Merge pull request 'Add optional harness hooks to the native turn loop' (#6) from feature/agent-harness-p1 into main

runtime/llm/sse.go

@@ -164,7 +164,7 @@ func emitParsed(ctx context.Context, r io.Reader, out chan<- Event) {
 	select {
 	case <-ctx.Done():
 		emitAbort(out, content.String(), ctx.Err())
-	case out <- Event{Kind: KindMessageEnd, Message: msg, Usage: usage}:
+	case out <- Event{Kind: KindMessageEnd, Message: msg, Usage: usage, FinishReason: finished}:
 	}
 }
 

runtime/llm/sse_test.go

@@ -38,6 +38,29 @@ func TestParseSSETextAndUsage(t *testing.T) {
 	if end.Usage.TotalTokens != 5 {
 		t.Fatalf("usage %#v", end.Usage)
 	}
+	if end.FinishReason != "stop" {
+		t.Fatalf("finish reason %q, want stop", end.FinishReason)
+	}
+}
+
+func TestParseSSEFinishReasonLengthKeepsStopEnd(t *testing.T) {
+	body := strings.Join([]string{
+		`data: {"choices":[{"delta":{"content":"半"},"finish_reason":"length"}]}`,
+		`data: [DONE]`,
+		"",
+	}, "\n")
+	var end Event
+	for ev := range ParseSSE(context.Background(), strings.NewReader(body)) {
+		if ev.Kind == KindMessageEnd {
+			end = ev
+		}
+	}
+	if end.FinishReason != "length" {
+		t.Fatalf("finish reason %q, want length", end.FinishReason)
+	}
+	if end.Message.StopReason != StopEnd {
+		t.Fatalf("stop reason %s, want end so existing callers stay unchanged", end.Message.StopReason)
+	}
 }
 
 func TestParseSSEToolCallIncrements(t *testing.T) {

runtime/llm/types.go

@@ -90,9 +90,12 @@ type Event struct {
 	ArgumentsDelta string
 	Message        Message
 	Usage          Usage
-	Err            error
-	Retryable      bool
-	ErrorMessage   string
+	// FinishReason 是供应商原始结束原因（stop、length、tool_calls 等）。
+	// 空表示这条事件不携带结束原因。StopReason 的映射不受此字段影响。
+	FinishReason string
+	Err          error
+	Retryable    bool
+	ErrorMessage string
 }
 
 // StreamOptions tunes one provider request.

runtime/loop.go

@@ -18,8 +18,15 @@ import (
 const (
 	ExecSequential = "sequential"
 	ExecParallel   = "parallel"
+
+	maxStopContinuations   = 3
+	maxLengthContinuations = 3
+	maxRepeatReminders     = 3
 )
 
+const repeatReminderText = "你已连续 3 次以相同参数调用同一工具。如果结果不符合预期，请更换参数或做法，不要再重复这一调用。"
+const lengthContinuePrompt = "上一段回复因为达到长度上限被截断。请从断点接着写，不要重复已经写出的内容。"
+
 // Event is a lifecycle notice for logs and UI.
 type Event struct {
 	Type    string
@@ -29,9 +36,11 @@ type Event struct {
 	// Detail carries tool arguments/results or reasoning deltas for
 	// process-visualization consumers; log writers may ignore it.
 	Detail string
-	// PromptTokens / CompletionTokens 只在 type=usage 时有意义。
+	// PromptTokens / CompletionTokens / TotalTokens 在 type=usage 或
+	// type=message_end 时有意义，供调用方累计用量。
 	PromptTokens     int
 	CompletionTokens int
+	TotalTokens      int
 }
 
