commit a906f1341598f10d3ce8915f7e36660b33013b00
parent 7790cfcdd0bf25535cca7850dce56c3531a79399
Author: I Mean I'm Just Saying <imeanimjustsaying@kiwifarms.st>
Date: Thu, 16 Jul 2026 21:26:56 -0400
Ask chat: opt-in Report mode (model-maintained living document)
An opt-in mode (native tool-capable providers only) where the model maintains
a persistent markdown report via an update_report tool while the user keeps
chatting, and the conversation is compacted into that report instead of
replaying every prior turn.
- update_report tool (section upsert) in shared.ts + all 3 native providers
via buildTools(includeFetch, includeReport); parsers extract section/content
conditionally (existing deep-equal assertions hold); independent per-turn
counter, off the search budget.
- applyReportPatch (pure, tested): case-insensitive ##-level section upsert,
preserves preamble/other sections, collapses blank runs.
- searchAgent: report/reportMode options, report in the result, report_update
event, runUpdateReport executor; COMPACTION — when report mode is on the
replayed history becomes a single "Report so far" message (the excerpt pool
still flows into grounding), rebuilt post-gather for the answer phase.
- useAskChat: report/reportMode/reportUpdating state, persisted + restored,
gated to tool-capable transports; ReportPanel (collapsible, Markdown, toggle
disabled on scripted with a hint, updating shimmer).
Off by default → all existing paths byte-for-byte unchanged.
Verified: tsc clean (export + common); +7 unit (applyReportPatch ×4,
update_report parse ×3) = 30 green; full e2e 108/108 (native report-mode +
compaction assertion + scripted-disabled cases); no new lint errors.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
Diffstat:
13 files changed, 704 insertions(+), 27 deletions(-)
diff --git a/export/CHANGELOG.md b/export/CHANGELOG.md
@@ -1,6 +1,7 @@
# Changelog
## [Unreleased]
+- **"Ask AI" chat — an opt-in Report mode for long research sessions.** Turn on **Report mode** (in the new **Report** panel; tool-capable providers only) and the assistant maintains a persistent Markdown **report** document — a running canvas it updates via an `update_report` tool as you keep chatting, upserting well-titled sections that persist across turns. Crucially, while it's on the conversation is **compacted into the report** instead of replaying every prior question and answer: each turn sends a single *"report so far"* summary plus the usual deduplicated excerpt pool, so a long back-and-forth stays bounded in tokens instead of growing every turn. The report renders live in the panel (with an *updating…* shimmer while a turn writes to it) and is saved with the conversation, so it survives reloads; *New chat* clears it. On the Scripted transport there's no tool to drive it, so the toggle is disabled with a hint. See `export/app/lib/nativeTools/{shared,anthropic,openai,gemini}.ts` (the `update_report` tool), `export/app/lib/askConversation.ts` (`applyReportPatch`), `export/app/lib/searchAgent.ts` (compaction + executor), `export/app/ask/{useAskChat.ts,ReportPanel.tsx,AskChat.tsx}`, and `export/e2e/ask-chat.spec.ts`.
- **"Ask AI" now grounds in your *current* search automatically.** With search and chat sharing one workspace, you no longer click "Ask AI about these results" to hand a frozen copy of your results to the chat — the chat reads the **live** search directly. Run a search, switch to Chat, and it's already grounded in exactly those results (with the same *answer only from these / may also search* toggle); change the search and the grounding follows. No active search → the chat searches on its own, as before. A **Detach** control lets you ask a free-form question without the current search grounding it (and a **Ground in my search** button re-attaches). The old "Ask AI about these results" button and its one-shot hand-off are retired. See `common/components/SearchSessionContext.tsx` (`liveGrounding`), `export/app/ask/useAskChat.ts`, `export/app/ask/{AskChat,PinnedResultsPanel}.tsx`, and `common/components/SearchResults.tsx`.
- **Search and "Ask AI" are now one workspace — the search bar stays put when you switch between them.** Previously `/` (search) and `/ask` (chat) were separate pages, and navigating from one to the other threw away your search. They now share a single shell: the search bar and its results live in a persistent layout, with a **Results ⇄ Chat** switch between the two views. Run a search, flip to Chat to ask about it, flip back — your search is exactly where you left it. Under the hood the search state was lifted out of the monolithic search component into a shared `SearchSession` (both views read the same committed search), so it's the one source of truth features build on. Hub mode is unchanged. See `common/components/{SearchSessionContext,WorkspaceSearchBar,SearchResults,TranscriptSearch}.tsx`, `export/app/(workspace)/*`, and `export/e2e/workspace-shell.spec.ts`.
- **Handing a search to "Ask AI" now passes the *whole* result set, not just the first 20.** Previously only the top 20 videos reached the chat, so it literally couldn't see the rest. Now every match is handed off (up to a generous cap) and presented in tiers: the model gets a compact **index of all matching videos** (title, channel, date, hit count) plus **full excerpts for the top ~12** — and it can pull excerpts for any other result on demand via the existing *fetch_context* tool, or you can click **Load context** on any of them. This keeps the payload bounded while letting the assistant reason over the complete set. See `common/lib/aiHandoff.ts` (tiered `buildSearchHandoff`) and `export/app/lib/askConversation.ts` (`buildTieredGrounding`).
diff --git a/export/app/ask/AskChat.tsx b/export/app/ask/AskChat.tsx
@@ -5,6 +5,7 @@ import { ArrowDownIcon, PlusIcon } from "lucide-react";
import { useAskChat } from "./useAskChat";
import { ProviderSettings } from "./ProviderSettings";
import { ContextPanel } from "./ContextPanel";
+import { ReportPanel } from "./ReportPanel";
import { PinnedResultsPanel } from "./PinnedResultsPanel";
import { MessageBubble } from "./MessageBubble";
import { Composer } from "./Composer";
@@ -160,6 +161,15 @@ export default function AskChat() {
/>
)}
+ <ReportPanel
+ report={s.report}
+ reportMode={s.reportMode}
+ setReportMode={s.setReportMode}
+ available={s.reportModeAvailable}
+ updating={s.reportUpdating}
+ busy={busy}
+ />
+
<div className="relative">
<div
ref={scrollRef}
diff --git a/export/app/ask/ReportPanel.tsx b/export/app/ask/ReportPanel.tsx
@@ -0,0 +1,95 @@
+"use client";
+
+import { useState } from "react";
+import { FileTextIcon, Loader2Icon } from "lucide-react";
+import { Markdown } from "yt-dlp-transcript-common/components/Markdown";
+
+// The running "report" surface for report mode: a persistent markdown document
+// the model maintains via the update_report tool while the user keeps chatting.
