxH, setDockMaxH] = useState<number | null>(null);
useEffect(() => {
const el = scrollRef.current;
if (!el) return;
const measure = () => {
setSbInset(Math.round((el.offsetWidth - el.clientWidth) / 2));
setDockMaxH(Math.max(el.clientHeight - 32, 160));
};
measure();
window.addEventListener("resize", measure);
return () => window.removeEventListener("resize", measure);
}, [scrollRef]);
useDockHeight({
scrollRef, blockRef, viewportRef, bodyRef, shellRef, contentRef,
expanded, enabled: !!pending || !!practice, mode, onScrollbarInset: setSbInset, onHeight: offer,
});
// GROWTH MUTE (owner round, 2026-10-05 — the generation churn): during a
// live turn the transcript content grows and stick-to-bottom chases it;
@@ -2057,48 +2163,11 @@ function DockOverlay({
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);
};
const onWheel = (e: WheelEvent) => forwardDockWheel(e, root, el);
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
// Reserve GROWTH is the phantom-scroll SOURCE (flicker round,
// 2026-09-26): the spacer's height lands in this commit, the smooth
// catch-up lags it by the growth amount, and the position machine
// used to read that lag as the learner leaving — form → bar → form
// (the dictation flush, reproduced d≈152 > COLLAPSE_DIST). Stamp the
// mute right where the spacer grows: it covers the scroll events of
// THIS growth no matter how long after the field stretched (the
// field grows on its own render; the offer lands on the NEXT
// DockOverlay render — ~1s apart live). Shrinks never stamp (the
// fold's own give-back must not self-mute the re-expand).
if (shouldStampGeometryMute(next, reserveRef.current, pendingRef.current)) geometryAtRef.current = Date.now();
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
@@ -2200,11 +2269,6 @@ function DockOverlay({
if (b) offer(b.getBoundingClientRect().height);
};
const mode = pending
? `question:${pending.toolCallId}`
: practice
? `practice:${practice.toolCallId}`
: "chat";
// The opening variant of a session appears instantly — the fade/rise is
// reserved for mid-session transitions.
const initialModeRef = useRef(mode);
@@ -2212,11 +2276,9 @@ function DockOverlay({
// Height morph: animate the shell from its previous height to its new one
// whenever the variant (chat ↔ question ↔ collapsed bar) changes. Typing
// growth keeps the same stateKey and is NOT animated.
const shellRef = useRef<HTMLDivElement>(null);
// The fold's CONTENT fade (owner round, 2026-10-05): the shell height
// morphs, but the bar↔form content used to hard-swap. A WAAPI fade/rise
// on the toggle — no remount, so the question form's state survives.
const contentRef = useRef<HTMLDivElement>(null);
const prevExpandedForFadeRef = useRef(expanded);
useIsoLayoutEffect(() => {
if (prevExpandedForFadeRef.current === expanded) return;
@@ -2264,8 +2326,10 @@ function DockOverlay({
return (
<div className="pointer-events-none absolute inset-x-0 bottom-0" style={{ right: sbInset }}>
<div className="h-24 bg-gradient-to-t from-background via-background/85 to-transparent" />
<div ref={blockRef} className="relative bg-background pb-5" style={{ "--dock-max-h": dockMaxH ? `${dockMaxH}px` : undefined } as React.CSSProperties}>
<div ref={blockRef} className="relative bg-background pb-5">
<div className={DOCK}>
<div ref={viewportRef} data-dock-viewport="" className="pointer-events-auto">
<div ref={bodyRef}>
{setupCta}
{/* The queued-steer indicator (owner spec 2026-10-03): quiet
one-liners above the shell - right-aligned to where its corner
@@ -2294,11 +2358,13 @@ function DockOverlay({
/>
)}
<div ref={shellRef} className="pointer-events-auto overflow-hidden rounded-2xl border bg-card shadow-lg">
{/* Include descendant margins (notably practice thinking's
bottom margin) in the natural content measurement. */}
<div
ref={contentRef}
key={mode}
className={mode === initialModeRef.current ? undefined : "dock-variant"}
>
ref={contentRef}
key={mode}
className={mode === initialModeRef.current ? "flow-root" : "flow-root dock-variant"}
>
{pending ? (
<QuestionDock
toolName={pending.tool as "quiz" | "probe" | "ask_user_question" | "explain" | "recall"}
@@ -2309,6 +2375,8 @@ function DockOverlay({
phase={phase}
onExpand={expandAndReveal}
onSubmit={onSubmitAnswer}
initialDraft={initialQuestionDraft}
onSaveDraft={onSaveQuestionDraft}
onEscape={dictationCancel}
dictationClaim={!revealPending}
compactionRunning={compactionRunning}
@@ -2331,6 +2399,8 @@ function DockOverlay({
)}
</div>
</div>
</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. */}
@@ -2340,6 +2410,7 @@ function DockOverlay({
pricing={pricing}
dictationSlot={dictationSlot}
chatId={chatId}
onModelChange={onModelChange}
/>
{/* The scroll-down disc (moved here from the Conversation,
owner refinement round 2026-09-24): anchored to the BLOCK's
components/dictation-language.tsx
+108
View File
@@ -0,0 +1,108 @@
"use client";
import { useEffect, useRef, useState, type KeyboardEvent as ReactKeyboardEvent } from "react";
import { DICTATION_LANGUAGES, type DictationLanguage } from "@/src/core";
import { MenuShell } from "@/components/menu-shell";
import { Hint } from "@/components/ui/hint";
import { SettingsMenuTrigger, SettingsRow } from "@/components/settings/controls";
import { SearchMenuList } from "@/components/settings/search-menu-list";
import { useDictationActive } from "@/components/dictation-state";
import { useDictationLanguage } from "@/components/dictation-preference";
type LanguagePreference = ReturnType<typeof useDictationLanguage>;
export function DictationLanguageSetting() {
const preference = useDictationLanguage();
const active = useDictationActive();
return <SettingsRow label="Language" testid="row-dictation-language" dimmed={active}>
<DictationLanguagePicker preference={preference} disabled={active} />
</SettingsRow>;
}
function DictationLanguagePicker({ preference, disabled }: {
preference: LanguagePreference;
disabled: boolean;
}) {
const { language, saving, error, load, save } = preference;
const [open, setOpen] = useState(false);
const triggerRef = useRef<HTMLButtonElement>(null);
const label = DICTATION_LANGUAGES.find((item) => item.code === language)?.label ?? "—";
useEffect(() => { if (disabled) setOpen(false); }, [disabled]);
useEffect(() => {
if (!open) return;
const onPointerDown = (event: PointerEvent) => {
const target = event.target as Element;
if (target instanceof Element && target.closest('[data-testid="dictation-language-menu"]')) return;
if (!triggerRef.current?.contains(target)) setOpen(false);
};
document.addEventListener("pointerdown", onPointerDown);
return () => document.removeEventListener("pointerdown", onPointerDown);
}, [open]);
const closeOnEscape = (event: ReactKeyboardEvent) => {
if (event.key !== "Escape") return;
event.preventDefault();
event.stopPropagation();
setOpen(false);
triggerRef.current?.focus();
};
const pick = async (code: DictationLanguage) => {
if (disabled || saving || language === undefined) return;
if (await save(code)) {
setOpen(false);
triggerRef.current?.focus();
}
};
return <>
<Hint label={error ?? `Dictation language: ${label}`} side="top">
<SettingsMenuTrigger
ref={triggerRef}
value={label}
data-testid="dictation-language"
aria-label={`Dictation language: ${label}`}
aria-expanded={open && !disabled}
disabled={disabled || saving}
onKeyDown={(event) => { if (open) closeOnEscape(event); }}
onClick={() => {
setOpen((value) => !value);
if (language === undefined) void load().catch(() => {});
}}
className="min-w-20 max-w-[65%]"
/>
</Hint>
{open && !disabled && <MenuShell
anchorRef={triggerRef}
testid="dictation-language-menu"
preferredSide="down"
menuClass="flex w-72 max-w-[calc(100vw-1rem)] flex-col rounded-xl border border-border bg-card p-2 text-xs shadow-xl"
>
<div onKeyDown={(event) => {
closeOnEscape(event);
if (event.key === "Enter") {
// Empty search results must not submit the composer, but a
// focused option button still needs Enter's native click.
if (event.target instanceof HTMLInputElement) event.preventDefault();
event.stopPropagation();
}
}}>
{error && <div role="alert" className="mb-1.5 text-destructive">{error}</div>}
<SearchMenuList
items={DICTATION_LANGUAGES}
itemKey={(item) => item.code}
searchTerms={(item) => [item.code, item.label]}
isSelected={(item) => item.code === language}
onPick={(item) => { void pick(item.code); }}
disabled={saving || language === undefined}
searchTestid="dictation-language-search"
listTestid="dictation-language-list"
optionTestid={(item) => `dictation-language-${item.code}`}
ariaLabel="Search dictation languages"
placeholder="Search languages…"
empty="No languages match."
renderItem={(item) => <span className="text-card-foreground">{item.label}</span>}
/>
</div>
</MenuShell>}
</>;
}
components/dictation-preference.ts
+132
View File
@@ -0,0 +1,132 @@
"use client";
import { useSyncExternalStore } from "react";
import {
FACTORY_SETTINGS, isDictationLanguage, sanitizeUserSettings,
type DictationLanguage, type UserSettings,
} from "@/src/core";
import { notifySettingsSaved, onSettingsSaved } from "@/components/settings-signal";
const INITIAL = { language: undefined as DictationLanguage | undefined, saving: false, error: null as string | null };
/** One confirmed preference and one write in flight, shared by every surface.
* The factory keeps transport tests isolated without a second persistence path.
*/
export function createDictationLanguageStore(request: typeof fetch, saved: () => void = () => {}) {
let snapshot = INITIAL;
let reading: Promise<DictationLanguage> | null = null;
let writing: Promise<boolean> | null = null;
let revision = 0;
let needsRead = true;
const listeners = new Set<() => void>();
const publish = (patch: Partial<typeof INITIAL>) => {
snapshot = { ...snapshot, ...patch };
for (const listener of listeners) listener();
};
function load(refresh = false): Promise<DictationLanguage> {
if (writing) return getLanguage();
if (reading && !refresh) return reading;
if (!refresh && !needsRead && snapshot.language !== undefined) return Promise.resolve(snapshot.language);
const version = ++revision;
needsRead = true;
reading = Promise.resolve().then(async () => {
try {
const res = await request("/api/settings");
if (!res.ok) throw new Error();
const body = await res.json() as { settings: UserSettings | null };
if (!("settings" in body)) throw new Error();
const record = sanitizeUserSettings(body.settings);
if (body.settings !== null && record === null) throw new Error();
if (version !== revision) return getLanguage();
const language = record?.dictation?.language ?? FACTORY_SETTINGS.dictation.language;
needsRead = false;
publish({ language, error: null });
return language;
} catch {
if (version !== revision) return getLanguage();
const message = "Could not load dictation language. Try again.";
publish({ error: message });
throw new Error(message);
} finally {
if (version === revision) reading = null;
}
});
return reading;
}
function save(language: DictationLanguage): Promise<boolean> {
// A different selection must never inherit the previous write's success.
if (writing || snapshot.language === undefined) return Promise.resolve(false);
revision++;
reading = null;
writing = Promise.resolve().then(async () => {
try {
const res = await request("/api/settings/dictation", {
method: "PUT", headers: { "Content-Type": "application/json" },
body: JSON.stringify({ language }),
});
if (!res.ok) throw new Error();
const body = await res.json() as { settings: UserSettings };
const confirmed = body.settings?.dictation?.language;
if (!isDictationLanguage(confirmed)) throw new Error();
needsRead = false;
publish({ language: confirmed, error: null });
saved();
return true;
} catch {
publish({ error: "Could not save dictation language. Try again." });
return false;
} finally {
writing = null;
publish({ saving: false });
}
});
publish({ saving: true, error: null });
return writing;
}
async function getLanguage(): Promise<DictationLanguage> {
for (;;) {
if (writing) await writing;
else if (reading) await reading;
else if (needsRead || snapshot.language === undefined) await load();
else return snapshot.language;
}
}
return {
load, save, getLanguage,
getSnapshot: () => snapshot,
getServerSnapshot: () => INITIAL,
subscribe(listener: () => void) { listeners.add(listener); return () => { listeners.delete(listener); }; },
};
}
let announcingSave = false;
const store = createDictationLanguageStore((...args) => fetch(...args), () => {
// All our readers already received the confirmed PUT response. Keep the
// repo's notification contract without depending on a redundant GET.
announcingSave = true;
try { notifySettingsSaved("dictation"); } finally { announcingSave = false; }
});
let readers = 0;
let stopSignals: (() => void) | undefined;
function subscribe(listener: () => void) {
const off = store.subscribe(listener);
if (readers++ === 0) {
stopSignals = onSettingsSaved((section) => {
if (section === "dictation" && !announcingSave) void store.load(true).catch(() => {});
});
// A confirmed preference lives for the window, not the page. Navigation
// and reopening Settings must not add a new network failure point.
void store.getLanguage().catch(() => {});
}
return () => { off(); if (--readers === 0) { stopSignals?.(); stopSignals = undefined; } };
}
export function useDictationLanguage() {
const snapshot = useSyncExternalStore(subscribe, store.getSnapshot, store.getServerSnapshot);
return { ...snapshot, load: store.load, save: store.save, getLanguage: store.getLanguage };
}
components/dictation-state.ts
+25
View File
@@ -0,0 +1,25 @@
"use client";
import { useSyncExternalStore } from "react";
// The keyboard chain needs a synchronous answer; settings also subscribe
// so the language control locks from connecting until cleanup completes.
let active = false;
const listeners = new Set<() => void>();
export function isDictationActive(): boolean { return active; }
export function setDictationActive(next: boolean): void {
if (active === next) return;
active = next;
for (const listener of listeners) listener();
}
function subscribe(listener: () => void): () => void {
listeners.add(listener);
return () => { listeners.delete(listener); };
}
export function useDictationActive(): boolean {
return useSyncExternalStore(subscribe, isDictationActive, () => false);
}
components/dictation.tsx
+40 -55
View File
@@ -4,6 +4,10 @@ import { useCallback, useEffect, useRef, useState } from "react";
import { Loader2Icon, MicIcon } from "lucide-react";
import { useSettingsDialog } from "@/components/settings/settings-dialog";
import { Hint } from "@/components/ui/hint";
import { deepgramListenUrl, type DictationLanguage } from "@/src/core";
import { useDictationLanguage } from "@/components/dictation-preference";
import { setDictationActive } from "@/components/dictation-state";
export { isDictationActive } from "@/components/dictation-state";
/**
* Voice dictation — the cloud engine (STT wave 1, 2026-09-17).
@@ -86,15 +90,7 @@ import { Hint } from "@/components/ui/hint";
* endpointing=300, no interims. Silence streams too (the worklet never
* pauses), so the socket never idles — no KeepAlive needed.
*/
const DG_URL =
"wss://api.deepgram.com/v1/listen" +
"?model=nova-3" +
"&encoding=linear16" +
"&sample_rate=16000" +
"&channels=1" +
"&smart_format=true" +
"&punctuate=true" +
"&endpointing=300";
// URL construction lives in the browser-safe core; only language varies.
/** Meter tuning — the owner's tested values, see the file header. */
const METER_FLOOR_DB = -50;
@@ -204,16 +200,6 @@ export function rmsToMeter(rms: number): number {
*/
let questionTarget: ((text: string) => void) | null = null;
/* Is a dictation run live on this page? The Esc chain (keyboard wave A,
* components/keyboard-shortcuts.tsx) needs a SYNCHRONOUS answer — the
* stop handler can't report back through a CustomEvent. Set at start(),
* cleared in cleanup() (every terminal path runs through cleanup). */
let dictationActiveFlag = false;
export function isDictationActive(): boolean {
return dictationActiveFlag;
}
/* The finisher registry: submit paths FAR from the hook's owner (the
* pending question's dock) need finish() without a prop bridge. Same
* module-level pattern as questionTarget — one window, one instance,
@@ -285,6 +271,7 @@ export type Dictation = {
* question waits.
*/
export function useDictation({ append }: { append: (text: string) => void }): Dictation {
const { getLanguage } = useDictationLanguage();
const [phase, setPhase] = useState<DictationPhase>("idle");
const [error, setError] = useState<string | null>(null);
/** Latest RMS from the worklet — the meter samples this (dictate pattern). */
@@ -327,9 +314,9 @@ export function useDictation({ append }: { append: (text: string) => void }): Di
}, []);
/** Full teardown of one run's machinery. Idempotent; generation-guarded by callers. */
const cleanup = useCallback(() => {
const cleanup = useCallback((warmNext = true) => {
genRef.current++;
dictationActiveFlag = false;
setDictationActive(false);
flush(); // the deferred transcript appears when the run ends (owner round 2)
if (safetyRef.current) {
clearTimeout(safetyRef.current);
@@ -357,16 +344,16 @@ export function useDictation({ append }: { append: (text: string) => void }): Di
for (const w of waitersRef.current.splice(0)) w();
// Pre-mint for the next run (mitigation D) — best-effort; a keyless
// page just doesn't warm.
warmTokens.refresh();
if (warmNext) warmTokens.refresh();
}, []);
/** Open the Deepgram socket and wire its messages. `gen` guards staleness. */
const openSocket = useCallback(
(gen: number, token: string) => {
(gen: number, token: string, language: DictationLanguage) => {
// Minted JWTs authenticate via the "bearer" subprotocol — ["token",
// jwt] is rejected at the handshake (verified live; "token" is for
// raw API keys only, as in the owner's pi dictate extension).
const ws = new WebSocket(DG_URL, ["bearer", token]);
const ws = new WebSocket(deepgramListenUrl(language), ["bearer", token]);
wsRef.current = ws;
ws.addEventListener("open", () => {
@@ -408,7 +395,7 @@ export function useDictation({ append }: { append: (text: string) => void }): Di
reconnectsRef.current++;
mintFreshToken() // always FRESH — a stale warm token is the likely culprit
.then((t) => {
if (gen === genRef.current) openSocket(gen, t);
if (gen === genRef.current) openSocket(gen, t, language);
})
.catch(() => {
if (gen === genRef.current) {
@@ -426,13 +413,23 @@ export function useDictation({ append }: { append: (text: string) => void }): Di
);
const start = useCallback(async () => {
dictationActiveFlag = true;
setDictationActive(true);
setError(null);
setPhase("connecting");
reconnectsRef.current = 0;
bufferRef.current = []; // fresh run, fresh buffer (round 2)
pendingRef.current = []; // fresh run, fresh queue (mitigation A)
const gen = ++genRef.current;
let language: DictationLanguage;
try {
language = await getLanguage();
} catch (e) {
if (gen !== genRef.current) return;
setError(e instanceof Error ? e.message : "Dictation language unavailable");
cleanup();
return;
}
if (gen !== genRef.current) return;
// Mitigation D: the warm token (pre-minted at mount / after the last
// run) resolves ~instantly; a cold/keyless mint surfaces below. The
@@ -509,8 +506,8 @@ export function useDictation({ append }: { append: (text: string) => void }): Di
if (gen !== genRef.current) return;
// Mitigation B: the handshake fires the moment the graph is live,
// overlapping any remaining token wait — mic-gated, never before it.
openSocket(gen, token);
}, [openSocket, cleanup]);
openSocket(gen, token, language);
}, [openSocket, cleanup, getLanguage]);
const stop = useCallback(() => {
const gen = genRef.current;
@@ -586,7 +583,13 @@ export function useDictation({ append }: { append: (text: string) => void }): Di
// attached.
useEffect(() => {
warmTokens.refresh(); // pre-mint at mount (mitigation D) — silent if keyless
}, []);
return () => {
// Navigation cancels capture: invalidate every pending await before
// releasing resources, and never append into a different page's input.
bufferRef.current = [];
cleanup(false);
};
}, [cleanup]);
useEffect(() => {
window.addEventListener("lh:kbd-toggle-dictation", toggle);
return () => window.removeEventListener("lh:kbd-toggle-dictation", toggle);
@@ -630,7 +633,6 @@ export function DictationBar({ dictation }: { dictation: Dictation }) {
* error keeps the old behavior (click = retry, which clears). */
const keyless = error === DICTATION_KEYLESS;
const [levels, setLevels] = useState<number[]>(() => new Array(METER_CELLS).fill(0));
const recording = phase === "recording" || phase === "stopping";
/** Adaptive cell count (owner round 2): derived from the meter's
* measured width — narrow rows show fewer bars instead of
* overflowing under neighbors. */
@@ -674,10 +676,7 @@ export function DictationBar({ dictation }: { dictation: Dictation }) {
if (!active) {
return (
<Hint
// Errored: the button span truncates the line — the tooltip is its
// full reveal (the keyless line KEEPS "Deepgram" — owner overrule:
// it's the Keys row label, the truth the user needs). Otherwise the
// quiet label + the live ⌘D binding.
// The error line truncates in the button; its tooltip reveals it.
label={error ?? "Dictate"}
action={error ? undefined : "toggle-dictation"}
side="top"
@@ -694,10 +693,8 @@ export function DictationBar({ dictation }: { dictation: Dictation }) {
}}
aria-label={error ? `Dictation failed: ${error}` : "Start dictation"}
data-testid="dictation-toggle"
// Same blurring as the floating chrome (owner, fix wave round 3),
// but lower opacity — the mic stays quieter than the model pills.
className="flex h-7 w-full items-center justify-center gap-1.5 rounded-md border border-border/50 bg-card/50 text-xs text-muted-foreground/70 shadow-lg backdrop-blur transition-colors hover:border-border hover:text-muted-foreground"
>
className="flex h-7 w-full items-center justify-center gap-1.5 rounded-md border border-border/50 bg-card/50 text-xs text-muted-foreground/70 shadow-lg backdrop-blur transition-colors hover:border-border hover:text-muted-foreground"
>
<MicIcon className="size-3.5" />
<span className="max-w-full truncate">{error ?? "Dictate"}</span>
</button>
@@ -707,32 +704,20 @@ export function DictationBar({ dictation }: { dictation: Dictation }) {
return (
<div className="flex w-full min-w-0 items-center gap-2" data-testid="dictation-bar">
<Hint
// The owner flagged this copy good ("Stop dictation (Esc)") —
// kept verbatim; only the machinery moved to the styled tooltip.
label="Stop dictation (Esc)"
side="top"
>
<Hint label="Stop dictation (Esc)" side="top">
<button
type="button"
onClick={toggle}
disabled={phase === "stopping"}
aria-label="Stop dictation"
data-testid="dictation-toggle"
className="flex h-7 shrink-0 items-center gap-1.5 rounded-md bg-destructive/10 px-2 text-xs font-medium text-destructive transition-colors hover:bg-destructive/20"
>
{phase === "stopping" ? (
<Loader2Icon className="size-3.5 animate-spin" />
) : (
<MicIcon className="size-3.5" />
)}
data-testid="dictation-toggle"
className="flex h-7 shrink-0 items-center gap-1.5 rounded-md bg-destructive/10 px-2 text-xs font-medium text-destructive transition-colors hover:bg-destructive/20"
>
{phase === "stopping" ? <Loader2Icon className="size-3.5 animate-spin" /> : <MicIcon className="size-3.5" />}
Stop
</button>
</Hint>
<span className="size-1.5 shrink-0 animate-pulse rounded-full bg-destructive" aria-hidden />
{/* Fixed width while recording — nothing ever appends text beside it
(no interims, owner round 1). Cells carry no minimum width; the
COUNT adapts to the measured row space (owner round 2). */}
<div ref={meterRef} className="flex h-4 min-w-0 flex-1 items-center gap-[2px]" data-testid="dictation-meter">
{levels.slice(-cellCount).map((v, i) => (
<span
components/dock-height.ts
+157
View File
@@ -0,0 +1,157 @@
"use client";
import { useCallback, useEffect, useLayoutEffect, type RefObject } from "react";
const useIsoLayoutEffect = typeof window === "undefined" ? useEffect : useLayoutEffect;
export interface DockHeightMeasurements {
columnHeight: number;
/** Stats/toolbar and block padding, outside the scrolling viewport. */
outsideHeight: number;
/** Everything in the viewport except the animated shell. */
beforeShellHeight: number;
shellBorderHeight: number;
/** Natural content height, not the shell's animated/clamped height. */
contentHeight: number;
/** Null when collapsed, in chat mode, or showing malformed input. */
questionHeight: number | null;
questionPaddingHeight: number;
}
export function dockHeightBudget(m: DockHeightMeasurements) {
const available = Math.max(m.columnHeight - 32, 0);
const viewportMaxHeight = Math.max(available - m.outsideHeight, 0);
const shellMaxHeight = Math.max(viewportMaxHeight - m.beforeShellHeight, 0);
const naturalHeight = m.beforeShellHeight + m.shellBorderHeight + m.contentHeight;
const questionMaxHeight = m.questionHeight === null
? null
: Math.max(viewportMaxHeight - (naturalHeight - m.questionHeight), 0);
// A cap smaller than the question's padding leaves no content viewport.
// Deliberately uncap the question and scroll the whole content instead;
// do not confuse a measured zero with an unmeasured CSS fallback.
const scrollWholeContent = questionMaxHeight === null
? naturalHeight > viewportMaxHeight
: questionMaxHeight <= m.questionPaddingHeight;
return { viewportMaxHeight, shellMaxHeight, questionMaxHeight, scrollWholeContent };
}
interface DockHeightElements {
column: HTMLElement;
block: HTMLElement;
viewport: HTMLElement;
body: HTMLElement;
shell: HTMLElement;
content: HTMLElement;
}
const pixels = (value: string) => Number.parseFloat(value) || 0;
/** Read natural geometry before writing any constraints. Cancel the shell's
* animated height out of both ancestor measurements, then substitute its
* natural content and borders. This stays valid during a WAAPI height morph
* and when the outer viewport is already capped. */
export function syncDockHeight(elements: DockHeightElements, expanded: boolean) {
const { column, block, viewport, body, shell, content } = elements;
// QuestionDock keeps a laid-out form in an invisible h-0 wrapper;
// PracticeDock removes its form. Neither hidden form is a height input.
const question = expanded ? content.querySelector<HTMLElement>("[data-dock-scroll-area]") : null;
const height = (el: HTMLElement) => el.getBoundingClientRect().height;
const shellStyle = getComputedStyle(shell);
const questionStyle = question ? getComputedStyle(question) : null;
const budget = dockHeightBudget({
columnHeight: column.clientHeight,
outsideHeight: Math.max(height(block) - height(viewport), 0),
beforeShellHeight: Math.max(height(body) - height(shell), 0),
shellBorderHeight: pixels(shellStyle.borderTopWidth) + pixels(shellStyle.borderBottomWidth),
contentHeight: height(content),
questionHeight: question ? height(question) : null,
questionPaddingHeight: questionStyle
? pixels(questionStyle.paddingTop) + pixels(questionStyle.paddingBottom)
: 0,
});
if (budget.questionMaxHeight === null) block.style.removeProperty("--dock-max-h");
else block.style.setProperty("--dock-max-h", budget.scrollWholeContent ? "none" : `${budget.questionMaxHeight}px`);
viewport.style.maxHeight = `${budget.viewportMaxHeight}px`;
// Overflow is visible in normal layouts: keep the shell's shadow intact.
// The stats/model toolbar is a sibling, never inside this overflow box.
viewport.style.overflowY = budget.scrollWholeContent ? "auto" : "visible";
// Clamp animated starting heights too, so a previous tall variant cannot
// paint above the column while morphing into a constrained layout.
shell.style.maxHeight = budget.scrollWholeContent ? "none" : `${budget.shellMaxHeight}px`;
return budget;
}
interface DockHeightRefs {
scrollRef: RefObject<HTMLElement | null>;
blockRef: RefObject<HTMLDivElement | null>;
viewportRef: RefObject<HTMLDivElement | null>;
bodyRef: RefObject<HTMLDivElement | null>;
shellRef: RefObject<HTMLDivElement | null>;
contentRef: RefObject<HTMLDivElement | null>;
}
export function useDockHeight({
scrollRef, blockRef, viewportRef, bodyRef, shellRef, contentRef,
expanded, enabled, mode, onScrollbarInset, onHeight,
}: DockHeightRefs & {
expanded: boolean;
enabled: boolean;
/** Reconnect observers when the keyed content node/scroll area changes. */
mode: string;
onScrollbarInset: (inset: number) => void;
/** Settled visible block height, fed through the existing reserve policy. */
onHeight: (height: number) => void;
}) {
const measure = useCallback(() => {
const column = scrollRef.current;
const block = blockRef.current;
const viewport = viewportRef.current;
const body = bodyRef.current;
const shell = shellRef.current;
const content = contentRef.current;
if (!column || !block || !viewport || !body || !shell || !content) return;
onScrollbarInset(Math.round((column.offsetWidth - column.clientWidth) / 2));
if (enabled) syncDockHeight({ column, block, viewport, body, shell, content }, expanded);
else {
block.style.removeProperty("--dock-max-h");
viewport.style.removeProperty("max-height");
viewport.style.removeProperty("overflow-y");
shell.style.removeProperty("max-height");
}
// ResizeObserver runs even when only a child field changed and React
// did not rerender DockOverlay. Keep its transcript reserve current,
// but never report a mid-morph height.
if (shell.getAnimations().length === 0) onHeight(block.getBoundingClientRect().height);
}, [scrollRef, blockRef, viewportRef, bodyRef, shellRef, contentRef, expanded, enabled, onScrollbarInset, onHeight]);
// Imperative constraints land before the overlay's later layout effect
// measures its morph target/reserve, and before paint (including reveal).
useIsoLayoutEffect(measure);
useIsoLayoutEffect(() => {
const observer = new ResizeObserver(measure);
for (const el of [scrollRef.current, blockRef.current, viewportRef.current, bodyRef.current, contentRef.current]) {
if (el) observer.observe(el);
}
const question = expanded ? contentRef.current?.querySelector("[data-dock-scroll-area]") : null;
if (question) observer.observe(question);
window.addEventListener("resize", measure);
return () => {
observer.disconnect();
window.removeEventListener("resize", measure);
};
}, [measure, mode, expanded, scrollRef, blockRef, viewportRef, bodyRef, contentRef]);
}
/** Overlay and transcript are siblings. Keep native scrolling in an actual
* scroll container (including the short-window viewport), then forward only
* at its edge. overflow:hidden/visible boxes cannot consume the wheel. */
export function forwardDockWheel(event: WheelEvent, root: HTMLElement, transcript: HTMLElement) {
if (event.deltaY === 0) return;
for (let n = event.target as HTMLElement | null; n && n !== root; n = n.parentElement) {
if (!(n instanceof HTMLElement) || !/^(auto|scroll)$/.test(getComputedStyle(n).overflowY)) continue;
if (n.scrollHeight - n.clientHeight < 2) continue;
if (event.deltaY < 0 ? n.scrollTop > 0 : n.scrollTop < n.scrollHeight - n.clientHeight - 0.5) return;
}
event.preventDefault();
transcript.scrollTop += event.deltaY * (event.deltaMode === 1 ? 16 : event.deltaMode === 2 ? transcript.clientHeight : 1);
}
components/home.tsx
+29 -30
View File
@@ -7,11 +7,13 @@ import { GraphView } from "@/components/graph-view";
import { MinimalHome } from "@/components/minimal-home";
import { SetupDialog } from "@/components/setup-dialog";
import { ComposerActions, NodePickerControl } from "@/components/composer-actions";
import { useStrongKeyless, useZeroKeys, ZeroKeysPill } from "@/components/setup-cta";
import { useModelSetup, selectedModelUnavailable, openModelPicker, ZeroKeysPill } from "@/components/setup-cta";
import { DictationBar, joinDictation, useDictation } from "@/components/dictation";
import { useFieldCycle } from "@/components/field-cycle";
import { SidebarToggle } from "@/components/sidebar";
import { StatsBar } from "@/components/stats-bar";
import { OPENING_PROMPT_FAILURE } from "@/components/opening-prompt";
import { SessionStarter } from "@/components/start-session";
import type { TurnUsage } from "@/src/core";
/**
@@ -41,15 +43,8 @@ export function Home({
// start until the user has picked all three slots — no defaults for
// 33 providers. The dock shows the CTA while unconfigured; submitting
// a topic opens the setup dialog WITH THE DRAFT PRESERVED (Save
// stores the slots and the queued message sends).
const [configured, setConfigured] = useState(initialConfigured);
// The zero-keys truth (zero-keys wave): no provider key = the pill
// takes precedence over the models CTA (keys come first in the chain).
const zeroKeys = useZeroKeys();
// The strong slot's keyless truth (pre-send key guards, owner interlude
// 2026-09-25): keys may exist generally while the STRONG slot's pinned
// provider has none — the same skip-everything treatment on send.
const strongKeyless = useStrongKeyless();
// stores slots only; the learner sends the message themselves).
const { catalog, configured, zeroKeys } = useModelSetup(initialConfigured);
const [setupOpen, setSetupOpen] = useState(false);
// Voice dictation (STT wave, 2026-09-17): finals append at the end of the
// topic draft; the strip lives in the stats bar's middle slot.
@@ -63,6 +58,9 @@ export function Home({
},
});
const [busy, setBusy] = useState(false);
const startingRef = useRef(false);
const starterRef = useRef(new SessionStarter());
const [startError, setStartError] = useState(false);
/* Run end focuses the topic input (owner, 2026-10-01): the hands are on
* the keyboard from the stop. Falling edge only; this surface's only
* dictation target IS this textarea, so no case-split is needed. */
@@ -112,7 +110,7 @@ export function Home({
};
const begin = async () => {
const t = topicRef.current.trim();
if (!t || busy) return;
if (!t || startingRef.current) return;
// Unconfigured: the send INTERCEPTS into the setup flow (the draft
// stays in the textarea; the learner sends it after Save).
if (!configured) {
@@ -123,36 +121,37 @@ export function Home({
// models case guards BEFORE creation; the key cases must too — a
// refused first send would otherwise leave an orphan error-shell
// chat in the sidebar (the 503 lands only after POST /api/chats
// already created the record). Same skip-everything-and-open-setup
// treatment; the server's 503 gate remains the belt for /chat/[id]
// sessions.
if (zeroKeys || strongKeyless) {
// already created the record). Zero providers opens connection setup;
// an unavailable selection with another provider opens its ordinary
// picker. The server's admission gate remains the belt.
if (zeroKeys) {
setSetupOpen(true);
return;
}
const strong = catalog?.slots.strong?.model;
if (selectedModelUnavailable(catalog, strong ? `${strong.provider}:${strong.id}` : undefined)) {
openModelPicker();
return; // Preserve the draft and never create an orphan session.
}
startingRef.current = true;
setBusy(true);
setStartError(false);
try {
// The map context (owner, 2026-10-04): picks in EITHER mode ride
// the CREATION as graphContext — the server writes a
// graph-context marker as message[0] (a separate custom message
// before the learner's first user message, propositions + the
// chains among them); the prompt stays the learner's own words.
const res = await fetch("/api/chats", {
method: "POST",
headers: { "Content-Type": "application/json" },
body: JSON.stringify({
title: t,
...(mode === "practice" ? { kind: "practice" } : {}),
...(practiceProps.length > 0 ? { graphContext: { propositions: practiceProps } } : {}),
}),
});
const { id } = (await res.json()) as { id: string };
// No settings PUT on create (per-session models, owner
// 2026-09-28): the SERVER stamps the newborn from the strong
// slot — the same default this surface's model control edits.
router.push(`/chat/${id}?prompt=${encodeURIComponent(t)}`);
const target = await starterRef.current.start({
title: t,
...(mode === "practice" ? { kind: "practice" } : {}),
...(practiceProps.length > 0 ? { graphContext: { propositions: practiceProps } } : {}),
}, fetch, JSON.stringify(catalog?.slots.strong));
router.push(target);
} catch {
startingRef.current = false;
setBusy(false);
setStartError(true);
}
};
@@ -241,6 +240,7 @@ export function Home({
)
)}
/>
{startError && <p role="alert" className="mb-2 text-sm text-destructive">{OPENING_PROMPT_FAILURE}</p>}
<form
className="flex items-end gap-1.5 rounded-2xl border bg-card p-1.5 shadow-lg"
onSubmit={(e) => {
@@ -326,7 +326,6 @@ export function Home({
open={setupOpen}
onClose={() => setSetupOpen(false)}
onDone={() => {
setConfigured(true);
setSetupOpen(false);
// Save only saves: the draft waits in the composer.
}}
components/model-catalog.ts
+54
View File
@@ -0,0 +1,54 @@
"use client";
import { useCallback, useEffect, useState } from "react";
import type { ModelCatalogResponse } from "@/src/core";
import { onSettingsSaved } from "@/components/settings-signal";
// Shared stale-while-revalidate catalog. Metadata refreshes retain the last
// good snapshot; connection changes clear selectable models immediately.
let catalogCache: ModelCatalogResponse | undefined;
let catalogError = false;
let catalogInflight: Promise<void> | undefined;
let catalogGeneration = 0;
let stopCatalogSignals: (() => void) | undefined;
const catalogReaders = new Set<() => void>();
const publishCatalog = () => { for (const reader of catalogReaders) reader(); };
function fetchCatalog(revalidate = false): void {
if (revalidate) { catalogGeneration++; catalogInflight = undefined; }
if (catalogInflight) return;
const generation = catalogGeneration;
catalogError = false;
publishCatalog();
catalogInflight = fetch("/api/models")
.then((r) => { if (!r.ok) throw new Error(`catalog request failed: ${r.status}`); return r.json() as Promise<ModelCatalogResponse>; })
.then((c) => {
if (generation === catalogGeneration) { catalogCache = c; catalogError = false; publishCatalog(); }
})
.catch(() => {
if (generation === catalogGeneration) { catalogError = true; publishCatalog(); }
})
.finally(() => { if (generation === catalogGeneration) catalogInflight = undefined; });
}
export function useModelCatalog(): { catalog: ModelCatalogResponse | undefined; failed: boolean; refresh: () => void } {
const [snapshot, setSnapshot] = useState(() => ({ catalog: catalogCache, failed: catalogError }));
useEffect(() => {
const first = catalogReaders.size === 0;
const read = () => setSnapshot({ catalog: catalogCache, failed: catalogError });
catalogReaders.add(read);
if (first) {
stopCatalogSignals = onSettingsSaved((section) => {
if (section === "keys") {
catalogGeneration++; catalogInflight = undefined; catalogCache = undefined;
fetchCatalog();
} else if (section === "models") fetchCatalog(true);
});
fetchCatalog(true);
} else fetchCatalog();
read();
return () => { catalogReaders.delete(read); if (!catalogReaders.size) { stopCatalogSignals?.(); stopCatalogSignals = undefined; } };
}, []);
const refresh = useCallback(() => fetchCatalog(true), []);
return { ...snapshot, refresh };
}
components/model-picker.tsx
+19 -59
View File
@@ -8,52 +8,11 @@ import { ModelMenuList, ThinkingMenuList } from "@/components/settings/model-men
import { composeSlotModel } from "@/components/settings/agent-section";
import { MenuShell } from "@/components/menu-shell";
import { Hint } from "@/components/ui/hint";
import { notifySettingsSaved, onSettingsSaved, useSettingsSaved } from "@/components/settings-signal";
import { notifySettingsSaved, useSettingsSaved } from "@/components/settings-signal";
import { useModelCatalog } from "@/components/model-catalog";
export { useModelCatalog } from "@/components/model-catalog";
import { defaultOrderLevel, populateLevel, THINKING_LABEL } from "@/components/thinking-default";
/** GET /api/models is cached per page load for instant menu opens, but
* the federation is LIVE per call — keys saved mid-session widen the
* catalog immediately server-side. ModelControl therefore re-fetches on
* every model-menu open (stale-while-revalidate: the cached list shows
* instantly, the fresh one swaps in while the menu is open). */
let catalogCache: ModelCatalogResponse | undefined;
let catalogInflight: Promise<void> | undefined;
let catalogGeneration = 0;
let stopCatalogSignals: (() => void) | undefined;
const catalogReaders = new Set<() => void>();
const publishCatalog = () => { for (const reader of catalogReaders) reader(); };
function fetchCatalog(revalidate = false): void {
if (revalidate) { catalogGeneration++; catalogInflight = undefined; }
if (catalogInflight) return;
const generation = catalogGeneration;
catalogInflight = fetch("/api/models")
.then((r) => { if (!r.ok) throw new Error(`catalog request failed: ${r.status}`); return r.json() as Promise<ModelCatalogResponse>; })
.then((c) => { if (generation === catalogGeneration) { catalogCache = c; publishCatalog(); } })
.catch(() => { /* existing unavailable/usage-only posture, no retry timer */ })
.finally(() => { if (generation === catalogGeneration) catalogInflight = undefined; });
}
export function useModelCatalog(): { catalog: ModelCatalogResponse | undefined; refresh: () => void } {
const [catalog, setCatalog] = useState(catalogCache);
useEffect(() => {
const first = catalogReaders.size === 0;
const read = () => setCatalog(catalogCache);
catalogReaders.add(read);
if (first) {
stopCatalogSignals = onSettingsSaved((section) => {
if (section === "keys") {
catalogGeneration++; catalogInflight = undefined; catalogCache = undefined;
publishCatalog(); fetchCatalog();
} else if (section === "models") fetchCatalog(true);
});
fetchCatalog(true);
} else fetchCatalog();
read();
return () => { catalogReaders.delete(read); if (!catalogReaders.size) { stopCatalogSignals?.(); stopCatalogSignals = undefined; } };
}, []);
const refresh = useCallback(() => fetchCatalog(true), []);
return { catalog, refresh };
}
/** provider:id → contextWindow over the cached catalog — the client-side
* twin of the server's catalogContextLimits join (context-meter
* amendment, owner 2026-09-23: pi's live catalog is the ONLY limit
@@ -84,7 +43,9 @@ export function catalogLimits(catalog: Pick<ModelCatalogResponse, "models">): Ma
export function ModelControl({ chatId, onModelChange }: { chatId?: string; onModelChange?: (model: string | undefined) => void }) {
const { catalog, refresh } = useModelCatalog();
const [models, setModels] = useState<ModelsSettings | undefined>();
const [session, setSession] = useState<ChatSettings | undefined>();
// A successful read with no settings is a legacy session; undefined
// means no session record is known yet, not permission to inherit Strong.
const [session, setSession] = useState<{ settings?: ChatSettings }>();
const [open, setOpen] = useState<"model" | "thinking" | null>(null);
const [error, setError] = useState<string | null>(null);
const containerRef = useRef<HTMLDivElement>(null);
@@ -99,8 +60,8 @@ export function ModelControl({ chatId, onModelChange }: { chatId?: string; onMod
const readRecord = useCallback(() => {
if (chatId) {
fetch(`/api/chats/${chatId}`)
.then((r) => r.json() as Promise<{ settings?: ChatSettings }>)
.then((body) => setSession(body.settings))
.then((r) => { if (!r.ok) throw new Error(`session read failed: ${r.status}`); return r.json() as Promise<{ settings?: ChatSettings }>; })
.then((body) => setSession({ settings: body.settings }))
.catch(() => {
/* same read-only posture as a models-section failure */
});
@@ -124,15 +85,17 @@ export function ModelControl({ chatId, onModelChange }: { chatId?: string; onMod
});
// Dual-bound truth: chat mode displays the session's own record,
// falling back to the strong slot for legacy unstamped sessions —
// EXACTLY what resolveConversationSpec runs. Draft mode is the strong
// slot itself.
// falling back to Strong only for a successfully read legacy unstamped
// session — EXACTLY what resolveConversationSpec runs. An unresolved or
// failed record read must not report a provisional global preference as
// this session's identity (including to the Send guard). Draft mode is
// the strong slot itself.
const strong = models?.strong;
const override: ChatSettings | undefined = chatId
? {
model: session?.model ?? strong?.model,
thinkingLevel: session?.thinkingLevel ?? strong?.thinkingLevel,
}
? session ? {
model: session.settings?.model ?? strong?.model,
thinkingLevel: session.settings?.thinkingLevel ?? strong?.thinkingLevel,
} : undefined
: strong;
const effectiveId = override?.model?.id;
const effectiveProvider = override?.model?.provider;
@@ -236,7 +199,7 @@ export function ModelControl({ chatId, onModelChange }: { chatId?: string; onMod
return;
}
const stored = (await res.json().catch(() => null)) as { settings?: ChatSettings } | null;
setSession(stored?.settings ?? next);
setSession({ settings: stored?.settings ?? next });
setOpen(null);
} catch {
setError("Could not reach the server — try again.");
@@ -315,10 +278,7 @@ export function ModelControl({ chatId, onModelChange }: { chatId?: string; onMod
return (
<div ref={containerRef} className="relative flex items-center gap-1.5" data-testid="model-control">
<Hint
// Trimmed (owner, subtlety round 2026-09-25): "Model selector"
// — the long sentence was way too much. ⌘K opens this same
// menu (registry, override-aware).
label="Model selector"
label={effectiveId ? `${effectiveProvider ? `${effectiveProvider}:` : ""}${effectiveId}` : "Model selector"}
action="open-model-picker"
side="bottom"
>
components/new-session.tsx
+45 -37
View File
@@ -8,10 +8,12 @@ import { DictationBar, joinDictation, useDictation } from "@/components/dictatio
import { useFieldCycle } from "@/components/field-cycle";
import { SidebarToggle } from "@/components/sidebar";
import { SetupDialog } from "@/components/setup-dialog";
import { useStrongKeyless, useZeroKeys, ZeroKeysPill } from "@/components/setup-cta";
import { useModelSetup, selectedModelUnavailable, openModelPicker, ZeroKeysPill } from "@/components/setup-cta";
import { StatsBar } from "@/components/stats-bar";
import { ChainTableOverlay } from "@/components/chain-table.tsx";
import { ComposerActions, NodePickerControl } from "@/components/composer-actions";
import { OPENING_PROMPT_FAILURE } from "@/components/opening-prompt";
import { SessionStarter } from "@/components/start-session";
import type { ChatSettings, GraphSnapshot } from "@/src/core";
/** The draft's pick set (owner round 10): the SAME shape the home pick
@@ -33,18 +35,18 @@ export type DraftPick = { id: string; name: string; statement: string };
const DOCK = "mx-auto w-full max-w-[46rem] px-4";
/**
* The draft's whole behavior as one pure seam: create the session (title =
* the first message), persist the lifted pre-chat settings, return the push
* target — /chat/<id>?prompt=<msg> — whose initialPrompt path delivers the
* message as turn 1. Injected fetch/router so tests pin the contract.
* Returns null on empty text; rethrows on transport failure (the caller's
* busy flag resets).
* Create, safely hand over the full opening text, then return a short
* /chat/<id> target. Storage/creation failures leave the caller's draft
* intact. A per-composer starter reuses an already-created session on an
* unchanged retry. The server stamps settings at birth; no settings PUT.
