Consume the typed RendererConfig surface

[hallerite]

Summary

Switches ClientConfig and RendererClient to consume the typed
renderers.RendererConfig discriminated union (rendered in
PrimeIntellect-ai/renderers#60).

• Removes the flat renderer, tool_parser, reasoning_parser, preserve_* fields from ClientConfig.
• Adds a single renderer_config: RendererConfig | None field. Pydantic dispatches on the name discriminator ("qwen3", "glm-5", …) and exposes exactly the kwargs that renderer supports.
• renderer_model_name / renderer_pool_size stay as separate fields — they're orthogonal to the renderer choice.
• _get_renderer_or_pool collapses to create_renderer_pool(model, config). Cache key shrinks from a 7-tuple of flat fields to (model, pool_size, config_json).

API

Before:

ClientConfig(
    client_type="renderer",
    renderer="qwen3.5",
    tool_parser=None,
    reasoning_parser=None,
    preserve_all_thinking=True,
)

After:

from renderers import Qwen35RendererConfig

ClientConfig(
    client_type="renderer",
    renderer_config=Qwen35RendererConfig(
        enable_thinking=False,
        add_vision_id=True,
        preserve_all_thinking=True,
    ),
)

Bogus combinations (e.g. add_vision_id under name="qwen3") raise a pydantic.ValidationError at config-load time instead of at renderer construction.

Stacking

Pins renderers to a git SHA on the renderers PR's head while it's in review. When that lands and a new dev release ships, this PR bumps to the PyPI release.

• Depends on: Typed renderer configs renderers#60
• Supersedes: Pass renderer chat template kwargs from sampling #1447 (which wired chat_template_kwargs through sampling.extra_body)

Validation

Static-only at the moment — the renderers branch hasn't shipped to PyPI, so the verifiers venv can't be re-synced to run pytest against the new typed surface yet.

• uv run ruff check ... — clean
• Test fixtures in test_renderer_client.py updated to construct typed configs; will run on CI after the renderers dep resolves.

Test plan

• CI runs once renderers typed-config branch is available (either merged + dev release, or the git pin resolves in the lockfile)
• Manual: round-trip a sample ClientConfig through TOML → pydantic → _get_renderer_or_pool and confirm the typed config flows through

---

Note

Medium Risk
Breaking change to renderer ClientConfig shape and shared-pool cache keys; rollout configs must migrate before deploy, and pool reuse behavior may differ for equivalent old flat settings.

Overview
Replaces flat renderer client settings with a typed renderers.RendererConfig object on ClientConfig, and threads that object through RendererClient pool creation.

ClientConfig drops renderer, tool_parser, reasoning_parser, preserve_all_thinking, and preserve_thinking_between_tool_calls in favor of renderer_config: RendererConfig | None (optional import when verifiers[renderers] is not installed). RendererClient now calls create_renderer_pool(model, renderer_config, …) and keys shared pools on (renderer_model_name, pool_size, renderer_config.model_dump_json()) instead of the old seven-field tuple—so pool sharing follows the full serialized config, not separate parser/preserve flags.

Dependency and test alignment: renderers is pinned to >=0.1.8.dev28 (with lockfile updates for fastokens, prime-pydantic-config, and removal of tyro/typeguard transitives). Tests construct typed configs (e.g. Qwen3VLRendererConfig) and use config_from_name() with create_renderer. E2E ScriptedVLLM returns an httpx-like response whose .content is JSON bytes for parse_generate_response.

Migration: configs must use discriminated renderer config objects (e.g. Qwen35RendererConfig(...)) instead of string renderer plus loose kwargs; invalid field combinations fail at Pydantic validation time.

^{Reviewed by Cursor Bugbot for commit fa4652a. Bugbot is set up for automated code reviews on this repo. Configure here.}

Note

Replace renderer string fields with typed RendererConfig in ClientConfig

• Replaces the renderer: str, tool_parser, reasoning_parser, preserve_all_thinking, and preserve_thinking_between_tool_calls fields in ClientConfig with a single renderer_config: RendererConfig | None field from the renderers package.
• Updates RendererClient._get_renderer_or_pool to pass the typed config object directly to create_renderer_pool, keying the shared pool cache on (renderer_model_name, pool_size, renderer_config_json) instead of the previous 7-tuple.
• Updates test utilities and e2e helpers to call config_from_name(renderer_name) when constructing renderers from a string name.
• Behavioral Change: requests that previously shared a renderer pool based on matching parser/preserve flags will now share pools based on the serialized renderer_config JSON, which may change pool-sharing behavior for existing configurations.

Changes since #1467 opened

• Made renderers package an optional dependency for verifiers.types module by conditionally importing RendererConfig [f322bc9]
• Changed renderers dependency specification from Git VCS URL to PyPI version constraint in optional dependency groups [1a406bf]
• Reformatted configuration key assignment in RendererClient [6c6af70]
• Updated dependency lock file [6c6af70]
• Updated renderers package dependency constraint to version >=0.1.8.dev28 [d107442]
• Introduced _ScriptedResponse class and modified ScriptedVLLM.post method to return response objects [fa4652a]

^{Macroscope summarized 5727298.}

[macroscopeapp]

Approvability

Verdict: Needs human review

This PR makes a breaking API change to ClientConfig, removing 5 fields (renderer, tool_parser, reasoning_parser, preserve_all_thinking, preserve_thinking_between_tool_calls) and replacing them with a typed renderer_config from an external dependency. Breaking interface changes warrant human review.

^{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 1a406bf. Configure here.}