+// The header carries the on/off toggle (disabled on transports without native
+// tools, since there is no tool to write the report), a live "updating…" shimmer
+// while a turn is writing, and the rendered report (or an empty-state hint).
+// Collapsible like the Context panel so it stays out of the way until wanted.
+export function ReportPanel({
+ report,
+ reportMode,
+ setReportMode,
+ available,
+ updating,
+ busy,
+}: {
+ report: string;
+ reportMode: boolean;
+ setReportMode: (on: boolean) => void;
+ available: boolean;
+ updating: boolean;
+ busy: boolean;
+}) {
+ const [open, setOpen] = useState(false);
+ const hasReport = report.trim() !== "";
+
+ return (
+ <details
+ className="rounded-lg border border-border bg-card/40"
+ open={open}
+ onToggle={(e) => setOpen((e.currentTarget as HTMLDetailsElement).open)}
+ >
+ <summary className="flex cursor-pointer items-center gap-2 px-4 py-2.5 text-sm font-medium text-foreground">
+ <FileTextIcon className="size-4 shrink-0 text-brand" />
+ Report
+ {updating && (
+ <span className="inline-flex items-center gap-1 text-xs font-normal text-brand">
+ <Loader2Icon className="size-3 animate-spin motion-reduce:animate-none" />
+ updating…
+ </span>
+ )}
+ {/* The toggle lives in the header so it's reachable without expanding. */}
+ <label
+ className="ml-auto flex items-center gap-1.5 text-xs font-normal text-muted-foreground"
+ title={
+ available
+ ? "Maintain a persistent report; the conversation is compacted into it."
+ : "Report mode needs a tool-capable provider (switch off Scripted)."
+ }
+ onClick={(e) => e.stopPropagation()}
+ >
+ <input
+ type="checkbox"
+ checked={reportMode && available}
+ disabled={!available || busy}
+ onChange={(e) => setReportMode(e.target.checked)}
+ className="accent-brand"
+ />
+ Report mode
+ </label>
+ </summary>
+
+ {open && (
+ <div className="flex flex-col gap-3 border-t border-border px-4 py-3">
+ {!available && (
+ <p className="text-xs text-warning">
+ Report mode needs a tool-capable provider. Switch the search mode off
+ Scripted (Auto or Native tools) to enable it.
+ </p>
+ )}
+ {hasReport ? (
+ <div
+ aria-live="polite"
+ className={
+ "rounded-md border border-border bg-background px-3 py-2 text-sm text-foreground transition-opacity " +
+ (updating ? "opacity-60" : "opacity-100")
+ }
+ >
+ <Markdown className="leading-relaxed">{report}</Markdown>
+ </div>
+ ) : (
+ <p className="text-xs text-muted-foreground">
+ Turn on report mode and ask a question — the assistant will build a
+ report here, keeping it current as the conversation grows.
+ </p>
+ )}
+ </div>
+ )}
+ </details>
+ );
+}
diff --git a/export/app/ask/useAskChat.ts b/export/app/ask/useAskChat.ts
@@ -11,7 +11,12 @@ import {
windowCues,
} from "yt-dlp-transcript-common/lib/transcriptWindow";
import { PROVIDERS, type Provider } from "../lib/askProvider";
-import { runAskTurn, type AgentEvent, type AgentMode } from "../lib/searchAgent";
+import {
+ runAskTurn,
+ supportsNativeTools,
+ type AgentEvent,
+ type AgentMode,
+} from "../lib/searchAgent";
import type { RetrievedVideo } from "../lib/askRetrieval";
import {
buildApiMessages,
@@ -43,6 +48,10 @@ type PersistedConvo = {
// Per-video extra transcript pulled in via "Load context", overlaid onto the
// live search grounding (keyed by video slug).
enrichments?: Record<string, Snip[]>;
+ // Report mode: the running report document + whether the mode is on. Persisted
+ // so a long research session's report survives reloads.
+ report?: string;
+ reportMode?: boolean;
};
const NON_TERMINAL_PHASES = new Set(["gathering", "answering", "streaming"]);
@@ -109,6 +118,15 @@ export function useAskChat() {
// as a starting point the AI may expand with its own searches.
const [strictGrounding, setStrictGroundingState] = useState(true);
+ // Report mode: an opt-in mode (native tool-capable providers only) where the
+ // model maintains a persistent markdown report via update_report, and the
+ // conversation is compacted into that report instead of replaying every turn.
+ const [report, setReport] = useState("");
+ const [reportMode, setReportModeState] = useState(false);
+ // True while a turn is writing to the report (drives the panel shimmer);
+ // cleared when the turn ends.
+ const [reportUpdating, setReportUpdating] = useState(false);
+
// The effective pinned grounding = the live search, overlaid with any per-video
// "Load context" enrichments, unless the user detached.
const pinned = useMemo<SearchHandoff | null>(() => {
@@ -153,6 +171,10 @@ export function useAskChat() {
pinnedRef.current = pinned;
const strictRef = useRef(strictGrounding);
strictRef.current = strictGrounding;
+ const reportRef = useRef(report);
+ reportRef.current = report;
+ const reportModeRef = useRef(reportMode);
+ reportModeRef.current = reportMode;
// Restore saved preferences + (if remembered) the provider's key/model, plus a
// persisted conversation.
@@ -189,6 +211,10 @@ export function useAskChat() {
if (parsed.enrichments && typeof parsed.enrichments === "object") {
setEnrichments(parsed.enrichments);
}
+ if (typeof parsed.report === "string") setReport(parsed.report);
+ if (typeof parsed.reportMode === "boolean") {
+ setReportModeState(parsed.reportMode);
+ }
}
} catch {
/* storage unavailable / corrupt */
@@ -208,7 +234,9 @@ export function useAskChat() {
messages.length === 0 &&
!contextOverride &&
!detached &&
- !hasEnrichments
+ !hasEnrichments &&
+ !report &&
+ !reportMode
) {
localStorage.removeItem(K_CONVO);
setStorageWarning(false);
@@ -226,6 +254,8 @@ export function useAskChat() {
strictGrounding,
detached,
enrichments,
+ report,
+ reportMode,
})
) {
setStorageWarning(msgs.length < messages.length);
@@ -249,7 +279,15 @@ export function useAskChat() {
return () => {
if (persistTimer.current) clearTimeout(persistTimer.current);
};
- }, [messages, contextOverride, strictGrounding, detached, enrichments]);
+ }, [
+ messages,
+ contextOverride,
+ strictGrounding,
+ detached,
+ enrichments,
+ report,
+ reportMode,
+ ]);
function loadProviderCreds(p: Provider, rememberOn: boolean) {
// Prefer a model the user typed for this provider earlier this session, so a
@@ -311,6 +349,17 @@ export function useAskChat() {
}
}
+ // Report mode is only meaningful on a tool-capable transport: Scripted has no
+ // tool to maintain the report, so the toggle is disabled there. (Persistence is
+ // handled by the conversation-persist effect, which lists reportMode as a dep.)