 // Config drives one agent session.
@@ -57,6 +66,41 @@ type Config struct {
 	// Budget 可选。任一上限命中即停止并返回 budget_exceeded 错误（StopError）；
 	// 零值字段不设限。截止时刻与外层 ctx 先到者生效。
 	Budget *Budget
+	// History 是会话转录为空时的初始转录。以 system 开头时，不再前置
+	// SystemPrompt 和工具清单。不回写进 Session。
+	History []llm.Message
+	// Stop 在模型不再调用工具、且本轮没有 follow-up 时询问是否续跑。
+	// nil 表示不询问。内核最多接受 3 次 Continue。
+	Stop func(last llm.Message, info StopInfo) StopDecision
+	// BeforeTool 返回错误时不执行工具，但仍写一条 IsError 的 tool 结果。
+	BeforeTool func(call llm.ToolCall) error
+	// UnknownTool 处理注册表里没有的调用。nil 时仍返回 unknown tool。
+	UnknownTool func(call llm.ToolCall) tools.Result
+	// AfterTool 只观察工具结果，不能改写。
+	AfterTool func(call llm.ToolCall, res llm.Message)
+	// Reminders 在每次调用模型前提供提醒。正文包在 system-reminder 里，
+	// 只进入本次请求，不写进 Session。
+	Reminders func(turn int) []string
+	// RepeatGuard 为真时，同名且参数按键排序后相同的调用连续 3 次，注入一条提醒。
+	// 每个 run 最多 3 条，不中断循环。
+	RepeatGuard bool
+	// ContinueOnLength 为真时，FinishReason=length 且没有工具调用就续写，最多 3 次。
+	// 续写提示不写进 Session。第 4 次截断返回错误。
+	ContinueOnLength bool
+}
+
+// StopInfo 是 Stop 钩子看到的本 run 进度。
+type StopInfo struct {
+	Continuations int
+	ToolCalls     int
+}
+
+// StopDecision 决定模型停下之后是否再跑一轮。
+type StopDecision struct {
+	Continue    bool
+	Replacement string
+	Reason      string
+	Notice      string
 }
 
 // Budget 是循环层的执行预算；与 exec.Budget 字段同构，由适配层换算。
@@ -73,6 +117,7 @@ type Result struct {
 	Messages   []llm.Message
 	StopReason llm.StopReason
 	Usage      llm.Usage
+	Notice     string
 	Err        error
 	Retryable  bool
 }
@@ -96,6 +141,9 @@ func Run(ctx context.Context, cfg Config, user llm.Message) (Result, error) {
 		}
 		messages = loaded
 	}
+	if len(messages) == 0 && len(cfg.History) > 0 {
+		messages = append([]llm.Message{}, cfg.History...)
+	}
 	if len(messages) == 0 || messages[0].Role != llm.RoleSystem {
 		sys := llm.Message{Role: llm.RoleSystem, Content: systemText(cfg)}
 		messages = append([]llm.Message{sys}, messages...)
@@ -112,6 +160,10 @@ func Run(ctx context.Context, cfg Config, user llm.Message) (Result, error) {
 	var last llm.Message
 	var usage llm.Usage
 	toolCalls := 0
+	stopContinues := 0
+	lengthContinues := 0
+	lengthNudge := false
+	var repeats repeatState
 	for turn := 0; turn < cfg.MaxTurns; turn++ {
 		if ctx.Err() != nil {
 			cfg.logger().Error("runtime turn aborted", ctx.Err())
@@ -122,7 +174,8 @@ func Run(ctx context.Context, cfg Config, user llm.Message) (Result, error) {
 		}
 		emit(cfg, Event{Type: "turn_start"})
 		emit(cfg, Event{Type: "message_start"})
-		end, errEv := consume(ctx, cfg, messages)
+		request := withEphemeral(messages, turnReminders(cfg, turn, &repeats, &lengthNudge))
+		end, errEv := consume(ctx, cfg, request)
 		if errEv != nil {
 			cfg.logger().Error("runtime turn failed", errEv, "retryable", end.Retryable)
 			return Result{Text: end.Message.Content, Messages: messages, StopReason: end.Message.StopReason, Err: errEv, Retryable: end.Retryable}, nil
@@ -135,7 +188,7 @@ func Run(ctx context.Context, cfg Config, user llm.Message) (Result, error) {
 			return Result{}, err
 		}
 		messages = append(messages, last)
-		emit(cfg, Event{Type: "message_end", Message: last})
+		emit(cfg, Event{Type: "message_end", Message: last, PromptTokens: end.Usage.PromptTokens, CompletionTokens: end.Usage.CompletionTokens, TotalTokens: end.Usage.TotalTokens})
 
