Use 4096-token release defaults

public/manifest/models.json

@@ -218,7 +218,7 @@
       ],
       "downloadBytes": 3803453440,
       "contextLength": 262144,
-      "defaultContext": 4000,
+      "defaultContext": 4096,
       "cpuFallback": false,
       "largestTensorBytes": 178790400,
       "requiredLimits": {

src/app/App.test.ts

@@ -1,7 +1,9 @@
 import { describe, expect, it } from 'vitest';
+import releaseManifest from '../../public/manifest/models.json';
 import { EngineClientError } from '../engine';
 import {
   DEFAULT_CONTEXT_SIZE,
+  DEFAULT_MAX_TOKENS,
   DEFAULT_MODEL_ID,
   modelArchitectureLabel,
   publishBackendReport,
@@ -39,8 +41,12 @@ describe('App release defaults', () => {
     expect(modelArchitectureLabel('qwen35')).toBe('Qwen3.5 hybrid');
   });
 
-  it('reserves enough context for a 2K completion plus prompts and tool schemas', () => {
-    expect(DEFAULT_CONTEXT_SIZE).toBe(4000);
+  it('uses one 4K ceiling for context allocation and completion requests', () => {
+    expect(DEFAULT_CONTEXT_SIZE).toBe(4096);
+    expect(DEFAULT_MAX_TOKENS).toBe(4096);
+    expect(DEFAULT_CONTEXT_SIZE).toBe(DEFAULT_MAX_TOKENS);
+    expect(releaseManifest.models.find((model) => model.id === '27b')?.defaultContext)
+      .toBe(DEFAULT_CONTEXT_SIZE);
   });
 });
 

src/app/App.tsx

@@ -32,6 +32,7 @@ import type {
   ToolRun,
 } from '../lib/contracts';
 import { formatBytes } from '../lib/format';
+import { DEFAULT_CONTEXT_SIZE, DEFAULT_MAX_TOKENS } from '../lib/runtime-defaults';
 import {
   conversationTitle,
   normalizeConversationModelId,
@@ -49,7 +50,7 @@ const MAX_TOOL_ROUNDS = 5;
 const MAX_TOOL_CALLS_PER_ROUND = 4;
 const MAX_TOOL_CALLS_TOTAL = 12;
 export const DEFAULT_MODEL_ID: ModelId = 'bonsai-27b';
-export const DEFAULT_CONTEXT_SIZE = 4000;
+export { DEFAULT_CONTEXT_SIZE, DEFAULT_MAX_TOKENS } from '../lib/runtime-defaults';
 
 const MODEL_DECOR: Record<ModelTierId, {
   id: ModelId;
@@ -87,7 +88,7 @@ const INITIAL_SETTINGS: RuntimeSettings = {
   backend: 'auto',
   contextSize: DEFAULT_CONTEXT_SIZE,
   temperature: 0.7,
-  maxTokens: 2000,
+  maxTokens: DEFAULT_MAX_TOKENS,
   systemPrompt: 'You are Bonsai, a private browser-local assistant. Be concise and evidence-led. Use a client tool only when it materially improves the answer, and never claim a tool succeeded before its response arrives.',
 };
 
@@ -932,7 +933,7 @@ export function App() {
         const calls = result.toolCalls.length > 0 ? result.toolCalls : fallbackToolCalls(thought.content);
         if (toolPlan.requiredToolName && calls.length === 0) {
           outputError = (result.toolCallError || result.finishReason === 'length')
-            ? `The ${toolPlan.requiredToolName} call ended before its arguments were complete. Regenerate, shorten the request, or clear earlier turns so the 2,000-token completion budget fits the context.`
+            ? `The ${toolPlan.requiredToolName} call ended before its arguments were complete. Regenerate, shorten the request, or clear earlier turns so the ${DEFAULT_MAX_TOKENS.toLocaleString()}-token completion ceiling fits the context.`
             : `The model did not call the required ${toolPlan.requiredToolName} tool for this artifact request.`;
         }
         if (toolPlan.requiredToolName && calls.some((call) => call.function.name !== toolPlan.requiredToolName)) {

src/app/components/SettingsSheet.tsx

@@ -1,5 +1,6 @@
 import { useEffect, useRef } from 'react';
 import type { ModelOption, RuntimeSettings, TelemetrySnapshot } from '../../lib/contracts';
+import { DEFAULT_MAX_TOKENS } from '../../lib/runtime-defaults';
 import type { ThemeMode } from '../theme';
 
 interface SettingsSheetProps {
@@ -103,8 +104,8 @@ export function SettingsSheet({
 
         <label className="range-setting">
           <span><small>Maximum completion</small><strong>{settings.maxTokens} tokens</strong></span>
-          <input type="range" min="64" max="2000" step="16" value={settings.maxTokens} onChange={(event) => update('maxTokens', Number(event.target.value))} />
-          <span className="range-limits"><i>64</i><i>2,000</i></span>
+          <input type="range" min="64" max={DEFAULT_MAX_TOKENS} step="16" value={settings.maxTokens} onChange={(event) => update('maxTokens', Number(event.target.value))} />
+          <span className="range-limits"><i>64</i><i>{DEFAULT_MAX_TOKENS.toLocaleString()}</i></span>
         </label>
 
