Merge pull request 'Split usage accounting end to end, retry by default, and take the WP6 pins (N3 WP6)' (#96) from feature/n3-usage-retry-wp6 into main

Split usage accounting end to end, retry by default, and take the WP6 pins (N3 WP6)

internal/runmgr/runmgr.go

@@ -101,6 +101,10 @@ type runHandle struct {
 	started          time.Time
 	promptTokens     int
 	completionTokens int
+	lastPromptTokens     int
+	lastCompletionTokens int
+	cacheReadTokens      int
+	cacheWriteTokens     int
 }
 
 type hitlAnswer struct {
@@ -236,10 +240,15 @@ func (c *runCtx) ConversationSandbox() string {
 }
 
 // AddUsage 累加本 run 的模型用量，finishRun 写入 run.Usage。
+// 计费拆分（N3 WP6）：Last* 覆盖（最近一次请求）、Cache* 累计。
 func (c *runCtx) AddUsage(u core.Usage) {
 	c.mgr.mu.Lock()
 	c.h.promptTokens += u.PromptTokens
 	c.h.completionTokens += u.CompletionTokens
+	c.h.lastPromptTokens = u.LastPromptTokens
+	c.h.lastCompletionTokens = u.LastCompletionTokens
+	c.h.cacheReadTokens += u.CacheReadTokens
+	c.h.cacheWriteTokens += u.CacheWriteTokens
 	c.mgr.mu.Unlock()
 }
 func (c *runCtx) Subtask() string { return c.subTask }
@@ -1080,6 +1089,10 @@ func (m *Manager) finishRun(rc *runCtx, status core.RunStatus, usage core.Usage,
 	usage.PromptTokens = rc.h.promptTokens
 	usage.CompletionTokens = rc.h.completionTokens
 	usage.TotalTokens = usage.PromptTokens + usage.CompletionTokens
+	usage.LastPromptTokens = rc.h.lastPromptTokens
+	usage.LastCompletionTokens = rc.h.lastCompletionTokens
+	usage.CacheReadTokens = rc.h.cacheReadTokens
+	usage.CacheWriteTokens = rc.h.cacheWriteTokens
 	m.mu.Unlock()
 	if m.shuttingDown.Load() && status == core.RunCancelled {
 		status = core.RunInterrupted
@@ -1395,7 +1408,17 @@ func (m *Manager) finishChild(rc *runCtx, runner runtime.Runner, emitResult bool
 		return runner.Execute(rc)
 	}()
 
+	// 子代理记账（N3 WP6）：token 用量随子 run 入账（此前只有 DurationMs）。
 	usage := core.Usage{DurationMs: float64(time.Since(rc.h.started).Milliseconds())}
+	m.mu.Lock()
+	usage.PromptTokens = rc.h.promptTokens
+	usage.CompletionTokens = rc.h.completionTokens
+	usage.TotalTokens = usage.PromptTokens + usage.CompletionTokens
+	usage.LastPromptTokens = rc.h.lastPromptTokens
+	usage.LastCompletionTokens = rc.h.lastCompletionTokens
+	usage.CacheReadTokens = rc.h.cacheReadTokens
+	usage.CacheWriteTokens = rc.h.cacheWriteTokens
+	m.mu.Unlock()
 	status := core.RunCompleted
 	summary := "完成"
 	payload := core.AgentResultPayload{AgentRun: rc.run.ID, Status: status, Summary: summary}

internal/runmgr/subagent_n2_test.go

@@ -207,3 +207,18 @@ func TestResumeSkipsCompletedMarker(t *testing.T) {
 		t.Fatalf("completed 终态不得续跑：before=%v after=%v", before, ids)
 	}
 }
+
+// N3 WP6：子代理记账链——AddUsage 累计拆分字段，finishChild 合并进 child.Usage。
+func TestAddUsageAccumulatesSplitFields(t *testing.T) {
+	c, _ := newSpawnEnv(t, &gateRunner{})
+	c.h = &runHandle{}
+	c.AddUsage(core.Usage{PromptTokens: 100, CompletionTokens: 10, LastPromptTokens: 100, LastCompletionTokens: 10, CacheReadTokens: 50, CacheWriteTokens: 4})
+	c.AddUsage(core.Usage{PromptTokens: 60, CompletionTokens: 6, LastPromptTokens: 60, LastCompletionTokens: 6, CacheReadTokens: 20, CacheWriteTokens: 2})
+	c.mgr.mu.Lock()
+	defer c.mgr.mu.Unlock()
+	if c.h.promptTokens != 160 || c.h.completionTokens != 16 ||
+		c.h.lastPromptTokens != 60 || c.h.lastCompletionTokens != 6 ||
+		c.h.cacheReadTokens != 70 || c.h.cacheWriteTokens != 6 {
+		t.Fatalf("记账链：%+v", c.h)
+	}
+}

