xonecas.io
64875
Mc

mini-coder / commit 472306b

perf: cache context usage for large sessions

xonecas2026-04-12 472306b97bbbffe6d6e263efe68a96b51149a6c4 parent 7a9e8e8

11 files changed

src/headless.test.ts+1−0
@@ -40,6 +40,7 @@ function createTestState(): AppState {
4040 effort: "medium",
4141 messages: [],
4242 stats: { totalInput: 0, totalOutput: 0, totalCost: 0 },
43+ contextTokens: 0,
4344 agentsMd: [],
4445 skills: [],
4546 plugins: [],
src/index.test.ts+1−0
@@ -36,6 +36,7 @@ function createTestState(plugins: LoadedPlugin[] = []): AppState {
3636 effort: "medium",
3737 messages: [],
3838 stats: { totalInput: 0, totalOutput: 0, totalCost: 0 },
39+ contextTokens: 0,
3940 agentsMd: [],
4041 skills: [],
4142 plugins,
src/index.ts+5−0
@@ -47,6 +47,7 @@ import {
4747 } from "./prompt.ts";
4848 import {
4949 appendMessage,
50+ computeContextTokens,
5051 computeStats,
5152 createSession,
5253 filterModelMessages,
@@ -526,6 +527,8 @@ export interface AppState {
526527 messages: ReturnType<typeof loadMessages>;
527528 /** Cumulative session input/output/cost stats for the status bar. */
528529 stats: SessionStats;
530+ /** Estimated model-visible context tokens for the next request. */
531+ contextTokens: number;
529532 /** Discovered AGENTS.md files. */
530533 agentsMd: AgentsMdFile[];
531534 /** Discovered skills. */
@@ -610,6 +613,7 @@ export async function init(): Promise<AppState> {
610613 const effort = startup.effort;
611614 const messages: ReturnType<typeof loadMessages> = [];
612615 const stats = computeStats(messages);
616+ const contextTokens = computeContextTokens(messages);
613617 const promptContext = await loadPromptContext(filterModelMessages(messages), {
614618 cwd,
615619 });
@@ -621,6 +625,7 @@ export async function init(): Promise<AppState> {
621625 effort,
622626 messages,
623627 stats,
628+ contextTokens,
624629 agentsMd: promptContext.agentsMd,
625630 skills: promptContext.skills,
626631 plugins: promptContext.plugins,
src/session.test.ts+48−0
@@ -9,9 +9,11 @@ import type {
99 UserMessage,
1010 } from "@mariozechner/pi-ai";
1111 import {
12+ addMessageToContextTokens,
1213 addMessageToStats,
1314 appendMessage,
1415 appendPromptHistory,
16+ computeContextTokens,
1517 computeStats,
1618 createSession,
1719 createUiMessage,
@@ -763,6 +765,52 @@ describe("cumulative stats", () => {
763765 expect(filtered[0]).toEqual(user);
764766 expect(filtered[1]).toEqual(assistant);
765767 });
768+
769+ test("computeContextTokens uses the latest valid assistant usage as the anchor", () => {
770+ const messages = [
771+ makeUser("first request"),
772+ makeAssistant("first reply", {
773+ input: 120,
774+ output: 30,
775+ cacheRead: 25,
776+ cacheWrite: 25,
777+ totalTokens: 0,
778+ }),
779+ makeUiMessage("ignored"),
780+ makeUser("follow-up"),
781+ makeToolResult("tc1", "done"),
782+ ];
783+
784+ expect(computeContextTokens(messages)).toBe(204);
785+ });
786+
787+ test("addMessageToContextTokens matches computeContextTokens across incremental updates", () => {
788+ const aborted = makeAssistant("partial reply");
789+ aborted.stopReason = "aborted";
790+
791+ const messages = [
792+ makeUser("hello world"),
793+ makeAssistant("anchored reply", {
794+ input: 200,
795+ output: 50,
796+ totalTokens: 250,
797+ }),
798+ makeToolResult("tc1", "ls\nREADME.md"),
799+ aborted,
800+ makeUiMessage("ignored"),
801+ makeUser("follow-up"),
802+ ];
803+
804+ let runningContextTokens = 0;
805+ for (const message of messages) {
806+ runningContextTokens = addMessageToContextTokens(
807+ runningContextTokens,
808+ message,
809+ );
810+ }
811+
812+ expect(runningContextTokens).toBe(computeContextTokens(messages));
813+ });
766814 });
767815
768816 describe("truncateSessions", () => {
