Qwen3.8-Flash-Next on any consumer hardware: one-click install for Windows / Linux. Strata inference engine, OpenAI/Anthropic API on localhost, optional image input.
C++
agent: output budget for a local model is the window (was 8192), the Reasoning menu reaches the request, fit_max_tokens for app-loaded models, plain outcome for silent commands
Niko1221be1a871349df
diff --git a/apps/host/server/agent/agent-bridge.mjs b/apps/host/server/agent/agent-bridge.mjsindex 6e423d756..6be1fcd26 100644--- a/apps/host/server/agent/agent-bridge.mjs+++ b/apps/host/server/agent/agent-bridge.mjs@@ -19,7 +19,7 @@ import { join } from 'node:path';import { randomUUID } from 'node:crypto';import { GatewayClient } from './gateway-client.mjs';import { HarnessProcess, resolveHarness, redact, stopStaleAgent } from './harness-runtime.mjs';-import { buildProfile, routeFor, MODE_PRESETS, modeOf } from './profile.mjs';+import { buildProfile, routeFor, MODE_PRESETS, modeOf, effortOf } from './profile.mjs';import { TurnProjector, parseArguments as parseArgs } from './events.mjs';import { ChildProjector } from './children.mjs';@@ -238,7 +238,7 @@ export function createAgentBridge({ dataDir, componentsDir = null, bundledDir =try { return await runTurnInner(c, options); } finally { c.starting = false; }}- async function runTurnInner(c, { chatId, message, workspacePath, choice = null, mode = null, history = [], onFrame, requestId = `req-${randomUUID()}`, signal, goal = null }) {+ async function runTurnInner(c, { chatId, message, workspacePath, choice = null, mode = null, history = [], onFrame, requestId = `req-${randomUUID()}`, signal, goal = null, thinking = null }) {let send = onFrame;signal?.addEventListener('abort', () => { send = () => {}; }, { once: true });const emit = frame => { try { send(frame); } catch { send = () => {}; } };@@ -249,14 +249,19 @@ export function createAgentBridge({ dataDir, componentsDir = null, bundledDir =const cancelled = () => { emit({ type: 'done', finishReason: 'aborted', cancelled: true, tokensIn: null, tokensOut: null, usageSource: 'unavailable', durationMs: 0, model: route.name, contextMax: route.contextWindow }); return { cancelled: true }; };if (c.cancelStart) return cancelled();- if (c.route !== `${route.route}/${route.modelId}` || c.routeGeneration !== runtime.generation) {+ // The model and the chat's thinking level (the Reasoning menu) are sent to the Harness when either changed. A level the route does+ // not have (an API model with its own vocabulary) is dropped and the model's own default stands: never an error for the person.+ const level = effortOf(thinking);+ const wanted = `${route.route}/${route.modelId}/${level ?? ''}`;+ if (c.route !== wanted || c.routeGeneration !== runtime.generation) {+ const select = effort => runtime.gw.rpc('session/selectModel', { request: { sessionId, provider: route.route, model: route.modelId, ...(effort ? { reasoningEffort: effort } : {}) } });try {- await runtime.gw.rpc('session/selectModel', { request: { sessionId, provider: route.route, model: route.modelId } });+ try { await select(level); } catch (e) { if (!level || e.code === 'session/not-found') throw e; await select(null); }} catch (e) {if (e.code === 'session/not-found') { delete sessionMap[c.id]; persist(); }throw new AgentUnavailable(`The coding assistant could not use ${route.name}. Pick the model again and retry.`, 'model-unavailable', e.message);}- c.route = `${route.route}/${route.modelId}`; c.routeGeneration = runtime.generation;+ c.route = wanted; c.routeGeneration = runtime.generation;}// The chat's permission mode is read once, here, for this whole turn. A change the user makes while thediff --git a/apps/host/server/agent/events.mjs b/apps/host/server/agent/events.mjsindex 7dfa80221..7862d4b3a 100644--- a/apps/host/server/agent/events.mjs+++ b/apps/host/server/agent/events.mjs@@ -91,6 +91,16 @@ export function withPlain(raw) {return plain ? `${plain}\n\nDetails: ${raw.replace(/^Error:\s*/, '')}` : raw;}+/**+ * A command that succeeded and printed nothing says so in words (the card used to show "(no output)" under a "Ran" that could not+ * tell the person whether anything happened). Opening a folder or a window is not something a command can prove it did, so the+ * sentence says only what is known.