Pipe chat_template_kwargs into typed RendererConfig

[eligotts]

Summary

Pipe sampling_args.extra_body.chat_template_kwargs through to the typed RendererConfig so users can set enable_thinking etc. in sampling and have it land on the right renderer — no [renderer] block needed.

• Pop chat_template_kwargs from sampling in RendererClient.get_native_response.
• Resolve None / AutoRendererConfig → concrete config via MODEL_RENDERER_MAP before merging, so kwargs validate against the actual renderer's schema (Qwen3RendererConfig.enable_thinking) instead of AutoRendererConfig's intentionally-minimal one. Mirrors renderers._resolve_auto's preserve_* carry.
• Validation comes for free: typed configs (extra="forbid") reject unknown keys with a field-path error; DefaultRendererConfig (extra="allow") keeps the escape hatch for arbitrary jinja kwargs.
• Kwargs override fields with the same name on the (resolved) base (last-write-wins, matches MITO's per-request override semantics).
• Pool cache already keys on effective renderer_config_json, so identical kwargs across requests reuse a single renderer.

Replaces #1447 (pre-typed-config era).

Effective truth table

ClientConfig.renderer_config	chat_template_kwargs	Result
Qwen3RendererConfig(...) (explicit)	{enable_thinking: False}	merges → pool built with enable_thinking=False
Qwen3RendererConfig(...) (explicit)	{enable_thinkng: False} (typo)	ValueError(... Qwen3RendererConfig ... extra_forbidden)
AutoRendererConfig() (prime-rl default)	{enable_thinking: False}	auto-resolves Qwen3 from MODEL_RENDERER_MAP[model], then merges
None (verifiers eval CLI default)	{enable_thinking: False}	same as above
any	none	unchanged from main

Companion PRs

• prime-rl: nothing required — already passes renderer + sampling_args through unchanged.
• renderers: nothing required — MODEL_RENDERER_MAP + config_from_name + extra="forbid" / extra="allow" cover everything we need.

Test plan

• PYTEST_DISABLE_PLUGIN_AUTOLOAD=1 uv run pytest tests/test_renderer_client.py -k "honors_configured_renderer_config or renderer_model_name_override or merges_chat_template_kwargs or rejects_invalid_chat_template or auto_resolves_for_chat_template" — 5 passed
• uvx ruff check verifiers/clients/renderer_client.py tests/test_renderer_client.py — clean

🤖 Generated with Claude Code

---

Note

Medium Risk
Changes per-request renderer pool selection and sampling param forwarding; mis-merged kwargs could affect tokenization across rollouts, though behavior is covered by new unit tests.

Overview
RendererClient now treats sampling_args.extra_body.chat_template_kwargs as renderer configuration instead of engine sampling: they are removed from params sent to /generate and merged into the effective RendererConfig used when building or selecting a shared renderer pool.

A new _resolve_renderer_config path resolves None / AutoRendererConfig to the concrete config for the tokenizer model (via MODEL_RENDERER_MAP, honoring ClientConfig.renderer_model_name when set), carries preserve_* fields from auto config, then applies kwargs with pydantic validation (unknown keys fail on strict configs). _get_renderer_or_pool accepts a per-request resolved config so pool cache keys reflect merged settings.

Tests cover explicit/auto/None bases, renderer_model_name override during auto-resolution, stripping kwargs from forwarded sampling params, and ValidationError on typos.

^{Reviewed by Cursor Bugbot for commit 14ab2a7. Bugbot is set up for automated code reviews on this repo. Configure here.}

Note

Pipe chat_template_kwargs from extra_body into a typed RendererConfig before generation

• Extracts chat_template_kwargs from sampling_params.extra_body in RendererClient.get_native_response and removes them before forwarding to /generate.
• Introduces _resolve_renderer_config in renderer_client.py to merge kwargs into a concrete RendererConfig, resolving Auto/None bases via MODEL_RENDERER_MAP and config_from_name.
• When ClientConfig.renderer_model_name is set, auto-resolution uses that override instead of the request model name.
• _get_renderer_or_pool now accepts an optional renderer_config parameter so the resolved config drives pool selection.
• Risk: unknown chat_template_kwargs keys now raise a pydantic ValidationError before generation, which is a breaking change for callers passing unrecognized keys.

^{Macroscope summarized 14ab2a7.}

[cursor]

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

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

^{Reviewed by Cursor Bugbot for commit dd02a30. Configure here.}

[macroscopeapp]

Approvability

Verdict: Needs human review

This PR introduces new feature capability by piping chat_template_kwargs into typed RendererConfig, enabling per-request configuration overrides with auto-resolution logic. Changes runtime behavior by creating new configuration propagation paths and validating against concrete renderer schemas.

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

[mikasenghaas]

as discussed, approving this to unblock quickly. short-term cleanup would be to expose resolve_renderer(model_name: str) -> RendererConfig from renderers so we can resolve 1. the renderer 2. the chat template kwargs in the prime-rl configs