fix(v1): forward sampling parameters across dialects

[xeophon]

Overview

Forward eval sampling parameters consistently across Chat Completions, Responses, Anthropic Messages, and training requests.

Details

• Flatten extra_body before building provider wire arguments, with explicit sampling fields taking precedence.
• Preserve provider-specific sampling fields while retaining protocol-specific max-token and reasoning mappings.
• Record service_tier in effective sampling metadata.
• Avoid conflicting Chat Completions max-token aliases when an eval supplies the override.

---

Note

Medium Risk
Changes how sampling overrides are merged into provider requests, which can alter generation (max tokens, reasoning, extra_body). Logic is localized to dialect override paths, not auth or data handling.

Overview
Eval sampling now reaches Chat Completions, Responses, Anthropic Messages, and training requests through a shared SamplingConfig.wire_args() helper. That flattens OpenAI-style extra_body into top-level wire args (typed knobs win), so extra provider fields are no longer dropped.

Dialect apply_overrides paths merge those flattened args instead of a small hardcoded subset, while still remapping protocol-specific keys (max_tokens → max_output_tokens, reasoning_effort into reasoning/output_config). Chat Completions drops conflicting max_tokens/max_completion_tokens aliases when the eval supplies either. service_tier is recorded in each dialect’s sampling metadata.

^{Reviewed by Cursor Bugbot for commit b4d847b. Bugbot is set up for automated code reviews on this repo. Configure here.}

Note

Forward sampling parameters across v1 dialects via wire_args

• Adds SamplingConfig.wire_args() to flatten non-None fields, lift extra_body values, and resolve max_tokens vs max_completion_tokens conflicts.
• Updates TrainClient and all dialects (Anthropic, Chat, Responses) to use wire_args() instead of manual model_dump logic.
• Adds service_tier to sampling_fields for all dialects.
• Improves dialect-specific handling: merges output_config/reasoning_effort for Anthropic and Responses, and maps max_tokens to max_output_tokens for Responses.
• Behavioral Change: AnthropicDialect now preserves body temperature and top_p unless explicitly overridden; ChatDialect drops both max_tokens and max_completion_tokens from the body when any max token override is provided.

^{Macroscope summarized b4d847b.}

[chatgpt-codex-connector]

💡 Codex Review

Here are some automated review suggestions for this pull request.

Reviewed commit: 02fd7a62d9

ℹ️ About Codex in GitHub

Your team has set up Codex to review pull requests in this repo. Reviews are triggered when you

• Open a pull request for review
• Mark a draft as ready
• Comment "@codex review".

If Codex has suggestions, it will comment; otherwise it will react with 👍.

Codex can also answer questions or update the PR. Try commenting "@codex address that feedback".

[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 ad95f52. Configure here.}

[macroscopeapp]

Approvability

Verdict: Not approved

Macroscope's review found this PR not approvable — This PR changes effective generation requests across three provider dialects and the training renderer by forwarding arbitrary sampling fields, flattening extra_body, and remapping token and reasoning settings. These cross-path runtime changes can materially alter model output and parameter precedence, warranting human review.

^{You can add or adjust custom eligibility rules. Learn more.}