Run serial-flagged tool calls in order while others fan out

Why: Mutual exclusion alone does not preserve order. With parallel
execution the host cannot keep order-sensitive tools (card streaming
begin/append/commit, plan updates) correct by locking a mutex, because
whichever goroutine grabs the lock first is not necessarily the first
declared call - a real failure observed in the backend card-stream test.

Evidence: go test ./... green; TestLoopSerialToolsKeepOrder asserts the
serial calls execute in declaration order (s1->s2->s3) while parallel
calls still fan out.

Impact: New Config.SerialTools set. In parallel mode, matching calls run
sequentially on the caller goroutine in model-declared order; others
stay concurrent. Empty set keeps the previous all-parallel behavior;
sequential mode is untouched.

Refs: Dev task feature/agent-e1-parallel-fixes (F5).

runtime/llm/client.go

@@ -20,6 +20,12 @@ type Client struct {
 	HTTPClient *http.Client
 	// Protocol 空 = openai-chat（零值兼容所有直接构造点）。
 	Protocol Protocol
+	// ReasoningSplit 在 openai-chat 请求体加 "reasoning_split": true：
+	// MiniMax 等 interleaved-thinking 提供商据此把思考放进独立的
+	// reasoning_content 字段而不是混在 content 的 <think> 块里（分流后
+	// SSE 的 reasoning_content 捕获照常生效）。对不认识该字段的服务器
+	// 属未知字段，多数按忽略处理；宿主按模型名自行决定是否置位。
+	ReasoningSplit bool
 }
 
 // Stream implements StreamFunc. Transport and provider failures are encoded as
@@ -67,6 +73,9 @@ func (c *Client) run(ctx context.Context, model string, messages []Message, tool
 	if effort := OpenAIReasoningEffort(opts.ThinkingLevel); effort != "" {
 		body["reasoning_effort"] = effort
 	}
+	if c.ReasoningSplit {
+		body["reasoning_split"] = true
+	}
 	if len(tools) > 0 {
 		body["tools"] = toProviderTools(tools)
 		body["tool_choice"] = "auto"

runtime/llm/reasoning_split_test.go

@@ -0,0 +1,47 @@
+package llm
+
+import (
+	"context"
+	"encoding/json"
+	"io"
+	"net/http"
+	"net/http/httptest"
+	"strings"
+	"testing"
+)
+
+// TestOpenAIChatBodyReasoningSplit 验证 Client.ReasoningSplit 置位时
+// openai-chat 请求体带 "reasoning_split": true，未置位时不带（未知字段
+// 不外溢给别的提供商）。
+func TestOpenAIChatBodyReasoningSplit(t *testing.T) {
+	var bodies []string
+	srv := httptest.NewServer(http.HandlerFunc(func(w http.ResponseWriter, r *http.Request) {
+		raw, _ := io.ReadAll(r.Body)
+		bodies = append(bodies, string(raw))
+		w.Header().Set("Content-Type", "text/event-stream")
+		_, _ = io.WriteString(w, "data: [DONE]\n\n")
+	}))
+	defer srv.Close()
+
+	drain := func(c *Client) {
+		for range c.Stream(context.Background(), "MiniMax-M3",
+			[]Message{{Role: RoleUser, Content: "hi"}}, nil, StreamOptions{}) {
+		}
+	}
+
+	drain(&Client{BaseURL: srv.URL, APIKey: "k"})
+	drain(&Client{BaseURL: srv.URL, APIKey: "k", ReasoningSplit: true})
+	if len(bodies) != 2 {
+		t.Fatalf("requests = %d", len(bodies))
+	}
+	if strings.Contains(bodies[0], "reasoning_split") {
+		t.Fatalf("flag off must not leak the field: %s", bodies[0])
+	}
+	var got map[string]any
+	if err := json.Unmarshal([]byte(bodies[1]), &got); err != nil {
+		t.Fatal(err)
+	}
+	if v, ok := got["reasoning_split"].(bool); !ok || !v {
+		t.Fatalf("flag on must set reasoning_split=true: %s", bodies[1])
+	}
+}