*/
export async function startDraftSession(
text: string,
fetchLike: typeof fetch,
mode?: "practice",
picks?: readonly DraftPick[],
starter = new SessionStarter(),
settingsIdentity?: string,
): Promise<string | null> {
const t = text.trim();
if (!t) return null;
@@ -52,19 +54,13 @@ export async function startDraftSession(
// from the pick bar creates its session WITH the graph context — the
// same contract as the home screen's practice send (the marker with
// [id] name: statement lines, target validation's truth set).
const res = await fetchLike("/api/chats", {
method: "POST",
headers: { "Content-Type": "application/json" },
body: JSON.stringify(
mode === "practice"
? { title: t, kind: "practice", ...(picks && picks.length > 0 ? { graphContext: { propositions: picks } } : {}) }
: { title: t },
),
});
const { id } = (await res.json()) as { id: string };
// No settings PUT on create (per-session models, owner 2026-09-28):
// the SERVER stamps the newborn from the strong slot.
return `/chat/${id}?prompt=${encodeURIComponent(t)}`;
return starter.start(
mode === "practice"
? { title: t, kind: "practice", ...(picks && picks.length > 0 ? { graphContext: { propositions: picks } } : {}) }
: { title: t },
fetchLike,
settingsIdentity,
);
}
/** The pure shell (no router) — the render-test seam: the draft view's
@@ -77,6 +73,7 @@ export function DraftShell({
text,
onChange,
onSend,
startError = false,
setupCta,
mode = "teach",
picks = [],
@@ -88,6 +85,7 @@ export function DraftShell({
text: string;
onChange: (v: string) => void;
onSend: () => void;
startError?: boolean;
/** The practice draft (owner, 2026-10-04): ?mode=practice — the
* created session carries kind:"practice" and the copy changes
* (the practice agent opens; no probing). */
@@ -190,6 +188,7 @@ export function DraftShell({
)}
setup={setupCta}
/>
{startError && <p role="alert" className="pointer-events-auto mb-2 text-sm text-destructive">{OPENING_PROMPT_FAILURE}</p>}
<div className="pointer-events-auto overflow-hidden rounded-2xl border bg-card shadow-lg">
<Composer
busy={busy}
@@ -243,6 +242,9 @@ export function NewSession({ initialConfigured, mode = "teach" }: { initialConfi
setText(v);
};
const [busy, setBusy] = useState(false);
const startingRef = useRef(false);
const starterRef = useRef(new SessionStarter());
const [startError, setStartError] = useState(false);
// THE PICK BAR (owner round 10): the draft's lifted pick set + the
// snapshot the standalone chain-table picker reads. Lazy — the graph
// is fetched on FIRST open only, then reused.
@@ -275,19 +277,12 @@ export function NewSession({ initialConfigured, mode = "teach" }: { initialConfi
// 33 providers, chats cannot start. The draft stays in the textarea;
// Save stores the slots and never sends (owner ruling 2026-10-09) —
// the learner sends the draft themselves.
const [configured, setConfigured] = useState(initialConfigured);
// The zero-keys truth (zero-keys wave): no provider key = the pill
// takes precedence over the models CTA (keys come first in the chain).
const zeroKeys = useZeroKeys();
// The strong slot's keyless truth (pre-send key guards, owner interlude
// 2026-09-25): keys may exist generally while the STRONG slot's pinned
// provider has none — the same skip-everything treatment on send.
const strongKeyless = useStrongKeyless();
const { catalog, configured, zeroKeys } = useModelSetup(initialConfigured);
const [setupOpen, setSetupOpen] = useState(false);
const send = useCallback(() => {
const draft = textRef.current.trim();
if (!draft || busy) return;
if (!draft || startingRef.current) return;
if (!configured) {
setSetupOpen(true);
return;
@@ -295,25 +290,39 @@ export function NewSession({ initialConfigured, mode = "teach" }: { initialConfi
// The pre-send KEY guards (owner interlude, no-orphan-chats): guard
// BEFORE startDraftSession creates the record — a refused first
// send would otherwise leave an orphan error-shell chat in the
// sidebar. Same skip-everything-and-open-setup treatment; the
// server's 503 gate remains the belt for /chat/[id] sessions.
if (zeroKeys || strongKeyless) {
// sidebar. Zero providers opens connection setup; an unavailable
// selection with another provider opens the ordinary Strong picker.
// The server's admission gate remains the belt.
if (zeroKeys) {
setSetupOpen(true);
return;
}
const strong = catalog?.slots.strong?.model;
if (selectedModelUnavailable(catalog, strong ? `${strong.provider}:${strong.id}` : undefined)) {
openModelPicker();
return; // The learner replaces it explicitly, then sends again.
}
startingRef.current = true;
setBusy(true);
void startDraftSession(draft, fetch, mode === "practice" ? "practice" : undefined, mode === "practice" ? picksRef.current : undefined)
setStartError(false);
void startDraftSession(draft, fetch, mode === "practice" ? "practice" : undefined, mode === "practice" ? picksRef.current : undefined,
starterRef.current, JSON.stringify(catalog?.slots.strong))
.then((target) => {
if (target) router.push(target);
else setBusy(false);
else { startingRef.current = false; setBusy(false); }
})
.catch(() => setBusy(false));
}, [busy, router, configured, zeroKeys, strongKeyless]);
.catch(() => {
startingRef.current = false;
setBusy(false);
setStartError(true);
});
}, [router, configured, zeroKeys, catalog, mode]);
return (
<>
<DraftShell
busy={busy}
startError={startError}
text={text}
onChange={applyText}
onSend={send}
@@ -369,7 +378,6 @@ export function NewSession({ initialConfigured, mode = "teach" }: { initialConfi
open={setupOpen}
onClose={() => setSetupOpen(false)}
onDone={() => {
setConfigured(true);
setSetupOpen(false);
// Save only saves (owner ruling 2026-10-09): the draft
// waits in the composer; the learner sends it.
components/opening-prompt.ts
+59
View File
@@ -0,0 +1,59 @@
/**
* Same-tab handover of a new session's full opening text. Never put the
* message in the navigation URL: long query payloads can be rejected before
* the chat page loads. Storage is required for automatic delivery, not an
* optional best-effort side effect of navigating away from the source draft.
*/
export const OPENING_PROMPT_FAILURE = "Couldn’t open the session. Your message is still here. Try sending again.";
export interface OpeningPrompt {
text: string;
/** Written BEFORE sending. A later mount restores a draft, never reposts. */
attempted: boolean;
}
const keyFor = (id: string) => `lh.openingPrompt.${id}`;
export function stashOpeningPrompt(id: string, text: string, attempted = false): boolean {
try {
const value = JSON.stringify({ text, attempted } satisfies OpeningPrompt);
// Catch both the storage getter and its methods (quota/security errors).
const storage = globalThis.sessionStorage;
storage.setItem(keyFor(id), value);
return storage.getItem(keyFor(id)) === value;
} catch {
return false;
}
}
export function readOpeningPromptEntry(id: string): OpeningPrompt | undefined {
try {
const value = globalThis.sessionStorage.getItem(keyFor(id));
if (!value) return undefined;
try {
const entry: unknown = JSON.parse(value);
if (entry && typeof entry === "object" && "text" in entry && typeof entry.text === "string" &&
entry.text.length > 0 && "attempted" in entry && typeof entry.attempted === "boolean") {
return { text: entry.text, attempted: entry.attempted };
}
} catch { /* Old raw-text stashes have no proof that sending never began. */ }
return { text: value, attempted: true };
} catch {
return undefined;
}
}
export function readOpeningPrompt(id: string): string | undefined {
return readOpeningPromptEntry(id)?.text;
}
export function clearOpeningPrompt(id: string): boolean {
try {
globalThis.sessionStorage.removeItem(keyFor(id));
return true;
} catch {
// Canonical conversation/acknowledgment remains the truth. A later
// mount with real conversation can try cleanup again; never crash it.
return false;
}
}
components/practice-cards.tsx
+5 -5
View File
@@ -1,8 +1,8 @@
"use client";
import { MessageResponse } from "@/components/ai-elements/message";
import { ToolContentResponse } from "@/components/ai-elements/message";
import { cn } from "@/lib/utils";
import { repairMathForDisplay, type ArtifactPartProps } from "@/src/core";
import type { ArtifactPartProps } from "@/src/core";
/**
* The practice transcript cards (practice sessions, owner design
@@ -66,7 +66,7 @@ export function ProblemCard({ part }: ArtifactPartProps) {
<div data-testid="problem-card" className="lh-card-enter my-1 rounded-xl border bg-card p-4">
<div className="mb-1 text-xs font-medium uppercase tracking-wide text-muted-foreground">Problem</div>
<div className="lh-prose text-[15px]">
<MessageResponse>{repairMathForDisplay(statement)}</MessageResponse>
<ToolContentResponse>{statement}</ToolContentResponse>
</div>
</div>
);
@@ -133,7 +133,7 @@ export function ConcludeCard({ part }: ArtifactPartProps) {
(owner round 3): the same quiet 50% register as quiz explanations. */}
{verdict.length > 0 && (
<div className="text-sm text-foreground/50">
<MessageResponse>{repairMathForDisplay(verdict)}</MessageResponse>
<ToolContentResponse>{verdict}</ToolContentResponse>
</div>
)}
{/* Then the divider, then the labeled model solution (Slice G: the
@@ -144,7 +144,7 @@ export function ConcludeCard({ part }: ArtifactPartProps) {
<div>
<div className="mb-0.5 text-xs font-medium uppercase tracking-wide text-muted-foreground">Model solution</div>
<div className="lh-prose text-[15px]">
<MessageResponse>{repairMathForDisplay(reveal)}</MessageResponse>
<ToolContentResponse>{reveal}</ToolContentResponse>
</div>
</div>
</>
components/practice-dock.tsx
+3 -4
View File
@@ -2,12 +2,11 @@
import { useCallback, useRef, useState } from "react";
import { ChevronRightIcon } from "lucide-react";
import { MessageResponse } from "@/components/ai-elements/message";
import { ToolContentResponse } from "@/components/ai-elements/message";
import { Button } from "@/components/ui/button";
import { joinDictation, useDictationTarget } from "@/components/dictation";
import { useEnterSubmits } from "@/components/enter-submits";
import { cn } from "@/lib/utils";
import { repairMathForDisplay } from "@/src/core";
/**
* The practice dock (practice sessions, owner design 2026-10-04): an
@@ -106,10 +105,10 @@ export function PracticeDock({ problemCallId, input, expanded, onExpand, onSubmi
cards' label row — owner round, 2026-10-05) + the statement,
markdown + LaTeX rendered, capped by the measured column
(--dock-max-h; the old fixed 40vh cut long problems short). */}
<div className="max-h-[var(--dock-max-h,40vh)] overflow-y-auto px-4 pt-3 pb-2">
<div data-dock-scroll-area="" className="max-h-[var(--dock-max-h,40vh)] overflow-y-auto px-4 pt-3 pb-2">
<div className="mb-1 text-xs font-medium uppercase tracking-wide text-muted-foreground">Problem</div>
<div className="lh-prose text-[15px]">
<MessageResponse>{typeof input.statement === "string" ? repairMathForDisplay(input.statement) : input.statement}</MessageResponse>
<ToolContentResponse>{input.statement}</ToolContentResponse>
</div>
</div>
components/question-dock.tsx
+35 -15
View File
@@ -5,10 +5,11 @@ import { ArrowDownIcon, ChevronRightIcon, PlayIcon } from "lucide-react";
import { Button } from "@/components/ui/button";
import { finishDictation, isDictationActive, joinDictation, useDictationTarget } from "@/components/dictation";
import { useEnterSubmits } from "@/components/enter-submits";
import { MessageResponse } from "@/components/ai-elements/message";
import { ToolContentResponse } from "@/components/ai-elements/message";
import { cn } from "@/lib/utils";
import {
buildAnswer,
repairMathForDisplay,
normalizeQuizType,
CONFIDENCE_STOPS,
DEFAULT_CONFIDENCE_INDEX,
@@ -28,6 +29,7 @@ import {
type DockSelection,
type ExplainToolInput,
type QuestionAnswerData,
type QuestionFormDraft,
type QuestionData,
type QuestionMode,
type QuizToolInput,
@@ -71,6 +73,9 @@ export interface QuestionDockProps {
reveal?: boolean;
onExpand: () => void;
onSubmit: (output: QuestionAnswerData) => void;
/** Restore only this call's unconfirmed submission after recovery. */
initialDraft?: QuestionFormDraft;
onSaveDraft?: (draft: QuestionFormDraft) => void;
/** Escape in a question field stops dictation (STT wave) — threaded
* from the page that owns the dictation hook. */
onEscape?: () => void;
@@ -100,6 +105,8 @@ export function QuestionDock({
reveal = false,
onExpand,
onSubmit,
initialDraft,
onSaveDraft,
onEscape,
phase = "probe",
dictationClaim = true,
@@ -155,6 +162,8 @@ export function QuestionDock({
phase={phase}
active={expanded}
onSubmit={onSubmit}
initialDraft={initialDraft}
onSaveDraft={onSaveDraft}
onEscape={onEscape}
dictationClaim={dictationClaim}
compactionRunning={compactionRunning}
@@ -166,7 +175,7 @@ export function QuestionDock({
/** A plain-text teaser of a question for the collapsed bar: unwrap math, flatten common LaTeX. */
function teaserText(q: string): string {
return q
return repairMathForDisplay(q)
.replace(/\\\(([\s\S]*?)\\\)/g, "$1") // \(...\) inline math
.replace(/\\\[([\s\S]*?)\\\]/g, "$1") // \[...\] display math
.replace(/\$\$?([^$]*)\$\$?/g, "$1") // standard $...$/$$...$$ math
@@ -250,6 +259,8 @@ function QuestionDockForm({
phase = "probe",
active,
onSubmit,
initialDraft,
onSaveDraft,
onEscape,
dictationClaim = true,
compactionRunning = false,
@@ -311,22 +322,22 @@ function QuestionDockForm({
// single-select displaces, multi-select coexists — picking the ask
// "Other…" row in a multi-select must never clear already-picked
// options, and picking an option must never clear a typed custom answer.
const [sel, setSel] = useState<DockSelection>(emptyDockSelection());
const [sel, setSel] = useState<DockSelection>(() => initialDraft?.selection ?? emptyDockSelection());
// The learner's transfer channel — thinking, never "a note".
const [thinking, setThinking] = useState("");
const [other, setOther] = useState("");
const [text, setText] = useState("");
const [thinking, setThinking] = useState(initialDraft?.thinking ?? "");
const [other, setOther] = useState(initialDraft?.other ?? "");
const [text, setText] = useState(initialDraft?.text ?? "");
/* Refs MIRROR the two dictation-target fields ("nothing gets lost"
* round, owner 2026-10-01): the run's flush lands via setState, which
* a closure waiting through `await finishDictation()` cannot observe
* in time — submit reads these, kept current at every write site. */
const thinkingTextRef = useRef("");
const answerTextRef = useRef("");
const thinkingTextRef = useRef(initialDraft?.thinking ?? "");
const answerTextRef = useRef(initialDraft?.text ?? "");
// Confidence (owner, 2026-10-01): an index into CONFIDENCE_STOPS,
// resting at 50 — a real coin flip by definition, untouched or
// deliberate (the same number, the same meaning). The dock unmounts
// between questions, so each question starts fresh at 50.
const [confIdx, setConfIdx] = useState<number>(DEFAULT_CONFIDENCE_INDEX);
const [confIdx, setConfIdx] = useState<number>(initialDraft?.confidenceIndex ?? DEFAULT_CONFIDENCE_INDEX);
const selectMode: SelectMode = qd.mode === "multi-select" ? "multi-select" : "single-select";
// The PRIMARY dictation input (owner, STT round 1): while a question
@@ -390,6 +401,15 @@ function QuestionDockForm({
// "untouched"); buildAnswer drops it on "I don't know".
confidence: kind === "quiz" ? CONFIDENCE_STOPS[confIdx] : undefined,
};
// Snapshot before optimistic submission unmounts the form. Ref values
// include the dictation finals that just landed, unlike stale closures.
onSaveDraft?.({
selection: { ...sel, selected: new Set(sel.selected) },
thinking: thinkingTextRef.current,
other,
text: answerTextRef.current,
confidenceIndex: confIdx,
});
onSubmit(buildAnswer(qd, req, toolName));
};
@@ -544,7 +564,7 @@ function QuestionDockForm({
{/* Question area — scrolls internally if very tall; the bottom row stays put.
Capped by the measured column (--dock-max-h; the old fixed 55dvh cut
long questions short — owner round, 2026-10-05). */}
<div data-session-find-root="" className="max-h-[var(--dock-max-h,55dvh)] overflow-y-auto px-4 pt-3">
<div data-dock-scroll-area="" data-session-find-root="" className="max-h-[var(--dock-max-h,55dvh)] overflow-y-auto px-4 pt-3">
<div className="mb-1 text-xs font-medium uppercase tracking-wide text-muted-foreground">
{toolName === "probe"
? "Probe"
@@ -560,11 +580,11 @@ function QuestionDockForm({
{qd.mode === "multi-select" && " · select all that apply"}
</div>
<div className="mb-3 font-medium">
<MessageResponse>{qd.question}</MessageResponse>
<ToolContentResponse>{qd.question}</ToolContentResponse>
</div>
{qd.context && (
<div className="mb-3 text-sm text-muted-foreground">
<MessageResponse>{qd.context}</MessageResponse>
<ToolContentResponse>{qd.context}</ToolContentResponse>
</div>
)}
@@ -600,7 +620,7 @@ function QuestionDockForm({
)}
</span>
<span>
<MessageResponse>{opt.label}</MessageResponse>
<ToolContentResponse>{opt.label}</ToolContentResponse>
{kind !== "quiz" && opt.description && (
// Option subtext renders in Source Sans 3 (owner's mockup);
// the "I don't know" copy stays JetBrains Mono. QUIZ options render
@@ -609,9 +629,9 @@ function QuestionDockForm({
// records must not leak either. Same MessageResponse path as the
// label (owner, 2026-10-12): descriptions are full markdown —
// LaTeX included — not plain text.
<MessageResponse className="lh-prose block text-xs text-muted-foreground">
<ToolContentResponse className="lh-prose block text-xs text-muted-foreground">
{opt.description}
</MessageResponse>
</ToolContentResponse>
)}
</span>
</button>
components/server-events.tsx
+3
View File
@@ -116,6 +116,9 @@ export function wireSource(source: NotifySource, hub: ReturnType<typeof createNo
const EventsContext = createContext<ReturnType<typeof createNotifyHub> | null>(null);
/** Transport adapters may also be mounted in isolation with an injected bridge. */
export function useNotifyHub() { return useContext(EventsContext); }
export function ServerEventsProvider({ children, connect }: { children: ReactNode; connect?: () => NotifySource }) {
const hub = useMemo(() => createNotifyHub(), []);
useEffect(() => wireSource((connect ?? openEventsChannel)(), hub), [hub, connect]);
components/session-core.ts
+220 -40
View File
@@ -17,6 +17,9 @@ import type { AgentEvent } from "@earendil-works/pi-agent-core";
import type { ToolResultMessage } from "@earendil-works/pi-ai";
import {
needsContinuation,
danglingQuestionOf,
answeredCallIds,
failedResponseAtEnd,
measuredContextTokens,
newAnswerResult,
newNextMessage,
@@ -30,6 +33,7 @@ import {
type LhViewMessage,
type LiveExecution,
type QuestionAnswerData,
type QuestionDraft,
} from "../src/core";
export type SessionStatus = "ready" | "submitted" | "streaming" | "error";
@@ -78,13 +82,25 @@ async function turnInFlight(model: SessionModel): Promise<boolean | null> {
}
}
/** One wire event off the SSE body: parsed JSON, or the [DONE] terminal. */
async function* sseEvents(reader: ReadableStreamDefaultReader<Uint8Array>): AsyncGenerator<Record<string, unknown>> {
const STREAM_DROPPED = "The connection dropped before the turn completed.";
/** Mark transport failures without confusing JSON/application bugs with drops. */
class StreamReadError extends Error {
constructor(cause: unknown) {
super(cause instanceof Error ? cause.message : String(cause), { cause });
}
}
/** Wire events; return true ONLY for the explicit completion marker. */
async function* sseEvents(reader: ReadableStreamDefaultReader<Uint8Array>): AsyncGenerator<Record<string, unknown>, boolean> {
const decoder = new TextDecoder();
let buf = "";
for (;;) {
const { done, value } = await reader.read();
if (done) return;
let chunk: ReadableStreamReadResult<Uint8Array>;
try { chunk = await reader.read(); }
catch (err) { throw new StreamReadError(err); }
const { done, value } = chunk;
if (done) return false;
buf += decoder.decode(value, { stream: true });
for (;;) {
const end = buf.indexOf("\n\n");
@@ -94,7 +110,7 @@ async function* sseEvents(reader: ReadableStreamDefaultReader<Uint8Array>): Asyn
const dataLine = event.split("\n").find((l) => l.startsWith("data: "));
if (!dataLine) continue;
const data = dataLine.slice("data: ".length);
if (data === "[DONE]") return;
if (data === "[DONE]") return true;
yield JSON.parse(data) as Record<string, unknown>;
}
}
@@ -164,6 +180,8 @@ function sameDetails(a: unknown, b: unknown): boolean {
export interface SessionState {
messages: LhViewMessage[];
nativeMessages: LhMessage[];
/** Server-read history only; optimistic bubbles are not delivery proof. */
hasConfirmedConversation: boolean;
/** Completed conversation usage; null after compaction or before first usage. */
contextTokens: number | null;
status: SessionStatus;
@@ -196,6 +214,8 @@ export function visibleQueuedSteers(
export interface SessionModelOptions extends SessionModelDeps {
chatId: string;
initialMessages: LhMessage[];
/** Retained, memory-only server failure when reopening a session. */
initialError?: string;
/** Fires on every state change (the hook re-renders). */
onChange?: () => void;
/** The reattach settled terminal (turn-settled): no turn running, the
@@ -220,6 +240,9 @@ export class SessionModel {
// at every full sync; the write-ahead absorbs the optimistic entries.
private base: LhMessage[];
private tail: LhMessage[] = [];
private questionDrafts = new Map<string, QuestionDraft>();
/** Only learner answers submitted through send(), never silent dismissals. */
private unconfirmedAnswers = new Map<string, string>();
// The live turn's execution layer (view model mid-turn states):
private executions = new Map<string, LiveExecution>();
private streamingIds = new Set<string>();
@@ -229,17 +252,24 @@ export class SessionModel {
private statusValue: SessionStatus = "ready";
/** Request/loop failures are transient; provider failures derive from messages. */
private errorValue: { message: string } | null = null;
/** Dismissing a banner must not authorize an automatic replacement turn. */
private automaticContinuationBlocked = false;
/** A completed failure from THIS consume, not the previous attempt's record. */
private liveFailureId: string | null = null;
// The single live consume (POST response or reattach stream): a newer
// consume cancels the older — one reader at a time.
private consumeAbort: AbortController | null = null;
/** Retain the latest request across stop(): an opening POST still owns
* its refusal after view detachment, but never after a newer request. */
private latestRequest: AbortController | null = null;
/** Synchronous ready check for the gate (send/read mount logic). */
readyRef = true;
constructor(opts: SessionModelOptions) {
this.chatId = opts.chatId;
this.base = opts.initialMessages;
this.errorValue = opts.initialError === undefined ? null : { message: opts.initialError };
this.automaticContinuationBlocked = opts.initialError !== undefined;
// BOUND: the browser's native fetch is a Window method — storing it
// unbound makes every call throw "Illegal invocation" in the browser
// (Node's undici doesn't care, so the tests never saw it). Always go
@@ -262,18 +292,19 @@ export class SessionModel {
? { executions: this.executions, inFlight: true, streamingIds: this.streamingIds }
: undefined;
const messages = [...this.base, ...this.tail];
const latest = messages.at(-1);
// One rule for live and restored failures: only the final completed
// assistant can own the banner. While trying again, hide the old
// failure from the first Thinking frame, but show a new failure at end.
const latest = failedResponseAtEnd(messages);
// Presentation/compaction markers do not resolve a failure. While
// trying again, hide the old banner from the first Thinking frame,
// but show a new completed failure from this consume immediately.
const providerError =
latest?.role === "assistant" && latest.stopReason === "error" &&
latest !== undefined &&
!this.incompleteUsageIds.has(latest.id) && (!live || latest.id === this.liveFailureId)
? { message: latest.errorMessage ?? "An error occurred." }
: null;
return {
messages: projectView(messages, live),
nativeMessages: messages,
hasConfirmedConversation: this.base.some(m => m.role === "user" || m.role === "assistant"),
contextTokens: measuredContextTokens(messages, this.incompleteUsageIds),
status: this.statusValue,
error: this.errorValue ?? providerError,
@@ -300,6 +331,7 @@ export class SessionModel {
private setError(message: string | null): void {
this.errorValue = message === null ? null : { message };
if (message !== null) this.automaticContinuationBlocked = true;
this.onChange?.();
}
@@ -346,6 +378,12 @@ export class SessionModel {
}
const agent = event as unknown as AgentEvent;
switch (agent.type) {
case "agent_start":
// Positive evidence of a new/replayed server-owned turn. Any
// later failure in that replay blocks continuation again.
this.automaticContinuationBlocked = false;
if (this.errorValue !== null) this.setError(null);
return false;
case "message_start":
case "message_update":
case "message_end": {
@@ -445,18 +483,28 @@ export class SessionModel {
/** Consume one SSE body to its [DONE] terminal. The turn-settled
* terminal (stream-route only) fires the settle hook and ends. */
async consumeStream(res: Response, controller = this.consumeAbort): Promise<void> {
async consumeStream(res: Response, controller = this.consumeAbort, allowRecovery = true): Promise<void> {
if (!res.body) throw new Error("The turn stream body is empty.");
const reader = res.body.getReader();
let sawEvent = false;
let sawDone = false;
try {
for await (const event of sseEvents(reader)) {
const events = sseEvents(reader);
for (;;) {
let next: IteratorResult<Record<string, unknown>, boolean>;
try { next = await events.next(); }
catch (err) {
if (err instanceof StreamReadError) break;
throw err;
}
if (this.consumeAbort !== controller) return;
if (next.done) { sawDone = next.value; break; }
const event = next.value;
if (event.type === "turn-settled") {
// The raced-settle terminal: no turn to attach to — the
// record is current; the settle hook refetches.
await reader.cancel().catch(() => {});
if (this.consumeAbort === controller) this.onReattachSettled?.();
sawDone = true;
this.onReattachSettled?.();
return;
}
const acted = this.applyWireEvent(event);
@@ -466,14 +514,32 @@ export class SessionModel {
}
}
} finally {
await reader.cancel().catch(() => {});
reader.releaseLock();
// A cancelled consume may finish after its replacement has entered
// Thinking. Only the current owner can settle shared state.
if (this.consumeAbort === controller) {
this.streamingIds = new Set();
this.executions = new Map();
if (this.statusValue !== "error") this.setStatus("ready");
if (sawDone && this.statusValue !== "error") this.setStatus("ready");
}
}
if (this.consumeAbort !== controller || sawDone) return;
if (!allowRecovery) throw new StreamReadError(new Error(STREAM_DROPPED));
// One read-only recovery: refresh the canonical base, then replay the
// SAME server-owned turn. Never resend the learner's input.
await this.resumeStream(true, false);
}
private async fetchView(url: string, signal: AbortSignal, recoveringDrop: boolean): Promise<Response> {
try {
const res = await this.fetch(url, { signal });
if (recoveringDrop && !res.ok) throw new Error((await res.text()) || STREAM_DROPPED);
return res;
} catch (err) {
if (recoveringDrop && !signal.aborted) throw new StreamReadError(err);
throw err;
}
}
/**
@@ -489,10 +555,11 @@ export class SessionModel {
* lands on top. The stream then replays the turn from its first event
* through the SAME consume path.
*/
async resumeStream(refetchRecordFirst = true): Promise<void> {
async resumeStream(refetchRecordFirst = true, allowRecovery = true): Promise<void> {
this.consumeAbort?.abort();
const controller = new AbortController();
this.consumeAbort = controller;
this.latestRequest = controller;
// REPLAY ONTO A CLEAN SLATE (owner, optimistic round 3, 2026-10-02):
// the reattach stream replays the turn from its FIRST event — an
// aborted earlier attempt (React StrictMode's simulated remount)
@@ -519,16 +586,27 @@ export class SessionModel {
this.setStatus("submitted");
try {
if (refetchRecordFirst) {
const recordRes = await this.fetch(`/api/chats/${encodeURIComponent(this.chatId)}`, {
signal: controller.signal,
});
const recordRes = await this.fetchView(`/api/chats/${encodeURIComponent(this.chatId)}`, controller.signal, !allowRecovery);
if (this.consumeAbort !== controller) return;
if (recordRes.ok) {
const record = (await recordRes.json()) as { messages?: LhMessage[] };
let text: string;
try { text = await recordRes.text(); }
catch (err) {
if (!allowRecovery && !controller.signal.aborted) throw new StreamReadError(err);
throw err;
}
// Parse outside the transport catch: malformed JSON is a
// contract bug, not a connection-drop diagnostic.
const record = JSON.parse(text) as { messages?: LhMessage[]; turnError?: string };
if (this.consumeAbort !== controller) return;
if (Array.isArray(record.messages)) {
this.mutate(() => {
this.base = record.messages!;
if (record.turnError !== undefined) {
this.errorValue = { message: record.turnError };
this.automaticContinuationBlocked = true;
}
this.forgetResolvedDrafts(record.messages!);
this.incompleteUsageIds.clear();
this.tail = [];
});
@@ -543,15 +621,19 @@ export class SessionModel {
});
}
}
const res = await this.fetch(`/api/chats/${encodeURIComponent(this.chatId)}/stream`, {
signal: controller.signal,
});
const res = await this.fetchView(`/api/chats/${encodeURIComponent(this.chatId)}/stream`, controller.signal, !allowRecovery);
if (this.consumeAbort !== controller) return;
if (!res.ok) throw new Error((await res.text()) || "Failed to attach to the running turn.");
await this.consumeStream(res, controller);
await this.consumeStream(res, controller, allowRecovery);
} catch (err) {
if (this.consumeAbort !== controller || (err as Error).name === "AbortError") return; // a newer consume won
this.setError((err as Error).message);
if (this.consumeAbort !== controller || controller.signal.aborted) return; // intentionally cancelled
// A failed record read can leave the optimistic answer in the tail.
// Restore its Submit affordance; the next attempt reconciles first.
this.mutate(() => {
this.tail = this.tail.filter(m => !this.unconfirmedAnswers.has(m.id));
this.unconfirmedAnswers.clear();
});
this.setError(!allowRecovery && err instanceof StreamReadError ? STREAM_DROPPED : (err as Error).message);
this.setStatus("error");
}
}
@@ -562,13 +644,44 @@ export class SessionModel {
* POST + consume. Ready-gated: one turn per client at a time (the
* server's 409 + identity retry covers the wake-race window).
*/
async send(input: ChatInput): Promise<void> {
if (!this.readyRef) return;
async send(input: ChatInput, onAccepted?: () => void): Promise<boolean> {
const recovering = this.statusValue === "error";
if (!this.readyRef && !recovering) return false;
this.consumeAbort?.abort();
const controller = new AbortController();
this.consumeAbort = controller;
this.latestRequest = controller;
this.setError(null);
this.setStatus("submitted");
if (recovering && input.type === "answer") {
try {
const res = await this.fetch(`/api/chats/${encodeURIComponent(this.chatId)}`, { signal: controller.signal });
if (this.consumeAbort !== controller) return false;
if (!res.ok) throw new Error((await res.text()) || "An error occurred.");
const record = await res.json() as { messages?: LhMessage[]; turnInFlight?: boolean; turnError?: string };
if (this.consumeAbort !== controller) return false;
// No verified record/liveness: retain the draft and never guess
// whether a lost acknowledgement meant the input was accepted.
if (!Array.isArray(record.messages) || typeof record.turnInFlight !== "boolean") throw new Error("An error occurred.");
this.setMessages(record.messages, record.turnError);
if (record.turnInFlight) {
await this.resumeStream(false);
return false;
}
if (input.type === "answer" && (answeredCallIds(record.messages).has(input.toolCallId) || danglingQuestionOf(record.messages)?.toolCallId !== input.toolCallId)) {
// The call already resolved, or no longer owns the dock. Adopt
// truth, never POST that answer a second time.
this.setStatus("ready");
return false;
}
} catch (err) {
if (this.consumeAbort !== controller || controller.signal.aborted) return false;
this.setError((err as Error).message);
this.setStatus("error");
return false;
}
}
if (input.type !== "continue") this.automaticContinuationBlocked = false;
// The optimistic appends — exactly the messages the server's
// write-ahead will persist (inputAppends), with LOCAL ids so
// adoption can tell them from the record's absorbed copies.
@@ -585,11 +698,14 @@ export class SessionModel {
optimistic.push({ ...newNextMessage(), id: localId() });
} else if (input.type === "answer") {
const tool = pendingCallTool([...this.base, ...this.tail], input.toolCallId) ?? "quiz";
optimistic.push({ ...newAnswerResult(input.toolCallId, tool, input.answer), id: localId() });
const result = { ...newAnswerResult(input.toolCallId, tool, input.answer), id: localId() };
optimistic.push(result);
this.unconfirmedAnswers.set(result.id, input.toolCallId);
}
if (optimistic.length > 0) this.mutate(() => {
this.tail = [...this.tail, ...optimistic];
});
let accepted = false;
try {
// SUBMITS ARE NOT ABORTABLE BY VIEW TEARDOWN (owner, 2026-10-10 —
// the dead-session-start postmortem): the POST must NOT ride
@@ -605,16 +721,24 @@ export class SessionModel {
// connection is Diet A's whole job.
const res = await this.postInputEvent(input);
if (res.status === 409) {
this.setError("A turn is already running — try again in a moment.");
this.setStatus("error");
return;
throw new Error("A turn is already running — try again in a moment.");
}
if (!res.ok) {
// Refusal envelopes (503 BYOK/key gates) and 400s: the RAW body
// text rides the error message — the banner shows it, and the
// page's refusal flow JSON-parses the envelope out of it.
throw new Error((await res.text()) || `The send failed (${res.status}).`);
// Keep model/key refusal envelopes verbatim: the page's refusal
// flow parses them. Retry's 400s are ordinary user-facing
// guidance, so show their error text without JSON syntax.
let message = (await res.text()) || `The send failed (${res.status}).`;
if (input.type === "retry" && res.status === 400) {
try {
const refusal = JSON.parse(message) as { error?: unknown } | null;
if (typeof refusal?.error === "string" && refusal.error.length > 0) message = refusal.error;
} catch { /* Non-envelope failures keep their original text. */ }
}
throw new Error(message);
}
accepted = true;
// Opening-message handover ends at acceptance, not stream completion.
onAccepted?.();
// CONNECTION DIET B (owner, 2026-10-10 — parallel-session freeze
// postmortem): the POST acknowledges and closes; there is no stream
// on this response to consume. Attach through the SAME reattach wire
@@ -626,12 +750,37 @@ export class SessionModel {
// A stop() or a newer send/consume won while the ack was in
// flight: the turn is live server-side either way — do not attach
// a stream from a view that is gone or superseded.
if (this.consumeAbort !== controller) return;
if (this.consumeAbort !== controller) return true;
await this.resumeStream(true);
return true;
} catch (err) {
if ((err as Error).name === "AbortError") return; // stop() or a newer consume won
const failedOpening = input.type === "text" && onAccepted !== undefined && !accepted;
if (failedOpening) {
// The POST outlives stop(), including StrictMode's cleanup. Remove
// only its own failed optimistic bubble even if a newer request
// won; never touch canonical messages or another attempt's tail.
const ids = new Set(optimistic.map(m => m.id));
this.mutate(() => { this.tail = this.tail.filter(m => !ids.has(m.id)); });
}
const detachedOpening = failedOpening && this.consumeAbort === null && this.latestRequest === controller;
if ((this.consumeAbort !== controller || controller.signal.aborted) && !detachedOpening) return accepted;
// A detached opening refusal must still reach the replayed mount's
// existing setup recovery. A newer send/reattach ALWAYS owns errors
// and status instead, even if that request has since been stopped.
// Ordinary text and answer teardown behavior stays unchanged.
if (input.type === "answer") {
const ids = new Set(optimistic.map(m => m.id));
this.mutate(() => {
this.tail = this.tail.filter(m => !ids.has(m.id));
for (const id of ids) this.unconfirmedAnswers.delete(id);
});
}
this.setError((err as Error).message);
this.setStatus("error");
// true means we observed this POST's acknowledgment. false does
// NOT prove non-delivery: a network failure may hide acceptance.
// Opening-message callers restore a draft, never automatically retry.
return accepted;
}
}
@@ -681,7 +830,7 @@ export class SessionModel {
* result (fork 1), riding as its input event. */
submitAnswer(tool: string, toolCallId: string, output: QuestionAnswerData): Promise<void> {
void tool;
return this.send({ type: "answer", toolCallId, answer: output });
return this.send({ type: "answer", toolCallId, answer: output }).then(() => undefined);
}
/** The dismissal path (owner design 2026-10-03, "A2"): the skipped
@@ -740,10 +889,32 @@ export class SessionModel {
this.setStatus("ready");
}
/** Submission snapshots outlive the dock, not the session mount. */
questionDraft(toolCallId: string): QuestionDraft | undefined {
return this.questionDrafts.get(toolCallId);
}
saveQuestionDraft(toolCallId: string, draft: Partial<QuestionDraft>): void {
this.questionDrafts.set(toolCallId, { ...this.questionDrafts.get(toolCallId), ...draft });
}
private forgetResolvedDrafts(record: readonly LhMessage[]): void {
const answered = answeredCallIds(record);
for (const id of answered) this.questionDrafts.delete(id);
for (const [id, callId] of this.unconfirmedAnswers) {
if (answered.has(callId)) this.unconfirmedAnswers.delete(id);
}
}
/** Full snapshot swap (the idle settle sync — server canon). */
setMessages(record: LhMessage[]): void {
setMessages(record: LhMessage[], turnError?: string): void {
this.mutate(() => {
this.base = record;
if (turnError !== undefined) {
this.errorValue = { message: turnError };
this.automaticContinuationBlocked = true;
}
this.forgetResolvedDrafts(record);
this.incompleteUsageIds.clear();
this.tail = [];
this.streamingMessageId = null; // the tail (and its slot) is gone
@@ -763,10 +934,19 @@ export class SessionModel {
if (baseAbsorbed) return; // not a pure append — the idle full sync is the net
this.mutate(() => {
this.base = record;
this.forgetResolvedDrafts(record);
this.tail = this.tail.filter((m) => !absorbed(m, record));
});
}
/** Automatic mount/sync continuation is never permission to retry a
* failed turn. Read current model truth, not a stale React closure. */
async continueIfOwed(record: readonly LhMessage[]): Promise<void> {
if (this.statusValue === "ready" && !this.automaticContinuationBlocked && this.getState().error === null && needsContinuation(record)) {
await this.send({ type: "continue" });
}
}
/** The mount/sync continuation fold (native-messages.ts). */
static needsContinuation = needsContinuation;
components/session-overlay.tsx
+19 -16
View File
@@ -7,6 +7,7 @@ import { fitGraphView, fitZoomFor, zoomedGraphView, type Viewport } from "@/comp
import { renderMermaid } from "@/components/mermaid-render";
import { MermaidError } from "@/components/mermaid-error";
import { Hint } from "@/components/ui/hint";
import { SvgImage } from "@/components/artifacts/svg-image";
/**
* SESSION-SCOPED OVERLAYS (owner wave, 2026-09-19): the "fullscreen"
@@ -269,22 +270,24 @@ export function MermaidOverlay({
: "transform 180ms ease-out, opacity 160ms ease-out",
}}
>
{/* data-lh-overlay-chart: globals.css pins the svg to the
wrapper's width (kills mermaid's inline max-width cap)
so the wrapper IS the natural-size box the fit math
centers. FIGURE MODE: a white stage (the figure's
backing law — models draw for a white page) with the
caller's invert filter riding it, exactly like the
in-card holder. */}
<div
data-lh-overlay-chart
className={cn(
"[&_svg]:h-auto [&_svg]:w-full",
rawSvg !== undefined && "bg-white",
)}
style={rawSvg !== undefined && inverted ? { filter: "invert(1) hue-rotate(180deg)" } : undefined}
dangerouslySetInnerHTML={{ __html: svg }}
/>
{/* Figures share the card's image boundary and white backing.
Mermaid keeps its inline renderer; globals.css removes
its max-width cap so it fills the natural-size box. */}
{rawSvg !== undefined ? (
<div
data-lh-overlay-chart
className="bg-white"
style={inverted ? { filter: "invert(1) hue-rotate(180deg)" } : undefined}
>
<SvgImage svg={svg} className="block h-auto w-full" />
</div>
) : (
<div
data-lh-overlay-chart
className="[&_svg]:h-auto [&_svg]:w-full"
dangerouslySetInnerHTML={{ __html: svg }}
/>
)}
</div>
<div className="pointer-events-none absolute inset-x-0 bottom-0 flex items-end justify-between p-2">
<span className="text-xs text-muted-foreground/60">Drag to pan · scroll to zoom</span>
components/session.ts
+13 -4
View File
@@ -22,16 +22,18 @@ export interface UseChatSessionOptions {
chatId: string;
/** The native record seed (server-rendered page or a full sync). */
initialMessages: LhMessage[];
/** Transient settlement failure retained by the server, not the transcript. */
initialError?: string;
/** Reattach on mount: a turn is running server-side with no live view. */
initiallyInFlight?: boolean;
}
export function useChatSession({ chatId, initialMessages, initiallyInFlight }: UseChatSessionOptions) {
export function useChatSession({ chatId, initialMessages, initialError, initiallyInFlight }: UseChatSessionOptions) {
// The model is per-mount (chatId-keyed component); onChange forces the
// re-render. The ref pattern keeps it stable across renders.
const modelRef = useRef<SessionModel | null>(null);
if (modelRef.current === null) {
modelRef.current = new SessionModel({ chatId, initialMessages });
modelRef.current = new SessionModel({ chatId, initialMessages, initialError });
}
const model = modelRef.current;
// The render VERSION — bumped by every model change, so the state memo
@@ -100,7 +102,8 @@ export function useChatSession({ chatId, initialMessages, initiallyInFlight }: U
[model],
);
const send = useCallback((input: ChatInput) => void model.send(input), [model]);
const send = useCallback((input: ChatInput, onAccepted?: () => void) => model.send(input, onAccepted), [model]);
const continueIfOwed = useCallback((record: readonly LhMessage[]) => void model.continueIfOwed(record), [model]);
const steer = useCallback((text: string) => void model.steer(text), [model]);
const submitAnswer = useCallback(
(tool: string, toolCallId: string, output: QuestionAnswerData) =>
@@ -115,7 +118,9 @@ export function useChatSession({ chatId, initialMessages, initiallyInFlight }: U
(problemCallId: string) => void model.dismissProblem(problemCallId),
[model],
);
const setMessages = useCallback((record: LhMessage[]) => model.setMessages(record), [model]);
const questionDraft = useCallback((toolCallId: string) => model.questionDraft(toolCallId), [model]);
const saveQuestionDraft = useCallback((toolCallId: string, draft: Parameters<SessionModel["saveQuestionDraft"]>[1]) => model.saveQuestionDraft(toolCallId, draft), [model]);
const setMessages = useCallback((record: LhMessage[], turnError?: string) => model.setMessages(record, turnError), [model]);
const adoptRecord = useCallback((record: LhMessage[]) => model.adoptRecord(record), [model]);
const stop = useCallback(() => model.stop(), [model]);
const clearError = useCallback(() => model.clearError(), [model]);
@@ -124,15 +129,19 @@ export function useChatSession({ chatId, initialMessages, initiallyInFlight }: U
return {
messages: state.messages,
nativeMessages: state.nativeMessages,
hasConfirmedConversation: state.hasConfirmedConversation,
contextTokens: state.contextTokens,
status: state.status,
error: state.error,
steerQueue: state.steerQueue,
send,
continueIfOwed,
steer,
submitAnswer,
dismissQuestion,
dismissProblem,
questionDraft,
saveQuestionDraft,
setMessages,
adoptRecord,
stop,
components/settings-signal.ts
+2 -2
View File
@@ -20,7 +20,7 @@ export const SETTINGS_SAVED_EVENT = "lh:settings-saved";
// "keys" joined the roster (zero-keys wave, 2026-09-25): keys-section
// dispatches it after save AND delete — deletions used to be invisible to
// the zero-keys pill and the setup dialog's auto-advance.
export type SettingsSection = "general" | "keys" | "agent" | "models" | "appearance" | "graph" | "compaction";
export type SettingsSection = "general" | "keys" | "agent" | "models" | "appearance" | "graph" | "compaction" | "dictation";
/** Writers call this after the PUT resolves ok — never before (a failed
* write notifies nothing; the record is unchanged). */
@@ -47,4 +47,4 @@ export function useSettingsSaved(handler: (section: SettingsSection) => void): v
const saved = useRef(handler);
saved.current = handler;
useEffect(() => onSettingsSaved((section) => saved.current(section)), []);
}
}
components/settings/agent-section.tsx
+19 -51
View File
@@ -55,6 +55,8 @@ import { useCallback, useEffect, useRef, useState } from "react";
import { MenuShell } from "@/components/menu-shell";
import { useSettingsSaved } from "@/components/settings-signal";
import { defaultOrderLevel, populateLevel } from "@/components/thinking-default";
import { useModelCatalog } from "@/components/model-catalog";
import { Hint } from "@/components/ui/hint";
import type React from "react";
import {
COMPACTION_THRESHOLD_FLOOR,
@@ -63,14 +65,13 @@ import {
type AgentSettings,
type CompactionSettings,
type GraphSettings,
type ModelCatalogEntry,
type ModelSlotName,
type ModelsSettings,
} from "@/src/core";
import type { ThinkingLevel } from "@earendil-works/pi-agent-core";
import { prettyModel } from "@/components/usage-format";
import { cn } from "@/lib/utils";
import { SettingsGroup, SettingsRow, Toggle, CommitInput } from "./controls.tsx";
import { SettingsGroup, SettingsRow, SettingsMenuTrigger, Toggle, CommitInput } from "./controls.tsx";
import { ModelMenuList, ThinkingMenuList } from "./model-menu";
import { ChatGPTBackgroundAccess } from "./chatgpt-connection";
import { NativeSubscriptionBackgroundAccess } from "./subscription-connections";
@@ -101,15 +102,6 @@ interface SegmentInfo {
shared: boolean;
}
/** The GET /api/models response fields this section consumes. */
interface ModelsBody {
models: ModelCatalogEntry[];
/** The three tiers' RESOLVED pins (BYOK: NULL per unset tier — no
* factory floor) — machinery only (the thinking list fallback); the
* rows' displays read the record override, never a phantom pin. */
slots: Record<ModelSlotName, { model: { provider: string; id: string }; thinkingLevel?: string } | null>;
}
/** Compose: one tool toggled, the rest of the section carried verbatim.
