Point the nested intent archive at the eval-v2 head

Why: Same cascade as the squad side (gitlink archive, no go.mod reference). Evidence: intent main head is the #8 merge. Impact: Nested pin only.
Merge pull request 'Point the nested agent-intent pin at the I1 command model' (#21 ) from feature/agent-stack-i1 into main
2026-10-07 08:23:05 +00:00 · 2026-10-07 07:47:57 +00:00 · 2026-10-07 07:46:45 +00:00 · 2026-10-07 06:25:43 +00:00 · 2026-10-07 06:25:13 +00:00 · 2026-10-07 06:21:56 +00:00
132 changed files with 11315 additions and 749 deletions
@@ -6,15 +6,48 @@ on:
    branches: [main]
 jobs:
  vet-and-test:
-    runs-on: nightly
+    runs-on: ubuntu-latest
+    env:
+      GOPROXY: https://goproxy.cn,direct
    steps:
-      - uses: actions/checkout@v4
-      - run: go vet ./... && go test -count=1 ./...
+      - name: Checkout
+        env:
+          TOKEN: ${{ secrets.GITHUB_TOKEN }}
+          SUBMODULE_TOKEN: ${{ secrets.SUBMODULE_TOKEN }}
+        run: |
+          set -euo pipefail
+          git init -q
+          git remote add origin "${GITHUB_SERVER_URL}/${GITHUB_REPOSITORY}.git"
+          git -c http.extraheader="AUTHORIZATION: token ${TOKEN}" fetch --depth=1 origin "${GITHUB_SHA}"
+          git checkout -q --detach FETCH_HEAD
+          if [ -z "${SUBMODULE_TOKEN}" ]; then
+            echo "SUBMODULE_TOKEN is not configured" >&2
+            exit 1
+          fi
+          git config --global url."https://weichen:${SUBMODULE_TOKEN}@git-home.u1680832.nyat.app:30319/".insteadOf "ssh://git@git-ssh.u1680832.nyat.app:31411/"
+          git -c fetch.recurseSubmodules=false submodule update --init --recursive
+      - name: Install Go 1.27.1
+        run: |
+          set -euo pipefail
+          curl -fsSL -o "$RUNNER_TEMP/go.tgz" https://mirrors.aliyun.com/golang/go1.27.1.linux-amd64.tar.gz
+          tar -C /usr/local -xzf "$RUNNER_TEMP/go.tgz"
+          ln -sfn /usr/local/go/bin/go /usr/local/bin/go
+          ln -sfn /usr/local/go/bin/gofmt /usr/local/bin/gofmt
+      - name: Vet, test, and race
+        run: go vet ./... && go test -count=1 ./... && go test -count=1 -race ./...
  docker-checkout-hostgateway:
    runs-on: ubuntu-latest
    options: --add-host gitea:host-gateway
    env:
      GITHUB_SERVER_URL: http://gitea:3000
    steps:
-      - uses: actions/checkout@v4
-      - run: test -f go.mod && echo "docker-mode checkout via host-gateway OK"
+      - name: Checkout without GitHub actions
+        env:
+          TOKEN: ${{ secrets.GITHUB_TOKEN }}
+        run: |
+          set -euo pipefail
+          git init -q
+          git remote add origin "${GITHUB_SERVER_URL}/${GITHUB_REPOSITORY}.git"
+          git -c http.extraheader="AUTHORIZATION: token ${TOKEN}" fetch --depth=1 origin "${GITHUB_SHA}"
+          git checkout -q --detach FETCH_HEAD
+          test -f go.mod && echo "docker-mode checkout via host-gateway OK"
@@ -1,56 +0,0 @@
-// Package intentgate 提供可选的意图闸门装饰器：在把 Prompt 交给被装饰的
-// 执行后端之前，用 agent-intent 做一次确定性识别；分支为守卫/命中/预期命中时
-// 交由宿主 Handler 解释，Ambiguous/Fallback 或 Handler 放弃时原样透传。
-//
-// 本包是 Agent-Native 仓内唯一 import agent-intent 的地方。不使用本包时，
-// 本仓与 agent-intent 无任何编译关系；未 Wrap 或快照零意图时，被装饰后端的
-// 行为与原来完全一致。
-package intentgate
-
-import (
-	"context"
-
-	"agent-intent/intent"
-	"agent-native/exec"
-)
-
-// Handler 解释"识别到意图"的执行请求。返回 handled=false 表示宿主放弃处理，
-// 装饰器回退为透传原后端。合成结果的语义（拒绝、引导、定向执行）由宿主定义，
-// 本包不发明结果语义。
-type Handler func(ctx context.Context, req exec.Request, res intent.Result) (result *exec.Result, err error, handled bool)
-
-type gated struct {
-	next    exec.Backend
-	rec     intent.Recognizer
-	snap    *intent.Snapshot
-	handler Handler
-}
-
-// Wrap 返回带意图闸门的后端。next/rec/snap/handler 任一为 nil 时直接返回 next
-// （等价于未启用，行为不变）。
-func Wrap(next exec.Backend, rec intent.Recognizer, snap *intent.Snapshot, h Handler) exec.Backend {
-	if next == nil || rec == nil || snap == nil || h == nil {
-		return next
-	}
-	return &gated{next: next, rec: rec, snap: snap, handler: h}
-}
-
-func (g *gated) Name() string { return g.next.Name() }
-
-func (g *gated) Execute(ctx context.Context, req exec.Request) (*exec.Result, error) {
-	res, err := g.rec.Recognize(ctx, req.Prompt, g.snap)
-	if err != nil {
-		// 识别系统错误 fail-closed：不静默放行可能应拦截的内容，也不吞错。
-		return nil, err
-	}
-	switch res.Branch {
-	case intent.BranchGuardHit, intent.BranchMatched, intent.BranchExpectationHit:
-	default:
-		return g.next.Execute(ctx, req) // 无裁决：透传原后端
-	}
-	out, herr, handled := g.handler(ctx, req, res)
-	if !handled {
-		return g.next.Execute(ctx, req)
-	}
-	return out, herr
-}
@@ -1,166 +0,0 @@
-package intentgate
-
-import (
-	"context"
-	"errors"
-	"reflect"
-	"testing"
-
-	"agent-intent/intent"
-	"agent-native/exec"
-)
-
-type stubBackend struct {
-	name  string
-	calls int
-}
-
-func (s *stubBackend) Name() string { return s.name }
-
-func (s *stubBackend) Execute(ctx context.Context, req exec.Request) (*exec.Result, error) {
-	s.calls++
-	return &exec.Result{Success: true, Logs: "stub:" + req.Prompt}, nil
-}
-
-func emptySnapshot(t *testing.T) *intent.Snapshot {
-	t.Helper()
-	reg, err := intent.NewRegistryBuilder(intent.Options{}).Build()
-	if err != nil {
-		t.Fatal(err)
-	}
-	return reg.Snapshot()
