fix(zai): set tool_stream=true to avoid api.z.ai 30s idle timeout

[aarongxa]

Summary

GLM models (4.6, 4.7, 5, 5.1) batch tool_call argument generation in a single chunk by default. When the tool args are large (long file writes, big shell commands), the model goes silent for 30+ seconds while generating them, and api.z.ai's 30-second server-side idle timeout kills the connection with ECONNRESET — surfacing as a "timeout" inside Hermes mid-tool-call.

Z.AI's documented fix is to send tool_stream=true alongside stream=true. That makes the model emit tool args incrementally and keeps chunks flowing within the 30s window.

What this changes

In _build_api_kwargs (chat-completions branch), inject extra_body["tool_stream"] = True whenever:

• self.provider == "zai", OR the base URL contains z.ai or bigmodel.cn, AND
• self.tools is non-empty (no benefit to streaming tool args when no tools are bound)

A user-supplied extra_body.tool_stream wins via setdefault, so it's overrideable.

Why it's safe

• Gated by provider/base-URL — no behavioral change for OpenRouter, Anthropic, OpenAI, Ollama, etc.
• Gated by self.tools — never sent on toolless calls.
• Z.AI ignores unknown extra_body fields gracefully on routes that don't support it (legacy GLM-4.5 and earlier), so even an over-broad match would be a no-op there. Per Z.AI docs, tool_stream is supported on GLM-4.6, 4.7, 5, and 5.1 — i.e. every currently-shipping GLM coding model.
• setdefault preserves any explicit override.

Test plan

• pytest tests/run_agent/test_run_agent.py -k "ZaiToolStream or zai or glm" — 8 passed (5 new + 3 pre-existing)
• pytest tests/run_agent/test_provider_parity.py tests/run_agent/test_strict_api_validation.py — 80 passed
• New tests cover: zai provider, api.z.ai global URL, bigmodel.cn URL, no-tools omission, non-zai omission

References

• Z.AI docs — Tool Streaming Output
• vercel/ai#12949 — root-cause analysis of the 30s timeout
• opencode#15350 — same issue under a different client

[teknium1]

Thanks for identifying the Z.AI idle-timeout failure mode. The requested wire behavior is still absent on current main, but this patch targets a request-builder body that was extracted after the branch was created.

Problems

• run_agent.py:5328-5331 is now only a forwarder to agent.chat_completion_helpers.build_api_kwargs (extracted in 4b25619bc4770396faf206429ddc180ad02231a9), so the proposed insertion point is no longer live.
• Direct Z.AI requests resolve through the registered profile path (agent/chat_completion_helpers.py:849-888; agent/transports/chat_completions.py:316-321). ZaiProfile.build_api_kwargs_extras currently emits thinking and reasoning_effort, but not tool_stream (plugins/model-providers/zai/__init__.py:88-108).

Suggested changes

• Salvage this into the current profile-aware transport path, preserving the tools-only gate and explicit request_overrides.extra_body precedence (agent/transports/chat_completions.py:537-606).
• Cover both direct provider="zai" and the PR's claimed custom Z.AI endpoint route with current-path tests.

Automated hermes-sweeper review.