feat(pr-review): add llm-extra-body input for provider-specific request fields

[arpit-acceleratedata]

Summary

Adds an optional llm-extra-body input to the pr-review action. It is a JSON object forwarded to the SDK LLM as litellm_extra_body.

Why

Some models think by default and expose the switch only as a request-body field. Qwen3.8 on Alibaba Model Studio is one: thinking is on by default and is controlled by enable_thinking or thinking_budget (docs). The action only passes model, key, and base URL, so there was no way to send it. On our repo a review with openai/qwen3.8-flash made ~28 model calls with full reasoning and hit the 20 minute job timeout without posting.

With llm-extra-body: '{"enable_thinking": false}' the same call drops from ~280 s to ~12 s in a direct API test.

Changes

• action.yml: new optional input llm-extra-body (default ''), passed as LLM_EXTRA_BODY to the run step.
• scripts/agent_script.py: parse LLM_EXTRA_BODY as JSON; must be an object; set litellm_extra_body. Invalid input exits with a clear error.
• README.md: input table row.

Notes

• Ignored in ACP mode, same as llm-base-url.
• No behaviour change when the input is empty.

[all-hands-bot]

👋 This PR needs a couple of things fixed before OpenHands can review it:

• the PR description's HUMAN: section needs at least 20 characters describing what you tested, not just the template placeholder

Push an update once this is addressed and this check re-runs automatically.

This is an automated check - no AI was used to generate this comment.

[all-hands-bot]

This review was posted by an AI agent (OpenHands).

Review of PR #544

Scope: action.yml, scripts/agent_script.py, README.md (+52/-0). The feature itself is reasonable and needed: litellm_extra_body is a real SDK field (openhands/sdk/llm/llm.py) and is the correct channel for enable_thinking / thinking_budget on Qwen-style endpoints, and max_iteration_per_run is a real Conversation parameter (default 500). JSON validation and the int / >= 1 checks on MAX_ITERATIONS are sound, and the ACP path is correctly unaffected because it returns before the OpenHands-only config is built.

Blocking

1. create_conversation now requires two new keys in every config dict and breaks a caller in this repo. The new reads use direct subscripting (config["extra_body"], config["max_iterations"]) instead of the optional-key pattern already used a few lines above (config.get("use_sub_agents", False)). validate_environment() always emits both keys, but create_conversation() is also called directly by tests with hand-built config dicts. The repository's own test tests/test_pr_review_review_context.py::test_create_conversation_uses_sdk_project_skill_loader (lines 243-249) passes a config without either key and now dies with KeyError: 'extra_body':

>       if config["extra_body"]:
E       KeyError: 'extra_body'
plugins/pr-review/scripts/agent_script.py:1107

I ran the CI command (uv run --group test pytest tests/) on this head: 1 failed, 799 passed, 24 skipped. The same test passes on origin/main, so this is a regression introduced by the PR, not pre-existing. Fix by using config.get("extra_body") / config.get("max_iterations") (consistent with use_sub_agents), or by updating the caller's dict.

Non-blocking, but the merge gate is red

2. The Validate PR description check fails on this head. Run 34249346569 reports the HUMAN: / "How to Test" section is the untouched template placeholder (the existing bot comment 5603558088 says the same). Per the repo's own gate (.github/workflows/pr-description-check.yml, .github/scripts/check_pr_description.py, REQUIRED_TEMPLATE_FIELDS), non-draft PRs must keep ## Why, ## Summary, and ## How to Test filled in. The PR body has ## Summary and ## Why but no ## How to Test, so this cannot be attached to a changed line and is reported here. The Tests, Check Extensions, and Deprecation deadlines runs are action_required (fork PR awaiting approval), so they have not yet exercised the failure in finding 1 on CI.

Notes (no action required)

• The llm-extra-body value flows from the workflow definition, not from PR content, so a fork PR cannot inject arbitrary request-body fields through it in pull_request context. Using the SDK's structured litellm_extra_body field (with drop_params: True) rather than a hand-rolled passthrough is the right call.
• README and action.yml descriptions match the implemented behavior, including the "ignored in ACP mode" caveat.

Please fix finding 1 (and fill in ## How to Test). Happy to re-review after a push.

🔄 CHANGES REQUESTED

[all-hands-bot]

Direct subscripting makes these two new keys mandatory for every caller of create_conversation(), not just validate_environment(). create_conversation() is also called with hand-built config dicts; tests/test_pr_review_review_context.py::test_create_conversation_uses_sdk_project_skill_loader (lines 243-249) omits both keys and now raises KeyError: 'extra_body'. Running the CI command uv run --group test pytest tests/ on this head gives 1 failed, 799 passed, 24 skipped; the same test passes on origin/main, so this is a regression from this PR.

Use the optional-key pattern already used a few lines above (config.get("use_sub_agents", False)):

if config.get("extra_body"):
    llm_config["litellm_extra_body"] = config["extra_body"]

The same applies to config["max_iterations"] at line 1147.