+ */+export function outcomeText(tool, text) {+ if ((tool === 'pwsh' || tool === 'bash') && /^\(no output\)$/.test(text.trim())) return 'The command finished without any error. It printed nothing, so there is nothing more to show.';+ return text;+}+/** Text of a tool-result message, bounded. */export function resultText(message) {const parts = Array.isArray(message?.content) ? message.content : [];@@ -204,7 +214,7 @@ export class TurnProjector {const frame = this.#card(callId, {state: denied ? 'denied' : isError ? 'error' : 'done',// A card the person (or the mode) refused keeps its plain sentence; the model's own error text ("the user rejected ...") is not for them.- result: this.byPolicy.has(callId) ? NOT_IN_MODE : denied ? (entry.card.result ?? 'Denied. Nothing ran.') : isError ? withPlain(resultText(d.message)) : resultText(d.message),+ result: this.byPolicy.has(callId) ? NOT_IN_MODE : denied ? (entry.card.result ?? 'Denied. Nothing ran.') : isError ? withPlain(resultText(d.message)) : outcomeText(entry.card.name, resultText(d.message)),durationMs: Math.max(0, (e.time ?? 0) - entry.calledAt),});if (frame) out.push(frame);diff --git a/apps/host/server/agent/profile.mjs b/apps/host/server/agent/profile.mjsindex 7bd71c6ba..1938106ce 100644--- a/apps/host/server/agent/profile.mjs+++ b/apps/host/server/agent/profile.mjs@@ -85,7 +85,29 @@ function fingerprint(provider, secret) {return createHash('sha256').update(JSON.stringify(provider)).update('|').update(secret ?? '').digest('hex').slice(0, 16);}-/** Output budget that leaves the prompt room: a quarter of the window, between 1K and 8K. */+/**+ * Output budget for an API model, whose real limit the app does not know: a quarter of the window, between 1K and 8K.+ * (A local model has its own, much larger budget: `localOutputBudget`.)+ */+/**+ * Output budget for a LOCAL model: nearly the whole window, so a thinking model is never cut off mid-thought by the app's own+ * limit (it was 8192, which stopped big coding answers at exactly 8192 tokens). The engine clamps the request to what is left+ * of the context (`fit_max_tokens`, which the app turns on for the models it loads), so a long prompt never turns this into an error.+ */+export function localOutputBudget(context) {+ const c = Number.isFinite(context) && context > 0 ? context : 32768;+ return Math.max(8192, c - 4096);+}++/**+ * The thinking levels the chat's Reasoning menu offers, as wire values for a local Strata model. `off` sends enable_thinking false;+ * the rest send enable_thinking true plus the effort (the server reads both; "Max" is the highest effort it has).+ */+export const LOCAL_EFFORTS = Object.freeze({ off: null, low: 'low', medium: 'medium', high: 'high', max: 'high' });+/** The menu's words to the Harness's level names. */+export const EFFORT_LEVEL = Object.freeze({ off: 'off', low: 'low', medium: 'medium', high: 'high', max: 'max' });+export const effortOf = value => EFFORT_LEVEL[String(value ?? '').toLowerCase()] ?? null;+export function outputBudget(context) {const c = Number.isFinite(context) && context > 0 ? context : 32768;return Math.max(1024, Math.min(8192, Math.floor(c / 4)));@@ -118,10 +140,12 @@ export function buildProfile({ local = [], connections = [], gate = true }) {api: 'openai-completions',baseURL: inst.baseUrl,apiKeyEnv: keyEnv,- compat: { supportsStore: false },+ // `qwen` thinking format: enable_thinking true/false plus reasoning_effort, which the Strata server reads.