+ const reportModeAvailable =
+ searchMode !== "scripted" &&
+ (searchMode === "native" || supportsNativeTools(provider, model));
+
+ function setReportMode(on: boolean) {
+ setReportModeState(on);
+ }
+
// Patch one message by index, immutably.
const patchAt = (index: number, fn: (m: UiMessage) => UiMessage) =>
setMessages((prev) => {
@@ -399,6 +448,12 @@ export function useAskChat() {
sources: e.videos,
}));
break;
+ case "report_update":
+ // The model wrote a section — flag the live "updating report" state
+ // (cleared when the turn ends). The updated report text arrives with
+ // the turn's result below.
+ setReportUpdating(true);
+ break;
case "delta":
patchAt(assistantIndex, (m) => ({
...m,
@@ -409,6 +464,11 @@ export function useAskChat() {
}
};
+ // Report mode only takes effect on a tool-capable transport; on scripted it
+ // must stay off (there's no update_report tool, and dropping the replayed
+ // history would lose context with nothing to compact into).
+ const effectiveReportMode = reportModeRef.current && reportModeAvailable;
+
try {
const result = await runAskTurn({
provider,
@@ -423,6 +483,8 @@ export function useAskChat() {
onEvent,
historyOverride,
seedVideos: args.seedVideos,
+ report: reportRef.current,
+ reportMode: effectiveReportMode,
precomputedGrounding: args.precomputedGrounding?.groundedContent
? {
videos: args.precomputedGrounding.videos ?? [],
@@ -431,6 +493,10 @@ export function useAskChat() {
}
: undefined,
});
+ // Persist the (possibly updated) report from this turn.
+ if (effectiveReportMode && result.report !== reportRef.current) {
+ setReport(result.report);
+ }
patchAt(assistantIndex, (m) => ({
...m,
content: m.content || result.answer,
@@ -463,11 +529,12 @@ export function useAskChat() {
}
} finally {
setBusy(false);
+ setReportUpdating(false);
sendingRef.current = false;
abortRef.current = null;
}
},
- [provider, apiKey, model, searchMode, summaries, aliases],
+ [provider, apiKey, model, searchMode, summaries, aliases, reportModeAvailable],
);
const send = useCallback(async () => {
@@ -581,6 +648,8 @@ export function useAskChat() {
setDetached(false);
setEnrichments({});
setStrictGroundingState(true);
+ setReport("");
+ setReportModeState(false);
};
// Detach from the live search → ask a free-form question (the AI searches on
@@ -690,6 +759,12 @@ export function useAskChat() {
setSearchMode,
markdownOn,
setMarkdownOn,
+ // report mode
+ report,
+ reportMode,
+ setReportMode,
+ reportModeAvailable,
+ reportUpdating,
// data
summariesReady,
corpusError,
diff --git a/export/app/lib/askConversation.test.ts b/export/app/lib/askConversation.test.ts
@@ -2,6 +2,7 @@ import { test } from "node:test";
import assert from "node:assert/strict";
import type { SearchAlias } from "yt-dlp-transcript-common/lib/searchAliases";
import {
+ applyReportPatch,
buildApiMessages,
buildGroundedContent,
collectPriorPool,
@@ -164,6 +165,51 @@ test("parseContext survives pruning and falls back to a single user block", () =
assert.deepEqual(parseContext(" "), []);
});
+test("applyReportPatch appends a new section to the end, preserving others", () => {
+ const before = "## Findings\n\nAlpha is discussed.";
+ const after = applyReportPatch(before, "Open Questions", "Who said beta?");
+ // The existing section is preserved…
+ assert.match(after, /## Findings\n\nAlpha is discussed\./);
+ // …and the new one is appended with a `## ` heading and its content.
+ assert.match(after, /## Open Questions\n\nWho said beta\?/);
+ // New section comes AFTER the existing one.
+ assert.ok(after.indexOf("## Findings") < after.indexOf("## Open Questions"));
+});
+
+test("applyReportPatch replaces an existing section's body, leaving neighbours", () => {
+ const before =
+ "## Findings\n\nOld body.\n\n## Timeline\n\n- 2020: a thing";
+ const after = applyReportPatch(before, "Findings", "New body with more detail.");
+ assert.match(after, /## Findings\n\nNew body with more detail\./);
+ // The old body is gone…
+ assert.doesNotMatch(after, /Old body/);
+ // …and the untouched neighbouring section survives intact.
+ assert.match(after, /## Timeline\n\n- 2020: a thing/);
+});
+
+test("applyReportPatch preserves a preamble and appends into empty reports", () => {
+ // First write into an empty report is just the section block.
+ const first = applyReportPatch("", "Findings", "Alpha.");
+ assert.equal(first, "## Findings\n\nAlpha.");
+ // A preamble ahead of any heading is kept when a new section is appended.
+ const withPreamble = applyReportPatch(
+ "Research notes on the archive.\n\n## Findings\n\nAlpha.",
+ "Sources",
+ "[1] video",
+ );
+ assert.match(withPreamble, /^Research notes on the archive\./);
+ assert.match(withPreamble, /## Sources\n\n\[1\] video/);
+});
+
+test("applyReportPatch matches the heading case-insensitively", () => {
+ const before = "## Findings\n\nOld.";
+ const after = applyReportPatch(before, "findings", "Updated.");
+ // Matched despite the case difference → replaced, not appended (one heading).
+ assert.equal((after.match(/## Findings/g) ?? []).length, 1);
+ assert.match(after, /## Findings\n\nUpdated\./);
+ assert.doesNotMatch(after, /Old\./);
+});
+
test("answerSystemPrompt asks for Markdown + citations and carries the glossary", () => {
const p = answerSystemPrompt(ALIASES);
assert.match(p, /Markdown/);
diff --git a/export/app/lib/askConversation.ts b/export/app/lib/askConversation.ts
@@ -90,6 +90,53 @@ export function collectPriorPool(
return order.map((k) => byKey.get(k)!);
}
+// ─── report mode (persistent markdown document, maintained via update_report) ───
+
+// Upsert a `## <section>` block into the running report. If a section with that
+// heading already exists (matched case-insensitively at the `##` level), its body
+// is replaced with `content`; otherwise a new `## <section>` block is appended to
+// the end. Any preamble and the other sections are preserved, and runs of 3+
+// newlines are collapsed to a single blank line so repeated edits stay tidy.