internal/runtime/llm.go

@@ -205,9 +205,10 @@ func (l *LLMRunner) Execute(rc RunCtx) error {
 
 	stream := l.Stream
 	if stream == nil {
+		// 重试经库默认（WP6：runtime.llmRetries → Config.RetryAttempts 映射），
+		// 此处不再显式包装（防双重）。
 		client := openLLMClient(route.Base, route.Key)
-		retries := int(eff.GetNumber("runtime.llmRetries", 2))
-		stream = agentllm.WithRetry(client.Stream, retries, nil)
+		stream = client.Stream
 	}
 
 	// 知识注入：当前消息包含 @knowledge 引用时检索注入（最小实现：显式引用）。
@@ -404,6 +405,16 @@ func outputTokenCap(budget core.RunBudget, eff core.EffectiveSettings, key strin
 // 组合根经 runCtx 注入，runmgr 的 storeMemory 实现之）。
 type memoryStore = agentmemory.Store
 
+// retryAttemptsOf 把 runtime.llmRetries 的设置语义（0=不重试，N=N 次）
+// 映射到库 Config.RetryAttempts（0=默认 2，负值=关，正数=次数）。
+func retryAttemptsOf(eff core.EffectiveSettings) int {
+	n := int(eff.GetNumber("runtime.llmRetries", 2))
+	if n <= 0 {
+		return -1
+	}
+	return n
+}
+
 // agentNameOf 是 OTel gen_ai.agent.name 的取值（画像 ID，空回退 agent-system）。
 func agentNameOf(prof core.AgentProfile) string {
 	if strings.TrimSpace(prof.ID) != "" {

internal/runtime/llm_direct.go

@@ -211,6 +211,7 @@ func (l *LLMRunner) runDirect(rc RunCtx, stream agentllm.StreamFunc, model strin
 		Trace:                  l.Trace,
 		AgentName:              agentNameOf(rc.Profile()),
 		Memory:                 st.memoryStore,
+		RetryAttempts:          retryAttemptsOf(eff),
 		UnknownTool: func(call agentllm.ToolCall) agenttools.Result {
 			return st.runTool(call.Name, call.Arguments)
 		},
@@ -275,6 +276,9 @@ func (st *directRun) noteToolDelta(ev agentllm.Event) {
 
 func (st *directRun) onEvent(ev nruntime.Event) {
 	switch ev.Type {
+	case "auto_retry_start":
+		// 重试可观测（N3 WP6）：沿用 process part 形态。
+		st.finalParts = append(st.finalParts, core.Part{Type: core.PartText, Text: "模型在输出前失败，正在重试（" + ev.Text + "）", Process: true})
 	case "turn_start":
 		st.responding = false
 		st.acting = false
@@ -314,6 +318,11 @@ func (st *directRun) onEvent(ev nruntime.Event) {
 		st.usageTotal.PromptTokens += ev.PromptTokens
 		st.usageTotal.CompletionTokens += ev.CompletionTokens
 		st.usageTotal.TotalTokens += ev.TotalTokens
+		// 计费拆分（N3 WP6）：上一请求覆盖、缓存命中累计。
+		st.usageTotal.LastPromptTokens = ev.PromptTokens
+		st.usageTotal.LastCompletionTokens = ev.CompletionTokens
+		st.usageTotal.CacheReadTokens += ev.CacheReadTokens
+		st.usageTotal.CacheWriteTokens += ev.CacheWriteTokens
 		st.queued = append([]agentllm.ToolCall(nil), ev.Message.ToolCalls...)
 		hasCalls := len(ev.Message.ToolCalls) > 0
 		if st.turnThink.Len() > 0 {

internal/runtime/llm_test.go

@@ -43,6 +43,7 @@ type fakeRunCtx struct {
 	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）
+	usageTotal          core.Usage  // AddUsage 累计（N3 WP6 记账断言用）
 	bgCalls             int
 }
 