-}
-
-// 带一个守卫 spec 与一个可路由 spec 的快照。
-func loadedSnapshot(t *testing.T) *intent.Snapshot {
-	t.Helper()
-	b := intent.NewRegistryBuilder(intent.Options{
-		GuardDomains:     []string{"guard"},
-		ProtectedDomains: []string{"guard"},
-	})
-	if err := b.Register(intent.IntentSpec{
-		ID: "guard.config_change", Domain: "guard", Version: 1,
-		Matcher: intent.MatchSpec{IncludesAny: []string{"系统配置", "网关"}},
-		Risk:    intent.RiskHigh,
-		Handler: "refuse.and_guide",
-		Metadata: map[string]string{
-			"guidance": "系统配置请在管理界面修改",
-		},
-	}); err != nil {
-		t.Fatal(err)
-	}
-	reg, err := b.Build()
-	if err != nil {
-		t.Fatal(err)
-	}
-	return reg.Snapshot()
-}
-
-func newRec() intent.Recognizer { return intent.NewRecognizer(intent.Options{}, nil, nil) }
-
-func TestZeroSpecSnapshotPassthroughIdentical(t *testing.T) {
-	// T0 行为解耦：零意图快照 + Wrap 后，与裸后端行为逐字节一致。
-	bare := &stubBackend{name: "native"}
-	gatedBe := Wrap(&stubBackend{name: "native"}, newRec(), emptySnapshot(t),
-		func(ctx context.Context, req exec.Request, res intent.Result) (*exec.Result, error, bool) {
-			t.Fatal("handler must not fire on empty snapshot")
-			return nil, nil, false
-		})
-	req := exec.Request{Prompt: "帮我改一下系统网关地址"}
-	want, werr := bare.Execute(context.Background(), req)
-	got, gerr := gatedBe.Execute(context.Background(), req)
-	if werr != nil || gerr != nil {
-		t.Fatalf("execute errors: %v %v", werr, gerr)
-	}
-	if !reflect.DeepEqual(want, got) {
-		t.Fatalf("results differ: %+v vs %+v", want, got)
-	}
-	if gatedBe.(*gated).next.(*stubBackend).calls != 1 {
-		t.Fatal("next backend not invoked exactly once")
-	}
-}
-
-func TestGuardHitInterceptsWithoutModel(t *testing.T) {
-	next := &stubBackend{name: "native"}
-	be := Wrap(next, newRec(), loadedSnapshot(t),
-		func(ctx context.Context, req exec.Request, res intent.Result) (*exec.Result, error, bool) {
-			if res.Branch != intent.BranchGuardHit || res.Top == nil || res.Top.ID != "guard.config_change" {
-				t.Fatalf("unexpected result: %+v", res)
-			}
-			// 宿主解释守卫意图：拒绝 + 引导（合成结果语义由宿主定义）。
-			return &exec.Result{Success: false, ErrorReason: "guard refused: " + res.Top.Metadata["guidance"]}, nil, true
-		})
-	out, err := be.Execute(context.Background(), exec.Request{Prompt: "帮我改一下系统网关地址"})
-	if err != nil {
-		t.Fatal(err)
-	}
-	if out.Success || out.ErrorReason == "" {
-		t.Fatalf("expected guard refusal, got %+v", out)
-	}
-	if next.calls != 0 {
-		t.Fatalf("model backend must not be called on guard hit, got %d calls", next.calls)
-	}
-}
-
-func TestFallbackPassesThroughToBackend(t *testing.T) {
-	next := &stubBackend{name: "native"}
-	be := Wrap(next, newRec(), loadedSnapshot(t),
-		func(ctx context.Context, req exec.Request, res intent.Result) (*exec.Result, error, bool) {
-			t.Fatal("handler must not fire on fallback")
-			return nil, nil, false
-		})
-	out, err := be.Execute(context.Background(), exec.Request{Prompt: "写一份周报"})
-	if err != nil || out == nil || !out.Success {
-		t.Fatalf("fallback must reach backend: %+v %v", out, err)
-	}
-	if next.calls != 1 {
-		t.Fatalf("backend calls = %d", next.calls)
-	}
-}
-
-func TestHandlerDeclineFallsBackToBackend(t *testing.T) {
-	next := &stubBackend{name: "native"}
-	be := Wrap(next, newRec(), loadedSnapshot(t),
-		func(ctx context.Context, req exec.Request, res intent.Result) (*exec.Result, error, bool) {
-			return nil, nil, false // 宿主放弃处理
-		})
-	out, err := be.Execute(context.Background(), exec.Request{Prompt: "帮我改一下系统网关地址"})
-	if err != nil || out == nil || !out.Success {
-		t.Fatalf("declined handling must pass through: %+v %v", out, err)
-	}
-	if next.calls != 1 {
-		t.Fatalf("backend calls = %d", next.calls)
-	}
-}
-
-func TestNilComponentsReturnNext(t *testing.T) {
-	next := &stubBackend{name: "native"}
-	if got := Wrap(next, nil, nil, nil); got != exec.Backend(next) {
-		t.Fatal("nil components must return next unchanged")
-	}
-}
-
-func TestRecognizeErrorFailsClosed(t *testing.T) {
-	next := &stubBackend{name: "native"}
-	b := intent.NewRegistryBuilder(intent.Options{MaxInputLen: 8})
-	reg, err := b.Build()
-	if err != nil {
-		t.Fatal(err)
-	}
-	be := Wrap(next, newRec(), reg.Snapshot(),
-		func(ctx context.Context, req exec.Request, res intent.Result) (*exec.Result, error, bool) {
-			t.Fatal("handler must not fire on recognize error")
-			return nil, nil, false
-		})
-	_, err = be.Execute(context.Background(), exec.Request{Prompt: "这一段输入明显超过八个字节"})
-	if err == nil {
-		t.Fatal("expected fail-closed error")
-	}
-	var ie *intent.Error
-	if !errors.As(err, &ie) {
-		t.Fatalf("expected *intent.Error, got %T", err)
-	}
-	if next.calls != 0 {
-		t.Fatal("backend must not run on recognize error")
-	}
-}
@@ -0,0 +1,26 @@
