fix(devin): aggregate tool calls by id and track cache write tokens

- Track and aggregate tool calls by call ID instead of slot index in streaming and buffered execution. - Support raw arguments from invalid JSON fields for custom tool calls. - Parse usage field 4 as cache write tokens instead of adding to prompt tokens. - Align client metadata with the default client name and drop deprecated tag 28. Closes: #5910
2026-10-05 07:12:06 +08:00 · 2026-09-18 01:04:43 +08:00
parent 0b55053944
commit 9e10db53ad
4 changed files with 619 additions and 72 deletions
--- a/internal/runtime/executor/devin_executor.go
+++ b/internal/runtime/executor/devin_executor.go
@@ -462,6 +462,9 @@ func (e *DevinExecutor) streamDevinFrames(
 		name      string
 	}
 	activeToolSlots := make(map[int]*devinActiveToolSlot)
+	activeCallByID := make(map[string]*devinActiveToolSlot)
+	var activeCallSlot *devinActiveToolSlot
+	toolCallCount := 0
 	thinkingBuf := &helps.UTF8SplitBuffer{}
 	contentBuf := &helps.UTF8SplitBuffer{}
 	var accumulatedThinking strings.Builder
@@ -583,10 +586,6 @@ func (e *DevinExecutor) streamDevinFrames(
 	}

 	emitToolCall := func(tc helps.DevinToolCallDelta) bool {
-		if tc.Index < 0 || tc.Index >= maxDevinToolCalls {
-			log.Warnf("devin executor: tool call index %d out of bounds (max %d), dropping", tc.Index, maxDevinToolCalls)
-			return true
-		}
 		if thoughtStarted {
 			stopEvent, _ := sjson.SetBytes([]byte(`{"event_type":"step.stop","index":0}`), "index", stepIndex)
 			_ = emitInteractionsEvent(stopEvent)
@@ -600,16 +599,24 @@ func (e *DevinExecutor) streamDevinFrames(
 			stepIndex++
 		}

-		slot, exists := activeToolSlots[tc.Index]
-		if exists && slot.id != "" && tc.ID != "" && tc.ID != slot.id {
-			stopEvent, _ := sjson.SetBytes([]byte(`{"event_type":"step.stop","index":0}`), "index", slot.stepIndex)
-			if !emitInteractionsEvent(stopEvent) {
-				return false
-			}
-			exists = false
+		argsChunk := tc.Arguments
+		if argsChunk == "" {
+			argsChunk = tc.InvalidJSONStr
 		}

-		if !exists {
+		var slot *devinActiveToolSlot
+		if tc.ID != "" {
+			slot = activeCallByID[tc.ID]
+		} else if activeCallSlot != nil {
+			slot = activeCallSlot
+		}
+
+		if slot == nil {
+			if toolCallCount >= maxDevinToolCalls {
+				log.Warnf("devin executor: total tool calls exceeded max %d, dropping", maxDevinToolCalls)
+				return true
+			}
+			toolCallCount++
 			sIdx := stepIndex
 			stepIndex++
 			slot = &devinActiveToolSlot{
@@ -617,7 +624,11 @@ func (e *DevinExecutor) streamDevinFrames(
 				id:        tc.ID,
 				name:      tc.Name,
 			}
-			activeToolSlots[tc.Index] = slot
+			activeToolSlots[sIdx] = slot
+			if tc.ID != "" {
+				activeCallByID[tc.ID] = slot
+			}
+			activeCallSlot = slot
 			startEvent, _ := sjson.SetBytes([]byte(`{"event_type":"step.start","index":0,"step":{"type":"function_call","name":"","id":"","call_id":"","arguments":{}}}`), "index", sIdx)
 			startEvent, _ = sjson.SetBytes(startEvent, "step.name", tc.Name)
 			startEvent, _ = sjson.SetBytes(startEvent, "step.id", tc.ID)
@@ -626,9 +637,11 @@ func (e *DevinExecutor) streamDevinFrames(
 				return false
 			}
 		} else {
+			activeCallSlot = slot
 			updated := false
 			if slot.id == "" && tc.ID != "" {
 				slot.id = tc.ID
+				activeCallByID[tc.ID] = slot
 				updated = true
 			}
 			if slot.name == "" && tc.Name != "" {
@@ -644,9 +657,9 @@ func (e *DevinExecutor) streamDevinFrames(
 			}
 		}

-		if tc.Arguments != "" {
+		if argsChunk != "" {
 			deltaEvent, _ := sjson.SetBytes([]byte(`{"event_type":"step.delta","index":0,"delta":{"type":"arguments_delta","arguments":""}}`), "index", slot.stepIndex)
-			deltaEvent, _ = translatorcommon.SetStringWithoutHTMLEscape(deltaEvent, "delta.arguments", tc.Arguments)
+			deltaEvent, _ = translatorcommon.SetStringWithoutHTMLEscape(deltaEvent, "delta.arguments", argsChunk)
 			if !emitInteractionsEvent(deltaEvent) {
 				return false
 			}
@@ -674,6 +687,8 @@ func (e *DevinExecutor) streamDevinFrames(
 				_ = emitInteractionsEvent(stopEvent)
 			}
 			clear(activeToolSlots)
+			clear(activeCallByID)
+			activeCallSlot = nil
 		}
 	}

@@ -729,6 +744,9 @@ func (e *DevinExecutor) streamDevinFrames(
 				if frameRes.Usage.CachedTokens > 0 {
 					finalUsage.CachedTokens = frameRes.Usage.CachedTokens
 				}
+				if frameRes.Usage.CacheWriteTokens > 0 {
+					finalUsage.CacheWriteTokens = frameRes.Usage.CacheWriteTokens
+				}
 				if frameRes.Usage.RequestID != "" {
 					finalUsage.RequestID = frameRes.Usage.RequestID
 				}
@@ -915,6 +933,9 @@ func (e *DevinExecutor) streamDevinFrames(
 		completedEvent, _ = sjson.SetBytes(completedEvent, "interaction.usage.total_input_tokens", totalInput)
 		completedEvent, _ = sjson.SetBytes(completedEvent, "interaction.usage.total_output_tokens", totalOutput)
 		completedEvent, _ = sjson.SetBytes(completedEvent, "interaction.usage.total_cached_tokens", finalUsage.CachedTokens)
+		if finalUsage.CacheWriteTokens > 0 {
+			completedEvent, _ = sjson.SetBytes(completedEvent, "interaction.usage.cache_write_tokens", finalUsage.CacheWriteTokens)
+		}
 		completedEvent, _ = sjson.SetBytes(completedEvent, "interaction.usage.total_tokens", totalTokens)
 		if detail, ok := helps.ParseInteractionsStreamUsage(completedEvent); ok {
 			if reporter != nil {
@@ -976,7 +997,8 @@ func consumeDevinFramesToInteractions(body io.Reader, model, chatModelUID string
 		args strings.Builder
 	}
 	var toolBuilders []*devinToolCallBuilder
-	slotToBuilderIndex := make(map[int]int)
+	callIDToBuilderIndex := make(map[string]int)
+	lastBuilderIdx := -1