 		if last.StopReason == llm.StopToolCalls && len(last.ToolCalls) > 0 {
 			if cfg.Budget != nil && cfg.Budget.MaxToolCalls > 0 && toolCalls+len(last.ToolCalls) > cfg.Budget.MaxToolCalls {
@@ -143,6 +196,9 @@ func Run(ctx context.Context, cfg Config, user llm.Message) (Result, error) {
 				return budgetStop(cfg, last, messages, usage, reason), nil
 			}
 			toolCalls += len(last.ToolCalls)
+			if cfg.RepeatGuard {
+				repeats.observe(last.ToolCalls)
+			}
 			results := executeTools(ctx, cfg, last)
 			for _, res := range results {
 				if err := persist(cfg, res); err != nil {
@@ -162,6 +218,16 @@ func Run(ctx context.Context, cfg Config, user llm.Message) (Result, error) {
 			continue
 		}
 		emit(cfg, Event{Type: "turn_end", Message: last})
+		if cfg.ContinueOnLength && end.FinishReason == "length" && len(last.ToolCalls) == 0 {
+			if lengthContinues >= maxLengthContinuations {
+				errCap := errString("length continuation exceeded")
+				cfg.logger().Error("runtime length continuation exceeded", errCap)
+				return Result{Text: last.Content, Messages: messages, StopReason: llm.StopError, Usage: usage, Err: errCap}, nil
+			}
+			lengthContinues++
+			lengthNudge = true
+			continue
+		}
 		if follow := drain(cfg.GetFollowUp); len(follow) > 0 {
 			for _, m := range follow {
 				if err := persist(cfg, m); err != nil {
@@ -171,10 +237,26 @@ func Run(ctx context.Context, cfg Config, user llm.Message) (Result, error) {
 			}
 			continue
 		}
-		emit(cfg, Event{Type: "agent_end", Message: last})
-		emit(cfg, Event{Type: "agent_settled", Message: last})
-		emitUsage(cfg, usage)
-		return Result{Text: last.Content, Messages: messages, StopReason: last.StopReason, Usage: usage}, nil
+		if cfg.Stop != nil {
+			decision := cfg.Stop(last, StopInfo{Continuations: stopContinues, ToolCalls: toolCalls})
+			if decision.Continue && stopContinues < maxStopContinuations {
+				stopContinues++
+				if decision.Replacement != "" && len(messages) > 0 {
+					messages[len(messages)-1].Content = decision.Replacement
+					last.Content = decision.Replacement
+				}
+				if decision.Reason != "" {
+					reasonMsg := llm.Message{Role: llm.RoleUser, Content: decision.Reason}
+					if err := persist(cfg, reasonMsg); err != nil {
+						return Result{}, err
+					}
+					messages = append(messages, reasonMsg)
+				}
+				continue
+			}
+			return settle(cfg, last, messages, usage, decision.Notice), nil
+		}
+		return settle(cfg, last, messages, usage, ""), nil
 	}
 	errCap := errString("max turns exceeded")
 	cfg.logger().Error("runtime turn cap hit", errCap)
@@ -290,19 +372,118 @@ func executeParallel(ctx context.Context, cfg Config, assistant llm.Message) []l
 
