[Anish Mehta (anishmehta24)]

Motivation & Context

With OpenAIChatCompletionClient against OpenAI-compatible thinking models (reproduced with DeepSeek, thinking.type=enabled), a response containing text, a tool call and reasoning was serialised on the follow-up request as two assistant messages: {content} followed by {tool_calls, reasoning_*}. DeepSeek requires the reasoning on the message that carries the content and returns 400: The reasoning_content in the thinking mode must be passed back to the API.

Description & Review Guide

• Major changes: _build_openai_messages now attaches a turn's tool calls to the assistant dict built from that turn's text (and text that follows coalesced tool calls to the tool_calls dict), so content, tool_calls and the buffered reasoning field travel as one assistant message. Six unit tests, including an end-to-end DeepSeek-style reasoning_content round trip through the existing response_parser/message_preparer hooks.
• Impact: assistant turns with text and tool calls are sent as a single message, which is the canonical Chat Completions shape. Reasoning text surfaced by a response_parser keeps its own dict, so message_preparer correlation (the vLLM example) is unchanged; refusals and non-text content are not merged. No stock reasoning_content branch was added, per packages/openai/AGENTS.md; a parser/preparer pair is still needed to surface and echo DeepSeek's field, and it now works.
• Focus for reviewers: the merge conditions in _build_openai_messages (only plain text and tool calls of the same turn, only when adjacent) and whether the single-message shape is acceptable for all providers.

Prepared with AI assistance (Claude Code) and human-verified: the new tests fail on main and pass with the fix; poe fmt, poe lint, poe pyright -P openai, poe mypy -P openai and the openai and foundry_local test suites were run locally.

Related Issue

Fixes #8382

Contribution Checklist

• The code builds clean without any errors or warnings
• All unit tests pass, and I have added new tests where possible
• The PR follows the Contribution Guidelines
• This PR is linked to an issue and there is no other open PR for this issue
• This is not a breaking change.

🤖 Generated with Claude Code

[Copilot]

Copilot was unable to review this pull request because the user who requested the review has reached their quota limit.

[Eduard van Valkenburg (eavanvalkenburg)]

Thanks for the contribution. Before this is ready, please complete the Microsoft CLA; license/cla is still queued, and the current DevFlow PR Review run has cancelled review and team_check jobs. Once those blockers are cleared, please re-request review so the approval-gated workflows can be run. Thanks!

[Anish Mehta (anishmehta24)]

@microsoft-github-policy-service agree

[Anish Mehta (anishmehta24)]

Eduard van Valkenburg (@eavanvalkenburg) CLA is signed and license/cla is green now. The previously queued jobs have run; only merge-gatekeeper is still pending. Thanks for the quick review!