fix(auxiliary): stop capping output with max_tokens by default (#34530) (#34845)

* fix(auxiliary): stop capping output with max_tokens by default

Auxiliary LLM calls (compression, titles, vision, etc.) no longer send
max_tokens on the OpenAI-compatible chat-completions path. Most providers
treat an omitted max_tokens as "use the model max", which is what we want;
an explicit cap only risks truncation or a wire-format 400.

This was surfaced by GitHub Copilot / GPT-5 (#34530): those models reject
max_tokens and require max_completion_tokens, so compression 400'd and fell
back to a static context marker. Omitting the param sidesteps that quirk
(and ZAI vision's error 1210) entirely.

The Anthropic Messages wire (MiniMax + /anthropic endpoints) keeps
max_tokens because it is a mandatory field there.

* test(auxiliary): update temperature-retry assertions for omitted max_tokens

The temperature-retry tests asserted retry_kwargs["max_tokens"] == 500 on an
api.openai.com endpoint. Now that auxiliary calls omit max_tokens on
OpenAI-compatible endpoints (#34530), that key is absent. Assert it's absent
in both first and retry kwargs and use model as the survives-the-retry witness.

tests/agent/test_auxiliary_client.py

@@ -88,6 +88,62 @@ class TestAuxiliaryMaxTokensParam:
             assert auxiliary_max_tokens_param(2048) == {"max_completion_tokens": 2048}
 
 
+class TestBuildCallKwargsMaxTokens:
+    """_build_call_kwargs should not cap output by default (#34530).
+
+    Most chat-completions providers treat an omitted max_tokens as "use the
+    model max", which is what we want for auxiliary tasks. An explicit cap only
+    risks truncation or a wire-format 400 (GitHub Copilot / GPT-5 reject
+    max_tokens; ZAI vision rejects it entirely). The Anthropic Messages wire is
+    the one exception — max_tokens is a mandatory field there.
+    """
+
+    @pytest.mark.parametrize(
+        "provider,model,base_url",
+        [
+            ("copilot", "gpt-5.4", "https://api.githubcopilot.com"),
+            ("copilot", "gpt-5.5", "https://api.githubcopilot.com"),
+            ("custom", "gpt-5", "https://api.openai.com/v1"),
+            ("openrouter", "anthropic/claude-sonnet-4.6", "https://openrouter.ai/api/v1"),
+            ("nous", "hermes-4", "https://inference-api.nousresearch.com/v1"),
+            ("custom", "qwen", "http://localhost:8080/v1"),
+            ("zai", "glm-4v-flash", "https://open.bigmodel.cn/api/paas/v4"),
+        ],
+    )
+    def test_omits_max_tokens_for_openai_compatible(self, provider, model, base_url):
+        from agent.auxiliary_client import _build_call_kwargs
+
+        kwargs = _build_call_kwargs(
+            provider=provider,
+            model=model,
+            messages=[{"role": "user", "content": "hi"}],
+            max_tokens=1234,
+            base_url=base_url,
+        )
+        assert "max_tokens" not in kwargs
+        assert "max_completion_tokens" not in kwargs
+
+    @pytest.mark.parametrize(
+        "provider,model,base_url",
+        [
+            ("minimax", "minimax-m2", "https://api.minimax.io/v1"),
+            ("custom", "claude", "https://proxy.example.com/anthropic/v1"),
+        ],
+    )
+    def test_keeps_max_tokens_on_anthropic_wire(self, provider, model, base_url):
+        from agent.auxiliary_client import _build_call_kwargs
+
+        kwargs = _build_call_kwargs(
+            provider=provider,
+            model=model,
+            messages=[{"role": "user", "content": "hi"}],
+            max_tokens=1234,
+            base_url=base_url,
+        )
+        assert kwargs["max_tokens"] == 1234
+        assert "max_completion_tokens" not in kwargs
+
+
 class TestNormalizeAuxProvider:
     def test_maps_github_copilot_aliases(self):
         assert _normalize_aux_provider("github") == "copilot"

tests/agent/test_unsupported_temperature_retry.py

@@ -112,8 +112,13 @@ class TestCallLlmUnsupportedTemperatureRetry:
         retry_kwargs = client.chat.completions.create.call_args_list[1].kwargs
         assert first_kwargs["temperature"] == 0.3
         assert "temperature" not in retry_kwargs
-        # other kwargs preserved
-        assert retry_kwargs["max_tokens"] == 500
+        # max_tokens is intentionally omitted on OpenAI-compatible endpoints
+        # (#34530) — auxiliary calls let the model max out its own output — so
+        # it must be absent in BOTH the first and retry kwargs. Use a kwarg that
+        # actually survives (model) to prove the retry preserves the rest.
+        assert "max_tokens" not in first_kwargs
+        assert "max_tokens" not in retry_kwargs
+        assert retry_kwargs["model"] == first_kwargs["model"]
 
     def test_non_temperature_400_does_not_retry_as_temperature(self):
         """Unrelated 400s (e.g. bad tool role) must not silently drop temp."""
@@ -207,7 +212,11 @@ class TestAsyncCallLlmUnsupportedTemperatureRetry:
         retry_kwargs = client.chat.completions.create.call_args_list[1].kwargs
         assert first_kwargs["temperature"] == 0.3
         assert "temperature" not in retry_kwargs
-        assert retry_kwargs["max_tokens"] == 500
+        # max_tokens is intentionally omitted on OpenAI-compatible endpoints
+        # (#34530); assert it's absent and that model survives the retry.
+        assert "max_tokens" not in first_kwargs
+        assert "max_tokens" not in retry_kwargs
+        assert retry_kwargs["model"] == first_kwargs["model"]
 
     @pytest.mark.asyncio
     async def test_async_non_temperature_400_does_not_retry(self):