 func runOne(ctx context.Context, cfg Config, call llm.ToolCall) llm.Message {
 	emit(cfg, Event{Type: "tool_execution_start", Tool: call.Name, Detail: call.Arguments})
+	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})
+			if cfg.AfterTool != nil {
+				cfg.AfterTool(call, msg)
+			}
+			return msg
+		}
+	}
 	emit(cfg, Event{Type: "tool_execution_update", Tool: call.Name, Detail: call.Arguments})
 	res := tools.Result{Content: "no tools configured", IsError: true}
-	if cfg.Tools != nil {
+	known := cfg.Tools != nil && cfg.Tools.Has(call.Name)
+	switch {
+	case known:
+		res = cfg.Tools.Execute(ctx, call.Name, json.RawMessage(call.Arguments))
+	case cfg.UnknownTool != nil:
+		res = cfg.UnknownTool(call)
+	case cfg.Tools != nil:
 		res = cfg.Tools.Execute(ctx, call.Name, json.RawMessage(call.Arguments))
 	}
-	emit(cfg, Event{Type: "tool_execution_end", Tool: call.Name, Text: res.Content})
-	return llm.Message{
+	msg := llm.Message{
 		Role:       llm.RoleTool,
 		Name:       call.Name,
 		ToolCallID: call.ID,
 		Content:    res.Content,
 		IsError:    res.IsError,
 	}
+	emit(cfg, Event{Type: "tool_execution_end", Tool: call.Name, Text: res.Content})
+	if cfg.AfterTool != nil {
+		cfg.AfterTool(call, msg)
+	}
+	return msg
+}
+
+func settle(cfg Config, last llm.Message, messages []llm.Message, usage llm.Usage, notice string) Result {
+	emit(cfg, Event{Type: "agent_end", Message: last})
+	emit(cfg, Event{Type: "agent_settled", Message: last})
+	emitUsage(cfg, usage)
+	return Result{Text: last.Content, Messages: messages, StopReason: last.StopReason, Usage: usage, Notice: notice}
+}
+
+func reminderMessage(text string) llm.Message {
+	return llm.Message{Role: llm.RoleUser, Content: "<system-reminder>\n" + text + "\n</system-reminder>"}
+}
+
+func turnReminders(cfg Config, turn int, repeats *repeatState, lengthNudge *bool) []llm.Message {
+	var extra []llm.Message
+	if cfg.Reminders != nil {
+		for _, text := range cfg.Reminders(turn) {
+			if strings.TrimSpace(text) == "" {
+				continue
+			}
+			extra = append(extra, reminderMessage(text))
+		}
+	}
+	if lengthNudge != nil && *lengthNudge {
+		extra = append(extra, llm.Message{Role: llm.RoleUser, Content: lengthContinuePrompt})
+		*lengthNudge = false
+	}
+	if cfg.RepeatGuard && repeats != nil {
+		due := repeats.streak/3 - repeats.reminders
+		room := maxRepeatReminders - repeats.reminders
+		if due > room {
+			due = room
+		}
+		for i := 0; i < due; i++ {
+			extra = append(extra, reminderMessage(repeatReminderText))
+			repeats.reminders++
+		}
+	}
+	return extra
+}
+
+func withEphemeral(messages, extra []llm.Message) []llm.Message {
+	if len(extra) == 0 {
+		return messages
+	}
+	out := make([]llm.Message, 0, len(messages)+len(extra))
+	out = append(out, messages...)
+	out = append(out, extra...)
+	return out
+}
+
+type repeatState struct {
+	sig       string
+	streak    int
+	reminders int
+}
+
+func (s *repeatState) observe(calls []llm.ToolCall) {
+	for _, call := range calls {
+		sig := call.Name + "\n" + canonicalArgs(call.Arguments)
+		if sig == s.sig {
+			s.streak++
+			continue
+		}
+		s.sig = sig
+		s.streak = 1
+	}
+}
+
+func canonicalArgs(raw string) string {
+	var v any
+	if err := json.Unmarshal([]byte(raw), &v); err != nil {
+		return raw
+	}
+	b, err := json.Marshal(v)
+	if err != nil {
+		return raw
+	}
+	return string(b)
 }
 
