fix(agent): stop cutting the transcript at the 800th word

electron/ai-edition/agent-tools.test.ts

@@ -205,6 +205,36 @@ describe("executeAgentTool", () => {
 		expect(payload.segments[1].kind).toBe("silence");
 	});
 
+	it("getTranscript returns a long transcript whole, word for word", () => {
+		// The regression test for a `.slice(0, 800)` that used to sit here. On the
+		// production path a segment is one WORD, so the cap cut a half-hour
+		// recording at roughly its fifth minute and reported nothing — the model
+		// trimmed the silences it could see and called the job done. 4000 words is
+		// about half an hour of speech.
+		const base = fixtureDocument();
+		const segments = Array.from({ length: 4000 }, (_, i) => ({
+			id: `seg_${i}`,
+			kind: "speech" as const,
+			startSec: i * 0.45,
+			endSec: i * 0.45 + 0.4,
+			text: `mot${i}`,
+			wordIds: [],
+		}));
+		const doc = {
+			...base,
+			transcript: null,
+			transcripts: [{ ...base.transcripts[0], segments }],
+		};
+
+		const result = executeAgentTool(doc, "getTranscript", "{}");
+		expect(result.ok).toBe(true);
+		const payload = JSON.parse(result.resultJson);
+		expect(payload.segments).toHaveLength(4000);
+		// The last word matters more than the count: a cap keeps the head and
+		// drops the tail, so the tail is what proves it is gone.
+		expect(payload.segments.at(-1).text).toBe("mot3999");
+	});
+
 	it("getTranscript fails cleanly when no transcript exists", () => {
 		const doc = { ...fixtureDocument(), transcripts: [], transcript: null };
 		const result = executeAgentTool(doc, "getTranscript", "{}");

electron/ai-edition/agent-tools.ts

@@ -854,9 +854,21 @@ export function executeAgentTool(
 			if (!transcript) {
 				return failure(`No transcript for asset ${assetId ?? "(none)"}.`);
 			}
-			// ponytail: segments only — words would blow the context for long
-			// recordings and the segment text already carries the content.
-			const segments = transcript.segments.slice(0, 800).map((s) => ({
+			// ponytail: no cap. There used to be a `.slice(0, 800)` here, guarded by
+			// "words would blow the context" — written believing a segment was a
+			// phrase. On the production path a segment IS one word
+			// (src/lib/captioning/transcribe.ts: whisper's word timings are mapped
+			// one-to-one), so the cap cut the transcript at the 800th WORD — around
+			// five minutes of speech — and said nothing about it. The model read a
+			// fifth of a half-hour recording, cut the silences it could see, and
+			// reported the job done, because nothing in the payload told it otherwise.
+			//
+			// A whole 30-minute transcript is ~285k characters, ~70k tokens: large,
+			// and well inside every model this app talks to. If a recording ever does
+			// get near a window, the honest fix is to know the window — the app has no
+			// per-model context budget today — not to guess a number here and drop the
+			// rest in silence.
+			const segments = transcript.segments.map((s) => ({
 				id: s.id,
 				kind: s.kind,
 				startSec: s.startSec,

technical-documentation/architecture/ai-agent.md

@@ -46,7 +46,7 @@ The model never free-writes the project document. It can only call the fixed set
 | Tool | What it does | What it mutates |
 |---|---|---|
 | `getCurrentDocument` | Reads a compact project, asset, clip, trim, and modifier snapshot with explicit time bases. Each asset reports `hasCameraTrack` / `cameraVisible` / `hasCursorTelemetry` beside `hasTranscript` (`hasCursorTelemetry` is three-valued: `true`, `false` when the asset was checked and has none, `null` when it was not checked — never `false` for something we failed to look at), the document reports `hasAnyCamera` and `autoFocusAll`, and each zoom reports the `renderedScale` the viewer will see plus `customScale` / `depthIsOverridden` when a custom scale makes its `depth` inert. | Nothing. |
-| `getTranscript` | Reads up to 800 transcript segments for an asset or the primary asset. | Nothing. |
+| `getTranscript` | Reads the transcript segments for an asset, or the primary asset, in full. On the production path a segment is one word, so a half-hour recording is a few thousand of them — there is no cap, and no per-model context budget to derive one from. | Nothing. |
 | `getCursorTrack` | Reads the recorded pointer telemetry for an asset as a DIGEST: the moments the cursor sat still or clicked, each with its hold, its average position, its click count, its source time and the `virtualSec` that `addZoom` takes — never the raw samples. Answers `available:false` with `reason:"no-sidecar"` (checked, this asset has none) or `reason:"unavailable"` (could not be read from here), and the two are never conflated. | Nothing. |
 | `addTrim` | Adds a source-time cut inside a clip. | `timeline.trimRanges`. |
 | `setTrim` | Moves or resizes an existing source-time trim. | The matching `timeline.trimRanges` entry. |

workbench/lib/fixtures.ts

@@ -356,8 +356,9 @@ export function multipleModifiers(options?: { projectId?: string }): AxcutDocume
 	});
 }
 
-/** More than 800 transcript segments — `agent-tools.ts:625` slices at 800 and
- * says nothing about it (DSL-6). */
+/** More than 800 transcript segments. `getTranscript` used to slice at 800 and
+ * say nothing about it; it no longer caps, and this fixture is what keeps that
+ * honest. */
 export function longTranscript(options?: { segments?: number; projectId?: string }): AxcutDocument {
 	const count = options?.segments ?? 900;
 	const durationSec = count * 2;