Nudge long-running turns onto the dual-speed path via the kernel reminder hook

Why: the dual-speed slow lane shipped as an observable capability plus soft
prompt guidance, but on the real (non-mock) path models routinely grind
through heavy document tasks synchronously for minutes with zero user-visible
output - the pattern only engages when the model happens to think of it.
Evidence: five new unit tests over an injected clock cover fire-after-threshold,
quiet-before-threshold, disabled-by-setting, and the replied/delegated guard
paths (reminders observed in the captured kernel requests); full go test
./... 32 pkgs green. Real-model run recorded in the Dev task card.
Impact: runDirect tracks run start, user-reply, and background-delegation
state; reminders() injects a one-shot dual-speed nudge (SR-B10) once elapsed
time crosses runtime.dualSpeedNudgeSeconds (default 75s, 0 disables) while
the user has neither a reply nor a background lane. New setting registered.
Refs: Dev task feature/dual-speed-real-path; SR-B10 (RDM requirements/01 V2.1).

internal/runtime/dual_speed_nudge_test.go

@@ -0,0 +1,142 @@
+package runtime
+
+import (
+	"context"
+	"encoding/json"
+	"fmt"
+	"strings"
+	"testing"
+	"time"
+
+	agentllm "agent-native/runtime/llm"
+	core "agent-system-core"
+)
+
+// 双速提醒经内核 Reminders 钩子进入下一请求的消息（不是响应部件）：
+// 这里捕获每次模型请求的原文，检索提醒标记。
+type dualSpeedRecorder struct {
+	clock   *stepClock
+	turns   [][]agentllm.Event
+	request []string
+	base    agentllm.StreamFunc
+}
+
+// stepClock 每 tick 前进 5 秒；runDirect 起 0 时刻，模型每被调用一次 tick 一格，
+// 20 次调用即 100 秒量级，足以压过默认 75 秒阈值。
+type stepClock struct{ step int64 }
+
+func (c *stepClock) now() time.Time { return time.Unix(0, c.step*int64(5*time.Second)) }
+func (c *stepClock) tick()          { c.step++ }
+
+func (rec *dualSpeedRecorder) stream(base agentllm.StreamFunc) agentllm.StreamFunc {
+	return func(ctx context.Context, model string, msgs []agentllm.Message, tools []agentllm.ToolDecl, opts agentllm.StreamOptions) <-chan agentllm.Event {
+		rec.clock.tick()
+		for _, m := range msgs {
+			if b, err := json.Marshal(m); err == nil {
+				rec.request = append(rec.request, string(b))
+			}
+		}
+		return base(ctx, model, msgs, tools, opts)
+	}
+}
+
+func (rec *dualSpeedRecorder) nudged() bool {
+	for _, r := range rec.request {
+		if strings.Contains(r, "任务快慢双速") {
+			return true
+		}
+	}
+	return false
+}
+
+func runDualSpeed(t *testing.T, turns [][]agentllm.Event, setup func(rc *fakeRunCtx)) *dualSpeedRecorder {
+	t.Helper()
+	rec := &dualSpeedRecorder{clock: &stepClock{}, turns: turns}
+	r := &LLMRunner{Stream: rec.stream(scriptStream(turns, nil)), Now: rec.clock.now}
+	rc := &fakeRunCtx{ctx: context.Background(), prof: verifyOn()}
+	if setup != nil {
+		setup(rc)
+	}
+	if err := r.Execute(rc); err != nil {
+		t.Fatalf("Execute: %v", err)
+	}
+	return rec
+}
+
+func toolOnlyTurns(n int) [][]agentllm.Event {
+	var turns [][]agentllm.Event
+	for i := 0; i < n; i++ {