 func persist(cfg Config, msg llm.Message) error {
@@ -320,7 +501,7 @@ func drain(fn func() []llm.Message) []llm.Message {
 }
 
 func emitUsage(cfg Config, usage llm.Usage) {
-	emit(cfg, Event{Type: "usage", PromptTokens: usage.PromptTokens, CompletionTokens: usage.CompletionTokens})
+	emit(cfg, Event{Type: "usage", PromptTokens: usage.PromptTokens, CompletionTokens: usage.CompletionTokens, TotalTokens: usage.TotalTokens})
 }
 
 func emit(cfg Config, ev Event) {

runtime/loop_harness_test.go

@@ -0,0 +1,332 @@
+package runtime
+
+import (
+	"context"
+	"encoding/json"
+	"errors"
+	"path/filepath"
+	"strings"
+	"sync"
+	"testing"
+
+	"agent-native/runtime/llm"
+	"agent-native/runtime/session"
+	"agent-native/runtime/tools"
+)
+
+func TestLoopStopContinuesUntilCap(t *testing.T) {
+	var calls int
+	res, err := Run(context.Background(), Config{
+		Model:    "fake",
+		MaxTurns: 8,
+		Stream: script(
+			llm.Message{Role: llm.RoleAssistant, Content: "forged-claim", StopReason: llm.StopEnd},
+			llm.Message{Role: llm.RoleAssistant, Content: "forged-claim", StopReason: llm.StopEnd},
+			llm.Message{Role: llm.RoleAssistant, Content: "forged-claim", StopReason: llm.StopEnd},
+			llm.Message{Role: llm.RoleAssistant, Content: "forged-claim", StopReason: llm.StopEnd},
+		),
+		Stop: func(last llm.Message, info StopInfo) StopDecision {
+			calls++
+			if info.Continuations != calls-1 {
+				t.Errorf("continuations = %d, want %d", info.Continuations, calls-1)
+			}
+			return StopDecision{Continue: true, Replacement: "已作废", Reason: "再试", Notice: "到顶了"}
+		},
+	}, llm.Message{Role: llm.RoleUser, Content: "go"})
+	if err != nil {
+		t.Fatal(err)
+	}
+	if calls != 4 {
+		t.Fatalf("stop calls = %d, want 4", calls)
+	}
+	if res.Err != nil || res.Notice != "到顶了" {
+		t.Fatalf("notice %q err %v", res.Notice, res.Err)
+	}
+	replaced, forged := 0, 0
+	for _, m := range res.Messages {
+		if m.Role != llm.RoleAssistant {
+			continue
+		}
+		if m.Content == "已作废" {
+			replaced++
+		}
+		if m.Content == "forged-claim" {
+			forged++
+		}
+	}
+	if replaced != 3 || forged != 1 {
+		t.Fatalf("replaced=%d forged=%d transcript=%s", replaced, forged, transcriptText(res.Messages))
+	}
+	if strings.Count(transcriptText(res.Messages), "再试") != 3 {
+		t.Fatalf("reasons = %s", transcriptText(res.Messages))
+	}
+}
+
+func TestLoopBeforeToolRejectPairsResult(t *testing.T) {
+	hits := 0
+	reg := tools.NewRegistry(t.TempDir())
+	reg.Register(hitTool{hits: &hits})
+	var seen llm.Message
+	res, err := Run(context.Background(), Config{
+		Model: "fake",
+		Tools: reg,
+		Stream: script(
+			llm.Message{Role: llm.RoleAssistant, StopReason: llm.StopToolCalls, ToolCalls: []llm.ToolCall{{
+				ID: "c1", Name: "echo", Arguments: `{}`,
+			}}},
+			llm.Message{Role: llm.RoleAssistant, Content: "done", StopReason: llm.StopEnd},
+		),
+		BeforeTool: func(call llm.ToolCall) error {
+			return errors.New("不允许")
+		},
+		AfterTool: func(call llm.ToolCall, msg llm.Message) {
+			seen = msg
+		},
+	}, llm.Message{Role: llm.RoleUser, Content: "go"})
+	if err != nil {
+		t.Fatal(err)
+	}
+	if hits != 0 {
+		t.Fatalf("tool executed %d times", hits)
+	}
+	if !seen.IsError || seen.Content != "不允许" || seen.ToolCallID != "c1" {
+		t.Fatalf("after = %#v", seen)
+	}
+	paired := false
+	for _, m := range res.Messages {
+		if m.Role == llm.RoleTool && m.ToolCallID == "c1" && m.IsError && m.Content == "不允许" {
+			paired = true
+		}
+	}
+	if !paired {
+		t.Fatalf("missing paired tool result: %s", transcriptText(res.Messages))
+	}
+}
+
+func TestLoopRemindersStayOutOfSession(t *testing.T) {