         <label className="system-prompt-setting">

src/engine/device-gate.test.ts

@@ -193,13 +193,13 @@ describe('evaluateModelContextPolicy', () => {
     ['f16 KV', { ...longContextExperiment, kvCacheType: 'f16' as const }],
     ['auto backend', { ...longContextExperiment, requestedBackend: 'auto' as const }],
     ['release defaults', { ...longContextExperiment, tuningScope: 'release-defaults' as const }],
-  ])('keeps 27B at the 4,000-token release limit for %s', (_label, input) => {
+  ])('keeps 27B at the 4,096-token release limit for %s', (_label, input) => {
     expect(evaluateModelContextPolicy(model27b, 8_448, input)).toEqual({
       allowed: false,
-      limit: 4_000,
+      limit: 4_096,
     });
-    expect(evaluateModelContextPolicy(model27b, 4_000, input).allowed).toBe(true);
-    expect(evaluateModelContextPolicy(model27b, 4_001, input).allowed).toBe(false);
+    expect(evaluateModelContextPolicy(model27b, 4_096, input).allowed).toBe(true);
+    expect(evaluateModelContextPolicy(model27b, 4_097, input).allowed).toBe(false);
   });
 
   it('leaves dense-tier manifest limits unchanged', () => {

src/engine/device-gate.ts

@@ -1,4 +1,5 @@
 import type { ManifestModelV2 } from './manifest';
+import { DEFAULT_CONTEXT_SIZE } from '../lib/runtime-defaults';
 import type {
   BenchmarkFlashMode,
   BenchmarkKvCacheType,
@@ -10,7 +11,7 @@ import type {
 } from './protocol';
 import type { BackendReport } from './native-log';
 
-export const MODEL_27B_RELEASE_CONTEXT_LIMIT = 4_000;
+export const MODEL_27B_RELEASE_CONTEXT_LIMIT = DEFAULT_CONTEXT_SIZE;
 export const MODEL_27B_BENCHMARK_CONTEXT_LIMIT = 8_448;
 
 export interface ModelContextPolicyInput {

src/engine/runtime.ts

@@ -8,6 +8,7 @@ import {
   type ResultTimings,
 } from '../../vendor/wllama-bonsai/esm/index.js';
 import runtimeSource from '../../vendor/wllama-bonsai/SOURCE.json';
+import { DEFAULT_MAX_TOKENS } from '../lib/runtime-defaults';
 import {
   assertBackendPolicy,
   estimateStorage,
@@ -1357,7 +1358,7 @@ export class BrowserEngineRuntime {
     try {
       const completionOptions = {
         messages: params.messages as ChatCompletionMessage[],
-        max_tokens: params.maxTokens ?? 2_000,
+        max_tokens: params.maxTokens ?? DEFAULT_MAX_TOKENS,
         temperature: params.temperature ?? 0,
         top_p: params.topP,
         top_k: params.topK ?? 1,

src/lib/runtime-defaults.ts

@@ -0,0 +1,2 @@
+export const DEFAULT_CONTEXT_SIZE = 4_096;
+export const DEFAULT_MAX_TOKENS = 4_096;

src/lib/tool-routing.test.ts

@@ -29,7 +29,7 @@ describe('HTML artifact intent routing', () => {
     expect(plan.tools?.map((tool) => tool.function.name)).toEqual(['html_artifact']);
     expect(plan.requiredToolName).toBe('html_artifact');
     expect(plan.minimumMaxTokens).toBe(HTML_ARTIFACT_MIN_TOKENS);
-    expect(plan.minimumMaxTokens).toBe(2_000);
+    expect(plan.minimumMaxTokens).toBe(4_096);
     expect(plan.systemInstruction).toContain('compact enough to finish');
     expect(plan.systemInstruction).toContain('Do not answer with a Markdown code block');
   });

src/lib/tool-routing.ts

@@ -1,7 +1,8 @@
 import type { EngineChatTool } from '../engine';
 import { AGENT_TOOLS, HTML_ARTIFACT_TOOL } from './agent-tools';
+import { DEFAULT_MAX_TOKENS } from './runtime-defaults';
 
-export const HTML_ARTIFACT_MIN_TOKENS = 2000;
+export const HTML_ARTIFACT_MIN_TOKENS = DEFAULT_MAX_TOKENS;
 
 const CREATION_VERB = /(?:\b(?:build|code|create|design|develop|generate|implement|make|write)\b|(?:созда(?:й|йте)|сдела(?:й|йте)|напиш(?:и|ите)|сгенерир(?:уй|уйте)|сверста(?:й|йте)|реализ(?:уй|уйте)|разработа(?:й|йте)|собер(?:и|ите)|постро(?:й|йте)))/iu;
 const HTML_PRODUCT_TARGET = /(?:\bweb[\s-]?(?:app|application|page|site)\b|\bwebsite\b|\blanding[\s-]?page\b|(?:веб|html)[\s-]?(?:приложени\w*|страниц\w*|сайт\w*|интерфейс\w*)|(?:^|[\s,.:;!?])(?:сайт\w*|лендинг\w*))/iu;