* Pure — exported for tests. */
export function composeTool(agent: AgentSettings | undefined, id: string, on: boolean): AgentSettings {
@@ -234,8 +226,10 @@ function MenuTrigger({
onClick,
widthClass,
disabled,
hint,
}: {
value: string;
hint?: string;
/** What the trigger shows when there is no override ("Default"). */
emptyLabel: string;
ariaLabel: string;
@@ -250,25 +244,20 @@ function MenuTrigger({
/** The no-key search state grays the whole row; the trigger follows. */
disabled?: boolean;
}) {
return (
<button
type="button"
const button = (
<SettingsMenuTrigger
value={value || emptyLabel}
aria-label={ariaLabel}
aria-expanded={open}
data-testid={testid}
onClick={() => {
if (!disabled) onClick();
}}
className={`flex h-7 items-center gap-1 rounded-sm border border-border bg-background px-1.5 font-ui text-xs text-foreground transition-colors ${
disabled ? "cursor-not-allowed" : "hover:border-muted-foreground/50"
} ${widthClass}`}
>
<span className="min-w-0 flex-1 truncate">{value || emptyLabel}</span>
<span aria-hidden className="text-muted-foreground">
▾
</span>
</button>
unavailable={disabled}
className={widthClass}
/>
);
return hint ? <Hint label={hint}>{button}</Hint> : button;
}
/* The section's dropdowns all mount the shared MenuShell (fix wave chunk
@@ -303,10 +292,9 @@ export function AgentSection({
/** Writes the whole compaction section (replace-PUT discipline). */
onCompactionChange: (next: CompactionSettings) => void;
}) {
const [catalog, setCatalog] = useState<ModelCatalogEntry[] | null>(null);
/** The three tiers' resolved pins — GET /api/models `slots`. The rows'
* Default hints and thinking lists resolve from it. */
const [slots, setSlots] = useState<ModelsBody["slots"] | null>(null);
const { catalog: body, failed: catalogFailed, refresh: refreshCatalog } = useModelCatalog();
const catalog = body?.models;
const slots = body?.slots;
/** The live search state (/api/search) — the row's reason line. */
const [search, setSearch] = useState<{ available: boolean; engine: string | null; reason: string } | null>(null);
/** Which search engines have a key (/api/keys) — the no-keys-anywhere
@@ -341,28 +329,9 @@ export function AgentSection({
.catch(() => {});
}, []);
// The catalog (federated: keyed providers only), the resolved slots,
// the prompt segments, and the search engines' key presence — one
// fetch each per section mount. The models read ALSO re-runs on the
// cross-surface signal (fix wave: another surface's write re-reads
// into the slot rows live).
const modelsRequest = useRef(0);
const modelsMounted = useRef(false);
const readModels = useCallback((connectionChanged = false) => {
const request = ++modelsRequest.current;
if (connectionChanged) setCatalog(null);
fetch("/api/models")
.then((r) => { if (!r.ok) throw new Error("Catalog unavailable"); return r.json(); })
.then((body: ModelsBody) => {
if (!modelsMounted.current || request !== modelsRequest.current) return;
setCatalog(body.models);
setSlots(body.slots);
})
.catch(() => { if (modelsMounted.current && request === modelsRequest.current) setCatalog(null); });
}, []);
// The shared catalog seeds the first frame and refreshes per mount and
// keys/models signal. Settings props still own saved selections/levels.
useSettingsSaved((section) => {
if (section === "models" || section === "keys") readModels(section === "keys");
// THE state-echo root cause: the prompt switches write through the
// agent section, but this view (the toggles' truth) never re-read —
// the settings-saved law (every reader re-reads) was missing this
@@ -371,8 +340,6 @@ export function AgentSection({
});
useEffect(() => {
modelsMounted.current = true;
readModels();
fetch("/api/settings/prompts")
.then((r) => r.json())
.then((body: { prober: SegmentInfo[]; teacher: SegmentInfo[] }) => setPrompts(body))
@@ -388,7 +355,6 @@ export function AgentSection({
setEngineKeys(present);
})
.catch(() => {});
return () => { modelsMounted.current = false; modelsRequest.current++; };
}, []);
// The reason line re-reads /api/search AFTER an engine write lands —
@@ -615,6 +581,7 @@ export function AgentSection({
return (
<>
<SettingsGroup label="Models" testid="group-models">
{catalogFailed && <p role="alert" className="text-xs text-red-400">Model catalog could not be loaded. Check the connection and retry. <button type="button" className="underline" onClick={refreshCatalog}>Retry</button></p>}
{MODEL_SLOTS.map((slot) => {
const override = models?.[slot];
const modelId = override?.model?.id ?? "";
@@ -652,6 +619,7 @@ export function AgentSection({
{slot !== "strong" && <><ChatGPTBackgroundAccess provider={coord?.provider} /><NativeSubscriptionBackgroundAccess provider={coord?.provider} /></>}
<MenuTrigger
value={modelId ? prettyModel(modelId) : ""}
hint={modelId ? `${provider}:${modelId}` : undefined}
emptyLabel="—"
ariaLabel={`${SLOT_META[slot].label} model`}
testid={`slot-model-${slot}`}
components/settings/appearance-section.tsx
+8 -15
View File
@@ -41,7 +41,7 @@ import { MenuShell } from "@/components/menu-shell";
import { Hint } from "@/components/ui/hint";
import { useAppearance } from "./appearance.tsx";
import { FontMenuList } from "./model-menu";
import { CommitInput, SettingsRow, ThinSelect, Toggle } from "./controls.tsx";
import { CommitInput, SettingsMenuTrigger, SettingsRow, ThinSelect, Toggle } from "./controls.tsx";
/** The factory reading family — the DEFAULT row's identity (unset =
* this). Exported for tests. */
@@ -210,29 +210,22 @@ export function AppearanceSection({
testid="row-font"
>
<span ref={fontMenuOpen ? menuAnchorRef : undefined} className="relative flex items-center gap-2">
{/* The MenuTrigger look (agent-section's collapsed
select — it lives unexported there, so the
classes ride inline; the searchable machine is
the shared part). The tooltip rides the CLOSED
button ONLY (owner ruling 2026-09-26: never while
{/* The shared Agent/Appearance collapsed-menu trigger.
The tooltip rides the CLOSED button ONLY
(owner ruling 2026-09-26: never while
the dropdown is out — it popped over the open
menu when the wrap covered the whole anchor
span). */}
{(() => {
const triggerButton = (
<button
type="button"
<SettingsMenuTrigger
value={appearance?.fontFamily || DEFAULT_READING_FONT}
aria-label="Reading font"
aria-expanded={fontMenuOpen}
data-testid="font-trigger"
onClick={openFontMenu}
className="flex h-7 min-w-20 items-center gap-1 rounded-sm border border-border bg-background px-1.5 font-ui text-xs text-foreground transition-colors hover:border-muted-foreground/50"
>
<span className="min-w-0 flex-1 truncate">{appearance?.fontFamily || DEFAULT_READING_FONT}</span>
<span aria-hidden className="text-muted-foreground">
▾
</span>
</button>
className="min-w-20"
/>
);
return fontMenuOpen ? (
triggerButton
components/settings/chatgpt-connection.tsx
+14 -13
View File
@@ -5,8 +5,10 @@ import type { ChatGPTStatus } from "@/src/core";
import { ConnectionDialog, ConnectionFooter, ConnectionRow } from "./connection-controls";
import { Button } from "@/components/ui/button";
import { Toggle } from "./controls";
import { useChatGPTBridge } from "./subscription-client";
export const CHATGPT_NOTICE = "LearnHarness stores your ChatGPT account details and credentials in an unencrypted local file on this device. When this connection is used, relevant conversation context, learning material, and tool results are sent directly to OpenAI. ChatGPT plan limits apply; any additional credits are authorized and managed in ChatGPT, not LearnHarness.";
export function chatGPTNotice(web = false): string { return web ? `${CHATGPT_NOTICE.replace("on this device", "on the machine running LearnHarness")} Complete ChatGPT sign-in in a browser on that same machine.` : CHATGPT_NOTICE; }
export const CHATGPT_BACKGROUND_HINT = "Background work counts toward your ChatGPT plan usage.";
const BACKGROUND_LABEL = "Allow background use for all ChatGPT accounts";
const REVOCATION_WARNING = "Account disconnected locally. Remote revocation could not be confirmed.";
@@ -41,6 +43,8 @@ const Context = createContext({
/** UI lifetime spans setup's keys → models transition. Only existing safe controls
* are used; account IDs remain control values, never displayed account labels. */
export function ChatGPTUIProvider({ children }: { children: React.ReactNode }) {
const bridge = useChatGPTBridge();
const web = typeof window !== "undefined" && bridge !== window.lh?.chatgpt;
const [status, setStatus] = useState<ChatGPTStatus>(empty);
const [dialog, setDialog] = useState<Dialog | null>(null);
const [busy, setBusy] = useState(false);
@@ -79,14 +83,14 @@ export function ChatGPTUIProvider({ children }: { children: React.ReactNode }) {
if (attempt.backgroundDirty && next.allowBackgroundUse !== attempt.background) {
operation.current = "change";
const request = ++statusSequence.current;
const result: unknown = await window.lh!.chatgpt!.consent(attempt.background);
const result: unknown = await bridge!.consent(attempt.background);
if (!owns(attempt)) return;
if (!isStatus(result)) { setError(replyError(result)); finish(attempt); return; }
replyFailure = result.error;
let saved: ChatGPTStatus | null;
if (request !== statusSequence.current) {
const refresh = ++statusSequence.current;
const value: unknown = await window.lh!.chatgpt!.status();
const value: unknown = await bridge!.status();
if (!owns(attempt)) return;
saved = refresh === statusSequence.current ? publishStatus(value) : statusRef.current;
} else saved = publishStatus(result);
@@ -120,7 +124,6 @@ export function ChatGPTUIProvider({ children }: { children: React.ReactNode }) {
};
useEffect(() => {
mounted.current = true;
const bridge = window.lh?.chatgpt;
const refresh = async () => {
const request = ++statusSequence.current;
try {
@@ -131,7 +134,7 @@ export function ChatGPTUIProvider({ children }: { children: React.ReactNode }) {
void refresh();
const off = bridge?.onChange(() => void refresh());
return () => { mounted.current = false; generation.current++; loginAttempt.current = null; off?.(); };
}, []);
}, [bridge]);
const managed = dialog?.kind === "manage" ? status.accounts.find((a) => a.id === dialog.accountId) ?? dialog.account : undefined;
const managedPresent = !!managed && status.accounts.some((a) => a.id === managed.id);
@@ -161,13 +164,13 @@ export function ChatGPTUIProvider({ children }: { children: React.ReactNode }) {
// Keep cancellation serialized even after login's pending IPC reply.
operation.current = "cancel"; setBusy(true);
const request = ++statusSequence.current;
void window.lh?.chatgpt?.cancel().then(async (value: unknown) => {
void bridge?.cancel().then(async (value: unknown) => {
if (!mounted.current || generation.current !== intent) return;
if (!isStatus(value)) { setError(replyError(value)); return; }
let next: ChatGPTStatus | null;
if (request !== statusSequence.current) {
const refresh = ++statusSequence.current;
const current: unknown = await window.lh!.chatgpt!.status();
const current: unknown = await bridge!.status();
if (!mounted.current || generation.current !== intent) return;
next = refresh === statusSequence.current ? publishStatus(current) : statusRef.current;
} else next = publishStatus(value);
@@ -195,7 +198,7 @@ export function ChatGPTUIProvider({ children }: { children: React.ReactNode }) {
let next: ChatGPTStatus | null;
if (request !== statusSequence.current) {
const refresh = ++statusSequence.current;
const current: unknown = await window.lh!.chatgpt!.status();
const current: unknown = await bridge!.status();
if (!mounted.current || intent !== generation.current) return;
next = refresh === statusSequence.current ? publishStatus(current) : statusRef.current;
} else next = publishStatus(value);
@@ -214,7 +217,6 @@ export function ChatGPTUIProvider({ children }: { children: React.ReactNode }) {
finally { if (intent === generation.current) { operation.current = null; if (mounted.current) setBusy(false); } }
}
async function login() {
const bridge = window.lh?.chatgpt;
const owner = dialogRef.current;
if (!bridge || operation.current || !status.available || status.pending || owner?.kind !== "connect") return;
const attempt: LoginAttempt = { generation: ++generation.current, dialog: owner, registration, background, backgroundDirty, pending: false, phase: "starting" };
@@ -241,7 +243,6 @@ export function ChatGPTUIProvider({ children }: { children: React.ReactNode }) {
// No finally: pending OAuth outlives this IPC promise. Only its status
// completion, failure, or explicit cancellation releases the attempt.
}
const bridge = typeof window !== "undefined" ? window.lh?.chatgpt : undefined;
const context = {
status, busy, error: warning ? null : visibleError, inlineError: dialog || warning ? null : visibleError, warning: dialog?.kind === "warning" ? null : warning,
connect: openConnect,
@@ -262,7 +263,7 @@ export function ChatGPTUIProvider({ children }: { children: React.ReactNode }) {
</> : dialog.kind === "manage" && dialog.confirm ? <div data-testid="chatgpt-disconnect-confirmation">
<p className="mt-3 text-sm font-semibold text-red-400">{CHATGPT_DISCONNECT_COPY.title}</p>
<p className="mt-3 text-xs text-foreground">{managed?.email || "ChatGPT account"}</p>
<p className="mt-3 text-xs text-muted-foreground">{CHATGPT_DISCONNECT_COPY.body}</p>
<p className="mt-3 text-xs text-muted-foreground">{web ? CHATGPT_DISCONNECT_COPY.body.replace("from this device", "from the machine running LearnHarness") : CHATGPT_DISCONNECT_COPY.body}</p>
{context.error && <p role="alert" className="mt-3 text-xs text-red-400">{context.error}</p>}
<div className="mt-4 flex items-center justify-end gap-2">
<Button variant="ghost" size="sm" disabled={busy} onClick={close}>Cancel</Button>
@@ -277,7 +278,7 @@ export function ChatGPTUIProvider({ children }: { children: React.ReactNode }) {
</select>
{registration && <Button variant="ghost" size="xs" disabled={busy} onClick={() => context.manage(registration)}>Manage…</Button>}
</div>}
<p className="mt-3 text-xs leading-relaxed text-muted-foreground" data-testid="chatgpt-privacy-notice">{CHATGPT_NOTICE}</p>
<p className="mt-3 text-xs leading-relaxed text-muted-foreground" data-testid="chatgpt-privacy-notice">{chatGPTNotice(web)}</p>
</> : null}
<div className={`mt-4 flex items-center gap-3 text-xs text-foreground ${busy && operation.current === "change" && status.available && !status.pending ? "[&>button:disabled]:cursor-auto" : ""}`} data-testid="chatgpt-global-background">
<Toggle label={BACKGROUND_LABEL} checked={dialog.kind === "connect" ? background : status.allowBackgroundUse} disabled={busy || status.pending || !status.available} onChange={(enabled) => {
@@ -327,12 +328,12 @@ function AccountDialog({ children, onClose, busy }: { children: React.ReactNode;
}
/** Shared compact settings/setup row; account selection precedes the quiet action. */
export function ChatGPTConnection(_props: { setup?: boolean } = {}) {
export function ChatGPTConnection({ labelClassName }: { setup?: boolean; labelClassName?: string } = {}) {
const ui = useContext(Context);
const active = ui.status.accounts.find((a) => a.id === ui.status.activeId && a.connected);
const accounts = ui.status.accounts.filter((a) => a.connected);
return <div className="flex flex-col gap-2" data-testid="chatgpt-connection">
<ConnectionRow label="ChatGPT" actionTestid={active ? undefined : "chatgpt-login"} action={active ? "Manage…" : "Connect"} disabled={!ui.status.available || ui.busy || ui.status.pending} onAction={() => active ? ui.manage(active.id) : ui.connect()}>
<ConnectionRow label="ChatGPT" labelClassName={labelClassName} actionTestid={active ? undefined : "chatgpt-login"} action={active ? "Manage…" : "Connect"} disabled={!ui.status.available || ui.busy || ui.status.pending} onAction={() => active ? ui.manage(active.id) : ui.connect()}>
{active && <select aria-label="ChatGPT account" value={active.id} disabled={ui.busy || ui.status.pending} onChange={(event) => ui.select(event.target.value)} className="min-w-0 rounded-sm border border-border bg-background px-2 py-1 text-xs text-foreground">
{accounts.map((a) => <option key={a.id} value={a.id}>{a.email || "ChatGPT account"}</option>)}
</select>}
components/settings/connection-controls.tsx
+2 -2
View File
@@ -4,9 +4,9 @@ import { createPortal } from "react-dom";
import { useBackdropDismiss } from "@/components/backdrop-dismiss";
import { Button } from "@/components/ui/button";
export function ConnectionRow({ label, action, disabled, onAction, children, testid, actionTestid }: { label: string; action: string; disabled: boolean; onAction: () => void; children?: React.ReactNode; testid?: string; actionTestid?: string }) {
export function ConnectionRow({ label, action, disabled, onAction, children, testid, actionTestid, labelClassName }: { label: string; action: string; disabled: boolean; onAction: () => void; children?: React.ReactNode; testid?: string; actionTestid?: string; labelClassName?: string }) {
return <div className="flex flex-wrap items-center gap-3" data-testid={testid}>
<span className="font-ui text-xs text-foreground">{label}</span>
<span className={labelClassName ?? "font-ui text-xs text-foreground"}>{label}</span>
{children}
<Button type="button" size="xs" variant={action === "Connect" ? "outline" : "ghost"} className={`cursor-pointer disabled:cursor-not-allowed ${action === "Connect" ? "" : "text-muted-foreground hover:text-foreground"}`} data-testid={actionTestid} disabled={disabled} onClick={onAction}>{action}</Button>
</div>;
components/settings/controls.tsx
+26 -2
View File
@@ -9,7 +9,7 @@
* control together) and the control itself locks — the row owns the dim.
*/
import { useState, type ReactNode } from "react";
import { useState, type ComponentPropsWithRef, type ReactNode } from "react";
import { cn } from "@/lib/utils";
/** A section block: pronounced mono heading + content. */
@@ -63,7 +63,7 @@ export function SettingsRow({
dimmed,
}: {
label: string;
hint?: string;
hint?: ReactNode;
children: React.ReactNode;
testid?: string;
/** Dependency graying, graph-view style: the whole row dims when the
@@ -81,6 +81,30 @@ export function SettingsRow({
);
}
/** The existing Agent/Appearance collapsed-menu look, shared with dictation.
* `unavailable` preserves callers that gate clicks without native disabling;
* `disabled` retains native button semantics where the caller requires them. */
export function SettingsMenuTrigger({
value, unavailable, disabled, className, ...props
}: Omit<ComponentPropsWithRef<"button">, "value" | "children"> & {
value: string;
unavailable?: boolean;
}) {
return <button
type="button"
{...props}
disabled={disabled}
className={cn(
"flex h-7 items-center gap-1 rounded-sm border border-border bg-background px-1.5 font-ui text-xs text-foreground transition-colors",
disabled || unavailable ? "cursor-not-allowed" : "hover:border-muted-foreground/50",
className,
)}
>
<span className="min-w-0 flex-1 truncate">{value}</span>
<span aria-hidden className="text-muted-foreground">▾</span>
</button>;
}
/** The on/off switch — a plain button in the switch role. */
export function Toggle({
checked,
components/settings/keys-section.tsx
+24 -1
View File
@@ -30,6 +30,7 @@ import { SubscriptionConnections } from "./subscription-connections.tsx";
import { ProviderMenuList } from "./model-menu.tsx";
import { MenuShell } from "@/components/menu-shell";
import { notifySettingsSaved } from "@/components/settings-signal";
import { DictationLanguageSetting } from "@/components/dictation-language";
/** One pi-ai provider with plain API-key auth (the Platform seam —
* derived from pi-ai's own env-key table; no drift possible). */
@@ -96,6 +97,14 @@ export function classifyKeys(
};
}
/** A key row names its credential, not whichever provider sorts first.
* Pi exposes Zen and Go separately but both read OPENCODE_API_KEY. */
export function providerKeyLabel(id: string, providers: readonly ProviderKeyInfo[] | undefined): string | undefined {
const matches = providers?.filter((p) => p.envVar === id);
if (matches?.some((p) => p.id === "opencode") && matches.some((p) => p.id === "opencode-go")) return "OpenCode (Zen / Go)";
return matches?.[0]?.label;
}
/** The minimal status line, e.g. "…ab12 · 64 chars" (" · env" marks an
* environment-owned value — the row itself is read-only then). Pure. */
export function keyStatusLine(status: AuthKeyStatus): string {
@@ -342,6 +351,9 @@ export function KeysSection({
if (openRow === null && adding === null) return;
const onKeyDown = (e: KeyboardEvent) => {
if (e.key !== "Escape") return;
// A language dropdown above this editor in the dismissal stack owns
// the first Escape; do not swallow it at document capture phase.
if (document.querySelector('[data-testid="dictation-language-menu"]')) return;
e.preventDefault();
e.stopPropagation();
setOpenRow(null);
@@ -417,7 +429,7 @@ export function KeysSection({
{providerRows.map((status) => (
<KeyTableRow
key={status.id}
label={providers?.find((p) => p.envVar === status.id)?.label ?? status.info?.label ?? status.id}
label={providerKeyLabel(status.id, providers) ?? status.info?.label ?? status.id}
docsUrl={status.info?.docsUrl}
status={status}
expanded={openRow === status.id}
@@ -478,6 +490,17 @@ export function KeysSection({
testid={`key-${status.id.toLowerCase()}`}
/>
))}
{label === "Dictation" && <>
<DictationLanguageSetting />
<p data-testid="dictation-setup-hint" className="px-1 text-[11px] leading-relaxed text-muted-foreground/70">
Create a <a
href="https://console.deepgram.com/"
target="_blank"
rel="noreferrer"
className="underline underline-offset-2 hover:text-foreground"
>Deepgram API key</a> with the Member role or higher, then add it above.
</p>
</>}
</SettingsGroup>)}
</>
components/settings/model-menu.tsx
+25 -80
View File
@@ -23,6 +23,7 @@ import { useEffect, useRef, useState } from "react";
import type { ModelCatalogEntry } from "@/src/core";
import { fuzzyScore, subscriptionLabel } from "@/src/core";
import { formatTokens } from "@/components/usage-format";
import { SearchMenuList } from "./search-menu-list";
function priceLabel(cost: ModelCatalogEntry["cost"]): string {
const fmt = (n: number) => `$${n < 1 ? n.toFixed(2).replace(/0$/, "") : n.toFixed(2)}`;
@@ -53,90 +54,34 @@ export function ModelMenuList({
/** Quiet footer note under the list (context-specific). */
footer?: string;
}) {
const [query, setQuery] = useState("");
const [menuIdx, setMenuIdx] = useState<number | null>(null);
const searchRef = useRef<HTMLInputElement>(null);
const listRef = useRef<HTMLUListElement>(null);
useEffect(() => {
searchRef.current?.focus();
}, []);
// Keep the arrow-walked result visible in the scrolling list.
useEffect(() => {
if (menuIdx === null) return;
listRef.current?.querySelector(`[data-idx="${menuIdx}"]`)?.scrollIntoView({ block: "nearest" });
}, [menuIdx]);
// The filtered, ranked model list.
const results = (models ?? [])
.map((m) => ({ m, s: fuzzyScore(query, m.provider, m.id, m.name) }))
.filter((r) => r.s !== null)
.sort((a, b) => b.s! - a.s! || a.m.id.localeCompare(b.m.id))
.map((r) => r.m);
return (
<>
<input
ref={searchRef}
data-testid="model-search"
value={query}
onChange={(e) => {
setQuery(e.target.value);
setMenuIdx(null); // new query → the walk restarts on Enter/↓
}}
onKeyDown={(e) => {
if (e.key === "ArrowDown" || e.key === "ArrowUp") {
e.preventDefault();
e.stopPropagation();
if (results.length === 0) return;
setMenuIdx((i) =>
e.key === "ArrowDown"
? Math.min((i ?? -1) + 1, results.length - 1)
: Math.max((i ?? results.length) - 1, 0),
);
} else if (e.key === "Enter") {
const chosen = results[menuIdx ?? 0];
if (chosen) {
e.preventDefault();
e.stopPropagation();
onPick(chosen);
}
}
}}
<SearchMenuList
items={models}
itemKey={(m) => `${m.provider}/${m.id}`}
searchTerms={(m) => [m.provider, m.id, m.name]}
compare={(a, b) => a.id.localeCompare(b.id)}
isSelected={(m) => m.id === effectiveId && m.provider === (effectiveProvider ?? m.provider)}
onPick={onPick}
searchTestid="model-search"
listTestid="model-list"
optionTestid={(m) => `model-option-${m.id.replace(/[^a-z0-9-]/gi, "_")}`}
placeholder="Search models…"
className="lh-input mb-1.5 w-full rounded-lg border bg-background px-2 py-1.5 text-foreground outline-none focus:ring-1 focus:ring-ring"
loading="Loading models…"
empty="No models match."
renderItem={(m) => <>
<span className="flex items-center justify-between gap-2">
<span className="truncate text-card-foreground">{m.name}</span>
<span className="shrink-0 tabular-nums text-muted-foreground" title={subscriptionLabel(`${m.provider}:${m.id}`) ? "Subscription connection" : "USD per 1M tokens in/out"}>
{subscriptionLabel(`${m.provider}:${m.id}`) ?? priceLabel(m.cost)}
</span>
</span>
<span className="flex items-center justify-between gap-2 text-muted-foreground">
<span className="truncate">{subscriptionLabel(`${m.provider}:${m.id}`) ? `${subscriptionLabel(`${m.provider}:${m.id}`)} (subscription)` : m.provider} · {m.id}</span>
<span className="shrink-0 tabular-nums">{formatTokens(m.contextWindow)}</span>
</span>
</>}
/>
{!models && <div className="px-2 py-3 text-center text-muted-foreground">Loading models…</div>}
{models && results.length === 0 && (
<div className="px-2 py-3 text-center text-muted-foreground">No models match.</div>
)}
<ul ref={listRef} data-testid="model-list" className="flex max-h-72 flex-col overflow-y-auto px-1">
{results.map((m, i) => (
<li key={`${m.provider}/${m.id}`}>
<button
type="button"
data-testid={`model-option-${m.id.replace(/[^a-z0-9-]/gi, "_")}`}
data-idx={i}
onClick={() => onPick(m)}
className={`flex w-full flex-col gap-0.5 rounded-lg px-2 py-1.5 text-left outline-none transition-colors hover:bg-muted ${
i === menuIdx ? "bg-accent/40 ring-1 ring-inset ring-border" : ""
}${m.id === effectiveId && m.provider === (effectiveProvider ?? m.provider) && i !== menuIdx ? " bg-muted/60" : ""}`}
>
<span className="flex items-center justify-between gap-2">
<span className="truncate text-card-foreground">{m.name}</span>
<span className="shrink-0 tabular-nums text-muted-foreground" title={subscriptionLabel(`${m.provider}:${m.id}`) ? "Subscription connection" : "USD per 1M tokens in/out"}>
{subscriptionLabel(`${m.provider}:${m.id}`) ?? priceLabel(m.cost)}
</span>
</span>
<span className="flex items-center justify-between gap-2 text-muted-foreground">
<span className="truncate">{subscriptionLabel(`${m.provider}:${m.id}`) ? `${subscriptionLabel(`${m.provider}:${m.id}`)} (subscription)` : m.provider} · {m.id}</span>
<span className="shrink-0 tabular-nums">{formatTokens(m.contextWindow)}</span>
</span>
</button>
</li>
))}
</ul>
{footer && (
<div className="mt-1 border-t px-2 pt-1.5 text-xs text-muted-foreground">{footer}</div>
)}
components/settings/search-menu-list.tsx
+93
View File
@@ -0,0 +1,93 @@
"use client";
import { useEffect, useRef, useState, type ReactNode } from "react";
import { fuzzyScore } from "@/src/core";
/** The model picker's search, arrow-walk, row styling and scrolling machine.
* Models and dictation languages supply only their catalog and row content.
* The caller owns the shared MenuShell and dismissal behavior. */
export function SearchMenuList<T>({
items, itemKey, searchTerms, compare, isSelected, onPick, renderItem,
searchTestid, listTestid, optionTestid, placeholder, ariaLabel, loading, empty,
disabled = false,
}: {
items: readonly T[] | undefined;
itemKey: (item: T) => string;
searchTerms: (item: T) => string[];
compare?: (a: T, b: T) => number;
isSelected: (item: T) => boolean;
onPick: (item: T) => void;
renderItem: (item: T) => ReactNode;
searchTestid: string;
listTestid: string;
optionTestid: (item: T) => string;
placeholder: string;
ariaLabel?: string;
loading?: string;
empty: string;
disabled?: boolean;
}) {
const [query, setQuery] = useState("");
const [menuIdx, setMenuIdx] = useState<number | null>(null);
const searchRef = useRef<HTMLInputElement>(null);
const listRef = useRef<HTMLUListElement>(null);
useEffect(() => { searchRef.current?.focus(); }, []);
useEffect(() => {
if (menuIdx === null) return;
listRef.current?.querySelector(`[data-idx="${menuIdx}"]`)?.scrollIntoView({ block: "nearest" });
}, [menuIdx]);
const results = (items ?? [])
.map((item) => ({ item, score: fuzzyScore(query, ...searchTerms(item)) }))
.filter((result) => result.score !== null)
.sort((a, b) => b.score! - a.score! || (compare?.(a.item, b.item) ?? 0))
.map(({ item }) => item);
return <>
<input
type="text"
ref={searchRef}
data-testid={searchTestid}
aria-label={ariaLabel}
value={query}
disabled={disabled}
onChange={(event) => { setQuery(event.target.value); setMenuIdx(null); }}
onKeyDown={(event) => {
if (event.key === "ArrowDown" || event.key === "ArrowUp") {
event.preventDefault();
event.stopPropagation();
if (disabled || results.length === 0) return;
setMenuIdx((i) => event.key === "ArrowDown"
? Math.min((i ?? -1) + 1, results.length - 1)
: Math.max((i ?? results.length) - 1, 0));
} else if (event.key === "Enter") {
const chosen = results[menuIdx ?? 0];
if (!disabled && chosen !== undefined) {
event.preventDefault();
event.stopPropagation();
onPick(chosen);
}
}
}}
placeholder={placeholder}
className="lh-input mb-1.5 w-full rounded-lg border bg-background px-2 py-1.5 text-foreground outline-none focus:ring-1 focus:ring-ring"
/>
{!items && loading && <div className="px-2 py-3 text-center text-muted-foreground">{loading}</div>}
{items && results.length === 0 && <div className="px-2 py-3 text-center text-muted-foreground">{empty}</div>}
<ul ref={listRef} data-testid={listTestid} className="flex max-h-72 flex-col overflow-y-auto px-1">
{results.map((item, i) => <li key={itemKey(item)}>
<button
type="button"
data-testid={optionTestid(item)}
data-idx={i}
disabled={disabled}
onClick={() => onPick(item)}
className={`flex w-full flex-col gap-0.5 rounded-lg px-2 py-1.5 text-left outline-none transition-colors hover:bg-muted ${
i === menuIdx ? "bg-accent/40 ring-1 ring-inset ring-border" : ""
}${isSelected(item) && i !== menuIdx ? " bg-muted/60" : ""}`}
>{renderItem(item)}</button>
</li>)}
</ul>
</>;
}
components/settings/settings-dialog.tsx
+9 -18
View File
@@ -40,7 +40,8 @@ import {
type UserSettings,
} from "@/src/core";
import { useAppearance } from "./appearance.tsx";
import { SettingsSection, SettingsRow, Toggle } from "./controls.tsx";
import { SettingsSection } from "./controls.tsx";
import { UpdateSettings } from "./update-settings.tsx";
import { AgentSection } from "./agent-section.tsx";
import type { ModelsSettings } from "@/src/core";
import { MENU_CLOSE_EVENT, MENU_OPEN_EVENT } from "@/components/menu-shell";
@@ -75,10 +76,6 @@ const DialogContext = createContext<{ open: (section?: string, row?: string) =>
/** AppShell-level mount: the sidebar row and the mod+, chord call this.
* `open(section)` lands the dialog on a specific section (the first-run
* setup flow deep-links to Keys when no provider key exists yet). */
/** The auto-update row's hint (tooltip wave): honest user language — no
* stream vocabulary. Exported for the copy pin (tooltip-mechanism test). */
export const AUTO_UPDATE_HINT = "Automatic updates arrive in an upcoming release";
export function SettingsDialogProvider({
children,
}: {
@@ -92,6 +89,7 @@ export function SettingsDialogProvider({
setLandOnRow(row ?? null);
setShow(true);
}, []);
useEffect(() => window.lh?.onOpenSettings?.((section) => open(section)), [open]);
return (
<ChatGPTUIProvider><NativeSubscriptionUIProvider><DialogContext.Provider value={{ open }}>
{children}
@@ -362,19 +360,12 @@ export function SettingsDialogBody({
title="General"
testid="section-general"
>
<SettingsRow
label="Auto-update"
hint={AUTO_UPDATE_HINT}
testid="row-auto-update"
>
<Toggle
checked={record.general?.autoUpdate ?? true}
label="auto update"
onChange={(on) =>
void write("general", { autoUpdate: on })
}
/>
</SettingsRow>
<UpdateSettings
autoUpdate={record.general?.autoUpdate ?? true}
onAutoUpdateChange={(on) =>
void write("general", { ...record.general, autoUpdate: on })
}
/>
</SettingsSection>
)}
{active === "keys" && (
components/settings/subscription-client.ts
+77
View File
@@ -0,0 +1,77 @@
"use client";
import { useMemo } from "react";
import type { ChatGPTControlBridge, ChatGPTStatus, NativeSubscriptionControlBridge, NativeSubscriptionProviderId, NativeSubscriptionStatus } from "@/src/core";
import { useNotifyHub, type createNotifyHub } from "@/components/server-events";
type Hub = ReturnType<typeof createNotifyHub> | null;
type Reply<T> = { value: T; browserAction?: string; error?: string };
const PATH = "/api/subscriptions";
const headers = { "Content-Type": "application/json", "x-learnharness-subscriptions": "1" };
async function status() {
const response = await fetch(PATH, { headers, cache: "no-store" });
if (!response.ok) throw new Error("Subscription status is unavailable.");
return response.json() as Promise<{ chatgpt: ChatGPTStatus; native: NativeSubscriptionStatus[] }>;
}
async function control<T>(providerId: string, op: string, fields: Record<string, unknown> = {}): Promise<Reply<T>> {
const response = await fetch(PATH, { method: "POST", headers, body: JSON.stringify({ providerId, op, ...fields }) });
if (!response.ok) throw new Error("The subscription connection could not be changed.");
return response.json() as Promise<Reply<T>>;
}
function onChange(hub: Hub, listener: () => void): () => void {
const offEvent = hub?.subscribe("subscriptions-changed", listener);
const offResync = hub?.onResync(listener);
return () => { offEvent?.(); offResync?.(); };
}
/** Reserve the new tab synchronously in the click gesture. Only a one-use
* same-origin action URL enters app JS; provider URLs remain in the owner.
* Electron dev has no tabs: its narrow main-process opener performs this step. */
async function withBrowser<T>(action: () => Promise<Reply<T>>): Promise<T> {
const external = window.lh?.subscriptionTransport === "web" ? window.lh.openSubscriptionBrowser : undefined;
const tab = external ? null : window.open("about:blank", "_blank");
if (!external && !tab) throw new Error("Allow a sign-in tab to open, then try again.");
if (tab) tab.opener = null;
try {
const reply = await action();
if (reply.browserAction) {
if (!/^\/api\/subscriptions\/browser\?id=[0-9a-f-]{36}$/.test(reply.browserAction)) throw new Error("Invalid sign-in action.");
if (external) await external(reply.browserAction);
else if (tab && !tab.closed) tab.location.href = reply.browserAction;
else throw new Error("The sign-in tab was closed. Cancel and try again.");
} else tab?.close();
return reply.value;
} catch (error) { tab?.close(); throw error; }
}
export function createWebChatGPTBridge(hub: Hub): ChatGPTControlBridge {
const call = async (op: string, fields?: Record<string, unknown>) => (await control<ChatGPTStatus>("chatgpt", op, fields)).value;
return {
status: async () => (await status()).chatgpt,
login: (id) => withBrowser(() => control<ChatGPTStatus>("chatgpt", "login", { id })),
cancel: () => call("cancel"), select: (id) => call("select", { id }),
disconnect: (id) => call("disconnect", { id }), logout: (id) => call("logout", { id }),
consent: (enabled) => call("consent", { enabled }),
onChange: (listener) => onChange(hub, listener),
};
}
export function createWebNativeBridge(hub: Hub): NativeSubscriptionControlBridge {
const call = async (providerId: NativeSubscriptionProviderId, op: string, fields?: Record<string, unknown>) => { await control(providerId, op, fields); };
return {
status: async () => (await status()).native,
login: (providerId, enabled) => call(providerId, "login", { enabled }),
cancel: (providerId) => call(providerId, "cancel"),
respond: (providerId, attemptId, promptId, value) => call(providerId, "respond", { attemptId, promptId, value }),
open: async (providerId, attemptId, actionId) => { await withBrowser(() => control(providerId, "open", { attemptId, actionId })); },
retry: (providerId) => call(providerId, "retry"), disconnect: (providerId) => call(providerId, "disconnect"),
consent: (providerId, enabled) => call(providerId, "consent", { enabled }),
onChange: (listener) => onChange(hub, listener),
};
}
export function useChatGPTBridge(): ChatGPTControlBridge | undefined {
const hub = useNotifyHub();
return useMemo(() => typeof window === "undefined" ? undefined : window.lh?.subscriptionTransport !== "web" && window.lh?.chatgpt ? window.lh.chatgpt : createWebChatGPTBridge(hub), [hub]);
}
export function useNativeSubscriptionBridge(): NativeSubscriptionControlBridge | undefined {
const hub = useNotifyHub();
return useMemo(() => typeof window === "undefined" ? undefined : window.lh?.subscriptionTransport !== "web" && window.lh?.subscriptions ? window.lh.subscriptions : createWebNativeBridge(hub), [hub]);
}
components/settings/subscription-connections.tsx
+16 -13
View File
@@ -6,16 +6,19 @@ import { notifySettingsSaved } from "@/components/settings-signal";
import { ConnectionDialog, ConnectionFooter, ConnectionRow } from "./connection-controls";
import { ChatGPTConnection } from "./chatgpt-connection";
import { Toggle } from "./controls";
import { useNativeSubscriptionBridge } from "./subscription-client";
export const COPILOT_POLICY_NOTICE = "Pi's native Copilot sign-in may enable additional models whose policy is not yet configured. Models explicitly disabled by policy are not enabled. Organization restrictions still apply.";
export const SUBSCRIPTION_BACKGROUND_HINT = "Allow LearnHarness helper agents and automatic follow-ups to use this connection.";
export function subscriptionNotice(label: string): string { return `Your prompts, tool results, and relevant session context are sent to ${label} to run the work you select. Credentials are stored locally in an owner-only plaintext file. Provider terms, data-use policies, and plan limits apply. This connection never falls back to an API key.`; }
export function subscriptionDisconnectCopy(label: string): string { return `Disconnect ${label} from LearnHarness? This stops work using this connection and removes its saved credentials from this device. It does not cancel your subscription or revoke access at the provider. Saved model choices will remain unavailable until you reconnect or change them.`; }
export function subscriptionNotice(label: string, web = false): string { return `Your prompts, tool results, and relevant session context are sent to ${label} to run the work you select. Credentials are stored ${web ? "on the machine running LearnHarness" : "locally"} in an owner-only plaintext file. Provider terms, data-use policies, and plan limits apply. This connection never falls back to an API key.`; }
export function subscriptionDisconnectCopy(label: string, web = false): string { return `Disconnect ${label} from LearnHarness? This stops work using this connection and removes its saved credentials from ${web ? "the machine running LearnHarness" : "this device"}. It does not cancel your subscription or revoke access at the provider. Saved model choices will remain unavailable until you reconnect or change them.`; }
export function emptySubscriptionStatus(providerId: NativeSubscriptionProviderId): NativeSubscriptionStatus { return { providerId, available: false, connected: false, pending: false, allowBackgroundUse: true, connectionId: null, epoch: 0, authorityRevision: 0, revision: 0 }; }
const Context = createContext({ statuses: [] as NativeSubscriptionStatus[], busy: false, open: (_providerId: NativeSubscriptionProviderId) => {} });
const CONTROL_ERROR = "The connection could not be changed. Try again.";
export function NativeSubscriptionUIProvider({ children }: { children: React.ReactNode }) {
const bridge = useNativeSubscriptionBridge();
const web = typeof window !== "undefined" && bridge !== window.lh?.subscriptions;
const [statuses, setStatuses] = useState<NativeSubscriptionStatus[]>([]);
const [dialog, setDialog] = useState<NativeSubscriptionProviderId | null>(null);
const [view, setView] = useState<"notice" | "connection" | "manage">("notice");
@@ -30,7 +33,7 @@ export function NativeSubscriptionUIProvider({ children }: { children: React.Rea
async function refresh() {
const request = ++sequence.current;
try {
const values = await window.lh?.subscriptions?.status();
const values = await bridge?.status();
if (!mounted.current || request !== sequence.current || !Array.isArray(values)) return;
const accepted = NATIVE_SUBSCRIPTIONS.map(({ providerId }) => {
const next = values.find((s) => s.providerId === providerId), old = statusesRef.current.find((s) => s.providerId === providerId);
@@ -41,10 +44,10 @@ export function NativeSubscriptionUIProvider({ children }: { children: React.Rea
}
useEffect(() => {
mounted.current = true; void refresh();
const off = window.lh?.subscriptions?.onChange(() => { void refresh(); notifySettingsSaved("keys"); });
const off = bridge?.onChange(() => { void refresh(); notifySettingsSaved("keys"); });
return () => { mounted.current = false; generation.current++; off?.(); };
// The root-lifetime subscriber reads current state through refs.
}, []);
}, [bridge]);
useEffect(() => { setPromptValue(""); }, [prompt?.id]);
function open(providerId: NativeSubscriptionProviderId) {
if (operation || statusesRef.current.some((s) => s.pending)) return;
@@ -62,7 +65,6 @@ export function NativeSubscriptionUIProvider({ children }: { children: React.Rea
function close() {
if (!dialog || busy) return;
if (confirm) { setConfirm(false); setError(null); return; }
const bridge = window.lh?.subscriptions;
if ((status?.pending || operation === "login") && bridge) { void control(() => bridge.cancel(dialog), true); return; }
generation.current++; setDialog(null); setError(null);
}
@@ -71,23 +73,21 @@ export function NativeSubscriptionUIProvider({ children }: { children: React.Rea
generation.current++; setError(null); setConfirm(false); setBackground(status.allowBackgroundUse); setBackgroundDirty(false); setView("notice");
}
async function login() {
const bridge = window.lh?.subscriptions;
if (!dialog || !bridge || operation || !status?.available) return;
const owner = ++generation.current; setOperation("login"); setError(null);
try { await bridge.login(dialog, backgroundDirty ? background : undefined); await refresh(); if (mounted.current && generation.current === owner) setDialog(null); }
catch { if (mounted.current && generation.current === owner) setError("Sign-in did not complete. Try connecting again."); }
finally { if (mounted.current && generation.current === owner) setOperation(null); }
}
const bridge = typeof window === "undefined" ? undefined : window.lh?.subscriptions;
return <Context.Provider value={{ statuses, busy: !!operation, open }}>
{children}
{dialog && descriptor && status && <ConnectionDialog key={`${dialog}:${view}:${confirm}`} label={`${descriptor.label} connection`} testid="subscription-connection-dialog" backdropTestid="subscription-dialog-backdrop" busy={busy} onClose={close}>
<h2 className="text-sm font-semibold text-foreground">{descriptor.label}</h2>
{confirm ? <>
<p className="mt-3 text-sm font-semibold text-red-400">Disconnect {descriptor.label}?</p>
<p className="mt-3 text-xs leading-relaxed text-muted-foreground">{subscriptionDisconnectCopy(descriptor.label)}</p>
<p className="mt-3 text-xs leading-relaxed text-muted-foreground">{subscriptionDisconnectCopy(descriptor.label, web)}</p>
</> : view === "notice" ? <>
<p className="mt-3 text-xs leading-relaxed text-muted-foreground" data-testid="subscription-privacy-notice">{subscriptionNotice(descriptor.label)}</p>
<p className="mt-3 text-xs leading-relaxed text-muted-foreground" data-testid="subscription-privacy-notice">{subscriptionNotice(descriptor.label, web)}</p>
{dialog === "github-copilot" && <p className="mt-3 text-xs leading-relaxed text-muted-foreground" data-testid="copilot-policy-notice">{COPILOT_POLICY_NOTICE}</p>}
</> : <>
<div className="mt-4 flex items-center gap-3 text-xs text-foreground">
@@ -119,16 +119,19 @@ export function NativeSubscriptionUIProvider({ children }: { children: React.Rea
</Context.Provider>;
}
export function NativeSubscriptionConnection({ providerId }: { providerId: NativeSubscriptionProviderId }) {
export function NativeSubscriptionConnection({ providerId, labelClassName }: { providerId: NativeSubscriptionProviderId; labelClassName?: string }) {
const ui = useContext(Context), descriptor = NATIVE_SUBSCRIPTIONS.find((p) => p.providerId === providerId)!;
const status = ui.statuses.find((s) => s.providerId === providerId) ?? emptySubscriptionStatus(providerId);
return <div className="flex flex-col gap-2" data-testid={`${providerId}-connection`}>
<ConnectionRow label={descriptor.label} action={status.connectionId ? "Manage…" : "Connect"} disabled={ui.busy || !status.available || ui.statuses.some((s) => s.pending)} onAction={() => ui.open(providerId)} />
<ConnectionRow label={descriptor.label} labelClassName={labelClassName} action={status.connectionId ? "Manage…" : "Connect"} disabled={ui.busy || !status.available || ui.statuses.some((s) => s.pending)} onAction={() => ui.open(providerId)} />
{status.error && <p role="alert" className="text-xs text-red-400">{status.error}</p>}
</div>;
}
export function SubscriptionConnections({ setup = false }: { setup?: boolean }) {
return <div className="space-y-3" data-testid="subscription-connections"><ChatGPTConnection setup={setup} />{NATIVE_SUBSCRIPTIONS.map((p) => <NativeSubscriptionConnection key={p.providerId} providerId={p.providerId} />)}</div>;
// Providers uses ordinary settings-label typography. The setup surface
// keeps its existing compact labels; dialogs and actions are untouched.
const labelClassName = setup ? undefined : "font-prose text-(length:--lh-prose-size) text-foreground";
return <div className="space-y-3" data-testid="subscription-connections"><ChatGPTConnection setup={setup} labelClassName={labelClassName} />{NATIVE_SUBSCRIPTIONS.map((p) => <NativeSubscriptionConnection key={p.providerId} providerId={p.providerId} labelClassName={labelClassName} />)}</div>;
}
export function NativeSubscriptionBackgroundAccess({ provider }: { provider: string | undefined }) {
const ui = useContext(Context), descriptor = NATIVE_SUBSCRIPTIONS.find((p) => p.logicalId === provider);
components/settings/update-settings.tsx
+113
View File
@@ -0,0 +1,113 @@
"use client";
import { useEffect, useState } from "react";
import { Button } from "@/components/ui/button";
import { BrailleSpinner } from "@/components/braille-spinner";
import { SettingsRow, Toggle } from "./controls.tsx";
/** Observational snapshot from electron/main.cjs; never updater permission. */
export type UpdateState = {
status: "idle" | "checking" | "downloading" | "up-to-date" | "downloaded" | "error" | "disabled";
version: string | null;
revision: number;
};
export const APP_VERSION = process.env.LH_VERSION ?? "";
const HOMEPAGE = process.env.LH_HOMEPAGE ?? "";
export const MANUAL_UPDATE_HINT = "Checking downloads available updates. Downloaded updates install when you quit the app.";
export const AUTO_UPDATE_HINT = "Check at startup and every 24 hours. Turning this off does not stop updates already in progress or ready to install.";
const UPDATE_ERROR = "Couldn’t check for updates. Try again later.";
/** Subscribe before requesting the snapshot. IPC replies and pushes can
* cross; a revision prevents an older reply from hiding a newer event.
