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++
API models: answer room is the connection's own maximum output when the provider lists one, else a generous 32K bounded by the window minus the prompt (was a quarter of the window, at most 8,192, which cut thinking models off with no answer)
Niko1221df52a2854828
diff --git a/apps/host/server/agent/profile.mjs b/apps/host/server/agent/profile.mjsindex 8bdbdab8c..3129c0edf 100644--- a/apps/host/server/agent/profile.mjs+++ b/apps/host/server/agent/profile.mjs@@ -116,9 +116,18 @@ export const LOCAL_EFFORTS = Object.freeze({ off: null, low: 'low', medium: 'medexport 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)));+/**+ * Output budget for an API model. The connection's own maximum output is used when it is known (`maxOutput`, read from the provider's model+ * list). Otherwise a generous 32K, never more than the window minus room for the prompt (a quarter of a small window at most is the prompt's).+ * It used to be a quarter of the window capped at 8,192, which cut a thinking model off while it was still thinking: `length` and no answer.+ * The budget is the answer's room: a model that thinks keeps what is left of it for the answer, and the Harness sends the window minus the prompt.+ */+export const DEFAULT_API_OUTPUT = 32768;+export function outputBudget(context, maxOutput = null) {+ const c = Number.isFinite(context) && context > 0 ? context : null;+ const window = c === null ? Infinity : Math.max(1024, c - Math.min(4096, Math.floor(c / 4)));+ if (Number.isFinite(maxOutput) && maxOutput > 0) return Math.max(1024, Math.min(Math.floor(maxOutput), window));+ return Math.max(1024, Math.min(DEFAULT_API_OUTPUT, window));}/**@@ -180,7 +189,7 @@ export function buildProfile({ local = [], connections = [], gate = true, tools...(conn.apiKey || keyless ? { apiKeyEnv: keyEnv } : {}),models: conn.models.map(m => ({id: m.id, name: m.name || m.id,- contextWindow: m.contextMax ?? ASSUMED_API_CONTEXT, maxTokens: outputBudget(m.contextMax ?? ASSUMED_API_CONTEXT),+ contextWindow: m.contextMax ?? ASSUMED_API_CONTEXT, maxTokens: outputBudget(m.contextMax ?? ASSUMED_API_CONTEXT, m.maxOutput),input: ['text'],})),};diff --git a/apps/host/server/model-providers.mjs b/apps/host/server/model-providers.mjsindex ce4208c53..e71c46df0 100644--- a/apps/host/server/model-providers.mjs+++ b/apps/host/server/model-providers.mjs@@ -75,7 +75,7 @@ export function createModelProviders({ directory, changed = () => {}, fetchImplif (rows.length > 100) fail('Add up to 100 models to one connection.');const models = rows.map(row => {const id = text(row.id, 256); if (!id || /[\r\n]/.test(id)) fail('Every model needs a model ID.');- return { id, name: text(row.name) || id, contextMax: numeric(row.contextMax) };+ return { id, name: text(row.name) || id, contextMax: numeric(row.contextMax), maxOutput: numeric(row.maxOutput) };});if (new Set(models.map(m => m.id)).size !== models.length) fail('Each model ID must be unique within this connection.');return { id: existing?.id ?? randomUUID(), kind, provider: catalog?.id ?? null, name, baseUrl, protocol, apiKey, models };@@ -128,7 +128,7 @@ export function createModelProviders({ directory, changed = () => {}, fetchImplif (connection.protocol !== 'anthropic' || !data.has_more) break;after = data.last_id; if (!after) break;}- const models = [...new Map(rows.filter(m => text(m.id, 256)).map(m => [m.id, { id: text(m.id, 256), name: text(m.display_name ?? m.name) || text(m.id, 256), contextMax: numeric(m.context_length ?? m.max_context_length ?? m.context_window) }])).values()];+ const models = [...new Map(rows.filter(m => text(m.id, 256)).map(m => [m.id, { id: text(m.id, 256), name: text(m.display_name ?? m.name) || text(m.id, 256), contextMax: numeric(m.context_length ?? m.max_context_length ?? m.context_window), maxOutput: numeric(m.max_output_tokens ?? m.max_completion_tokens ?? m.top_provider?.max_completion_tokens) }])).values()];return { models: models.sort((a, b) => a.name.localeCompare(b.name)) };}async function stream({ selection, message, history = [], systemPrompt = '', signal, onFrame }) {diff --git a/apps/host/tests/agent-events.test.mjs b/apps/host/tests/agent-events.test.mjsindex 68a293abb..71ad09837 100644--- a/apps/host/tests/agent-events.test.mjs+++ b/apps/host/tests/agent-events.test.mjs@@ -49,7 +49,13 @@ test('routeFor picks the chat\'s own choice', () => {assert.equal(routeFor(p, { connectionId: conn.id, modelId: 'm2' }).modelId, 'm2');assert.equal(routeFor(p, { instanceId: 'inst-1' }).route, 'strata-local');assert.equal(routeFor(p, { connectionId: 'gone', modelId: 'm1' }), null);- assert.equal(outputBudget(2048), 1024); assert.equal(outputBudget(1_000_000), 8192);+ assert.equal(outputBudget(2048), 1536, 'a small window keeps a quarter for the prompt');+ assert.equal(outputBudget(1_000_000), 32768, 'a big window with no known maximum gets a generous 32K, not 8K');+ assert.equal(outputBudget(128_000), 32768); assert.equal(outputBudget(null), 32768);+ assert.equal(outputBudget(16_384), 16_384 - 4096, 'never more than the window minus room for the prompt');+ assert.equal(outputBudget(200_000, 64_000), 64_000, 'the connection own maximum is used when known');+ assert.equal(outputBudget(8192, 64_000), 8192 - 2048, 'a known maximum is still bounded by the window');+ assert.equal(outputBudget(200_000, 4096), 4096, 'a small known maximum is respected');});test('the agent environment keeps what a shell needs and drops ambient credentials', () => {@@ -161,3 +167,12 @@ test('the browser and the person\'s own tools have plain sentences, translated iassert.equal(presentTool({ callId: 'c4', name: 'mcp__browser__browser_snapshot', args: {} }).preview, '', 'no empty braces on a card');assert.equal(presentTool({ callId: 'c5', name: 'mcp__browser__browser_evaluate', args: { function: '() => document.title' } }).preview, '', 'a script in the page is not shown as code on the card');});++test('an API model\'s route gets a generous answer room, or the connection\'s own maximum when the provider listed one', () => {+ const profile = buildProfile({ local: [], connections: [{ id: 'c1', name: 'C', baseUrl: 'https://x/v1', protocol: 'openai-chat', apiKey: 'k', models: [+ { id: 'unknown', contextMax: null }, { id: 'big', contextMax: 200_000, maxOutput: 64_000 }, { id: 'small', contextMax: 16_384 }] }] });+ const providers = JSON.parse(profile.patch.split('\n').slice(1).join('\n')).find(e => e.id === 'llm-pi-ai').config.providers;+ const models = Object.values(providers)[0].models;+ const by = Object.fromEntries(models.map(m => [m.id, m.maxTokens]));+ assert.equal(by.unknown, 32768); assert.equal(by.big, 64_000); assert.equal(by.small, 16_384 - 4096);+});