+{"id":"025357339a97c04d","timestamp":"2026-10-07T06:11:15.521547827Z","type":"message","message":{"role":"system","content":"## Environment\ndate: 2026-10-07（周三）\nmodel: fake\nworkspace: /home/weichen/projects/Agent-System-Dev-wt-feature-agent-stack-n1/deps/agent-native/adapters/native\ntool profile: workspace\n\nDeliverable files: output.md. Use workspace tools to write those files. Do not invent paths outside the workspace.\n\nTools available in this workspace:\n- read_file: Read a UTF-8 text file inside the workspace. Without offset/limit returns the whole file (truncated at 200 KiB); with offset/limit returns numbered lines from a 1-based offset.\n- write_file: Create or replace a UTF-8 text file inside the workspace.\n- edit_file: Replace one unique exact text span in a workspace file.\n- multi_edit: Apply several unique-span replacements to one workspace file atomically: either all edits land or the file is left untouched.\n- list_dir: List names in a workspace directory.\n- grep: Search workspace text files. Literal substring by default; regex=true switches to Go regexp, include filters file names (e.g. *.go), ignore_case folds case. Caps at 50 hits.\n- glob: Find workspace files matching a glob pattern (** crosses directories, * within one segment). Returns up to 500 relative paths."}}
+{"id":"46a70655ee65ddc3","parentId":"025357339a97c04d","timestamp":"2026-10-07T06:11:15.524296265Z","type":"message","message":{"role":"user","content":"hi"}}
+{"id":"ab242da6bd626cae","parentId":"46a70655ee65ddc3","timestamp":"2026-10-07T06:11:15.524748291Z","type":"message","message":{"role":"assistant","content":"hi","stop_reason":"end"}}
+{"id":"f637d2872ab9f34b","parentId":"ab242da6bd626cae","timestamp":"2026-10-07T06:11:15.525362933Z","type":"message","message":{"role":"user"}}
+{"id":"c1bada8c14ef7ec0","parentId":"f637d2872ab9f34b","timestamp":"2026-10-07T06:11:15.525828249Z","type":"message","message":{"role":"assistant","content":"hi","stop_reason":"end"}}
+{"id":"a01a79f9d2d2c219","parentId":"c1bada8c14ef7ec0","timestamp":"2026-10-07T06:11:15.52644899Z","type":"message","message":{"role":"user","content":"hi"}}
+{"id":"37ce369360ed5bc3","parentId":"a01a79f9d2d2c219","timestamp":"2026-10-07T06:11:22.666535507Z","type":"message","message":{"role":"user","content":"hi"}}
+{"id":"64a2479c0d206bf1","parentId":"37ce369360ed5bc3","timestamp":"2026-10-07T06:11:22.668999313Z","type":"message","message":{"role":"assistant","content":"hi","stop_reason":"end"}}
+{"id":"61593f441df28345","parentId":"64a2479c0d206bf1","timestamp":"2026-10-07T06:11:22.669756111Z","type":"message","message":{"role":"user"}}
+{"id":"c5aa1aebdaa45000","parentId":"61593f441df28345","timestamp":"2026-10-07T06:11:22.670242386Z","type":"message","message":{"role":"assistant","content":"hi","stop_reason":"end"}}
+{"id":"3a9f8790c99f9e94","parentId":"c5aa1aebdaa45000","timestamp":"2026-10-07T06:11:22.670926846Z","type":"message","message":{"role":"user","content":"hi"}}
+{"id":"064604f3b2f1ad65","parentId":"3a9f8790c99f9e94","timestamp":"2026-10-07T06:13:48.448571493Z","type":"message","message":{"role":"user","content":"hi"}}
+{"id":"9d3b5edce163fdbe","parentId":"064604f3b2f1ad65","timestamp":"2026-10-07T06:13:48.449399929Z","type":"message","message":{"role":"assistant","content":"hi","stop_reason":"end"}}
+{"id":"f280fb470f6e90f5","parentId":"9d3b5edce163fdbe","timestamp":"2026-10-07T06:13:48.450112717Z","type":"message","message":{"role":"user"}}
+{"id":"258b37ad6c79c52d","parentId":"f280fb470f6e90f5","timestamp":"2026-10-07T06:13:48.450611112Z","type":"message","message":{"role":"assistant","content":"hi","stop_reason":"end"}}
+{"id":"b6462f3dbbd35ce9","parentId":"258b37ad6c79c52d","timestamp":"2026-10-07T06:13:48.451303652Z","type":"message","message":{"role":"user","content":"hi"}}
+{"id":"1f742b0f9b7c92a1","parentId":"b6462f3dbbd35ce9","timestamp":"2026-10-07T06:13:48.971444939Z","type":"message","message":{"role":"user","content":"hi"}}
+{"id":"f472774cc9cb019d","parentId":"1f742b0f9b7c92a1","timestamp":"2026-10-07T06:13:48.973639624Z","type":"message","message":{"role":"assistant","content":"hi","stop_reason":"end"}}
+{"id":"27e484d323f079d3","parentId":"f472774cc9cb019d","timestamp":"2026-10-07T06:13:48.97478444Z","type":"message","message":{"role":"user"}}
+{"id":"9a97876386e882d4","parentId":"27e484d323f079d3","timestamp":"2026-10-07T06:13:48.97545809Z","type":"message","message":{"role":"assistant","content":"hi","stop_reason":"end"}}
+{"id":"a8a4c88afce69517","parentId":"9a97876386e882d4","timestamp":"2026-10-07T06:13:48.976690973Z","type":"message","message":{"role":"user","content":"hi"}}
+{"id":"c2865055fe240f43","parentId":"a8a4c88afce69517","timestamp":"2026-10-07T06:14:02.592947456Z","type":"message","message":{"role":"user","content":"hi"}}
+{"id":"a02f22b1760c8363","parentId":"c2865055fe240f43","timestamp":"2026-10-07T06:14:02.595218749Z","type":"message","message":{"role":"assistant","content":"hi","stop_reason":"end"}}
+{"id":"c3321f8e2d467eda","parentId":"a02f22b1760c8363","timestamp":"2026-10-07T06:14:02.59619084Z","type":"message","message":{"role":"user"}}
+{"id":"7d0eb0b5d39283b2","parentId":"c3321f8e2d467eda","timestamp":"2026-10-07T06:14:02.596897779Z","type":"message","message":{"role":"assistant","content":"hi","stop_reason":"end"}}
+{"id":"f99d01430932d356","parentId":"7d0eb0b5d39283b2","timestamp":"2026-10-07T06:14:02.598248098Z","type":"message","message":{"role":"user","content":"hi"}}
@@ -0,0 +1,7 @@
+{
+  "exec_id": "cx",
+  "leaf": "f99d01430932d356",
+  "model": "fake",
+  "scope": "",
+  "updated": "2026-10-07T06:14:02Z"
+}
@@ -0,0 +1 @@
+stream ended without a message
@@ -0,0 +1 @@
+hi
@@ -14,6 +14,9 @@ import (

