NVIDIA NIM reasoning models: chat_template_kwargs never injected, breaking streamed reasoning_content and causing inconsistent latency

[opassoca]

Root cause and fix for inconsistent streaming and missing reasoning traces when running NIM backed reasoning models (deepseek-ai/deepseek-v4-pro, z-ai/glm-5.2) through OpenClaude.

Symptom: response text arrives in irregular bursts, sometimes near instant, sometimes stalling for a second or more, and the reasoning or thinking block that NIM is capable of streaming never renders in the CLI, even on models where it should.

I reverse engineered the shipped bundle (@gitlawb/openclaude, dist/cli.mjs, npm build dated Jul 7 2026, 21MB minified) rather than guessing. chat_template_kwargs and enable_thinking have zero occurrences anywhere in that file. Both wire formats that cover NIM routed models, deepseek_compatible and zai_compatible, only ever populate body.thinking or body.reasoning_effort in the OpenAI compatible request shim. Neither path ever sets chat_template_kwargs, which is the parameter the NIM template layer actually reads to gate bounded, streamed reasoning_content. Absent that signal, the upstream model still reasons, just without a budget or streaming contract on the NIM side, which lines up exactly with the erratic delivery and the silent drop of the reasoning trace.

The fix is a single conditional inserted immediately after the existing reasoning_effort handling in both wire format branches: when request.baseUrl contains integrate.api.nvidia.com, set body.chat_template_kwargs equal to thinking true and enable_thinking true, before the request is dispatched. Verified against deepseek-ai/deepseek-v4-pro on NIM: streaming is now continuous and the reasoning trace renders correctly, matching the documented contract for the DeepSeek and GLM builds on NIM.

This is currently a runtime patch against the minified dist bundle only, since the npm package ships without the matching TypeScript source for the provider shim. It needs to land in the actual openai-compat shim source to survive the next release. The exact patch, anchor string and injected snippet, is ready to hand off if a maintainer wants to take it from here.