+// Pure + testable — the executor in searchAgent applies it as the model calls the
+// tool, threading the result through the turn.
+export function applyReportPatch(
+ report: string,
+ section: string,
+ content: string,
+): string {
+ const heading = /^##\s+(.+?)\s*$/;
+ const target = section.trim().toLowerCase();
+ const body = content.trim();
+ const lines = report.split("\n");
+
+ // Index every `##`-level section heading (exactly two hashes: `### …` and
+ // deeper don't match, so sub-headings inside a section body are left intact).
+ const heads: { line: number; title: string }[] = [];
+ lines.forEach((l, i) => {
+ const m = heading.exec(l);
+ if (m) heads.push({ line: i, title: m[1] });
+ });
+
+ const hit = heads.findIndex((h) => h.title.trim().toLowerCase() === target);
+
+ let out: string;
+ if (hit === -1) {
+ // Append a new section, keeping any existing preamble/sections ahead of it.
+ const block = `## ${section.trim()}\n\n${body}`;
+ out = report.trim() === "" ? block : `${report.replace(/\s+$/, "")}\n\n${block}`;
+ } else {
+ // Replace the matched section's body, preserving its original heading line.
+ const start = heads[hit].line;
+ const end = hit + 1 < heads.length ? heads[hit + 1].line : lines.length;
+ const before = lines.slice(0, start);
+ const after = lines.slice(end);
+ const block = `${lines[start]}\n\n${body}`;
+ out = [...before, ...block.split("\n"), ...after].join("\n");
+ }
+
+ return out.replace(/\n{3,}/g, "\n\n").trim();
+}
+
// ─── editable context (view / prune / forward to a new session) ───
const ROLE_MARKER = /^(USER|ASSISTANT):\s*$/i;
@@ -241,6 +288,7 @@ export function gatherSystemPrompt(
mode: "native" | "scripted",
budget: number,
canFetch = false,
+ canReport = false,
): string {
const base =
"You are helping answer a question about a video-transcript archive. " +
@@ -256,10 +304,17 @@ export function gatherSystemPrompt(
"fetch_context tool with the video's ref (and optionally a timestamp in " +
"seconds) to read the surrounding transcript before answering."
: "";
+ const reportLine = canReport
+ ? " You are maintaining a running report document that persists across turns. " +
+ "After gathering evidence, call the update_report tool to record findings in " +
+ "well-titled sections (pass the section heading and its full markdown " +
+ "content); keep the report the source of truth. Then answer briefly."
+ : "";
const tail =
mode === "native"
? "Call the search_transcripts tool to search." +
fetchLine +
+ reportLine +
" Call the finish tool as soon as you have enough excerpts, or " +
"immediately if no search is needed."
: "Reply with EXACTLY one line and nothing else: either " +
diff --git a/export/app/lib/nativeTools/anthropic.ts b/export/app/lib/nativeTools/anthropic.ts
@@ -13,6 +13,10 @@ import {
FINISH_TOOL_NAME,
type ParsedToolCall,
postJson,
+ REPORT_BUDGET,
+ REPORT_PARAMS,
+ REPORT_TOOL_DESCRIPTION,
+ REPORT_TOOL_NAME,
SEARCH_PARAMS,
SEARCH_TOOL_DESCRIPTION,
SEARCH_TOOL_NAME,
@@ -20,8 +24,9 @@ import {
} from "./shared";
// The fetch_context tool is offered only when the turn can read transcripts
-// (expand mode with a pinned set); search + finish are always present.
-function buildTools(includeFetch: boolean) {
+// (expand mode with a pinned set); update_report only in report mode; search +
+// finish are always present.
+function buildTools(includeFetch: boolean, includeReport: boolean) {
const tools: { name: string; description: string; input_schema: unknown }[] = [
{
name: SEARCH_TOOL_NAME,
@@ -36,6 +41,13 @@ function buildTools(includeFetch: boolean) {
input_schema: FETCH_PARAMS,
});
}
+ if (includeReport) {
+ tools.push({
+ name: REPORT_TOOL_NAME,
+ description: REPORT_TOOL_DESCRIPTION,
+ input_schema: REPORT_PARAMS,
+ });
+ }
tools.push({
name: FINISH_TOOL_NAME,
description: FINISH_TOOL_DESCRIPTION,
@@ -54,7 +66,13 @@ export function parseAnthropicToolUses(json: unknown): ParsedToolCall[] {
type?: string;
id?: string;
name?: string;
- input?: { query?: unknown; video?: unknown; aroundSeconds?: unknown };
+ input?: {
+ query?: unknown;
+ video?: unknown;
+ aroundSeconds?: unknown;
+ section?: unknown;
+ content?: unknown;
+ };
};
if (b.type === "tool_use" && typeof b.name === "string") {
out.push({
@@ -66,6 +84,12 @@ export function parseAnthropicToolUses(json: unknown): ParsedToolCall[] {
typeof b.input?.aroundSeconds === "number"
? b.input.aroundSeconds
: undefined,
+ ...(typeof b.input?.section === "string"
+ ? { section: b.input.section }
+ : {}),
+ ...(typeof b.input?.content === "string"
+ ? { content: b.input.content }
+ : {}),
});
}
}
@@ -74,14 +98,20 @@ export function parseAnthropicToolUses(json: unknown): ParsedToolCall[] {
export async function anthropicGather(ctx: NativeGatherContext): Promise<void> {
const canFetch = !!ctx.runFetchContext;
- const tools = buildTools(canFetch);
+ const canReport = !!ctx.runUpdateReport;
+ const tools = buildTools(canFetch, canReport);
const messages: { role: string; content: unknown }[] = [
...ctx.history.map((m) => ({ role: m.role, content: m.content })),
{ role: "user", content: ctx.question },
];
let searchesRun = 0;
let fetchesRun = 0;
- const maxRounds = ctx.budget + 1 + (canFetch ? FETCH_BUDGET : 0);
+ let reportsRun = 0;
+ const maxRounds =
+ ctx.budget +
+ 1 +
+ (canFetch ? FETCH_BUDGET : 0) +
+ (canReport ? REPORT_BUDGET : 0);
for (let round = 0; round <= maxRounds; round++) {
if (ctx.signal?.aborted) throw abortError();
@@ -110,8 +140,13 @@ export async function anthropicGather(ctx: NativeGatherContext): Promise<void> {
const wantsFetch =
canFetch &&
calls.some((c) => c.name === FETCH_TOOL_NAME && c.video.trim() !== "");
+ const wantsReport =
+ canReport &&
+ calls.some(
+ (c) => c.name === REPORT_TOOL_NAME && (c.section ?? "").trim() !== "",
+ );
// No actionable tool call → the model finished (or answered directly). Done.