src/session.ts+151−1
@@ -107,7 +107,7 @@ export interface UiMessage {
107107 }
108108
109109 /** Any message persisted in session history. */
110-type PersistedMessage = Message | UiMessage;
110+export type PersistedMessage = Message | UiMessage;
111111
112112 /** Options for creating a new session. */
113113 interface CreateSessionOpts {
@@ -1071,3 +1071,153 @@ export function computeStats(
10711071
10721072 return stats;
10731073 }
1074+
1075+// ---------------------------------------------------------------------------
1076+// Context estimation
1077+// ---------------------------------------------------------------------------
1078+
1079+/** Conservative fixed estimate for an image block's token footprint. */
1080+const ESTIMATED_IMAGE_TOKENS = 1_200;
1081+
1082+/** Calculate context tokens from assistant usage, falling back when `totalTokens` is zero. */
1083+function calculateUsageTokens(usage: AssistantMessage["usage"]): number {
1084+ return (
1085+ usage.totalTokens ||
1086+ usage.input + usage.output + usage.cacheRead + usage.cacheWrite
1087+ );
1088+}
1089+
1090+/** Estimate token usage from a character count using a conservative chars/4 heuristic. */
1091+function estimateCharacterTokens(charCount: number): number {
1092+ return Math.ceil(charCount / 4);
1093+}
1094+
1095+type UserMultipartContent = Exclude<
1096+ Extract<Message, { role: "user" }>["content"],
1097+ string
1098+>;
1099+type TextOrImageContentBlock =
1100+ | UserMultipartContent[number]
1101+ | Extract<Message, { role: "toolResult" }>["content"][number];
1102+
1103+function estimateTextOrImageContentTokens(
1104+ content: readonly TextOrImageContentBlock[],
1105+): number {
1106+ let chars = 0;
1107+ let imageTokens = 0;
1108+
1109+ for (const block of content) {
1110+ if (block.type === "text") {
1111+ chars += block.text.length;
1112+ continue;
1113+ }
1114+ if (block.type === "image") {
1115+ imageTokens += ESTIMATED_IMAGE_TOKENS;
1116+ }
1117+ }
1118+
1119+ return estimateCharacterTokens(chars) + imageTokens;
1120+}
1121+
1122+function estimateUserMessageTokens(
1123+ message: Extract<Message, { role: "user" }>,
1124+): number {
1125+ if (typeof message.content === "string") {
1126+ return estimateCharacterTokens(message.content.length);
1127+ }
1128+ return estimateTextOrImageContentTokens(message.content);
1129+}
1130+
1131+function estimateAssistantBlockCharacters(
1132+ block: Extract<Message, { role: "assistant" }>["content"][number],
1133+): number {
1134+ if (block.type === "text") {
1135+ return block.text.length;
1136+ }
1137+ if (block.type === "thinking") {
1138+ return block.thinking.length;
1139+ }
1140+ return block.name.length + JSON.stringify(block.arguments).length;
1141+}
1142+
1143+function estimateAssistantMessageTokens(
1144+ message: Extract<Message, { role: "assistant" }>,
1145+): number {
1146+ const chars = message.content.reduce((total, block) => {
1147+ return total + estimateAssistantBlockCharacters(block);
1148+ }, 0);
1149+ return estimateCharacterTokens(chars);
1150+}
1151+
1152+function estimateToolResultMessageTokens(
1153+ message: Extract<Message, { role: "toolResult" }>,
1154+): number {
1155+ return estimateTextOrImageContentTokens(message.content);
1156+}
1157+
1158+/** Estimate token usage for a model-visible message. */
1159+function estimateMessageTokens(message: Message): number {
1160+ switch (message.role) {
1161+ case "user":
1162+ return estimateUserMessageTokens(message);
1163+ case "assistant":
1164+ return estimateAssistantMessageTokens(message);
1165+ case "toolResult":
1166+ return estimateToolResultMessageTokens(message);
1167+ }
1168+}
1169+
1170+/**
1171+ * Fold one persisted message into the running context estimate for the next request.
