🎯 fix: Route the Label Cap Per Model Family

Round-twenty review: the 256-token label cap set maxTokens
unconditionally, but GPT-5+ rejects max_tokens (the OpenAI builder
routes its cap into modelKwargs.max_completion_tokens /
max_output_tokens) and o-series models reject it with no stable kwargs
alternative — every label on those models would have failed. The cap
now mirrors the builder: modelKwargs for GPT-5+ (responses-API aware),
no cap for o-series (title parity; the 200-char persistence bound
still applies), maxOutputTokens for Google, maxTokens otherwise.
Pinned in host.spec for both reasoning families.

The round's other finding is the tenth instance of the documented
edited+reconnect index limitation, answered on-thread.

packages/api/src/agents/activityLabels/__tests__/host.spec.ts

@@ -197,6 +197,31 @@ describe('resolveActivityLabelModel model precedence', () => {
     expect((resolved.clientOptions as Record<string, unknown>).maxTokens).toBe(256);
   });
 
+  /** GPT-5+ rejects `max_tokens`: the label cap must ride in modelKwargs
+   *  exactly as the OpenAI builder routes primary caps. */
+  it('routes the label cap into modelKwargs for GPT-5-family models', async () => {
+    mockGetOptions.mockResolvedValueOnce({
+      llmConfig: { model: 'gpt-5.2' },
+    } as never);
+    const resolved = await resolve({ activityLabel: true, activityModel: 'gpt-5.2' });
+    const clientOptions = resolved.clientOptions as Record<string, unknown>;
+    expect(clientOptions.maxTokens).toBeUndefined();
+    expect(clientOptions.modelKwargs).toEqual({ max_completion_tokens: 256 });
+  });
+
+  /** o-series models reject `max_tokens` and get NO cap — title parity;
+   *  the 200-char persistence bound still applies. */
+  it('sets no cap at all for o-series reasoning models', async () => {
+    mockGetOptions.mockResolvedValueOnce({
+      llmConfig: { model: 'o3-mini' },
+    } as never);
+    const resolved = await resolve({ activityLabel: true, activityModel: 'o3-mini' });
+    const clientOptions = resolved.clientOptions as Record<string, unknown>;
+    expect(clientOptions.maxTokens).toBeUndefined();
+    expect(clientOptions.modelKwargs).toBeUndefined();
+    expect(clientOptions.maxOutputTokens).toBeUndefined();
+  });
+
   /** The Anthropic carrier holds client CONSTRUCTION options — for
    *  user-provided base URLs that includes the SSRF-safe fetch dispatcher —
    *  so it must survive the strip even with no custom headers, and by the

packages/api/src/agents/activityLabels/host.ts

@@ -334,17 +334,35 @@ export async function resolveActivityLabelModel({
   const anthropicCarrier = rawOptions.clientOptions;
   const clientOptions = Object.fromEntries(
     Object.entries(rawOptions).filter(([key]) => !omitTitleOptions.has(key)),
-  ) as MaybeAzureConfig & { clientOptions?: { defaultHeaders?: unknown } };
+  ) as MaybeAzureConfig & {
+    clientOptions?: { defaultHeaders?: unknown };
+    modelKwargs?: Record<string, unknown>;
+  };
   if (anthropicCarrier != null && clientOptions.clientOptions == null) {
     clientOptions.clientOptions = anthropicCarrier;
   }
   /** Replace the stripped primary caps with a SMALL one — see
-   *  {@link LABEL_MAX_OUTPUT_TOKENS}. Installed after the filter so the
-   *  omit set cannot remove it; keyed per provider family (Google-style
-   *  wrappers read `maxOutputTokens`, the rest `maxTokens`). */
+   *  {@link LABEL_MAX_OUTPUT_TOKENS} — routed with the SAME model/API
+   *  conversion as the OpenAI builder (endpoints/openai/llm.ts): GPT-5+
+   *  rejects `max_tokens`, so its cap goes to `modelKwargs`
+   *  (responses-API aware); o-series reasoning models get NO cap at all,
+   *  matching the title path, since they reject `max_tokens` and have no
+   *  stable kwargs cap across API surfaces (the 200-char persistence bound
+   *  still applies); Google-family wrappers read `maxOutputTokens`. */
+  const isGpt5Plus = model != null && /\bgpt-[5-9](?:\.\d+)?\b/i.test(model);
+  const isOSeries = model != null && /\bo[1-9](?:[-.]|\b)/i.test(model);
   if (provider === Providers.GOOGLE || provider === Providers.VERTEXAI) {
     (clientOptions as { maxOutputTokens?: number }).maxOutputTokens = LABEL_MAX_OUTPUT_TOKENS;
-  } else {
+  } else if (isGpt5Plus) {
+    const paramName =
+      (rawOptions as { useResponsesApi?: boolean }).useResponsesApi === true
+        ? 'max_output_tokens'
+        : 'max_completion_tokens';
+    clientOptions.modelKwargs = {
+      ...(clientOptions.modelKwargs ?? {}),
+      [paramName]: LABEL_MAX_OUTPUT_TOKENS,
+    };
+  } else if (!isOSeries) {
     (clientOptions as { maxTokens?: number }).maxTokens = LABEL_MAX_OUTPUT_TOKENS;
   }
   if (options.configOptions) {