+	path := filepath.Join(t.TempDir(), "transcript.jsonl")
+	st, err := session.Open(path)
+	if err != nil {
+		t.Fatal(err)
+	}
+	var saw bool
+	_, err = Run(context.Background(), Config{
+		Model:   "fake",
+		Session: st,
+		Stream: func(ctx context.Context, model string, messages []llm.Message, decls []llm.ToolDecl, opts llm.StreamOptions) <-chan llm.Event {
+			for _, m := range messages {
+				if strings.Contains(m.Content, "<system-reminder>") && strings.Contains(m.Content, "先核对交付物") {
+					saw = true
+				}
+			}
+			ch := make(chan llm.Event, 1)
+			ch <- llm.Event{Kind: llm.KindMessageEnd, Message: llm.Message{Role: llm.RoleAssistant, Content: "ok", StopReason: llm.StopEnd}}
+			close(ch)
+			return ch
+		},
+		Reminders: func(turn int) []string {
+			if turn != 0 {
+				t.Errorf("turn %d", turn)
+			}
+			return []string{"先核对交付物"}
+		},
+	}, llm.Message{Role: llm.RoleUser, Content: "ping"})
+	if err != nil {
+		t.Fatal(err)
+	}
+	if !saw {
+		t.Fatal("reminder missing from the provider request")
+	}
+	loaded, err := st.Load()
+	if err != nil {
+		t.Fatal(err)
+	}
+	if strings.Contains(transcriptText(loaded), "先核对交付物") {
+		t.Fatalf("reminder persisted: %#v", loaded)
+	}
+}
+
+func TestLoopRepeatGuardFiresAndCaps(t *testing.T) {
+	var mu sync.Mutex
+	var requests []string
+	turns := make([]llm.Message, 0, 13)
+	for i := 0; i < 12; i++ {
+		args := `{"b":1,"a":2}`
+		if i%2 == 1 {
+			args = `{"a":2,"b":1}`
+		}
+		turns = append(turns, llm.Message{Role: llm.RoleAssistant, StopReason: llm.StopToolCalls, ToolCalls: []llm.ToolCall{{
+			ID: "c", Name: "echo", Arguments: args,
+		}}})
+	}
+	turns = append(turns, llm.Message{Role: llm.RoleAssistant, Content: "done", StopReason: llm.StopEnd})
+	var i int
+	stream := func(ctx context.Context, model string, messages []llm.Message, decls []llm.ToolDecl, opts llm.StreamOptions) <-chan llm.Event {
+		mu.Lock()
+		requests = append(requests, transcriptText(messages))
+		n := i
+		i++
+		mu.Unlock()
+		ch := make(chan llm.Event, 1)
+		msg := turns[len(turns)-1]
+		if n < len(turns) {
+			msg = turns[n]
+		}
+		ch <- llm.Event{Kind: llm.KindMessageEnd, Message: msg}
+		close(ch)
+		return ch
+	}
+	res, err := Run(context.Background(), Config{
+		Model:       "fake",
+		MaxTurns:    20,
+		RepeatGuard: true,
+		Stream:      stream,
+	}, llm.Message{Role: llm.RoleUser, Content: "go"})
+	if err != nil {
+		t.Fatal(err)
+	}
+	if res.Text != "done" {
+		t.Fatalf("text %q", res.Text)
+	}
+	reminders := 0
+	for _, req := range requests {
+		reminders += strings.Count(req, repeatReminderText)
+	}
+	if reminders != 3 {
+		t.Fatalf("reminders = %d, want 3; requests=%d", reminders, len(requests))
+	}
+	if strings.Contains(transcriptText(res.Messages), repeatReminderText) {
+		t.Fatal("repeat reminder was written into the transcript")
+	}
+	if !strings.Contains(requests[3], repeatReminderText) {
+		t.Fatal("first reminder should arrive on the request after the third identical call")
+	}
+}
+
+func TestLoopContinueOnLengthCaps(t *testing.T) {
+	var requests []string
+	var n int
+	stream := func(ctx context.Context, model string, messages []llm.Message, decls []llm.ToolDecl, opts llm.StreamOptions) <-chan llm.Event {
+		requests = append(requests, transcriptText(messages))
+		n++
+		ch := make(chan llm.Event, 1)
+		ch <- llm.Event{
+			Kind:         llm.KindMessageEnd,
+			FinishReason: "length",
+			Message:      llm.Message{Role: llm.RoleAssistant, Content: "part", StopReason: llm.StopEnd},
+		}
+		close(ch)
+		return ch
+	}
+	res, err := Run(context.Background(), Config{
+		Model:            "fake",
+		MaxTurns:         8,
+		ContinueOnLength: true,
