internal/providers/anthropic/anthropic.go (2)

609-650: ⚠️ Potential issue | 🟠 Major

Skip streamed tool_use blocks that never provide a function name.

Line 636 emits a tool_calls delta before validating event.ContentBlock.Name. That diverges from extractToolCalls(), which drops nameless tool blocks, and it means the stream path can produce function.name:"" while still setting emittedToolCalls=true and later upgrading the final finish_reason to tool_calls.

Proposed fix

-func (sc *streamConverter) getToolCallState(blockIndex int, block *anthropicContent) streamToolCallState {
+func (sc *streamConverter) getToolCallState(blockIndex int, block *anthropicContent) (streamToolCallState, bool) {
 	if state, ok := sc.toolCalls[blockIndex]; ok {
-		return state
+		return state, true
+	}
+	if block == nil || strings.TrimSpace(block.Name) == "" {
+		return streamToolCallState{}, false
 	}

 	state := streamToolCallState{
 		Ordinal: sc.nextToolCallIdx,
 	}
-	if block != nil {
-		state.ID = block.ID
-		state.Name = block.Name
-	}
+	state.ID = block.ID
+	state.Name = block.Name
 	sc.toolCalls[blockIndex] = state
 	sc.nextToolCallIdx++
-	return state
+	return state, true
 }
@@
 	case "content_block_start":
 		if event.ContentBlock != nil && event.ContentBlock.Type == "tool_use" {
-			state := sc.getToolCallState(event.Index, event.ContentBlock)
+			state, ok := sc.getToolCallState(event.Index, event.ContentBlock)
+			if !ok {
+				return ""
+			}
 			sc.emittedToolCalls = true
 			return sc.emitChatChunk(map[string]interface{}{

🤖 Prompt for AI Agents

Verify each finding against the current code and only fix it if needed.

In `@internal/providers/anthropic/anthropic.go` around lines 609 - 650, The
convertEvent path is emitting a tool_calls delta for content_block_start
tool_use even when event.ContentBlock.Name is empty, which diverges from
extractToolCalls(); modify convertEvent so that when handling
"content_block_start" for a "tool_use" it first verifies event.ContentBlock.Name
(or state.Name) is non-empty and if empty simply return "" (do not call
getToolCallState, do not set sc.emittedToolCalls, and do not emit any delta);
otherwise proceed as now (call getToolCallState, mark emittedToolCalls and emit
the tool_calls chunk). Ensure the checks reference convertEvent,
getToolCallState, sc.emittedToolCalls and event.ContentBlock.Name/Name to locate
the change.

---

395-425: ⚠️ Potential issue | 🟠 Major

Mirror the malformed tool_use fallback in the non-streaming path.

Line 399 always maps tool_use to tool_calls, even when extractToolCalls(resp.Content) returned nil. That leaves clients with finish_reason:"tool_calls" and no message.tool_calls, which is the same malformed-payload inconsistency the stream path now avoids via mapStreamStopReason().

Proposed fix

 	content := extractTextContent(resp.Content)
 	toolCalls := extractToolCalls(resp.Content)

-	finishReason := mapAnthropicStopReasonToOpenAI(resp.StopReason)
+	finishReason := resp.StopReason
+	if resp.StopReason != "tool_use" || len(toolCalls) > 0 {
+		finishReason = mapAnthropicStopReasonToOpenAI(resp.StopReason)
+	}

 	usage := core.Usage{

🤖 Prompt for AI Agents

Verify each finding against the current code and only fix it if needed.

In `@internal/providers/anthropic/anthropic.go` around lines 395 - 425,
convertFromAnthropicResponse currently sets finishReason via
mapAnthropicStopReasonToOpenAI(resp.StopReason) and always maps tool_use to
tool_calls even when extractToolCalls(resp.Content) returns nil, causing
finish_reason:"tool_calls" with no message.tool_calls; change
convertFromAnthropicResponse to mirror the streaming fallback: after calling
toolCalls := extractToolCalls(resp.Content) and finishReason :=
mapAnthropicStopReasonToOpenAI(resp.StopReason), if finishReason indicates
"tool_calls" but toolCalls is nil/empty, set finishReason to the fallback string
used in streaming (e.g., "tool_use") so clients don’t receive a finish_reason
that promises missing message.tool_calls; use the same symbol names
(convertFromAnthropicResponse, extractToolCalls, mapAnthropicStopReasonToOpenAI)
to locate where to apply this check and mutation.