 	getToolCalls := func() []helps.DevinToolCall {
 		if len(toolBuilders) == 0 {
@@ -1078,6 +1100,9 @@ func consumeDevinFramesToInteractions(body io.Reader, model, chatModelUID string
 				if frameRes.Usage.CachedTokens > 0 {
 					finalUsage.CachedTokens = frameRes.Usage.CachedTokens
 				}
+				if frameRes.Usage.CacheWriteTokens > 0 {
+					finalUsage.CacheWriteTokens = frameRes.Usage.CacheWriteTokens
+				}
 				if frameRes.Usage.RequestID != "" {
 					finalUsage.RequestID = frameRes.Usage.RequestID
 				}
@@ -1123,32 +1148,47 @@ func consumeDevinFramesToInteractions(body io.Reader, model, chatModelUID string
 			textParts = append(textParts, frameRes.ContentText)
 		}
 		for _, tc := range frameRes.ToolCallDeltas {
-			slotIdx := tc.Index
-			if slotIdx < 0 || slotIdx >= maxDevinToolCalls {
-				log.Warnf("devin executor: tool call index %d out of bounds (max %d), dropping", slotIdx, maxDevinToolCalls)
-				continue
+			argsChunk := tc.Arguments
+			if argsChunk == "" {
+				argsChunk = tc.InvalidJSONStr
 			}
-			bIdx, exists := slotToBuilderIndex[slotIdx]
-			if exists && tc.ID != "" && toolBuilders[bIdx].id != "" && tc.ID != toolBuilders[bIdx].id {
-				exists = false
+
+			var bIdx int
+			var exists bool
+			if tc.ID != "" {
+				bIdx, exists = callIDToBuilderIndex[tc.ID]
+			} else if lastBuilderIdx >= 0 {
+				bIdx = lastBuilderIdx
+				exists = true
 			}
+
 			if !exists {
 				if len(toolBuilders) >= maxDevinToolCalls {
 					log.Warnf("devin executor: total tool calls exceeded max %d, dropping", maxDevinToolCalls)
 					continue
 				}
 				bIdx = len(toolBuilders)
-				toolBuilders = append(toolBuilders, &devinToolCallBuilder{})
-				slotToBuilderIndex[slotIdx] = bIdx
+				toolBuilders = append(toolBuilders, &devinToolCallBuilder{
+					id:   tc.ID,
+					name: tc.Name,
+				})
+				if tc.ID != "" {
+					callIDToBuilderIndex[tc.ID] = bIdx
+				}
+				lastBuilderIdx = bIdx
+			} else {
+				lastBuilderIdx = bIdx
+				if toolBuilders[bIdx].id == "" && tc.ID != "" {
+					toolBuilders[bIdx].id = tc.ID
+					callIDToBuilderIndex[tc.ID] = bIdx
+				}
+				if tc.Name != "" {
+					toolBuilders[bIdx].name = tc.Name
+				}
 			}
-			if tc.ID != "" {
-				toolBuilders[bIdx].id = tc.ID
-			}
-			if tc.Name != "" {
-				toolBuilders[bIdx].name = tc.Name
-			}
-			if tc.Arguments != "" {
-				toolBuilders[bIdx].args.WriteString(tc.Arguments)
+
+			if argsChunk != "" {
+				toolBuilders[bIdx].args.WriteString(argsChunk)
 			}
 		}
 	}
@@ -1221,8 +1261,12 @@ func consumeDevinFramesToInteractions(body io.Reader, model, chatModelUID string
 		fnStep, _ = sjson.SetBytes(fnStep, "name", tc.Name)
 		fnStep, _ = sjson.SetBytes(fnStep, "id", tc.ID)
 		fnStep, _ = sjson.SetBytes(fnStep, "call_id", tc.ID)
-		if tc.Arguments != "" && json.Valid([]byte(tc.Arguments)) {
-			fnStep, _ = sjson.SetRawBytes(fnStep, "arguments", []byte(tc.Arguments))
+		if tc.Arguments != "" {
+			if json.Valid([]byte(tc.Arguments)) {
+				fnStep, _ = sjson.SetRawBytes(fnStep, "arguments", []byte(tc.Arguments))
+			} else {
+				fnStep, _ = translatorcommon.SetStringWithoutHTMLEscape(fnStep, "arguments", tc.Arguments)
+			}
 		}
 		steps = append(steps, fnStep)
 	}
@@ -1248,6 +1292,9 @@ func consumeDevinFramesToInteractions(body io.Reader, model, chatModelUID string
 		out, _ = sjson.SetBytes(out, "usage.total_input_tokens", totalInput)
 		out, _ = sjson.SetBytes(out, "usage.total_output_tokens", totalOutput)
 		out, _ = sjson.SetBytes(out, "usage.total_cached_tokens", finalUsage.CachedTokens)
+		if finalUsage.CacheWriteTokens > 0 {
+			out, _ = sjson.SetBytes(out, "usage.cache_write_tokens", finalUsage.CacheWriteTokens)
+		}
 		out, _ = sjson.SetBytes(out, "usage.total_tokens", totalTokens)
 	}

