fix(context): persist NVIDIA DeepSeek endpoint limit

agent/conversation_loop.py

@@ -4746,6 +4746,13 @@ def run_conversation(
                             provider=agent.provider,
                             api_mode=agent.api_mode,
                         )
+                        # Persist an explicit provider-reported limit before
+                        # compression/retry. The next request can be rate
+                        # limited, omit usage, or the process can restart; none
+                        # of those should discard metadata the provider already
+                        # confirmed. Keep the probe flags as a best-effort
+                        # post-success retry if this write cannot complete.
+                        save_context_length(agent.model, agent.base_url, new_ctx)
                         # Context probing flags — only set on built-in
                         # compressor (plugin engines manage their own).  This
                         # value came from the provider, so it is safe to cache.

agent/model_metadata.py

@@ -634,12 +634,17 @@ def _is_known_provider_base_url(base_url: str) -> bool:
 
 
 def _endpoint_scoped_context_length(model: str, base_url: str) -> Optional[int]:
-    """Return metadata confirmed only for the Kimi Coding endpoint.
+    """Return context metadata confirmed for one provider endpoint.
 
     Kimi Coding serves K3 under the bare slug ``k3``, but users may also
     configure or select the public-facing aliases ``kimi-k3`` and
     ``kimi-k3-cot``. Only canonical ``https://api.kimi.com/coding`` endpoints
     (legacy Moonshot keys do not serve K3) get the 1 Mi context window.
+
+    NVIDIA NIM serves ``deepseek-ai/deepseek-v4-pro`` with a 262,144-token
+    window even though DeepSeek's native endpoint serves the V4 family with a
+    1M window. Keep the lower limit scoped to NVIDIA instead of weakening the
+    global model-family metadata.
     """
     normalized = _normalize_base_url(base_url)
     try:
@@ -659,6 +664,18 @@ def _endpoint_scoped_context_length(model: str, base_url: str) -> Optional[int]:
         and model.strip().lower() in {"k3", "kimi-k3", "kimi-k3-cot"}
     ):
         return 1_048_576
+    if (
+        parsed.scheme.lower() == "https"
+        and (parsed.hostname or "").lower() == "integrate.api.nvidia.com"
+        and port in (None, 443)
+        and parsed.username is None
+        and parsed.password is None
+        and parsed.path.rstrip("/") == "/v1"
+        and not parsed.query
+        and not parsed.fragment
+        and model.strip().lower() == "deepseek-ai/deepseek-v4-pro"
+    ):
+        return 262_144
     return None
 
 

tests/agent/test_model_metadata.py

@@ -134,6 +134,42 @@ class TestEstimateRequestTokensRough:
 # =========================================================================
 
 class TestDefaultContextLengths:
+    def test_nvidia_deepseek_v4_pro_context_is_endpoint_scoped(self):
+        """NVIDIA's 262K NIM window must not lower DeepSeek V4 globally."""
+        with patch("agent.model_metadata.get_cached_context_length", return_value=None), \
+             patch("agent.model_metadata.fetch_model_metadata", return_value={}), \
+             patch("agent.model_metadata.fetch_endpoint_model_metadata", return_value={}), \
+             patch("agent.model_metadata._query_ollama_api_show", return_value=None), \
+             patch("agent.models_dev.lookup_models_dev_context", return_value=None):
+            accepted_urls = (
+                "https://integrate.api.nvidia.com/v1",
+                "https://INTEGRATE.API.NVIDIA.COM/v1/",
+                "https://integrate.api.nvidia.com:443/v1",
+            )
+            rejected_urls = (
+                "http://integrate.api.nvidia.com/v1",
+                "https://integrate.api.nvidia.com:8443/v1",
+                "https://integrate.api.nvidia.com/v1/other",
+                "https://integrate.api.nvidia.com/v1?route=other",
+                "https://example.invalid/v1",
+                "https://api.deepseek.com/v1",
+                "https://openrouter.ai/api/v1",
+            )
+
+            for base_url in accepted_urls:
+                assert get_model_context_length(
+                    "deepseek-ai/deepseek-v4-pro",
+                    provider="nvidia",
+                    base_url=base_url,
+                ) == 262_144
+
+            for base_url in rejected_urls:
+                assert get_model_context_length(
+                    "deepseek-ai/deepseek-v4-pro",
+                    provider="nvidia",
+                    base_url=base_url,
+                ) == 1_000_000
+
     def test_k3_context_is_scoped_to_confirmed_coding_endpoint(self):
         """The bare ``k3`` slug's 1 Mi context must not leak to unverified endpoints.
 
@@ -1076,4 +1112,3 @@ class TestMoAContextLength:
         assert compressor.context_length == configured_context
         assert compressor.threshold_tokens == configured_context // 2
         endpoint_probe.assert_not_called()
-

tests/run_agent/test_413_compression.py

@@ -327,6 +327,62 @@ class TestHTTP413Compression:
         assert result["final_response"] == "Recovered after compression"
 
 
+    def test_provider_context_limit_is_cached_before_retry_succeeds(self, agent):
+        """A confirmed limit survives when the recovery response omits usage."""
+        err_400 = Exception(
+            "Error code: 400 - {'error': {'message': "
+            "\"This model's maximum context length is 262144 tokens. "
+            "However, your messages resulted in 271877 tokens.\", 'code': 400}}"
+        )
+        err_400.status_code = 400
+        # NVIDIA-compatible endpoints can omit usage. Before the fix, caching
+        # happened only in the successful-response usage block, so this lost
+        # the provider-confirmed limit across a restart.
+        ok_resp = _mock_response(
+            content="Recovered without usage metadata",
+            finish_reason="stop",
+            usage=None,
+        )
+        agent.model = "deepseek-ai/deepseek-v4-pro"
+        agent.provider = "nvidia"
+        agent.base_url = "https://integrate.api.nvidia.com/v1"
+        agent.context_compressor.update_model(
+            model=agent.model,
+            context_length=1_000_000,
+            base_url=agent.base_url,
+            api_key=agent.api_key,
+            provider=agent.provider,
+            api_mode=agent.api_mode,
+        )
+        agent.client.chat.completions.create.side_effect = [err_400, ok_resp]
+
+        with (
+            patch.object(agent, "_compress_context") as mock_compress,
+            patch.object(agent, "_persist_session"),
+            patch.object(agent, "_save_trajectory"),
+            patch.object(agent, "_cleanup_task_resources"),
+            patch("agent.conversation_loop.save_context_length") as mock_save,
+        ):
+            mock_compress.return_value = (
+                [{"role": "user", "content": "compressed summary"}],
+                "compressed prompt",
+            )
+            result = agent.run_conversation(
+                "continue",
+                conversation_history=[
+                    {"role": "user", "content": "previous question"},
+                    {"role": "assistant", "content": "previous answer"},
+                ],
+            )
+
+        assert result["completed"] is True
+        mock_save.assert_called_once_with(
+            "deepseek-ai/deepseek-v4-pro",
+            "https://integrate.api.nvidia.com/v1",
+            262_144,
+        )
+
+
     def test_context_length_retry_rebuilds_request_after_compression(self, agent):
         """Retry must send the compressed transcript, not the stale oversized payload."""
         err_400 = Exception(
@@ -1132,5 +1188,3 @@ class TestOverflowWithCompactionDisabled:
         assert result.get("failed") is True
         assert result.get("compaction_disabled") is True
         assert "auto-compaction is disabled" in result["error"]
-
-