+		// 每次调用用独立 ID（toolCall 帮手的 ID 固定，重复会撞内核的工具槽位）。
+		call := agentllm.ToolCall{ID: fmt.Sprintf("c-ds-%d", i), Name: "make_report", Arguments: "{}"}
+		turns = append(turns, textDeltaTurn("", call))
+	}
+	return turns
+}
+
+// 长时间纯工具轮、用户零答复 → 阈值后恰好注入一次双速提醒。
+func TestDualSpeedNudgeFiresOnceAfterThreshold(t *testing.T) {
+	turns := append(toolOnlyTurns(20), textDeltaTurn("完成。"))
+	rec := runDualSpeed(t, turns, nil)
+	if !rec.nudged() {
+		t.Fatal("dual-speed nudge missing after threshold with no user reply")
+	}
+	count := 0
+	for _, r := range rec.request {
+		count += strings.Count(r, "任务快慢双速")
+	}
+	if count > 1 {
+		t.Fatalf("nudge injected %d times, want at most once per request window", count)
+	}
+}
+
+// 短任务（阈值内完成）不提醒。
+func TestDualSpeedNudgeQuietBeforeThreshold(t *testing.T) {
+	turns := append(toolOnlyTurns(1), textDeltaTurn("完成。"))
+	rec := runDualSpeed(t, turns, nil)
+	if rec.nudged() {
+		t.Fatal("nudge fired before the threshold")
+	}
+}
+
+// 设置 runtime.dualSpeedNudgeSeconds=0 → 关闭，超时也不提醒。
+func TestDualSpeedNudgeDisabledBySetting(t *testing.T) {
+	turns := append(toolOnlyTurns(20), textDeltaTurn("完成。"))
+	rec := runDualSpeed(t, turns, func(rc *fakeRunCtx) {
+		rc.runtimeExtra = map[string]any{"dualSpeedNudgeSeconds": float64(0)}
+	})
+	if rec.nudged() {
+		t.Fatal("nudge fired although disabled via runtime.dualSpeedNudgeSeconds=0")
+	}
+}
+
+// 模型已给出面向用户的答复（快通道成立）→ 即使后续事实核查续写拖过阈值也不提醒。
+func TestDualSpeedNudgeSkippedWhenReplied(t *testing.T) {
+	turns := [][]agentllm.Event{
+		textDeltaTurn("", toolCall("make_report", `{"content":"准确率 95%"}`)),
+		textDeltaTurn("完整总结：产出物与要点。"),
+		textDeltaTurn("已修正并重新登记。"),
+	}
+	rec := runDualSpeed(t, turns, func(rc *fakeRunCtx) {
+		rc.verify = true
+		rc.execArtifact = &core.Artifact{ID: "art_0123456789ab", Name: "报告.md"}
+		rc.execArtifactContent = []byte("# 报告\n\n准确率 95%。")
+	})
+	if rec.nudged() {
+		t.Fatal("nudge fired although the model already gave its user reply")
+	}
+}
+
+// 已把深度工作放后台（慢通道成立）→ 不再提醒。
+func TestDualSpeedNudgeSkippedWhenDelegated(t *testing.T) {
+	turns := [][]agentllm.Event{
+		textDeltaTurn("", toolCall("delegate_task", `{"name":"深度解析","task":"逐页解析","background":true}`)),
+	}
+	turns = append(turns, toolOnlyTurns(19)...)
+	turns = append(turns, textDeltaTurn("后台进行中，先给结论。"))
+	rec := runDualSpeed(t, turns, func(rc *fakeRunCtx) {
+		rc.bgID = "run_bg_1"
+	})
+	if rec.nudged() {
+		t.Fatal("nudge fired although background delegation already happened")
+	}
+}

internal/runtime/llm_direct.go

@@ -65,6 +65,12 @@ type directRun struct {
 	servedModel      string
 	switches         []string
 	switchMu         sync.Mutex
+	// 任务快慢双速（SR-B10 实路径）：started 用注入时钟记 run 起点，
+	// nudged/replied/delegated 三个标志驱动「超时仍未反馈则提醒转双速」的一次性触发。
+	started            time.Time
+	nudgedDualSpeed    bool
+	replied            bool
+	delegatedBackground bool
 }
 
 type toolEvidence struct {
@@ -97,7 +103,7 @@ func (t directTool) Execute(ctx context.Context, args json.RawMessage) (agenttoo
 }
 