--- a/internal/runtime/executor/devin_executor_test.go
+++ b/internal/runtime/executor/devin_executor_test.go
@@ -780,8 +780,6 @@ func TestConsumeDevinFramesToInteractions_MultiToolCallsNoPanic(t *testing.T) {
 	tc0 = protowire.AppendString(tc0, "tool_0")
 	tc0 = protowire.AppendTag(tc0, 3, protowire.BytesType)
 	tc0 = protowire.AppendString(tc0, `{"a":`)
-	tc0 = protowire.AppendTag(tc0, 4, protowire.VarintType)
-	tc0 = protowire.AppendVarint(tc0, 0) // index 0

 	var f1 []byte
 	f1 = protowire.AppendTag(f1, 6, protowire.BytesType)
@@ -796,8 +794,6 @@ func TestConsumeDevinFramesToInteractions_MultiToolCallsNoPanic(t *testing.T) {
 	tc1 = protowire.AppendString(tc1, "tool_1")
 	tc1 = protowire.AppendTag(tc1, 3, protowire.BytesType)
 	tc1 = protowire.AppendString(tc1, `{"b": 2}`)
-	tc1 = protowire.AppendTag(tc1, 4, protowire.VarintType)
-	tc1 = protowire.AppendVarint(tc1, 1) // index 1

 	var f2 []byte
 	f2 = protowire.AppendTag(f2, 6, protowire.BytesType)
@@ -806,10 +802,10 @@ func TestConsumeDevinFramesToInteractions_MultiToolCallsNoPanic(t *testing.T) {

 	// Frame 3: tool call 0 continuation
 	var tc0Cont []byte
+	tc0Cont = protowire.AppendTag(tc0Cont, 1, protowire.BytesType)
+	tc0Cont = protowire.AppendString(tc0Cont, "call_0")
 	tc0Cont = protowire.AppendTag(tc0Cont, 3, protowire.BytesType)
 	tc0Cont = protowire.AppendString(tc0Cont, `1}`)
-	tc0Cont = protowire.AppendTag(tc0Cont, 4, protowire.VarintType)
-	tc0Cont = protowire.AppendVarint(tc0Cont, 0) // index 0

 	var f3 []byte
 	f3 = protowire.AppendTag(f3, 6, protowire.BytesType)
@@ -954,14 +950,12 @@ func TestStreamDevinFrames_SequentialToolCallsSameIndexDifferentID(t *testing.T)
 	tc0 = protowire.AppendString(tc0, "title")
 	tc0 = protowire.AppendTag(tc0, 3, protowire.BytesType)
 	tc0 = protowire.AppendString(tc0, `{"title": "Triage issue 5802"}`)
-	tc0 = protowire.AppendTag(tc0, 4, protowire.VarintType)
-	tc0 = protowire.AppendVarint(tc0, 0) // index 0

 	var f1 []byte
 	f1 = protowire.AppendTag(f1, 6, protowire.BytesType)
 	f1 = protowire.AppendBytes(f1, tc0)

-	// Frame 2: bash_1 (same index 0, but different ID)
+	// Frame 2: bash_1
 	var tc1 []byte
 	tc1 = protowire.AppendTag(tc1, 1, protowire.BytesType)
 	tc1 = protowire.AppendString(tc1, "bash_1")
@@ -969,8 +963,6 @@ func TestStreamDevinFrames_SequentialToolCallsSameIndexDifferentID(t *testing.T)
 	tc1 = protowire.AppendString(tc1, "bash")
 	tc1 = protowire.AppendTag(tc1, 3, protowire.BytesType)
 	tc1 = protowire.AppendString(tc1, `{"command": "gh issue view 5802 2>&1 | head -100"}`)
-	tc1 = protowire.AppendTag(tc1, 4, protowire.VarintType)
-	tc1 = protowire.AppendVarint(tc1, 0) // index 0

 	var f2 []byte
 	f2 = protowire.AppendTag(f2, 6, protowire.BytesType)
@@ -1072,8 +1064,6 @@ func TestConsumeDevinFramesToInteractions_SequentialToolCallsSameIndexDifferentI
 	tc0 = protowire.AppendString(tc0, "title")
 	tc0 = protowire.AppendTag(tc0, 3, protowire.BytesType)
 	tc0 = protowire.AppendString(tc0, `{"title": "Triage issue 5802"}`)
-	tc0 = protowire.AppendTag(tc0, 4, protowire.VarintType)
-	tc0 = protowire.AppendVarint(tc0, 0)

 	var f1 []byte
 	f1 = protowire.AppendTag(f1, 6, protowire.BytesType)
@@ -1086,8 +1076,6 @@ func TestConsumeDevinFramesToInteractions_SequentialToolCallsSameIndexDifferentI
 	tc1 = protowire.AppendString(tc1, "bash")
 	tc1 = protowire.AppendTag(tc1, 3, protowire.BytesType)
 	tc1 = protowire.AppendString(tc1, `{"command": "gh issue view 5802 2>&1 | head -100"}`)
-	tc1 = protowire.AppendTag(tc1, 4, protowire.VarintType)
-	tc1 = protowire.AppendVarint(tc1, 0)

 	var f2 []byte
 	f2 = protowire.AppendTag(f2, 6, protowire.BytesType)
@@ -1129,7 +1117,7 @@ func TestConsumeDevinFramesToInteractions_SequentialToolCallsSameIndexDifferentI

 func TestConsumeDevinFramesToInteractions_ToolCallsLimit128(t *testing.T) {
 	var buf bytes.Buffer
-	// Create 135 tool calls across sequential ID switches on index 0
+	// Create 135 tool calls across sequential ID switches
 	for i := 0; i < 135; i++ {
 		var tc []byte
 		tc = protowire.AppendTag(tc, 1, protowire.BytesType)
@@ -1138,8 +1126,6 @@ func TestConsumeDevinFramesToInteractions_ToolCallsLimit128(t *testing.T) {
 		tc = protowire.AppendString(tc, fmt.Sprintf("tool_%d", i))
 		tc = protowire.AppendTag(tc, 3, protowire.BytesType)
 		tc = protowire.AppendString(tc, `{"param":1}`)
-		tc = protowire.AppendTag(tc, 4, protowire.VarintType)
-		tc = protowire.AppendVarint(tc, 0)

 		var f []byte
 		f = protowire.AppendTag(f, 6, protowire.BytesType)
@@ -1166,8 +1152,6 @@ func TestStreamDevinFrames_SameIDDoesNotDuplicateStart(t *testing.T) {
 	tc0 = protowire.AppendString(tc0, "tool_1")
 	tc0 = protowire.AppendTag(tc0, 3, protowire.BytesType)
 	tc0 = protowire.AppendString(tc0, `{"a":`)
-	tc0 = protowire.AppendTag(tc0, 4, protowire.VarintType)
-	tc0 = protowire.AppendVarint(tc0, 0)