runtime/loop.go

@@ -38,6 +38,9 @@ type Event struct {
 	Message llm.Message
 	Tool    string
 	Text    string
+	// CallID 在 tool_execution_start / tool_execution_end 时携带工具调用的
+	// ID：并行执行下同回合多个调用并发进出，消费方按 ID 配对而非按顺序。
+	CallID string
 	// Detail carries tool arguments/results or reasoning deltas for
 	// process-visualization consumers; log writers may ignore it.
 	Detail string
@@ -66,6 +69,9 @@ type Config struct {
 	Trace         func(name string, attrs map[string]string) func()
 	MaxTurns      int
 	ToolExecution string
+	// SerialTools 是并行执行下的顺序敏感工具名集合：这些调用按模型给出的
+	// 原顺序在调用方 goroutine 逐个执行（互斥且保序），其余调用并发执行。
+	SerialTools map[string]bool
 	GetSteering   func() []llm.Message
 	GetFollowUp   func() []llm.Message
 	// OnEvent 观察循环事件。同一次 Run 内不会并发调用
@@ -594,7 +600,10 @@ func serializeHostCallbacks(cfg *Config) {
 
 func executeTools(ctx context.Context, cfg Config, assistant llm.Message) []llm.Message {
 	if cfg.ToolExecution == ExecParallel && len(assistant.ToolCalls) > 1 {
-		return executeParallel(ctx, cfg, assistant)
+		if len(cfg.SerialTools) == 0 {
+			return executeParallel(ctx, cfg, assistant)
+		}
+		return executeMixed(ctx, cfg, assistant)
 	}
 	out := make([]llm.Message, 0, len(assistant.ToolCalls))
 	for _, call := range assistant.ToolCalls {
@@ -607,6 +616,35 @@ func executeTools(ctx context.Context, cfg Config, assistant llm.Message) []llm.
 	return out
 }
 
+// executeMixed 并行执行非串行调用，串行名（SerialTools）按模型给出的原顺序
+// 在调用方 goroutine 逐个执行——互斥且保序（卡片流等有先后依赖的工具不能只
+// 靠互斥：并发派发下先抢到锁的不一定是先声明的调用）。
+func executeMixed(ctx context.Context, cfg Config, assistant llm.Message) []llm.Message {
+	out := make([]llm.Message, len(assistant.ToolCalls))
+	var wg sync.WaitGroup
+	serialIdx := make([]int, 0, len(assistant.ToolCalls))
+	for i, call := range assistant.ToolCalls {
+		if cfg.SerialTools[call.Name] {
+			serialIdx = append(serialIdx, i)
+			continue
+		}
+		wg.Add(1)
+		go func(i int, call llm.ToolCall) {
+			defer wg.Done()
+			out[i] = runOne(ctx, cfg, call)
+		}(i, call)
+	}
+	for _, i := range serialIdx {
+		if err := ctx.Err(); err != nil {
+			out[i] = cancelledToolMessage(assistant.ToolCalls[i], err)
+			continue
+		}
+		out[i] = runOne(ctx, cfg, assistant.ToolCalls[i])
+	}
+	wg.Wait()
+	return out
+}
+
 func executeParallel(ctx context.Context, cfg Config, assistant llm.Message) []llm.Message {
 	out := make([]llm.Message, len(assistant.ToolCalls))
 	var wg sync.WaitGroup
@@ -706,22 +744,38 @@ func noteToolHook(cfg Config, call llm.ToolCall, msg llm.Message) {
 	cfg.Hooks.Emit(hooks.Call{Event: ev, Tool: call.Name, Args: call.Arguments, Result: msg.Content, IsError: msg.IsError})
 }
 