+		Stream:           stream,
+	}, llm.Message{Role: llm.RoleUser, Content: "go"})
+	if err != nil {
+		t.Fatal(err)
+	}
+	if res.Err == nil || !strings.Contains(res.Err.Error(), "length continuation exceeded") {
+		t.Fatalf("err = %v", res.Err)
+	}
+	if n != 4 {
+		t.Fatalf("model calls = %d, want 4", n)
+	}
+	if len(res.Messages) == 0 || !strings.Contains(transcriptText(res.Messages), "part") {
+		t.Fatal("truncated transcript was dropped")
+	}
+	if strings.Contains(transcriptText(res.Messages), lengthContinuePrompt) {
+		t.Fatal("continuation prompt was persisted")
+	}
+	prompts := 0
+	for _, req := range requests {
+		prompts += strings.Count(req, lengthContinuePrompt)
+	}
+	if prompts != 3 {
+		t.Fatalf("continuation prompts = %d, want 3", prompts)
+	}
+}
+
+func TestLoopHistorySkipsGeneratedSystem(t *testing.T) {
+	var first string
+	_, err := Run(context.Background(), Config{
+		Model:        "fake",
+		SystemPrompt: "LEAK",
+		History: []llm.Message{
+			{Role: llm.RoleSystem, Content: "FROM_HISTORY"},
+			{Role: llm.RoleUser, Content: "earlier"},
+		},
+		Stream: func(ctx context.Context, model string, messages []llm.Message, decls []llm.ToolDecl, opts llm.StreamOptions) <-chan llm.Event {
+			if len(messages) > 0 {
+				first = messages[0].Content
+			}
+			ch := make(chan llm.Event, 1)
+			ch <- llm.Event{Kind: llm.KindMessageEnd, Message: llm.Message{Role: llm.RoleAssistant, Content: "ok", StopReason: llm.StopEnd}}
+			close(ch)
+			return ch
+		},
+	}, llm.Message{Role: llm.RoleUser, Content: "now"})
+	if err != nil {
+		t.Fatal(err)
+	}
+	if first != "FROM_HISTORY" || strings.Contains(first, "LEAK") {
+		t.Fatalf("system = %q", first)
+	}
+}
+
+func TestLoopUnknownToolFallsThrough(t *testing.T) {
+	res, err := Run(context.Background(), Config{
+		Model: "fake",
+		UnknownTool: func(call llm.ToolCall) tools.Result {
+			return tools.Result{Content: "via " + call.Name}
+		},
+		Stream: script(
+			llm.Message{Role: llm.RoleAssistant, StopReason: llm.StopToolCalls, ToolCalls: []llm.ToolCall{{
+				ID: "c1", Name: "make_report", Arguments: `{}`,
+			}}},
+			llm.Message{Role: llm.RoleAssistant, Content: "done", StopReason: llm.StopEnd},
+		),
+	}, llm.Message{Role: llm.RoleUser, Content: "go"})
+	if err != nil {
+		t.Fatal(err)
+	}
+	found := false
+	for _, m := range res.Messages {
+		if m.Role == llm.RoleTool && m.Content == "via make_report" && !m.IsError {
+			found = true
+		}
+	}
+	if !found {
+		t.Fatalf("unknown tool was not delegated: %s", transcriptText(res.Messages))
+	}
+}
+
+func TestLoopTurnCapKeepsTranscript(t *testing.T) {
+	res, err := Run(context.Background(), Config{
+		Model:    "fake",
+		MaxTurns: 1,
+		Stream: script(llm.Message{Role: llm.RoleAssistant, StopReason: llm.StopToolCalls, ToolCalls: []llm.ToolCall{{
+			ID: "c1", Name: "missing", Arguments: `{}`,
+		}}}),
+	}, llm.Message{Role: llm.RoleUser, Content: "ping"})
+	if err != nil {
+		t.Fatal(err)
+	}
+	if res.Err == nil || !strings.Contains(res.Err.Error(), "max turns exceeded") {
+		t.Fatalf("err = %v", res.Err)
+	}
+	if !strings.Contains(transcriptText(res.Messages), "ping") {
+		t.Fatalf("transcript dropped: %s", transcriptText(res.Messages))
+	}
+}
+
+type hitTool struct{ hits *int }
+
+func (h hitTool) Name() string                { return "echo" }
+func (h hitTool) Description() string         { return "echo" }
+func (h hitTool) Parameters() json.RawMessage { return json.RawMessage(`{"type":"object"}`) }
+func (h hitTool) Execute(context.Context, json.RawMessage) (tools.Result, error) {
+	*h.hits++
+	return tools.Result{Content: "ran"}, nil
+}