+ compat: { supportsStore: false, thinkingFormat: 'qwen', supportsReasoningEffort: true },models: [{- id: inst.modelId, name: inst.label || inst.modelId, contextWindow: context, maxTokens: outputBudget(context),+ id: inst.modelId, name: inst.label || inst.modelId, contextWindow: context, maxTokens: localOutputBudget(context),input: inst.vision ? ['text', 'image'] : ['text'],+ reasoning: true, reasoningEfforts: LOCAL_EFFORTS,}],};slots = Math.max(slots, Number.isFinite(inst.slots) && inst.slots >= 1 ? Math.floor(inst.slots) : 1);diff --git a/apps/host/server/index.mjs b/apps/host/server/index.mjsindex 06e008435..3b7cd276f 100644--- a/apps/host/server/index.mjs+++ b/apps/host/server/index.mjs@@ -478,7 +478,7 @@ async function runAgentTurn({ res, controller, chatId, body, message, workspaceItry {// "/goal ..." is the Harness's goal command (set, show, pause, resume, clear); everything else is a message.const goalLine = /^\/goal(?:\s+([\s\S]*))?$/i.exec(message.trim());- const turnOptions = { chatId, message, workspacePath: cwd, choice, mode: typeof body.mode === 'string' ? body.mode : null, history: Array.isArray(body.history) ? body.history : [], onFrame: write, signal: controller.signal };+ const turnOptions = { chatId, message, workspacePath: cwd, choice, mode: typeof body.mode === 'string' ? body.mode : null, thinking: typeof body.thinking === 'string' ? body.thinking : null, history: Array.isArray(body.history) ? body.history : [], onFrame: write, signal: controller.signal };const result = goalLine ? await agent.goalCommand({ ...turnOptions, line: goalLine[1] ?? '' }) : await agent.runTurn(turnOptions);if (result?.detached) void result.ended.then(r => settle(r?.tokensOut ?? null), () => settle(null));else settle(result?.tokensOut ?? null, result?.error ? 'The agent turn failed.' : undefined);diff --git a/apps/host/server/native-control.mjs b/apps/host/server/native-control.mjsindex c3a7baed0..968635a9c 100644--- a/apps/host/server/native-control.mjs+++ b/apps/host/server/native-control.mjs@@ -103,6 +103,8 @@ export function createNativeControl({directory,requests=()=>[],changed=()=>{},sp/* Installed app: several requests at once by default (4, fewer where the card has no room), unless a number was chosen. */const parallelChoice=packaged?chooseParallel({config:cfg,values:args.values??{},hardware:snapshot.hardware?.gpus?.length?snapshot.hardware:{gpus:hardwareProbe()??[]}}):null;let parallelNote=null;+ /* The agent asks for long answers (a thinking model can use most of the window): let the engine clamp the output to what is left of the context instead of refusing the request. */+ if(packaged&&cfg.fit_max_tokens===undefined)cfg.fit_max_tokens=true;if(parallelChoice){const applied=applyParallel(cfg,parallelChoice);for(const k of Object.keys(cfg))if(!(k in applied.cfg))delete cfg[k];Object.assign(cfg,applied.cfg);parallelNote=applied.note;}const fingerprint=createHash('sha256').update(JSON.stringify({...cfg,api_key:undefined,port:undefined})).digest('hex');const placement=Array.isArray(cfg.gpu)?cfg.gpu:cfg.gpu===undefined?[]:[cfg.gpu];const matching=[...instances.values()].find(i=>i.model===args.model&&JSON.stringify(i.gpus)===JSON.stringify(placement));diff --git a/apps/host/tests/agent-events.test.mjs b/apps/host/tests/agent-events.test.mjsindex 2cdefffd2..3eeb540c2 100644--- a/apps/host/tests/agent-events.test.mjs+++ b/apps/host/tests/agent-events.test.mjs@@ -18,7 +18,9 @@ test('the generated profile names credentials, never carries them, and declaresconst providers = entries.find(e => e.id === 'llm-pi-ai').config.providers;assert.deepEqual(Object.keys(providers).sort(), ['api-c0nn-1234-ab', 'strata-local']);assert.equal(providers['strata-local'].models[0].contextWindow, 32768);- assert.equal(providers['strata-local'].models[0].maxTokens, 8192);+ assert.equal(providers['strata-local'].models[0].maxTokens, 28672, 'a local model may use nearly the whole window (it was 8192, which cut big answers off)');+ assert.equal(providers['strata-local'].models[0].reasoning, true);+ assert.equal(providers['strata-local'].compat.thinkingFormat, 'qwen');assert.equal(providers['api-c0nn-1234-ab'].models[1].contextWindow, ASSUMED_API_CONTEXT, 'an unknown API window is a stated assumption, not