1172+ *
1173+ * Assistant messages with valid usage anchor the full model-visible context for
1174+ * that point in the transcript, so they replace the running estimate. All other
1175+ * model-visible messages are added incrementally using the same conservative
1176+ * estimation logic used before the first valid assistant usage appears.
1177+ *
1178+ * @param contextTokens - Running estimate before this message.
1179+ * @param message - Persisted message to fold into the estimate.
1180+ * @returns Updated context-token estimate.
1181+ */
1182+export function addMessageToContextTokens(
1183+ contextTokens: number,
1184+ message: PersistedMessage,
1185+): number {
1186+ if (message.role === "ui") {
1187+ return contextTokens;
1188+ }
1189+
1190+ const usage = getAssistantUsage(message);
1191+ if (
1192+ message.role === "assistant" &&
1193+ usage &&
1194+ message.stopReason !== "aborted" &&
1195+ message.stopReason !== "error"
1196+ ) {
1197+ return calculateUsageTokens(usage);
1198+ }
1199+
1200+ return contextTokens + estimateMessageTokens(message);
1201+}
1202+
1203+/**
1204+ * Estimate the current model-visible context size for the next request.
1205+ *
1206+ * Recomputed on session-load boundaries and maintained incrementally during an
1207+ * active turn so render-time status-bar updates do not need to rescan the full
1208+ * message history.
1209+ *
1210+ * @param messages - Full persisted session history.
1211+ * @returns Estimated context tokens visible to the next model request.
1212+ */
1213+export function computeContextTokens(
1214+ messages: readonly PersistedMessage[],
1215+): number {
1216+ let contextTokens = 0;
1217+
1218+ for (const message of messages) {
1219+ contextTokens = addMessageToContextTokens(contextTokens, message);
1220+ }
1221+
1222+ return contextTokens;
1223+}
src/submit.ts+13−0
@@ -19,6 +19,7 @@ import {
1919 } from "./index.ts";
2020 import { parseInput } from "./input.ts";
2121 import {
22+ addMessageToContextTokens,
2223 addMessageToStats,
2324 appendMessage,
2425 appendPromptHistory,
@@ -212,9 +213,17 @@ function handleAgentEvent(event: AgentEvent, state: AppState): void {
212213 case "assistant_message":
213214 state.messages.push(event.message);
214215 state.stats = addMessageToStats(state.stats, event.message);
216+ state.contextTokens = addMessageToContextTokens(
217+ state.contextTokens,
218+ event.message,
219+ );
215220 break;
216221 case "tool_result":
217222 state.messages.push(event.message);
223+ state.contextTokens = addMessageToContextTokens(
224+ state.contextTokens,
225+ event.message,
226+ );
218227 break;
219228 case "text_delta":
220229 case "thinking_delta":
@@ -277,6 +286,10 @@ export async function submitResolvedInput(
277286
278287 const turn = appendMessage(state.db, session.id, userMessage);
279288 state.messages.push(userMessage);
289+ state.contextTokens = addMessageToContextTokens(
290+ state.contextTokens,
291+ userMessage,
292+ );
280293 hooks?.onUserMessage?.(state);
281294
282295 state.git = await loadGitState(state.cwd);
src/ui.test.ts+1−0
@@ -221,6 +221,7 @@ function createTestState(): AppState {
221221 effort: "medium",
222222 messages: [],
223223 stats: { totalInput: 0, totalOutput: 0, totalCost: 0 },
224+ contextTokens: 0,
224225 agentsMd: [],
225226 skills: [],
226227 plugins: [],
src/ui/commands.test.ts+5−0
@@ -8,6 +8,7 @@ import type { AppState } from "../index.ts";