 	var f1 []byte
 	f1 = protowire.AppendTag(f1, 6, protowire.BytesType)
@@ -1181,8 +1165,6 @@ func TestStreamDevinFrames_SameIDDoesNotDuplicateStart(t *testing.T) {
 	tc1 = protowire.AppendString(tc1, "tool_1")
 	tc1 = protowire.AppendTag(tc1, 3, protowire.BytesType)
 	tc1 = protowire.AppendString(tc1, `1}`)
-	tc1 = protowire.AppendTag(tc1, 4, protowire.VarintType)
-	tc1 = protowire.AppendVarint(tc1, 0)

 	var f2 []byte
 	f2 = protowire.AppendTag(f2, 6, protowire.BytesType)
@@ -2836,3 +2818,402 @@ func TestDevinExecutor_SupplementImagesEdgeCases(t *testing.T) {
 		}
 	})
 }
+
+func TestRegressionIssue5910_ToolCallAggregationByCallID(t *testing.T) {
+	// Frame 1: call_1 start + partial args
+	var tc1 []byte
+	tc1 = protowire.AppendTag(tc1, 1, protowire.BytesType)
+	tc1 = protowire.AppendString(tc1, "call_1")
+	tc1 = protowire.AppendTag(tc1, 2, protowire.BytesType)
+	tc1 = protowire.AppendString(tc1, "tool_1")
+	tc1 = protowire.AppendTag(tc1, 3, protowire.BytesType)
+	tc1 = protowire.AppendString(tc1, `{"a":`)
+
+	var f1 []byte
+	f1 = protowire.AppendTag(f1, 6, protowire.BytesType)
+	f1 = protowire.AppendBytes(f1, tc1)
+
+	// Frame 2: call_2 start + args
+	var tc2 []byte
+	tc2 = protowire.AppendTag(tc2, 1, protowire.BytesType)
+	tc2 = protowire.AppendString(tc2, "call_2")
+	tc2 = protowire.AppendTag(tc2, 2, protowire.BytesType)
+	tc2 = protowire.AppendString(tc2, "tool_2")
+	tc2 = protowire.AppendTag(tc2, 3, protowire.BytesType)
+	tc2 = protowire.AppendString(tc2, `{"b":2}`)
+
+	var f2 []byte
+	f2 = protowire.AppendTag(f2, 6, protowire.BytesType)
+	f2 = protowire.AppendBytes(f2, tc2)
+
+	// Frame 3: call_1 continuation args
+	var tc1Cont []byte
+	tc1Cont = protowire.AppendTag(tc1Cont, 1, protowire.BytesType)
+	tc1Cont = protowire.AppendString(tc1Cont, "call_1")
+	tc1Cont = protowire.AppendTag(tc1Cont, 3, protowire.BytesType)
+	tc1Cont = protowire.AppendString(tc1Cont, `1}`)
+
+	var f3 []byte
+	f3 = protowire.AppendTag(f3, 6, protowire.BytesType)
+	f3 = protowire.AppendBytes(f3, tc1Cont)
+
+	var buf bytes.Buffer
+	buf.Write(helps.WrapConnectEnvelope(f1))
+	buf.Write(helps.WrapConnectEnvelope(f2))
+	buf.Write(helps.WrapConnectEnvelope(f3))
+	buf.Write(helps.WrapConnectEnvelopeWithFlag(helps.ConnectFlagEndStream, []byte(`{}`)))
+
+	_, respLog, err := consumeDevinFramesToInteractions(&buf, "devin/swe-2", "chat-model-uid")
+	if err != nil {
+		t.Fatalf("consumeDevinFramesToInteractions failed: %v", err)
+	}
+
+	if len(respLog.ToolCalls) != 2 {
+		t.Fatalf("expected 2 tool calls, got %d", len(respLog.ToolCalls))
+	}
+	if respLog.ToolCalls[0].ID != "call_1" || respLog.ToolCalls[0].Arguments != `{"a":1}` {
+		t.Errorf("tool call 0 = %+v, want call_1 with args {\"a\":1}", respLog.ToolCalls[0])
+	}
+	if respLog.ToolCalls[1].ID != "call_2" || respLog.ToolCalls[1].Arguments != `{"b":2}` {
+		t.Errorf("tool call 1 = %+v, want call_2 with args {\"b\":2}", respLog.ToolCalls[1])
+	}
+}
+
+func TestRegressionIssue5910_CustomToolCallInvalidJSONStr(t *testing.T) {
+	// Upstream Devin sends raw arguments for custom tool calls in field 4 (invalid_json_str),
+	// parse error in field 5 (invalid_json_err), and custom flag in field 6 (is_custom_tool_call).
+	var tc []byte
+	tc = protowire.AppendTag(tc, 1, protowire.BytesType)
+	tc = protowire.AppendString(tc, "call_custom_1")
+	tc = protowire.AppendTag(tc, 2, protowire.BytesType)
+	tc = protowire.AppendString(tc, "bash")
+	// Field 4: invalid_json_str = "ls -la"
+	tc = protowire.AppendTag(tc, 4, protowire.BytesType)
+	tc = protowire.AppendString(tc, "ls -la")
+	// Field 5: invalid_json_err = "not valid json"
+	tc = protowire.AppendTag(tc, 5, protowire.BytesType)
+	tc = protowire.AppendString(tc, "not valid json")
+	// Field 6: is_custom_tool_call = true
+	tc = protowire.AppendTag(tc, 6, protowire.VarintType)
+	tc = protowire.AppendVarint(tc, 1)
+
+	var f1 []byte
+	f1 = protowire.AppendTag(f1, 6, protowire.BytesType)
+	f1 = protowire.AppendBytes(f1, tc)
+
+	var buf bytes.Buffer
+	buf.Write(helps.WrapConnectEnvelope(f1))
+	buf.Write(helps.WrapConnectEnvelopeWithFlag(helps.ConnectFlagEndStream, []byte(`{}`)))
+
+	interactionsJSON, respLog, err := consumeDevinFramesToInteractions(&buf, "devin/swe-2", "chat-model-uid")
+	if err != nil {
+		t.Fatalf("consumeDevinFramesToInteractions failed: %v", err)
+	}
+
+	if len(respLog.ToolCalls) != 1 {
+		t.Fatalf("expected 1 tool call, got %d", len(respLog.ToolCalls))
+	}
+	if respLog.ToolCalls[0].Arguments != "ls -la" {
+		t.Errorf("expected Arguments %q from invalid_json_str, got %q", "ls -la", respLog.ToolCalls[0].Arguments)
+	}
+
+	steps := gjson.GetBytes(interactionsJSON, "steps").Array()
+	if len(steps) != 1 {
+		t.Fatalf("expected 1 step in interactions JSON, got %d", len(steps))
+	}
+	if steps[0].Get("name").String() != "bash" || steps[0].Get("id").String() != "call_custom_1" {
+		t.Errorf("unexpected step 0: %s", steps[0].Raw)
+	}
+	if steps[0].Get("arguments").String() != "ls -la" {
+		t.Errorf("expected step 0 arguments %q, got %q", "ls -la", steps[0].Get("arguments").String())
+	}
+
+	// Also verify streaming receives raw arguments for custom tool call
