Fit OpenCode generation budget to context

[bingran-you]

Summary

• measure the irreducible OpenCode prompt before reserving completion tokens
• reduce only the per-turn generation allowance when the prompt needs part of that reserve
• reject prompts only when no generation token can fit
• prevent extra_body from overriding enforced max_tokens, n, or logprobs
• cover the real 12,306-token GRPO regression plus exact boundary and full-evaluation-context cases

Verification

• pytest pipelines/benchflow-task-posttrain/tests -q (228 passed)
• ruff check pipelines/benchflow-task-posttrain/src pipelines/benchflow-task-posttrain/tests
• python -m compileall -q pipelines/benchflow-task-posttrain/src
• git diff --check

[chatgpt-codex-connector]

💡 Codex Review

Here are some automated review suggestions for this pull request.

Reviewed commit: 3dbf7d80e2

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".

[chatgpt-codex-connector]

Honor max_tokens carried in extra_body

When a client sends the completion cap only in extra_body (for example the request shape added in the new tests, but with a smaller value), this filter drops max_tokens before the cap is read from only top-level max_completion_tokens/max_tokens. Those requests now fall back to config.max_tokens_per_call (4096 by default), so a caller asking for a short generation can be forwarded to TRL with thousands of tokens and exceed the expected per-call or idle budget; the reserved value should still be consumed as the requested cap before it is omitted from generation_kwargs.

Useful? React with 👍 / 👎.