* The banner shares this read-only observer so late loads recover too. */
export function useUpdateState() {
const [state, setState] = useState<UpdateState | null>(null);
const [failed, setFailed] = useState(false);
useEffect(() => {
const lh = window.lh;
if (!lh) return;
let live = true;
let revision = -1;
const accept = (next: UpdateState) => {
if (!live || next.revision < revision) return;
revision = next.revision;
setState(next);
setFailed(false);
};
const off = lh.onUpdateStatus?.(accept);
void lh.getUpdateState?.().then(accept).catch(() => {
if (live && revision === -1) setFailed(true);
});
return () => { live = false; off?.(); };
}, []);
return { state, failed };
}
export function updateStatusText(state: UpdateState | null): string | null {
switch (state?.status) {
case "checking": return "Checking for updates…";
case "downloading": return state.version ? `Downloading version ${state.version}…` : "Downloading an update…";
case "up-to-date": return "No newer version was found at the last check.";
case "downloaded": return `${state.version ? `Version ${state.version}` : "An update"} is ready. It will install when you quit the app.`;
case "error": return UPDATE_ERROR;
default: return null;
}
}
/** Version information is useful everywhere; update controls only belong
* to a shell with an available updater. Opening this row never checks. */
export function UpdateSettings({
autoUpdate,
onAutoUpdateChange,
version = APP_VERSION,
homepage = HOMEPAGE,
}: {
autoUpdate: boolean;
onAutoUpdateChange: (on: boolean) => void;
version?: string;
homepage?: string;
}) {
const { state, failed } = useUpdateState();
const [checkFailed, setCheckFailed] = useState(false);
useEffect(() => setCheckFailed(false), [state]);
const active = state !== null && state.status !== "disabled";
const busy = state?.status === "checking" || state?.status === "downloading";
const status = state?.status === "downloaded" ? updateStatusText(state)
: failed || checkFailed ? UPDATE_ERROR : updateStatusText(state);
const releaseNotes = version && homepage ? `${homepage}/releases/tag/v${version}` : null;
const check = () => {
setCheckFailed(false);
void window.lh?.checkForUpdates?.().catch(() => setCheckFailed(true));
};
return (
<>
<SettingsRow
label={`Version ${version || "unknown"}`}
testid="row-version"
hint={
<>
{status && <span className="block" role="status" data-testid="update-status">{status}</span>}
{active && <span className="block">{MANUAL_UPDATE_HINT}</span>}
{releaseNotes && <a href={releaseNotes} target="_blank" rel="noopener noreferrer" className="underline-offset-2 hover:text-foreground hover:underline" data-testid="release-notes">Release notes</a>}
</>
}
>
{active && (state.status === "downloaded" ? (
<Button type="button" size="xs" className="cursor-pointer" data-testid="update-restart" onClick={() => window.lh?.installUpdate?.()}>Restart</Button>
) : (
<Button type="button" size="xs" variant="outline" className="cursor-pointer disabled:cursor-not-allowed" data-testid="update-check" disabled={busy} aria-busy={busy} onClick={check}>
{busy && <BrailleSpinner label={state.status === "downloading" ? "Downloading update" : "Checking for updates"} />}
Check for updates
</Button>
))}
</SettingsRow>
{active && (
<SettingsRow label="Automatic updates" hint={AUTO_UPDATE_HINT} testid="row-auto-update">
<Toggle checked={autoUpdate} label="auto update" onChange={onAutoUpdateChange} />
</SettingsRow>
)}
</>
);
}
components/setup-cta.tsx
+30 -2
View File
@@ -15,9 +15,10 @@
* its own save), the same re-read pattern chat.tsx uses for "models".
*/
import type { AuthKeyStatus } from "@/src/core";
import { useEffect, useState } from "react";
import { MODEL_SLOTS, type AuthKeyStatus } from "@/src/core";
import { classifyKeys, type ProviderKeyInfo } from "@/components/settings/keys-section";
import { useModelCatalog } from "@/components/model-picker";
import { useModelCatalog } from "@/components/model-catalog";
import type { ModelCatalogResponse } from "@/src/core";
/** Pure: does a MODEL-PROVIDER key exist among the statuses? Reuses
@@ -32,6 +33,33 @@ export function hasProviderKey(
return classifyKeys(statuses, providers).providerRows.some((s) => s.present);
}
/** Completeness is saved identity, not current provider availability. */
export function savedSlotsConfigured(slots: ModelCatalogResponse["slots"] | undefined): boolean {
return MODEL_SLOTS.every((slot) => !!slots?.[slot]?.model);
}
/** Keep the last known completeness during connection invalidation, but
* follow ordinary slot repairs as soon as their shared snapshot arrives. */
export function useModelSetup(initialConfigured: boolean) {
const { catalog } = useModelCatalog();
const [lastConfigured, setLastConfigured] = useState(initialConfigured);
const configured = catalog ? savedSlotsConfigured(catalog.slots) : lastConfigured;
useEffect(() => { if (catalog) setLastConfigured(savedSlotsConfigured(catalog.slots)); }, [catalog]);
return { catalog, configured, zeroKeys: catalog ? !catalog.providers.some((p) => p.keyPresent) : false };
}
/** A known unavailable route needs an explicit replacement, never an
* automatic provider/model substitution. Unknown metadata defers to the
* server's admission gate instead of declaring a route disconnected. */
export function selectedModelUnavailable(catalog: ModelCatalogResponse | undefined, selected: string | undefined): boolean {
return !!catalog && !!selected && !catalog.models.some((m) => `${m.provider}:${m.id}` === selected);
}
/** Reuse the ordinary control (also opened by the existing keyboard action). */
export function openModelPicker(): void {
window.dispatchEvent(new CustomEvent("lh:kbd-open-model-picker"));
}
/** The zero-keys truth for the calling page: TRUE only after a
* successful read says no provider key exists (pending and failed
* reads render nothing — the pill must never flash on a gap, and a
components/setup-dialog.tsx
+123 -99
View File
@@ -13,7 +13,8 @@
* STAGED BY DEPENDENCY (owner, 2026-09-20: "first keys, then models —
* there's no point presenting the model picker before the user has set a
* key because there are no models to pick then"):
* Stage 1 — no keyed provider: a SELF-CONTAINED first-key screen
* Unknown catalog — neutral loading (or error/Retry), never a key screen.
* Stage 1 — no keyed provider or saved selection: a SELF-CONTAINED first-key screen
* (zero-keys wave, owner 2026-09-25): provider picker + one key
* input + save via PUT /api/keys/:id. The dialog STAYS MOUNTED —
* saving dispatches the "keys" signal, the fetch effect re-runs on
@@ -23,6 +24,9 @@
* Stage 2 — Strong → Cheap → Vision pickers over the FEDERATED catalog
* (keyed providers only), the vision slot filtered to image-capable
* models, in the app's shared dropdown language (ModelMenuList).
* Saved disconnected selections and edits are retained internally in
* the exclusive provider stage. Connecting with all slots already saved
* closes setup; otherwise the pickers require only the missing slots.
* All three required (owner: no skip, no defaults); Save PUTs the
* models section and reports onDone(models).
*
@@ -41,6 +45,8 @@ import { prettyModel } from "@/components/usage-format";
import { ModelMenuList } from "@/components/settings/model-menu";
import { MenuShell } from "@/components/menu-shell";
import { notifySettingsSaved, useSettingsSaved } from "@/components/settings-signal";
import { useModelCatalog } from "@/components/model-catalog";
import { savedSlotsConfigured } from "@/components/setup-cta";
import { Hint } from "@/components/ui/hint";
import { type ProviderKeyInfo } from "@/components/settings/keys-section";
import { ProviderMenuList } from "@/components/settings/model-menu";
@@ -73,27 +79,26 @@ interface ModelsBody {
* (owner's dependency ordering). Exported for tests. */
export function setupStage(
providers: { id: string; keyPresent: boolean }[] | null,
): "keys" | "pickers" {
if (providers === null) return "keys"; // fetch failed → the honest fallback
): "loading" | "keys" | "pickers" {
if (providers === null) return "loading";
return providers.some((p) => p.keyPresent) ? "pickers" : "keys";
}
/** Prefill (owner, BYOK wave 2026-09-22): each pinned tier opens with its
* pick — matched in the FEDERATED catalog (a lapsed key's pin does not
* prefill); unpinned tiers stay empty and carry the "needed" mark.
* First-run = nothing pinned → all three empty + marked. Exported pure
* for tests. */
type ModelSelection = Pick<ModelCatalogEntry, "provider" | "id">;
/** Saved identity is independent of selectable metadata. A disconnected
* pin still prefills; only the menu is restricted to the live catalog. */
export function prefillFromSlots(
models: ModelCatalogEntry[],
slots: Record<ModelSlotName, { model: { provider: string; id: string } } | null> | undefined,
): Partial<Record<ModelSlotName, ModelCatalogEntry>> {
const pre: Partial<Record<ModelSlotName, ModelCatalogEntry>> = {};
): Partial<Record<ModelSlotName, ModelSelection>> {
const pre: Partial<Record<ModelSlotName, ModelSelection>> = {};
for (const slot of SLOT_ORDER) {
const pin = slots?.[slot];
const entry = pin?.model
? models.find((m) => m.provider === pin.model!.provider && m.id === pin.model!.id)
: undefined;
if (entry) pre[slot] = entry;
if (pin?.model) pre[slot] = entry ?? pin.model;
}
return pre;
}
@@ -217,24 +222,38 @@ export function SetupKeysBody({
);
}
export function SetupDialog({
open,
onClose,
/** Saved: the models section landed; the caller marks configured (the
* draft stays put — save never sends). */
onDone,
}: {
function prefillLevels(body: ModelsBody | undefined): Partial<Record<ModelSlotName, ThinkingLevel>> {
const levels: Partial<Record<ModelSlotName, ThinkingLevel>> = {};
for (const slot of SLOT_ORDER) {
const pin = body?.slots?.[slot];
const entry = body?.models.find((m) => m.provider === pin?.model.provider && m.id === pin.model.id);
if (pin?.thinkingLevel !== undefined) levels[slot] = pin.thinkingLevel;
else if (entry) levels[slot] = prefillLevelFromPin(entry, pin);
}
return levels;
}
interface SetupDialogProps {
open: boolean;
onClose: () => void;
/** Saving never sends the caller's draft. */
onDone: (models: ModelsSettings) => void;
}) {
}
export function SetupDialog(props: SetupDialogProps) {
// An open owns its draft. Mount-time cache seeding avoids a key-screen
// first frame, and reopening cannot reuse a previous open's edits.
return props.open ? <OpenSetupDialog {...props} /> : null;
}
function OpenSetupDialog({ onClose, onDone }: SetupDialogProps) {
const backdropDismiss = useBackdropDismiss(onClose);
const [body, setBody] = useState<ModelsBody | null>(null);
const [failed, setFailed] = useState(false);
const [picked, setPicked] = useState<Partial<Record<ModelSlotName, ModelCatalogEntry>>>({});
/** Per-slot thinking level (chunk 2): populated from the ordering on
* every fresh model pick, editable — an explicit pick always wins. */
const [levels, setLevels] = useState<Partial<Record<ModelSlotName, ThinkingLevel>>>({});
const { catalog, failed, refresh } = useModelCatalog();
const providerStageSeen = useRef(false);
const connectionCloseRequested = useRef(false);
const [picked, setPicked] = useState(() => prefillFromSlots(catalog?.models ?? [], catalog?.slots));
const [levels, setLevels] = useState(() => prefillLevels(catalog));
const editedSlots = useRef(new Set<ModelSlotName>());
const [saving, setSaving] = useState(false);
const [error, setError] = useState<string | null>(null);
const [openMenu, setOpenMenu] = useState<{ slot: ModelSlotName; kind: "model" | "thinking" } | null>(null);
@@ -249,64 +268,65 @@ export function SetupDialog({
const [keyError, setKeyError] = useState<string | null>(null);
const [keyMenuOpen, setKeyMenuOpen] = useState(false);
// Stage 1's provider-picker trigger wrapper — the shared shell's anchor.
// Declared with the other hooks: the early `if (!open) return null` sits
// below them all. (Stage 2's slot menus anchor by anchorTestid instead —
// see the slot MenuShell below for why a ref cannot work there.)
// Stage 2's slot menus anchor by anchorTestid instead — see the slot
// MenuShell below for why a freshly attached ref cannot work there.
const keyTriggerRef = useRef<HTMLDivElement>(null);
// The keys signal (zero-keys wave): keys-section saves/deletes AND
// THIS dialog's own first-key save dispatch it; the fetch effect below
// re-runs on it, so saving a key flips the stage to the pickers with
// the widened catalog — no restart, no Settings round-trip.
// re-runs on it, so saving a key widens the catalog. Incomplete saved
// slots advance to pickers; complete saved slots just close setup.
const [keysTick, setKeysTick] = useState(0);
useSettingsSaved((section) => {
if (section === "keys") setKeysTick((t) => t + 1);
});
// Fresh /api/models per open AND per keys change — a key saved a
// moment ago (stage 1's own screen) widens the pickers immediately.
useEffect(() => {
if (!open) return;
let cancelled = false;
setBody(null);
setPicked({});
setLevels({});
setError(null);
fetch("/api/keys")
.then((r) => r.json())
.then((b: { providers?: ProviderKeyInfo[] }) => { if (!cancelled) setKeyProviders(b.providers ?? []); })
.catch(() => { if (!cancelled) setKeyProviders([]); });
fetch("/api/models")
.then((r) => { if (!r.ok) throw new Error("Catalog unavailable"); return r.json(); })
.then((b: ModelsBody) => {
if (cancelled) return;
setBody(b);
setFailed(false);
// Prefill (owner, 2026-09-22): each pinned tier opens with its
// pick (matched in the FEDERATED catalog — a lapsed key's pin
// does not prefill); unpinned tiers stay empty and carry the
// "needed" mark. First-run = nothing pinned → all three marked.
const pre = prefillFromSlots(b.models, b.slots);
setPicked(pre);
// Levels prefill with the pins (chunk 2): stored level if still
// supported, else the ordering's level for the pinned model.
const preLevels: Partial<Record<ModelSlotName, ThinkingLevel>> = {};
for (const slot of SLOT_ORDER) {
const entry = pre[slot];
if (entry) preLevels[slot] = prefillLevelFromPin(entry, b.slots?.[slot]);
}
setLevels(preLevels);
})
.catch(() => { if (!cancelled) setFailed(true); });
return () => { cancelled = true; };
}, [open, keysTick]);
}, [keysTick]);
useEffect(() => {
if (!catalog) return;
const pre = prefillFromSlots(catalog.models, catalog.slots);
const preLevels = prefillLevels(catalog);
// Fresh metadata is always used, but a response cannot reset an
// edited slot, including a level chosen while the request was pending.
setPicked((prev) => {
const next = { ...prev };
for (const slot of SLOT_ORDER) if (!editedSlots.current.has(slot)) {
if (pre[slot]) next[slot] = pre[slot]; else delete next[slot];
}
return next;
});
setLevels((prev) => {
const next = { ...prev };
for (const slot of SLOT_ORDER) if (!editedSlots.current.has(slot)) {
if (preLevels[slot] !== undefined) next[slot] = preLevels[slot]; else delete next[slot];
}
return next;
});
}, [catalog]);
const stage = setupStage(catalog?.providers ?? null);
const closeAfterConnection = stage === "pickers" && providerStageSeen.current && savedSlotsConfigured(catalog?.slots);
useEffect(() => {
if (stage === "keys") providerStageSeen.current = true;
if (closeAfterConnection && !connectionCloseRequested.current) {
connectionCloseRequested.current = true;
onClose(); // Connection recovery is not a model-settings save.
}
}, [stage, closeAfterConnection, onClose]);
// Esc + outside-click — layered dismissal, closable WITHOUT
// configuring (browsing stays possible; the gate is the send
// interception, not this dialog). An open slot menu eats Esc first
// (and any click outside it); only a bare Esc closes the dialog.
useEffect(() => {
if (!open) return;
const onKey = (e: KeyboardEvent) => {
if (e.key !== "Escape") return;
// THE ONE-ESC LAW (owner, keyboard round 2026-10-06): claim at
@@ -341,17 +361,18 @@ export function SetupDialog({
document.removeEventListener("keydown", onKey);
document.removeEventListener("pointerdown", onPointerDown);
};
}, [open, onClose, openMenu, keyMenuOpen]);
}, [onClose, openMenu, keyMenuOpen]);
if (!open) return null;
// Do not briefly paint model pickers between successful connection and
// closing an already-configured setup. No saved model/thinking writes.
if (closeAfterConnection) return null;
const stage = failed ? "keys" : setupStage(body ? body.providers : null);
const allPicked = SLOT_ORDER.every((s) => picked[s]);
/** Stage 1's save: PUT the pasted key, dispatch the keys signal, and
* CLEAN the local screen — the signal re-fetch widens the catalog and
* the stage flips to the pickers on its own; the dialog stays mounted
* the whole way. Shape-gate 400s keep their user-facing line inline. */
* CLEAN the local key field. The fresh catalog decides whether to
* close recovery or advance to the missing model slots. Shape-gate
* 400s keep their user-facing line inline. */
async function saveKey() {
if (!keyProvider || !keyValue.trim() || keySaving) return;
setKeySaving(true);
@@ -398,6 +419,20 @@ export function SetupDialog({
onDone(models);
}
const keysBody = <SetupKeysBody
providers={keyProviders}
provider={keyProvider}
value={keyValue}
saving={keySaving}
error={keyError}
menuOpen={keyMenuOpen}
menuAnchor={keyTriggerRef}
onToggleMenu={() => setKeyMenuOpen((v) => !v)}
onPick={(p) => { setKeyProvider(p); setKeyMenuOpen(false); setKeyError(null); }}
onValue={setKeyValue}
onSave={() => void saveKey()}
/>;
return createPortal(
<div
className="fixed inset-0 z-50 flex items-center justify-center bg-background/80"
@@ -427,32 +462,15 @@ export function SetupDialog({
</Hint>
</header>
{stage === "keys" ? (
<div className="flex flex-col px-6 py-5">
{failed ? (
<p className="text-sm text-foreground">
Model catalog could not be loaded. Check the connection and retry.
</p>
) : (
<SetupKeysBody
providers={keyProviders}
provider={keyProvider}
value={keyValue}
saving={keySaving}
error={keyError}
menuOpen={keyMenuOpen}
menuAnchor={keyTriggerRef}
onToggleMenu={() => setKeyMenuOpen((v) => !v)}
onPick={(p) => {
setKeyProvider(p);
setKeyMenuOpen(false);
setKeyError(null);
}}
onValue={setKeyValue}
onSave={() => void saveKey()}
/>
)}
</div>
{failed && (
<p role="alert" data-testid="setup-catalog-error" className="px-6 py-3 text-sm text-foreground">
Model catalog could not be loaded. Check the connection and retry. <button type="button" data-testid="setup-retry" className="underline" onClick={refresh}>Retry</button>
</p>
)}
{stage === "loading" ? (
!failed && <p data-testid="setup-stage-loading" role="status" className="px-6 py-5 text-xs text-muted-foreground">Loading models…</p>
) : stage === "keys" ? (
<div className="flex flex-col px-6 py-5">{keysBody}</div>
) : (
<div className="flex flex-col gap-4 overflow-y-auto px-6 py-5" data-testid="setup-stage-pickers">
<p className="text-xs text-muted-foreground">
@@ -461,14 +479,16 @@ export function SetupDialog({
{SLOT_ORDER.map((slot) => {
const entry = picked[slot];
const level = levels[slot];
const levelsForEntry = entry?.supportedLevels ?? [];
const catalog =
slot === "vision" ? body?.models.filter((m) => m.input.includes("image")) : body?.models;
const metadata = catalog?.models.find((m) => m.id === entry?.id && m.provider === entry.provider);
// Unavailable metadata cannot erase or clamp a saved preference.
const levelsForEntry = metadata?.supportedLevels ?? (entry ? Object.keys(THINKING_LABEL) : []);
const available = catalog?.models ?? (failed ? [] : undefined);
const slotCatalog = slot === "vision" ? available?.filter((m) => m.input.includes("image")) : available;
return (
<div key={slot} data-testid={`setup-slot-${slot}`} className="flex flex-col gap-1.5">
<span className="flex items-center gap-1.5">
<span className="font-ui text-sm text-foreground">{SLOT_META[slot].label}</span>
{!body?.slots?.[slot]?.model && !picked[slot] && (
{!catalog?.slots?.[slot]?.model && !picked[slot] && (
<span
data-testid={`setup-needed-${slot}`}
className="rounded-sm border border-border px-1 py-0.5 text-[10px] font-normal tracking-wide text-muted-foreground"
@@ -481,6 +501,7 @@ export function SetupDialog({
{slot !== "strong" && <><ChatGPTBackgroundAccess provider={entry?.provider} /><NativeSubscriptionBackgroundAccess provider={entry?.provider} /></>}
<div className="flex items-center gap-2">
<div className="relative flex-1">
<Hint label={entry ? `${entry.provider}:${entry.id}` : "Pick a model…"}>
<button
type="button"
aria-label={`${SLOT_META[slot].label} model`}
@@ -494,6 +515,7 @@ export function SetupDialog({
▾
</span>
</button>
</Hint>
{/* Anchor by TESTID, not a ref (the settings-section pattern):
React attaches a host element's ref AFTER its descendants'
layout effects run, so a ref freshly attached in this same
@@ -512,10 +534,11 @@ export function SetupDialog({
menuClass="flex w-[22rem] flex-col rounded-xl border border-border bg-card p-2 text-xs shadow-xl"
>
<ModelMenuList
models={catalog}
models={slotCatalog}
effectiveId={entry?.id}
effectiveProvider={entry?.provider}
onPick={(m) => {
editedSlots.current.add(slot);
setPicked((prev) => ({ ...prev, [slot]: m }));
// Auto-populate (chunk 2, owner: "populate
// everywhere"): no explicit level → the
@@ -527,7 +550,7 @@ export function SetupDialog({
}));
setOpenMenu(null);
}}
footer={catalog && catalog.length === 0 ? "No image-capable models. Add a provider key with access to a vision model." : undefined}
footer={slotCatalog && slotCatalog.length === 0 && slot === "vision" ? "No image-capable models. Add a provider key with access to a vision model." : undefined}
/>
</MenuShell>
)}
@@ -563,6 +586,7 @@ export function SetupDialog({
levels={levelsForEntry.map(String)}
effectiveLevel={level}
onPick={(lv) => {
editedSlots.current.add(slot);
setLevels((prev) => ({ ...prev, [slot]: lv as ThinkingLevel }));
setOpenMenu(null);
}}
components/sidebar.tsx
+10 -2
View File
@@ -49,6 +49,7 @@ const STORAGE_KEY = "lh:sidebar-open";
* Mirrors SIDEBAR_AUTOCLOSE in electron/main.cjs — keep in sync.
*/
import type { ChatGPTControlBridge, NativeSubscriptionControlBridge } from "@/src/core";
import type { UpdateState } from "@/components/settings/update-settings.tsx";
const SIDEBAR_AUTOCLOSE = 928;
// Module scope, guarded for SSR (the server has no navigator).
@@ -60,15 +61,22 @@ declare global {
lh?: {
/** Read-only host OS; determines the native-control gutter. */
platform?: string;
/** External Next desktop-dev uses the server's web connections. */
subscriptionTransport?: "desktop" | "web";
openSubscriptionBrowser?: (action: string) => Promise<void>;
chatgpt?: ChatGPTControlBridge;
subscriptions?: NativeSubscriptionControlBridge;
expandWindow: () => void;
/** The zoom truth (owner round 7): subscribes to "lh:zoom" pushes; returns unsubscribe. */
onZoom?: (cb: (factor: number) => void) => () => void;
getZoom?: () => Promise<number>;
/** Auto-update channel (Wave C, 2026-10-10): "lh:update-status" pushes once a staged update exists. */
onUpdateStatus?: (cb: (status: { status: string; version: string | null }) => void) => () => void;
/** Observational update events and a snapshot for late subscribers. */
onUpdateStatus?: (cb: (status: UpdateState) => void) => () => void;
getUpdateState?: () => Promise<UpdateState>;
/** Explicit manual check also downloads and stages available updates. */
checkForUpdates?: () => Promise<void>;
installUpdate?: () => void;
onOpenSettings?: (cb: (section: string) => void) => () => void;
};
}
}
components/start-session.ts
+34
View File
@@ -0,0 +1,34 @@
import { clearOpeningPrompt, OPENING_PROMPT_FAILURE, stashOpeningPrompt } from "./opening-prompt";
interface SessionCreation {
title: string;
kind?: "practice";
graphContext?: { propositions: readonly { id: string; name: string; statement: string }[] };
}
/** Per source-composer instance. Retain a successful creation if storage
* fails, so an unchanged retry does not create another empty session.
* Changed text/context/settings is a different start, not a stale title.
*/
export class SessionStarter {
private created?: { identity: string; id: string };
async start(body: SessionCreation, fetchLike: typeof fetch, settingsIdentity?: string): Promise<string> {
const identity = JSON.stringify([body, settingsIdentity]);
if (this.created?.identity !== identity) {
if (this.created) clearOpeningPrompt(this.created.id);
this.created = undefined;
const res = await fetchLike("/api/chats", {
method: "POST",
headers: { "Content-Type": "application/json" },
body: JSON.stringify(body),
});
if (!res.ok) throw new Error(OPENING_PROMPT_FAILURE);
const result = await res.json() as { id?: unknown };
if (typeof result.id !== "string" || !/^[A-Za-z0-9_-]+$/.test(result.id)) throw new Error(OPENING_PROMPT_FAILURE);
this.created = { identity, id: result.id };
}
if (!stashOpeningPrompt(this.created.id, body.title)) throw new Error(OPENING_PROMPT_FAILURE);
return `/chat/${this.created.id}`;
}
}
components/stats-bar.tsx
+9 -2
View File
@@ -1,6 +1,6 @@
"use client";
import { useEffect, useMemo, useRef, useState, type ReactNode } from "react";
import { useCallback, useEffect, useMemo, useRef, useState, type ReactNode } from "react";
import { subscriptionLabels, subscriptionLabel, type SessionPricing, type TurnUsage } from "@/src/core";
import { formatTokens, formatUsd, planCostLabel, prettyModel } from "@/components/usage-format";
import { ModelControl, catalogLimits, useModelCatalog } from "@/components/model-picker";
@@ -34,6 +34,7 @@ export function StatsBar({
dictationSlot,
chatId,
leadingSlot,
onModelChange,
}: {
turns: TurnUsage[];
contextTokens?: number | null;
@@ -43,6 +44,8 @@ export function StatsBar({
* record; absent on the draft surfaces, where it binds the strong
* slot (the new-session default). */
chatId?: string;
/** The same effective identity the visible model control owns. */
onModelChange?: (model: string | undefined) => void;
/** The dictation strip (STT wave, 2026-09-17): the row's middle is dead
* space between the session data and the model control — the mic
* button lives there, and while recording the meter unfurls to fill
@@ -55,6 +58,10 @@ export function StatsBar({
}) {
const [open, setOpen] = useState(false);
const [selectedModel, setSelectedModel] = useState<string | undefined>();
const modelChanged = useCallback((model: string | undefined) => {
setSelectedModel(model);
onModelChange?.(model);
}, [onModelChange]);
// Same dropdown discipline as the model control (owner, df4433f treatment,
// stats round 1): close on any pointerdown outside the bar or on Escape.
const barRef = useRef<HTMLDivElement>(null);
@@ -192,7 +199,7 @@ export function StatsBar({
{/* The model + thinking control — one controller for the conversation
agents (prober/teacher); the record's `turns` above still folds
the ACTUAL models billed, regardless of what the picker says. */}
<ModelControl chatId={chatId} onModelChange={setSelectedModel} />
<ModelControl chatId={chatId} onModelChange={modelChanged} />
</div>
);
}
components/update-banner.tsx
+5 -18
View File
@@ -3,32 +3,19 @@
/**
* The auto-update banner (Wave C, owner 2026-10-10: silent download +
* restart-banner UX). The shell pushes "lh:update-status" once a staged
* update exists (electron/main.cjs update-downloaded); this is the whole
* renderer surface — a small fixed card, bottom-right, offering Restart.
* update exists (electron/main.cjs update-downloaded). Snapshot retrieval
* also covers a late load. The card offers Restart alongside Settings.
* Browser builds have no window.lh and never see a push: renders null.
*
* No dismiss: a downloaded update installs on quit anyway
* (autoInstallOnAppQuit); the banner is a courtesy, not a gate.
*/
import { useEffect, useState } from "react";
// The window.lh bridge type lives in components/sidebar.tsx (the incumbent
// declare-global) — onUpdateStatus/installUpdate are declared there too.
type UpdateStatus = { status: string; version: string | null };
import { useUpdateState } from "@/components/settings/update-settings.tsx";
export default function UpdateBanner() {
const [ready, setReady] = useState<UpdateStatus | null>(null);
useEffect(() => {
// Absent bridge (browser / dev) → stays hidden forever.
return window.lh?.onUpdateStatus?.((s) => {
if (s?.status === "downloaded") setReady(s);
});
}, []);
if (!ready) return null;
const { state: ready } = useUpdateState();
if (ready?.status !== "downloaded") return null;
return (
<div
data-testid="update-banner"
components/usage-format.ts
+5 -19
View File
@@ -29,25 +29,11 @@ export function splitQualified(id: string): { provider: string | null; id: strin
return { provider: id.slice(0, colon), id: id.slice(colon + 1) };
}
/**
* "z-ai/glm-5.3-flash" → "glm-5.3-flash" — vendor prefix dropped, the
* name KEPT EXACTLY as its provider writes it (dashes stay, no
* auto-caps — owner rounds 4-5, 2026-10-05); a trailing all-digit
* segment (OpenRouter's date suffix, "qwen3.8-max-0902") is dropped.
* Quiet-chip friendly.
*
* With `qualified` (usage stamps are "provider:id", model-slots wave):
* the provider rides as a quiet suffix — "glm-5.3-flash · openrouter" —
* because with federation the same model via two providers is genuinely
* different bookkeeping. Bare catalog ids are unchanged.
*/
/** Model-only labels omit the vendor path, never versions, dates or variants.
* Only usage callers pass `qualified`: a bare catalog id's colon belongs to
* the model (e.g. :free), not a provider. Usage retains its provider suffix. */
export function prettyModel(id: string, qualified = false): string {
const { provider, id: bare } = splitQualified(id);
const { provider, id: bare } = qualified ? splitQualified(id) : { provider: null, id };
const short = bare.includes("/") ? bare.slice(bare.lastIndexOf("/") + 1) : bare;
const parts = short.split("-");
const trimmed = /^\d+$/.test(parts[parts.length - 1] ?? "") ? parts.slice(0, -1) : parts;
// No auto-capitalization, dashes KEPT (owner rounds 4-5, 2026-10-05):
// model names render exactly as their providers write them.
const pretty = trimmed.join("-");
return qualified && provider ? `${pretty} · ${provider}` : pretty;
return provider ? `${short} · ${provider}` : short;
}
docs/architecture.md
+25 -15
View File
@@ -1,6 +1,6 @@
# Learning Harness — Architecture
Status: **the canonical architecture spec** — describes the as-built system, verified against the code by a full audit on 2026-09-18 (every file under `src/`, `app/`, `components/`, `electron/`, `scripts/`, the lint config, and the test suite; `lint` and `typecheck` clean on the audited tree). The pre-audit spec is archived at [`archive/architecture-2026-09-08.md`](archive/architecture-2026-09-08.md) — history only. Product reasoning lives in [`docs/vision.md`](vision.md); file-by-file layout lives in the README.
Status: **the canonical architecture spec** — describes the as-built system, verified against the code by a full audit on 2026-09-18 (every file under `src/`, `app/`, `components/`, `electron/`, `scripts/`, the lint config, and the test suite; `lint` and `typecheck` clean on the audited tree). The pre-audit spec is archived at [`docs/archive/architecture-2026-09-08.md`](archive/architecture-2026-09-08.md) — history only. Product reasoning lives in [`docs/vision.md`](vision.md); file-by-file layout lives in the README.
**What this is.** A standalone learning app (web + Electron desktop): the learner's knowledge accumulates into one persistent, cross-session knowledge graph, built from tutoring sessions in which a teacher agent authors a formal, visible dependency tree (the plan tree) and verifies understanding with structured questions. The graph is the home page; the session is the classroom; everything else serves those two.
@@ -307,26 +307,34 @@ sequenceDiagram
- **The prober→teacher handoff**: the first turn(s) run the prober (floor/ceiling via quiz, goal via ask); `probe_done` (server-executed) files the assessment and queues a handoff steer; from the first `tool-probe_done` part in the transcript, the handler builds the teacher agent for every later turn. Phases (`core/phases.ts`) fold strictly: probe → plan → teach.
- **The formal plan tree is author-once**: the teacher authors the session's dependency map with a single schema-validated `plan_init` tool (a mermaid-subset graph, parsed deterministically with line-numbered errors, validated as a DAG). Progress is never mutated — statuses (`current`/`visited`/upcoming) are **derived** from tagged quiz calls in the transcript; the tree panel and the plan blocks render that derivation.
- **Job tools are server-executed, ack-and-steer**: `research`/`demo`/`visual` resolve immediately mid-turn with a fire-and-forget ack (the teacher keeps teaching); `spawn(job, input)` fires the background runner.
- Each turn service is a named, error-contained seam taking explicit deps; failures log one `[chat <id>]` line and never touch the response. Swapping any of them touches nothing else.
- Turn services are named seams taking explicit dependencies. Extraction contains its errors, but **canonical persistence must not fail silently**: a failed input write refuses acceptance without starting an agent; `persistTurn` throws on a failed reply write. The runner reports settlement failure before `[DONE]` and releases liveness even when normal settlement fails. Steer reconciliation uses the durable record, so a later bookkeeping failure cannot replay already-saved learner input.
- **Late attachment must still see settlement failure.** A small, memory-only notice per session precedes the settled stream terminal when the failing turn has already left the live registry. The next accepted turn, successful session deletion, or server exit clears it. Reopening seeds the existing banner from that notice and waits for ordinary learner input; dismissing the banner is not permission to automatically retry the model. No synthetic failure entry is added to the saved transcript.
- **Recovery never blindly resends an answer.** Submitted question drafts survive optimistic unmount in memory; the existing Submit action reconciles uncertain acceptance against the canonical record and running-turn state before sending again. EOF or a rejected stream read without a terminal triggers one read-only reattachment; another transport failure shows the existing error banner rather than claiming completion. Explicit completion, cancellation, and the literal-final-entry saved-error rule remain unchanged.
### 5.1 The wake pipeline (Workbench → Classroom)
On sub-agent completion the runner builds a **steer message** — a user-role `UIMessage` whose text part is the model-facing result and whose `data-subagent-artifact` part carries the full result for the client (provider conversion drops data parts) — and hands it to the wake pipeline:
**Worker completion and session delivery are separate lifecycles.** Dispatch persists the tool input on the job before spawning. On completion, the runner saves its result, terminal worker status (`done`/`error`), and `delivery: {kind: "result", status: "pending"}` in one job-record update. It then queues the native custom **steer message** without awaiting delivery. Its content is model-facing text; details carry the UI artifact. Worker resources are released and the Background agents row disappears even if an unanswered question holds the result indefinitely. Dispatch cards say “finished,” not “delivered.” Only computing workers can be aborted; abort saves an independent pending abort notice.
```mermaid
flowchart TB
R["job runner finishes sub-agent"] -->|"deliver(steer) — injectable seam"| ENQ["enqueue steer in per-chat queue (globalThis)"]
ENQ --> PUMP["pump(chatId) — one loop per chat (pumping set serializes)"]
PUMP --> GATE{"turn-free AND no pending question? (400ms poll)"}
GATE -->|no| PUMP
GATE -->|yes| REFRESH["re-fetch fresh record, re-check steerDelivered (claim race)"]
REFRESH --> RUN["turnRunner(chatId, record + steer) — the SAME handleChatRequest path"]
RUN -->|"200"| FIN["finalizeSteer → finalizeJob (done/error)<br/>— status flips only after the wake turn persists"]
RUN -->|"409"| BACKOFF["backoff 500→1000→2000→5000ms → retry"]
CRASH["crash recovery: listJobs → recoverStaleJobs (stale >15min)<br/>re-deliver results / respawn via setJobRespawn"] --> PUMP
R["save worker outcome + pending delivery; release worker"] --> ENQ["one registry entry per steer ID, including wake cargo"]
ENQ --> LIVE{"live agent and background consent?"}
LIVE -->|yes| STEER["pi steering at native boundaries; answer run takes parked results"]
LIVE -->|no| GATE["wait for turn-free, question-free, unpaused, consented session"]
GATE --> RUN["idle wake through the normal turn composition"]
RUN -->|"409"| BACKOFF["existing backoff 500→1000→2000→5000ms"]
BACKOFF --> GATE
RUN -->|"403"| HOLD["retain ownership until consent permits delivery"]
RUN -->|"other failed start"| QUIET["quiet pending; next learner run or restart recovery"]
RUN --> SAVE["canonical marker persistence, not merely HTTP 200"]
STEER --> SAVE
SAVE --> FIN["acknowledge delivery; worker outcome unchanged"]
CRASH["lazy recovery + live-agent registration restore known pending jobs"] --> ENQ
```
Job status flips done/error only after the wake turn persists. Delivery is **exactly-once**: deterministic steer ids (`steer_<jobId>`, `steer_probe_<chatId>`), `steerDelivered` checks at every claim point, and `claimSteers` finalizing in-transcript steers instead of re-appending. The client adopts the wake turn through the same `turnInFlight` machinery as returning mid-stream (busy shimmer, composer gate, append-only `mergeAdopted` adoption) and renders steers as artifacts — never a bubble.
The per-chat registry retains ownership across live handoff, idle-wake flight, and 409/403 return; repeated submissions share one entry/promise. A deterministic marker ID in the **saved conversation** completes delivery. Streaming visibility alone does not. A failed bookkeeping acknowledgement cannot replay a saved marker, and later teaching/model failure does not reopen delivery or retry the accepted turn. Failed starts leave results durably pending without a new error UI or retry timer; the in-memory entry stays parked until a learner run or process restart. Normal question/dismissal waits and consent deferrals are not failures.
Recovery reads saved job input before legacy transcript lookup. Missing research questions fail honestly without invoking the researcher. Legacy running jobs with saved results become finished workers with pending delivery; legacy terminal jobs without delivery metadata remain untouched, rather than guessing that old outcomes should be replayed. Pending result and abort notices are restored before a newly registered agent's first steering poll, even without a preceding jobs-feed read. Deleting a session discards its pending delivery obligations; only live turns and computing workers block deletion. These guarantees use the existing JSON storage and single-process delivery registry, not a transactional queue or a promise of exactly-once model execution.
Sub-agents **verify, never one-shot**: the demo-builder headlessly executes its HTML in system Chrome before delivery; the visual-maker renders each SVG and looks at it (needs a vision-capable model — the `vision` slot in settings).
@@ -423,9 +431,11 @@ The Electron main is a launcher + credential/security guard, never a host for te
### 8.1 Subscription connection boundary
`electron/chatgpt.cjs` retains the existing verified-account ChatGPT authentication, storage and revocation behavior. `electron/subscriptions.cjs` delegates GitHub Copilot/Grok login and refresh to the installed Pi SDK through its public provider/Models interfaces. Their single saved connections live independently in owner-only, atomically written `github-copilot-credentials.json` / `xai-credentials.json` beside the existing app credential files. Electron main alone owns durable OAuth credentials and serialized refresh; no global Pi files or API-key credential precedence are involved.
`electron/chatgpt.cjs` retains the existing verified-account ChatGPT authentication, storage and revocation behavior. `electron/subscriptions.cjs` delegates GitHub Copilot/Grok login and refresh to the installed Pi SDK through its public provider/Models interfaces. Their single saved connections live independently in owner-only, atomically written `github-copilot-credentials.json` / `xai-credentials.json` beside the existing app credential files. In packaged desktop mode Electron main alone owns durable OAuth credentials and serialized refresh; no global Pi files or API-key credential precedence are involved.
The renderer gets projected status, public device user codes, prompts and controls over origin-pinned IPC. The privately launched server child receives credential-free catalogs and scoped request auth (including native headers/account endpoints), never refresh tokens. `src/agent/subscription-bridge.ts` rejects framework-only IPC and invalidates stale connection/catalog scopes. `subscription-stream.ts` dispatches directly through native providers, retaining native protocol identity on the wire and restoring logical app identity in assistant history. Persisted routes `chatgpt`, `github-copilot-subscription` and `xai-subscription` are distinct from ordinary API routes and never fall back to them.
The packaged desktop renderer gets projected status, public device user codes, prompts and controls over origin-pinned IPC. The privately launched server child receives credential-free catalogs and scoped request auth (including native headers/account endpoints), never refresh tokens. `src/agent/subscription-bridge.ts` rejects framework-only IPC and invalidates stale connection/catalog scopes. `subscription-stream.ts` dispatches directly through native providers, retaining native protocol identity on the wire and restoring logical app identity in assistant history. Persisted routes `chatgpt`, `github-copilot-subscription` and `xai-subscription` are distinct from ordinary API routes and never fall back to them.
For ordinary Next web/dev, `src/server/web-subscriptions.ts` composes these same plain-Node controllers into one lazy process-global credential owner beside configured `AUTH_FILE`. The server injects the private transport through `src/agent/subscription-runtime.ts`; the agent layer never imports the server layer. Existing scoped leases, authority revisions and foreground/background abort listeners work through that transport, without exposing an HTTP token/lease operation. `src/server/web-subscription-http.ts` exposes only same-origin status/control operations. Generic `subscriptions-changed` notifications use the existing SSE notify→fetch channel; opaque one-use browser actions redirect to provider pages without putting authorization URLs in app JSON. Browser adapters preserve the UI bridge contracts; Electron dev uses the web owner with a narrow system-browser opener. Desktop markers fail closed rather than falling back to web credentials. This is the existing shared developer-server trust model, not user authentication or DNS-rebinding protection; LAN behavior is unchanged. Only one server process may own an auth directory. See [development](development.md#subscription-connections-in-webdev) for storage, browser locality and access limitations.
`subscription-request-cycle.ts` is the shared admission/automatic-input recovery state machine; the old ChatGPT module is its compatibility export. Background consent is independent per provider, default ON, with foreground-preserving cancellation and queue ownership retained on no-dispatch. Native usage rejected before dispatch is transiently excluded from helper usage collectors, without adding persisted provenance fields. Catalogs and read-time pricing preserve the selected connection: subscription use displays compact provider labels, API dollars remain dollars, and mixed usage joins both without estimating subscription dollars or claiming detected tiers/allowances.
docs/compaction-accounting.md
+11
View File
@@ -24,6 +24,17 @@ prompt, plan continuity, persistence, threshold, scheduling, and authorization
lifecycle. It does not adopt Pi's coding summary or separate split-turn summary
calls. A cut that drops no active history does not launch a summarizer.
When the active cutoff is applied, conversation requests carry the latest summary
once, followed by the retained transcript without older compaction markers. When
a subsequent compaction supplies that prior summary separately, it includes it
once, plus the newly superseded transcript without summary markers. All markers
and original messages remain stored. The existing full-history fallback for an
unresolved cutoff is unchanged, including its summary text.
A provider response stopped at its output limit (`length`) is an incomplete
summary, not a successful checkpoint. Compaction fails without changing the
existing history or prior summary; a later attempt can retry normally.
## Conversation meter
The numerator is the latest qualifying completed conversation response's native
docs/development.md
+20 -9
View File
@@ -2,13 +2,13 @@
The developer README (split out of the storefront README, 2026-10-10). Everything below is for people working ON LearnHarness, not using it.
A standalone learning app (web + Electron desktop): a **teacher agent** (Next.js; the Vercel AI SDK's `UIMessage` protocol end-to-end, LLM engine [pi-ai](docs/pi-engine-swap.md) via OpenRouter), probe→teach lessons with graded quiz/ask interactions, a formal **session plan tree** the teacher authors once (`plan_init`) and whose progress is derived from tagged quiz calls, and a persistent cross-session **knowledge graph** of what the learner knows — the graph is the app's home page.
A standalone learning app (web + Electron desktop): a **teacher agent** (Next.js; the Vercel AI SDK's `UIMessage` protocol end-to-end, LLM engine [pi-ai](pi-engine-swap.md) via OpenRouter), probe→teach lessons with graded quiz/ask interactions, a formal **session plan tree** the teacher authors once (`plan_init`) and whose progress is derived from tagged quiz calls, and a persistent cross-session **knowledge graph** of what the learner knows — the graph is the app's home page.