+	var streamBuf bytes.Buffer
+	streamBuf.Write(helps.WrapConnectEnvelope(f1))
+	streamBuf.Write(helps.WrapConnectEnvelopeWithFlag(helps.ConnectFlagEndStream, []byte(`{}`)))
+
+	exec := NewDevinExecutor(&config.Config{})
+	out := make(chan cliproxyexecutor.StreamChunk, 50)
+	opts := cliproxyexecutor.Options{
+		SourceFormat: sdktranslator.FormatInteractions,
+	}
+
+	go func() {
+		defer close(out)
+		exec.streamDevinFrames(
+			context.Background(),
+			&streamBuf,
+			cliproxyexecutor.Request{Model: "devin/swe-2"},
+			opts,
+			"chat-model-uid",
+			sdktranslator.FormatInteractions,
+			nil,
+			out,
+		)
+	}()
+
+	var streamArgs strings.Builder
+	for chunk := range out {
+		if chunk.Err != nil {
+			t.Fatalf("unexpected stream error: %v", chunk.Err)
+		}
+		lines := strings.Split(string(chunk.Payload), "\n")
+		for _, line := range lines {
+			if strings.HasPrefix(line, "data: ") {
+				data := strings.TrimPrefix(line, "data: ")
+				if gjson.Get(data, "event_type").String() == "step.delta" && gjson.Get(data, "delta.type").String() == "arguments_delta" {
+					streamArgs.WriteString(gjson.Get(data, "delta.arguments").String())
+				}
+			}
+		}
+	}
+
+	if streamArgs.String() != "ls -la" {
+		t.Errorf("expected stream delta arguments %q, got %q", "ls -la", streamArgs.String())
+	}
+}
+
+func TestRegressionIssue5910_StreamInterleavedToolCallsByCallID(t *testing.T) {
+	// Frame 1: call_1 start + partial args
+	var tc1 []byte
+	tc1 = protowire.AppendTag(tc1, 1, protowire.BytesType)
+	tc1 = protowire.AppendString(tc1, "call_1")
+	tc1 = protowire.AppendTag(tc1, 2, protowire.BytesType)
+	tc1 = protowire.AppendString(tc1, "tool_1")
+	tc1 = protowire.AppendTag(tc1, 3, protowire.BytesType)
+	tc1 = protowire.AppendString(tc1, `{"a":`)
+
+	var f1 []byte
+	f1 = protowire.AppendTag(f1, 6, protowire.BytesType)
+	f1 = protowire.AppendBytes(f1, tc1)
+
+	// Frame 2: call_2 start + full args
+	var tc2 []byte
+	tc2 = protowire.AppendTag(tc2, 1, protowire.BytesType)
+	tc2 = protowire.AppendString(tc2, "call_2")
+	tc2 = protowire.AppendTag(tc2, 2, protowire.BytesType)
+	tc2 = protowire.AppendString(tc2, "tool_2")
+	tc2 = protowire.AppendTag(tc2, 3, protowire.BytesType)
+	tc2 = protowire.AppendString(tc2, `{"b":2}`)
+
+	var f2 []byte
+	f2 = protowire.AppendTag(f2, 6, protowire.BytesType)
+	f2 = protowire.AppendBytes(f2, tc2)
+
+	// Frame 3: call_1 continuation args
+	var tc1Cont []byte
+	tc1Cont = protowire.AppendTag(tc1Cont, 1, protowire.BytesType)
+	tc1Cont = protowire.AppendString(tc1Cont, "call_1")
+	tc1Cont = protowire.AppendTag(tc1Cont, 3, protowire.BytesType)
+	tc1Cont = protowire.AppendString(tc1Cont, `1}`)
+
+	var f3 []byte
+	f3 = protowire.AppendTag(f3, 6, protowire.BytesType)
+	f3 = protowire.AppendBytes(f3, tc1Cont)
+
+	var buf bytes.Buffer
+	buf.Write(helps.WrapConnectEnvelope(f1))
+	buf.Write(helps.WrapConnectEnvelope(f2))
+	buf.Write(helps.WrapConnectEnvelope(f3))
+	buf.Write(helps.WrapConnectEnvelopeWithFlag(helps.ConnectFlagEndStream, []byte(`{}`)))
+
+	exec := NewDevinExecutor(&config.Config{})
+	out := make(chan cliproxyexecutor.StreamChunk, 50)
+	opts := cliproxyexecutor.Options{
+		SourceFormat: sdktranslator.FormatInteractions,
+	}
+
+	go func() {
+		defer close(out)
+		exec.streamDevinFrames(
+			context.Background(),
+			&buf,
+			cliproxyexecutor.Request{Model: "devin/swe-2"},
+			opts,
+			"chat-model-uid",
+			sdktranslator.FormatInteractions,
+			nil,
+			out,
+		)
+	}()
+
+	var chunks []cliproxyexecutor.StreamChunk
+	for chunk := range out {
+		if chunk.Err != nil {
+			t.Fatalf("unexpected chunk error: %v", chunk.Err)
+		}
+		chunks = append(chunks, chunk)
+	}
+
+	var events []gjson.Result
+	for _, chunk := range chunks {
+		lines := strings.Split(string(chunk.Payload), "\n")
+		for _, line := range lines {
+			if strings.HasPrefix(line, "data: ") {
+				data := strings.TrimPrefix(line, "data: ")
+				if strings.TrimSpace(data) != "[DONE]" {
+					events = append(events, gjson.Parse(data))
+				}
+			}
+		}
+	}
+
+	var toolCallsStarted []string
+	var toolCallsStopped []int64
+	stepArgs := make(map[int64]*strings.Builder)
+	for _, ev := range events {
+		eventType := ev.Get("event_type").String()
+		if eventType == "step.start" && ev.Get("step.type").String() == "function_call" {
+			idx := ev.Get("index").Int()
+			toolCallsStarted = append(toolCallsStarted, fmt.Sprintf("index:%d,id:%s,name:%s", idx, ev.Get("step.id").String(), ev.Get("step.name").String()))
+			stepArgs[idx] = &strings.Builder{}
+		} else if eventType == "step.delta" && ev.Get("delta.type").String() == "arguments_delta" {
+			idx := ev.Get("index").Int()
+			if b, ok := stepArgs[idx]; ok {
+				b.WriteString(ev.Get("delta.arguments").String())
+			}
+		} else if eventType == "step.stop" {
+			toolCallsStopped = append(toolCallsStopped, ev.Get("index").Int())
+		}
+	}
+
+	if len(toolCallsStarted) != 2 {
+		t.Fatalf("expected exactly 2 tool calls started, got %d: %v", len(toolCallsStarted), toolCallsStarted)
+	}
+	if toolCallsStarted[0] != "index:0,id:call_1,name:tool_1" {