@@ -62,6 +63,15 @@ func (f *fakeRunCtx) Emit(typ core.StreamEventType, payload any) error {
 	return nil
 }
 func (f *fakeRunCtx) SetAssistantParts(parts []core.Part) error { f.assistant = parts; return nil }
+func (f *fakeRunCtx) AddUsage(u core.Usage) {
+	f.usageTotal.PromptTokens += u.PromptTokens
+	f.usageTotal.CompletionTokens += u.CompletionTokens
+	f.usageTotal.TotalTokens += u.TotalTokens
+	f.usageTotal.LastPromptTokens = u.LastPromptTokens
+	f.usageTotal.LastCompletionTokens = u.LastCompletionTokens
+	f.usageTotal.CacheReadTokens += u.CacheReadTokens
+	f.usageTotal.CacheWriteTokens += u.CacheWriteTokens
+}
 func (f *fakeRunCtx) AskHITL(card core.HITLCard) (core.HITLResponse, error) {
 	return core.HITLResponse{}, fmt.Errorf("not implemented")
 }

internal/runtime/squad_runner.go

@@ -53,7 +53,11 @@ func (s *SquadRunner) Execute(rc RunCtx) error {
 			toolParts = append(toolParts, core.Part{Type: core.PartToolResult, Process: true, ToolResult: &core.ToolResultPart{CallID: ev.Tool, Status: "ok", Summary: truncateRun(ev.Detail, 160)}})
 		case "usage":
 			if sink, ok := rc.(interface{ AddUsage(core.Usage) }); ok {
-				sink.AddUsage(core.Usage{PromptTokens: ev.PromptTokens, CompletionTokens: ev.CompletionTokens, TotalTokens: ev.PromptTokens + ev.CompletionTokens})
+				sink.AddUsage(core.Usage{
+					PromptTokens: ev.PromptTokens, CompletionTokens: ev.CompletionTokens, TotalTokens: ev.PromptTokens + ev.CompletionTokens,
+					LastPromptTokens: ev.LastPromptTokens, LastCompletionTokens: ev.LastCompletionTokens,
+					CacheReadTokens: ev.CacheReadTokens, CacheWriteTokens: ev.CacheWriteTokens,
+				})
 			}
 		}
 	}

internal/runtime/sub_loop.go

@@ -111,7 +111,12 @@ func (l *SubLLMRunner) Execute(rc RunCtx) error {
 		PermissionAsk:          PermissionAskOf(rc),
 		Trace:                  l.Trace,
 		AgentName:              agentNameOf(rc.Profile()),
+		RetryAttempts:          retryAttemptsOf(rc.Settings()),
 	}, user)
+	// 子代理记账（N3 WP6）：token 用量入账（此前只有 DurationMs）。
+	if sink, ok := rc.(interface{ AddUsage(core.Usage) }); ok {
+		sink.AddUsage(st.usageTotal)
+	}
 	if len(st.parts) == 0 {
 		st.parts = append(st.parts, core.Part{Type: core.PartText, Text: "（子代理未返回内容）"})
 	}
@@ -169,11 +174,12 @@ type subSlot struct {
 }
 
 type subRun struct {
-	rc       RunCtx
-	parts    []core.Part
-	queued   []agentllm.ToolCall
-	inflight *subSlot
-	turnText strings.Builder
+	rc         RunCtx
+	parts      []core.Part
+	queued     []agentllm.ToolCall
+	inflight   *subSlot
+	turnText   strings.Builder
+	usageTotal core.Usage
 }
 
 func (st *subRun) onEvent(ev nruntime.Event) {
@@ -184,6 +190,13 @@ func (st *subRun) onEvent(ev nruntime.Event) {
 		st.turnText.WriteString(ev.Text)
 		_ = st.rc.Emit(core.EvAgentProgress, core.AgentProgressPayload{AgentRun: st.rc.Run().ID, Note: truncateRun(st.turnText.String(), 80)})
 	case "message_end":
+		st.usageTotal.PromptTokens += ev.PromptTokens
+		st.usageTotal.CompletionTokens += ev.CompletionTokens
+		st.usageTotal.TotalTokens += ev.TotalTokens
+		st.usageTotal.LastPromptTokens = ev.PromptTokens
+		st.usageTotal.LastCompletionTokens = ev.CompletionTokens
+		st.usageTotal.CacheReadTokens += ev.CacheReadTokens
+		st.usageTotal.CacheWriteTokens += ev.CacheWriteTokens
 		st.queued = append([]agentllm.ToolCall(nil), ev.Message.ToolCalls...)
 		hasCalls := len(ev.Message.ToolCalls) > 0
 		if body := strings.TrimSpace(st.turnText.String()); body != "" {