 	"agent-native/exec"
 	"agent-native/runtime"
+	"agent-native/runtime/checkpoint"
+	"agent-native/runtime/filestate"
+	"agent-native/runtime/hooks"
 	"agent-native/runtime/llm"
 	"agent-native/runtime/mcp"
 	"agent-native/runtime/promptx"
@@ -25,39 +28,68 @@ import (
 	"logging-system/logsys"
 )

-// Options 构造后端所需的模型接入与执行参数（由组合根注入，无产品依赖）。
-type Options struct {
+// ModelOptions 是模型接入。Stream 非空时不再按 APIBase/APIKey 构造客户端。
+type ModelOptions struct {
 	APIBase  string
 	APIKey   string
 	Model    string
 	MaxTurns int
 	Timeout  time.Duration
-	// Stream 测试注入的流函数；缺省按 APIBase/APIKey 构造 llm.Client。
-	Stream llm.StreamFunc
-	// Search 可选：提供时注册 tavily_search / web_search（密钥由组合根注入）。
-	Search tools.SearchProvider
-	// Exec 可选：提供时注册 run_command / run_python（解释器留空项不注册）。
-	// nil = 不注册执行类工具（fail-closed，与 BashPermission 默认 deny 对齐）。
-	Exec *tools.ExecConfig
-	// Fetch 可选：提供时注册 fetch_url；nil = 不注册（fail-closed）。
-	Fetch *tools.FetchConfig
-	// MCP 可选：提供时把各 server 工具以 mcp__<server>__<tool> 注册进工作区。
-	// nil = 不注册（fail-closed：server 不可达时整个步骤失败，不静默跳过）。
-	MCP *mcp.Manager
-	// SkillsRoots 可选：全局技能根目录；执行时自动并入工作区 .agent/skills/。
-	SkillsRoots []string
-	// Compact 可选：转录超过阈值时把头部历史压成摘要（长会话上下文治理）。
-	Compact *CompactConfig
-	// Subagent 可选：提供时注册 delegate_task（委派由宿主实现，内核只开端口）。
-	Subagent subagent.Factory
-	// HostTools 是宿主产品工具（登记工件、卡片、知识库）。与内核工具同名时失败，不覆盖。
-	HostTools []tools.Tool
-	// MaxToolCalls 是单步工具调用上限。0 表示不限制。请求自带预算时以请求为准。
+	Stream   llm.StreamFunc
+}
+
+// ToolOptions 是可选工具装配。nil 字段不注册对应工具。
+type ToolOptions struct {
+	Search           tools.SearchProvider
+	Exec             *tools.ExecConfig
+	Fetch            *tools.FetchConfig
+	MCP              *mcp.Manager
+	SkillsRoots      []string
+	Subagent         subagent.Factory
+	HostTools        []tools.Tool
+	Checkpoints      *checkpoint.Log
+	CheckpointAnchor func() string
+	// CodeOrchestration 为真时注册 run_tool_script。默认关。
+	CodeOrchestration bool
+}
+
+// SessionOptions 是单次执行的会话预算与压缩。
+type SessionOptions struct {
+	Compact      *CompactConfig
 	MaxToolCalls int
-	// Log 可选。nil 时丢弃日志。
-	Log *logsys.Logger
-	// ExecutionID 写入日志，不参与执行控制。
+}
+
+// ObserveOptions 是日志。Log 为 nil 时丢弃。
+type ObserveOptions struct {
+	Log         *logsys.Logger
 	ExecutionID string
+	// Trace 交给运行循环。nil 不采集。
+	Trace func(name string, attrs map[string]string) func()
+}
+
+// GateOptions 把直连已经使用的停止、提醒和工具钩子交给小队成员。零值不改变行为。
+type GateOptions struct {
+	BeforeTool  func(call llm.ToolCall) error
+	AfterTool   func(call llm.ToolCall, res llm.Message)
+	Stop        func(last llm.Message, info runtime.StopInfo) runtime.StopDecision
+	Reminders   func(turn int) []string
+	RepeatGuard bool
+	Hooks       *hooks.Bus
+	// Ask 在 MCP 工具要求用户输入时调用。nil 则把问题作为错误返回。
+	Ask func(ctx context.Context, question string) (string, error)
+	// PermissionAsk 在内核钩子对工具表态 ask 时向宿主请求裁决。
+	// nil 时内核保持旧行为：ask 与 deny 一样按拒绝处理。
+	PermissionAsk func(ctx context.Context, call llm.ToolCall) runtime.AskVerdict
+}
+
+// Options 构造后端所需的模型接入与执行参数（由组合根注入，无产品依赖）。
+// 字段经嵌入提升，读取仍写 opts.Model；复合字面量要写子结构名。
+type Options struct {
+	ModelOptions
+	ToolOptions
+	SessionOptions
+	ObserveOptions
+	GateOptions
 }

 // CompactConfig 声明长会话压缩策略（优化性质：压缩失败不阻断执行，但留日志）。
@@ -122,6 +154,9 @@ func (b *Backend) Execute(ctx context.Context, req exec.Request) (*exec.Result,
 	if b.opts.Stream == nil && (strings.TrimSpace(b.opts.APIBase) == "" || strings.TrimSpace(b.opts.APIKey) == "") {
 		return &exec.Result{Success: false, ErrorReason: "agent-native backend requires api base/key (or injected stream)"}, nil
 	}
+	if b.opts.Ask != nil {
+		ctx = mcp.WithAsk(ctx, b.opts.Ask)
+	}
 	model := strings.TrimSpace(b.opts.Model)
 	if strings.TrimSpace(req.Model) != "" {
 		model = strings.TrimSpace(req.Model)
@@ -164,7 +199,7 @@ func (b *Backend) Execute(ctx context.Context, req exec.Request) (*exec.Result,

 	var reg *tools.Registry
 	if profile == exec.ToolProfileWorkspace {
-		reg = tools.WorkspaceTools(workDir)
+		reg = tools.WorkspaceToolsCheckpointed(workDir, filestate.New(), b.opts.Checkpoints, b.opts.CheckpointAnchor)
 		skills.RegisterLoader(reg, skillStore)
 		if b.opts.Search != nil {
 			tools.RegisterSearch(reg, b.opts.Search)
@@ -179,9 +214,19 @@ func (b *Backend) Execute(ctx context.Context, req exec.Request) (*exec.Result,
 			}
 			tools.RegisterExecTools(reg, workDir, execCfg)
 		}
+		if b.opts.CodeOrchestration {
+			tools.RegisterCodeTool(reg)
+		}
+		reg.SetActivationFile(filepath.Join(workDir, ".agentx", "activated-tools.json"))
+		_ = reg.RestoreActivated()
 		if b.opts.MCP != nil {
-			if err := b.opts.MCP.RegisterInto(ctx, reg); err != nil {
-				return &exec.Result{Success: false, ErrorReason: err.Error()}, nil
+			switch {
+			case len(req.ToolAllowlist) > 0 || b.opts.MCP.Ready():
+				if err := b.opts.MCP.RegisterInto(ctx, reg); err != nil {
+					return &exec.Result{Success: false, ErrorReason: err.Error()}, nil
+				}
+			default:
+				b.opts.MCP.StartBackground(ctx, reg)
 			}
 		}
 		subagent.Register(reg, b.opts.Subagent)
@@ -276,9 +321,7 @@ func (b *Backend) Execute(ctx context.Context, req exec.Request) (*exec.Result,

