mirror of
https://github.com/mudler/LocalAI.git
synced 2026-10-28 09:08:54 -04:00
Stefan Walczandlocalai-org-maint-bot
9c0266e506
* fix(openresponses): keep server-side MCP tool items on synchronous paths When a model runs MCP tools server-side, /v1/responses streams or builds function_call and function_call_output items for each call, but the two synchronous paths dropped them from the final response: - non-streaming: the output items are local to each MCP iteration, which recurses, so only the last iteration's items were returned and stored. - streaming: response.completed and the stored response were rebuilt from the last iteration (reasoning, message, unexecuted tool calls) only. The background paths already keep them. Without the trail, a follow-up with previous_response_id, or a client that rebuilds history from the completed output, never shows the model which tools ran and what they returned. Measured with gemma4-26b behind a model with mcp.auto: on 12 follow-up questions that need fresh tool data, the model called the tool 0/12 times when the history carried only text (and invented numbers), and 12/12 times when the earlier tool calls and results were included. The non-streaming handler now passes the items of earlier iterations into the recursive call; the streaming handler collects each finished MCP iteration's reasoning, function_call and function_call_output items and puts them in front of the final items in response.completed and in the store. Requests without server-side tools are unchanged. No new state: the items travel in the already replicated stored response. Tests: an in-memory MCP server behind the real LocalToolExecutor and a stubbed backend. For both synchronous paths the completed and the stored output are function_call, function_call_output, message (with reasoning in front when the model reasoned before the call), and a follow-up through ResponsesEndpoint with previous_response_id hands the backend the tool call and its result. Counter-checks: requests without tools still return a single message, and both background paths are pinned as they were. Assisted-by: Claude Code:claude-opus-5-5 Signed-off-by: Stefan Walcz <stefan.walcz@walcz.de> * fix(responses): preserve MCP trail associations Keep tool result call IDs when replaying stored responses. Clear consumed streaming state so the iteration limit cannot duplicate the final round. Add regression coverage for replay and iteration exhaustion. Assisted-by: Codex:GPT-6 Signed-off-by: Stefan Walcz <stefan.walcz@walcz.de> --------- Signed-off-by: Stefan Walcz <stefan.walcz@walcz.de> Co-authored-by: localai-org-maint-bot <306269227+localai-org-maint-bot@users.noreply.github.com>