+		t.Errorf("tool call 0 = %q, want index:0,id:call_1,name:tool_1", toolCallsStarted[0])
+	}
+	if toolCallsStarted[1] != "index:1,id:call_2,name:tool_2" {
+		t.Errorf("tool call 1 = %q, want index:1,id:call_2,name:tool_2", toolCallsStarted[1])
+	}
+
+	if len(toolCallsStopped) != 2 {
+		t.Fatalf("expected exactly 2 tool calls stopped, got %d: %v", len(toolCallsStopped), toolCallsStopped)
+	}
+	if toolCallsStopped[0] != 0 || toolCallsStopped[1] != 1 {
+		t.Errorf("tool calls stopped indices = %v, want [0, 1]", toolCallsStopped)
+	}
+
+	if stepArgs[0].String() != `{"a":1}` {
+		t.Errorf("stepArgs[0] = %q, want {\"a\":1}", stepArgs[0].String())
+	}
+	if stepArgs[1].String() != `{"b":2}` {
+		t.Errorf("stepArgs[1] = %q, want {\"b\":2}", stepArgs[1].String())
+	}
+}
+
+func TestRegressionIssue5910_UsageStatsCacheWriteTokensInResponses(t *testing.T) {
+	// Frame 1 with Usage Field 7
+	var f7Bytes []byte
+	// Field 2: prompt tokens = 3
+	f7Bytes = protowire.AppendTag(f7Bytes, 2, protowire.VarintType)
+	f7Bytes = protowire.AppendVarint(f7Bytes, 3)
+	// Field 3: completion tokens = 10
+	f7Bytes = protowire.AppendTag(f7Bytes, 3, protowire.VarintType)
+	f7Bytes = protowire.AppendVarint(f7Bytes, 10)
+	// Field 4: cache_write_tokens = 14361
+	f7Bytes = protowire.AppendTag(f7Bytes, 4, protowire.VarintType)
+	f7Bytes = protowire.AppendVarint(f7Bytes, 14361)
+	// Field 5: cached tokens = 50
+	f7Bytes = protowire.AppendTag(f7Bytes, 5, protowire.VarintType)
+	f7Bytes = protowire.AppendVarint(f7Bytes, 50)
+
+	var frame []byte
+	frame = protowire.AppendTag(frame, 7, protowire.BytesType)
+	frame = protowire.AppendBytes(frame, f7Bytes)
+
+	var buf bytes.Buffer
+	buf.Write(helps.WrapConnectEnvelope(frame))
+	buf.Write(helps.WrapConnectEnvelopeWithFlag(helps.ConnectFlagEndStream, []byte(`{}`)))
+
+	// 1. Non-streaming test
+	interactionsJSON, respLog, err := consumeDevinFramesToInteractions(&buf, "devin/swe-2", "chat-model-uid")
+	if err != nil {
+		t.Fatalf("consumeDevinFramesToInteractions failed: %v", err)
+	}
+
+	if respLog.Usage == nil {
+		t.Fatal("expected non-nil respLog.Usage")
+	}
+	if respLog.Usage.PromptTokens != 3 {
+		t.Errorf("respLog.Usage.PromptTokens = %d, want 3", respLog.Usage.PromptTokens)
+	}
+	if respLog.Usage.CacheWriteTokens != 14361 {
+		t.Errorf("respLog.Usage.CacheWriteTokens = %d, want 14361", respLog.Usage.CacheWriteTokens)
+	}
+
+	root := gjson.ParseBytes(interactionsJSON)
+	if root.Get("usage.total_input_tokens").Int() != 53 {
+		t.Errorf("usage.total_input_tokens = %d, want 53", root.Get("usage.total_input_tokens").Int())
+	}
+	if root.Get("usage.cache_write_tokens").Int() != 14361 {
+		t.Errorf("usage.cache_write_tokens = %d, want 14361", root.Get("usage.cache_write_tokens").Int())
+	}
+	detail := helps.ParseInteractionsUsage(interactionsJSON)
+	if detail.CacheCreationTokens != 14361 {
+		t.Errorf("ParseInteractionsUsage CacheCreationTokens = %d, want 14361", detail.CacheCreationTokens)
+	}
+	if detail.InputTokens != 53 {
+		t.Errorf("ParseInteractionsUsage InputTokens = %d, want 53", detail.InputTokens)
+	}
+
+	// 2. Streaming test
+	var streamBuf bytes.Buffer
+	streamBuf.Write(helps.WrapConnectEnvelope(frame))
+	streamBuf.Write(helps.WrapConnectEnvelopeWithFlag(helps.ConnectFlagEndStream, []byte(`{}`)))
+
+	exec := NewDevinExecutor(&config.Config{})
+	out := make(chan cliproxyexecutor.StreamChunk, 50)
+	opts := cliproxyexecutor.Options{
+		SourceFormat: sdktranslator.FormatInteractions,
+	}
+
+	go func() {
+		defer close(out)
+		exec.streamDevinFrames(
+			context.Background(),
+			&streamBuf,
+			cliproxyexecutor.Request{Model: "devin/swe-2"},
+			opts,
+			"chat-model-uid",
+			sdktranslator.FormatInteractions,
+			nil,
+			out,
+		)
+	}()
+
+	var completedEvent []byte
+	for chunk := range out {
+		if chunk.Err != nil {
+			t.Fatalf("unexpected chunk error: %v", chunk.Err)
+		}
+		lines := strings.Split(string(chunk.Payload), "\n")
+		for _, line := range lines {
+			if strings.HasPrefix(line, "data: ") {
+				data := []byte(strings.TrimPrefix(line, "data: "))
+				if gjson.GetBytes(data, "event_type").String() == "interaction.completed" {
+					completedEvent = data
+				}
+			}
+		}
+	}
+
+	if len(completedEvent) == 0 {
+		t.Fatal("expected interaction.completed event in stream")
+	}
+	cRoot := gjson.ParseBytes(completedEvent)
+	if cRoot.Get("interaction.usage.cache_write_tokens").Int() != 14361 {
+		t.Errorf("interaction.usage.cache_write_tokens = %d, want 14361", cRoot.Get("interaction.usage.cache_write_tokens").Int())
+	}
+	sDetail, ok := helps.ParseInteractionsStreamUsage(completedEvent)
+	if !ok {
+		t.Fatal("ParseInteractionsStreamUsage returned false")
+	}
+	if sDetail.CacheCreationTokens != 14361 {
+		t.Errorf("ParseInteractionsStreamUsage CacheCreationTokens = %d, want 14361", sDetail.CacheCreationTokens)
+	}
+	if sDetail.InputTokens != 53 {
+		t.Errorf("ParseInteractionsStreamUsage InputTokens = %d, want 53", sDetail.InputTokens)
+	}
+}
--- a/internal/runtime/executor/helps/devin_wire.go
+++ b/internal/runtime/executor/helps/devin_wire.go
@@ -68,10 +68,12 @@ type DevinToolCall struct {

