Add DeepSeek context budget diagnostics

apps/desktop/src/main/main.ts

@@ -142,6 +142,7 @@ import {
  testConnection,
 } from '@maka/runtime';
 import type { BotIncomingMessage, ToolArtifactRecorderInput } from '@maka/runtime';
+import type { ContextBudgetPolicy } from '@maka/runtime';
 import { testProxyConnection } from '@maka/runtime/network/proxy-test';
 import { fetchWeChatQrcode, pollWeChatQrcodeStatus } from './wechat-scan-login.js';
 import {
@@ -735,6 +736,7 @@ backends.register('ai-sdk', async (ctx) => {
  modelFactory: (input) => getAIModel({ ...input, fetch: modelFetch }),
  tools: builtinTools,
  providerOptions: buildProviderOptions(connection, model),
+ contextBudget: buildContextBudgetPolicy(connection),
  systemPrompt: ({ cwd }) => buildSystemPrompt(ctx.header, cwd),
  turnTailPrompt: ({ cwd }) => buildTurnTailPrompt(cwd),
  recordLlmCall: (event) => recordLlmCall({ repo: telemetryRepo, lookupPricing }, event),
@@ -756,6 +758,38 @@ backends.register('ai-sdk', async (ctx) => {
  });
 });
 
+function buildContextBudgetPolicy(connection: LlmConnection): ContextBudgetPolicy | undefined {
+ if (process.env.MAKA_CONTEXT_BUDGET === 'off') return undefined;
+ const maxHistoryEstimatedTokens =
+ parseOptionalPositiveInt(process.env.MAKA_CONTEXT_HISTORY_BUDGET_TOKENS) ??
+ defaultHistoryBudgetTokens(connection);
+ if (maxHistoryEstimatedTokens === undefined) return undefined;
+ const maxHistoryTurns = parseOptionalPositiveInt(process.env.MAKA_CONTEXT_HISTORY_BUDGET_TURNS);
+ const minRecentTurns = parsePositiveInt(process.env.MAKA_CONTEXT_MIN_RECENT_TURNS, 2);
+ return {
+ name: 'desktop-default-history-budget',
+ maxHistoryEstimatedTokens,
+ ...(maxHistoryTurns !== undefined ? { maxHistoryTurns } : {}),
+ minRecentTurns,
+ };
+}
+
+function defaultHistoryBudgetTokens(connection: LlmConnection): number | undefined {
+ if (connection.providerType === 'deepseek') return undefined;
+ return 32_000;
+}
+
+function parsePositiveInt(value: string | undefined, fallback: number): number {
+ const parsed = parseOptionalPositiveInt(value);
+ return parsed ?? fallback;
+}
+
+function parseOptionalPositiveInt(value: string | undefined): number | undefined {
+ if (!value) return undefined;
+ const parsed = Number.parseInt(value, 10);
+ return Number.isFinite(parsed) && parsed > 0 ? parsed : undefined;
+}
+
 function buildSubscriptionModelFetch(
  connection: LlmConnection,
  sessionId: string,

package.json

@@ -19,7 +19,8 @@
  "check:stale": "node scripts/check-stale-dist.mjs",
  "prepare:officecli": "node scripts/prepare-officecli.mjs",
  "check:officecli-bundle": "node scripts/check-officecli-bundle.mjs",
- "check:release": "npm run check:stale && npm run check:officecli-bundle"
+ "check:release": "npm run check:stale && npm run check:officecli-bundle",
+ "cost:deepseek-baseline": "node scripts/deepseek-live-cost-baseline.mjs"
  },
  "devDependencies": {
  "@types/node": "^25.0.0",

packages/core/src/events.ts

@@ -10,7 +10,11 @@
  */
 
 import type { PermissionRequest, PermissionResponse, ToolCategory } from './permission.js';
-import type { PrefixChangeReason } from './usage-stats/types.js';
+import type {
+ ContextBudgetDiagnostic,
+ PrefixChangeReason,
+ PromptSegmentEstimate,
+} from './usage-stats/types.js';
 
 export const TOOL_OUTPUT_STREAMS = ['stdout', 'stderr'] as const;
 export const TOOL_OUTPUT_DELTA_MAX_CHARS = 8192;
@@ -335,6 +339,8 @@ export interface TokenUsageEvent extends BaseEvent {
  contextRemaining?: number;
  prefixHash?: string;
  prefixChangeReason?: PrefixChangeReason;
+ promptSegments?: PromptSegmentEstimate[];
+ contextBudget?: ContextBudgetDiagnostic;
 }
 
