Skip to content
TheChatArchiveThe Chat Archive
gpt-5.6-sol· openaiHugging Face ibragim-bad/swe_rebench_07_2026_trajectories

Subagent toolcalls not handled when using LiteLLM proxy Summary When using cagent with a…

  1. User

    Subagent tool_calls not handled when using LiteLLM proxy

    Summary

    When using cagent with a LiteLLM proxy (OpenAI-compatible API), a subagent can return a valid tool_calls response, but the host agent reports “no response from subagent” and does not execute the tool call.

    Environment

    • cagent: v1.19.7
    • OS: Linux (WSL2) x86_64
    • LLM: gemini-pro via LiteLLM proxy (OpenAI-compatible)

    Expected

    Host agent receives the subagent response and executes the tool call(s), then continues the conversation with the tool result.

    Actual

    Host agent reports there is “no response from subagent” even though the subagent response contains a valid tool_calls entry. The requested tool call is not executed.

    Minimal reproduction

    1. Configure cagent to use LiteLLM as the model provider (OpenAI-compatible API).
    2. Run a workflow where the subagent is asked to use tools.
    3. The subagent returns a tool call (e.g., company_search with JSON arguments) and finishes with finish_reason: "stop".
    4. The host agent logs “no response from subagent” and does not invoke the tool.

    Notes

    • The subagent response includes a valid tool call object (type: function, name: company_search, arguments: {"datasources": ["confluence", "github", "gdrive"], "query": "LiteLLM usage teams integrations standards"}).
    • I can reproduce consistently with LiteLLM as the proxy in front of an OpenAI-compatible endpoint.

    Please let me know if you want a sanitized example payload or a minimal agent config.

    Interface

    No new interfaces are introduced.

  2. Assistant

    I’ll trace the OpenAI-compatible response conversion and subagent execution path, reproduce the dropped tool_calls case in a focused test, then patch and run the project validations.

  3. Assistant

    rg isn’t installed in this workspace, so I’m using the repository’s available search tools and narrowing the trace from the exact error message.

  4. Assistant

    The failure is in stream aggregation, not delegation itself: a chunk carrying both delta.tool_calls and finish_reason: "stop" is returned immediately before its tool-call delta is processed. That matches LiteLLM’s response shape. I’m adding a regression at this boundary and reordering terminal handling so the call is retained and normalized to tool_calls.

  5. Assistant

    The focused regression is now running. The implementation also preserves content/reasoning if a proxy coalesces those deltas into the terminal chunk, instead of discarding them alongside the tool call.

  6. Assistant

    The first compile attempt hit the container’s process/thread limit while Go was building the newly downloaded dependency graph. I’m rerunning with constrained build parallelism; this is an environment resource failure, not a test failure.

  7. Assistant

    The focused regression and the full runtime package now pass. Next I’m running the repository-required build, full test suite, and linter with low Go parallelism to stay within the container’s process ceiling.

  8. Assistant

    The repository’s task runner is not installed in this environment. I’m reading the Taskfile targets and invoking their underlying local commands directly so the same checks are still performed.

  9. Assistant

    The full suite is still progressing cleanly under the constrained runner; the earlier command output was only the dependency/setup phase. I’m continuing until the process exits and reports its final status.

  10. Assistant

    The fix-specific tests pass and the binary builds. The repository-wide suite completes but has pre-existing environment/baseline failures unrelated to these files (stale example model IDs, root-only permission expectations, and hook fixture output); I’ve verified the changed runtime path independently and am doing the final diff check now.

Responses were produced in different, uncontrolled contexts (system prompts, settings, prior turns). This is not a benchmark.