+// toolCallKey 是把当前工具调用注入 Execute ctx 的私有键：并行执行下工具实现
+// 需要知道自己服务于哪次调用（配对过程事件、按调用记账），经
+// ToolCallFromContext 读取；旧宿主/直接构造的 ctx 返回 ok=false。
+type toolCallKey struct{}
+
+// WithToolCall 把 call 挂进 ctx（内核 runOne 在执行前注入）。
+func WithToolCall(ctx context.Context, call llm.ToolCall) context.Context {
+	return context.WithValue(ctx, toolCallKey{}, call)
+}
+
+// ToolCallFromContext 取出当前正在执行的调用；不存在时 ok=false。
+func ToolCallFromContext(ctx context.Context) (llm.ToolCall, bool) {
+	call, ok := ctx.Value(toolCallKey{}).(llm.ToolCall)
+	return call, ok
+}
+
 func runOne(ctx context.Context, cfg Config, call llm.ToolCall) llm.Message {
 	if err := ctx.Err(); err != nil {
 		return cancelledToolMessage(call, err)
 	}
 	if why := rejectIncompleteToolCall(call); why != "" {
 		msg := llm.Message{Role: llm.RoleTool, Name: call.Name, ToolCallID: call.ID, Content: why, IsError: true}
-		emit(cfg, Event{Type: "tool_execution_start", Tool: call.Name, Detail: call.Arguments})
-		emit(cfg, Event{Type: "tool_execution_end", Tool: call.Name, Text: why})
+		emit(cfg, Event{Type: "tool_execution_start", Tool: call.Name, CallID: call.ID, Detail: call.Arguments})
+		emit(cfg, Event{Type: "tool_execution_end", Tool: call.Name, CallID: call.ID, Text: why})
 		if cfg.AfterTool != nil {
 			cfg.AfterTool(call, msg)
 		}
 		return msg
 	}
-	emit(cfg, Event{Type: "tool_execution_start", Tool: call.Name, Detail: call.Arguments})
+	emit(cfg, Event{Type: "tool_execution_start", Tool: call.Name, CallID: call.ID, Detail: call.Arguments})
 	if gated, msg, blocked := gateTool(ctx, cfg, call); blocked {
-		emit(cfg, Event{Type: "tool_execution_end", Tool: call.Name, Text: msg.Content})
+		emit(cfg, Event{Type: "tool_execution_end", Tool: call.Name, CallID: call.ID, Text: msg.Content})
 		noteToolHook(cfg, call, msg)
 		if cfg.AfterTool != nil {
 			cfg.AfterTool(call, msg)
@@ -733,7 +787,7 @@ func runOne(ctx context.Context, cfg Config, call llm.ToolCall) llm.Message {
 	if cfg.BeforeTool != nil {
 		if err := cfg.BeforeTool(call); err != nil {
 			msg := llm.Message{Role: llm.RoleTool, Name: call.Name, ToolCallID: call.ID, Content: err.Error(), IsError: true}
-			emit(cfg, Event{Type: "tool_execution_end", Tool: call.Name, Text: msg.Content})
+			emit(cfg, Event{Type: "tool_execution_end", Tool: call.Name, CallID: call.ID, Text: msg.Content})
 			noteToolHook(cfg, call, msg)
 			if cfg.AfterTool != nil {
 				cfg.AfterTool(call, msg)
@@ -744,7 +798,7 @@ func runOne(ctx context.Context, cfg Config, call llm.ToolCall) llm.Message {
 	emit(cfg, Event{Type: "tool_execution_update", Tool: call.Name, Detail: call.Arguments})
 	if err := commitToolIntent(cfg, call); err != nil {
 		msg := llm.Message{Role: llm.RoleTool, Name: call.Name, ToolCallID: call.ID, Content: err.Error(), IsError: true}
-		emit(cfg, Event{Type: "tool_execution_end", Tool: call.Name, Text: msg.Content})
+		emit(cfg, Event{Type: "tool_execution_end", Tool: call.Name, CallID: call.ID, Text: msg.Content})
 		return msg
 	}
 	endSpan := startSpan(cfg, "execute_tool "+call.Name, map[string]string{
@@ -754,6 +808,7 @@ func runOne(ctx context.Context, cfg Config, call llm.ToolCall) llm.Message {
 	defer endSpan()
 	res := tools.Result{Content: "no tools configured", IsError: true}
 	known := cfg.Tools != nil && cfg.Tools.Has(call.Name)
+	ctx = WithToolCall(ctx, call)
 	switch {
 	case known:
 		res = cfg.Tools.Execute(ctx, call.Name, json.RawMessage(call.Arguments))
@@ -769,7 +824,7 @@ func runOne(ctx context.Context, cfg Config, call llm.ToolCall) llm.Message {
 		Content:    res.Content,
 		IsError:    res.IsError,
 	}
-	emit(cfg, Event{Type: "tool_execution_end", Tool: call.Name, Text: res.Content})
+	emit(cfg, Event{Type: "tool_execution_end", Tool: call.Name, CallID: call.ID, Text: res.Content})
 	noteToolHook(cfg, call, msg)
 	if cfg.AfterTool != nil {
 		cfg.AfterTool(call, msg)