Not a chat app with a graph bolted on: sessions are probe→teach lessons, the plan tree is a structured, schema-validated object, sessions accumulate into the graph through the extraction pipeline (propositions + reasoning chains — the stored truth; the canvas renders a derived image), and the teacher reaches back into the graph through read-only prober tools on its own judgment so probing compounds across sessions.
- [`docs/architecture.md`](docs/architecture.md) is the **canonical architecture spec** — layers, the import law, the five systems, the invariants, the seam directory. (The pre-audit spec is archived under `docs/archive/` — history only.)
- Living design doc (all decisions): [`docs/vision.md`](docs/vision.md)
- Build plan + wave/gate results: [`docs/plan.md`](docs/plan.md)
- [`docs/architecture.md`](architecture.md) is the **canonical architecture spec** — layers, the import law, the five systems, the invariants, the seam directory. (The pre-audit spec is archived under `docs/archive/` — history only.)
- Living design doc (all decisions): [`docs/vision.md`](vision.md)
- Build plan + wave/gate results: [`docs/plan.md`](plan.md)
## Commands
@@ -22,6 +22,7 @@ npm start # serve the production build
npm test # the full suite (node:test via tsx; deterministic, no LLM, no network)
npm run lint # eslint — includes the import law (architecture §2)
npm run typecheck # tsc --noEmit
node scripts/smoke-web-subscriptions.mjs # after build: offline Next start/dev + reload gate
npm run desktop:dev # next dev + Electron shell pointed at it
npm run desktop:dist # local signed/notarized macOS arm64 build (LH_NOTARY keychain)
@@ -32,7 +33,7 @@ npx tsx --env-file=.env.local scripts/backfill-graph.ts # wipe + re-extrac
npx tsx --env-file=.env.local scripts/replay-layout.ts # re-derive layout from the ledgers
```
For Windows/Linux packages and the three-OS CI/release flow, see [`desktop-release.md`](desktop-release.md). Native renderer binaries are traced per platform, so release builds run on the target OS, not as cross-builds from macOS.
For Windows/Linux packages and the three-OS CI/release flow, see [`docs/desktop-release.md`](desktop-release.md). Native renderer binaries are traced per platform, so release builds run on the target OS, not as cross-builds from macOS.
## Environment (`.env.local`)
@@ -47,9 +48,19 @@ Secrets and machine-specific paths only. Model choice and thinking budgets are t
The packaged desktop app takes keys from `userData/auth.json` (BYOK, mode 0600, read/written directly by the server via `LH_AUTH_FILE`). PDFs render natively with pdfjs/Skia; SVGs use resvg.
### Subscription connections in web/dev
Ordinary `npm run dev` and `npm start` support the existing **ChatGPT, GitHub Copilot and Grok** connections in Settings → Providers. No separate launcher or change to Next's listening/LAN behavior is required. `npm run desktop:dev` uses these same dev-server connections; the packaged desktop app keeps its separate main-process credentials and private IPC transport.
Web/dev saves `chatgpt-credentials.json`, `github-copilot-credentials.json` and `xai-credentials.json` **beside the configured API-key `auth.json`**: the project directory by default, or the directory containing `LH_AUTH_FILE`. These are owner-only plaintext files, not files under `data/`, desktop Application Support, or a global Pi profile. Git and standalone-build exclusions cover both the files and atomic-write temporary files. Run only one server process per credential directory: refresh serialization is process-local, not a cross-process lock.
Sign-in opens a new tab in the browser that initiated it; permit that tab if your browser blocks popups. **ChatGPT sign-in must be completed in a browser on the server's machine**, because its callback listener remains `127.0.0.1` on that machine. Copilot/Grok use device codes. In the Electron dev shell, provider pages open in the system browser instead. Once connected, credentials and refresh remain server-owned; other browsers using that server share its connections and background-use consent, just as they share its API keys. Disconnect, storage-failure handling, and provider-specific background consent retain the existing behavior; subscription routes never fall back to API-key billing.
This remains a **trusted developer/shared-server mode**, not multi-user hosting. Subscription controls require same-origin JSON/custom-header requests; browser-opening actions are short-lived, origin-bound and one-use. These checks prevent ordinary cross-origin browser controls, **not user authentication or DNS-rebinding protection**. Do not expose an unprotected dev/server instance to untrusted users or the public Internet; a public deployment needs an independent access/host gate. Remote ChatGPT callback routing, per-user accounts and credential import are not implemented.
## Architecture
Five layers, dependency arrows pointing inward only, machine-enforced as four ESLint zones (shell → barrels only; core and agent never import below themselves). Two composition roots, one writer per store. The full spec — layers, systems, runtime flows, invariants, seams: [`docs/architecture.md`](docs/architecture.md).
Five layers, dependency arrows pointing inward only, machine-enforced as four ESLint zones (shell → barrels only; core and agent never import below themselves). Two composition roots, one writer per store. The full spec — layers, systems, runtime flows, invariants, seams: [`docs/architecture.md`](architecture.md).
```mermaid
flowchart TB
@@ -165,7 +176,7 @@ pricing-generation step: Pi AI's catalog and public calculator are the authority
**Practice (problem sets).** A node selection resolves to the STORED reasoning chains it fully contains (a chain rides iff every member is selected; unchained nodes ride as singletons); the `/practice` creation dialog picks a birth-session cluster (the graph view's own groups) and freezes its propositions + chains into a self-contained `ProblemSet`. Three named-level axes steer generation (ladder rung 0–4, cue distance 0–2, Theory|Application — never a number to interpolate; defaults Direct exercise · Adjacent context · Theory), plus an optional free-form note the agents weigh but need not follow. Creation has two entries: the cluster picker on `/practice`, or the graph view's chain table — a full-bleed sortable table of every reasoning chain, grouped by birth cluster under the shared top bar, selecting chips, chains, or whole clusters, whose floating selection bar shows the exact `N nodes → M chains` the agents will receive — handing a free-form node selection to the same dialog — and carries the fenced Delete action, making it the graph's management view (whole-cluster cleanup through the same full-death confirm dialog). A partitioner agent covers every chain exactly once, one generator per partition writes problems against the owner's problem-engineering framework (commit-before-pose: statement + solution + rubric together), and grading is **strengthening, not assessment** — one attempt per problem with its own submit button, instant solution/rubric reveal, then a score 0–10 with brief reasoning landing live; nothing ever written back to the graph. Runs are registry-backed background jobs on the activity dock (no steer, never wakes the teacher); sets live in `data/practice`, crash-swept on boot.
**Dictation.** Optional speech-to-text: the server mints short-lived Deepgram tokens; audio streams browser→Deepgram over WebSocket, never through the server.
**Dictation.** Optional speech-to-text: the server mints short-lived Deepgram tokens; audio streams browser→Deepgram over WebSocket, never through the server. Settings → Providers → Dictation → Language (below the Deepgram key) offers Nova-3's documented monolingual languages/locales or `language=multi` for automatic multilingual transcription. The app-wide `dictation.language` preference persists through the existing settings record and `PUT /api/settings/dictation`; absent preferences resolve to English (`en`). The settings control locks from connecting through stopping, and reconnects keep the run's original language. A shared preference controller publishes confirmed PUT responses to every reader immediately; the confirmed preference is cached across page navigation and reopening Settings, with no polling or per-recording read. Explicit external dictation-settings notifications invalidate the cache, and capture waits for a successful refresh. Navigating away cancels the run, invalidates pending startup/reconnect work and releases its microphone/socket without appending into another page's input. All other streaming parameters remain unchanged. The picker uses the model dropdown's shared `SearchMenuList` and `MenuShell`: the same fuzzy search, arrow navigation, Enter selection, row styling and scrolling. Its compact mono `SettingsMenuTrigger` is shared with Agent and Appearance; a muted setup line below the key and language rows links the Deepgram Console and states the Member-or-higher key-role requirement.
## Testing
@@ -179,10 +190,10 @@ All deterministic, no LLM, no network (engines are injected — pi-ai's faux pro
## Lineage
The teaching methodology comes from an Obsidian-based POC (its `knowledge graph design.md` was the original epistemic-model source, since amended — see the amendments table in `docs/vision.md`). The harness was originally built on the pi agent SDK, pivoted to standalone Next.js + the Vercel AI SDK, and later swapped its LLM engine back to pi-ai while keeping the AI SDK's `UIMessage` as the client protocol ([`docs/pi-engine-swap.md`](docs/pi-engine-swap.md)).
The teaching methodology comes from an Obsidian-based POC (its `knowledge graph design.md` was the original epistemic-model source, since amended — see the amendments table in `docs/vision.md`). The harness was originally built on the pi agent SDK, pivoted to standalone Next.js + the Vercel AI SDK, and later swapped its LLM engine back to pi-ai while keeping the AI SDK's `UIMessage` as the client protocol ([`docs/pi-engine-swap.md`](pi-engine-swap.md)).
## Data
Nothing under `data/` is ever tracked by git. Chats, graph ledgers, jobs, positions, and the settings records are plain files (JSON/JSONL) — inspect them freely; they are the canonical records. The desktop app relocates the whole directory via `LH_DATA_DIR`.
**Schema stamps (2026-10-10):** chat records carry `version` (stamped at the store's write choke point; absence reads as v1 — lazy migration, no mass rewrite), and the graph family (ledgers, sources, render-settings) shares one sidecar `graph/format.json`, created on the family's first write. `positions.json` is deliberately unstamped (rebuildable). A record stamped NEWER than the running build is refused on read — honest refusal for the downgrade scenario: a read-modify-write store never rewrites what it doesn't understand.
**Schema stamps (2026-10-10):** chat records carry `version` (stamped at the store's write choke point; absence reads as v1 — lazy migration, no mass rewrite), and the graph family (ledgers, sources, render-settings) shares one sidecar `graph/format.json`, created on the family's first write. `positions.json` is deliberately unstamped (rebuildable). A record stamped NEWER than the running build is refused on read — honest refusal for the downgrade scenario: a read-modify-write store never rewrites what it doesn't understand.
docs/probe-teach-split.md
+9 -2
View File
@@ -21,7 +21,8 @@ philosophy so they share a view of the whole process).
field it must not use. No plan tools, no jobs, no demos.
- **The teacher** runs everything from planning onward: full tool set
(teach quiz with nodeId, plan_init, jobs, demos). Constructed the moment the
transcript carries a `probe_done`.
transcript carries a concluded `probe_done`: a non-failed/non-aborted
assistant call followed by a matching successful tool result.
- **The swap happens between probing and planning** — the planning phase
needs full knowledge of how teaching is done, so the teacher agent starts
at planning.
@@ -43,6 +44,12 @@ The prober concludes explicitly:
- **Refuses while a question is open** ("resolve the open question first") —
same doctrine as nodeId validation.
- **Refuses a second conclusion** (idempotent: the probe concludes once).
Conclusion requires a later `probe_done` result matching the call's ID
and tool name. The calling generation must not have failed or aborted;
the result must not be `isError` or explicitly contain `details.ok:false`.
Clean acknowledgments without an explicit `ok:true` remain valid. Bare,
unmatched, errored, or refused calls keep probing open; saved attempts
are never deleted or replay-executed by this detection rule.
- The assessment is the tool call's input, so it streams to the learner live
and is durable in the transcript by the tool result itself. A dedicated
assessment card/viewer comes later (UI).
@@ -53,7 +60,7 @@ The prober concludes explicitly:
- **Side effect: it queues the handoff.** When the prober's turn ends, the
wake pipeline (the same machinery that delivers sub-agent results and
abort steers) fires a server-side turn with a handoff steer — the teacher
is constructed (probe_done is in the transcript) and authors the plan.
is constructed (probe_done has successfully concluded) and authors the plan.
The learner experiences assessment → plan as one continuous flow. If the
wake is lost (crash window), the next client message still constructs the
teacher — the transcript is the source of truth.
docs/tool-content-recovery.md
+63
View File
@@ -0,0 +1,63 @@
# Tool-content presentation recovery
## Contract
Model-authored tool content is parsed against its schema before presentation. Math or document recovery must not change persisted tool arguments, option identities, learner answers, shuffling, grading, or future answer snapshots. Those use the source-derived, schema-parsed content.
`ToolContentResponse` projects text through `repairMathForDisplay`, then uses the ordinary Markdown renderer. Teaching/probing questions, asks, recalls/explanations, and practice statements/solutions share this path. Ordinary assistant prose does not receive tool-content recovery.
There are three separate responsibilities:
1. **Argument/shape parsing:** decode encoded argument shapes, then schema conversion, validation and cleaning. No LaTeX recovery here.
2. **Content projection:** distinguish literal code, prose/document structure and math; recover only supported escaping defects.
3. **Presentation:** Markdown/KaTeX rendering, including paragraph layout. In labeled Why/answer reveals, only the first direct paragraph is inline with its label; subsequent paragraphs retain their normal layout.
## Recovery rules
### Literal code
Code and inline-code source ranges come from the rendering stack's CommonMark parser with its math extensions. This covers fences, indentation, quotes and lists without maintaining separate container heuristics. The projection copies those source ranges verbatim; it does not serialize an AST back into Markdown.
### Document escapes
A literal two-character `\n` is different from a physical newline. Document-layout recovery requires evidence such as repeated document-escape candidates or a candidate alongside over-escaped known TeX commands. Complete known n-prefixed TeX words, including `\nu`, `\nabla`, `\neq` and declared n-prefixed macros, take precedence over interpreting their prefix as document layout. Code, genuine TeX comments and recognized path tokens are protected.
Recovered double-dollar display framing puts the formula and closing fence on their own lines. Decoding the opening line alone can otherwise turn closing dollars into erroneous TeX content.
### JSON-swallowed command prefixes
Inside math, a control character can be the result of JSON consuming the start of a command: TAB + `ext` instead of `\text`, for example. Recovery reconstructs a command only if the control character plus the **complete following alphabetic word** names a known KaTeX command. All supported occurrences recover in one projection.
The owner-approved typical-model-authoring policy is:
- **Inline math:** recover recognized prefixes, including physical LF + `u` to `\nu`, and TAB + `o` to `\to`. There is no word-length cutoff.
- **Display math:** keep physical LF and CRLF as source formatting; recover recognized non-newline prefixes.
- **Correctly written commands:** remain commands in both modes. `\nu` is never decoded into LF + `u`.
- **Unrecognized whitespace:** remains unchanged. `x<TAB>+<TAB>y` does not become invented `\t`/`\ty` commands.
The control-character case is inherently ambiguous: whitespace followed by a variable may happen to spell a command. This policy targets typical model-authored LaTeX, not every syntactically possible human whitespace arrangement. TeX parse success alone cannot distinguish the two meanings.
### Extra command escaping, percentages and rows
Recovery examines the original TeX context before changing backslash runs:
- Complete native environment pairs and native `\substack` arguments preserve genuine row separators.
- Consistently extra-escaped environment pairs establish an escaping factor for commands **and rows**. For example, an encoded four-backslash row becomes two backslashes, not one.
- Outside protected native row/comment contexts, extra prefixes on known commands can recover. Unknown command words are not invented.
- Doubled percent escapes in ordinary numeric expressions can recover to `\%`, including numeric percentages at the end of a display formula. Genuine following comment text or equation rows remain literal. This is not a blanket `%` addition to an even-backslash rule.
- As an additional guard, partial command recovery does not turn parseable source into a TeX parse failure. Parse success is not the primary correctness test: both a corrupted matrix and a lost percentage can parse successfully.
Unbalanced, unsupported or ambiguous content is not generally corrected into new mathematics. This is an escaping recovery projection, not a general TeX authoring repair system. There are no language-specific decoders.
## Historical answers
`answerForPresentation` joins model-authored display fields to the original tool call without changing recorded answer facts. Option labels bind only through unambiguous selected values, exact/source-recovered labels, or one uniquely remaining pair in a complete bijection. Ambiguous rows keep their saved snapshots. Saved order, indices, selected values, verdicts, confidence and learner text are retained. Dismissed answers are unchanged. No migration or record rewrite is performed.
## Verification
- `tests/content-recovery.test.ts`: recovery/preservation rules and adversarial row/comment/code examples.
- `tests/tool-content-presentation.test.tsx`: provenance-labeled synthetic corpus regressions, historical joins and ambiguity.
- `tests/tool-content-interaction.test.tsx`: select/submit/persist/reveal flow and paragraph structure.
- Existing Markdown, card, practice, tool-shape and grading suites remain applicable.
The local frozen corpus can be replayed offline with `scripts/replay-tool-rendering-corpus.ts`. Replay verifies source hashes and preserves baseline findings; it reports glyph loss, matrix rows, rendering errors, code bytes, schema regressions, answer facts and projection convergence. Corpus collection and larger diagnostic batches are separate from product recovery and are not prerequisites for running these unit tests.
electron/main.cjs
+93 -21
View File
@@ -24,12 +24,13 @@
* would shadow every settings-page key change until restart.
*/
const { app, BrowserWindow, Menu, ipcMain, session, shell, dialog } = require("electron");
const { app, BrowserWindow, Menu, ipcMain, session, shell } = require("electron");
const { createChatGPT, fileStore } = require("./chatgpt.cjs");
const { attachChatGPTChild, trustedSender, chatGPTControl } = require("./chatgpt-ipc.cjs");
const { createSubscriptions, loadSubscriptionSDK } = require("./subscriptions.cjs");
const { attachSubscriptionChild, subscriptionControl } = require("./subscription-ipc.cjs");
const { desktopAPIHeaders } = require("./desktop-api.cjs");
const { openWebSubscriptionBrowser } = require("./web-subscription-browser.cjs");
const { randomBytes } = require("node:crypto");
const { spawn } = require("node:child_process");
const fs = require("node:fs");
@@ -140,6 +141,7 @@ let subscriptions = null;
let subscriptionChild = null;
// Per-launch request capability, never OAuth tokens and never renderer state.
let desktopAPISecret = null;
let reopenWindow = null;
const trustedAppWindows = new Set();
// A GUI shell must never die to a stray exception (EPIPE dialogs, a race
@@ -314,6 +316,7 @@ function createWindow(url) {
contextIsolation: true,
sandbox: true,
nodeIntegration: false,
...(process.env.LH_DEV_URL ? { additionalArguments: ["--lh-web-subscriptions"] } : {}),
preload: path.join(__dirname, "preload.cjs"),
},
// Live-resize + first-paint flash (owner, 2026-09-17): during a resize
@@ -436,31 +439,57 @@ const ZOOM_LADDER = [0.75, 0.9, 1, 1.1, 1.25, 1.5, 1.75, 2];
* and before-quit alone would race the swap. */
const UPDATE_CHECK_INTERVAL_MS = 24 * 60 * 60 * 1000;
let autoUpdater = null; // electron-updater's, packaged builds only
// Observational only: this snapshot NEVER gates a check, download or install.
// The revision lets late subscribers reject an older in-flight IPC snapshot.
let updateState = { status: "idle", version: null, revision: 0 };
function broadcastUpdateStatus(status) {
for (const win of BrowserWindow.getAllWindows()) {
try {
win.webContents.send("lh:update-status", status);
} catch {
/* window mid-teardown — the next push retries */
function reportUpdateStatus(status, version = null) {
try {
// A later check/error must not hide an update already staged for quit.
if (updateState.status === "downloaded" && status !== "downloaded") return;
updateState = { status, version: typeof version === "string" ? version : null, revision: updateState.revision + 1 };
for (const win of BrowserWindow.getAllWindows()) {
try {
win.webContents.send("lh:update-status", updateState);
} catch {
/* window mid-teardown — snapshot retrieval covers a later load */
}
}
} catch {
/* Status reporting must never interrupt the updater. */
}
}
/** Read the SAME data directory passed to the packaged server. Only an
* explicit false opts out; absent/corrupt/unreadable settings keep updates ON.
* Re-read for each automatic check so saving the toggle needs no restart. */
function automaticUpdatesEnabled() {
try {
const settings = JSON.parse(fs.readFileSync(path.join(DATA_DIR(), "settings.json"), "utf8"));
return settings?.general?.autoUpdate !== false;
} catch {
return true;
}
}
function checkForUpdates({ manual = false } = {}) {
if (!autoUpdater) return;
if (!manual && !automaticUpdatesEnabled()) {
log("auto-update: automatic check skipped (turned off in Settings)");
return;
}
// Do not gate on display state: a stale spinner/error must never disable
// future background checks. electron-updater owns its in-flight work.
autoUpdater
.checkForUpdates()
.then((result) => {
// The download itself is silent (autoDownload); the banner fires on
// update-downloaded. Manual checks additionally answer directly.
if (!manual) return;
const version = result?.updateInfo?.version;
if (!version) dialog.showMessageBox({ type: "info", message: "You're up to date.", detail: `LearnHarness ${app.getVersion()} is the latest version.` });
// A declined check (e.g. an unsupported package) establishes no
// latest-version claim. Real results arrive through updater events.
if (!result) reportUpdateStatus("idle");
})
.catch((err) => {
log(`update check failed: ${err?.message ?? err}`);
if (manual) dialog.showMessageBox({ type: "warning", message: "Couldn't check for updates.", detail: String(err?.message ?? err) });
reportUpdateStatus("error");
});
}
@@ -477,12 +506,14 @@ function installUpdateNow() {
function setupAutoUpdater() {
if (!app.isPackaged) {
log("auto-update: dev build — updater disabled");
reportUpdateStatus("disabled");
return;
}
try {
autoUpdater = require("electron-updater").autoUpdater;
} catch (err) {
log(`auto-update: electron-updater unavailable (${err?.message ?? err}) — updates disabled`);
reportUpdateStatus("disabled");
return;
}
const feed = process.env.LH_UPDATE_FEED;
@@ -500,11 +531,37 @@ function setupAutoUpdater() {
debug: (m) => {}, // too chatty for the persistent log
error: (m) => log(`updater error: ${m}`),
};
autoUpdater.on("checking-for-update", () => reportUpdateStatus("checking"));
autoUpdater.on("update-available", (info) => reportUpdateStatus("downloading", info?.version));
autoUpdater.on("update-not-available", () => reportUpdateStatus("up-to-date"));
autoUpdater.on("update-downloaded", (info) => {
log(`update downloaded: ${info?.version} — banner shown`);
broadcastUpdateStatus({ status: "downloaded", version: info?.version ?? null });
reportUpdateStatus("downloaded", info?.version);
});
autoUpdater.on("error", (err) => log(`updater error: ${err?.message ?? err}`));
autoUpdater.on("error", (err) => {
log(`updater error: ${err?.message ?? err}`);
reportUpdateStatus("error");
});
}
/** One result surface for manual checks. Preload buffers the navigation
* request until React subscribes, including when all macOS windows closed. */
function checkForUpdatesFromMenu() {
try {
const win = BrowserWindow.getAllWindows()[0] ?? reopenWindow?.();
if (win) {
if (win.isMinimized()) win.restore();
win.focus();
const openSettings = () => {
try { win.webContents.send("lh:open-settings", "general"); } catch { /* window closed */ }
};
if (win.webContents.isLoadingMainFrame()) win.webContents.once("did-finish-load", openSettings);
else openSettings();
}
} catch {
/* Navigation failure must not prevent the manual check. */
}
checkForUpdates({ manual: true });
}
function zoomStep(win, dir) {
@@ -530,9 +587,8 @@ function installMenu() {
label: app.name,
submenu: [
{ role: "about" },
// Manual check (Wave C): answers via native dialog — no renderer
// round-trip. The 24h automatic checks stay silent.
{ label: "Check for Updates…", click: () => checkForUpdates({ manual: true }) },
// Manual checks report their result in Settings → General.
{ label: "Check for Updates…", click: () => checkForUpdatesFromMenu() },
{ type: "separator" },
{ role: "hide" },
{ role: "hideOthers" },
@@ -646,6 +702,13 @@ async function main() {
return subscriptionControl(command, subscriptions);
});
// The dev shell has no browser tabs. Only a one-use action on its own
// Next server may be resolved here; never accept an arbitrary external URL.
ipcMain.handle("lh:subscriptions:web-browser", async (event, action) => {
if (!devUrl || !trustedAppWindows.has(BrowserWindow.fromWebContents(event.sender)) || !trustedSender(event, new URL(url).origin, BrowserWindow)) throw new Error("Untrusted development request.");
await openWebSubscriptionBrowser(action, new URL(url).origin, { openExternal: (target) => shell.openExternal(target) });
});
installMenu();
// Auto-update (Wave C): wire the updater before the window so the
@@ -654,6 +717,14 @@ async function main() {
// are log lines — this can never block boot.
setupAutoUpdater();
ipcMain.on("lh:update-install", () => installUpdateNow());
ipcMain.handle("lh:update:get", (event) => {
if (!trustedAppWindows.has(BrowserWindow.fromWebContents(event.sender)) || !trustedSender(event, new URL(url).origin, BrowserWindow)) throw new Error("Untrusted desktop request.");
return updateState;
});
ipcMain.handle("lh:update:check", (event) => {
if (!trustedAppWindows.has(BrowserWindow.fromWebContents(event.sender)) || !trustedSender(event, new URL(url).origin, BrowserWindow)) throw new Error("Untrusted desktop request.");
checkForUpdates({ manual: true });
});
// Sidebar-at-narrow-window (owner, 2026-09-17): the renderer auto-closes
// the sidebar below SIDEBAR_AUTOCLOSE (928px); opening it again at that
@@ -670,13 +741,14 @@ async function main() {
const target = SIDEBAR_AUTOCLOSE;
if (width < target) win.setBounds({ x, y, width: target, height });
});
reopenWindow = () => { chatgpt?.resume(); subscriptions?.resume(); return createWindow(url); };
app.on("activate", () => {
if (BrowserWindow.getAllWindows().length === 0) { chatgpt?.resume(); subscriptions?.resume(); createWindow(url); }
if (BrowserWindow.getAllWindows().length === 0) reopenWindow();
});
createWindow(url);
log(`window open: ${url}`);
// First update check AFTER the window exists (the downloaded-banner
// push needs a listener); dev builds no-op (no autoUpdater).
// Startup + 24h checks stay main-owned and read the current preference.
// No renderer subscription is required; late loads retrieve the snapshot.
if (autoUpdater) {
checkForUpdates();
setInterval(() => checkForUpdates(), UPDATE_CHECK_INTERVAL_MS);
electron/preload.cjs
+31 -3
View File
@@ -19,16 +19,31 @@
*
* window.lh.onUpdateStatus(cb) / window.lh.installUpdate() — the
* auto-update channel (Wave C, 2026-10-10): "lh:update-status" pushes
* {status:"downloaded", version} once an update is staged; installUpdate
* asks main to kill the server and quit-and-install. Browser builds:
* both absent, the banner never renders.
* {status, version, revision} on updater events; installUpdate asks main
* to kill the server and quit-and-install. getUpdateState retrieves the
* snapshot for a late subscriber; checkForUpdates requests a manual
* check (including download/install-on-quit, even with automatic checks OFF).
*
* window.lh.onOpenSettings(cb) — menu navigation, buffered until React
* subscribes after a window is reopened. Browser builds have no bridge.
*
* Channel names are mirrored in electron/main.cjs — keep them in sync.
*/
const { contextBridge, ipcRenderer } = require("electron");
let openSettingsListener = null;
let pendingOpenSettings = null;
ipcRenderer.on("lh:open-settings", (_event, section) => {
if (openSettingsListener) openSettingsListener(section);
else pendingOpenSettings = section;
});
contextBridge.exposeInMainWorld("lh", {
platform: process.platform,
...(process.argv?.includes("--lh-web-subscriptions") ? {
subscriptionTransport: "web",
openSubscriptionBrowser: (action) => ipcRenderer.invoke("lh:subscriptions:web-browser", action),
} : {}),
chatgpt: {
status: () => ipcRenderer.invoke("lh:chatgpt:control", { op: "status" }),
login: (id) => ipcRenderer.invoke("lh:chatgpt:control", { op: "login", id }),
@@ -71,4 +86,17 @@ contextBridge.exposeInMainWorld("lh", {
return () => ipcRenderer.removeListener("lh:update-status", listener);
},
installUpdate: () => ipcRenderer.send("lh:update-install"),
getUpdateState: () => ipcRenderer.invoke("lh:update:get"),
checkForUpdates: () => ipcRenderer.invoke("lh:update:check"),
onOpenSettings: (cb) => {
openSettingsListener = cb;
if (pendingOpenSettings !== null) {
const section = pendingOpenSettings;
pendingOpenSettings = null;
cb(section);
}
return () => {
if (openSettingsListener === cb) openSettingsListener = null;
};
},
});
electron/web-subscription-browser.cjs
+19
View File
@@ -0,0 +1,19 @@
"use strict";
/** Desktop-dev has no tabs. Resolve only an opaque action on its own Next
* server, then open the owner-validated provider page outside the renderer.
* The caller must authenticate its trusted app main frame first. */
async function openWebSubscriptionBrowser(action, origin, { fetch: request = globalThis.fetch, openExternal }) {
if (typeof action !== "string" || !/^\/api\/subscriptions\/browser\?id=[0-9a-f-]{36}$/.test(action)) throw new Error("Invalid sign-in action.");
const response = await request(new URL(action, origin), {
method: "POST", redirect: "manual", credentials: "omit",
headers: { origin, "x-learnharness-subscriptions": "1" },
signal: AbortSignal.timeout(15000),
});
const target = response.headers.get("location");
if (response.status !== 303 || !target) throw new Error("Sign-in action expired.");
const parsed = new URL(target);
if (parsed.protocol !== "https:" || parsed.username || parsed.password) throw new Error("Invalid provider page.");
await openExternal(target);
}
module.exports = { openWebSubscriptionBrowser };
next.config.ts
+8
View File
@@ -1,6 +1,11 @@
import { readFileSync } from "node:fs";
import type { NextConfig } from "next";
const pkg = JSON.parse(readFileSync(new URL("./package.json", import.meta.url), "utf8")) as { version: string; homepage: string };
const nextConfig: NextConfig = {
// Public build metadata only; no new runtime Electron version API.
env: { LH_VERSION: pkg.version, LH_HOMEPAGE: pkg.homepage },
// Desktop target (Electron wrapper, 2026-09-17): standalone output = a
// self-contained .next/standalone/server.js the Electron shell spawns
// on 127.0.0.1. Harmless for plain web deploys (extra build output).
@@ -43,6 +48,9 @@ const nextConfig: NextConfig = {
"*": [
"**/data/**/*",
"auth.json",
"chatgpt-credentials.json*",
"github-copilot-credentials.json*",
"xai-credentials.json*",
".env*",
],
},
package-lock.json
Generated
+9 -6
View File
@@ -1,12 +1,12 @@
{
"name": "learnharness",
"version": "0.3.7",
"version": "0.3.9",
"lockfileVersion": 3,
"requires": true,
"packages": {
"": {
"name": "learnharness",
"version": "0.3.7",
"version": "0.3.9",
"license": "AGPL-3.0-only",
"dependencies": {
"@earendil-works/pi-agent-core": "1.4.0",
@@ -36,7 +36,10 @@
"katex": "0.18.2",
"linkedom": "^0.18.13",
"lucide-react": "^1.33.0",
"mdast-util-from-markdown": "2.0.3",
"mdast-util-math": "3.0.0",
"media-chrome": "^4.19.2",
"micromark-extension-math": "3.1.0",
"mermaid": "^11.17.0",
"motion": "^13.1.1",
"nanoid": "^6.0.1",
@@ -66,7 +69,7 @@
"@types/turndown": "^5.0.6",
"@typescript-eslint/parser": "^8.68.0",
"concurrently": "^10.0.5",
"electron": "^44.4.1",
"electron": "^44.9.3",
"electron-builder": "26.16.1",
"eslint": "^9.39.5",
"eslint-import-resolver-typescript": "^4.4.5",
@@ -8969,9 +8972,9 @@
}
},
"node_modules/electron": {
"version": "44.4.1",
"resolved": "https://registry.npmjs.org/electron/-/electron-44.4.1.tgz",
"integrity": "sha512-gXJarQ2Utfn8G6bp/IqLXvckOSCtrxPWjFqQaqm02Gf4zYn1aP+RN9PKimoLH/DMqjKLfBrNC4C4CviPv0cXnQ==",
"version": "44.9.3",
"resolved": "https://registry.npmjs.org/electron/-/electron-44.9.3.tgz",
"integrity": "sha512-Cv9/yMnZxNWsMlIe8T4v//NKHdUvbeyoTJoXE3BIcRxus2meeFaO9h/P3zB5d4cBXum51IgAYrT2QjpIDWKpJA==",
"dev": true,
"license": "MIT",
"dependencies": {
package.json
+5 -2
View File
@@ -1,6 +1,6 @@
{
"name": "learnharness",
"version": "0.3.7",
"version": "0.3.9",
"license": "AGPL-3.0-only",
"author": {
"name": "Amos Blomqvist",
@@ -51,7 +51,10 @@
"katex": "0.18.2",
"linkedom": "^0.18.13",
"lucide-react": "^1.33.0",
"mdast-util-from-markdown": "2.0.3",
"mdast-util-math": "3.0.0",
"media-chrome": "^4.19.2",
"micromark-extension-math": "3.1.0",
"mermaid": "^11.17.0",
"motion": "^13.1.1",
"nanoid": "^6.0.1",
@@ -84,7 +87,7 @@
"@types/turndown": "^5.0.6",
"@typescript-eslint/parser": "^8.68.0",
"concurrently": "^10.0.5",
"electron": "^44.4.1",
"electron": "^44.9.3",
"electron-builder": "26.16.1",
"eslint": "^9.39.5",
"eslint-import-resolver-typescript": "^4.4.5",
scripts/build-server.mjs
+2 -1
View File
@@ -52,7 +52,8 @@ if (exists(leakedData)) {
// (the DEV BYOK key store, 0600) byte-identical into the DMG (caught in
// the 2026-10-08 payload audit; excluded in next.config.ts — this guard
// makes a regression loud instead of silent).
for (const secret of ["auth.json", ".env", ".env.local"]) {
const subscriptionSecrets = fs.readdirSync(STANDALONE).filter((name) => /^(chatgpt|github-copilot|xai)-credentials\.json/.test(name));
for (const secret of ["auth.json", ...subscriptionSecrets, ".env", ".env.local"]) {
if (exists(path.join(STANDALONE, secret))) {
console.error(`[build-server] REFUSING: standalone contains ${secret} — dev secrets would ship in the app.`);
console.error(" Check outputFileTracingExcludes in next.config.ts.");
scripts/smoke-opening-prompt.mjs
+251
View File
@@ -0,0 +1,251 @@
/** Focused synthetic live-UI check. No model/provider calls: /api/chat is
* intercepted, catalog is a fixture, app HOME/data/auth are throwaway, and
* the browser uses a fresh profile. Creates only synthetic session records.
* Run: node scripts/smoke-opening-prompt.mjs
*/
import assert from "node:assert/strict";
import { spawn } from "node:child_process";
import { once } from "node:events";
import fs from "node:fs";
import net from "node:net";
import { createRequire } from "node:module";
import os from "node:os";
import path from "node:path";
import puppeteer from "puppeteer-core";
const root = path.resolve(import.meta.dirname, "..");
const home = fs.mkdtempSync(path.join(os.tmpdir(), "lh154-ui-"));
fs.mkdirSync(path.join(home, "data"));
fs.writeFileSync(path.join(home, "auth.json"), "{}\n");
// Next's supported font fixture avoids external font downloads in this check.
const require = createRequire(import.meta.url);
const { getFontAxes } = require("next/dist/compiled/@next/font/dist/google/get-font-axes.js");
const { getGoogleFontsUrl } = require("next/dist/compiled/@next/font/dist/google/get-google-fonts-url.js");
const fonts = Object.fromEntries(["JetBrains Mono", "Source Sans 3"].map(family => [
getGoogleFontsUrl(family, getFontAxes(family, ["variable"], ["normal"]), "swap"),
`@font-face { font-family: '${family}'; src: local('Arial'); font-weight: 100 900; }`,
]));
fs.writeFileSync(path.join(home, "fonts.cjs"), `module.exports = ${JSON.stringify(fonts)};\n`);
const listener = net.createServer();
listener.listen(0, "127.0.0.1");
await once(listener, "listening");
const port = listener.address().port;
await new Promise(resolve => listener.close(resolve));
const origin = `http://127.0.0.1:${port}`;
const env = {
PATH: process.env.PATH, HOME: home, TMPDIR: home, NODE_ENV: "development",
LH_DATA_DIR: path.join(home, "data"), LH_AUTH_FILE: path.join(home, "auth.json"),
LH_IGNORE_ENV_KEYS: "1", NEXT_TELEMETRY_DISABLED: "1",
NEXT_FONT_GOOGLE_MOCKED_RESPONSES: path.join(home, "fonts.cjs"),
};
const server = spawn(process.execPath, ["node_modules/next/dist/bin/next", "dev", "--hostname", "127.0.0.1", "--port", String(port)], {
cwd: root, env, stdio: ["ignore", "pipe", "pipe"],
});
let logs = "";
const capture = data => { logs = (logs + data).slice(-24000); };
server.stdout.on("data", capture); server.stderr.on("data", capture);
let browser;
let page;
const errors = [];
const checks = [];
try {
await new Promise((resolve, reject) => {
const deadline = setTimeout(() => reject(new Error("app startup timed out")), 60000);
const ready = () => { if (logs.includes("Ready in")) { clearTimeout(deadline); server.stdout.off("data", ready); resolve(); } };
server.stdout.on("data", ready);
server.once("exit", code => { clearTimeout(deadline); reject(new Error(`app exited ${code}`)); });
ready();
});
browser = await puppeteer.launch({
executablePath: process.env.LH_TEST_CHROME || "/Applications/Google Chrome.app/Contents/MacOS/Google Chrome",
headless: true, userDataDir: path.join(home, "browser"), env,
args: ["--disable-background-networking", "--disable-sync", "--no-first-run", "--disable-default-apps"],
});
page = await browser.newPage();
page.setDefaultTimeout(45000);
page.on("pageerror", err => errors.push(err.message));
const model = { id: "synthetic", name: "Synthetic (no calls)", provider: "fixture", input: ["text"], reasoning: false, supportedLevels: ["off"], cost: { input: 0, output: 0, cacheRead: 0, cacheWrite: 0 }, contextWindow: 8192, maxTokens: 1024 };
const slot = { model: { provider: model.provider, id: model.id }, thinkingLevel: "off" };
const catalog = { models: [model], providers: [{ id: "fixture", label: "Fixture", keyPresent: true }], defaults: null,
slots: { strong: slot, cheap: slot, vision: slot }, rolePins: {} };
const posts = [];
const creates = [];
const accepted = new Map();
let mode = "accepted";
let held = [];
await page.setRequestInterception(true);
page.on("request", async request => {
try {
const url = new URL(request.url());
if (url.origin !== origin) { await request.abort(); return; }
const pathname = url.pathname;
const respond = body => request.respond({ status: 200, contentType: "application/json", body: JSON.stringify(body) });
if (pathname === "/api/models") return await respond(catalog);
// Notifications are outside this handover check. 204 tells native
// EventSource not to reconnect or hold a test connection open.
if (pathname === "/api/events") return await request.respond({ status: 204 });
if (pathname === "/api/chat") {
const body = JSON.parse(request.postData()); posts.push(body);
if (mode === "held-post") { held.push(request); return; }
if (mode === "lost-ack") return await request.abort("failed");
if (mode === "refused") return await request.respond({ status: 503, contentType: "application/json", body: JSON.stringify({ code: "models_not_configured", missing: ["strong"] }) });
accepted.set(body.chatId, [{ id: "saved_user", role: "user", content: body.input.text, timestamp: Date.now() }]);
return await respond({ ok: true });
}
if (/^\/api\/chats\/[^/]+\/stream$/.test(pathname)) {
// Keep teacher delivery unfinished: cleanup must NOT wait for it.
held.push(request); return;
}
const match = pathname.match(/^\/api\/chats\/([^/]+)$/);
if (match && request.method() === "GET" && accepted.has(match[1])) {
return await respond({ messages: accepted.get(match[1]), turns: [], turnInFlight: false });
}
if (pathname === "/api/chats" && request.method() === "POST") creates.push(JSON.parse(request.postData()));
await request.continue();
} catch (err) { if (!/Target closed|Session closed|Invalid InterceptionId/.test(String(err))) errors.push(String(err)); }
});
async function open(route) {
// CDP-intercepted requests do not settle themselves on navigation.
// Release synthetic streams from prior cases rather than pin slots.
await Promise.all(held.splice(0).map(request => request.abort("aborted").catch(() => {})));
await page.goto(origin + route, { waitUntil: "domcontentloaded", timeout: 120000 });
await page.waitForSelector(route === "/" ? "textarea" : '[data-testid="composer"] textarea');
await page.waitForFunction(() => !document.querySelector('[data-testid="setup-cta"]') && !document.querySelector('[data-testid="setup-cta-keys"]'));
}
async function fill(text) {
await page.evaluate(value => {
const textarea = document.querySelector('[data-testid="composer"] textarea') || document.querySelector("textarea");
Object.getOwnPropertyDescriptor(HTMLTextAreaElement.prototype, "value").set.call(textarea, value);
textarea.dispatchEvent(new Event("input", { bubbles: true }));
}, text);
}
async function submit() {
await page.waitForFunction(() => {
const button = document.querySelector('button[aria-label="Start session"], [data-testid="composer"] button[aria-label="Send"]');
return button && !button.disabled;
});
await page.locator('button[aria-label="Start session"], [data-testid="composer"] button[aria-label="Send"]').click();
}
async function stash(id) { return page.evaluate(chat => sessionStorage.getItem(`lh.openingPrompt.${chat}`), id); }
async function assertDraft(text) { await page.waitForFunction(value => document.querySelector('[data-testid="composer"] textarea')?.value === value, {}, text); }
const long = "A full synthetic syllabus — 世界 & ? #. ".repeat(1200).trim();
for (const route of ["/", "/chat/new"]) {
for (const text of ["a short synthetic topic", long]) {
await open(route);
const count = posts.length;
await fill(text); await submit();
await page.waitForFunction(() => /^\/chat\/chat_/.test(location.pathname) && !location.search);
await page.waitForFunction(() => document.querySelector('[data-testid="composer"]'));
await page.waitForFunction(() => !Object.keys(sessionStorage).some(k => k.startsWith("lh.openingPrompt.")));
assert.equal(posts.length, count + 1);
assert.equal(posts.at(-1).input.text, text);
assert.equal(new URL(page.url()).search, "");
assert.equal(await stash(posts.at(-1).chatId), null);
checks.push(`${route}: ${text.length > 1000 ? "long" : "short"} input delivered intact through short URL; stash removed while stream held`);
}
}
// The old long-query mechanism really is rejected by this isolated Node server.
const oldResponse = await fetch(`${origin}${new URL(page.url()).pathname}?prompt=${encodeURIComponent(long)}`);
assert.equal(oldResponse.status, 431); await oldResponse.text();
checks.push("old long-query request returned431 (browser-shell blank appearance not asserted)");
for (const route of ["/", "/chat/new"]) {
await open(route);
await page.evaluate(() => {
const setItem = Storage.prototype.setItem;
Storage.prototype.setItem = function(key, value) {
if (window.failOpeningStorage && key.startsWith("lh.openingPrompt.")) throw new DOMException("synthetic quota", "QuotaExceededError");
return setItem.call(this, key, value);
};
window.failOpeningStorage = true;
});
const createCount = creates.length; const sendCount = posts.length;
await fill(long); await submit();
await page.waitForSelector('[role="alert"]');
assert.equal(new URL(page.url()).pathname, route);
assert.equal(await page.$eval("textarea", element => element.value), long);
assert.equal(posts.length, sendCount);
assert.equal(creates.length, createCount + 1);
await page.evaluate(() => { window.failOpeningStorage = false; });
await submit();
await page.waitForFunction(() => /^\/chat\/chat_/.test(location.pathname));
await page.waitForFunction(() => !Object.keys(sessionStorage).some(k => k.startsWith("lh.openingPrompt.")));
assert.equal(creates.length, createCount + 1, "unchanged retry reuses creation");
assert.equal(posts.at(-1).input.text, long);
checks.push(`${route}: quota failure preserved visible draft/no navigation; retry reused creation`);
}
for (const failure of ["lost-ack", "refused", "held-post"]) {
mode = failure;
await open("/chat/new");
const text = `recover interrupted ${failure}`;
await fill(text); await submit();
await page.waitForFunction(() => /^\/chat\/chat_/.test(location.pathname));
await page.waitForFunction(() => Object.values(sessionStorage).some(v => v.includes('"attempted":true')));
if (failure !== "held-post") await assertDraft(text);
const count = posts.length;
await page.reload({ waitUntil: "domcontentloaded" });
await assertDraft(text);
assert.equal(posts.length, count, "interrupted start must not automatically repost");
checks.push(`${failure}: reload restored full draft, no automatic repost`);
if (failure === "lost-ack") {
mode = "accepted";
await submit();
const id = new URL(page.url()).pathname.split("/").at(-1);
await page.waitForFunction(chat => !sessionStorage.getItem(`lh.openingPrompt.${chat}`), {}, id);
assert.equal(posts.length, count + 1);
assert.equal(posts.at(-1).input.text, text);
checks.push("restored draft sent only after explicit click and cleaned up on acceptance");
}
}
mode = "accepted";
// Marker-only session must still automatically deliver a fresh opening.
await open("/chat/new");
const markerId = await page.evaluate(async () => {
const res = await fetch("/api/chats", { method: "POST", headers: { "Content-Type": "application/json" }, body: JSON.stringify({ title: "marker fixture", graphContext: { propositions: [{ id: "synthetic_node", name: "Fixture", statement: "Synthetic" }] } }) });
return (await res.json()).id;
});
await page.evaluate(id => sessionStorage.setItem(`lh.openingPrompt.${id}`, JSON.stringify({ text: "marker opening", attempted: false })), markerId);
const markerCount = posts.length;
await page.goto(`${origin}/chat/${markerId}`, { waitUntil: "domcontentloaded" });
await page.waitForFunction(id => !sessionStorage.getItem(`lh.openingPrompt.${id}`), {}, markerId);
assert.equal(posts.length, markerCount + 1);
assert.equal(posts.at(-1).input.text, "marker opening");
checks.push("graph-marker-only record still receives opening text");
// A real persisted user prevents delivery and cleans a stale copy on mount.
const chatFile = path.join(home, "data", "chats", `${markerId}.json`);
const record = JSON.parse(fs.readFileSync(chatFile, "utf8"));
record.messages.push({ id: "confirmed_fixture", role: "user", content: "already received", timestamp: Date.now() });
fs.writeFileSync(chatFile, JSON.stringify(record));
await page.evaluate(id => sessionStorage.setItem(`lh.openingPrompt.${id}`, JSON.stringify({ text: "stale opening", attempted: true })), markerId);
const completedCount = posts.length;
await page.reload({ waitUntil: "domcontentloaded" });
await page.waitForFunction(id => !sessionStorage.getItem(`lh.openingPrompt.${id}`), {}, markerId);
assert.equal(posts.length, completedCount);
checks.push("persisted real conversation removed stale stash without reposting");
await open("/chat/new");
const legacyId = await page.evaluate(async () => {
const response = await fetch("/api/chats", { method: "POST", headers: { "Content-Type": "application/json" }, body: JSON.stringify({ title: "legacy fixture" }) });
return (await response.json()).id;
});
const legacyCount = posts.length;
await page.goto(`${origin}/chat/${legacyId}?prompt=${encodeURIComponent("legacy opening")}`, { waitUntil: "domcontentloaded" });
await page.waitForFunction(() => document.body.innerText.includes("legacy opening"));
await page.waitForFunction(id => !sessionStorage.getItem(`lh.openingPrompt.${id}`), {}, legacyId);
assert.equal(posts.length, legacyCount + 1);
assert.equal(posts.at(-1).input.text, "legacy opening");
checks.push("legacy short ?prompt link still sends its complete opening message");
assert.deepEqual(errors, []);
console.log(JSON.stringify({ checks, pageErrors: errors, isolation: "synthetic catalog/POST+streams; empty auth; no providers", origin }, null, 2));
} catch (err) {
console.error(err);
console.error(JSON.stringify({ checks, errors, url: page?.url(), dom: await page?.evaluate(() => ({ textareas: [...document.querySelectorAll('textarea')].map(e => ({ length: e.value.length, placeholder: e.placeholder })), buttons: [...document.querySelectorAll('button[aria-label="Send"], button[aria-label="Start session"]')].map(e => ({ disabled: e.disabled, label: e.getAttribute('aria-label') })), alerts: [...document.querySelectorAll('[role="alert"]')].map(e => e.textContent), failStorage: window.failOpeningStorage, text: document.body.innerText.slice(-1200) })).catch(() => null) }, null, 2));
console.error(logs); process.exitCode = 1;
} finally {
await browser?.close();
const exited = once(server, "exit");
server.kill("SIGTERM");
const killTimer = setTimeout(() => server.kill("SIGKILL"), 5000); killTimer.unref();
await exited; clearTimeout(killTimer);
fs.rmSync(home, { recursive: true, force: true });
}
scripts/smoke-web-subscriptions.mjs
+125
View File
@@ -0,0 +1,125 @@
/** Offline Next runtime gate (run after next build, with no dev server running).