- if (!wantsSearch && !wantsFetch) return;
+ if (!wantsSearch && !wantsFetch && !wantsReport) return;
// Replay the assistant's tool_use turn verbatim, then answer each tool_use.
messages.push({
@@ -139,6 +174,20 @@ export async function anthropicGather(ctx: NativeGatherContext): Promise<void> {
} else {
content = "Reached the limit on transcript reads this turn.";
}
+ } else if (
+ canReport &&
+ call.name === REPORT_TOOL_NAME &&
+ (call.section ?? "").trim() !== ""
+ ) {
+ if (reportsRun < REPORT_BUDGET) {
+ reportsRun += 1;
+ content = await ctx.runUpdateReport!(
+ call.section!,
+ call.content ?? "",
+ );
+ } else {
+ content = "Reached the limit on report updates this turn.";
+ }
} else {
// finish (or any other tool): acknowledge so the block is satisfied.
content = "Acknowledged.";
diff --git a/export/app/lib/nativeTools/gemini.ts b/export/app/lib/nativeTools/gemini.ts
@@ -13,13 +13,17 @@ import {
FINISH_TOOL_NAME,
type ParsedToolCall,
postJson,
+ REPORT_BUDGET,
+ REPORT_PARAMS,
+ REPORT_TOOL_DESCRIPTION,
+ REPORT_TOOL_NAME,
SEARCH_PARAMS,
SEARCH_TOOL_DESCRIPTION,
SEARCH_TOOL_NAME,
type NativeGatherContext,
} from "./shared";
-function buildTools(includeFetch: boolean) {
+function buildTools(includeFetch: boolean, includeReport: boolean) {
const functionDeclarations: {
name: string;
description: string;
@@ -38,6 +42,13 @@ function buildTools(includeFetch: boolean) {
parameters: FETCH_PARAMS,
});
}
+ if (includeReport) {
+ functionDeclarations.push({
+ name: REPORT_TOOL_NAME,
+ description: REPORT_TOOL_DESCRIPTION,
+ parameters: REPORT_PARAMS,
+ });
+ }
functionDeclarations.push({
name: FINISH_TOOL_NAME,
description: FINISH_TOOL_DESCRIPTION,
@@ -60,7 +71,13 @@ export function parseGeminiFunctionCalls(json: unknown): ParsedToolCall[] {
part as {
functionCall?: {
name?: string;
- args?: { query?: unknown; video?: unknown; aroundSeconds?: unknown };
+ args?: {
+ query?: unknown;
+ video?: unknown;
+ aroundSeconds?: unknown;
+ section?: unknown;
+ content?: unknown;
+ };
};
}
).functionCall;
@@ -74,6 +91,12 @@ export function parseGeminiFunctionCalls(json: unknown): ParsedToolCall[] {
typeof fc.args?.aroundSeconds === "number"
? fc.args.aroundSeconds
: undefined,
+ ...(typeof fc.args?.section === "string"
+ ? { section: fc.args.section }
+ : {}),
+ ...(typeof fc.args?.content === "string"
+ ? { content: fc.args.content }
+ : {}),
});
}
}
@@ -82,7 +105,8 @@ export function parseGeminiFunctionCalls(json: unknown): ParsedToolCall[] {
export async function geminiGather(ctx: NativeGatherContext): Promise<void> {
const canFetch = !!ctx.runFetchContext;
- const tools = buildTools(canFetch);
+ const canReport = !!ctx.runUpdateReport;
+ const tools = buildTools(canFetch, canReport);
const model = ctx.model || PROVIDERS.gemini.defaultModel;
const url =
`https://generativelanguage.googleapis.com/v1beta/models/` +
@@ -96,7 +120,12 @@ export async function geminiGather(ctx: NativeGatherContext): Promise<void> {
];
let searchesRun = 0;
let fetchesRun = 0;
- const maxRounds = ctx.budget + 1 + (canFetch ? FETCH_BUDGET : 0);
+ let reportsRun = 0;
+ const maxRounds =
+ ctx.budget +
+ 1 +
+ (canFetch ? FETCH_BUDGET : 0) +
+ (canReport ? REPORT_BUDGET : 0);
for (let round = 0; round <= maxRounds; round++) {
if (ctx.signal?.aborted) throw abortError();
@@ -119,7 +148,12 @@ export async function geminiGather(ctx: NativeGatherContext): Promise<void> {
const wantsFetch =
canFetch &&
calls.some((c) => c.name === FETCH_TOOL_NAME && c.video.trim() !== "");
- if (!wantsSearch && !wantsFetch) return; // finish or a plain answer → done
+ const wantsReport =
+ canReport &&
+ calls.some(
+ (c) => c.name === REPORT_TOOL_NAME && (c.section ?? "").trim() !== "",
+ );
+ if (!wantsSearch && !wantsFetch && !wantsReport) return; // finish/answer → done
// Replay the model's functionCall turn, then send functionResponse parts.