internal/runtime/usage_n3_test.go

@@ -0,0 +1,71 @@
+package runtime
+
+import (
+	"context"
+	"testing"
+
+	agentllm "agent-native/runtime/llm"
+	core "agent-system-core"
+)
+
+// ---- N3 WP6：计费拆分与重试收口（backend 侧）----
+
+// retryAttemptsOf 映射：设置 0=不重试→-1；N→N；默认 2→2。
+func TestRetryAttemptsMapping(t *testing.T) {
+	if got := retryAttemptsOf(core.EffectiveSettings{}); got != 2 {
+		t.Fatalf("默认应 2，得到 %d", got)
+	}
+	if got := retryAttemptsOf(core.EffectiveSettings{"runtime": map[string]any{"llmRetries": float64(0)}}); got != -1 {
+		t.Fatalf("设置 0 应映射为关闭（-1），得到 %d", got)
+	}
+	if got := retryAttemptsOf(core.EffectiveSettings{"runtime": map[string]any{"llmRetries": float64(5)}}); got != 5 {
+		t.Fatalf("设置 5 应映射为 5，得到 %d", got)
+	}
+}
+
+// 直连 run 的用量拆分：message_end 到达后 Last* 覆盖、cache 累计入账。
+func TestDirectRunUsageSplitOnRecord(t *testing.T) {
+	rc := &fakeRunCtx{ctx: context.Background(), prof: core.AgentProfile{Model: core.ModelRoute{Chat: "test-model"}}}
+	stream := func(ctx context.Context, model string, msgs []agentllm.Message, _ []agentllm.ToolDecl, _ agentllm.StreamOptions) <-chan agentllm.Event {
+		ch := make(chan agentllm.Event, 1)
+		ch <- agentllm.Event{Kind: agentllm.KindMessageEnd,
+			Message: agentllm.Message{Role: agentllm.RoleAssistant, Content: "x", StopReason: agentllm.StopEnd},
+			Usage:   agentllm.Usage{PromptTokens: 120, CompletionTokens: 12, TotalTokens: 132, CacheReadTokens: 80, CacheWriteTokens: 6},
+		}
+		close(ch)
+		return ch
+	}
+	r := &LLMRunner{Stream: stream}
+	if err := r.Execute(rc); err != nil {
+		t.Fatalf("Execute: %v", err)
+	}
+	u := rc.usageTotal
+	if u.PromptTokens != 120 || u.CompletionTokens != 12 ||
+		u.LastPromptTokens != 120 || u.LastCompletionTokens != 12 ||
+		u.CacheReadTokens != 80 || u.CacheWriteTokens != 6 {
+		t.Fatalf("用量拆分未入账：%+v", u)
+	}
+}
+
+// 子代理记账：subRun 的 message_end 累计 token 并经 AddUsage 入账。
+func TestSubRunAccountsTokens(t *testing.T) {
+	rc := &fakeRunCtx{ctx: context.Background(), prof: core.AgentProfile{Model: core.ModelRoute{Chat: "test-model"}}}
+	stream := func(ctx context.Context, model string, msgs []agentllm.Message, _ []agentllm.ToolDecl, _ agentllm.StreamOptions) <-chan agentllm.Event {
+		ch := make(chan agentllm.Event, 1)
+		ch <- agentllm.Event{Kind: agentllm.KindMessageEnd,
+			Message: agentllm.Message{Role: agentllm.RoleAssistant, Content: "子结果", StopReason: agentllm.StopEnd},
+			Usage:   agentllm.Usage{PromptTokens: 66, CompletionTokens: 6, TotalTokens: 72, CacheReadTokens: 30},
+		}
+		close(ch)
+		return ch
+	}
+	r := &SubLLMRunner{Stream: stream}
+	if err := r.Execute(rc); err != nil {
+		t.Fatalf("Execute: %v", err)
+	}
+	u := rc.usageTotal
+	if u.PromptTokens != 66 || u.CompletionTokens != 6 ||
+		u.LastPromptTokens != 66 || u.LastCompletionTokens != 6 || u.CacheReadTokens != 30 {
+		t.Fatalf("子代理用量未入账：%+v", u)
+	}
+}

internal/runtime/verify.go

@@ -33,7 +33,10 @@ func (l *LLMRunner) runVerifier(rc RunCtx, body string, evidence []toolEvidence,
 			return gates.Verdict{}, fmt.Errorf("验证器调用失败")
 		}
 		client := openLLMClient(route.Base, route.Key)
-		stream = client.Stream
+		// 验证器不经内核循环：显式接重试（WP6 同映射）与 5xx 回退链。
+		stream = agentllm.WithFallback(
+			agentllm.WithRetry(client.Stream, retryAttemptsOf(rc.Settings()), nil),
+			ChatFallbacks(rc.Settings()), nil)
 	}
 	model := route.Model
 	if model == "" {