* Proves cold-start registration, cross-route owner/events and dev reloads using
* synthetic credentials in a fresh profile. No OAuth, provider traffic or real
* auth/data. The dev reload changes only a source mtime, restoring it afterward.
* This is a test, not an alternative app launcher. */
import assert from "node:assert/strict";
import fs from "node:fs/promises";
import path from "node:path";
import os from "node:os";
import { pathToFileURL } from "node:url";
import { createRequire } from "node:module";
import { spawn } from "node:child_process";
import { createServer } from "node:net";
const ROOT = path.resolve(import.meta.dirname, "..");
const require = createRequire(import.meta.url);
const { SCOPES } = require("../electron/chatgpt.cjs");
const sdk = (file) => import(pathToFileURL(path.join(ROOT, "node_modules/@earendil-works/pi-ai/dist", file)));
const { createModels } = await sdk("index.js");
const { builtinProviders } = await sdk("providers/all.js");
const { githubCopilotProvider } = await sdk("providers/github-copilot.js");
const models = createModels();
for (const provider of builtinProviders()) models.setProvider(provider);
const model = models.getModels("openai").find((entry) => entry.api === "openai-responses");
const copilot = githubCopilotProvider().getModels()[0];
assert(model && copilot, "Installed SDK catalogs required.");
const profile = await fs.mkdtemp(path.join(os.tmpdir(), "lh-web-runtime-"));
await fs.chmod(profile, 0o700);
for (const name of ["home", "config", "cache", "xdg-data", "data", "pi", "claude", "codex", "tmp"]) await fs.mkdir(path.join(profile, name), { mode: 0o700 });
await fs.writeFile(path.join(profile, "auth.json"), "{}\n", { mode: 0o600 });
const preload = path.join(profile, "offline-fetch.mjs");
await fs.writeFile(preload, `const original = globalThis.fetch;
globalThis.fetch = async (input, init) => {
const url = new URL(typeof input === 'string' || input instanceof URL ? input : input.url);
if (url.href === 'https://api.openai.com/v1/models') return Response.json({ models: [{ slug: ${JSON.stringify(model.id)}, display_name: 'Offline fixture ChatGPT', visibility: 'list' }] });
if (['localhost', '127.0.0.1', '[::1]'].includes(url.hostname)) return original(input, init);
throw new Error('Offline subscription gate forbids external network: ' + url.origin + url.pathname);
};\n`, { mode: 0o600 });
// Intentionally an allowlist, not a copy of the calling shell's environment.
const env = {
PATH: process.env.PATH ?? "/usr/bin:/bin", HOME: path.join(profile, "home"),
XDG_CONFIG_HOME: path.join(profile, "config"), XDG_CACHE_HOME: path.join(profile, "cache"), XDG_DATA_HOME: path.join(profile, "xdg-data"),
PI_CODING_AGENT_DIR: path.join(profile, "pi"), PI_AGENT_DIR: path.join(profile, "pi"), CLAUDE_CONFIG_DIR: path.join(profile, "claude"), CODEX_HOME: path.join(profile, "codex"),
TMPDIR: path.join(profile, "tmp"), LH_AUTH_FILE: path.join(profile, "auth.json"), LH_DATA_DIR: path.join(profile, "data"),
NODE_OPTIONS: `--import=${pathToFileURL(preload).href}`, NEXT_TELEMETRY_DISABLED: "1",
};
const saved = {
version: 1, hostId: "offline-host", allowBackgroundUse: true, activeId: "offline-account",
registrations: [{ id: "offline-account", clientId: "oaiapp_fixture", issuer: "https://auth.openai.com", subject: "offline-subject", email: "fixture@example.invalid", tokens: { accessToken: "fixture-chatgpt-access", refreshToken: "fixture-chatgpt-refresh", idToken: "fixture-id-token", scopes: SCOPES, expiresAt: Date.now() + 3600000 } }],
};
function timeout(promise, milliseconds, message) {
let timer;
return Promise.race([promise, new Promise((_, reject) => { timer = setTimeout(() => reject(new Error(message)), milliseconds); })]).finally(() => clearTimeout(timer));
}
async function freePort() {
const socket = createServer();
await new Promise((resolve, reject) => { socket.once("error", reject); socket.listen(0, "127.0.0.1", resolve); });
const { port } = socket.address(); await new Promise((resolve) => socket.close(resolve)); return port;
}
async function run(mode) {
await fs.writeFile(path.join(profile, "chatgpt-credentials.json"), JSON.stringify(saved), { mode: 0o600 });
for (const id of ["github-copilot", "xai"]) await fs.writeFile(path.join(profile, `${id}-credentials.json`), JSON.stringify({
version: 1, allowBackgroundUse: true, connectionId: "00000000-0000-4000-8000-000000000001",
credential: { type: "oauth", access: id === "xai" ? "fixture-xai-access" : "tid=fixture;proxy-ep=proxy.individual.githubcopilot.com;", refresh: "fixture-refresh", expires: Date.now() + 3600000, ...(id === "github-copilot" ? { availableModelIds: [copilot.id] } : {}) },
}), { mode: 0o600 });
const port = await freePort(), origin = `http://localhost:${port}`;
let logs = "", exited = false;
const child = spawn(process.execPath, [path.join(ROOT, "node_modules/next/dist/bin/next"), mode, "-p", String(port)], { cwd: ROOT, env, stdio: ["ignore", "pipe", "pipe"] });
const exit = new Promise((resolve) => child.once("exit", () => { exited = true; resolve(); }));
try {
await timeout(new Promise((resolve, reject) => {
const receive = (data) => { logs = (logs + data).slice(-16000); if (logs.includes("Ready in")) resolve(); };
child.stdout.on("data", receive); child.stderr.on("data", receive);
child.once("error", reject); child.once("exit", (code) => reject(new Error(`Next ${mode} exited ${code}.`)));
}), 60000, `Next ${mode} readiness timeout.`);
const request = (url, command) => fetch(origin + url, {
...(command ? { method: "POST", body: JSON.stringify(command) } : {}),
headers: { origin, "sec-fetch-site": "same-origin", "x-learnharness-subscriptions": "1", "content-type": "application/json" },
signal: AbortSignal.timeout(60000),
});
// Controls have not been imported through an HTTP request yet.
const catalog = (await (await request("/api/models")).json()).providers;
for (const id of ["chatgpt", "github-copilot-subscription", "xai-subscription"]) assert(catalog.find((provider) => provider.id === id)?.keyPresent, `${mode}: cold ${id} unavailable.`);
let state = await (await request("/api/subscriptions")).json();
assert.equal(state.chatgpt.accounts[0].email, "fixture@example.invalid"); assert.equal(state.native.length, 2);
for (const secret of ["fixture-chatgpt-access", "fixture-chatgpt-refresh", "fixture-id-token", "fixture-xai-access", "fixture-refresh"]) assert(!JSON.stringify(state).includes(secret));
const stream = await fetch(origin + "/api/events", { signal: AbortSignal.timeout(60000) });
const reader = stream.body.getReader();
try {
const first = await reader.read(); assert(first.value);
assert.equal((await request("/api/subscriptions", { providerId: "chatgpt", op: "consent", enabled: false })).status, 200);
let events = new TextDecoder().decode(first.value);
while (!events.includes("subscriptions-changed")) {
const frame = await timeout(reader.read(), 5000, "Cross-route subscription SSE timeout.");
assert(!frame.done, "Subscription event stream ended early."); events += new TextDecoder().decode(frame.value);
}
} finally { await reader.cancel(); }
state = await (await request("/api/subscriptions")).json(); assert.equal(state.chatgpt.allowBackgroundUse, false);
const login = await (await request("/api/subscriptions", { providerId: "chatgpt", op: "login" })).json();
assert(login.browserAction); assert(!JSON.stringify(login).includes("auth.openai.com"));
if (mode === "dev") {
const file = path.join(ROOT, "src/server/web-subscriptions.ts"), times = await fs.stat(file);
await fs.utimes(file, new Date(), new Date());
try { assert.equal((await request("/api/models")).status, 200); assert.equal((await request("/api/subscriptions")).status, 200); }
finally { await fs.utimes(file, times.atime, times.mtime); }
}
const navigate = () => fetch(origin + login.browserAction, { redirect: "manual", headers: { "sec-fetch-site": "same-origin" }, signal: AbortSignal.timeout(30000) });
const browser = await navigate(); assert.equal(browser.status, 303, `${mode}: owner/action lost across routes or reload.`);
const location = new URL(browser.headers.get("location")); assert.equal(location.origin, "https://auth.openai.com"); assert.equal(new URL(location.searchParams.get("redirect_uri")).hostname, "127.0.0.1");
assert.equal((await navigate()).status, 410);
assert.equal((await request("/api/subscriptions", { providerId: "chatgpt", op: "cancel" })).status, 200);
assert.equal((await fetch(origin + "/api/subscriptions", { method: "POST", headers: { origin: "https://unrelated.invalid", "content-type": "application/json", "x-learnharness-subscriptions": "1" }, body: JSON.stringify({ providerId: "chatgpt", op: "disconnect", id: "offline-account" }) })).status, 403);
assert.equal((await request("/api/subscriptions", { providerId: "chatgpt", op: "disconnect", id: "offline-account" })).status, 200);
assert.equal((await (await request("/api/models")).json()).providers.find((provider) => provider.id === "chatgpt").keyPresent, false);
console.log(`[smoke-web-subscriptions] ${mode}: cold catalogs, safe status, cross-route SSE/consent/disconnect, opaque redirect/replay, CSRF${mode === "dev" ? ", owner retained across source reload" : ""} passed`);
} catch (error) { console.error(logs); throw error; }
finally {
if (!exited) child.kill("SIGTERM");
try { await timeout(exit, 5000, "Next shutdown timeout."); }
catch { child.kill("SIGKILL"); await exit; }
}
}
try { await run("start"); await run("dev"); }
finally { await fs.rm(profile, { recursive: true, force: true }); }
src/agent/chatgpt-bridge.ts
+13 -7
View File
@@ -1,4 +1,5 @@
import type { Model } from "@earendil-works/pi-ai";
import { onWebSubscriptionPacket, webSubscriptionRequest, webSubscriptionsEnabled } from "./subscription-runtime.ts";
import type { ChatGPTStatus } from "../core/contracts/chatgpt.ts";
import type { AdmissionScope, NoDispatchKind } from "./chatgpt-request-cycle.ts";
export type { ChatGPTStatus } from "../core/contracts/chatgpt.ts";
@@ -20,14 +21,15 @@ const validStamp = (s: Stamp) => s && [s.epoch, s.authorityRevision, s.revision]
function desktopIPCConnected(): boolean {
return process.env.LH_DESKTOP_SERVER === "1" && (process.env.LH_DESKTOP_API_SECRET?.length ?? 0) >= 64 && typeof process.send === "function" && process.connected === true;
}
function runtimeConnected(): boolean { return desktopIPCConnected() || webSubscriptionsEnabled(); }
function bridge(): BridgeState {
const global = globalThis as typeof globalThis & { [key]?: BridgeState };
if (!global[key]) {
const state: BridgeState = { status: unavailable(), stamp: { epoch: 0, authorityRevision: 0, revision: 0 }, models: [], catalogAccount: null, pending: new Map(), controllers: new Map(), listeners: new Set(), counter: 0, catalogRequest: 0 };
global[key] = state;
process.on("message", (raw: unknown) => {
const receive = (raw: unknown) => {
const m = raw as { channel?: string; kind?: string; id?: string; op?: string; ok?: boolean; value?: unknown; stamp: Stamp; status: ChatGPTStatus; invalidate?: string; error?: string };
if (!desktopIPCConnected() || m?.channel !== "lh:chatgpt") return;
if (m?.channel !== "lh:chatgpt") return;
if (m.kind === "state" && validStamp(m.stamp) && m.stamp.revision >= state.stamp.revision) {
acceptStatus(state, { ...m.stamp, status: m.status });
if (m.invalidate === "all") { state.catalogRequest++; clearCatalog(state); }
@@ -43,8 +45,11 @@ function bridge(): BridgeState {
if (m.ok === true) call.resolve(m.value);
else call.reject(new Error(m.error ?? "ChatGPT connection failed."));
}
});
};
process.on("message", (raw: unknown) => { if (desktopIPCConnected()) receive(raw); });
onWebSubscriptionPacket("lh:chatgpt", (packet) => { if (webSubscriptionsEnabled()) receive(packet); });
process.on("disconnect", () => {
if (webSubscriptionsEnabled()) return;
state.status = unavailable(); clearCatalog(state); state.catalogRequest++;
for (const controller of state.controllers.keys()) controller.abort("all");
for (const call of state.pending.values()) { clearTimeout(call.timer); call.reject(new Error("The desktop connection closed.")); }
@@ -60,9 +65,10 @@ function acceptStatus(state: BridgeState, value: StatusValue) {
if (value.epoch !== state.stamp.epoch || value.authorityRevision !== state.stamp.authorityRevision || value.status.activeId !== state.status.activeId) clearCatalog(state);
state.status = value.status; state.stamp = { epoch: value.epoch, authorityRevision: value.authorityRevision, revision: value.revision };
}
/** Runtime-only inherited IPC. No HTTP credential route or renderer method. */
/** Runtime-only: inherited desktop IPC or direct web owner. Never HTTP leases. */
export function chatGPTRequest<T>(op: "status" | "models" | "lease", fields: Record<string, unknown> = {}): Promise<T> {
const state = bridge();
if (webSubscriptionsEnabled()) return webSubscriptionRequest<T>("lh:chatgpt", op, fields);
if (!desktopIPCConnected()) return Promise.reject(new Error("ChatGPT sign-in is available only in the desktop app with a private server connection."));
return new Promise<T>((resolve, reject) => {
const id = `${process.pid}:${++state.counter}`;
@@ -74,7 +80,7 @@ export function chatGPTRequest<T>(op: "status" | "models" | "lease", fields: Rec
});
}
export function chatGPTStatus(): ChatGPTStatus { return bridge().status; }
export function chatGPTConnected(): boolean { const s = chatGPTStatus(); return desktopIPCConnected() && s.available && s.accounts.some((a) => a.id === s.activeId && a.connected && a.planEnabled === true); }
export function chatGPTConnected(): boolean { const s = chatGPTStatus(); return runtimeConnected() && s.available && s.accounts.some((a) => a.id === s.activeId && a.connected && a.planEnabled === true); }
export function chatGPTBackgroundAllowed(): boolean { const s = chatGPTStatus(); return chatGPTConnected() && s.allowBackgroundUse === true; }
export function onChatGPTChange(listener: () => void): () => void { bridge().listeners.add(listener); return () => { bridge().listeners.delete(listener); }; }
export function chatGPTModels(): Model<"openai-responses">[] { return bridge().models; }
@@ -86,7 +92,7 @@ export function chatGPTScope(selectedModelId: string): AdmissionScope {
}
export function chatGPTScopeCurrent(scope: AdmissionScope): boolean {
const s = bridge();
return desktopIPCConnected() && chatGPTConnected() && scope.accountId === s.catalogAccount && scope.accountId === s.status.activeId && scope.epoch === s.stamp.epoch && scope.authorityRevision === s.stamp.authorityRevision && !!chatGPTModel(scope.selectedModelId);
return runtimeConnected() && chatGPTConnected() && scope.accountId === s.catalogAccount && scope.accountId === s.status.activeId && scope.epoch === s.stamp.epoch && scope.authorityRevision === s.stamp.authorityRevision && !!chatGPTModel(scope.selectedModelId);
}
export async function chatGPTLease(scope: AdmissionScope, background: boolean): Promise<LeaseValue> {
if (!chatGPTScopeCurrent(scope)) throw new Error("ChatGPT catalog authority changed");
@@ -108,7 +114,7 @@ export async function chatGPTDenial(scope: AdmissionScope | undefined, controlle
/** Preserve account order; unknown capabilities/windows are never fabricated. */
export async function refreshChatGPTModels(supported: Model<"openai-responses">[]): Promise<void> {
const state = bridge();
if (!desktopIPCConnected()) { clearCatalog(state); return; }
if (!runtimeConnected()) { clearCatalog(state); return; }
const sequence = ++state.catalogRequest;
try {
const value = await chatGPTRequest<ModelsValue>("models");
src/agent/subagents/aborts.ts
+5 -7
View File
@@ -41,18 +41,16 @@ export function unregisterJobAbort(jobId: string): void {
* the liveness gate for the stale-job sweep: a "running" record with a
* live registration belongs to a loop this process owns (hung or slow, it
* is the one truth — never re-spawn beside it); without one the record is
* an orphan of a process death and recovers immediately. The span is
* loop-start to delivery-completion: the runner's unregistering finally
* runs only after deliver() resolves, by which point finalizeJob has
* flipped the record — so absence of registration is unambiguous. */
* an orphan of a process death and recovers immediately. The registration
* spans computation and cleanup only. Finished outcomes have independent,
* persisted delivery obligations; no worker waits for their consumption. */
export function isJobLive(jobId: string): boolean {
return registry().has(jobId);
}
/** Abort a running job's loop. Returns false when no controller is live —
* the job's loop is not executing (crashed, or in the delivering-result
* phase where the runner has already returned); the caller then owns the
* whole abort on the record side. */
* the job's loop is not executing. The endpoint separately refuses abort
* for a finished worker, even while its session delivery remains pending. */
export function fireJobAbort(jobId: string): boolean {
const controller = registry().get(jobId);
if (!controller) return false;
src/agent/subagents/extraction.ts
+13 -8
View File
@@ -28,7 +28,7 @@ import type { GraphStore } from "../../core/ports.ts";
import { emit } from "../../core/events.ts";
import { settleGraphLayout } from "../../core/layout-settle.ts";
import { batchRuns, coverageOf, messageViewLines } from "../../core/coverage.ts";
import type { LhMessage } from "../../core/native-messages.ts";
import { failedOutputCallIds, type LhMessage } from "../../core/native-messages.ts";
import { auditorGraphTools } from "../tools/graph.ts";
import { runSubmitAgent, healStringifiedArrays } from "./submit-agent.ts";
@@ -124,16 +124,20 @@ export function transcriptView(record: ChatRecord): string {
}
/** The extraction view under the teacher-prose ruling (owner, 2026-09-28):
* ONLY the teacher agent's text blocks — no learner messages, no
* thinking, no tool calls or results, no quiz blocks — each labeled with
* ONLY learner-facing teacher prose (text or send_to_user) — no learner
* messages, thinking, other tool calls/results, or quiz blocks — labeled with
* its ACTUAL message id (owner, 2026-10-13): the view lines ARE
* core/coverage.ts messageViewLines ("TEACHER [id] text"), the exact
* lines the coverage fingerprints hash, so the extractor can anchor
* every proposition to the passage it came from (the session submit's
* required messageId). The shared markdown engine stays reserved for the
* future user-facing export. */
function proseView(messages: ReadonlyArray<LhMessage>, title?: string): string {
const lines = messages.flatMap((message) => messageViewLines(message, message.id));
function proseView(
messages: ReadonlyArray<LhMessage>,
title?: string,
failedOutputCalls: ReadonlySet<string> = failedOutputCallIds(messages),
): string {
const lines = messages.flatMap((message) => messageViewLines(message, message.id, failedOutputCalls));
const head = title && title.length > 0 ? `# ${title}\n\n` : "";
return head + lines.join("\n\n");
}
@@ -1113,6 +1117,7 @@ export function createExtractionPipeline(deps: ExtractionPipelineDeps): Extracti
// stage runs (typed — the trigger skips, routes surface 503).
refuseIfStagesUnready();
const covered = coverageOf(record.messages, graphStore.readCoverage(), chatId);
const failedOutputCalls = failedOutputCallIds(record.messages);
if (covered.uncovered.length === 0) {
console.info(`[extraction ${chatId}] nothing uncovered — no-op`);
return { nodes: 0, chains: 0 };
@@ -1148,14 +1153,14 @@ export function createExtractionPipeline(deps: ExtractionPipelineDeps): Extracti
// fingerprints' messageViewLines ("TEACHER [id] text"), one
// shared seam.
const messages = batch.map((slot) => slot.message);
const view = proseView(messages, record.title);
const view = proseView(messages, record.title, failedOutputCalls);
// The batch's teacher-prose ids — the ONLY anchors this
// job's submit may name (validateSessionAnchors at the
// submit seam refuses anything else: ids from other
// batches — even real ones of this chat — other chats,
// and non-prose messages).
const batchIds = new Set(
messages.filter((m) => messageViewLines(m, m.id).length > 0).map((m) => m.id),
messages.filter((m) => messageViewLines(m, m.id, failedOutputCalls).length > 0).map((m) => m.id),
);
setPhase(jobKey, "queued", "chat", undefined, part);
const findings = await withExtractSlot(jobKey, async () => {
@@ -1188,7 +1193,7 @@ export function createExtractionPipeline(deps: ExtractionPipelineDeps): Extracti
const batchStats = mergeFindings(chatId, findings, auditResult, { chatStore, graphStore }, {
foldedAt,
coverage: batch
.filter((slot) => messageViewLines(slot.message, slot.message.id).length > 0)
.filter((slot) => slot.tokens > 0)
.map((slot) => ({ chatId, fp: slot.fp, tokens: slot.tokens, foldedAt })),
});
// The layout ledger evolves with the merge (sync — the
src/agent/subagents/pdf-reader.ts
+2 -1
View File
@@ -2,6 +2,7 @@ import { randomUUID } from "node:crypto";
import type { AssistantMessage, Message, ToolResultMessage, Usage, Tool } from "@earendil-works/pi-ai";
import { normalizeContext } from "@earendil-works/pi-ai";
import { Type } from "typebox";
import { resolveSourcePath } from "../tools/fs-shared.ts";
import { piEngineFor, type PiEngine, type ThinkingLevel } from "../models.ts";
import { providerEventObserver, type ProviderEventObserver } from "../provider-events.ts";
import type { PdfResult, PdfToolInput } from "../../core/contracts.ts";
@@ -166,7 +167,7 @@ export async function runPdfReaderAgent(
): Promise<PdfResult> {
const headless = opts.headless ?? getHeadless();
const maxIterations = opts.maxIterations ?? MAX_ITERATIONS;
const pdfPath = input.path;
const pdfPath = resolveSourcePath(input.path);
// Deterministic upfront failure: unreadable / unparseable
// for renders — the brief never starts on a broken source.
src/agent/subagents/runner.ts
+35 -27
View File
@@ -3,6 +3,7 @@ import type { SearchProvider } from "../../core/search.ts";
import {
accumulateJobUsage,
buildSubagentSteer,
isSubagentFailure,
piUsageToStep,
type DemoResult,
type DemoToolInput,
@@ -40,13 +41,13 @@ import { runVisualMaker } from "./visual-maker.ts";
* (the job tools' execute fires this via the spawn seam, fire-and-forget),
* updating the job's activity line on every iteration. On completion the
* runner builds the STEER MESSAGE (buildSubagentSteer) and hands it to
* `deliver`; the wake pipeline fires the server-side teacher turn and
* flips the job to done/error when the lesson has actually resumed.
* `deliver` without waiting. The worker's final outcome and pending
* delivery are saved together; the session owns delivery from then on.
*
* On failure the result is the failure shape (empty html/svg + "FAILED: …"
* notes; research: failure brief) — and it STILL steers, because the
* teacher must learn the outcome. If delivery itself fails, the job is
* marked error (the learner's activity card shows it).
* teacher must learn the outcome. Failed delivery starts stay quietly
* pending and never rewrite the worker's computation outcome.
*/
/**
@@ -67,9 +68,9 @@ export interface JobRunnerDeps {
chatStore: ChatStore;
jobStore: JobStore;
/**
* Hand a finished job's steer to the wake pipeline — resolves when the
* steer is delivered (the wake turn completed and the job status
* flipped); rejects when delivery fails unrecoverably.
* Hand a finished job's steer to the session delivery registry. The
* promise tracks durable marker acceptance, but the worker never waits
* on it. Its persisted pending obligation survives process restart.
*/
deliver: (steer: SteerMessage) => Promise<void>;
/** Settings-resolved engine per role (model slots wave, 2026-09-20):
@@ -96,8 +97,8 @@ export interface JobRunnerDeps {
export interface JobRunner {
/** Run one job's sub-agent loop to completion (background; never throws).
* The tool input rides along from dispatch (the spawn fires mid-stream,
* before the tool part is persisted — only crash-recovery re-spawns
* resolve the input from the transcript). */
* before the tool part is persisted). Recovery reads saved job input,
* with transcript fallback for legacy jobs. Delivery is not awaited. */
runJob(jobId: string, input?: JobToolInput): Promise<void>;
}
@@ -270,14 +271,15 @@ export function createJobRunner(deps: JobRunnerDeps): JobRunner {
}
}
/** Resolve the job's tool input: dispatch passes it directly; only a
* crash-recovery re-spawn (no input) falls back to scanning the
* transcript (by then onFinish has long persisted the part). */
/** Dispatch input wins, then the job's saved arguments. Legacy jobs
* without saved input fall back to the durable transcript, which may
* lack a call if the dispatching turn died before settlement. */
function resolveInput(
job: JobRecord,
input: JobToolInput | undefined,
): JobToolInput | undefined {
if (input) return input;
if (job.input !== undefined) return job.input as JobToolInput;
const chat = chatStore.getChat(job.chatId);
if (!chat) return undefined;
// Wave D: the record is native history — the scan reads the native
@@ -295,27 +297,28 @@ export function createJobRunner(deps: JobRunnerDeps): JobRunner {
}
/**
* The tail of every job: store the result, build the steer, hand it to
* the wake pipeline. The pipeline flips the status when the lesson has
* actually resumed — until then the job reads "running" so the client's
* activity card keeps polling and re-syncs once, at the end.
* Save the computation outcome and session delivery obligation together.
* Worker liveness ends after cleanup, not after a lesson consumes the result.
*/
async function finishJob(jobId: string, result: ResearchResult | DemoResult | VisualResult | PdfResult): Promise<void> {
const job = jobStore.getJob(jobId);
if (!job) return;
// Aborted mid-run or mid-delivery (abortJob flips the record before
// Aborted mid-run (abortJob flips the record before
// the loop unwinds): no result write, no steer — the abort path owns
// the record and queues the abort steer itself.
if (job.status !== "running") return;
jobStore.updateJob(jobId, { result, activity: "delivering result" });
const steer = buildSubagentSteer(job, result);
try {
await deliver(steer);
} catch (err) {
const reason = err instanceof Error ? err.message : String(err);
console.error(`[job ${jobId}] steer delivery failed:`, err);
jobStore.updateJob(jobId, { status: "error", error: reason, activity: "delivery failed" });
}
const failed = isSubagentFailure(job.kind, result);
jobStore.updateJob(jobId, {
result,
status: failed ? "error" : "done",
activity: failed ? "failed" : "finished",
delivery: { kind: "result", status: "pending" },
});
// The saved result now belongs to the session. Do not keep the worker,
// abort registration, or temporary resources alive behind a question.
void deliver(buildSubagentSteer(job, result)).catch((err: unknown) => {
console.error(`[job ${jobId}] pending delivery:`, err);
});
}
/** Chat-record + input resolution shared by all three runners. */
@@ -343,7 +346,12 @@ export function createJobRunner(deps: JobRunnerDeps): JobRunner {
const { job } = ctx;
const input = (passedInput ?? resolveInput(job, passedInput)) as ResearchToolInput | undefined;
const question = typeof input?.question === "string" && input.question.trim() ? input.question : "(no question)";
if (!(typeof input?.question === "string" && input.question.trim())) {
jobStore.updateJob(jobId, { error: "research question unavailable" });
await finishJob(jobId, { brief: "Research failed: the research question is unavailable.", sources: [] });
return;
}
const question = input.question;
const context = typeof input?.context === "string" && input.context.trim() ? input.context : undefined;
const prompt = [
src/agent/subscription-bridge.ts
+10 -3
View File
@@ -1,4 +1,5 @@
import type { Api, Model, ModelAuth } from "@earendil-works/pi-ai";
import { onWebSubscriptionPacket, webSubscriptionRequest, webSubscriptionsEnabled } from "./subscription-runtime.ts";
import { NATIVE_SUBSCRIPTIONS, type NativeSubscriptionProviderId, type NativeSubscriptionScope, type NativeSubscriptionSnapshot, type NativeSubscriptionStatus } from "../core/contracts/subscriptions.ts";
import type { AdmissionScope, NoDispatchKind } from "./chatgpt-request-cycle.ts";
@@ -18,6 +19,7 @@ export function nativeSubscriptionId(logicalId: string): NativeSubscriptionProvi
function desktopConnected(): boolean {
return process.env.LH_DESKTOP_SERVER === "1" && (process.env.LH_DESKTOP_API_SECRET?.length ?? 0) >= 64 && typeof process.send === "function" && process.connected === true;
}
function runtimeConnected(): boolean { return desktopConnected() || webSubscriptionsEnabled(); }
function notify(state: State) { for (const listener of state.listeners) { try { listener(); } catch { /* observer isolation */ } } }
function invalidate(state: State, providerId?: NativeSubscriptionProviderId) {
for (const [controller, owner] of state.controllers) if (!providerId || owner.providerId === providerId) controller.abort("all");
@@ -60,7 +62,11 @@ function bridge(): State {
if (message.ok) pending.resolve(message.value);
else pending.reject(new Error(message.error ?? "Subscription connection failed."));
});
onWebSubscriptionPacket("lh:subscriptions", (packet) => {
if (webSubscriptionsEnabled() && packet.snapshot) accept(state, packet.snapshot as NativeSubscriptionSnapshot);
});
process.on("disconnect", () => {
if (webSubscriptionsEnabled()) return;
invalidate(state);
for (const pending of state.pending.values()) { clearTimeout(pending.timer); pending.reject(new Error("The desktop connection closed.")); }
state.pending.clear(); notify(state);
@@ -70,6 +76,7 @@ function bridge(): State {
}
export function subscriptionRequest<T>(op: "snapshot" | "lease", fields: Record<string, unknown> = {}): Promise<T> {
const state = bridge();
if (webSubscriptionsEnabled()) return webSubscriptionRequest<T>("lh:subscriptions", op, fields);
if (!desktopConnected()) return Promise.reject(new Error("Subscription sign-in is available only in the desktop app with a private server connection."));
return new Promise<T>((resolve, reject) => {
const id = `${process.pid}:${++state.counter}`;
@@ -82,17 +89,17 @@ export function subscriptionRequest<T>(op: "snapshot" | "lease", fields: Record<
}
export async function refreshSubscriptionModels(): Promise<void> {
const state = bridge();
if (!desktopConnected()) { invalidate(state); return; }
if (!runtimeConnected()) { invalidate(state); return; }
try { accept(state, await subscriptionRequest<NativeSubscriptionSnapshot>("snapshot")); }
catch { invalidate(state); notify(state); }
}
export function nativeSubscriptionStatus(providerId: NativeSubscriptionProviderId): NativeSubscriptionStatus | undefined { return bridge().statuses.get(providerId); }
export function nativeSubscriptionConnected(providerId: NativeSubscriptionProviderId): boolean {
const status = nativeSubscriptionStatus(providerId);
return desktopConnected() && status?.available === true && status.connected && !!status.connectionId;
return runtimeConnected() && status?.available === true && status.connected && !!status.connectionId;
}
export function nativeSubscriptionBackgroundAllowed(providerId: NativeSubscriptionProviderId): boolean { return nativeSubscriptionConnected(providerId) && nativeSubscriptionStatus(providerId)?.allowBackgroundUse === true; }
export function nativeSubscriptionModels(): Model<Api>[] { return desktopConnected() ? [...bridge().models.values()].flat() : []; }
export function nativeSubscriptionModels(): Model<Api>[] { return runtimeConnected() ? [...bridge().models.values()].flat() : []; }
export function nativeSubscriptionModel(logicalId: string, modelId: string): Model<Api> | undefined { return nativeSubscriptionModels().find((m) => m.provider === logicalId && m.id === modelId); }
export function onNativeSubscriptionChange(listener: () => void): () => void { const state = bridge(); state.listeners.add(listener); return () => { state.listeners.delete(listener); }; }
export function subscriptionScope(providerId: NativeSubscriptionProviderId, selectedModelId: string): NativeSubscriptionScope {
src/agent/subscription-runtime.ts
+30
View File
@@ -0,0 +1,30 @@
/** Server-injected private web transport. The agent layer never imports a
* filesystem owner or server composition root. Desktop IPC remains separate. */
export type SubscriptionRuntimeChannel = "lh:chatgpt" | "lh:subscriptions";
export type SubscriptionRuntimePacket = Record<string, unknown>;
interface Runtime {
request?: <T>(channel: SubscriptionRuntimeChannel, op: string, fields: Record<string, unknown>) => Promise<T>;
listeners: Map<SubscriptionRuntimeChannel, Set<(packet: SubscriptionRuntimePacket) => void>>;
}
const KEY = Symbol.for("lh.webSubscriptionRuntime");
function runtime(): Runtime {
const global = globalThis as typeof globalThis & { [KEY]?: Runtime };
return global[KEY] ??= { listeners: new Map() };
}
export function configureWebSubscriptionRuntime(request: NonNullable<Runtime["request"]>): void { runtime().request = request; }
export function webSubscriptionsEnabled(): boolean {
return process.env.LH_DESKTOP_SERVER !== "1" && !process.env.LH_DESKTOP_API_SECRET && !!runtime().request;
}
export function onWebSubscriptionPacket(channel: SubscriptionRuntimeChannel, listener: (packet: SubscriptionRuntimePacket) => void): void {
const state = runtime(); const listeners = state.listeners.get(channel) ?? new Set();
listeners.add(listener); state.listeners.set(channel, listeners);
}
export function publishWebSubscriptionPacket(channel: SubscriptionRuntimeChannel, packet: SubscriptionRuntimePacket): void {
for (const listener of runtime().listeners.get(channel) ?? []) {
try { listener(packet); } catch { /* observer isolation */ }
}
}
export function webSubscriptionRequest<T>(channel: SubscriptionRuntimeChannel, op: string, fields: Record<string, unknown>): Promise<T> {
if (!webSubscriptionsEnabled()) return Promise.reject(new Error("The web subscription owner is unavailable."));
return runtime().request!<T>(channel, op, fields);
}
src/agent/teacher-agent.ts
+26 -12
View File
@@ -15,10 +15,11 @@ import {
type TeacherToolContext,
} from "./tools.ts";
import { practiceTools, type PracticeToolContext } from "./tools/practice.ts";
import { healToolRecord } from "./tools/heal.ts";
import { healToolRecord, withArgumentHeal } from "./tools/heal.ts";
import { sendToUserTool } from "./tools/send-to-user.ts";
import { quizInputObject, quizInputProblems } from "../core/question-core.ts";
import { QUESTION_TOOLS, type EnabledTools } from "../core/pending.ts";
import { toLlmMessages, nativeInputCause } from "../core/native-messages.ts";
import { toLlmMessages, nativeInputCause, PRACTICE_PROBLEM_TOOL } from "../core/native-messages.ts";
import type { ChatGPTRequestCycle } from "./chatgpt-request-cycle.ts";
import { usageWasAdmitted } from "./subscription-request-cycle.ts";
@@ -145,22 +146,27 @@ export interface AgentBuildOptions {
* belt-and-suspenders HALF for the mixed batch (a question + other calls
* in one message) that terminate-only semantics would let run on — the
* learner cannot answer while a turn is in flight (per-chat 409), so a
* cleanly-posed question MUST always end the turn. maxTurns is gone
* cleanly-posed question MUST always end the turn. Practice's problem pose
* has the same wait-for-the-learner contract, including mixed output/problem
* batches; failed poses must still retry. maxTurns is gone
* (owner ruling 2026-09-25: no maximum-agent-turns net on the teaching
* turn).
*/
function makeQuestionStopper() {
function makeQuestionStopper(kind: AgentKind) {
return (context: {
message: { content: Array<{ type: string; id?: string; name?: string }> };
toolResults: Array<{ toolCallId: string; isError?: boolean }>;
}): boolean => {
const called = new Set(
context.message.content
.filter((c) => c.type === "toolCall" && c.name !== undefined && (QUESTION_TOOLS as readonly string[]).includes(c.name!))
.filter((c) => c.type === "toolCall" && c.name !== undefined && (
(QUESTION_TOOLS as readonly string[]).includes(c.name) ||
(kind === "practice" && c.name === PRACTICE_PROBLEM_TOOL)
))
.map((c) => c.id!),
);
if (called.size === 0) return false;
// Stop only when every question call of this turn resolved cleanly.
// Stop only when every question/problem pose resolved cleanly.
for (const result of context.toolResults) {
if (called.has(result.toolCallId) && result.isError) return false;
}
@@ -174,7 +180,7 @@ function baseAgentOptions(kind: AgentKind, opts: AgentBuildOptions, tools: Agent
// conversation-spec resolution as teaching); the cast keeps AgentKind
// ("practice" is not a ModelRole) honest about that.
const engine = opts.engine ?? piEngineFor(kind as "prober" | "teacher");
const stopper = makeQuestionStopper();
const stopper = makeQuestionStopper(kind);
const systemPrompt = opts.promptSwitches
? composeInstructions(kind, opts.promptSwitches)
: composedInstructions[kind]!;
@@ -190,7 +196,16 @@ function baseAgentOptions(kind: AgentKind, opts: AgentBuildOptions, tools: Agent
systemPrompt,
model: engine.model,
thinkingLevel: engine.level,
tools,
// Conditional registry inclusion, like the graph-tool gates: when
// closed, neither the tool nor its usage instructions reach the model.
// Match Claude across providers (including OpenRouter), not the API
// transport. Explicit test registries remain caller-owned.
tools: [
...tools,
...(opts.tools === undefined && engine.model.id.toLowerCase().includes("claude")
? [withArgumentHeal(sendToUserTool)]
: []),
],
messages: opts.messages ?? [],
},
streamFn: engine.streamFn,
@@ -312,10 +327,9 @@ export function bareQuestionTools(): Record<string, AgentTool<any>> {
* from turn 1 of a kind:"practice" session — no prober→teacher split,
* no plan tools, no question tools. Its ENTIRE registry is
* problem + conclude (tools/practice.ts); the doctrine prompt
* (prompts/practice/) carries the method. The question-stop rule does
* not apply (problem rides terminate: true natively —
* QUESTION_TOOLS does not include it), so baseAgentOptions's stopper is
* a no-op here; it stays wired for uniformity.
* (prompts/practice/) carries the method. A successful problem pose ends
* the run, natively for a problem-only batch and via the shared stop hook
* for a mixed batch (for example, send_to_user + problem).
*/
export function createPracticeAgent(ctx: PracticeToolContext, opts: AgentBuildOptions = {}) {
const tools = opts.tools ?? practiceTools(ctx);
src/agent/tools/jobs.ts
+1 -1
View File
@@ -78,7 +78,7 @@ function dispatch(
input: ResearchInput | DemoInput | VisualInput | PdfInput,
): string {
const existing = ctx.jobStore.findJobByToolCallId(ctx.chatId, toolCallId);
const job = existing ?? ctx.jobStore.createJob({ kind, chatId: ctx.chatId, toolCallId });
const job = existing ?? ctx.jobStore.createJob({ kind, chatId: ctx.chatId, toolCallId, input });
if (!existing) ctx.spawn(job, input);
return `Dispatched (job ${job.id}). ${ack}`;
}
src/agent/tools/probe.ts
+3 -2
View File
@@ -42,8 +42,9 @@ function probeConcluded(messages: readonly LhMessage[]): boolean {
* live), so the agent sees the fixed PROBE_DONE_ACK_TEXT — never an
* echo. details = {ok:true} only.
*
* The first probe_done in the transcript is also the construction seam:
* from that message on, every request builds the TEACHER agent.
* The first concluded probe_done (non-failed call plus matching successful
* result) is also the construction seam: from that result on, every request
* builds the TEACHER agent. Bare calls and refused results keep probing.
*/
export function probeDoneTool(ctx: ProbeDoneContext): Record<string, AgentTool<any>> {
return {
src/agent/tools/send-to-user.ts
+36
View File
@@ -0,0 +1,36 @@
import { Type } from "typebox";
import type { AgentTool } from "@earendil-works/pi-agent-core";
import { SEND_TO_USER_TOOL } from "../../core/native-messages.ts";
import { serverExecute } from "../tool-execute.ts";
const sendToUserInput = Type.Object({
message: Type.String({
minLength: 1,
pattern: "\\S",
description: "The complete content to display to the learner. Markdown and LaTeX are supported.",
}),
});
/** Claude's documented verbatim mid-turn output channel. Availability is
* model-gated at the shared conversation-agent construction boundary;
* the native call remains in history, and the view renders its message. */
export const sendToUserTool: AgentTool<typeof sendToUserInput> = {
name: SEND_TO_USER_TOOL,
label: SEND_TO_USER_TOOL,
description:
"## REQUIRED OUTPUT CHANNEL — ALWAYS USE `send_to_user`\n\n" +
"**ALL learner-facing messages MUST be sent using `send_to_user`.** This includes explanations, teaching, " +
"feedback, corrections, hints, introductions, progress messages, and final responses. " +
"There is no exception for short messages or the end of a turn.\n\n" +
"Ordinary response text may be converted into summarized thinking instead of reaching the learner. " +
"**Never rely on ordinary response text to communicate with the learner. Thinking is not delivery.**\n\n" +
"Put the complete message in `message`. Markdown and LaTeX are supported. " +
"Do not duplicate it in ordinary response text.\n\n" +
"Question content still belongs in the appropriate question tool; send any accompanying teaching or " +
"commentary through `send_to_user` before calling that tool.\n\n" +
"This tool does not pause or end your turn. Continue with other tools when needed. " +
"After delivering your final message, end the turn without additional ordinary response text.",
parameters: sendToUserInput,
executionMode: "sequential",
execute: serverExecute(SEND_TO_USER_TOOL, sendToUserInput, () => "Message received for display."),
};
src/core/answer-presentation.ts
+57
View File
@@ -0,0 +1,57 @@
/** Read-only provenance for historical question cards. Never persist this
* projection: answer identities, ordering, grading, and learner text remain
* the recorded facts; only model-authored display text comes from the call. */
import { normalizeOptions, type QuestionAnswerData } from "./question-core.ts";
import { parseAskInput, parseExplainInput, parseQuizInput } from "./tool-parts.ts";
import { repairMathForDisplay } from "./math-delims.ts";
export function answerForPresentation(answer: QuestionAnswerData, input: unknown): QuestionAnswerData {
if (answer.dismissed) return answer;
const quiz = answer.kind === "quiz" ? parseQuizInput(input) : undefined;
const explain = answer.kind === "explain" ? parseExplainInput(input) : undefined;
const source = quiz ?? explain ?? (answer.kind === "ask" ? parseAskInput(input) : undefined);
if (!source) return answer;
const projected = { ...answer, questionText: source.question };
if (quiz) projected.explanation = quiz.explanation;
if (explain) projected.committed = explain.correctAnswer;
if (!answer.options?.length || !("options" in source) || !source.options?.length) return projected;
const original = normalizeOptions(source.options);
const saved = answer.options;
const bindings = new Map<number, number>(); // saved position -> original position
const used = new Set<number>();
const bindUnique = (candidates: number[][]) => {
for (let si = 0; si < saved.length; si++) {
if (bindings.has(si)) continue;
const possible = candidates[si]!.filter((oi) => !used.has(oi));
if (possible.length !== 1) continue;
const oi = possible[0]!;
// Two saved rows claiming the same source are ambiguous, even if
// each row independently has just one candidate.
if (candidates.some((other, sj) => sj !== si && !bindings.has(sj) && other.includes(oi))) continue;
bindings.set(si, oi);
used.add(oi);
}
};
// Selected-answer values are stable identities when the record retained
// them. Do not normalize those values, recompute a shuffle, or regrade.
bindUnique(saved.map((option) => {
const values = answer.answers.filter((a) => a.index === option.index).map((a) => a.value);
return original.flatMap((o, oi) => values.includes(o.value) ? [oi] : []);
}));
const recoveredLabels = original.map((o) => repairMathForDisplay(o.label));
bindUnique(saved.map((option) => original.flatMap((o, oi) =>
o.label === option.label || recoveredLabels[oi] === option.label ? [oi] : [])));
// A bijection with one remaining pair has no ordering ambiguity. With
// multiple unmatched rows, keep their snapshots rather than guess.
if (saved.length === original.length && bindings.size === saved.length - 1) {
const si = saved.findIndex((_, i) => !bindings.has(i));
const oi = original.findIndex((_, i) => !used.has(i));
bindings.set(si, oi);
}
projected.options = saved.map((option, si) => {
const oi = bindings.get(si);
return oi === undefined ? option : { ...option, label: original[oi]!.label };
});
return projected;
}
src/core/compaction.ts
+3 -3
View File
@@ -272,9 +272,9 @@ export function newCompactionMarker(summary: string, details: CompactionDetails)
};
}
/** The ACTIVE compaction marker: the newest one in the record (chained
* compactions — the last one rules; older markers fall inside the
* dropped span and never render). */
/** The ACTIVE compaction marker: the newest one with a retained-message
* id. Its summary supersedes older markers even when their stored
* positions remain inside its retained tail. */
export function activeCompactionOf(messages: readonly LhMessage[]): (LhMarkerMessage & { details: CompactionDetails }) | null {
for (let i = messages.length - 1; i >= 0; i--) {
const m = messages[i]!;
src/core/content-reader.ts
+215
View File
@@ -0,0 +1,215 @@
/** Source-preserving boundaries for tool-authored Markdown. Recovery never
* mutates input, reserializes Markdown, or enters protected literal content. */
import katex from "katex";
import { fromMarkdown } from "mdast-util-from-markdown";
import { mathFromMarkdown } from "mdast-util-math";
import { math } from "micromark-extension-math";
export type ContentSpan =
| { kind: "prose" | "code"; raw: string }
| { kind: "math"; raw: string; open: string; body: string; close: string; display: boolean };
const blankLine = /^\r?\n[ \t]*\r?\n/;
const escaped = (text: string, i: number) => {
let n = 0; while (i > 0 && text[--i] === "\\") n++;
return n % 2 === 1;
};
const runLength = (text: string, i: number, char: string) => { let n = 0; while (text[i + n] === char) n++; return n; };
type MarkdownNode = { type: string; position?: { start: { offset?: number }; end: { offset?: number } }; children?: MarkdownNode[] };
/** Use the rendering stack's actual CommonMark+math grammar for literal
* ranges, including nested quotes/lists and implicit container endings.