 // DevinToolCallDelta represents a streaming tool call chunk from response Field 6.
 type DevinToolCallDelta struct {
-	ID        string
-	Name      string
-	Arguments string
-	Index     int
+	ID               string
+	Name             string
+	Arguments        string
+	InvalidJSONStr   string
+	InvalidJSONErr   string
+	IsCustomToolCall bool
 }

 // DevinImage represents an image attachment in a DevinPrompt (Prompt #10).
@@ -98,6 +100,7 @@ type DevinUsage struct {
 	PromptTokens     int64             `json:"prompt_tokens"`
 	CompletionTokens int64             `json:"completion_tokens"`
 	CachedTokens     int64             `json:"cached_tokens"`
+	CacheWriteTokens int64             `json:"cache_write_tokens,omitempty"`
 	StatusCode       uint64            `json:"status_code,omitempty"`
 	RequestID        string            `json:"request_id,omitempty"`
 	ModelName        string            `json:"model_name,omitempty"`
@@ -253,7 +256,7 @@ func BuildDevinClientMetadataBytes(sessionToken, deviceSeed, osName string) []by

 	var f1Bytes []byte
 	f1Bytes = protowire.AppendTag(f1Bytes, 1, protowire.BytesType)
-	f1Bytes = protowire.AppendString(f1Bytes, "devin-cli")
+	f1Bytes = protowire.AppendString(f1Bytes, DevinDefaultClientName)

 	f1Bytes = protowire.AppendTag(f1Bytes, 2, protowire.BytesType)
 	f1Bytes = protowire.AppendString(f1Bytes, DevinDefaultClientVersion)
@@ -273,9 +276,6 @@ func BuildDevinClientMetadataBytes(sessionToken, deviceSeed, osName string) []by
 	f1Bytes = protowire.AppendTag(f1Bytes, 12, protowire.BytesType)
 	f1Bytes = protowire.AppendString(f1Bytes, DevinDefaultClientName)

-	f1Bytes = protowire.AppendTag(f1Bytes, 28, protowire.BytesType)
-	f1Bytes = protowire.AppendString(f1Bytes, DevinDefaultClientName)
-
 	f1Bytes = protowire.AppendTag(f1Bytes, 31, protowire.BytesType)
 	f1Bytes = protowire.AppendString(f1Bytes, deviceFingerprint)
 	return f1Bytes
@@ -674,8 +674,8 @@ func parseDevinToolCallDelta(data []byte) (DevinToolCallDelta, error) {
 				return tc, protowire.ParseError(vn)
 			}
 			pos += vn
-			if num == 4 {
-				tc.Index = int(v)
+			if num == 6 {
+				tc.IsCustomToolCall = (v != 0)
 			}
 		case protowire.BytesType:
 			val, bn := protowire.ConsumeBytes(data[pos:])
@@ -690,6 +690,10 @@ func parseDevinToolCallDelta(data []byte) (DevinToolCallDelta, error) {
 				tc.Name = string(val)
 			case 3:
 				tc.Arguments = string(val)
+			case 4:
+				tc.InvalidJSONStr = string(val)
+			case 5:
+				tc.InvalidJSONErr = string(val)
 			}
 		default:
 			nSkip := protowire.ConsumeFieldValue(num, typ, data[pos:])
@@ -785,8 +789,8 @@ func parseDevinUsageField(data []byte) *DevinUsage {
 				u.PromptTokens += int64(v)
 			case 3: // Output tokens
 				u.CompletionTokens = int64(v)
-			case 4: // Additional context/system prompt tokens in OpenAI-family models (total prompt = 2 + 4)
-				u.PromptTokens += int64(v)
+			case 4: // Cache write tokens
+				u.CacheWriteTokens += int64(v)
 			case 5: // Cache read tokens
 				u.CachedTokens = int64(v)
 			case 6: // Status code
--- a/internal/runtime/executor/helps/devin_wire_test.go
+++ b/internal/runtime/executor/helps/devin_wire_test.go
@@ -679,9 +679,11 @@ func TestParseDevinUsageField_HeadersAndField4(t *testing.T) {
 		t.Fatal("expected non-nil usage")
 	}