const modelParts = calls.map((c) => ({
@@ -135,7 +169,9 @@ export async function geminiGather(ctx: NativeGatherContext): Promise<void> {
? { aroundSeconds: c.aroundSeconds }
: {}),
}
- : {},
+ : c.name === REPORT_TOOL_NAME
+ ? { section: c.section ?? "", content: c.content ?? "" }
+ : {},
},
}));
contents.push({ role: "model", parts: modelParts });
@@ -161,6 +197,17 @@ export async function geminiGather(ctx: NativeGatherContext): Promise<void> {
} else {
result = "Reached the limit on transcript reads this turn.";
}
+ } else if (
+ canReport &&
+ call.name === REPORT_TOOL_NAME &&
+ (call.section ?? "").trim() !== ""
+ ) {
+ if (reportsRun < REPORT_BUDGET) {
+ reportsRun += 1;
+ result = await ctx.runUpdateReport!(call.section!, call.content ?? "");
+ } else {
+ result = "Reached the limit on report updates this turn.";
+ }
} else {
result = "Acknowledged.";
}
diff --git a/export/app/lib/nativeTools/openai.ts b/export/app/lib/nativeTools/openai.ts
@@ -13,13 +13,17 @@ import {
FINISH_TOOL_NAME,
type ParsedToolCall,
postJson,
+ REPORT_BUDGET,
+ REPORT_PARAMS,
+ REPORT_TOOL_DESCRIPTION,
+ REPORT_TOOL_NAME,
SEARCH_PARAMS,
SEARCH_TOOL_DESCRIPTION,
SEARCH_TOOL_NAME,
type NativeGatherContext,
} from "./shared";
-function buildTools(includeFetch: boolean) {
+function buildTools(includeFetch: boolean, includeReport: boolean) {
const tools: {
type: "function";
function: { name: string; description: string; parameters: unknown };
@@ -43,6 +47,16 @@ function buildTools(includeFetch: boolean) {
},
});
}
+ if (includeReport) {
+ tools.push({
+ type: "function",
+ function: {
+ name: REPORT_TOOL_NAME,
+ description: REPORT_TOOL_DESCRIPTION,
+ parameters: REPORT_PARAMS,
+ },
+ });
+ }
tools.push({
type: "function",
function: {
@@ -69,21 +83,34 @@ export function parseOpenAIToolCalls(json: unknown): ParsedToolCall[] {
let query = "";
let video = "";
let aroundSeconds: number | undefined;
+ let section: string | undefined;
+ let content: string | undefined;
try {
const args = JSON.parse(c.function?.arguments ?? "{}");
if (typeof args.query === "string") query = args.query;
if (typeof args.video === "string") video = args.video;
if (typeof args.aroundSeconds === "number") aroundSeconds = args.aroundSeconds;
+ if (typeof args.section === "string") section = args.section;
+ if (typeof args.content === "string") content = args.content;
} catch {
/* malformed args → empty */
}
- return { id: c.id ?? "", name: c.function?.name ?? "", query, video, aroundSeconds };
+ return {
+ id: c.id ?? "",
+ name: c.function?.name ?? "",
+ query,
+ video,
+ aroundSeconds,
+ ...(section !== undefined ? { section } : {}),
+ ...(content !== undefined ? { content } : {}),
+ };
});
}
export async function openaiGather(ctx: NativeGatherContext): Promise<void> {
const canFetch = !!ctx.runFetchContext;
- const tools = buildTools(canFetch);
+ const canReport = !!ctx.runUpdateReport;
+ const tools = buildTools(canFetch, canReport);
const messages: unknown[] = [
{ role: "system", content: ctx.system },
...ctx.history.map((m) => ({ role: m.role, content: m.content })),
@@ -91,7 +118,12 @@ export async function openaiGather(ctx: NativeGatherContext): Promise<void> {
];
let searchesRun = 0;
let fetchesRun = 0;
- const maxRounds = ctx.budget + 1 + (canFetch ? FETCH_BUDGET : 0);
+ let reportsRun = 0;
+ const maxRounds =
+ ctx.budget +
+ 1 +
+ (canFetch ? FETCH_BUDGET : 0) +
+ (canReport ? REPORT_BUDGET : 0);
for (let round = 0; round <= maxRounds; round++) {
if (ctx.signal?.aborted) throw abortError();
@@ -118,8 +150,13 @@ export async function openaiGather(ctx: NativeGatherContext): Promise<void> {
const wantsFetch =
canFetch &&
calls.some((c) => c.name === FETCH_TOOL_NAME && c.video.trim() !== "");
+ const wantsReport =
+ canReport &&
+ calls.some(
+ (c) => c.name === REPORT_TOOL_NAME && (c.section ?? "").trim() !== "",
+ );
// No actionable tool call → done gathering.
- if (!wantsSearch && !wantsFetch) return;
+ if (!wantsSearch && !wantsFetch && !wantsReport) return;
// Replay the assistant message with its tool_calls, then answer EVERY call
// (OpenAI requires a tool response for each tool_call id).
@@ -144,6 +181,17 @@ export async function openaiGather(ctx: NativeGatherContext): Promise<void> {
} else {
content = "Reached the limit on transcript reads this turn.";
}
+ } else if (
+ canReport &&
+ call.name === REPORT_TOOL_NAME &&
+ (call.section ?? "").trim() !== ""
+ ) {
+ if (reportsRun < REPORT_BUDGET) {
+ reportsRun += 1;
+ content = await ctx.runUpdateReport!(call.section!, call.content ?? "");
+ } else {
+ content = "Reached the limit on report updates this turn.";
+ }
} else {
content = "Acknowledged.";
}
diff --git a/export/app/lib/nativeTools/parsers.test.ts b/export/app/lib/nativeTools/parsers.test.ts
@@ -3,7 +3,7 @@ import assert from "node:assert/strict";
import { parseAnthropicToolUses } from "./anthropic";
import { parseOpenAIToolCalls } from "./openai";
import { parseGeminiFunctionCalls } from "./gemini";
-import { FETCH_TOOL_NAME, SEARCH_TOOL_NAME } from "./shared";
+import { FETCH_TOOL_NAME, REPORT_TOOL_NAME, SEARCH_TOOL_NAME } from "./shared";
test("parseAnthropicToolUses reads search and fetch_context calls", () => {
const calls = parseAnthropicToolUses({
@@ -69,6 +69,78 @@ test("parseOpenAIToolCalls tolerates malformed arguments", () => {
assert.equal(calls[0].video, "");
});
+test("parseAnthropicToolUses reads update_report calls (section + content)", () => {
+ const calls = parseAnthropicToolUses({
+ content: [
+ {
+ type: "tool_use",
+ id: "r1",
+ name: REPORT_TOOL_NAME,
+ input: { section: "Findings", content: "Alpha is discussed." },
+ },
+ ],
+ });
+ assert.equal(calls.length, 1);
+ assert.equal(calls[0].name, REPORT_TOOL_NAME);
+ assert.equal(calls[0].section, "Findings");
+ assert.equal(calls[0].content, "Alpha is discussed.");
+ // A search call carries no section/content keys (deep-equality with the older
+ // shape still holds).
