Add renderer chat template kwargs passthrough

[eligotts]

Summary

• Add chat_template_kwargs to create_renderer and create_renderer_pool.
• Let DefaultRenderer forward arbitrary template kwargs, while model-specific renderers declare the kwargs they support and fail fast on unsupported keys.
• Wire enable_thinking and reasoning_effort for the renderers that model those template controls.

We are extracting renderer template controls from chat_template_kwargs for now. Longer term, we should move this toward typed pydantic config models with explicit renderer-specific template args, so users cannot dump arbitrary keys into config. Until then, unsupported kwargs fail at renderer construction.

Companion PRs

• Prime RL: Use sampling chat template kwargs for renderer RL prime-rl#2605
• Verifiers: Pass renderer chat template kwargs from sampling verifiers#1447

Validation

• uvx ruff check renderers/base.py renderers/qwen3.py renderers/qwen35.py renderers/nemotron3.py renderers/gpt_oss.py renderers/glm5.py renderers/glm45.py renderers/deepseek_v3.py renderers/kimi_k2.py renderers/kimi_k25.py renderers/laguna_xs2.py tests/test_chat_template_kwargs.py
• PYTHONPATH=. uvx --with pytest --with transformers --with openai-harmony --with tiktoken --with numpy --with jinja2 --with fastokens --with openai pytest tests/test_chat_template_kwargs.py

Note

Add chat_template_kwargs passthrough to renderer constructors

• Adds a chat_template_kwargs: dict[str, Any] | None parameter to create_renderer and create_renderer_pool in base.py, forwarding fixed template toggles (e.g. enable_thinking, reasoning_effort) directly to renderer constructors.
• Each renderer class now declares a CHAT_TEMPLATE_KWARGS class attribute listing its supported keys; create_renderer validates against this set and raises ValueError for unsupported or reserved constructor kwargs.
• Supported renderers updated: Qwen3, Qwen35, DeepSeekV3, GLM4.5, GLM5, GptOss, KimiK25, LagunaXS2, Nemotron3.
• Behavioral Change: passing a key that overlaps with renderer constructor kwargs (e.g. tool_parser) inside chat_template_kwargs now raises ValueError instead of silently overriding constructor behavior.

^{Macroscope summarized b543277.}

---

Note

Medium Risk
Moderate risk because it changes renderer construction and argument validation, which can break existing integrations that pass unexpected kwargs or rely on permissive behavior.

Overview
Adds a new chat_template_kwargs parameter to create_renderer and create_renderer_pool, binding chat-template controls at renderer construction time and forwarding them into DefaultRenderer/model-specific renderer constructors.

Introduces validation that (1) rejects reserved constructor kwargs inside chat_template_kwargs and (2) enforces per-renderer allowlists via each renderer’s new CHAT_TEMPLATE_KWARGS attribute (e.g. enable_thinking for several model renderers, reasoning_effort for gpt-oss), raising ValueError on unsupported keys; includes new unit tests and README documentation for the behavior.

^{Reviewed by Cursor Bugbot for commit b543277. Bugbot is set up for automated code reviews on this repo. Configure here.}

[macroscopeapp]

Approvability

Verdict: Approved

Additive feature adding an optional chat_template_kwargs parameter with proper validation to prevent misuse. Backwards compatible, well-tested, and self-contained with no impact to existing behavior.

^{You can customize Macroscope's approvability policy. Learn more.}

[cursor]

Cursor Bugbot has reviewed your changes and found 1 potential issue.

There are 2 total unresolved issues (including 1 from previous review).

^{❌ Bugbot Autofix is OFF. To automatically fix reported issues with cloud agents, enable autofix in the Cursor dashboard.}

^{Reviewed by Cursor Bugbot for commit 7fbf390. Configure here.}