* Keep source offsets: never serialize the AST back into Markdown. */
function literalRanges(text: string): Array<{ start: number; end: number }> {
const pending: MarkdownNode[] = [fromMarkdown(text, { extensions: [math({ singleDollarTextMath: true })], mdastExtensions: [mathFromMarkdown()] })];
const ranges: Array<{ start: number; end: number }> = [];
while (pending.length) {
const node = pending.pop()!;
if (node.type === "code" || node.type === "inlineCode") {
const start = node.position?.start.offset, end = node.position?.end.offset;
if (start !== undefined && end !== undefined) ranges.push({ start, end });
} else if (node.children) pending.push(...node.children);
}
return ranges.sort((a, b) => a.start - b.start);
}
/** Mirrors the app's native dollar-run and legacy-delimiter conventions,
* with code and blank paragraphs as inline boundaries. */
export function readContentSpans(text: string): ContentSpan[] {
const spans: ContentSpan[] = [];
const literals = literalRanges(text);
let literal = 0, i = 0, plain = 0;
const opaque = (end: number) => {
if (plain < i) spans.push({ kind: "prose", raw: text.slice(plain, i) });
spans.push({ kind: "code", raw: text.slice(i, end) }); i = plain = end;
};
while (i < text.length) {
while (literal < literals.length && literals[literal]!.end <= i) literal++;
const barrier = literals[literal]?.start ?? Infinity;
if (i >= barrier) { opaque(literals[literal]!.end); continue; }
let open = "", close = "", display = false, after = -1, at = -1;
if (text[i] === "\\" && !escaped(text, i) && ["(", "["].includes(text[i + 1] ?? "")) {
open = text.slice(i, i + 2); close = open === "\\(" ? "\\)" : "\\]"; display = open === "\\[";
for (let j = i + 2; j < text.length - 1; j++) {
if (j >= barrier || (!display && blankLine.test(text.slice(j)))) break;
if (text.startsWith(close, j) && !escaped(text, j)) { at = j; after = j + 2; break; }
}
} else if (text[i] === "$" && !escaped(text, i)) {
const size = runLength(text, i, "$"); open = "$".repeat(size);
const eol = text.indexOf("\n", i + size);
const before = text.slice(text.lastIndexOf("\n", i - 1) + 1, i);
display = size >= 2 && /^[ \t]{0,3}$/.test(before) && !text.slice(i + size, eol < 0 ? text.length : eol).includes("$");
if (display) {
let start = eol;
while (start >= 0 && start < text.length) {
start++;
const next = text.indexOf("\n", start);
const match = /^([ \t]{0,3})(\$+)[ \t]*\r?$/.exec(text.slice(start, next < 0 ? text.length : next));
if (match && match[2]!.length >= size) { at = start + match[1]!.length; close = match[2]!; after = at + close.length; break; }
start = next;
}
} else {
for (let j = i + size; j < text.length;) {
if (j >= barrier || blankLine.test(text.slice(j))) break;
if (text[j] === "$" && !escaped(text, j)) {
const n = runLength(text, j, "$");
if (n === size) { at = j; close = open; after = j + n; break; }
j += n;
} else j++;
}
}
}
if (at >= 0) {
if (plain < i) spans.push({ kind: "prose", raw: text.slice(plain, i) });
spans.push({ kind: "math", raw: text.slice(i, after), open, body: text.slice(i + open.length, at), close, display });
i = plain = after;
} else i += open.length || 1;
}
if (plain < text.length) spans.push({ kind: "prose", raw: text.slice(plain) });
return spans;
}
export function mapMathContent(text: string, transform: (body: string, display: boolean) => string): string {
return readContentSpans(text).map(span => span.kind === "math" ? span.open + transform(span.body, span.display) + span.close : span.raw).join("");
}
const knownCommands = new Map<string, boolean>();
/** Probe only a command name, never arbitrary content. Missing-argument errors
* still identify a known command. No trust, macros, IO, or persistent TeX state. */
export function knownTexCommand(name: string): boolean {
if (!/^[A-Za-z]+$/.test(name)) return false;
const cached = knownCommands.get(name); if (cached !== undefined) return cached;
let known = true;
try { katex.renderToString("\\" + name, { throwOnError: true, trust: false, strict: false }); }
catch (error) { known = !String(error).includes("Undefined control sequence: \\" + name); }
knownCommands.set(name, known);
return known;
}
export function canParseTex(body: string, display: boolean): boolean {
try { katex.renderToString(body, { throwOnError: true, trust: false, strict: false, displayMode: display }); return true; }
catch { return false; }
}
export interface TexRange { start: number; end: number; kind: "row" | "encoded" | "comment"; factor?: number }
export function numericPercentEscape(inner: string, percent: number): boolean {
let start = percent; while (start > 0 && inner[start - 1] === "\\") start--;
if (percent - start < 2 || (percent - start) % 2 !== 0 || !/\d[ \t]*$/.test(inner.slice(0, start))) return false;
const tail = inner.slice(percent + 1);
// A final source-formatting LF before the closing fence is not a next
// equation row. A meaningful following line or comment text remains literal.
if (/^[\s.)\]}]*$/.test(tail)) return true;
return !/[\r\n]/.test(tail) && !/^[ \t]*[A-Za-z]/.test(tail);
}
/** TeX boundaries shared by document decoding and command recovery. Native
* rows are literal; consistently extra-escaped environment pairs supply an
* explicit escaping factor for both commands AND row separators. */
export function texRanges(inner: string): TexRange[] {
const ranges: TexRange[] = [];
const stack: Array<{ name: string; start: number; factor: number }> = [];
let i = 0;
while (i < inner.length) {
if (inner[i] === "%" && !escaped(inner, i)) {
let start = i; while (start > 0 && inner[start - 1] === "\\") start--;
const factor = stack.at(-1)?.factor ?? ranges.filter(r => r.kind === "encoded" && i >= r.start && i < r.end).at(-1)?.factor;
const encodedLiteral = factor && factor > 1 && (i - start) % factor === 0 && ((i - start) / factor) % 2 === 1;
const native = stack.some(frame => frame.factor === 1) || ranges.some(r => r.kind === "row" && i >= r.start && i < r.end);
if (encodedLiteral || (!native && numericPercentEscape(inner, i))) { i++; continue; }
const newline = inner.indexOf("\n", i);
const end = newline < 0 ? inner.length : newline + 1;
ranges.push({ start: i, end, kind: "comment" }); i = end; continue;
}
if (inner[i] !== "\\") { i++; continue; }
let end = i + 1; while (inner[end] === "\\") end++;
const count = end - i;
const marker = /^(begin|end)\s*\{([^{}]+)\}/.exec(inner.slice(end));
if (marker) {
const factor = count % 2 === 1 ? 1 : count;
if (marker[1] === "begin") stack.push({ name: marker[2]!, start: i, factor });
else if (stack.at(-1)?.name === marker[2] && stack.at(-1)?.factor === factor) {
const open = stack.pop()!;
ranges.push({ start: open.start, end: end + marker[0].length, kind: factor === 1 ? "row" : "encoded", factor });
}
i = end + marker[0].length; continue;
}
// KaTeX's substack macro expands its braced argument into subarray:
// these are row separators even without literal begin/end markers.
const substack = /^substack\s*\{/.exec(inner.slice(end));
if (substack) {
let depth = 1, j = end + substack[0].length;
for (; j < inner.length && depth > 0; j++) {
if (inner[j] === "%" && !escaped(inner, j)) { const lf = inner.indexOf("\n", j); j = lf < 0 ? inner.length : lf; continue; }
if (!escaped(inner, j)) { if (inner[j] === "{") depth++; else if (inner[j] === "}") depth--; }
}
if (depth === 0) ranges.push({ start: i, end: j, kind: count % 2 === 1 ? "row" : "encoded", factor: count % 2 === 1 ? 1 : count });
}
i = end;
}
return ranges.sort((a, b) => a.start - b.start);
}
export function outsideTexComments(body: string, transform: (part: string) => string): string {
let from = 0; const out: string[] = [];
for (const range of texRanges(body).filter(r => r.kind === "comment")) {
out.push(transform(body.slice(from, range.start)), body.slice(range.start, range.end)); from = range.end;
}
out.push(transform(body.slice(from))); return out.join("");
}
/** Document escapes are not arbitrary string decoding. Complete known TeX
* words win over the prefix `\\n`; code is never inspected or rewritten. */
export function recoverDocumentLayout(text: string): string {
const spans = readContentSpans(text);
const candidates = spans.filter(span => span.kind !== "code").map(span => span.raw).join("\u0000");
const declared = new Set([
...candidates.matchAll(/\\(?:newcommand|renewcommand|providecommand)\*?\s*(?:\{\s*)?\\(n[A-Za-z]+)/g),
...candidates.matchAll(/\\(?:def|gdef|edef)\s*\\(n[A-Za-z]+)/g),
].map(m => m[1]!));
const knownN = (name: string) => declared.has(name) || knownTexCommand(name);
const tokens = [...candidates.matchAll(/(\\+)(n[A-Za-z]*)/g)];
const evidence = tokens.filter(m => m[1]!.length % 2 === 1 && !knownN(m[2]!));
const structural = evidence.length >= 2 || candidates.includes("\\n\\n") || (evidence.length > 0 &&
[...candidates.matchAll(/(\\+)([A-Za-z]+)/g)].some(m => m[1]!.length % 2 === 0 && knownTexCommand(m[2]!)));
if (!structural) return text;
const decode = (part: string) => part.replace(/(\\+)(n[A-Za-z]*)/g, (whole: string, slashes: string, name: string, offset: number) => {
if (slashes.length % 2 === 0 || knownN(name)) return whole;
const word = part.slice(0, offset).split(/\s/).at(-1) ?? "";
if (/^(?:[A-Za-z]:[\\/]|https?:\/\/|\/)/.test(word + whole)) return whole;
return slashes.slice(0, -1) + "\n" + name.slice(1);
});
const projected = spans.map(span => ({ span, body: span.kind === "math" ? outsideTexComments(span.body, decode) : span.kind === "code" ? span.raw : decode(span.raw) }));
let out = "";
for (let i = 0; i < projected.length; i++) {
const { span, body } = projected[i]!;
if (span.kind !== "math") { out += body; continue; }
// Decoding document layout can expose a flow opener whose close was
// written on the formula line. Frame BOTH ends; otherwise Markdown
// consumes the closing dollars as TeX and introduces a rendering error.
if (/^\${2,}$/.test(span.open) && body !== span.body && body.includes("\n")) {
if (!/^[ \t]{0,3}$/.test(out.slice(out.lastIndexOf("\n") + 1))) out += "\n\n";
out += span.open + (body.startsWith("\n") ? "" : "\n") + body + (body.endsWith("\n") ? "" : "\n") + span.close;
const next = projected[i + 1];
if (next && !/^[\r\n]/.test(next.span.kind === "math" ? next.span.open : next.body)) out += "\n\n";
} else out += span.open + body + span.close;
}
return out;
}
src/core/contracts/jobs.ts
+28 -32
View File
@@ -21,6 +21,12 @@ export interface JobRecord {
chatId: string;
/** The tool part this job fulfills. */
toolCallId: string;
/** Dispatch-time arguments, saved before spawn. Older jobs resolve from history. */
input?: unknown;
/** Session-owned delivery, independent of the worker's terminal outcome.
* Absent on legacy records: old terminal jobs must never be replayed. */
delivery?: { kind: "result" | "abort"; status: "pending" | "delivered" };
/** Worker computation only; a pending delivery does not keep it running. */
status: JobStatus;
/** Latest human-readable activity line, e.g. `web_search: "sigma.js v3"`. */
activity: string;
@@ -148,42 +154,32 @@ export function accumulateJobUsage(
* sourced briefs) — tracking a live activity line in the job record.
* ONE central client poll of GET /api/jobs?chatId=… (in Chat) drives every
* job surface: the transcript dispatch cards (live activity by
* toolCallId, quiet one-liner after delivery), the sub-agents sidebar
* toolCallId, quiet one-liner after worker completion), the sub-agents sidebar
* panel, and the single re-sync when a job settles.
*
* 3. On completion the runner builds a STEER MESSAGE (buildSubagentSteer)
* and hands it to the wake pipeline (src/server/jobs.ts): a user-role
* result-carrying native marker — the content is what the MODEL needs
* (the research brief, the builder's notes), plus a data-subagent-artifact
* part with what the UI renders (the full result: brief + sources /
* SVG / demo HTML). Provider conversion drops data parts, so artifact
* payloads never enter the teacher's context. metadata.subagentResult
* identifies the job; the UI renders these as sub-agent result cards,
* NEVER as learner interjections.
* 3. Dispatch saves input on the job before spawn. On completion the
* runner saves result + terminal worker status + pending delivery in
* one job update, then queues buildSubagentSteer WITHOUT waiting for
* delivery. The native custom marker's content is model-facing text;
* details carry the full artifact payload for the UI. Worker resources
* are released even if a question keeps the result pending indefinitely.
* Only computing workers appear in Background agents or accept abort.
*
* 4. The wake pipeline delivers the steer by firing a SERVER-SIDE teacher
* turn through the same turn composition as a client request
* (handleChatRequest) — write-ahead persist, turn liveness, extraction
* trigger: one code path. A per-chat serialized queue: a steer waits
* while a turn is in flight OR a question tool part is pending (an
* unanswered client-side tool call breaks provider conversion — the
* steer fires right after the learner's continuation), then fires; a 409
* (a client turn raced in) retries. Delivery is idempotent per jobId, so
* a crashed wake can be re-entered by the stale-job recovery sweep. The
* job's status flips to done/error only AFTER the wake turn completes —
* a polling client that sees the flip re-syncs once and gets the
* finished record.
* 4. Session delivery owns one registry entry per deterministic steer ID,
* including while a wake is in flight or a 409/403 returns its cargo.
* A live agent takes it via pi steering; registration restores saved
* pending jobs before the answer run's first poll. Otherwise the idle
* wake uses the normal turn composition and respects pending questions,
* dismissal, and background consent. A saved transcript marker completes
* delivery; HTTP 200 or streamed visibility alone does not.
*
* CONTINUATION COALESCING: the idle window above can starve — a
* quiz-chaining teacher ends every turn with a new pending check, so
* the gates never stay open long enough for a wake to fire. The fix:
* when a client turn wins the turn registry, the handler CLAIMS every
* queued, undelivered steer (jobs.ts claimSteers) and appends it after
* the learner's answer in that very turn — conversion-safe, because the
* answered quiz carries its output. Every answer flushes pending
* steers immediately; the wake pump remains for turns that end with no
* pending question. The pump re-checks steerDelivered right before
* firing, so a claim racing the pump converges to exactly one delivery.
* Failed starts remain quietly pending until a learner run or restart
* recovery, without new timers, visible delivery errors, or worker-status
* changes. Existing 409 backoff and consent deferral remain. Accepted
* turns are never automatically replayed. Legacy running records with a
* result become finished/pending; legacy terminal records with no
* delivery metadata remain untouched. Session deletion discards pending
* deliveries rather than treating them as running work.
*
* 5. No pending job parts ever exist, so the composer never gates on jobs
* (the learner can interject mid-job; the steer queues behind their
src/core/coverage.ts
+25 -27
View File
@@ -1,9 +1,9 @@
/**
* Extraction coverage — the pure core. THE TEACHER'S PROSE IS THE UNIT
* (owner ruling, 2026-09-28): the extractor sees ONLY the teacher agent's
* text blocks — no learner messages, no thinking, no tool calls or
* results, no quiz blocks. Everything else is INVISIBLE to coverage: it
* neither opens nor joins a run, so answering a quiz or editing a learner
* prose — ordinary text blocks and explicit send_to_user messages. No
* learner messages, thinking, other tool calls/results, or quiz blocks.
* Everything else is INVISIBLE to coverage: it neither opens nor joins a run, so answering a quiz or editing a learner
* message triggers nothing. "Covered" is by construction exactly "what
* the extractor saw": the fingerprint (sha256 of a message's view lines)
* and the extractor's prompt share ONE selection — the teacher prose.
@@ -22,33 +22,27 @@
import { createHash } from "node:crypto";
import type { CoverageEntry } from "./contracts/graph.ts";
import type { LhMessage } from "./native-messages.ts";
import { failedOutputCallIds, userFacingTextOf, type LhMessage } from "./native-messages.ts";
export type { CoverageEntry };
/** The user-facing text of a native message (string or text blocks). */
function textOf(content: unknown): string[] {
if (typeof content === "string") return content.trim() ? [content] : [];
if (!Array.isArray(content)) return [];
return content
.filter((c): c is { type: "text"; text: string } => (c as { type?: string }).type === "text")
.map((c) => c.text)
.filter((t) => t.trim().length > 0);
}
/**
* The extraction-relevant lines of ONE native message — the shared seam
* between the extractor's prompt and messageFingerprint (coverage).
* TEACHER PROSE ONLY (owner, 2026-09-28): an assistant message yields one
* line per text block; everything else (user messages, tool results,
* markers) yields NOTHING and is invisible to coverage. `mid` is the
* message id — part of the hash input, so the same prose in two messages
* line per learner-facing prose block; everything else (user messages,
* tool results, markers) yields NOTHING and is invisible to coverage.
* `mid` is the message id — part of the hash input, so the same prose in two messages
* is two slots (per-message coverage).
*/
export function messageViewLines(message: LhMessage, mid: string): string[] {
export function messageViewLines(
message: LhMessage,
mid: string,
failedOutputCalls: ReadonlySet<string> = failedOutputCallIds([message]),
): string[] {
const lines: string[] = [];
if ((message as { role?: string }).role !== "assistant") return lines;
for (const text of textOf((message as { content: unknown }).content)) {
for (const text of userFacingTextOf((message as { content: unknown }).content, failedOutputCalls)) {
lines.push(`TEACHER [${mid}] ${text.trim()}`);
}
return lines;
@@ -62,9 +56,10 @@ export function messageViewLines(message: LhMessage, mid: string): string[] {
*/
export function teacherProseBlocks(messages: ReadonlyArray<LhMessage>): string[] {
const out: string[] = [];
const failedOutputCalls = failedOutputCallIds(messages);
for (const message of messages) {
if ((message as { role?: string }).role !== "assistant") continue;
for (const text of textOf((message as { content: unknown }).content)) out.push(text.trim());
for (const text of userFacingTextOf((message as { content: unknown }).content, failedOutputCalls)) out.push(text.trim());
}
return out;
}
@@ -79,9 +74,9 @@ export function viewTokens(lines: string[]): number {
* teacher prose, with the message id as the tag). Changes exactly when
* the teacher prose of the message changes — nothing else can re-open it.
*/
export function messageFingerprint(message: LhMessage): string {
export function messageFingerprint(message: LhMessage, failedOutputCalls?: ReadonlySet<string>): string {
const h = createHash("sha256");
h.update(JSON.stringify([(message as { role?: string }).role, messageViewLines(message, message.id)]));
h.update(JSON.stringify([(message as { role?: string }).role, messageViewLines(message, message.id, failedOutputCalls)]));
return h.digest("hex");
}
@@ -95,7 +90,12 @@ export interface CoverageSlot {
/** Fingerprint a whole transcript, in order (used by the trigger, the run,
* and the coverage view — one derivation everywhere). */
export function coverageSlots(messages: ReadonlyArray<LhMessage>): CoverageSlot[] {
return messages.map((message) => ({ message, fp: messageFingerprint(message), tokens: viewTokens(messageViewLines(message, message.id)) }));
const failedOutputCalls = failedOutputCallIds(messages);
return messages.map((message) => ({
message,
fp: messageFingerprint(message, failedOutputCalls),
tokens: viewTokens(messageViewLines(message, message.id, failedOutputCalls)),
}));
}
/** The covered-fingerprint set for one chat, from the ledger. */
@@ -131,16 +131,14 @@ export function coverageOf(
let uncoveredTokens = 0;
for (let i = 0; i < slots.length; i++) {
const slot = slots[i]!;
if (messageViewLines(slot.message, slot.message.id).length === 0) continue; // invisible
if (slot.tokens === 0) continue; // invisible (including failed output calls)
if (covered.has(slot.fp)) coveredTokens += slot.tokens;
else {
uncovered.push(i);
uncoveredTokens += slot.tokens;
}
}
const visibleCount = slots.length - slots.filter(
(s) => messageViewLines(s.message, s.message.id).length === 0,
).length;
const visibleCount = slots.filter((s) => s.tokens > 0).length;
return {
slots,
uncovered,
src/core/dictation.ts
+90
View File
@@ -0,0 +1,90 @@
/** Nova-3 monolingual language codes, including documented locale aliases.
* Source: https://developers.deepgram.com/docs/models-languages-overview
* Multilingual is a separate mode; its supported language set differs.
*/
const NOVA_3_LANGUAGES = [
["Afrikaans", ["af", "af-ZA"]],
["Arabic", ["ar", "ar-AE", "ar-SA", "ar-QA", "ar-KW", "ar-SY", "ar-LB", "ar-PS", "ar-JO", "ar-EG", "ar-SD", "ar-TD", "ar-MA", "ar-DZ", "ar-TN", "ar-IQ", "ar-IR"]],
["Armenian", ["hy"]],
["Assamese", ["as", "as-IN"]],
["Belarusian", ["be"]],
["Bengali", ["bn"]],
["Bosnian", ["bs"]],
["Bulgarian", ["bg"]],
["Catalan", ["ca"]],
["Chinese (Cantonese, Traditional)", ["zh-HK"]],
["Chinese (Mandarin, Simplified)", ["zh", "zh-CN", "zh-Hans"]],
["Chinese (Mandarin, Traditional)", ["zh-TW", "zh-Hant"]],
["Croatian", ["hr"]],
["Czech", ["cs", "cs-CZ"]],
["Danish", ["da", "da-DK"]],
["Dutch", ["nl"]],
["English", ["en", "en-US", "en-AU", "en-GB", "en-IN", "en-NZ"]],
["Estonian", ["et"]],
["Finnish", ["fi"]],
["Flemish", ["nl-BE"]],
["French", ["fr", "fr-CA"]],
["Georgian", ["ka", "ka-GE"]],
["German", ["de"]],
["German (Switzerland)", ["de-CH"]],
["Greek", ["el"]],
["Gujarati", ["gu", "gu-IN"]],
["Hebrew", ["he"]],
["Hindi", ["hi"]],
["Hungarian", ["hu"]],
["Indonesian", ["id"]],
["Italian", ["it"]],
["Japanese", ["ja"]],
["Kannada", ["kn"]],
["Kazakh", ["kk", "kk-KZ"]],
["Korean", ["ko", "ko-KR"]],
["Latvian", ["lv"]],
["Lithuanian", ["lt"]],
["Macedonian", ["mk"]],
["Malay", ["ms"]],
["Marathi", ["mr"]],
["Mongolian", ["mn"]],
["Nepali", ["ne"]],
["Norwegian", ["no"]],
["Pashto", ["ps", "ps-AF"]],
["Persian", ["fa"]],
["Polish", ["pl"]],
["Portuguese", ["pt", "pt-BR", "pt-PT"]],
["Punjabi", ["pa", "pa-IN"]],
["Romanian", ["ro"]],
["Russian", ["ru"]],
["Serbian", ["sr"]],
["Slovak", ["sk"]],
["Slovenian", ["sl"]],
["Spanish", ["es", "es-419"]],
["Swedish", ["sv", "sv-SE"]],
["Tagalog", ["tl"]],
["Tamil", ["ta"]],
["Telugu", ["te"]],
["Thai", ["th", "th-TH"]],
["Turkish", ["tr", "tr-TR"]],
["Ukrainian", ["uk"]],
["Urdu", ["ur"]],
["Vietnamese", ["vi"]],
] as const;
export type DictationLanguage = "multi" | (typeof NOVA_3_LANGUAGES)[number][1][number];
export const DICTATION_LANGUAGES: readonly { code: DictationLanguage; label: string }[] = [
{ code: "multi", label: "Multilingual" },
...NOVA_3_LANGUAGES.flatMap(([name, codes]) =>
codes.map((code, i) => ({ code, label: i === 0 ? name : `${name} (${code})` })),
).sort((a, b) => a.label.localeCompare(b.label, "en")),
];
export function isDictationLanguage(value: unknown): value is DictationLanguage {
return DICTATION_LANGUAGES.some(({ code }) => code === value);
}
/** Only language varies. Keep the owner's tuned streaming parameters. */
export function deepgramListenUrl(language: DictationLanguage): string {
return "wss://api.deepgram.com/v1/listen" +
"?model=nova-3&encoding=linear16&sample_rate=16000&channels=1" +
"&smart_format=true&punctuate=true&endpointing=300" +
`&language=${encodeURIComponent(language)}`;
}
src/core/events.ts
+3
View File
@@ -38,6 +38,8 @@ export interface ServerEvents {
"activity-changed": Record<string, never>;
/** The session list changed (create, delete, title). */
"chats-changed": Record<string, never>;
/** Web subscription state changed. No account or credential data on the bus. */
"subscriptions-changed": Record<string, never>;
/** A chat's record changed WITHOUT a turn (the silent dismissal write).
* Same notify-then-fetch contract as the turn events: open views of
* the chat refetch and adopt the durable record. */
@@ -61,6 +63,7 @@ export const SERVER_EVENT_TYPES = [
"jobs-changed",
"activity-changed",
"chats-changed",
"subscriptions-changed",
"record-changed",
"compaction-changed",
] as const satisfies readonly ServerEventType[];
src/core/index.ts
+2
View File
@@ -47,6 +47,7 @@ export * from "./input-events.ts";
// the deleted server projection to the client side. Pure, browser-safe.
export * from "./view.ts";
export * from "./tool-parts.ts";
export * from "./answer-presentation.ts";
export * from "./prices.ts";
export * from "./context-usage.ts";
export type { ChatGPTStatus, ChatGPTControlBridge } from "./contracts/chatgpt.ts";
@@ -58,6 +59,7 @@ export * from "./fuzzy.ts";
// the user-settings record's shape/sanitizer (pure) and its fully-resolved
// view. Browser-safe — the Wave-4 panel imports these directly.
export * from "./settings.ts";
export * from "./dictation.ts";
// graph-deletion.ts (wave 6): the whole module is PURE but storage-port
// driven — the deletion RUNS server-side only. The shell needs exactly one
// two type-only things from it: the ChatDeleteMode union (the delete
src/core/input-events.ts
+66
View File
@@ -16,6 +16,9 @@
* continue — run the next turn on the current record (mount
* continuations, job-settle re-syncs: the record ends at an
* answered question with no teacher reaction yet).
* retry — retry a failed model response (issue #110), without new
* input. Failed attempts stay in the record; only the retry's
* temporary agent history omits the trailing failures.
*
* (The plan verdict input died 2026-09-26 with the verdict itself — the
* learner's post-plan reply is a natural text message.)
@@ -45,6 +48,8 @@ export type ChatInput =
| { type: "text"; text: string }
| { type: "answer"; toolCallId: string; answer: QuestionAnswerData }
| { type: "continue" }
/** The failed-turn retry (issue #110): the banner's Retry button. */
| { type: "retry" }
/** The practice dock's two buttons (owner, 2026-10-04): the learner's
* thinking on the ACTIVE problem — one lane (owner 2026-10-05):
* whether it claims an answer is read from the text. One
@@ -170,6 +175,8 @@ export function parseChatInput(input: unknown): ChatInput {
return { type: "answer", toolCallId: i.toolCallId, answer: i.answer as QuestionAnswerData };
case "continue":
return { type: "continue" };
case "retry":
return { type: "retry" };
case "thinking":
if (typeof i.problemCallId !== "string" || i.problemCallId.length === 0) {
throw new ChatInputError("thinking input requires problemCallId");
@@ -210,6 +217,10 @@ export function inputAppends(
}
case "continue":
return dismissalConclusion(record);
case "retry":
// The input already landed write-ahead. retryDrops affects only
// the temporary agent history — nothing new appends here.
return [];
case "thinking": {
const active = activeProblemOf(record.messages);
if (!active || active.toolCallId !== input.problemCallId) {
@@ -233,6 +244,61 @@ export function inputAppends(
}
}
/** The failure owning the ephemeral banner. Plan presentation and
* compaction do not resolve an error; substantive later input/output does.
* This is deliberately broader than Retry eligibility below. */
export function failedResponseAtEnd(messages: readonly LhMessage[]): Extract<LhMessage, { role: "assistant" }> | undefined {
for (let i = messages.length - 1; i >= 0; i--) {
const m = messages[i]!;
if (m.role === "custom" && (m.customType === "plan-block" || m.customType === "compaction")) continue;
return m.role === "assistant" && m.stopReason === "error" ? m : undefined;
}
return undefined;
}
/** Does the record end in a failure with input that can be retried?
* Aborts are deliberate stops; compaction after a failure requires new
* learner input rather than a retry using a changed context. */
export function endsInFailedTurn(messages: readonly LhMessage[]): boolean {
const { ids, before } = failedTail(messages);
return ids.size > 0 && before !== undefined && before.role !== "assistant";
}
/** The ids omitted ONLY from the retry's temporary agent history.
* Failed attempts (including partial text and their ids) stay on disk.
* Pi's provider conversion already skips error/aborted responses, but
* continue() rejects an assistant tail BEFORE conversion. Omitting the
* trailing failures lets that continuation start from the saved input or
* completed tool results without replaying tools or resubmitting input.
* Plan blocks may follow the failure and stay; later compaction blocks
* Retry without changing compaction itself. */
export function retryDrops(messages: readonly LhMessage[]): Set<string> {
const { ids, before } = failedTail(messages);
if (ids.size === 0) {
const failure = failedResponseAtEnd(messages);
if (failure) {
throw new ChatInputError(`${failure.errorMessage ?? "An error occurred."}\n\nThis response cannot be retried after compaction. Send a new message to continue.`);
}
throw new ChatInputError("retry input requires a failed turn to retry");
}
if (!before || before.role === "assistant") throw new ChatInputError("retry input leaves no input to continue from");
return ids;
}
/** Walk back over failed assistants and UI-only plan blocks. A compaction
* is a boundary, not a marker Retry may cross. A failure AFTER a compaction
* can still retry: that compacted context was already its starting state. */
function failedTail(messages: readonly LhMessage[]): { ids: Set<string>; before: LhMessage | undefined } {
const ids = new Set<string>();
let i = messages.length - 1;
for (; i >= 0; i--) {
const m = messages[i]!;
if (m.role === "assistant" && m.stopReason === "error") ids.add(m.id);
else if (!(m.role === "custom" && m.customType === "plan-block")) break;
}
return { ids, before: messages[i] };
}
/** The deferred conclude synthesis (owner ruling, 2026-10-04): a
* problem-dismissal marker waits for the user's next input — the pair
* (synthetic conclude call + ack) rides ahead of it, so the agent reads
src/core/math-delims.ts
+107 -558
View File
@@ -1,572 +1,121 @@
/**
* Legacy math delimiter preprocessing — the \(...\)/\[...\] reader.
*
* Current authoring uses `$...$` inline and `$$...$$` on separate lines for
* display (owner, 2026-10-12). This pure transform keeps correctly formed
* old backslash-delimited content readable; dollar math passes through to
* Streamdown's native parser. Escape restoration below recognizes both
* syntaxes while retaining the existing recovery rules.
*
* Rules:
* - `\(...\)` → `$$...$$` (two-dollar inline, still accepted by micromark),
* `\[...\]` → `$$\n...\n$$` (block form: flow math).
* - An escaped delimiter (`\\(` = literal backslash + paren) is left alone.
* - Nothing inside fenced code blocks or inline code spans is touched.
* - Unbalanced delimiters are left untouched: Streamdown streams chunks, and
* an incomplete `\[` completes on a later chunk (the whole text is
* re-transformed each render).
*
* Pure string → string. No markdown parsing — a single linear scan with a
* code-state machine, cheap enough to run in the render path of a memo'd
* component.
*/
/** Shared Markdown/TeX boundaries for legacy delimiter compatibility and
* display-only tool-content recovery. Source arguments are never rewritten. */
import {
canParseTex, knownTexCommand, mapMathContent, numericPercentEscape, outsideTexComments,
readContentSpans, recoverDocumentLayout, texRanges,
} from "./content-reader.ts";
const INLINE_OPEN = "\\(";
const INLINE_CLOSE = "\\)";
const DISPLAY_OPEN = "\\[";
const DISPLAY_CLOSE = "\\]";
const FENCE_OPEN_RE = /^ {0,3}(`{3,}|~{3,})/;
/**
* Is the backslash at `i` itself escaped (i.e. preceded by an odd number of
* backslashes)? If so, the `\(`/`\[` starting there is a literal backslash
* followed by a bracket — not a delimiter.
*/
function isEscaped(text: string, i: number): boolean {
let count = 0;
for (let j = i - 1; j >= 0 && text[j] === "\\"; j--) count += 1;
return count % 2 === 1;
}
/** Length of the backtick run starting at `i` (0 if `text[i]` is not a tick). */
function backtickRun(text: string, i: number): number {
let count = 0;
while (text[i + count] === "`") count += 1;
return count;
}
/**
* Try to find the closing delimiter for the opening one at `i` (2 chars,
* already verified unescaped). The close scan stops at inline code spans and
* fence starts (barriers — math never contains them), and returns -1 when the
* delimiter is unbalanced (mid-stream chunk).
*/
function findClose(
text: string,
from: number,
close: string,
): number {
for (let j = from; j < text.length - 1; j++) {
if (text[j] === "`") return -1; // inline code barrier
if (
text[j] === "\\" &&
text[j + 1] === close[1] &&
!isEscaped(text, j)
) {
return j;
}
}
return -1;
}
/** True when `i` sits at the start of a line (position 0 or right after \n). */
function atLineStart(text: string, i: number): boolean {
return i === 0 || text[i - 1] === "\n";
}
/**
* Copy a fenced code block starting at `i` (line start, fence regex matched)
* verbatim to the output. Returns the index after the closing fence line —
* or the end of the text when the fence is unclosed (mid-stream: everything
* after is literal code until the chunk completes).
*/
function copyFencedBlock(text: string, i: number, out: string[]): number {
const open = FENCE_OPEN_RE.exec(text.slice(i));
const marker = open?.[1] ?? "```";
const closeRe = new RegExp(`^ {0,3}\\${marker[0]}{${marker.length},}[ \\t]*$`);
let j = i;
// Copy lines until the matching closing fence.
while (j < text.length) {
const lineEnd = text.indexOf("\n", j);
const next = lineEnd === -1 ? text.length : lineEnd + 1;
const line = text.slice(j, lineEnd === -1 ? text.length : lineEnd);
out.push(text.slice(j, next));
if (j > i && closeRe.test(line)) return next; // closing fence line
j = next;
}
return text.length;
}
/** Is the `\[` at `i` block-positioned (own line, only whitespace before)? */
function displayIsBlockPositioned(text: string, i: number): boolean {
for (let j = i - 1; j >= 0; j--) {
const ch = text[j];
if (ch === "\n") return true;
if (ch !== " " && ch !== "\t") return false;
}
return true;
}
/** Is the `\[` at `i` indented (whitespace before it on the line)? */
function displayIsIndented(text: string, i: number): boolean {
let sawWhitespace = false;
for (let j = i - 1; j >= 0; j--) {
const ch = text[j];
if (ch === "\n") return sawWhitespace;
if (ch !== " " && ch !== "\t") return false; // mid-paragraph
sawWhitespace = true;
}
return sawWhitespace;
}
/** Does the closing delimiter end its line (only \n or EOF after it)? */
function closeEndsLine(text: string, afterClose: number): boolean {
return afterClose >= text.length || text[afterClose] === "\n";
}
/**
* Rewrite balanced `\(...\)` → `$$...$$` and `\[...\]` → `$$\n...\n$$`.
* Everything else passes through byte-for-byte.
*/
/** The escaped-dollar-aware inline closer (owner, 2026-10-12): micromark's
* mathText closer is code-span logic — ANY size-matched `$` run closes,
* `\$` included, so `$x = \$5$` closes at the `\$` and errors. KaTeX's
* own auto-render scanner skips escaped dollars, and KaTeX renders `\$`
* fine — the author's intent is unambiguous: an escaped dollar (odd number
* of backslashes before it) never closes the span. Mirrors findDollarClose's
* inline rules otherwise (exact size match, whole-run skip). */
function findInlineDollarCloseEscapeAware(text: string, from: number): number {
for (let j = from; j < text.length;) {
if (text[j] !== "$") {
j += 1;
continue;
}
if (isEscaped(text, j)) {
j += 1;
continue; // `\$` never closes
}
const count = dollarRun(text, j);
if (count === 1) return j;
j += count;
}
return -1;
}
/** Replace `\$` (odd-backslash dollars) inside a math span with the
* dollar-free `\text{\textdollar}` — KaTeX-safe in math mode (bare
* `\textdollar` is text-mode-only there). Even-backslash dollars (`\\$`,
* row break + dollar) are left alone. */
function replaceEscapedDollars(content: string): string {
let replaced = "";
for (let j = 0; j < content.length;) {
if (content[j] === "\\" && content[j + 1] === "$" && !isEscaped(content, j)) {
replaced += "\\text{\\textdollar}";
j += 2;
} else {
replaced += content[j];
j += 1;
}
}
return replaced;
}
export function transformMathDelimiters(markdown: string): string {
const out: string[] = [];
let i = 0;
while (i < markdown.length) {
const ch = markdown[i]!;
// Fenced code blocks: copied verbatim (only detectable at line start).
if (atLineStart(markdown, i) && FENCE_OPEN_RE.test(markdown.slice(i))) {
i = copyFencedBlock(markdown, i, out);
continue;
}
// Inline code spans: copied verbatim.
if (ch === "`") {
const run = backtickRun(markdown, i);
let j = i + run;
// Find the matching run (same length, per CommonMark).
while (j < markdown.length) {
const nextRun = backtickRun(markdown, j);
if (nextRun === run) break;
j += Math.max(nextRun, 1);
}
if (j < markdown.length) {
out.push(markdown.slice(i, j + run));
i = j + run;
} else {
// Unclosed run (mid-stream): literal, keep scanning after it.
out.push(markdown.slice(i, i + run));
i += run;
}
continue;
}
// Math delimiters.
if (ch === "\\" && !isEscaped(markdown, i)) {
const next = markdown[i + 1];
if (next === "(" || next === "[") {
const isDisplay = next === "[";
const open = isDisplay ? DISPLAY_OPEN : INLINE_OPEN;
const close = isDisplay ? DISPLAY_CLOSE : INLINE_CLOSE;
const closeAt = findClose(markdown, i + open.length, close);
if (closeAt !== -1) {
const content = markdown.slice(i + open.length, closeAt).trim();
if (content.length > 0) {
if (isDisplay && displayIsIndented(markdown, i)) {
// Indented display math (inside a list item, quote, …):
// an indented $$ flow block half-parses inside list
// continuations (empty math, eaten items), so degrade to
// the inline form — renders inline, stays inside the item.
// mathText cannot span line endings, so collapse them.
out.push(`$$${content.replace(/\s*\n\s*/g, " ")}$$`);
} else if (isDisplay) {
// Flow math ($$ alone on its lines) renders as display
// KaTeX. Mid-paragraph \[...\] (or trailing text after the
// closer) would half-parse into literal dollars + KaTeX
// errors — pad with blank lines to force a clean block.
// Block-positioned math passes through in place.
const afterClose = closeAt + close.length;
const padBefore =
displayIsBlockPositioned(markdown, i) ? "" : "\n\n";
const padAfter =
closeEndsLine(markdown, afterClose) ? "" : "\n\n";
out.push(`${padBefore}$$\n${content}\n$$${padAfter}`);
} else {
// Preserve the legacy reader's two-dollar inline output
// (renders inline, not display).
out.push(`$$${content}$$`);
}
i = closeAt + close.length;
continue;
}
}
// Unbalanced or empty: leave both characters untouched.
}
}
// Escaped dollars inside inline math (owner, 2026-10-12): normalize
// BEFORE tokenization — inside a balanced inline span, `\$` becomes the
// dollar-free `\text{\textdollar}` so the scanner's closer is the one
// the author meant (micromark would otherwise close at the `\$`, error
// out, and cascade). Display runs are copied whole and left untouched —
// a lone `$` can't close `$$`, so `\$` already survives there — and
// copying the run stops its second `$` posing as an inline opener.
// Spans without escaped dollars emit byte-identical text; prose `\$`,
// code, and unbalanced spans are untouched.
if (ch === "$" && !isEscaped(markdown, i)) {
const run = dollarRun(markdown, i);
if (run === 1) {
const closeAt = findInlineDollarCloseEscapeAware(markdown, i + 1);
if (closeAt !== -1) {
const content = markdown.slice(i + 1, closeAt);
const replaced = replaceEscapedDollars(content);
if (replaced !== content) {
out.push(`$${replaced}$`);
i = closeAt + 1;
continue;
}
}
} else {
out.push(markdown.slice(i, i + run));
i += run;
continue;
}
}
out.push(ch);
i += 1;
}
return out.join("");
}
/**
* LaTeX-escape restoration (owner ruling 2026-09-26) — the JSON-debris fix.
*
* When a model emits LaTeX with SINGLE backslashes inside JSON tool-call
* arguments (`\forall`, not the properly escaped `\\forall`), JSON escape
* resolution eats the command's first letter: `\f` → form feed + "orall",
* `\v` → vertical tab + "arphi", `\t` → tab + "o". The debris survives into
* the persisted RAW args and renders as garbage. Inside a MATH SPAN these
* control chars can never be legitimate content, so restoring the two-char
* escape reconstructs the model's intent byte-perfectly. Outside spans the
* same chars might be real prose (pasted tables, formatted text), so only
* balanced math spans are touched. The legacy `\(…\)` / `\[…\]` pass stays
* unchanged; the dollar reader adds `$…$` and fenced `$$…$$` recognition,
* excluding code and escaped opening dollars. Unbalanced spans pass through.
*
* ASYMMETRIC NEWLINE RULE (owner ruling 2026-09-27, Option A): inline spans
* restore all six debris chars — real newlines/tabs inside INLINE math
* essentially never occur (options and inline math are single-line by
* nature). Display spans restore five and KEEP real newlines — multi-line
* display math (aligned environments, equation systems) is common, and a
* real line break is byte-identical on the wire to sloppy `\neq`/`\nu`/
* `\neg`/`\nabla` (JSON `\n` resolves to a newline either way). Corrupting
* legitimate content is strictly worse than failing to repair that rarest
* slice, so display-span n-command sloppiness stays broken by design.
*/
const ESCAPE_DEBRIS: Record<string, string> = {
"\b": "\\b",
"\t": "\\t",
"\n": "\\n",
"\v": "\\v",
"\f": "\\f",
"\r": "\\r",
const escaped = (text: string, i: number) => {
let n = 0; while (i > 0 && text[--i] === "\\") n++;
return n % 2 === 1;
};
const DEBRIS_RE = /[\b\t\n\v\f\r]/;
/** Display spans keep real newlines (multi-line equations) — debris there is
* everything except \n (a lone real \r never occurs in authored text, and
* `\right`/`\rho`/`\rangle` corruption is common in display math). */
const DISPLAY_DEBRIS_RE = /[\b\t\v\f\r]/;
/** Shared display-only repair. Keep raw arguments and model history unchanged. */
export function repairMathForDisplay(text: string): string {
return healOverescapedCommands(restoreMathEscapes(text));
/** Preserve the legacy list-continuation policy: indented display math
* degrades to inline rather than disrupting list structure. Literal code
* indentation is already protected by the shared reader. */
function indented(text: string, i: number): boolean {
const prefix = text.slice(text.lastIndexOf("\n", i - 1) + 1, i);
return prefix.length > 0 && /^[ \t]+$/.test(prefix);
}
function blockPosition(text: string, i: number): boolean {
return /^[ \t]*$/.test(text.slice(text.lastIndexOf("\n", i - 1) + 1, i));
}
/** micromark closes native single-dollar spans at escaped dollars too.