+ const search = parseAnthropicToolUses({
+ content: [{ type: "tool_use", id: "s", name: SEARCH_TOOL_NAME, input: { query: "x" } }],
+ });
+ assert.equal("section" in search[0], false);
+ assert.equal("content" in search[0], false);
+});
+
+test("parseOpenAIToolCalls reads update_report calls (section + content)", () => {
+ const calls = parseOpenAIToolCalls({
+ choices: [
+ {
+ message: {
+ tool_calls: [
+ {
+ id: "r1",
+ function: {
+ name: REPORT_TOOL_NAME,
+ arguments: JSON.stringify({
+ section: "Timeline",
+ content: "- 2020: a thing",
+ }),
+ },
+ },
+ ],
+ },
+ },
+ ],
+ });
+ assert.equal(calls[0].name, REPORT_TOOL_NAME);
+ assert.equal(calls[0].section, "Timeline");
+ assert.equal(calls[0].content, "- 2020: a thing");
+});
+
+test("parseGeminiFunctionCalls reads update_report calls (section + content)", () => {
+ const calls = parseGeminiFunctionCalls({
+ candidates: [
+ {
+ content: {
+ parts: [
+ {
+ functionCall: {
+ name: REPORT_TOOL_NAME,
+ args: { section: "Sources", content: "[1] video" },
+ },
+ },
+ ],
+ },
+ },
+ ],
+ });
+ assert.equal(calls[0].name, REPORT_TOOL_NAME);
+ assert.equal(calls[0].section, "Sources");
+ assert.equal(calls[0].content, "[1] video");
+});
+
test("parseGeminiFunctionCalls reads search and fetch_context calls", () => {
const calls = parseGeminiFunctionCalls({
candidates: [
diff --git a/export/app/lib/nativeTools/shared.ts b/export/app/lib/nativeTools/shared.ts
@@ -24,17 +24,25 @@ export type NativeGatherContext = {
// window text. Present only in expand mode with a pinned set (native only);
// when undefined the fetch_context tool is not offered.
runFetchContext?: (video: string, aroundSeconds?: number) => Promise<string>;
+ // Upserts a section of the running report document (report mode, native only).
+ // When undefined the update_report tool is not offered.
+ runUpdateReport?: (section: string, content: string) => Promise<string>;
signal?: AbortSignal;
};
export const SEARCH_TOOL_NAME = "search_transcripts";
export const FETCH_TOOL_NAME = "fetch_context";
+export const REPORT_TOOL_NAME = "update_report";
export const FINISH_TOOL_NAME = "finish";
// Bounds transcript reads per turn (independent of the search budget) so the
// user's key isn't spent on unbounded fetching.
export const FETCH_BUDGET = 4;
+// Bounds report writes per turn (independent of the search budget). Report mode
+// only needs a handful of section upserts per turn to keep the document current.
+export const REPORT_BUDGET = 6;
+
export const SEARCH_TOOL_DESCRIPTION =
"Search the video-transcript archive for excerpts relevant to a query. " +
"Returns matching videos with timestamped snippet lines. Use focused keyword " +
@@ -46,6 +54,11 @@ export const FETCH_TOOL_DESCRIPTION =
"result) and optionally a timestamp in seconds to centre on. The returned " +
"lines are added to your citable excerpts for that video.";
+export const REPORT_TOOL_DESCRIPTION =
+ "Create or update a section of the running report document that persists " +
+ "across turns. Pass the section heading and its full new markdown content; " +
+ "the section is replaced (or appended if new).";
+
export const FINISH_TOOL_DESCRIPTION =
"Call this when you have gathered enough excerpts (or none are needed) and " +
"are ready to answer.";
@@ -81,16 +94,36 @@ export const FETCH_PARAMS = {
required: ["video"],
} as const;
+// JSON-Schema for the update_report tool's input.
+export const REPORT_PARAMS = {
+ type: "object",
+ properties: {
+ section: {
+ type: "string",
+ description: "The section heading (without the leading ##).",
+ },
+ content: {
+ type: "string",
+ description: "The full new markdown content for the section.",
+ },
+ },
+ required: ["section", "content"],
+} as const;
+
export const EMPTY_PARAMS = { type: "object", properties: {} } as const;
// Shape returned by every provider's tool-call parser: `query` for search,
-// `video`/`aroundSeconds` for fetch_context (all optional, filled per tool).
+// `video`/`aroundSeconds` for fetch_context, `section`/`content` for
+// update_report (all optional, filled per tool — omitted keys stay absent so
+// deep-equality against a search/finish call still matches).
export type ParsedToolCall = {
id: string;
name: string;
query: string;
video: string;
aroundSeconds?: number;
+ section?: string;
+ content?: string;
};
// Thrown when a native tool-calling request fails in a way that suggests the
diff --git a/export/app/lib/searchAgent.ts b/export/app/lib/searchAgent.ts
@@ -17,6 +17,7 @@ import {
import { retrieve, type RetrievedVideo } from "./askRetrieval";
import {
answerSystemPrompt,
+ applyReportPatch,
buildApiMessages,
buildGroundedContent,
collectPriorPool,
@@ -49,6 +50,8 @@ export type AgentEvent =
// The model read more of a pinned video's transcript (fetch_context tool).
| { type: "fetch_start"; ref: string; label: string }
| { type: "fetch_done"; ref: string; count: number }
+ // The model upserted a section of the running report (update_report tool).
+ | { type: "report_update"; section: string }
// Fired once gather is done, BEFORE the answer streams — carries the grounding
// so the UI can persist it and a failed answer stream can be retried without
// re-searching.
@@ -225,6 +228,12 @@ export type RunAskTurnOptions = {
// Human-edited context: replaces the history reconstructed from `prior` (used
// when the user has pruned the context in the context panel).
historyOverride?: ChatMessage[];
+ // Report mode (native providers only): the running report document carried in,
+ // and a flag turning the mode on. When on, the prior Q&A text replay is dropped
+ // in favour of a compact "report so far" history (the excerpt pool still flows),
+ // and the update_report tool is offered so the model keeps the report current.
+ report?: string;
+ reportMode?: boolean;
};
export type AskTurnResult = {
@@ -232,6 +241,9 @@ export type AskTurnResult = {
videos: RetrievedVideo[];
groundedContent: string;
truncated: boolean;
+ // The report after this turn (possibly updated via update_report). Echoes the
+ // input report unchanged when report mode is off or nothing was written.
+ report: string;
};
export async function runAskTurn(
@@ -250,7 +262,28 @@ export async function runAskTurn(
onEvent,
} = opts;
const budget = opts.budget ?? DEFAULT_BUDGET;
- const history = opts.historyOverride ?? buildApiMessages(prior);
+ const reportMode = opts.reportMode === true;
+ // The running report — mutated in place by the update_report executor below and
+ // returned in the result so the caller can persist it.
+ let report = opts.report ?? "";
+
+ // Report mode COMPACTION: instead of replaying every prior Q&A turn as text,
+ // send a single "report so far" assistant message (empty report → no history at
+ // all). The cumulative excerpt pool still flows into the grounding unchanged —
+ // only the prior Q&A text replay is dropped, keeping long sessions bounded.
+ const compactHistory = (): ChatMessage[] =>
+ report.trim() === ""
+ ? []
+ : [
+ {
+ role: "assistant",
+ content:
+ "Report so far (keep it current with update_report):\n\n" + report,
+ },
+ ];
+ const history = reportMode
+ ? compactHistory()
+ : opts.historyOverride ?? buildApiMessages(prior);
// Retry path: grounding already gathered on a prior attempt — skip gather and
// go straight to (re)streaming the answer over the same excerpts.