88 import { COMMANDS } from "../input.ts";
99 import {
1010 appendPromptHistory,
11+ computeContextTokens,
1112 createSession,
1213 openDatabase,
1314 } from "../session.ts";
@@ -75,6 +76,7 @@ function createTestState(): AppState {
7576 effort: "medium",
7677 messages: [],
7778 stats: { totalInput: 0, totalOutput: 0, totalCost: 0 },
79+ contextTokens: 0,
7880 agentsMd: [],
7981 skills: [],
8082 plugins: [],
@@ -377,6 +379,7 @@ describe("ui/commands", () => {
377379 totalOutput: 0,
378380 totalCost: 0,
379381 });
382+ expect(state.contextTokens).toBe(computeContextTokens(state.messages));
380383 } finally {
381384 state.db.close();
382385 }
@@ -436,6 +439,7 @@ describe("ui/commands", () => {
436439 { role: "ui", kind: "info", content: "old", timestamp: 1 },
437440 ];
438441 state.stats = { totalInput: 10, totalOutput: 20, totalCost: 0.5 };
442+ state.contextTokens = 123;
439443
440444 try {
441445 expect(controller.handleCommand("new", state)).toBe(true);
@@ -449,6 +453,7 @@ describe("ui/commands", () => {
449453 totalOutput: 0,
450454 totalCost: 0,
451455 });
456+ expect(state.contextTokens).toBe(0);
452457 expect(state.agentsMd).toEqual([
453458 { path: "/tmp/reloaded/AGENTS.md", content: "Reloaded context" },
454459 ]);
src/ui/commands.ts+5−0
@@ -19,6 +19,7 @@ import type { AppState } from "../index.ts";
1919 import { getAvailableModels, saveOAuthCredentials } from "../index.ts";
2020 import { COMMANDS } from "../input.ts";
2121 import {
22+ computeContextTokens,
2223 computeStats,
2324 forkSession,
2425 listPromptHistory,
@@ -387,6 +388,7 @@ export function createCommandController(
387388 state.session = picked;
388389 state.messages = loadMessages(state.db, picked.id);
389390 state.stats = computeStats(state.messages);
391+ state.contextTokens = computeContextTokens(state.messages);
390392 runtime.scrollConversationToBottom();
391393 }
392394 }
@@ -402,6 +404,7 @@ export function createCommandController(
402404 state.session = null;
403405 state.messages = [];
404406 state.stats = { totalInput: 0, totalOutput: 0, totalCost: 0 };
407+ state.contextTokens = 0;
405408 await runtime.reloadPromptContext(state);
406409 runtime.scrollConversationToBottom();
407410 runtime.render();
@@ -415,6 +418,7 @@ export function createCommandController(
415418 state.session = forked;
416419 state.messages = loadMessages(state.db, forked.id);
417420 state.stats = computeStats(state.messages);
421+ state.contextTokens = computeContextTokens(state.messages);
418422 runtime.appendInfoMessage("Forked session.", state);
419423 };
420424
@@ -434,6 +438,7 @@ export function createCommandController(
434438 if (removed) {
435439 state.messages = loadMessages(state.db, state.session.id);
436440 state.stats = computeStats(state.messages);
441+ state.contextTokens = computeContextTokens(state.messages);
437442 runtime.scrollConversationToBottom();
438443 runtime.render();
439444 }
src/ui/status.test.ts+58−24
@@ -4,7 +4,11 @@ import {
44 registerFauxProvider,
55 } from "@mariozechner/pi-ai";
66 import type { AppState } from "../index.ts";
7-import { createUiMessage, openDatabase } from "../session.ts";
7+import {
8+ computeContextTokens,
9+ createUiMessage,
10+ openDatabase,
11+} from "../session.ts";
812 import { DEFAULT_SHOW_REASONING, DEFAULT_VERBOSE } from "../settings.ts";
913 import { DEFAULT_THEME } from "../theme.ts";
1014 import { renderStatusBar } from "./status.ts";