* Replace an intended literal dollar with equivalent dollar-free TeX before
* tokenization. Code, prose, unbalanced spans and display dollars stay raw. */
function replaceEscapedDollars(content: string): string {
let out = "";
for (let i = 0; i < content.length; i++) {
if (content[i] === "\\" && content[i + 1] === "$" && !escaped(content, i)) {
out += "\\text{\\textdollar}"; i++;
} else out += content[i];
}
return out;
}
/** Read old \(...\)/\[...\] syntax without touching literal content or
* reparsing native math as Markdown. Dollar authoring remains native. */
export function transformMathDelimiters(markdown: string): string {
let offset = 0;
return readContentSpans(markdown).map(span => {
const start = offset; offset += span.raw.length;
if (span.kind !== "math") return span.raw;
if (span.open === "$") return "$" + replaceEscapedDollars(span.body) + span.close;
if (span.open !== "\\(" && span.open !== "\\[") return span.raw;
const body = span.body.trim();
if (!body) return span.raw;
if (span.open === "\\(") return `$$${body}$$`;
if (indented(markdown, start)) return `$$${body.replace(/\s*\n\s*/g, " ")}$$`;
const before = blockPosition(markdown, start) ? "" : "\n\n";
const end = start + span.raw.length;
const after = end === markdown.length || markdown[end] === "\n" ? "" : "\n\n";
return `${before}$$\n${body}\n$$${after}`;
}).join("");
}
const DEBRIS: Record<string, string> = {
"\b": "\\b", "\t": "\\t", "\v": "\\v", "\f": "\\f", "\r": "\\r",
};
/** Recover JSON-swallowed command prefixes only in recognized math, never
* in code or genuine TeX comments. Restore all debris in one pass. Physical
* display newlines and CRLF remain source whitespace. Inline LF recovery is
* limited to complete known n-command names. Under the approved typical-
* model-authoring policy this includes newline+u→nu in inline math only;
* already-correct commands are never decoded into whitespace. */
export function restoreMathEscapes(text: string): string {
if (!DEBRIS_RE.test(text)) return text;
return restoreDollarMathEscapes(restoreLegacyMathEscapes(text));
if (!/[\b\t\n\v\f\r]/.test(text)) return text;
return mapMathContent(text, (body, display) => outsideTexComments(body, part => {
const restored = part.replace(/[\b\t\v\f\r]/g, (char, offset: number) => {
if (char === "\r" && part[offset + 1] === "\n") return char;
const suffix = /^[A-Za-z]+/.exec(part.slice(offset + 1))?.[0];
return suffix && knownTexCommand(DEBRIS[char]!.slice(1) + suffix) ? DEBRIS[char]! : char;
});
if (display) return restored;
return restored.replace(/\n([A-Za-z]+)/g, (whole, suffix: string) =>
knownTexCommand("n" + suffix) ? "\\n" + suffix : whole);
}));
}
/** Keep legacy recovery byte-for-byte; only the recognized syntax expands. */
function restoreLegacyMathEscapes(text: string): string {
if (!DEBRIS_RE.test(text)) return text;
const out: string[] = [];
let i = 0;
while (i < text.length) {
const ch = text[i]!;
if (ch === "\\" && !isEscaped(text, i)) {
const next = text[i + 1];
if (next === "(" || next === "[") {
const isDisplay = next === "[";
const open = isDisplay ? DISPLAY_OPEN : INLINE_OPEN;
const close = isDisplay ? DISPLAY_CLOSE : INLINE_CLOSE;
const closeAt = findClose(text, i + open.length, close);
if (closeAt !== -1) {
const inner = text.slice(i + open.length, closeAt);
const debris = isDisplay ? DISPLAY_DEBRIS_RE : DEBRIS_RE;
out.push(open + inner.replace(debris, (c) => ESCAPE_DEBRIS[c]!) + close);
i = closeAt + close.length;
continue;
}
}
}
out.push(ch);
i += 1;
}
return out.join("");
}
const commandSymbol = /[,;:!|{}%]/;
function dollarRun(text: string, i: number): number {
let count = 0;
while (text[i + count] === "$") count += 1;
return count;
}
/** Native math fences start at a line's block position; one-line $$…$$
* remains inline. A flow opener may have metadata, but never another $. */
function dollarIsDisplay(text: string, i: number, size: number): boolean {
if (size < 2 || !displayIsBlockPositioned(text, i)) return false;
const end = text.indexOf("\n", i + size);
return !text.slice(i + size, end === -1 ? text.length : end).includes("$");
}
/** Match micromark's delimiter runs: exact size inline, at least the
* opener's size on an otherwise-empty line for display. Inside math,
* backslashes are TeX data, not Markdown escapes for closing dollars. */
function findDollarClose(
text: string,
from: number,
size: number,
isDisplay: boolean,
): { at: number; after: number } | undefined {
if (isDisplay) {
let lineEnd = text.indexOf("\n", from);
while (lineEnd !== -1) {
const start = lineEnd + 1;
lineEnd = text.indexOf("\n", start);
const line = text.slice(start, lineEnd === -1 ? text.length : lineEnd);
const close = /^([ \t]{0,3})(\$+)[ \t]*\r?$/.exec(line);
if (close && close[2]!.length >= size) {
const at = start + close[1]!.length;
return { at, after: at + close[2]!.length };
}
/** Decode prefixes and row separators according to ORIGINAL TeX context.
* A native pair is literal; an encoded pair supplies a factor, so an encoded
* four-backslash row becomes TWO, never a command prefix. Unknown commands
* without that structural evidence are left alone. */
function recoverCommandRuns(inner: string, display: boolean): string {
const ranges = texRanges(inner);
const recovered = inner.replace(/\\+/g, (run: string, offset: number) => {
const contexts = ranges.filter(r => offset >= r.start && offset < r.end);
if (contexts.some(r => r.kind === "row" || r.kind === "comment") || run.length % 2 !== 0) return run;
const next = inner[offset + run.length] ?? "";
const encoded = contexts.filter(r => r.kind === "encoded").at(-1);
const word = /^[A-Za-z]+/.exec(inner.slice(offset + run.length))?.[0];
if (encoded) {
const factor = encoded.factor!;
const rows = run.length % (2 * factor) === 0;
const command = word ? knownTexCommand(word) : commandSymbol.test(next);
if (run.length % factor === 0 && (rows || command)) return "\\".repeat(run.length / factor);
return run;
}
return undefined;
}
for (let j = from; j < text.length;) {
if (text[j] !== "$") {
j += 1;
continue;
}
const count = dollarRun(text, j);
if (count === size) return { at: j, after: j + count };
j += count;
}
return undefined;
}
/** Dollar recognition only — the same six/five-character repair maps,
* with no over-escaping normalization or mathematical validation. */
function restoreDollarMathEscapes(text: string): string {
if (!DEBRIS_RE.test(text) || !text.includes("$")) return text;
const out: string[] = [];
let i = 0;
while (i < text.length) {
if (atLineStart(text, i) && FENCE_OPEN_RE.test(text.slice(i))) {
i = copyFencedBlock(text, i, out);
continue;
}
if (text[i] === "`") {
const run = backtickRun(text, i);
let j = i + run;
while (j < text.length) {
const nextRun = backtickRun(text, j);
if (nextRun === run) break;
j += Math.max(nextRun, 1);
}
const after = j < text.length ? j + run : i + run;
out.push(text.slice(i, after));
i = after;
continue;
}
// Legacy spans have already received their own inline/display policy.
// Do not reinterpret dollars inside them in this second pass.
if (text[i] === "\\" && !isEscaped(text, i)) {
const next = text[i + 1];
if (next === "(" || next === "[") {
const closeAt = findClose(text, i + 2, next === "[" ? DISPLAY_CLOSE : INLINE_CLOSE);
if (closeAt !== -1) {
out.push(text.slice(i, closeAt + 2));
i = closeAt + 2;
continue;
}
}
}
if (text[i] === "$" && !isEscaped(text, i)) {
const size = dollarRun(text, i);
const isDisplay = dollarIsDisplay(text, i, size);
const close = findDollarClose(text, i + size, size, isDisplay);
if (close) {
const debris = isDisplay ? DISPLAY_DEBRIS_RE : DEBRIS_RE;
out.push(
text.slice(i, i + size) +
text.slice(i + size, close.at).replace(debris, (c) => ESCAPE_DEBRIS[c]!) +
text.slice(close.at, close.after),
);
i = close.after;
continue;
}
out.push(text.slice(i, i + size));
i += size;
continue;
}
out.push(text[i]!);
i += 1;
}
return out.join("");
}
/**
* Over-escaped command heal (owner ruling 2026-10-15) — the doubled-
* backslash fix.
*
* Some models stochastically emit DOUBLED backslashes inside JSON tool-call
* arguments: the saved record holds `\\nabla` (two real backslashes) where
* `\nabla` was meant. KaTeX reads the `\\` as a row break and then renders
* the command name as plain letters — the learner sees "the word nabla, not
* the symbol". Field evidence (2026-10-15, real data): xiaomi/mimo-v2.6-pro
* doubled per-call, per-field (one probe's question doubled while its
* explanation and options in the SAME call stayed clean — authoring, not
* transport), and deepseek-v4.1-flash once; quadrupling appeared when the
* in-session agent "compensated" for its own doubled first attempt.
*
* THE RULE. Inside a balanced math span ONLY (legacy `\(…\)`/`\[…\]` and
* dollar `$…$`/`$$…$$`, code spans and fences excluded), an EVEN-length
* backslash run (2, 4, …) immediately followed by a command start — a letter
* (any command name; bare letters after a row break are never authored
* content) or a punctuation command (`,` `;` `:` `!` `|` — thin spaces,
* norm bars) — collapses to a single backslash. Even runs before anything
* else (space, newline, `&`, `[`, `$`, digits) are legitimate row breaks and
* stay; ODD runs (3, 5, …) are row break(s) plus a real command with its own
* backslash — e.g. matrix cells written `a \\\nabla` — and stay untouched,
* so matrix content cannot be corrupted. Outside spans nothing is touched:
* doubled backslashes in prose or code are not ours to judge. Unbalanced
* spans pass through, exactly like the debris passes.
*/
const COMMAND_START_CLASS = "[A-Za-z,;:!|]";
/** Collapse even backslash runs before a command start; keep odd runs. */
function halveOverescapedCommandRuns(inner: string): string {
return inner.replace(
new RegExp("(\\\\+)(?=" + COMMAND_START_CLASS + ")", "g"),
(run: string) => (run.length % 2 === 0 ? "\\" : run),
);
if (next === "%") return numericPercentEscape(inner, offset + run.length) ? "\\" : run;
if (word) return word.length > 1 && knownTexCommand(word) ? "\\" : run;
return next && commandSymbol.test(next) ? "\\" : run;
});
// Structural evidence is primary (parse success cannot prove row
// preservation). As an additional guard, never turn parseable source into
// a parse failure through partial command recovery.
if (recovered !== inner && !canParseTex(recovered, display) && canParseTex(inner, display)) return inner;
return recovered;
}
export function healOverescapedCommands(text: string): string {
if (!text.includes("\\\\")) return text;
const out: string[] = [];
let i = 0;
while (i < text.length) {
if (atLineStart(text, i) && FENCE_OPEN_RE.test(text.slice(i))) {
i = copyFencedBlock(text, i, out);
continue;
}
if (text[i] === "`") {
const run = backtickRun(text, i);
let j = i + run;
while (j < text.length) {
const nextRun = backtickRun(text, j);
if (nextRun === run) break;
j += Math.max(nextRun, 1);
}
const after = j < text.length ? j + run : i + run;
out.push(text.slice(i, after));
i = after;
continue;
}
if (text[i] === "\\" && !isEscaped(text, i)) {
const next = text[i + 1];
if (next === "(" || next === "[") {
const isDisplay = next === "[";
const open = isDisplay ? DISPLAY_OPEN : INLINE_OPEN;
const close = isDisplay ? DISPLAY_CLOSE : INLINE_CLOSE;
const closeAt = findClose(text, i + open.length, close);
if (closeAt !== -1) {
const inner = text.slice(i + open.length, closeAt);
out.push(open + halveOverescapedCommandRuns(inner) + close);
i = closeAt + close.length;
continue;
}
}
}
if (text[i] === "$" && !isEscaped(text, i)) {
const size = dollarRun(text, i);
const isDisplay = dollarIsDisplay(text, i, size);
const close = findDollarClose(text, i + size, size, isDisplay);
if (close) {
out.push(
text.slice(i, i + size) +
halveOverescapedCommandRuns(text.slice(i + size, close.at)) +
text.slice(close.at, close.after),
);
i = close.after;
continue;
}
out.push(text.slice(i, i + size));
i += size;
continue;
}
out.push(text[i]!);
i += 1;
}
return out.join("");
return text.includes("\\\\") ? mapMathContent(text, recoverCommandRuns) : text;
}
/** Presentation-only, language-independent recovery. Shape parsing, saved
* identities, grading and model history must use the unprojected source. */
export function repairMathForDisplay(text: string): string {
return healOverescapedCommands(restoreMathEscapes(recoverDocumentLayout(text)));
}
src/core/native-messages.ts
+91 -20
View File
@@ -98,6 +98,55 @@ export function asAgentMessages(messages: readonly LhMessage[]): AgentMessage[]
return messages as unknown as AgentMessage[];
}
// ─── explicit learner-facing output ─────────────────────────────────────────
export const SEND_TO_USER_TOOL = "send_to_user";
/** Read the output tool's message without changing the native call. Also
* accepts partial arguments while streaming; malformed/empty inputs carry
* no prose. This projection is independent of the currently selected model. */
export function sendToUserText(block: unknown): string | undefined {
if (!block || typeof block !== "object") return undefined;
const call = block as { type?: unknown; name?: unknown; arguments?: unknown };
if (call.type !== "toolCall" || call.name !== SEND_TO_USER_TOOL) return undefined;
if (!call.arguments || typeof call.arguments !== "object") return undefined;
const message = (call.arguments as { message?: unknown }).message;
return typeof message === "string" && message.trim() ? message : undefined;
}
/** Learner-facing prose, in content order. Only the explicit output tool
* joins ordinary text here — thinking and all other tool inputs stay out.
* Shared by coverage/extraction and Markdown; native history stays intact. */
export function userFacingTextOf(content: unknown, failedOutputCalls?: ReadonlySet<string>): string[] {
if (typeof content === "string") return content.trim() ? [content] : [];
if (!Array.isArray(content)) return [];
const out: string[] = [];
for (const block of content) {
if (!block || typeof block !== "object") continue;
if (block.type === "toolCall" && failedOutputCalls?.has(block.id)) continue;
const text = block.type === "text" ? block.text : sendToUserText(block);
if (typeof text === "string" && text.trim()) out.push(text);
}
return out;
}
/** Failed output calls remain retry context, not delivered teaching.
* Truncated calls can have valid-looking partial arguments. Failed/aborted
* generations do not execute their calls at all, so may have no result. */
export function failedOutputCallIds(messages: readonly LhMessage[]): ReadonlySet<string> {
const failed = new Set<string>();
for (const message of messages) {
if (message.role === "toolResult" && message.toolName === SEND_TO_USER_TOOL && message.isError) {
failed.add(message.toolCallId);
} else if (message.role === "assistant" && (message.stopReason === "error" || message.stopReason === "aborted")) {
for (const block of message.content) {
if (block.type === "toolCall" && block.name === SEND_TO_USER_TOOL) failed.add(block.id);
}
}
}
return failed;
}
// ─── id minting ──────────────────────────────────────────────────────────────
/** The app-wide message id shape (same as the UIMessage-era mint). */
@@ -502,9 +551,10 @@ function resultIdsFrom(messages: readonly AgentMessage[]): Set<string> {
* - markers flatten to a user message carrying their content (the
* canonical text single-sourced in core/contracts); plan blocks are
* UI-only and filtered.
* - dangling tool calls (an aborted turn's unanswered questions) get the
* synthesized interrupted result — the provider protocol's hard
* invariant: a tool call must be followed by its result.
* - dangling calls from successfully generated assistant messages (for
* example, an unanswered question) get a synthesized interrupted result.
* Error/aborted generations do not: Pi drops those assistant messages,
* so manufacturing their results would leave orphan results on the wire.
*
* Contract (pi-agent-core): never throws — an unknown shape passes
* through untouched rather than killing the loop.
@@ -513,8 +563,9 @@ export function toLlmMessages(messages: readonly AgentMessage[]): Message[] {
// THE COMPACTION CUT (owner round 14, 2026-10-05): the ACTIVE marker
// (the newest — chained compactions rule last) names the first kept
// message; everything before it is replaced by the marker's content
// (header + summary) rendered ONCE, up top. The marker itself is then
// skipped in place — it must never render twice. A cut naming a
// (header + summary) rendered ONCE, up top. All compaction markers
// in the retained tail are then skipped, including older summaries
// whose stored positions still lie after the cut. A cut naming a
// missing id degrades honestly (everything renders — never throw).
const lhMessages = messages as readonly LhMessage[];
const active = activeCompactionOf(lhMessages);
@@ -555,11 +606,10 @@ export function toLlmMessages(messages: readonly AgentMessage[]): Message[] {
// UI-only markers: plan blocks, and problem dismissals (the
// synthesized conclude pair carries the model's face).
if (marker.customType === "plan-block" || marker.customType === "problem-dismissed") continue;
// The ACTIVE compaction marker already rendered as the head —
// never twice. (Stale markers before the cut are unreachable; a
// marker whose cut could not resolve has no head and renders as
// its user text — honest degradation.)
if (marker.customType === "compaction" && head !== null && marker.id === active!.id) continue;
// The head supersedes every compaction marker, not only its own
// stored position. Without a resolved head, preserve the existing
// full-history fallback, including the summary text.
if (marker.customType === "compaction" && head !== null) continue;
out.push({
role: "user",
content: [{ type: "text", text: marker.content }],
@@ -570,11 +620,12 @@ export function toLlmMessages(messages: readonly AgentMessage[]): Message[] {
// Strip the record id: the wire carries the native shape only.
const { id: _id, ...native } = m as LhMessage;
out.push(native as Message);
// Dangling-call synthesis: an assistant message whose toolCalls have
// no results in the remaining tail (abort mid-question) — the
// provider rejects unanswered calls outright.
// Repair unanswered calls from completed generations, not calls in
// failed generations. Pi filters error/aborted assistants itself;
// adding results for their unexecuted calls would leave orphans.
if (role === "assistant") {
const assistant = m as AssistantMessage;
if (assistant.stopReason === "error" || assistant.stopReason === "aborted") continue;
const calls = assistant.content.filter((b) => b.type === "toolCall");
if (calls.length === 0) continue;
const answered = resultIdsFrom(messages.slice(i + 1));
@@ -618,15 +669,35 @@ export function danglingQuestionOf(
return null;
}
/** The probe→teach construction seam (chat-handler): a probe_done call in
* the transcript — any state — means the probe is concluded: teach. */
export function hasProbeDone(messages: readonly LhMessage[]): boolean {
/** Successful probe_done result MESSAGE ids, in transcript order. A call
* from a failed/aborted generation grants no completion authority. Its
* later result must match both call id and tool name, must not be an
* error, and must not explicitly refuse with details.ok:false. Older
* clean acknowledgments without ok:true still count. No history changes. */
export function probeDoneResultIds(messages: readonly LhMessage[]): Set<string> {
const calls = new Set<string>();
const results = new Set<string>();
for (const m of messages) {
if ((m as { role?: string }).role !== "assistant") continue;
const assistant = m as AssistantMessage;
if (assistant.content.some((b) => b.type === "toolCall" && b.name === "probe_done")) return true;
if (m.role === "assistant") {
if (m.stopReason === "error" || m.stopReason === "aborted") continue;
for (const block of m.content) {
if (block.type === "toolCall" && block.name === "probe_done") calls.add(block.id);
}
} else if (
m.role === "toolResult" && m.toolName === "probe_done" &&
calls.has(m.toolCallId) && !m.isError &&
(m.details as { ok?: unknown } | null | undefined)?.ok !== false
) {
results.add(m.id);
}
}
return false;
return results;
}
/** The probe→teach construction seam (chat-handler and the tool guard):
* only a concluded call plus a matching successful result means teach. */
export function hasProbeDone(messages: readonly LhMessage[]): boolean {
return probeDoneResultIds(messages).size > 0;
}
/** Every toolResult id in the record — answer-event validation (the call
src/core/phases.ts
+8 -11
View File
@@ -33,7 +33,7 @@
* Both are strictly sequential; the two derivations agree on every settled
* record (a call's part projects into the same view boundary).
*
* Pure: no imports, no runtime deps — safe on server, client, and tests.
* Pure: only browser-safe core helpers — safe on server, client, and tests.
*/
/** The three phases of a session, as the picker / header / dividers see them. */
@@ -80,7 +80,10 @@ function advance(
// plan → teach only via the first nodeId-tagged quiz CALL (the call,
// not its answer — an errored-and-retried call still proves teaching
// began, and the retry lands in the same group anyway).
if (phase === "probe" && part.type === "tool-probe_done" && part.state === "output-available") {
if (
phase === "probe" && part.type === "tool-probe_done" && part.state === "output-available" &&
(part.output as { ok?: unknown } | null | undefined)?.ok !== false
) {
phase = "plan";
}
// plan → teach via the first nodeId-tagged quizzing CALL — quiz OR
@@ -153,7 +156,7 @@ export function phaseTimeline(messages: readonly PhaseMessage[]): Phase3[] {
// ─── the native fold (Wave D: record-side phase over LhMessage[]) ───────────
import type { LhMessage } from "./native-messages.ts";
import { probeDoneResultIds, type LhMessage } from "./native-messages.ts";
/** The phase at `upToIndex` (inclusive; default: the last message) of a
* NATIVE record. probe→plan advances at the probe_done call's clean
@@ -163,18 +166,12 @@ import type { LhMessage } from "./native-messages.ts";
export function phaseAt3Lh(messages: readonly LhMessage[], upToIndex?: number): Phase3 {
let phase: Phase3 = "probe";
const last = Math.min(upToIndex ?? messages.length - 1, messages.length - 1);
const concluded = probeDoneResultIds(messages.slice(0, last + 1));
for (let i = 0; i <= last; i++) {
const m = messages[i]!;
const role = (m as { role?: string }).role;
if (role === "toolResult") {
const result = m as { toolName?: string; isError?: boolean };
if (
phase === "probe" &&
result.toolName === "probe_done" &&
!result.isError
) {
phase = "plan";
}
if (phase === "probe" && concluded.has(m.id)) phase = "plan";
} else if (role === "assistant") {
const content = (m as { content?: ReadonlyArray<{ type?: string; name?: string; arguments?: unknown }> }).content ?? [];
if (phase === "plan") {
src/core/ports.ts
+2
View File
@@ -136,6 +136,8 @@ export interface CreateJobInit {
toolCallId: string;
/** First activity line (defaults to "starting"). */
activity?: string;
/** Saved on the job before its worker is spawned. */
input?: unknown;
}
/** Mergeable fields — identity (id, kind, chatId, toolCallId, createdAt) is immutable. */
src/core/question-core.ts
+15
View File
@@ -511,6 +511,21 @@ export interface DockSelection {
export const emptyDockSelection = (): DockSelection => ({ selected: new Set(), otherSelected: false, dontKnow: false });
/** In-memory submission snapshot: survives the question form's optimistic
* unmount, never written to storage. Indices use the call's stable shuffle. */
export interface QuestionFormDraft {
selection: DockSelection;
thinking: string;
other: string;
text: string;
confidenceIndex: number;
}
export interface QuestionDraft {
form?: QuestionFormDraft;
readingThinking?: string;
}
/**
* Toggling a listed option. Single-select: the option becomes THE answer and
* displaces "Other". Multi-select: the option toggles alongside its siblings
src/core/session-markdown.ts
+8 -14
View File
@@ -22,22 +22,15 @@
import type { LhMessage } from "./native-messages.ts";
import {
failedOutputCallIds,
PRACTICE_CONCLUDE_TOOL,
PRACTICE_PROBLEM_TOOL,
problemInputForConclude,
SEND_TO_USER_TOOL,
userFacingTextOf,
} from "./native-messages.ts";
import type { ToolResultMessage } from "@earendil-works/pi-ai";
/** The user-facing text blocks of a native message (string or text parts). */
function textOf(content: unknown): string[] {
if (typeof content === "string") return content.trim() ? [content] : [];
if (!Array.isArray(content)) return [];
return content
.filter((c): c is { type: "text"; text: string } => (c as { type?: string }).type === "text")
.map((c) => c.text)
.filter((t) => t.trim().length > 0);
}
/** The practice join context (owner-approved, 2026-10-06 export round):
* practice tool RESULTS are acks — the statement, outcome, and verdict
* all live in the assistant's call arguments, and the model solution
@@ -48,6 +41,7 @@ export interface MarkdownContext {
messages: ReadonlyArray<LhMessage>;
/** toolCallId → the call part's name + arguments. */
calls: Map<string, { name: string; args: Record<string, unknown> | undefined }>;
failedOutputCalls?: ReadonlySet<string>;
}
function markdownContext(messages: ReadonlyArray<LhMessage>): MarkdownContext {
@@ -66,7 +60,7 @@ function markdownContext(messages: ReadonlyArray<LhMessage>): MarkdownContext {
}
}
}
return { messages, calls };
return { messages, calls, failedOutputCalls: failedOutputCallIds(messages) };
}
/** solvedWithHints → "Solved with hints" — the outcome enum reads as prose. */
@@ -172,16 +166,16 @@ export function messageMarkdown(
const parts: string[] = [];
if (role === "user") {
for (const text of textOf((message as { content: unknown }).content)) {
for (const text of userFacingTextOf((message as { content: unknown }).content)) {
parts.push(`### Learner${anchor}\n\n${text.trim()}`);
}
} else if (role === "assistant") {
for (const text of textOf((message as { content: unknown }).content)) {
for (const text of userFacingTextOf((message as { content: unknown }).content, ctx?.failedOutputCalls ?? failedOutputCallIds([message]))) {
parts.push(`### Teacher${anchor}\n\n${text.trim()}`);
}
} else if (role === "toolResult") {
const result = message as ToolResultMessage;
if (result.isError) return null; // retry context, not learning evidence
if (result.isError || result.toolName === SEND_TO_USER_TOOL) return null; // retry context or output ack, not learning evidence
const { heading, body } = toolResultBody(result, ctx);
parts.push(`### ${heading}${anchor}\n\n${body.join("\n\n")}`);
}
src/core/settings-resolution.ts
+5
View File
@@ -35,6 +35,7 @@ import {
type BackgroundModelRole,
type FactorySettings,
type GeneralSettings,
type DictationSettings,
type CompactionSettings,
type GraphSettings,
type ModelSlot,
@@ -69,6 +70,7 @@ export type EffectiveAppearance = FactorySettings["appearance"];
/** The concrete view: every knob resolved, no absent fields. */
export interface EffectiveSettings {
general: Required<GeneralSettings>;
dictation: Required<DictationSettings>;
agent: {
models: Record<BackgroundModelRole, EffectiveModelSpec | undefined>;
/** Delta map as stored — consumers resolve absent key → enabled. */
@@ -176,6 +178,9 @@ export function resolveSettings(record: UserSettings | null): EffectiveSettings
general: {
autoUpdate: record?.general?.autoUpdate ?? FACTORY_SETTINGS.general.autoUpdate,
},
dictation: {
language: record?.dictation?.language ?? FACTORY_SETTINGS.dictation.language,
},
agent: {
models,
tools: agent.tools ?? {},
src/core/settings.ts
+17 -1
View File
@@ -32,6 +32,7 @@ import { AUTH_PROVIDER_KEYS } from "./auth-keys.ts";
import type { ModelRole } from "../../config.ts";
import type { ChatSettings } from "./contracts/chat.ts";
import type { SearchEngineId } from "./search.ts";
import { isDictationLanguage, type DictationLanguage } from "./dictation.ts";
/** The schema version this code understands. Bump ONLY for breaking shape
* changes (a field renamed/meaning changed); pure additions never bump. */
@@ -231,10 +232,15 @@ export const ROLE_SLOT: Readonly<Record<ModelRole, ModelSlotName>> = {
/** General settings — app-level behavior. Version DISPLAY and plans login
* live outside the record (package.json / auth.json) by owner decision. */
export interface GeneralSettings {
/** Auto-update the app (default ON). Mechanism ships in a later wave. */
/** Automatic startup/24h update checks (default ON). Manual checks bypass this. */
autoUpdate?: boolean;
}
/** One app-wide dictation preference, shared by every composer. */
export interface DictationSettings {
language?: DictationLanguage;
}
/** Agent settings — everything about the non-conversation machinery. */
export interface AgentSettings {
/** Background-role model/thinking overrides (absent role = config pin). */
@@ -361,6 +367,7 @@ export interface CompactionSettings {
export interface UserSettings {
version: number;
general?: GeneralSettings;
dictation?: DictationSettings;
agent?: AgentSettings;
/** The strong/cheap/vision model slots (model slots wave, 2026-09-20). */
models?: ModelsSettings;
@@ -378,6 +385,7 @@ export interface UserSettings {
export interface FactorySettings {
version: number;
general: Required<GeneralSettings>;
dictation: Required<DictationSettings>;
appearance: {
mode: "dark" | "light";
background: string | null;
@@ -397,6 +405,7 @@ export interface FactorySettings {
export const FACTORY_SETTINGS: FactorySettings = {
version: CURRENT_SETTINGS_VERSION,
general: { autoUpdate: true },
dictation: { language: "en" },
appearance: {
mode: "dark",
background: null, // the dark default (#141414) lives in the mode's derivation curve
@@ -615,6 +624,13 @@ export function sanitizeUserSettings(raw: unknown): UserSettings | null {
if (r.general !== null && typeof r.general === "object" && !Array.isArray(r.general)) {
out.general = sanitizeGeneral(r.general as Record<string, unknown>);
}
if (r.dictation !== undefined) {
const raw = r.dictation;
out.dictation = raw !== null && typeof raw === "object" && !Array.isArray(raw) &&
isDictationLanguage((raw as Record<string, unknown>).language)
? { language: (raw as DictationSettings).language }
: {};
}
if (r.agent !== null && typeof r.agent === "object" && !Array.isArray(r.agent)) {
out.agent = sanitizeAgent(r.agent as Record<string, unknown>);
}
src/core/tool-parts.ts
+6 -26
View File
@@ -30,7 +30,6 @@
import { Value } from "typebox/value";
import type { TSchema } from "typebox";
import { quizDisplayInputSchema, askInputSchema, askOptionProblems, quizInputProblems, explainInputProblems, explainDisplayInputSchema, type AskToolInput, type QuizToolInput, type ExplainToolInput } from "./question-core.ts";
import { repairMathForDisplay } from "./math-delims.ts";
import { healEncodedArguments } from "./argument-heal.ts";
import { probeDoneInputSchema, type ProbeDoneInput } from "./contracts/probe.ts";
@@ -67,28 +66,9 @@ function boundaryParse<T>(schema: TSchema, input: unknown): T | undefined {
* never have been posed (e.g. `correctAnswer: undefined`, which Convert
* widens to "") renders as a quiet fallback, not a broken quiz.
*/
/** LaTeX-escape restoration, deep over every string field (the JSON-debris
* fix, owner ruling 2026-09-26): records persist RAW call args, so
* single-backslash LaTeX eaten by JSON escape resolution (FF ate the 'f' of
* \forall, VT the 'v' of \varphi, TAB the 't' of \to) survives to the
* display path. This is the ONE seam every rendered question passes
* through — new turns and already-poisoned records alike heal here.
* healOverescapedCommands (owner ruling 2026-10-15) rides the same walk:
* the same models also stochastically DOUBLE command backslashes
* (`\\nabla` saved where `\nabla` was meant — KaTeX shows the row break
* plus "nabla" as letters), so even runs before a command start inside
* balanced math spans collapse to one backslash. */
function restoreMathEscapesDeep(value: unknown): unknown {
if (typeof value === "string") return repairMathForDisplay(value);
if (Array.isArray(value)) return value.map(restoreMathEscapesDeep);
if (typeof value === "object" && value !== null) {
const out: Record<string, unknown> = {};
for (const [k, v] of Object.entries(value)) out[k] = restoreMathEscapesDeep(v);
return out;
}
return value;
}
/** Shape parsing never repairs authored text. The dock builds its answer
* data from this source copy; recovery belongs exclusively to presentation
* (ToolContentResponse), never option identities, grading, or snapshots. */
export function parseQuizInput(input: unknown): QuizToolInput | undefined {
// quizDisplayInputSchema (question-core, owner 2026-10-03): base + the
// OPTIONAL `type` — new teach-phase records carry predict|check and the
@@ -100,7 +80,7 @@ export function parseQuizInput(input: unknown): QuizToolInput | undefined {
// wobble must heal HERE too or a posed quiz renders invisible (the
// tool-parts law). healEncodedArguments is the same pure pass the
// boundary's withArgumentHeal runs.
const healedInput = healEncodedArguments(restoreMathEscapesDeep(input), quizDisplayInputSchema);
const healedInput = healEncodedArguments(input, quizDisplayInputSchema);
const cleaned = boundaryParse<QuizToolInput>(quizDisplayInputSchema, healedInput);
if (!cleaned) return undefined;
if (quizInputProblems(cleaned).length > 0) return undefined;
@@ -114,7 +94,7 @@ export function parseQuizInput(input: unknown): QuizToolInput | undefined {
* keying, so they parse to undefined exactly like the zod seam did).
*/
export function parseAskInput(input: unknown): AskToolInput | undefined {
const cleaned = boundaryParse<AskToolInput>(askInputSchema, restoreMathEscapesDeep(input));
const cleaned = boundaryParse<AskToolInput>(askInputSchema, input);
if (!cleaned) return undefined;
if (askOptionProblems(cleaned.options ?? []).length > 0) return undefined;
return cleaned;
@@ -127,7 +107,7 @@ export function parseAskInput(input: unknown): AskToolInput | undefined {
* variants share the same fields; the teacher's nodeId is an unknown
* key Value.Clean strips, same as the quiz seam. */
export function parseExplainInput(input: unknown): ExplainToolInput | undefined {
const cleaned = boundaryParse<ExplainToolInput>(explainDisplayInputSchema, restoreMathEscapesDeep(input));
const cleaned = boundaryParse<ExplainToolInput>(explainDisplayInputSchema, input);
if (!cleaned) return undefined;
if (explainInputProblems(cleaned).length > 0) return undefined;
return cleaned;
src/core/view.ts
+27 -1
View File
@@ -34,7 +34,7 @@
import type { AssistantMessage, ToolResultMessage } from "@earendil-works/pi-ai";
import type { LhMarkerMessage, LhMessage } from "./native-messages.ts";
import { modelSolutionForConclude, PRACTICE_CONCLUDE_TOOL, problemInputForConclude } from "./native-messages.ts";
import { modelSolutionForConclude, PRACTICE_CONCLUDE_TOOL, problemInputForConclude, SEND_TO_USER_TOOL, sendToUserText } from "./native-messages.ts";
import { activeCompactionOf, type CompactionDetails } from "./compaction.ts";
import { QUESTION_TOOLS } from "./pending.ts";
import type { SubagentArtifactData, SubagentResultMeta } from "./contracts/jobs.ts";
@@ -124,6 +124,14 @@ function lastSuccessfulPlanInit(parts: LhViewPart[]): number {
* view messages. Pure: the same input always projects to the same list. */
export function projectView(messages: readonly LhMessage[], live?: LiveTurnState): LhViewMessage[] {
const out: LhViewMessage[] = [];
// Output calls become text rather than tool parts, so read settled
// failures up front: a failed call must retain the normal error path.
const outputErrors = new Map<string, ToolResultMessage>();
for (const message of messages) {
if (message.role === "toolResult" && message.toolName === SEND_TO_USER_TOOL && message.isError) {
outputErrors.set(message.toolCallId, message);
}
}
const lastAssistant = (): LhViewMessage | undefined => {
const last = out[out.length - 1];
@@ -218,6 +226,24 @@ export function projectView(messages: readonly LhMessage[], live?: LiveTurnState
reasoning.push({ type: "reasoning", id: `${m.id}-r${reasoningN++}`, text: block.thinking, state: "done" });
} else if (block.type === "toolCall") {
const part = toolPartOf(block, live, streamingMessage);
if (block.name === SEND_TO_USER_TOOL) {
// Failed/aborted generations never execute their calls;
// the provider error/abort UI already owns that outcome.
if (assistant.stopReason === "error" || assistant.stopReason === "aborted") continue;
const error = outputErrors.get(block.id);
if (error) applyResult(part, error);
if (part.state === "output-error") {
rest.push(part);
} else {
const text = sendToUserText(block);
if (text !== undefined) {
rest.push({ type: "text", id: `${m.id}-t${textN++}`, text, state: "done" });
}
}
// No live tool card or acknowledgement: this channel is
// ordinary prose, including while its arguments stream.
continue;
}
// The conclusion's reveal join (owner round, 2026-10-05): the
// hero of a conclude card is the modelSolution committed at POSE
// time — a record fold, enriched here so the card renders pure.
src/server/chat-handler.ts
+84 -22
View File
@@ -17,7 +17,7 @@ import { createExtractionPipeline } from "../agent/subagents/extraction.ts";
import type { ModelRole } from "../../config.ts";
import {
deliverJobResult,
noteSteerVisible,
noteSteersDurable,
reconcileUserSteers,
registerLiveAgent,
restoreUncommittedAutomatic,
@@ -42,8 +42,8 @@ import {
import { modelsNotConfigured, providerKeyMissing as providerKeyMissingEnvelope, type ProviderKeyMissing } from "../core/settings.ts";
import { beginTurn, endTurn } from "./turn-liveness.ts";
import { applyParkedDismissal } from "./dismissal-park.ts";
import { registerTurn, unregisterTurn } from "./turn-runs.ts";
import { persistTurn, triggerExtraction } from "./turn-service.ts";
import { registerTurn, rememberTurnFailure, unregisterTurn } from "./turn-runs.ts";
import { persistTurn, triggerExtraction, TurnPersistenceError } from "./turn-service.ts";
import { hasProbeDone, danglingQuestionOf } from "../core/native-messages.ts";
import { composeInstructions } from "../agent/teacher-agent.ts";
import { ensureCompacted, kickCompaction, effectiveThresholdTokens } from "./compaction.ts";
@@ -55,10 +55,11 @@ import {
ChatInputError,
inputAppends,
parseChatInput,
retryDrops,
type TurnInput,
} from "../core/input-events.ts";
import { steerDelivered } from "../core/contracts.ts";
import type {CallUsage, SteerMessage} from "../core/contracts.ts";
import type {CallUsage} from "../core/contracts.ts";
/**
* One chat turn over the server-owned native record. Framework-free (plain
@@ -79,6 +80,10 @@ import type {CallUsage, SteerMessage} from "../core/contracts.ts";
* for write-ahead persistence, turn liveness, and the turn services.
*/
function settlementErrorText(error: unknown): string {
return error instanceof TurnPersistenceError ? error.message : "An error occurred.";
}
const jobRunner = createJobRunner({
chatStore: stores.chatStore,
jobStore: stores.jobStore,
@@ -182,6 +187,17 @@ export async function handleChatInput(
return Response.json({ error: "Unknown chat" }, { status: 404 });
}
// Refuse an ineligible Retry before catalog refresh, compaction, or any
// turn/write effects. Revalidate after the awaited gates below too.
if (input.type === "retry") {
try {
retryDrops(record.messages);
} catch (err) {
if (err instanceof ChatInputError) return Response.json({ error: err.message }, { status: 400 });
throw err;
}
}
// THE UNIVERSAL TURN GATE (BYOK, owner, 2026-09-22): a turn needs the
// CONVERSATION chain — the chat's own model override over the strong
// slot — and there is no factory floor; when both are empty the turn
@@ -255,7 +271,20 @@ export async function handleChatInput(
chatId,
);
const compactedRecord = stores.chatStore.getChat(chatId);
if (compactedRecord) record = compactedRecord;
if (!compactedRecord) return Response.json({ error: "Unknown chat" }, { status: 404 });
record = compactedRecord;
// Recompute Retry eligibility on current truth, still before beginTurn.
// A compaction marker or another writer may have landed during awaits.
// These omissions affect ONLY the agent context, never the saved record.
let retryOmissions = new Set<string>();
if (input.type === "retry") {
try {
retryOmissions = retryDrops(record.messages);
} catch (err) {
if (err instanceof ChatInputError) return Response.json({ error: err.message }, { status: 400 });
throw err;
}
}
// The input's native appends (pure; the honest 400 for an answer with no
// pending question call). Computed BEFORE beginTurn so a refused input
@@ -289,7 +318,11 @@ export async function handleChatInput(
// the composer gates), so adjacency holds by construction.
if (input.type === "answer" && danglingQuestionOf([...record.messages, ...appends]) !== null) {
const merged = [...record.messages, ...appends];
stores.chatStore.saveMessages(chatId, merged);
try { stores.chatStore.saveMessages(chatId, merged); }
catch (err) {
console.error(`[chat ${chatId}] write-ahead persist failed:`, err);
return new Response("Your input could not be saved. Please try again.", { status: 500 });
}
emit("record-changed", { chatId });
return settledTurnResponse(); // the cheap terminal — no beginTurn, no turn
}
@@ -309,18 +342,26 @@ export async function handleChatInput(
// ride this same write-ahead - the pump de-queued them at fire, so this
// is the ONLY copy ever appended.)
const merged = [...record.messages, ...appends];
// Retry preserves every persisted attempt. Pi needs a non-assistant
// tail to start continue(), so its temporary history omits the trailing
// failures only. Count that SAME temporary prefix at settlement, or the
// retained failures would make persistTurn skip genuinely new output.
const agentMessages = input.type === "retry" ? merged.filter((m) => !retryOmissions.has(m.id)) : merged;
const stagedAutomatic = isSubscriptionProvider(conversation.provider) && input.type === "steers";
let persistBaseCount = stagedAutomatic ? record.messages.length : merged.length;
let persistBaseCount = stagedAutomatic ? record.messages.length : agentMessages.length;
try {
if (!stagedAutomatic) stores.chatStore.saveMessages(chatId, merged);
// Dock fix (owner ruling 2026-09-26): persisted markers are visible
// in the record — flip their jobs now (the wake's steers ride this
// write-ahead; the delivering turn may run for minutes).
for (const appended of stagedAutomatic ? [] : appends) {
if ((appended as { role?: string }).role === "custom") noteSteerVisible(appended as SteerMessage);
}
} catch (err) {
console.error(`[chat ${chatId}] write-ahead persist failed:`, err);
endTurn(chatId);
return new Response("Your input could not be saved. Please try again.", { status: 500 });
}
// Visibility bookkeeping is not the canonical save. Its failure must
// neither misreport a saved input as lost nor prevent the turn running.
try {
if (!stagedAutomatic) noteSteersDurable(chatId, merged);
} catch (err) {
console.error(`[chat ${chatId}] input visibility bookkeeping failed:`, err);
}
// MONEY CHANNEL (wave 4): one collector per request, fed by the agent's
@@ -346,7 +387,7 @@ export async function handleChatInput(
// the session is still probing: build the PROBER (quiz/ask + probe_done).
// Both share this whole composition (write-ahead persist, liveness,
// steers, usage, extraction) — only the agent differs.
const teaching = hasProbeDone(merged);
const teaching = hasProbeDone(agentMessages);
// BYOK (owner, 2026-09-22): the conversation chain was resolved by the
// turn gate above (chat.model override → strong slot → refused) — the
// engine build is pure from here; thinking levels clamp natively.
@@ -365,19 +406,23 @@ export async function handleChatInput(
// sub-agent JOB tools (research/demo/visual/pdf; adjudicated with the
// user: quiz/ask/fs/plan/probe_done are always-on).
if (isSubscriptionProvider(conversation.provider)) requestCycle = new SubscriptionRequestCycle({
initialMessages: asAgentMessages(merged),
initialMessages: asAgentMessages(agentMessages),
staged: stagedAutomatic ? asAgentMessages(appends) : undefined,
foregroundOwed: !callPurpose.background,
purpose: callPurpose,
scopeCurrent: (scope) => connectionScopeCurrent(conversation.provider, scope),
commitStaged: () => { stores.chatStore.saveMessages(chatId, merged); persistBaseCount = merged.length; },
commitStaged: () => {
stores.chatStore.saveMessages(chatId, merged);
persistBaseCount = merged.length;
noteSteersDurable(chatId, merged);
},
restoreAutomatic: (automaticIds, retainedIds) => restoreUncommittedAutomatic(chatId, constructed, automaticIds, retainedIds),
});
const agentOpts: AgentBuildOptions = {
engine,
callPurpose,
requestCycle,
messages: asAgentMessages(merged),
messages: asAgentMessages(agentMessages),
onUsage: collectUsage,
toolToggles: effective.agent.tools,
// Debug (owner, 2026-10-04): teacher/prober raw provider events under
@@ -468,11 +513,12 @@ export async function handleChatInput(
// run's first LLM call (the native replacement for continuation
// coalescing). Settlement still cannot beat the flow below: it needs
// a streamed response, which never happens inside the constructor.
registerLiveAgent(chatId, constructed);
registerLiveAgent(chatId, constructed, input.type !== "steers" && input.type !== "continue",
input.type === "steers" ? new Set(appends.map((message) => message.id)) : undefined);
// Debug recorder run marker (owner, 2026-10-04): correlates the
// JSONL event stream with the record's turns; no-op unless
// LH_PROVIDER_EVENTS=1.
providerEventNote(chatId, "turn", `run start (base: ${merged.length} messages)`);
providerEventNote(chatId, "turn", `run start (base: ${agentMessages.length} messages)`);
const handle = runPiTurn({
agent: constructed,
requestCycle,
@@ -496,10 +542,26 @@ export async function handleChatInput(
console.error(`[chat ${chatId}] turn error:`, error);
return "An error occurred.";
},
onSettleError: settlementErrorText,
onSettleFailure: (error, reportedErrorText) => {
// POST only acknowledges acceptance; the submitter may attach
// after this turn has already failed and left the registry.
rememberTurnFailure(chatId, reportedErrorText ?? settlementErrorText(error));
endTurn(chatId);
unregisterTurn(chatId);
unregisterLiveAgent(chatId);
// The canonical save may have succeeded before later bookkeeping
// threw. Reconcile against actual durability, never replay input
// merely because usage/extraction/other settlement work failed.
const persisted = stores.chatStore.getChat(chatId)?.messages ?? [];
reconcileUserSteers(chatId, persisted);
settleSteers(chatId, persisted, true, constructed);
},
onFinish: (turnMessages, isAborted) => {
// Wave B: the record write is the agent's own final state —
// the new messages are the slice past the merged base.
// Wave B: append only messages past the temporary agent prefix.
// Retry's saved record still includes its earlier failed attempts.
const stamped = persistTurn(stores.chatStore, chatId, persistBaseCount, turnMessages, isAborted);
noteSteersDurable(chatId, stores.chatStore.getChat(chatId)?.messages ?? stamped);
// usage log append (idempotent per turn id) — the stats bar
// refolds this on the client after the turn settles. The
// provenance message is the turn's FIRST new assistant message.
@@ -545,7 +607,7 @@ export async function handleChatInput(
// resolve; stranded markers return to pending), then let
// the pump take any stragglers down the idle wake.
unregisterLiveAgent(chatId);
settleSteers(chatId, stamped);
settleSteers(chatId, stamped, false, constructed);
// Composer steers (2026-10-03): delivered ids dropped, stranded
// ones stay parked for the chat's next run (the registry is the
// stranding safety net — the agent's undrained copy died above).

Some files were not shown because too many files have changed in this diff Show More