CAI hangs indefinitely when a single turn issues multiple tool_calls (not covered by #370 idle-timeout)

[mallangmollangquokka]

Environment

• cai-framework version: 0.5.10 (latest on PyPI, confirmed no newer version available)
• OS: Windows 11 + WSL2 (Ubuntu 24.04, native ext4 filesystem, not /mnt/c)
• Model: deepseek/deepseek-v4-flash and deepseek/deepseek-v4-pro (both reproduced)
• Agent: default one_tool_agent

Summary

When the model returns multiple tool_calls in a single assistant turn, CAI sometimes hangs indefinitely with no output, no error, and no timeout — even well past the 10s idle-kill added in #360 / PR #370. The process just sits there until an external wrapper timeout (we used a 500s timeout wrapper) kills it.

I want to flag upfront: I have not isolated a minimal reproduction case, and I'm not 100% sure this is a distinct bug vs. some interaction with our specific setup (WSL2, DeepSeek provider, long-running session with many prior turns). Posting in case it's useful signal / matches something you're already tracking.

What we observed

Across ~9 separate long-running sessions solving a CTF challenge, the hang consistently coincided with turns where the assistant issued 2+ tool_calls at once, e.g.:

• Two generic_linux_command calls in one turn (ls -la /tmp/ + ls -la ~/challenge/) — hung ~500s, no idle-kill message, no output.
• A generic_linux_command using a python3 << 'EOF' ... EOF heredoc alongside another call in the same turn — same symptom.

By contrast, single-command turns (including ones that legitimately trigger the existing idle-timeout, e.g. find / scanning a slow mount) behaved correctly and printed [Terminated: idle 10s, likely waiting for input] as expected from PR #370.

This suggests whatever coordinates multiple tool_calls within one turn may not be covered by the per-command idle-detection added in #370 (which looks like it's scoped to a single command's I/O loop).

Expected behavior

Either:

• The same idle-detection/timeout protection applies uniformly regardless of how many tool_calls are in a turn, or
• Multiple tool_calls in one turn are executed strictly sequentially with each one individually protected by the existing idle timeout.

Notes

• Confirmed unrelated to /mnt/* (9p) slowness — reproduced on pure native WSL2 filesystem.
• Confirmed not model-specific to us, though we only tested DeepSeek; noting CAI hangs when using Anthropic provided Claude model #401 ("CAI hangs when using Anthropic provided Claude model") looked like it could be the same underlying class of issue with a different provider.
• Happy to share full JSONL session logs if useful.

Thanks for the great project — apologies in advance if this turns out to be a config issue on our end rather than a CAI bug.

[mumtajali788m-design]

Dilshad