a catalog default');assert.equal(entries.find(e => e.id === 'agent-default-model').config.provider, 'strata-local', 'local stays the default');assert.equal(entries.find(e => e.id === 'session-title-llm').disabled, true);diff --git a/apps/host/tests/agent-tools.test.mjs b/apps/host/tests/agent-tools.test.mjsindex 4236aaf41..a7901c2f8 100644--- a/apps/host/tests/agent-tools.test.mjs+++ b/apps/host/tests/agent-tools.test.mjs@@ -29,7 +29,7 @@ test('the gate: only commands that provably just look are allowed without a quesfor (const ok of ['Get-ChildItem', 'Get-ChildItem -Recurse src | Select-Object -First 5', 'git status', 'git log --oneline -5', 'git diff', 'cat notes.txt', 'ls -la', 'node --version', 'Get-Content a.txt | Select-String needle', 'git branch', 'git remote -v']) {assert.equal(isReadOnlyShell(ok), true, ok);}- for (const bad of ['', 'Set-Content a.txt x', 'echo hi > a.txt', 'Get-ChildItem; Remove-Item a.txt', 'rm -rf x', 'git commit -m x', 'git push', 'git branch -D x', 'git config user.name x', 'git remote add o u', 'curl http://x', 'Get-ChildItem | ForEach-Object { $_ }', 'powershell -Command x', 'node script.js', 'python x.py', 'ls && rm a', 'find . -delete', 'echo $(whoami)', 'Invoke-WebRequest u', 'Start-Process x', 'New-Item a', 'cat a | tee b']) {+ for (const bad of ['', 'Set-Content a.txt x', 'echo hi > a.txt', 'Get-ChildItem; Remove-Item a.txt', 'rm -rf x', 'git commit -m x', 'git push', 'git branch -D x', 'git config user.name x', 'git remote add o u', 'curl http://x', 'Get-ChildItem | ForEach-Object { $_ }', 'Start-Process explorer.exe', 'explorer.exe C:/Users/x/Downloads', 'ii ~/Downloads', 'Invoke-Item $env:USERPROFILE/Downloads', 'start .', 'open .', 'xdg-open .', 'notepad a.txt', 'powershell -Command x', 'node script.js', 'python x.py', 'ls && rm a', 'find . -delete', 'echo $(whoami)', 'Invoke-WebRequest u', 'Start-Process x', 'New-Item a', 'cat a | tee b']) {assert.equal(isReadOnlyShell(bad), false, bad);}assert.equal(isReadOnlyShell('x'.repeat(700)), false, 'a long command is never trusted');@@ -220,3 +220,29 @@ function readdirDeep(dir) {for (const e of readdirSync(dir, { withFileTypes: true })) { const p = join(dir, e.name); if (e.isDirectory()) out.push(...readdirDeep(p)); else out.push(p); }return out;}++test('the output budget is not the old 8192, and the Reasoning menu reaches the request (off sends enable_thinking false)', { skip }, async (t) => {+ const worker = await startScriptedWorker({});+ const dir = mkdtempSync(join(tmpdir(), 'strata-think-'));+ const ws = join(dir, 'ws'); mkdirSync(ws, { recursive: true });+ const local = [{ id: 'i', baseUrl: worker.url, modelId: 'scripted-model', context: 65536 }];+ const bridge = createAgentBridge({ dataDir: dir, localInstances: () => local, connections: () => [], env: process.env, startTimeoutMs: 90_000 });+ t.after(async () => { await bridge.close(); await worker.close(); cleanup(dir); });+ const ask = async (chatId, thinking) => {+ worker.requests.length = 0;+ await bridge.runTurn({ chatId, message: 'SCRIPT:echo hi', workspacePath: ws, mode: 'bypass', thinking, onFrame: () => {} });+ return worker.requests.find(r => r.toolNames.length).params;+ };+ const off = await ask('think-off', 'Off');+ assert.equal(off.enable_thinking, false);+ assert.ok(off.max_completion_tokens > 40000, `budget ${off.max_completion_tokens} (the window minus the prompt, never the old 8192)`);+ const high = await ask('think-high', 'High');+ assert.equal(high.enable_thinking, true); assert.equal(high.reasoning_effort, 'high');+ const low = await ask('think-low', 'Low');+ assert.equal(low.enable_thinking, true); assert.equal(low.reasoning_effort, 'low');+ const max = await ask('think-max', 'Max');+ assert.equal(max.enable_thinking, true); assert.equal(max.reasoning_effort, 'high');+ // the same chat switching its level between turns+ assert.equal((await ask('think-off', 'High')).enable_thinking, true);+ assert.equal((await ask('think-off', 'Off')).enable_thinking, false);+});