runtime/session_agent_test.go

@@ -8,6 +8,30 @@ import (
 	"agent-native/runtime/llm"
 )
 
+func TestAgentSessionPromptForwardsStop(t *testing.T) {
+	called := false
+	sess := NewAgentSession(Config{
+		Model: "m",
+		Stream: func(ctx context.Context, model string, messages []llm.Message, td []llm.ToolDecl, opts llm.StreamOptions) <-chan llm.Event {
+			ch := make(chan llm.Event, 1)
+			ch <- llm.Event{Kind: llm.KindMessageEnd, Message: llm.Message{Role: llm.RoleAssistant, Content: "hi", StopReason: llm.StopEnd}}
+			close(ch)
+			return ch
+		},
+		Stop: func(last llm.Message, info StopInfo) StopDecision {
+			called = true
+			return StopDecision{}
+		},
+	})
+	res, err := sess.Prompt(context.Background(), llm.Message{Role: llm.RoleUser, Content: "go"})
+	if err != nil {
+		t.Fatal(err)
+	}
+	if !called || res.Text != "hi" {
+		t.Fatalf("called=%v text=%q", called, res.Text)
+	}
+}
+
 func TestAgentSessionSteerFollowAndAbort(t *testing.T) {
 	var mu sync.Mutex
 	step := 0