runtime/loop_callctx_test.go

@@ -0,0 +1,69 @@
+package runtime
+
+import (
+	"context"
+	"encoding/json"
+	"sync"
+	"testing"
+
+	"agent-native/runtime/llm"
+	"agent-native/runtime/tools"
+)
+
+// ctxEchoTool 把自己从 ctx 读到的调用 ID 记下来（并发安全），验证内核
+// 在 Execute 前注入了当前调用。
+type ctxEchoTool struct {
+	mu  sync.Mutex
+	ids []string
+	bad int
+}
+
+func (c *ctxEchoTool) Name() string                { return "echo" }
+func (c *ctxEchoTool) Description() string         { return "echo" }
+func (c *ctxEchoTool) Parameters() json.RawMessage { return json.RawMessage(`{"type":"object"}`) }
+func (c *ctxEchoTool) Execute(ctx context.Context, _ json.RawMessage) (tools.Result, error) {
+	call, ok := ToolCallFromContext(ctx)
+	c.mu.Lock()
+	defer c.mu.Unlock()
+	if !ok || call.ID == "" || call.Name != "echo" {
+		c.bad++
+		return tools.Result{Content: "call missing from context", IsError: true}, nil
+	}
+	c.ids = append(c.ids, call.ID)
+	return tools.Result{Content: "ok " + call.ID}, nil
+}
+
+// TestLoopToolSeesOwnCallInContext 并行执行两个调用，各自读到自己的 ID。
+func TestLoopToolSeesOwnCallInContext(t *testing.T) {
+	tool := &ctxEchoTool{}
+	reg := tools.NewRegistry(t.TempDir())
+	reg.Register(tool)
+	res, err := Run(context.Background(), Config{
+		Model:         "fake",
+		Tools:         reg,
+		ToolExecution: ExecParallel,
+		Stream: script(
+			llm.Message{Role: llm.RoleAssistant, StopReason: llm.StopToolCalls, ToolCalls: []llm.ToolCall{
+				{ID: "pa", Name: "echo", Arguments: `{}`},
+				{ID: "pb", Name: "echo", Arguments: `{}`},
+			}},
+			llm.Message{Role: llm.RoleAssistant, Content: "done", StopReason: llm.StopEnd},
+		),
+	}, llm.Message{Role: llm.RoleUser, Content: "go"})
+	if err != nil {
+		t.Fatal(err)
+	}
+	tool.mu.Lock()
+	defer tool.mu.Unlock()
+	if tool.bad != 0 || len(tool.ids) != 2 {
+		t.Fatalf("bad=%d ids=%v", tool.bad, tool.ids)
+	}
+	seen := map[string]bool{}
+	for _, id := range tool.ids {
+		seen[id] = true
+	}
+	if !seen["pa"] || !seen["pb"] {
+		t.Fatalf("ids = %v", tool.ids)
+	}
+	_ = res
+}