@@ -87,6 +91,7 @@ function createTestState(): AppState {
8791 effort: "medium",
8892 messages: [],
8993 stats: { totalInput: 0, totalOutput: 0, totalCost: 0 },
94+ contextTokens: 0,
9095 agentsMd: [],
9196 skills: [],
9297 plugins: [],
@@ -109,6 +114,11 @@ function createTestState(): AppState {
109114 };
110115 }
111116
117+function setMessages(state: AppState, messages: AppState["messages"]): void {
118+ state.messages = messages;
119+ state.contextTokens = computeContextTokens(messages);
120+}
121+
112122 function makeAssistantWithUsage(
113123 text: string,
114124 usage: Partial<ReturnType<typeof fauxAssistantMessage>["usage"]>,
@@ -217,11 +227,11 @@ describe("ui/status", () => {
217227 ] as const;
218228
219229 for (const testCase of cases) {
220- state.messages = [
230+ setMessages(state, [
221231 makeAssistantWithUsage("context anchor", {
222232 totalTokens: testCase.totalTokens,
223233 }),
224- ];
234+ ]);
225235 expect(getUsagePill(state)).toMatchObject({
226236 text: `in:0 out:0 · ${testCase.totalTokens.toFixed(1)}%/100 · $0.00`,
227237 bgColor: testCase.bgColor,
@@ -234,6 +244,24 @@ describe("ui/status", () => {
234244 }
235245 });
236246
247+ test("renderStatusBar uses cached contextTokens instead of rescanning messages", () => {
248+ const faux = registerFauxProvider();
249+ const state = createTestState();
250+ state.model = {
251+ ...faux.getModel(),
252+ contextWindow: 100,
253+ };
254+ state.messages = [{ role: "user", content: "x".repeat(400), timestamp: 1 }];
255+ state.contextTokens = 25;
256+
257+ try {
258+ expect(getUsageSummary(state)).toBe("in:0 out:0 · 25.0%/100 · $0.00");
259+ } finally {
260+ faux.unregister();
261+ state.db.close();
262+ }
263+ });
264+
237265 test("renderStatusBar colors the model and usage pills independently", () => {
238266 const faux = registerFauxProvider();
239267 const state = createTestState();
@@ -244,9 +272,9 @@ describe("ui/status", () => {
244272
245273 try {
246274 state.effort = "xhigh";
247- state.messages = [
275+ setMessages(state, [
248276 makeAssistantWithUsage("cold context", { totalTokens: 10 }),
249- ];
277+ ]);
250278 expect(getModelPill(state)).toMatchObject({
251279 bgColor: "color09",
252280 fgColor: "color00",
@@ -257,9 +285,9 @@ describe("ui/status", () => {
257285 });
258286
259287 state.effort = "low";
260- state.messages = [
288+ setMessages(state, [
261289 makeAssistantWithUsage("hot context", { totalTokens: 95 }),
262- ];
290+ ]);
263291 expect(getModelPill(state)).toMatchObject({
264292 bgColor: "color02",
265293 fgColor: "color00",
@@ -288,26 +316,29 @@ describe("ui/status", () => {
288316 };
289317
290318 try {
291- state.messages = [
319+ setMessages(state, [
292320 { role: "user", content: "first", timestamp: 1 },
293321 makeAssistantWithUsage("Second.", {
294322 input: 200,
295323 output: 50,
296324 totalTokens: 250,
297325 }),
298- ];
326+ ]);
299327 const anchoredPercent = getUsagePercent(state);
300328
301- state.messages = [...state.messages, createUiMessage("x".repeat(5_000))];
329+ setMessages(state, [
330+ ...state.messages,
331+ createUiMessage("x".repeat(5_000)),
332+ ]);
302333 const withUiOnlyPercent = getUsagePercent(state);
303334
304- state.messages = [
335+ setMessages(state, [
305336 ...state.messages,
306337 { role: "user", content: "12345678", timestamp: 2 },
307- ];
338+ ]);
308339 const withTrailingUserPercent = getUsagePercent(state);
309340
310- state.messages = [