 export interface ErrorEvent extends BaseEvent {

packages/core/src/index.ts

@@ -693,7 +693,10 @@ export {
 // usage-stats/types.ts
 export type {
  LlmCallRecord,
+ ContextBudgetDiagnostic,
  PricingConfig,
+ PromptSegmentEstimate,
+ PromptSegmentKind,
  TimeRange,
  ToolInvocationRecord,
  UsageBucket,

packages/core/src/runtime-event.ts

@@ -16,7 +16,11 @@
 
 import type { AttachmentRef } from './events.js';
 import type { PermissionRequest, PermissionResponse } from './permission.js';
-import type { PrefixChangeReason } from './usage-stats/types.js';
+import type {
+ ContextBudgetDiagnostic,
+ PrefixChangeReason,
+ PromptSegmentEstimate,
+} from './usage-stats/types.js';
 
 // ============================================================================
 // Role / Author / Status
@@ -180,6 +184,8 @@ export interface RuntimeEventTokenUsage {
  contextRemaining?: number;
  prefixHash?: string;
  prefixChangeReason?: PrefixChangeReason;
+ promptSegments?: PromptSegmentEstimate[];
+ contextBudget?: ContextBudgetDiagnostic;
 }
 
 /**

packages/core/src/session.ts

@@ -11,7 +11,11 @@
 
 import type { AttachmentRef, ToolResultContent } from './events.js';
 import type { PermissionMode } from './permission.js';
-import type { PrefixChangeReason } from './usage-stats/types.js';
+import type {
+ ContextBudgetDiagnostic,
+ PrefixChangeReason,
+ PromptSegmentEstimate,
+} from './usage-stats/types.js';
 
 export const SESSION_STATUSES = [
  'active',
@@ -243,6 +247,8 @@ export interface TokenUsageMessage {
  costUsd?: number;
  prefixHash?: string;
  prefixChangeReason?: PrefixChangeReason;
+ promptSegments?: PromptSegmentEstimate[];
+ contextBudget?: ContextBudgetDiagnostic;
 }
 
 export interface TurnStateMessage {

packages/core/src/usage-stats/types.ts

@@ -72,6 +72,8 @@ export interface UsageLogRow {
  turnId?: string;
  prefixHash?: string;
  prefixChangeReason?: PrefixChangeReason;
+ promptSegments?: PromptSegmentEstimate[];
+ contextBudget?: ContextBudgetDiagnostic;
 }
 
 export interface PricingConfig {
@@ -117,6 +119,8 @@ export interface LlmCallRecord {
  startedAt: number;
  prefixHash?: string;
  prefixChangeReason?: PrefixChangeReason;
+ promptSegments?: PromptSegmentEstimate[];
+ contextBudget?: ContextBudgetDiagnostic;
 }
 
 export type PrefixChangeReason =
@@ -129,6 +133,35 @@ export type PrefixChangeReason =
  | 'stable'
  | 'unknown';
 
+export type PromptSegmentKind =
+ | 'system_prompt'
+ | 'tool_schema'
+ | 'prior_history'
+ | 'current_user'
+ | 'turn_tail';
+
+export interface PromptSegmentEstimate {
+ kind: PromptSegmentKind;
+ chars: number;
+ estimatedTokens: number;
+ messageCount?: number;
+ eventCount?: number;
+ toolCount?: number;
+}
+
+export interface ContextBudgetDiagnostic {
+ enabled: boolean;
+ policyName?: string;
+ maxHistoryEstimatedTokens?: number;
+ maxHistoryTurns?: number;
+ estimatedTokensBefore: number;
+ estimatedTokensAfter: number;
+ keptTurns: number;
+ droppedTurns: number;
+ keptEvents: number;
+ droppedEvents: number;
+}
+
 export interface ToolInvocationRecord {
  sessionId?: string;
  turnId?: string;

packages/runtime/package.json

@@ -15,6 +15,7 @@
  "./tool-output-delta": "./dist/tool-output-delta.js",
  "./stream-watchdog": "./dist/stream-watchdog.js",
  "./model-factory": "./dist/model-factory.js",
+ "./context-budget": "./dist/context-budget.js",
  "./test-connection": "./dist/test-connection.js",
  "./model-fetcher": "./dist/model-fetcher.js",
  "./materializer": "./dist/materializer.js",

packages/runtime/src/__tests__/ai-sdk-backend.test.ts

@@ -24,6 +24,7 @@ import {
  canonicalizeToolSet,
  computeRequestShapeDiagnostic,
 } from '../request-shape.js';
+import { applyRuntimeEventContextBudget } from '../context-budget.js';
 
 describe('AiSdkBackend model history', () => {
  test('prefers RuntimeEvent prior messages and appends current user once', async () => {
@@ -800,6 +801,86 @@ describe('AiSdkBackend request-shape diagnostics', () => {
  });
 });
 