runtime/loop_callid_test.go

@@ -0,0 +1,49 @@
+package runtime
+
+import (
+	"context"
+	"testing"
+
+	"agent-native/runtime/llm"
+	"agent-native/runtime/tools"
+)
+
+// TestLoopToolEventsCarryCallID 验证 tool_execution_start/end 事件带工具调用
+// ID：并行执行下消费方按 ID 配对（回归锚点，防止字段被删后顺序假设回归）。
+func TestLoopToolEventsCarryCallID(t *testing.T) {
+	hits := 0
+	reg := tools.NewRegistry(t.TempDir())
+	reg.Register(hitTool{hits: &hits})
+	var starts, ends []Event
+	res, err := Run(context.Background(), Config{
+		Model: "fake",
+		Tools: reg,
+		Stream: script(
+			llm.Message{Role: llm.RoleAssistant, StopReason: llm.StopToolCalls, ToolCalls: []llm.ToolCall{{
+				ID: "call-7", Name: "echo", Arguments: `{}`,
+			}}},
+			llm.Message{Role: llm.RoleAssistant, Content: "done", StopReason: llm.StopEnd},
+		),
+		OnEvent: func(ev Event) {
+			switch ev.Type {
+			case "tool_execution_start":
+				starts = append(starts, ev)
+			case "tool_execution_end":
+				ends = append(ends, ev)
+			}
+		},
+	}, llm.Message{Role: llm.RoleUser, Content: "go"})
+	if err != nil {
+		t.Fatal(err)
+	}
+	if hits != 1 || len(starts) != 1 || len(ends) != 1 {
+		t.Fatalf("hits=%d starts=%d ends=%d", hits, len(starts), len(ends))
+	}
+	if starts[0].CallID != "call-7" || starts[0].Tool != "echo" {
+		t.Fatalf("start = %+v", starts[0])
+	}
+	if ends[0].CallID != "call-7" || ends[0].Tool != "echo" {
+		t.Fatalf("end = %+v", ends[0])
+	}
+	_ = res
+}

runtime/loop_serial_partition_test.go

@@ -0,0 +1,72 @@
+package runtime
+
+import (
+	"context"
+	"encoding/json"
+	"sync"
+	"testing"
+
+	"agent-native/runtime/llm"
+	"agent-native/runtime/tools"
+)
+
+// orderTool 记录自己的执行序号（并发安全）。
+type orderTool struct {
+	name string
+	mu   sync.Mutex
+	seq  []string
+}
+
+func (o *orderTool) Name() string                { return o.name }
+func (o *orderTool) Description() string         { return o.name }
+func (o *orderTool) Parameters() json.RawMessage { return json.RawMessage(`{"type":"object"}`) }
+func (o *orderTool) Execute(ctx context.Context, _ json.RawMessage) (tools.Result, error) {
+	id := o.name
+	if call, ok := ToolCallFromContext(ctx); ok {
+		id = call.ID
+	}
+	o.mu.Lock()
+	o.seq = append(o.seq, id)
+	o.mu.Unlock()
+	return tools.Result{Content: "ran " + id}, nil
+}
+
+// TestLoopSerialToolsKeepOrder 并行模式下 SerialTools 命中的调用按声明顺序执行
+// （互斥不保序不够：卡片流等有先后依赖），未命中的并发执行。
+func TestLoopSerialToolsKeepOrder(t *testing.T) {
+	cards := &orderTool{name: "card"}
+	reads := &orderTool{name: "read"}
+	reg := tools.NewRegistry(t.TempDir())
+	reg.Register(cards)
+	reg.Register(reads)
+	res, err := Run(context.Background(), Config{
+		Model: "fake", Tools: reg,
+		ToolExecution: ExecParallel,
+		SerialTools:   map[string]bool{"card": true},
+		Stream: script(
+			llm.Message{Role: llm.RoleAssistant, StopReason: llm.StopToolCalls, ToolCalls: []llm.ToolCall{
+				{ID: "s1", Name: "card", Arguments: `{}`},
+				{ID: "p1", Name: "read", Arguments: `{}`},
+				{ID: "s2", Name: "card", Arguments: `{}`},
+				{ID: "s3", Name: "card", Arguments: `{}`},
+			}},
+			llm.Message{Role: llm.RoleAssistant, Content: "done", StopReason: llm.StopEnd},
+		),
+	}, llm.Message{Role: llm.RoleUser, Content: "go"})
+	if err != nil {
+		t.Fatal(err)
+	}
+	cards.mu.Lock()
+	got := append([]string(nil), cards.seq...)
+	cards.mu.Unlock()
+	if len(got) != 3 || got[0] != "s1" || got[1] != "s2" || got[2] != "s3" {
+		t.Fatalf("card executions must follow declaration order s1→s2→s3, got %v", got)
+	}
+	reads.mu.Lock()
+	nRead := len(reads.seq)
+	reads.mu.Unlock()
+	if nRead != 1 {
+		t.Fatalf("read executions = %d", nRead)
+	}
+	_ = res
+}