 	if runRes.Err != nil {
 		reason := runRes.Err.Error()
-		b.mu.Lock()
-		b.lastFailure[workDir] = reason
-		b.mu.Unlock()
+		b.noteFailure(workDir, reason)
 		assessment := exec.AssessStepFailure(exec.StepFailureSignals{
 			Err:     runRes.Err,
 			WorkDir: workDir,
@@ -293,9 +336,7 @@ func (b *Backend) Execute(ctx context.Context, req exec.Request) (*exec.Result,
 		}
 		return res, nil
 	}
-	b.mu.Lock()
-	delete(b.lastFailure, workDir)
-	b.mu.Unlock()
+	b.clearFailure(workDir)

 	patches, matErr := exec.MaterializeTargets(workDir, req.Targets, nil)
 	if matErr != nil {
@@ -319,7 +360,7 @@ func (b *Backend) run(ctx context.Context, model string, budget *runtime.Budget,
 	if strings.TrimSpace(execID) == "" {
 		execID = b.opts.ExecutionID
 	}
-	runRes, err := runtime.Run(ctx, runtime.Config{
+	agent := runtime.NewAgentSession(runtime.Config{
 		Model:        model,
 		SystemPrompt: b.systemPrompt(req, model, profile, workDir, skillIndex),
 		Tools:        reg,
@@ -329,6 +370,11 @@ func (b *Backend) run(ctx context.Context, model string, budget *runtime.Budget,
 		Log:          b.logger(),
 		ExecutionID:  execID,
 		Budget:       budget,
+		Reminders:    b.opts.Reminders,
+		RepeatGuard:  b.opts.RepeatGuard,
+		Hooks:        gateBus(b.opts.GateOptions),
+		PermissionAsk: b.opts.PermissionAsk,
+		Trace:        b.opts.Trace,
 		OnEvent: func(ev runtime.Event) {
 			logs.WriteString(ev.Type)
 			if ev.Tool != "" {
@@ -351,8 +397,8 @@ func (b *Backend) run(ctx context.Context, model string, budget *runtime.Budget,
 				})
 			}
 		},
-	}, llm.Message{Role: llm.RoleUser, Content: req.Prompt})
-	return runRes, err
+	})
+	return agent.Prompt(ctx, llm.Message{Role: llm.RoleUser, Content: req.Prompt})
 }

 // maybeCompact 在转录超过阈值时压缩头部历史。压缩是优化性质：
@@ -419,6 +465,40 @@ func defaultSummarizer(stream llm.StreamFunc, model string) func(ctx context.Con
 	}
 }

+func gateBus(g GateOptions) *hooks.Bus {
+	bus := g.Hooks
+	if g.BeforeTool == nil && g.AfterTool == nil && g.Stop == nil {
+		return bus
+	}
+	if bus == nil {
+		bus = hooks.New()
+	}
+	if g.BeforeTool != nil {
+		fn := g.BeforeTool
+		bus.On(hooks.AdaptBefore(func(name, args string) error { return fn(llm.ToolCall{Name: name, Arguments: args}) }))
+	}
+	if g.AfterTool != nil {
+		fn := g.AfterTool
+		bus.On(hooks.AdaptAfter(func(name, args, result string, isError bool) {
+			fn(llm.ToolCall{Name: name, Arguments: args}, llm.Message{Content: result, IsError: isError})
+		}))
+	}
+	if g.Stop != nil {
+		fn := g.Stop
+		bus.On(func(c hooks.Call) hooks.Outcome {
+			if c.Event != hooks.Stop {
+				return hooks.Outcome{}
+			}
+			d := fn(llm.Message{Role: llm.RoleAssistant, Content: c.LastText}, runtime.StopInfo{Continuations: c.Continuations, ToolCalls: c.ToolCalls})
+			if d.Continue {
+				return hooks.Outcome{Decision: hooks.Block, Reason: d.Reason, Replacement: d.Replacement, Notice: d.Notice}
+			}
+			return hooks.Outcome{Decision: hooks.Allow, Notice: d.Notice}
+		})
+	}
+	return bus
+}
+
 type errString string

 func (e errString) Error() string { return string(e) }
@@ -433,9 +513,7 @@ var resumeUnsafeMarkers = []string{

 // rotateIfPoisoned 上次失败属 resume-unsafe 且存在旧会话时归档旧会话。
 func (b *Backend) rotateIfPoisoned(workDir string) {
-	b.mu.Lock()
-	reason := strings.ToLower(b.lastFailure[workDir])
-	b.mu.Unlock()
+	reason := strings.ToLower(b.failureReason(workDir))
 	if reason == "" {
 		return
 	}
@@ -469,6 +547,36 @@ func (b *Backend) logger() *logsys.Logger {
 	return log.With("execution_id", b.opts.ExecutionID)
 }

+const lastFailureFile = ".native-last-failure"
+
+func (b *Backend) noteFailure(workDir, reason string) {
+	b.mu.Lock()
+	b.lastFailure[workDir] = reason
+	b.mu.Unlock()
+	_ = os.WriteFile(filepath.Join(workDir, lastFailureFile), []byte(reason), 0o644)
+}
+
+func (b *Backend) clearFailure(workDir string) {
+	b.mu.Lock()
+	delete(b.lastFailure, workDir)
+	b.mu.Unlock()
+	_ = os.Remove(filepath.Join(workDir, lastFailureFile))
+}
+
+func (b *Backend) failureReason(workDir string) string {
+	b.mu.Lock()
+	reason := b.lastFailure[workDir]
+	b.mu.Unlock()
+	if reason != "" {
+		return reason
+	}
+	raw, err := os.ReadFile(filepath.Join(workDir, lastFailureFile))
+	if err != nil {
+		return ""
+	}
+	return string(raw)
+}
+
 // systemPrompt 经 promptx 组装：角色指令 + 环境事实 + 技能索引 + 交付物约束。
 func (b *Backend) systemPrompt(req exec.Request, model, profile, workDir string, skillIndex []skills.Skill) string {
 	var index []promptx.IndexEntry
@@ -2,6 +2,7 @@ package native