 func (l *LLMRunner) runDirect(rc RunCtx, stream agentllm.StreamFunc, model string, msgs []agentllm.Message, decls []agentllm.ToolDecl, streamOpts agentllm.StreamOptions, cardflow bool, eff core.EffectiveSettings, maxTurns, maxTools int, refresh turnRefresh) error {
-	st := &directRun{l: l, rc: rc, eff: eff, cardflow: cardflow}
+	st := &directRun{l: l, rc: rc, eff: eff, cardflow: cardflow, started: l.now()}
 	// 记忆端口（N3 WP5）：runCtx 暴露 store 时接线（注入由循环按请求完成）。
 	if ms, ok := rc.(interface{ StoreMemory() memoryStore }); ok {
 		st.memoryStore = ms.StoreMemory()
@@ -473,6 +479,10 @@ func (st *directRun) onEvent(ev nruntime.Event) {
 			if body != "" {
 				st.finalParts = append(st.finalParts, core.Part{Type: core.PartText, Text: body, Process: hasCalls})
 				st.turnBody = body
+				if !hasCalls {
+					// 面向用户的答复段落已出现：快通道成立，双速提醒不再需要。
+					st.replied = true
+				}
 				st.textIdx++
 			}
 		}
@@ -693,6 +703,7 @@ func (st *directRun) runTool(name, args string) agenttools.Result {
 				if serr != nil {
 					out = "子代理失败：" + serr.Error()
 				} else {
+					st.delegatedBackground = true
 					out = fmt.Sprintf("已在后台启动 %s（慢通道）：收件箱已建进度卡，完成后结果回投该卡并自动汇报；先把快通道的即时反馈写给用户，不要等待", id)
 				}
 				break

internal/runtime/llm_test.go

@@ -42,9 +42,10 @@ type fakeRunCtx struct {
 	bgID                string
 	bgErr               error
 	sandbox             string     // 非空时 ConversationSandbox 返回之（N3 WP2 转录测试）
-	memEntries          []memEntry // 非空时 StoreMemory 返回之（N3 WP5 记忆注入测试）
+	memEntries          []memEntry // 非空时 StoreMemory 返回之（N3 WP5 记忆注入断言用）
 	usageTotal          core.Usage // AddUsage 累计（N3 WP6 记账断言用）
 	bgCalls             int
+	runtimeExtra        map[string]any // 覆盖 runtime 组设置（双速提醒阈值等测试用）
 }
 
 func (f *fakeRunCtx) Context() context.Context   { return f.ctx }
@@ -55,7 +56,11 @@ func (f *fakeRunCtx) Settings() core.EffectiveSettings {
 	if f.cardflow {
 		eff["ui"] = map[string]any{"cardflow": true}
 	}
-	eff["runtime"] = map[string]any{"verify": f.verify, "composeFinal": f.compose}
+	runtimeMap := map[string]any{"verify": f.verify, "composeFinal": f.compose}
+	for k, v := range f.runtimeExtra {
+		runtimeMap[k] = v
+	}
+	eff["runtime"] = runtimeMap
 	return eff
 }
 func (f *fakeRunCtx) Emit(typ core.StreamEventType, payload any) error {

internal/runtime/verify.go

@@ -149,6 +149,13 @@ func (st *directRun) reminders(turn int) []string {
 		st.closedToolTurn = false
 		st.checklistTouched = false
 	}
+	// 任务快慢双速（SR-B10 实路径触发）：run 已运行超过阈值、用户还没看到任何
+	// 答复、也没把深度工作放后台时，一次性提醒转双速——先把判断/计划写给用户，
+	// 重活 delegate 后台。已经回复或已委派则永不打扰；接近完成的任务可按文案
+	// 自行忽略。这是内核每回合的 Reminders 注入点（agent-native 循环原生钩子）。
+	if note := st.dualSpeedNudge(); note != "" {
+		return []string{note}
+	}
 	if st.toolTurnsSince < gates.ReminderToolGap {
 		return nil
 	}
@@ -175,6 +182,24 @@ func (st *directRun) reminders(turn int) []string {
 	return got
 }
 