341+ setMessages(state, [
311342 ...state.messages,
312343 {
313344 role: "toolResult",
@@ -317,7 +348,7 @@ describe("ui/status", () => {
317348 isError: false,
318349 timestamp: 3,
319350 },
320- ];
351+ ]);
321352 const withToolResultPercent = getUsagePercent(state);
322353
323354 expect(getUsageSummary(state)).toBe(
@@ -341,16 +372,19 @@ describe("ui/status", () => {
341372 };
342373
343374 try {
344- state.messages = [{ role: "user", content: "12345678", timestamp: 1 }];
375+ setMessages(state, [{ role: "user", content: "12345678", timestamp: 1 }]);
345376 const initialPercent = getUsagePercent(state);
346377
347- state.messages = [createUiMessage("x".repeat(5_000)), ...state.messages];
378+ setMessages(state, [
379+ createUiMessage("x".repeat(5_000)),
380+ ...state.messages,
381+ ]);
348382 const withUiOnlyPercent = getUsagePercent(state);
349383
350- state.messages = [
384+ setMessages(state, [
351385 ...state.messages,
352386 { role: "user", content: "more visible text", timestamp: 2 },
353- ];
387+ ]);
354388 const withSecondUserPercent = getUsagePercent(state);
355389
356390 expect(initialPercent).toBeGreaterThan(0);
@@ -383,7 +417,7 @@ describe("ui/status", () => {
383417 ...faux.getModel(),
384418 contextWindow: 1_000,
385419 };
386- state.messages = [
420+ setMessages(state, [
387421 { role: "user", content: "first", timestamp: 1 },
388422 makeAssistantWithUsage("valid", {
389423 input: 150,
@@ -392,7 +426,7 @@ describe("ui/status", () => {
392426 }),
393427 aborted,
394428 { role: "user", content: "1234", timestamp: 2 },
395- ];
429+ ]);
396430 return state;
397431 };
398432
@@ -422,7 +456,7 @@ describe("ui/status", () => {
422456 ...faux.getModel(),
423457 contextWindow: 1_000,
424458 };
425- fromComponents.messages = [
459+ setMessages(fromComponents, [
426460 makeAssistantWithUsage("fallback", {
427461 input: 120,
428462 output: 30,
@@ -430,8 +464,8 @@ describe("ui/status", () => {
430464 cacheWrite: 25,
431465 totalTokens: 0,
432466 }),
433- ];
434- fromTotalTokens.messages = [
467+ ]);
468+ setMessages(fromTotalTokens, [
435469 makeAssistantWithUsage("fallback", {
436470 input: 120,
437471 output: 30,
@@ -439,7 +473,7 @@ describe("ui/status", () => {
439473 cacheWrite: 25,
440474 totalTokens: 200,
441475 }),
442- ];
476+ ]);
443477
444478 try {
445479 expect(getUsagePercent(fromComponents)).toBe(
src/ui/status.ts+2−139
@@ -2,8 +2,7 @@
22 * Status-bar formatting and rendering for the terminal UI.
33 *
44 * Computes the cumulative usage pill, abbreviates the working directory, and
5- * estimates current model-visible context usage from persisted conversation
6- * history.
5+ * formats the cached current-context estimate maintained on application state.
76 *
87 * @module
98 */
@@ -12,14 +11,9 @@ import { homedir } from "node:os";
1211 import { Spacer } from "@cel-tui/components";
1312 import { HStack, Text, visibleWidth } from "@cel-tui/core";
1413 import type { Node } from "@cel-tui/types";
15-import type { AssistantMessage, Message } from "@mariozechner/pi-ai";
1614 import type { AppState } from "../index.ts";
17-import { filterModelMessages, getAssistantUsage } from "../session.ts";
1815 import type { StatusTone, Theme } from "../theme.ts";
1916
20-/** Conservative fixed estimate for an image block's token footprint. */
21-const ESTIMATED_IMAGE_TOKENS = 1_200;
22-
2317 /**
2418 * Abbreviate a path with `~` for the home directory.