 import (
 	"context"
+	"errors"
 	"os"
 	"path/filepath"
 	"strings"
@@ -39,12 +40,17 @@ func script(turns ...llm.Message) llm.StreamFunc {

 func TestBackendWorkspaceProfileWritesDeliverable(t *testing.T) {
 	dir := t.TempDir()
-	b := New(Options{Model: "fake", Stream: script(
-		llm.Message{Role: llm.RoleAssistant, StopReason: llm.StopToolCalls, ToolCalls: []llm.ToolCall{{
-			ID: "c1", Name: "write_file", Arguments: `{"path":"output.md","content":"# 交付\n\n工具写入的交付物内容。","reason":"产出交付物"}`,
-		}}},
-		llm.Message{Role: llm.RoleAssistant, Content: "已写入 output.md", StopReason: llm.StopEnd},
-	)})
+	b := New(Options{
+		ModelOptions: ModelOptions{
+			Model: "fake",
+			Stream: script(
+				llm.Message{Role: llm.RoleAssistant, StopReason: llm.StopToolCalls, ToolCalls: []llm.ToolCall{{
+					ID: "c1", Name: "write_file", Arguments: `{"path":"output.md","content":"# 交付\n\n工具写入的交付物内容。"}`,
+				}}},
+				llm.Message{Role: llm.RoleAssistant, Content: "已写入 output.md", StopReason: llm.StopEnd},
+			),
+		},
+	})
 	var events []exec.ProcessEvent
 	res, err := b.Execute(context.Background(), exec.Request{
 		WorkDir: dir, Prompt: "写交付物", Instructions: "你是测试角色",
@@ -81,9 +87,14 @@ func TestBackendWorkspaceProfileWritesDeliverable(t *testing.T) {

 func TestBackendReasoningProfileWritesText(t *testing.T) {
 	dir := t.TempDir()
-	b := New(Options{Model: "fake", Stream: script(
-		llm.Message{Role: llm.RoleAssistant, Content: "纯文本交付内容（无工具）。", StopReason: llm.StopEnd},
-	)})
+	b := New(Options{
+		ModelOptions: ModelOptions{
+			Model: "fake",
+			Stream: script(
+				llm.Message{Role: llm.RoleAssistant, Content: "纯文本交付内容（无工具）。", StopReason: llm.StopEnd},
+			),
+		},
+	})
 	res, err := b.Execute(context.Background(), exec.Request{
 		WorkDir: dir, Prompt: "直接回答", ToolProfile: exec.ToolProfileReasoning,
 		Targets: []string{"notes.md"},
@@ -109,14 +120,56 @@ func TestBackendFailClosedWithoutConfig(t *testing.T) {
 	if res.Success || !strings.Contains(res.ErrorReason, "requires") {
 		t.Fatalf("want fail-closed config error, got %+v", res)
 	}
-	res2, _ := New(Options{APIBase: "http://x", APIKey: "k"}).Execute(context.Background(),
+	res2, _ := New(Options{
+		ModelOptions: ModelOptions{
+			APIBase: "http://x",
+			APIKey:  "k",
+		},
+	}).Execute(context.Background(),
 		exec.Request{WorkDir: t.TempDir(), Targets: []string{"x.md"}})
 	if res2.Success || !strings.Contains(res2.ErrorReason, "model") {
 		t.Fatalf("want model required error, got %+v", res2)
 	}
-	res3, _ := New(Options{Stream: script()}).Execute(context.Background(),
+	res3, _ := New(Options{
+		ModelOptions: ModelOptions{
+			Stream: script(),
+		},
+	}).Execute(context.Background(),
 		exec.Request{WorkDir: t.TempDir(), Targets: []string{"x.md"}, ToolProfile: "bogus"})
 	if res3.Success || !strings.Contains(res3.ErrorReason, "tool profile") {
 		t.Fatalf("want profile error, got %+v", res3)
 	}
 }
+
+func TestBeforeToolReachesSquadMembers(t *testing.T) {
+	dir := t.TempDir()
+	var seen string
+	b := New(Options{
+		ModelOptions: ModelOptions{
+			Model: "fake",
+			Stream: script(
+				llm.Message{Role: llm.RoleAssistant, StopReason: llm.StopToolCalls, ToolCalls: []llm.ToolCall{{
+					ID: "c1", Name: "write_file", Arguments: `{"path":"output.md","content":"nope"}`,
+				}}},
+				llm.Message{Role: llm.RoleAssistant, Content: "blocked", StopReason: llm.StopEnd},
+			),
+		},
+		GateOptions: GateOptions{BeforeTool: func(call llm.ToolCall) error {
+			seen = call.Name
+			return errors.New("计划模式拒绝了写工具")
+		}},
+	})
+	res, err := b.Execute(context.Background(), exec.Request{WorkDir: dir, Prompt: "写", Targets: []string{"output.md"}})
+	if err != nil {
+		t.Fatal(err)
+	}
+	if seen != "write_file" {
+		t.Fatalf("before tool saw %q", seen)
+	}
+	if res.Success {
+		t.Fatal("denied write still succeeded")
+	}
+	if _, err := os.Stat(filepath.Join(dir, "output.md")); err == nil {
+		t.Fatal("write_file ran after deny")
+	}
+}
@@ -30,11 +30,15 @@ func TestBackendCompactBeforeRun(t *testing.T) {
 	var sawSummary bool
 	// 注入式捕获：验证 Run 收到的消息里有摘要标记。
 	bk2 := New(Options{
-		Model:  "fake",
-		Stream: captureFirstStream(&sawSummary),
-		Compact: &CompactConfig{ThresholdChars: 1000, KeepLastN: 3, Summarizer: func(ctx context.Context, messages []llm.Message) (string, error) {
-			return "预置历史摘要", nil
-		}},
+		ModelOptions: ModelOptions{
+			Model:  "fake",
+			Stream: captureFirstStream(&sawSummary),
+		},
+		SessionOptions: SessionOptions{
+			Compact: &CompactConfig{ThresholdChars: 1000, KeepLastN: 3, Summarizer: func(ctx context.Context, messages []llm.Message) (string, error) {
+				return "预置历史摘要", nil
+			}},
+		},
 	})
 	res, err := bk2.Execute(context.Background(), exec.Request{
 		WorkDir: dir, Prompt: "继续", Targets: []string{"output.md"},
@@ -47,8 +51,11 @@ func TestBackendCompactBeforeRun(t *testing.T) {
 		t.Fatal("compaction should have run before the loop")
 	}
 	data, _ := os.ReadFile(transcript)
-	if !strings.Contains(string(data), "[conversation summary]") {
-		t.Fatalf("transcript should contain the summary marker, got:\n%.200s", data)
+	if !strings.Contains(string(data), "预置历史摘要") || !strings.Contains(string(data), `"type":"compaction"`) {
+		t.Fatalf("transcript should keep history and append a compaction entry, got:\n%.400s", data)
+	}
+	if !strings.Contains(string(data), `"role":"system"`) {
+		t.Fatal("compaction must not rewrite the original transcript")
 	}
 }