+// dualSpeedNudgeSeconds：双速提醒的运行时长阈值（秒）。0 = 关闭。
+const dualSpeedNudgeDefaultSeconds = 75
+
+// dualSpeedNudge 返回双速提醒文案；不满足条件或已提醒过返回空串。
+// 条件用注入时钟（l.now）计量，测试可控。
+func (st *directRun) dualSpeedNudge() string {
+	after := int(st.eff.GetNumber("runtime.dualSpeedNudgeSeconds", dualSpeedNudgeDefaultSeconds))
+	if after <= 0 || st.nudgedDualSpeed || st.replied || st.delegatedBackground {
+		return ""
+	}
+	if st.started.IsZero() || st.l.now().Sub(st.started) < time.Duration(after)*time.Second {
+		return ""
+	}
+	st.nudgedDualSpeed = true
+	elapsed := int(st.l.now().Sub(st.started).Round(time.Second).Seconds())
+	return fmt.Sprintf("本次任务已连续运行约 %d 秒，用户还没有收到任何答复（任务快慢双速，SR-B10）。现在二选一：若剩余工作量大（预计还要一分钟以上），先把你的判断、概要或执行计划作为即时反馈回复给用户，同时用 delegate_task 的 background=true 把深度处理放到后台执行（收件箱有进度卡，完成后系统自动回投结果并汇报）；若已接近完成，把这句话忽略，继续做完后在最终答复里给出完整结论。不要默默长时间工作。", elapsed)
+}
+
 // unfinishedItems 返回计划与待办的未完成项正文（状态过滤留在产品侧）。
 func (st *directRun) unfinishedItems() (todos, plans []string) {
 	for _, item := range st.todos.Items {

internal/settings/settings.go

@@ -97,6 +97,7 @@ func Schema() []core.SettingDesc {
 		{Key: "runtime.llmRetries", Group: "runtime", Label: "模型重试次数", Type: core.SettingNumber, Default: 2, Effect: core.EffectNewRun, Description: "直连模型在还没有任何输出时，遇到超时、限流或服务端错误的额外重试次数。0 = 不重试。已经开始输出后不再重试"},
 		{Key: "runtime.codeOrchestration", Group: "runtime", Label: "代码编排", Type: core.SettingBool, Default: false, Effect: core.EffectNewRun, Description: "允许模型用 run_tool_script 写脚本批量调用工具。默认关闭"},
 		{Key: "runtime.factCheck", Group: "runtime", Label: "终轮事实核查", Type: core.SettingBool, Default: true, Effect: core.EffectNewRun, Description: "交付物里的百分比/日期/大额数字回溯本轮材料，找不到出处的给模型一轮修正（修正过仍失败不拦截，只如实记录）"},
+		{Key: "runtime.dualSpeedNudgeSeconds", Group: "runtime", Label: "快慢双速提醒阈值（秒）", Type: core.SettingNumber, Default: 75, Effect: core.EffectNewRun, Description: "任务连续运行超过该秒数、用户还没收到任何答复、也没把深度工作放后台时，一次性提醒模型转快慢双速：先把判断/概要/计划回复给用户，重活用 delegate_task 的 background 放后台（SR-B10）。0 = 关闭"},
 		{Key: "runtime.verifierSameModel", Group: "runtime", Label: "允许同源验证器", Type: core.SettingBool, Default: false, Effect: core.EffectNewRun, Description: "轻量档未单独配置时，默认跳过与主模型同源的独立验证（硬事实与事实核查仍执行）；显式开启才用主模型做验证"},
 		{Key: "runtime.composeFinal", Group: "runtime", Label: "终轮整理回复", Type: core.SettingBool, Default: false, Effect: core.EffectNewRun, Description: "用过工具的回合结束后，再用主模型重写一段最终回复。默认关闭（裁定 4：转录权威后工作模型已看到完整证据，重写只见过 160 字摘要、实测会虚构章节与承诺未做之事）；异常收尾时另有兜底"},
 		{Key: "runtime.intentSemantic", Group: "runtime", Label: "意图语义裁决（慢通道）", Type: core.SettingBool, Default: false, Effect: core.EffectRestart, Description: "规则歧义时用 embedding 语义相似度裁决。需已配置 embedding 端点。默认关闭（行为与 I1 一致）"},