+describe('AiSdkBackend context budget and prompt attribution', () => {
+ test('context budget keeps whole recent turns and drops older turns', () => {
+ const events = [
+ runtimeTextEvent({ id: 'old-u', turnId: 'old', role: 'user', author: 'user', text: 'old user text' }),
+ runtimeTextEvent({ id: 'old-a', turnId: 'old', role: 'model', author: 'agent', text: 'old assistant text' }),
+ runtimeTextEvent({ id: 'new-u', turnId: 'new', role: 'user', author: 'user', text: 'new user text' }),
+ runtimeTextEvent({ id: 'new-a', turnId: 'new', role: 'model', author: 'agent', text: 'new assistant text' }),
+ ];
+
+ const budgeted = applyRuntimeEventContextBudget(events, {
+ name: 'test-budget',
+ maxHistoryEstimatedTokens: 1,
+ minRecentTurns: 1,
+ charsPerToken: 1,
+ });
+
+ assert.ok(budgeted);
+ assert.deepEqual([...new Set(budgeted.events.map((event) => event.turnId))], ['new']);
+ assert.equal(budgeted.diagnostic.droppedTurns, 1);
+ assert.equal(budgeted.diagnostic.keptTurns, 1);
+ assert.equal(budgeted.diagnostic.droppedEvents, 2);
+ });
+
+ test('usage events include prompt segments and context budget diagnostics', async () => {
+ const model = completionModel();
+ const events: SessionEvent[] = [];
+ const backend = new AiSdkBackend({
+ sessionId: 'session-1',
+ header: header(),
+ appendMessage: async () => {},
+ connection: connection(),
+ apiKey: 'sk-test',
+ modelId: 'mock-model-id',
+ permissionEngine: new PermissionEngine({ newId: () => 'permission-id', now: () => 1 }),
+ modelFactory: () => model,
+ tools: [testTool('Read', z.object({ path: z.string() }))],
+ newId: idGenerator(),
+ now: monotonicClock(),
+ systemPrompt: 'durable system',
+ turnTailPrompt: 'volatile tail',
+ contextBudget: {
+ name: 'test-budget',
+ maxHistoryEstimatedTokens: 1,
+ minRecentTurns: 1,
+ charsPerToken: 1,
+ },
+ });
+
+ for await (const event of backend.send({
+ turnId: 'turn-current',
+ text: 'current user',
+ context: [],
+ runtimeContext: [
+ runtimeTextEvent({ id: 'old-u', turnId: 'old', role: 'user', author: 'user', text: 'old user text' }),
+ runtimeTextEvent({ id: 'old-a', turnId: 'old', role: 'model', author: 'agent', text: 'old assistant text' }),
+ runtimeTextEvent({ id: 'new-u', turnId: 'new', role: 'user', author: 'user', text: 'new user text' }),
+ runtimeTextEvent({ id: 'new-a', turnId: 'new', role: 'model', author: 'agent', text: 'new assistant text' }),
+ ],
+ })) {
+ events.push(event);
+ }
+
+ assert.deepEqual(compactPrompt(model), [
+ { role: 'system', content: 'durable system' },
+ { role: 'user', content: [{ type: 'text', text: 'new user text' }] },
+ { role: 'assistant', content: [{ type: 'text', text: 'new assistant text' }] },
+ { role: 'user', content: [{ type: 'text', text: 'current user\n\nvolatile tail' }] },
+ ]);
+ const usage = events.find((event): event is Extract<SessionEvent, { type: 'token_usage' }> =>
+ event.type === 'token_usage'
+ );
+ assert.ok(usage);
+ assert.equal(usage.contextBudget?.policyName, 'test-budget');
+ assert.equal(usage.contextBudget?.droppedTurns, 1);
+ assert.equal(usage.promptSegments?.some((segment) => segment.kind === 'prior_history'), true);
+ assert.equal(usage.promptSegments?.some((segment) => segment.kind === 'tool_schema'), true);
+ assert.equal(usage.promptSegments?.some((segment) => segment.kind === 'turn_tail'), true);
+ });
+});
+
 describe('AiSdkBackend RunTrace', () => {
  test('records turn, model, usage, and completion trace events without changing SessionEvents', async () => {
  const trace: RunTraceEvent[] = [];