@@ -0,0 +1,89 @@
+package native
+
+import (
+	"context"
+	"errors"
+	"strings"
+
+	"agent-native/exec"
+)
+
+// contractBackend 把 Backend 适配为 exec.ContractBackend。
+// 独立类型的原因：旧门面 Execute(ctx, exec.Request) 与契约的
+// Execute(ctx, exec.ContractRequest) 签名冲突，同名方法不能并存；
+// 包装让现有 squad 消费（旧门面）与契约消费（conformance/Core）互不影响。
+type contractBackend struct{ b *Backend }
+
+// Contract 返回该后端的契约形态（N1 WP6；旧门面保持不变）。
+func (b *Backend) Contract() exec.ContractBackend {
+	return contractBackend{b: b}
+}
+
+func (c contractBackend) Descriptor() exec.BackendDescriptor {
+	return exec.BackendDescriptor{
+		Name:         c.b.Name(),
+		Capabilities: []exec.Capability{exec.CapText, exec.CapStream, exec.CapTools},
+	}
+}
+
+func (c contractBackend) Execute(ctx context.Context, req exec.ContractRequest) (*exec.ContractResult, error) {
+	var onEvent func(ev exec.ProcessEvent)
+	if req.OnEvent != nil {
+		next := req.OnEvent
+		onEvent = func(ev exec.ProcessEvent) {
+			next(exec.Event{Kind: exec.EventKind(ev.Type), Text: ev.Text, Tool: ev.Tool, Detail: ev.Detail})
+		}
+	}
+	// 契约语义：空 ToolPolicy.Profile 等同 reasoning（无工具纯文本任务）。
+	profile := strings.TrimSpace(req.ToolPolicy.Profile)
+	if profile == "" {
+		profile = exec.ToolProfileReasoning
+	}
+	old := exec.Request{
+		Prompt:        req.Prompt,
+		Instructions:  req.Instructions,
+		ExecutionID:   req.ExecutionID,
+		Targets:       append([]string(nil), req.Targets...),
+		ToolProfile:   profile,
+		ToolAllowlist: append([]string(nil), req.ToolPolicy.Allowlist...),
+		WorkDir:       req.Workspace.Root,
+		OnEvent:       onEvent,
+	}
+	if b := req.ContractBudget; b.MaxPromptTokens > 0 || b.MaxCompletionTokens > 0 || !b.Deadline.IsZero() {
+		old.Budget = &exec.Budget{
+			MaxPromptTokens:     b.MaxPromptTokens,
+			MaxCompletionTokens: b.MaxCompletionTokens,
+			Deadline:            b.Deadline,
+		}
+	}
+	res, err := c.b.Execute(ctx, old)
+	if err != nil {
+		// 传输/实现级错误：按契约映射为 Unknown，不经 Go error 返回业务终态。
+		return &exec.ContractResult{
+			Outcome: exec.OutcomeUnknown,
+			Err:     &exec.Error{Code: exec.ErrProviderUnavailable, Reason: err.Error(), SideEffects: exec.SideEffectPossible},
+		}, nil
+	}
+	out := &exec.ContractResult{
+		Logs: res.Logs,
+		Usage: exec.Usage{
+			PromptTokens:     res.PromptTokens,
+			CompletionTokens: res.CompletionTokens,
+		},
+	}
+	switch {
+	case res.Success:
+		out.Outcome = exec.OutcomeSucceeded
+		out.Files = res.Patches
+	case errors.Is(ctx.Err(), context.Canceled):
+		out.Outcome = exec.OutcomeCanceled
+		out.Err = &exec.Error{Code: exec.ErrCanceled, Reason: res.ErrorReason, SideEffects: exec.SideEffectPossible}
+	case errors.Is(ctx.Err(), context.DeadlineExceeded):
+		out.Outcome = exec.OutcomeFailed
+		out.Err = &exec.Error{Code: exec.ErrDeadlineExceeded, Reason: res.ErrorReason, SideEffects: exec.SideEffectPossible}
+	default:
+		out.Outcome = exec.OutcomeFailed
+		out.Err = &exec.Error{Code: exec.ErrProviderRejected, Reason: res.ErrorReason, Retryable: res.Retryable, SideEffects: exec.SideEffectPossible}
+	}
+	return out, nil
+}
@@ -0,0 +1,67 @@
+package native
+
+import (
+	"context"
+	"testing"
+	"time"
+
+	"agent-native/conformance"
+	"agent-native/exec"
+	"agent-native/runtime/llm"
+)
+
+// TestContractBackendConformance 把 native 后端的契约形态挂上最小契约套件
+// （N1 WP6：conformance 首次获得真实实现调用方）。
+func TestContractBackendConformance(t *testing.T) {
+	conformance.Suite(t, func() exec.ContractBackend {
+		return New(Options{
+			ModelOptions: ModelOptions{
+				Model: "fake",
+				Stream: script(
+					llm.Message{Role: llm.RoleAssistant, Content: "hi", StopReason: llm.StopEnd},
+				),
+			},
+		}).Contract()
+	})
+}
+
+// TestContractExecuteCancel 取消路径：流阻塞到 ctx 结束，Outcome=Canceled 且
+// SideEffects 显式。
+func TestContractExecuteCancel(t *testing.T) {
+	b := New(Options{
+		ModelOptions: ModelOptions{
+			Model: "fake",
+			Stream: func(ctx context.Context, model string, messages []llm.Message, decls []llm.ToolDecl, opts llm.StreamOptions) <-chan llm.Event {
+				ch := make(chan llm.Event)
+				go func() {
+					<-ctx.Done()
+					close(ch)
+				}()
+				return ch
+			},
+		},
+	}).Contract()
+	ctx, cancel := context.WithTimeout(context.Background(), 20*time.Millisecond)
+	defer cancel()
+	res, err := b.Execute(ctx, exec.ContractRequest{ExecutionID: "cx", Prompt: "hi"})
+	if err != nil {
+		t.Fatalf("transport error: %v", err)
+	}
+	if res.Outcome != exec.OutcomeCanceled && res.Outcome != exec.OutcomeFailed {
+		t.Fatalf("outcome = %q", res.Outcome)
+	}
+	if res.Err == nil || res.Err.SideEffects == "" {
+		t.Fatalf("err = %+v", res.Err)
+	}
+}
+
+// TestContractDescriptor 描述的身份与能力。
+func TestContractDescriptor(t *testing.T) {
+	d := New(Options{}).Contract().Descriptor()
+	if d.Name != "agent-native" {
+		t.Fatalf("name = %q", d.Name)
+	}
+	if !d.Has(exec.CapText) || !d.Has(exec.CapStream) || !d.Has(exec.CapTools) {
+		t.Fatalf("capabilities = %v", d.Capabilities)
+	}
+}
@@ -22,19 +22,23 @@ func TestBackendExecToolsWiredWithOptionsAndRequestEnv(t *testing.T) {
 	}
 	dir := t.TempDir()
 	b := New(Options{
-		Model: "fake",
-		Stream: script(
-			llm.Message{Role: llm.RoleAssistant, StopReason: llm.StopToolCalls, ToolCalls: []llm.ToolCall{{
-				ID: "c1", Name: "run_command",
-				Arguments: `{"command":"echo role=$ROLE_FLAG step=$STEP_FLAG > exec_out.txt"}`,
-			}}},
-			llm.Message{Role: llm.RoleAssistant, StopReason: llm.StopToolCalls, ToolCalls: []llm.ToolCall{{
-				ID: "c2", Name: "write_file",
-				Arguments: `{"path":"output.md","content":"done"}`,
-			}}},
-			llm.Message{Role: llm.RoleAssistant, Content: "ok", StopReason: llm.StopEnd},
-		),
-		Exec: &tools.ExecConfig{Bash: bash, Env: []string{"PATH=/usr/bin:/bin", "ROLE_FLAG=backend"}},
+		ModelOptions: ModelOptions{
+			Model: "fake",
+			Stream: script(
+				llm.Message{Role: llm.RoleAssistant, StopReason: llm.StopToolCalls, ToolCalls: []llm.ToolCall{{
+					ID: "c1", Name: "run_command",
+					Arguments: `{"command":"echo role=$ROLE_FLAG step=$STEP_FLAG > exec_out.txt"}`,
+				}}},
+				llm.Message{Role: llm.RoleAssistant, StopReason: llm.StopToolCalls, ToolCalls: []llm.ToolCall{{
+					ID: "c2", Name: "write_file",
+					Arguments: `{"path":"output.md","content":"done"}`,
+				}}},
+				llm.Message{Role: llm.RoleAssistant, Content: "ok", StopReason: llm.StopEnd},
+			),
+		},
+		ToolOptions: ToolOptions{
+			Exec: &tools.ExecConfig{Bash: bash, Env: []string{"PATH=/usr/bin:/bin", "ROLE_FLAG=backend"}},
+		},
 	})
 	res, err := b.Execute(context.Background(), exec.Request{
 		WorkDir: dir, Prompt: "run", Targets: []string{"output.md"},
@@ -54,7 +58,12 @@ func TestBackendExecToolsWiredWithOptionsAndRequestEnv(t *testing.T) {

