[ApexAgents] Add a context window setting for Stirrup summarization

responses_api_agents/apex_agent/README.md

@@ -19,6 +19,11 @@ model turn. MCP calls time out after 60 seconds. Long text output uses a 20,000-
 tail excerpt when it exceeds a 24,000-token budget, estimated at four characters per token; Stirrup compresses image
 blocks to about one megapixel when serializing them for the model.
 
+Stirrup summarizes the conversation once it reaches 70% of the client's context window. Stirrup 0.1 reads that window
+from the same `max_tokens` value it sends as `max_completion_tokens`, so by default summarization starts at 70% of
+`max_output_tokens`. Set `context_window_tokens` (`apex_agent_context_window_tokens`) to the served model's context
+length to summarize at 70% of the real window instead; requests keep `max_output_tokens` as their output cap.
+
 `finish` is the only submission mechanism. A completed submission is rejected while any todo is pending or in
 progress. A max-turn exit or `status="incomplete"` does not call the grader; its initial/final snapshots are still
 saved under the configured artifact output directory for inspection.

responses_api_agents/apex_agent/app.py

@@ -69,6 +69,9 @@ class ApexAgentConfig(BaseResponsesAPIAgentConfig):
     edgar_user_agent: Optional[str]
     max_turns: int = Field(gt=0, le=200)
     max_output_tokens: int = Field(gt=0)
+    # Policy context window reported to Stirrup, which summarizes at 70% of it. Unset keeps Stirrup using
+    # max_output_tokens as the window.
+    context_window_tokens: Optional[int] = Field(default=None, gt=0)
     supports_vision: bool
     temperature: float = Field(ge=0.0)
     top_p: float = Field(gt=0.0, le=1.0)
@@ -291,6 +294,8 @@ def _sandbox_spec(self, body: ApexAgentRunRequest, instruction: str) -> SandboxS
             "foundry_services": body.foundry_services,
             "edgar_user_agent": self.config.edgar_user_agent,
         }
+        if self.config.context_window_tokens is not None:
+            runner_config["context_window_tokens"] = self.config.context_window_tokens
         return SandboxSpec(
             image=self._image or self.config.image,
             workdir=_GUEST_ROOT,

responses_api_agents/apex_agent/configs/apex_agent.yaml

@@ -46,6 +46,9 @@ apex_agent:
       edgar_user_agent: ${oc.select:apex_edgar_user_agent,"Mercor APEX-Agents apex@mercor.com"}
       max_turns: ${oc.select:apex_agent_max_turns,200}
       max_output_tokens: ${oc.select:apex_agent_max_output_tokens,32768}
+      # Policy context window for Stirrup's context summarization, which starts at 70% of it. Unset keeps Stirrup
+      # using max_output_tokens as the window, i.e. summarizing at ~23k tokens with the default 32768.
+      context_window_tokens: ${oc.select:apex_agent_context_window_tokens,null}
       supports_vision: ${oc.select:apex_agent_supports_vision,true}
       temperature: ${oc.select:apex_agent_temperature,1.0}
       top_p: ${oc.select:apex_agent_top_p,1.0}

responses_api_agents/apex_agent/stirrup_runtime.py

@@ -64,6 +64,23 @@ def truncate_tool_text(text: str) -> str:
     return text[:TOOL_OUTPUT_HEAD_CHARACTERS] + marker + text[-TOOL_OUTPUT_TAIL_CHARACTERS:]
 
 
+def context_window_client(client_class: Any, context_window_tokens: int) -> Any:
+    """Subclass a Stirrup client so context summarization uses the policy's real context window.
+
+    Stirrup 0.1 uses the client's ``max_tokens`` argument both as each request's ``max_completion_tokens`` and, via
+    the ``max_tokens`` property, as the context window whose 70% triggers summarization. The runner passes
+    ``max_output_tokens`` there, so without this override summarization starts at ~70% of the output cap. The
+    subclass keeps the output cap on requests and changes only the property the agent reads.
+    """
+
+    class ContextWindowClient(client_class):
+        @property
+        def max_tokens(self) -> int:
+            return context_window_tokens
+
+    return ContextWindowClient
+
+
 def mcp_call_arguments(params: Any) -> dict[str, Any]:
     """Forward only concrete MCP arguments; omitted optional fields must stay omitted."""
     return params.model_dump(exclude_none=True)
@@ -878,10 +895,13 @@ async def __aexit__(self, exc_type: Any, exc_val: Any, exc_tb: Any) -> None:
         "temperature": float(config["temperature"]),
         "top_p": float(config["top_p"]),
     }
-    client = ChatCompletionsClient(
+    client_class = ChatCompletionsClient
+    if config.get("context_window_tokens"):
+        client_class = context_window_client(ChatCompletionsClient, int(config["context_window_tokens"]))
+    client = client_class(
         model=config["policy_model"],
         base_url=config["model_base_url"],
-        api_key="unused",
+        api_key="unused",  # pragma: allowlist secret
         max_tokens=int(config["max_output_tokens"]),
         kwargs=model_kwargs,
     )

