Skip to content

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++

thinking: the Strata server reads chat_template_kwargs, not a top-level enable_thinking (Off had no effect); plain output for read-only PowerShell noise

Niko1221daa45b1e37fb
diff --git a/apps/host/server/agent/events.mjs b/apps/host/server/agent/events.mjs
index 7862d4b3a..7e0c30480 100644
--- a/apps/host/server/agent/events.mjs
+++ b/apps/host/server/agent/events.mjs
@@ -97,6 +97,12 @@ export function withPlain(raw) {
* sentence says only what is known.
*/
export function outcomeText(tool, text) {
+ if (tool === 'pwsh' || tool === 'bash') {
+ // In the read-only modes PowerShell runs in Constrained Language Mode and complains about the Harness's own output-encoding line:
+ // noise appended after a command that worked. It says nothing about the command, so it is not shown.
+ text = text.replace(/\r?\n?\[stderr\]\r?\n(?:Cannot create type\. Only core types are supported in this language mode\.[\s\S]*?CannotCreateTypeConstrainedLanguage\s*)+$/, '');
+ if (!text.trim()) text = '(no output)';
+ }
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;
}
diff --git a/apps/host/server/agent/profile.mjs b/apps/host/server/agent/profile.mjs
index 1938106ce..c71c45235 100644
--- a/apps/host/server/agent/profile.mjs
+++ b/apps/host/server/agent/profile.mjs
@@ -100,8 +100,8 @@ export function localOutputBudget(context) {
}
/**
- * 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).
+ * The thinking levels the chat's Reasoning menu offers, as wire values for a local Strata model. `off` sends
+ * chat_template_kwargs.enable_thinking false; the rest send it true plus reasoning_effort (the server maps high and max to its top level).
*/
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. */
@@ -140,8 +140,9 @@ export function buildProfile({ local = [], connections = [], gate = true }) {
api: 'openai-completions',
baseURL: inst.baseUrl,
apiKeyEnv: keyEnv,
- // `qwen` thinking format: enable_thinking true/false plus reasoning_effort, which the Strata server reads.
- compat: { supportsStore: false, thinkingFormat: 'qwen', supportsReasoningEffort: true },
+ // The Strata server reads the thinking level from `chat_template_kwargs` (enable_thinking, reasoning_effort) or from a top-level
+ // reasoning_effort; a top-level `enable_thinking` (pi-ai's `qwen` format) is ignored, so the model went on thinking with the menu on Off.
+ compat: { supportsStore: false, thinkingFormat: 'chat-template', chatTemplateKwargs: { enable_thinking: { $var: 'thinking.enabled' }, reasoning_effort: { $var: 'thinking.effort', omitWhenOff: true } } },
models: [{
id: inst.modelId, name: inst.label || inst.modelId, contextWindow: context, maxTokens: localOutputBudget(context),
input: inst.vision ? ['text', 'image'] : ['text'],
diff --git a/apps/host/tests/agent-events.test.mjs b/apps/host/tests/agent-events.test.mjs
index 3eeb540c2..025046560 100644
--- a/apps/host/tests/agent-events.test.mjs
+++ b/apps/host/tests/agent-events.test.mjs
@@ -20,7 +20,7 @@ test('the generated profile names credentials, never carries them, and declares
assert.equal(providers['strata-local'].models[0].contextWindow, 32768);
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['strata-local'].compat.thinkingFormat, 'chat-template');
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.mjs
index a7901c2f8..ad91601bd 100644
--- a/apps/host/tests/agent-tools.test.mjs
+++ b/apps/host/tests/agent-tools.test.mjs
@@ -233,16 +233,18 @@ test('the output budget is not the old 8192, and the Reasoning menu reaches the
await bridge.runTurn({ chatId, message: 'SCRIPT:echo hi', workspacePath: ws, mode: 'bypass', thinking, onFrame: () => {} });
return worker.requests.find(r => r.toolNames.length).params;
};
+ const ctk = p => p.chat_template_kwargs ?? {};
const off = await ask('think-off', 'Off');
- assert.equal(off.enable_thinking, false);
+ assert.equal(ctk(off).enable_thinking, false); assert.equal(ctk(off).reasoning_effort, undefined);
+ assert.equal(off.enable_thinking, undefined, 'a top-level enable_thinking is the field the Strata server ignores');
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');
+ assert.equal(ctk(high).enable_thinking, true); assert.equal(ctk(high).reasoning_effort, 'high');
const low = await ask('think-low', 'Low');
- assert.equal(low.enable_thinking, true); assert.equal(low.reasoning_effort, 'low');
+ assert.equal(ctk(low).enable_thinking, true); assert.equal(ctk(low).reasoning_effort, 'low');
const max = await ask('think-max', 'Max');
- assert.equal(max.enable_thinking, true); assert.equal(max.reasoning_effort, 'high');
+ assert.equal(ctk(max).enable_thinking, true); assert.equal(ctk(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);
+ assert.equal(ctk(await ask('think-off', 'High')).enable_thinking, true);
+ assert.equal(ctk(await ask('think-off', 'Off')).enable_thinking, false);
});