2519 *
@@ -84,143 +78,12 @@ function formatModelInfo(state: AppState): string {
8478 return `${state.model.provider}/${state.model.id} · ${formatEffort(state.effort)}`;
8579 }
8680
87-/** Calculate context tokens from assistant usage, falling back when `totalTokens` is zero. */
88-function calculateUsageTokens(usage: AssistantMessage["usage"]): number {
89- return (
90- usage.totalTokens ||
91- usage.input + usage.output + usage.cacheRead + usage.cacheWrite
92- );
93-}
94-
95-/** Estimate token usage from a character count using a conservative chars/4 heuristic. */
96-function estimateCharacterTokens(charCount: number): number {
97- return Math.ceil(charCount / 4);
98-}
99-
100-type UserMultipartContent = Exclude<
101- Extract<Message, { role: "user" }>["content"],
102- string
103->;
104-type TextOrImageContentBlock =
105- | UserMultipartContent[number]
106- | Extract<Message, { role: "toolResult" }>["content"][number];
107-
108-function estimateTextOrImageContentTokens(
109- content: readonly TextOrImageContentBlock[],
110-): number {
111- let chars = 0;
112- let imageTokens = 0;
113-
114- for (const block of content) {
115- if (block.type === "text") {
116- chars += block.text.length;
117- continue;
118- }
119- if (block.type === "image") {
120- imageTokens += ESTIMATED_IMAGE_TOKENS;
121- }
122- }
123-
124- return estimateCharacterTokens(chars) + imageTokens;
125-}
126-
127-function estimateUserMessageTokens(
128- message: Extract<Message, { role: "user" }>,
129-): number {
130- if (typeof message.content === "string") {
131- return estimateCharacterTokens(message.content.length);
132- }
133- return estimateTextOrImageContentTokens(message.content);
134-}
135-
136-function estimateAssistantBlockCharacters(
137- block: Extract<Message, { role: "assistant" }>["content"][number],
138-): number {
139- if (block.type === "text") {
140- return block.text.length;
141- }
142- if (block.type === "thinking") {
143- return block.thinking.length;
144- }
145- return block.name.length + JSON.stringify(block.arguments).length;
146-}
147-
148-function estimateAssistantMessageTokens(
149- message: Extract<Message, { role: "assistant" }>,
150-): number {
151- const chars = message.content.reduce((total, block) => {
152- return total + estimateAssistantBlockCharacters(block);
153- }, 0);
154- return estimateCharacterTokens(chars);
155-}
156-
157-function estimateToolResultMessageTokens(
158- message: Extract<Message, { role: "toolResult" }>,
159-): number {
160- return estimateTextOrImageContentTokens(message.content);
161-}
162-
163-/** Estimate token usage for a model-visible message. */
164-function estimateMessageTokens(message: Message): number {
165- switch (message.role) {
166- case "user":
167- return estimateUserMessageTokens(message);
168- case "assistant":
169- return estimateAssistantMessageTokens(message);
170- case "toolResult":
171- return estimateToolResultMessageTokens(message);
172- }
173-}
174-
175-/** Find the latest assistant usage that can anchor context estimation. */
176-function getLatestValidAssistantUsage(
177- messages: readonly Message[],
178-): { index: number; tokens: number } | null {
179- for (let i = messages.length - 1; i >= 0; i--) {
180- const message = messages[i];
181- if (
182- message?.role === "assistant" &&
183- message.stopReason !== "aborted" &&
184- message.stopReason !== "error"
185- ) {
186- const usage = getAssistantUsage(message);
187- if (!usage) {
188- continue;
189- }
190- return {
191- index: i,
192- tokens: calculateUsageTokens(usage),
193- };
194- }
195- }
196- return null;
197-}
198-
199-/** Estimate the current model-visible context size for the next request. */
200-function estimateCurrentContextTokens(state: AppState): number {
201- const messages = filterModelMessages(state.messages);
202- const latestUsage = getLatestValidAssistantUsage(messages);
203-
204- if (!latestUsage) {
205- return messages.reduce((total, message) => {
206- return total + estimateMessageTokens(message);
207- }, 0);
208- }
209-
210- let total = latestUsage.tokens;
211- for (let i = latestUsage.index + 1; i < messages.length; i++) {
212- total += estimateMessageTokens(messages[i]!);
213- }
214- return total;
215-}
216-
21781 /** Estimate current context usage as a percentage of the active model window. */
21882 function getContextPercentage(state: AppState): number {
21983 if (!state.model || state.model.contextWindow <= 0) {
22084 return 0;
22185 }
222- const contextTokens = estimateCurrentContextTokens(state);
223- return (contextTokens / state.model.contextWindow) * 100;
86+ return (state.contextTokens / state.model.contextWindow) * 100;
22487 }
22588
22689 /** Format cumulative session totals plus estimated current context usage for the status bar. */