@@ -272,6 +305,7 @@ export async function runAskTurn(
videos: pv,
groundedContent,
truncated: opts.precomputedGrounding.truncated ?? false,
+ report,
};
}
@@ -359,6 +393,19 @@ export async function runAskTurn(
);
};
+ // Report mode: apply the model's section upsert to the running report, emit an
+ // event so the UI can show a live "updating report" indicator, and hand back a
+ // short confirmation as the tool result. Report writes don't consume the search
+ // budget (they have their own REPORT_BUDGET inside each provider loop).
+ const runUpdateReport = async (
+ section: string,
+ content: string,
+ ): Promise<string> => {
+ report = applyReportPatch(report, section, content);
+ onEvent({ type: "report_update", section });
+ return `Updated the "${section}" section.`;
+ };
+
const gatherCtx: NativeGatherContext = {
apiKey,
model,
@@ -372,12 +419,14 @@ export async function runAskTurn(
const doGather = async (kind: "native" | "scripted"): Promise<void> => {
const canFetch = kind === "native" && canFetchThisTurn;
- let system = gatherSystemPrompt(aliases, kind, budget, canFetch);
+ const canReport = kind === "native" && reportMode;
+ let system = gatherSystemPrompt(aliases, kind, budget, canFetch, canReport);
if (canFetchThisTurn) system += `\n\n${buildSeedDigest(seedVideos)}`;
const ctx: NativeGatherContext = {
...gatherCtx,
system,
runFetchContext: canFetch ? runFetchContext : undefined,
+ runUpdateReport: canReport ? runUpdateReport : undefined,
};
if (kind === "scripted") return scriptedGather(ctx, provider);
switch (provider) {
@@ -420,17 +469,21 @@ export async function runAskTurn(
].slice(0, MAX_CONTEXT_VIDEOS);
const groundedContent = buildGroundedContent(question, finalVideos);
+ // In report mode, rebuild the compact history from the NOW-updated report so the
+ // answer phase sees the freshest report the gather loop wrote.
+ const answerHistory = reportMode ? compactHistory() : history;
+
onEvent({ type: "answer_start", videos: finalVideos, groundedContent });
const answer = await askStream({
provider,
apiKey,
model,
system: answerSystemPrompt(aliases),
- messages: [...history, { role: "user", content: groundedContent }],
+ messages: [...answerHistory, { role: "user", content: groundedContent }],
maxTokens: 2048,
signal,
onDelta: (chunk) => onEvent({ type: "delta", text: chunk }),
});
- return { answer, videos: finalVideos, groundedContent, truncated };
+ return { answer, videos: finalVideos, groundedContent, truncated, report };
}
diff --git a/export/e2e/ask-chat.spec.ts b/export/e2e/ask-chat.spec.ts
@@ -610,6 +610,99 @@ test.describe("ask chat", () => {
).toBeVisible();
});
+ test("report mode: the model maintains a report and the history is compacted", async ({
+ page,
+ }) => {
+ await installRoutes(page);
+ let reportToolOffered = false;
+ const answerBodies: string[] = [];
+ await page.route("https://api.anthropic.com/**", async (route) => {
+ if (route.request().method() === "OPTIONS") {
+ await route.fulfill({ status: 204, headers: CORS });
+ return;
+ }
+ const body = route.request().postDataJSON() as {
+ system?: string;
+ tools?: { name?: string }[];
+ messages?: { role: string; content: unknown }[];
+ };
+ const system = body.system ?? "";
+ // Answer phase → capture the messages so we can assert compaction.
+ if (system.includes("Markdown")) {
+ answerBodies.push(JSON.stringify(body.messages ?? []));
+ await route.fulfill({
+ status: 200,
+ headers: { ...CORS, "content-type": "text/event-stream" },
+ body: sse("Report-mode answer [1]"),
+ });
+ return;
+ }
+ // Native gather turn: write a report section once, then finish.
+ if (Array.isArray(body.tools)) {
+ if (body.tools.some((t) => t.name === "update_report")) {
+ reportToolOffered = true;
+ }
+ const msgs = body.messages ?? [];
+ const lastUser = [...msgs].reverse().find((m) => m.role === "user");
+ const isToolResult = Array.isArray(lastUser?.content);
+ await route.fulfill({
+ status: 200,
+ headers: { ...CORS, "content-type": "application/json" },
+ body: isToolResult
+ ? toolUse("finish", {})
+ : toolUse("update_report", {
+ section: "Findings",
+ content: "Alpha finding recorded from the archive.",
+ }),
+ });
+ return;
+ }
+ await route.fulfill({
+ status: 200,
+ headers: { ...CORS, "content-type": "text/event-stream" },
+ body: sse("DONE"),
+ });
+ });
+ await page.goto("/ask/");
+ await keyIn(page, "Native tools");
+
+ // Enable report mode (the toggle lives in the Report panel header).
+ await page.getByLabel("Report mode").check();
+
+ // Turn 1 establishes the report + some history.
+ await ask(page, "first question about the alpha topic");
+ await expect(page.getByText(/Report-mode answer/)).toBeVisible();
+ expect(reportToolOffered).toBe(true);
+
+ // The Report panel shows the section the model wrote.
+ await page.locator("summary", { hasText: "Report" }).click();
+ await expect(
+ page.getByText("Alpha finding recorded from the archive."),
+ ).toBeVisible();
+
+ // Turn 2: the answer request must be COMPACTED — it carries the "Report so
+ // far" summary and does NOT replay turn 1's full question text.
+ await ask(page, "second question about the beta topic");
+ await expect(page.getByText(/Report-mode answer/).nth(1)).toBeVisible();
+
+ const lastAnswer = answerBodies[answerBodies.length - 1];
+ expect(lastAnswer).toContain("Report so far");
+ expect(lastAnswer).toContain("Alpha finding recorded from the archive.");
+ expect(lastAnswer).not.toContain("first question about the alpha topic");
+ });
+
+ test("report mode is disabled on the scripted transport", async ({ page }) => {
+ await installRoutes(page);
+ await page.goto("/ask/");
+ // keyIn fills the key first (proving hydration) before switching to Scripted,
+ // so the mode click isn't lost to a pre-hydration no-op.
+ await keyIn(page, "Scripted");
+ // The toggle is present but disabled (no tool to maintain the report).
+ const toggle = page.getByLabel("Report mode");
+ await expect(toggle).toBeVisible();
+ await expect(toggle).toBeDisabled();
+ });
+
test("editing the context changes what the next turn sends", async ({
page,
}) => {
