mirror of
https://github.com/mudler/LocalAI.git
synced 2026-04-30 12:08:13 -04:00
f877942d97
Two bugs in MergeOpenResponsesConfig (/v1/responses + WebSocket, *not*
/v1/chat/completions — that has a separate, working path via Tool
unmarshal + SetFunctionCallNameString):
1. **Shape mismatch.** OpenAI's specific-function tool_choice nests the
name under "function":
{"type": "function", "function": {"name": "my_function"}}
The legacy flat shape was:
{"type": "function", "name": "my_function"}
Only the flat shape was handled. OpenAI-compliant clients that reach
/v1/responses (openai-python with the Responses API, Stainless-generated
SDKs, …) silently failed to force the function.
2. **Wrong setter.** The code called SetFunctionCallString(name), which
writes the mode field (functionCallString: "none"/"auto"/"required").
The specific-function name lives in a separate field
(functionCallNameString), read by ShouldCallSpecificFunction and
FunctionToCall. Net effect: a correctly-formed tool_choice never
engaged grammar-based forcing.
The fix preserves backward compatibility by accepting both shapes
(nested preferred, flat as fallback) and routes to the correct setter.
Note: The same "wrong setter" pattern appears at three other sites —
anthropic/messages.go:883, openai/realtime_model.go:171, and
openresponses/responses.go:776 — and /v1/chat/completions has its own
issue parsing tool_choice="required" as a string (json.Unmarshal on a
raw string fails silently). Those are filed as a tracking issue rather
than bundled here to keep this PR focused.
## Test plan
9 new Ginkgo specs under "MergeOpenResponsesConfig tool_choice parsing":
- string modes: "required" / "auto" / "none"
- OpenAI-spec nested shape: {type:function, function:{name}}
- Legacy Anthropic-compat flat shape: {type:function, name}
- Shape-preference: nested wins over flat when both present
- Malformed: missing type, wrong type, missing name, empty name, nil
$ go test ./core/http/middleware/ -count=1 -run TestMiddleware
Ran 28 of 28 Specs in 0.003 seconds -- PASS
## Repro (against /v1/responses)
curl -N http://localai/v1/responses \
-H 'Content-Type: application/json' \
-d '{"model":"qwen3.6-35b-a3b-apex",
"input":"Weather in Berlin?",
"tools":[{"type":"function","name":"get_weather",
"parameters":{"type":"object",
"properties":{"city":{"type":"string"}},
"required":["city"]}}],
"tool_choice":{"type":"function",
"function":{"name":"get_weather"}}}'
Before: grammar-based forcing silently inactive; model free-texts.
After : grammar forces get_weather invocation; output contains
tool_calls with function:{name:"get_weather", arguments:{...}}.