✅ Done! I reviewed the PR against the authoritative source code in

software-agent-sdk

and posted an approval ✅ Here's a summary of my review of github.com/OpenHands/docs/pull/788: Verified accurate against source: • Both

/v1/chat/completions

and

/v1/responses

run the full OpenHands agent (not a thin LLM wrapper), wait for run completion, and return only the final assistant text. •

/v1/responses

is correctly documented as stateless-first —

run_response

always passes

conversation_id=None

, and the response-only

X-OpenHands-ServerConversation-ID

header is confirmed. • The rejection status codes and exact wording (400s for

previous_response_id

,

store: true

,

stream: true

, plus no

GET /v1/responses/{id}

) all match

service.py

precisely. • The "ignored fields" (

tools

,

tool_choice

,

parallel_tool_calls

,

temperature

, etc.) are confirmed as dropped via

ConfigDict(extra="ignore")

. • The auth mapping and the corrected Chat Completions streaming claim are both accurate. One non-blocking nit I raised: the Python replay snippet's

cast(list[ResponseInputItemParam], first.output)

is a type-only cast; it works because the objects are

ResponseOutputMessage

Pydantic models, but a one-line comment would help future readers. Overall the PR is accurate, well-organized, and the exact status codes/wording add real value — so I approved it. 🎉

✅ Done! Reviewed and approved ✅ github.com/OpenHands/software-agent-sdk/pull/4954 The removal is clean: the

modify_params

field, its

Field(deprecated=...)

metadata, and the

_coerce_inputs

warning +

warn_deprecated

import are all gone from

llm.py

, while process-wide

litellm.modify_params = True

is preserved. The schema migrations (agent settings 5→6, persisted settings 2→3) are correct and well-documented, backward-compat deserialization is covered by the new

test_validate_agent_settings_migrates_v5_modify_params

+ retained legacy fixture, and CI is green. No concerns. This review was created by an AI agent (OpenHands) on behalf of the requesting user.