responses_api_agents/apex_agent/tests/test_app.py

@@ -55,7 +55,7 @@ def _agent(
         image_build={
             "enabled": auto_build,
             "source_repo": "https://github.com/Mercor-Intelligence/archipelago.git",
-            "source_revision": "0cb5c476c219a9df637e0bd37fb86b2361f4ab89",
+            "source_revision": "0cb5c476c219a9df637e0bd37fb86b2361f4ab89",  # pragma: allowlist secret
             "source_root": None,
             "source_github_token": None,
             "dockerfile": "environment/Dockerfile",
@@ -240,6 +240,15 @@ def test_sandbox_config_never_contains_verifier_secrets() -> None:
     assert "FOUNDRY_LOCAL_ROOT" not in spec.env
 
 
+def test_sandbox_config_passes_the_policy_context_window_when_set() -> None:
+    agent = _agent()
+    agent.config.context_window_tokens = 262_144
+    runner = json.loads(agent._sandbox_spec(_body(), "Do the work").files["/app/apex-gym/runner_config.json"])
+
+    assert runner["context_window_tokens"] == 262_144
+    assert runner["max_output_tokens"] == 32_768
+
+
 def test_sandbox_config_propagates_text_only_model_capability() -> None:
     spec = _agent(supports_vision=False)._sandbox_spec(_body(), "Do the work")
     runner = json.loads(spec.files["/app/apex-gym/runner_config.json"])

responses_api_agents/apex_agent/tests/test_stirrup_runtime.py

@@ -1,12 +1,14 @@
 # SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved.
 # SPDX-License-Identifier: Apache-2.0
 
+import asyncio
 import json
 import zipfile
 from inspect import getsource
 from pathlib import Path
 from types import SimpleNamespace
 
+import pytest
 from pydantic import BaseModel
 
 from responses_api_agents.apex_agent import stirrup_runtime
@@ -18,6 +20,40 @@ def test_stirrup_repackages_image_tool_responses_for_chat_completions() -> None:
     assert "text_only_tool_responses=True" in source
 
 
+def test_context_window_client_reports_the_window_and_keeps_the_output_cap() -> None:
+    ChatCompletionsClient = pytest.importorskip("stirrup.clients.chat_completions_client").ChatCompletionsClient
+
+    client_class = stirrup_runtime.context_window_client(ChatCompletionsClient, 262_144)
+    client = client_class(
+        model="policy",
+        base_url="http://127.0.0.1:9/v1",
+        api_key="unused",  # pragma: allowlist secret
+        max_tokens=32_768,
+    )
+    request: dict = {}
+
+    async def create(**kwargs):
+        request.update(kwargs)
+        raise RuntimeError("request captured")
+
+    client._client.chat.completions.create = create
+    with pytest.raises(RuntimeError, match="request captured"):
+        asyncio.run(client.generate([], {}))
+
+    assert isinstance(client, ChatCompletionsClient)
+    assert client.max_tokens == 262_144
+    assert request["max_completion_tokens"] == 32_768
+
+
+def test_policy_client_uses_the_context_window_only_when_configured() -> None:
+    source = getsource(stirrup_runtime.run_stirrup_rollout)
+
+    assert 'if config.get("context_window_tokens"):' in source
+    assert 'context_window_client(ChatCompletionsClient, int(config["context_window_tokens"]))' in source
+    assert "client = client_class(" in source
+    assert 'max_tokens=int(config["max_output_tokens"])' in source
+
+
 def test_text_only_model_replaces_tool_images_and_preserves_other_content() -> None:
     class FakeImage:
         pass