-	// 3 + 58 = 61
-	if usage.PromptTokens != 61 {
-		t.Errorf("PromptTokens = %d, want 61 (3 turn + 58 context)", usage.PromptTokens)
+	if usage.PromptTokens != 3 {
+		t.Errorf("PromptTokens = %d, want 3", usage.PromptTokens)
+	}
+	if usage.CacheWriteTokens != 58 {
+		t.Errorf("CacheWriteTokens = %d, want 58", usage.CacheWriteTokens)
 	}
 	if usage.CompletionTokens != 39 {
 		t.Errorf("CompletionTokens = %d, want 39", usage.CompletionTokens)
@@ -875,3 +877,116 @@ func TestBuildDevinGetChatMessageRequest_FiltersAutomationUpdateAndObfuscatesDes
 		t.Fatalf("wire bytes should contain 'to a existing unified'")
 	}
 }
+
+func TestRegressionIssue5910_ClientMetadata(t *testing.T) {
+	b := BuildDevinClientMetadataBytes("test-session-token", "device-seed", "linux")
+	pos := 0
+	var ideName string
+	hasTag28 := false
+
+	for pos < len(b) {
+		num, typ, n := protowire.ConsumeTag(b[pos:])
+		if n <= 0 {
+			t.Fatalf("corrupt tag at %d", pos)
+		}
+		pos += n
+
+		if num == 1 && typ == protowire.BytesType {
+			val, bn := protowire.ConsumeBytes(b[pos:])
+			if bn <= 0 {
+				t.Fatalf("corrupt bytes at %d", pos)
+			}
+			pos += bn
+			ideName = string(val)
+		} else if num == 28 {
+			hasTag28 = true
+			nSkip := protowire.ConsumeFieldValue(num, typ, b[pos:])
+			if nSkip <= 0 {
+				t.Fatalf("corrupt field at %d", pos)
+			}
+			pos += nSkip
+		} else {
+			nSkip := protowire.ConsumeFieldValue(num, typ, b[pos:])
+			if nSkip <= 0 {
+				t.Fatalf("corrupt field at %d", pos)
+			}
+			pos += nSkip
+		}
+	}
+
+	if ideName != DevinDefaultClientName {
+		t.Errorf("BuildDevinClientMetadataBytes field 1 = %q, want %q", ideName, DevinDefaultClientName)
+	}
+	if hasTag28 {
+		t.Errorf("BuildDevinClientMetadataBytes should not emit field 28")
+	}
+}
+
+func TestRegressionIssue5910_UsageStatsCacheWriteTokens(t *testing.T) {
+	var f7Bytes []byte
+	// Field 2: input_tokens = 3
+	f7Bytes = protowire.AppendTag(f7Bytes, 2, protowire.VarintType)
+	f7Bytes = protowire.AppendVarint(f7Bytes, 3)
+
+	// Field 4: cache_write_tokens = 14361
+	f7Bytes = protowire.AppendTag(f7Bytes, 4, protowire.VarintType)
+	f7Bytes = protowire.AppendVarint(f7Bytes, 14361)
+
+	usage := parseDevinUsageField(f7Bytes)
+	if usage == nil {
+		t.Fatal("expected non-nil usage")
+	}
+
+	if usage.PromptTokens != 3 {
+		t.Errorf("PromptTokens = %d, want 3 (cache_write_tokens must not inflate prompt_tokens)", usage.PromptTokens)
+	}
+	if usage.CacheWriteTokens != 14361 {
+		t.Errorf("CacheWriteTokens = %d, want 14361", usage.CacheWriteTokens)
+	}
+}
+
+func TestRegressionIssue5910_ToolCallDeltaFields(t *testing.T) {
+	var tcBytes []byte
+	// Field 1: id
+	tcBytes = protowire.AppendTag(tcBytes, 1, protowire.BytesType)
+	tcBytes = protowire.AppendString(tcBytes, "call_999")
+	// Field 2: name
+	tcBytes = protowire.AppendTag(tcBytes, 2, protowire.BytesType)
+	tcBytes = protowire.AppendString(tcBytes, "custom_bash")
+	// Field 3: arguments
+	tcBytes = protowire.AppendTag(tcBytes, 3, protowire.BytesType)
+	tcBytes = protowire.AppendString(tcBytes, `{"cmd":"pwd"}`)
+	// Field 4: invalid_json_str
+	tcBytes = protowire.AppendTag(tcBytes, 4, protowire.BytesType)
+	tcBytes = protowire.AppendString(tcBytes, `pwd && ls`)
+	// Field 5: invalid_json_err
+	tcBytes = protowire.AppendTag(tcBytes, 5, protowire.BytesType)
+	tcBytes = protowire.AppendString(tcBytes, "syntax error near unexpected token")
+	// Field 6: is_custom_tool_call
+	tcBytes = protowire.AppendTag(tcBytes, 6, protowire.VarintType)
+	tcBytes = protowire.AppendVarint(tcBytes, 1)
+
+	tc, err := parseDevinToolCallDelta(tcBytes)
+	if err != nil {
+		t.Fatalf("parseDevinToolCallDelta failed: %v", err)
+	}
+
+	if tc.ID != "call_999" {
+		t.Errorf("tc.ID = %q, want call_999", tc.ID)
+	}
+	if tc.Name != "custom_bash" {
+		t.Errorf("tc.Name = %q, want custom_bash", tc.Name)
+	}
+	if tc.Arguments != `{"cmd":"pwd"}` {
+		t.Errorf("tc.Arguments = %q, want {\"cmd\":\"pwd\"}", tc.Arguments)
+	}
+	if tc.InvalidJSONStr != "pwd && ls" {
+		t.Errorf("tc.InvalidJSONStr = %q, want 'pwd && ls'", tc.InvalidJSONStr)
+	}
+	if tc.InvalidJSONErr != "syntax error near unexpected token" {
+		t.Errorf("tc.InvalidJSONErr = %q, want 'syntax error near unexpected token'", tc.InvalidJSONErr)
+	}
+	if !tc.IsCustomToolCall {
+		t.Errorf("tc.IsCustomToolCall = %v, want true", tc.IsCustomToolCall)
+	}
+}