MindWorkAI/AI-Studio
MindWorkAI/AI-Studio
1
0
Fork 0
mirror of https://github.com/MindWorkAI/AI-Studio.git synced 2026-09-27 02:33:38 +00:00
Code Issues Actions Packages Projects Releases Wiki Activity

Fixed the tool documentation about sequential tool calls

Browse Source
This commit is contained in:
Thorsten Sommer 2026-09-24 12:07:15 +02:00
parent 35a9f1732e
commit ed2d807bc5
Signed by untrusted user who does not match committer: tsommer
GPG Key ID: 371BBA77A02C0108
1 changed files with 3 additions and 1 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

4
documentation/Tools.md
View File

@ -36,7 +36,9 @@ Tool result handling also differs by API, and this is what the adapters exist fo
- **Responses** returns `function_call` output items and receives results as `function_call_output` input items correlated by `call_id`. There the ID comes from the provider, so a call without one cannot be answered at all and ends the conversation. The whole output of a round has to be sent back for the next one, reasoning items included. - **Responses** returns `function_call` output items and receives results as `function_call_output` input items correlated by `call_id`. There the ID comes from the provider, so a call without one cannot be answered at all and ends the conversation. The whole output of a round has to be sent back for the next one, reasoning items included.
- **Anthropic** works in content blocks: the model's turn is one assistant message whose blocks may mix `text`, `thinking`, and `tool_use`, and it has to be returned unchanged — thinking blocks in particular. All results of a round belong in a **single** user message as `tool_result` blocks; splitting them across several messages teaches the model to stop asking for more than one tool at a time. It is also the only one of the three with an error flag on a result (`is_error`), which the harness sets for failed and blocked calls. - **Anthropic** works in content blocks: the model's turn is one assistant message whose blocks may mix `text`, `thinking`, and `tool_use`, and it has to be returned unchanged — thinking blocks in particular. All results of a round belong in a **single** user message as `tool_result` blocks; splitting them across several messages teaches the model to stop asking for more than one tool at a time. It is also the only one of the three with an error flag on a result (`is_error`), which the harness sets for failed and blocked calls.
AI Studio currently executes local tool calls sequentially. Therefore, Chat Completions requests with tools always set `parallel_tool_calls` to `false`, limiting each model response to at most one tool call. Requests without tools omit the parameter, and additional API parameters cannot override this behavior. Models can still request additional tools across subsequent responses. AI Studio currently executes local tool calls sequentially. Therefore, Chat Completions requests with tools set `parallel_tool_calls` to `false`, limiting each model response to at most one tool call. Requests without tools omit the parameter, and additional API parameters cannot override this behavior. Models can still request additional tools across subsequent responses.
The exception is a provider which rejects the parameter. Hugging Face answers it with a bad request, so its provider passes `mayAskForSequentialToolCalls: false` to `StreamOpenAICompatibleChatCompletion`, and its requests omit the parameter even when they offer tools. Its models may then ask for several calls in one response, which the loop works through one by one, checking the limits per call. Not every provider honors the parameter either — OpenRouter let Qwen ask for two calls at once — and such a response is handled the same way.
The OpenAI Responses API may continue to return multiple function calls in one response. AI Studio processes those calls sequentially as well; concurrent execution of separate local tool calls is not currently implemented. This does not restrict concurrency used internally by an individual tool. The OpenAI Responses API may continue to return multiple function calls in one response. AI Studio processes those calls sequentially as well; concurrent execution of separate local tool calls is not currently implemented. This does not restrict concurrency used internally by an individual tool.