[Kh4L]

Stirrup's ChatCompletionsClient(max_tokens=...) is both the per-request output cap (sent as max_completion_tokens) and the context window the agent uses to decide when to summarize (token_usage.total / client.max_tokens, at 70%). The Apex agent passes max_output_tokens (32768), so Stirrup summarizes at ~23k tokens whatever the policy's context window is. On long tasks the agent then loops read -> summarize -> re-read until it reaches the turn cap.

• stirrup_runtime.py: context_window_client() subclasses the Stirrup client and overrides only the max_tokens property, so summarization uses the configured window while requests keep max_completion_tokens = max_output_tokens.
• app.py, configs/apex_agent.yaml: context_window_tokens (apex_agent_context_window_tokens), unset by default, which keeps the current behavior.
• README and tests.

With a 262k-context policy and apex_agent_context_window_tokens: 262144, the first summary moves from ~23k tokens to ~185k.

#3668 changes the same client setup; this commit rebased onto it is on spanev/apex-context-window-on-prebuilt-world.

[copy-pr-bot]

This pull request requires additional validation before any workflows can run on NVIDIA's runners.

Pull request vetters can view their responsibilities here.

Contributors can view more details about this message here.