 func TestBackendExecToolsAbsentWithoutOptions(t *testing.T) {
 	stream, c := captureStream()
-	b := New(Options{Model: "fake", Stream: stream})
+	b := New(Options{
+		ModelOptions: ModelOptions{
+			Model:  "fake",
+			Stream: stream,
+		},
+	})
 	res, err := b.Execute(context.Background(), exec.Request{
 		WorkDir: t.TempDir(), Prompt: "x", Targets: []string{"output.md"},
 		ToolProfile: exec.ToolProfileReasoning,
@@ -64,7 +73,12 @@ func TestBackendExecToolsAbsentWithoutOptions(t *testing.T) {
 	}
 	// reasoning 档本无工具；改用 workspace 档断言 decls 不含执行工具。
 	stream2, c2 := captureStream()
-	b2 := New(Options{Model: "fake", Stream: stream2})
+	b2 := New(Options{
+		ModelOptions: ModelOptions{
+			Model:  "fake",
+			Stream: stream2,
+		},
+	})
 	_, _ = b2.Execute(context.Background(), exec.Request{
 		WorkDir: t.TempDir(), Prompt: "x", Targets: []string{"output.md"},
 		ToolProfile: exec.ToolProfileWorkspace,
@@ -83,8 +97,13 @@ func TestBackendAllowlistCanDisableExecTools(t *testing.T) {
 	bash, _ := osexec.LookPath("bash")
 	stream, c := captureStream()
 	b := New(Options{
-		Model: "fake", Stream: stream,
-		Exec: &tools.ExecConfig{Bash: bash, Env: []string{"PATH=/usr/bin:/bin"}},
+		ModelOptions: ModelOptions{
+			Model:  "fake",
+			Stream: stream,
+		},
+		ToolOptions: ToolOptions{
+			Exec: &tools.ExecConfig{Bash: bash, Env: []string{"PATH=/usr/bin:/bin"}},
+		},
 	})
 	res, err := b.Execute(context.Background(), exec.Request{
 		WorkDir: t.TempDir(), Prompt: "x", Targets: []string{"output.md"},
@@ -125,7 +144,15 @@ func TestBackendSearchToolsWiredViaOptions(t *testing.T) {
 		}}},
 		llm.Message{Role: llm.RoleAssistant, Content: "ok", StopReason: llm.StopEnd},
 	)
-	b := New(Options{Model: "fake", Stream: stream, Search: provider})
+	b := New(Options{
+		ModelOptions: ModelOptions{
+			Model:  "fake",
+			Stream: stream,
+		},
+		ToolOptions: ToolOptions{
+			Search: provider,
+		},
+	})
 	dir := t.TempDir()
 	res, err := b.Execute(context.Background(), exec.Request{
 		WorkDir: dir, Prompt: "search", Targets: []string{"output.md"},
@@ -152,7 +179,12 @@ func TestBackendSearchToolsWiredViaOptions(t *testing.T) {

 func TestBackendSearchToolsAbsentWithoutOptions(t *testing.T) {
 	stream, c := captureStream()
-	b := New(Options{Model: "fake", Stream: stream})
+	b := New(Options{
+		ModelOptions: ModelOptions{
+			Model:  "fake",
+			Stream: stream,
+		},
+	})
 	_, _ = b.Execute(context.Background(), exec.Request{
 		WorkDir: t.TempDir(), Prompt: "x", Targets: []string{"output.md"},
 		ToolProfile: exec.ToolProfileWorkspace,
@@ -180,6 +212,9 @@ func TestBackendMCPToolsWiredViaOptions(t *testing.T) {
 		t.Fatal(err)
 	}
 	defer mgr.Close()
+	if err := mgr.RegisterInto(context.Background(), tools.NewRegistry(t.TempDir())); err != nil {
+		t.Fatal(err)
+	}
 	stream, c := captureStream(
 		llm.Message{Role: llm.RoleAssistant, StopReason: llm.StopToolCalls, ToolCalls: []llm.ToolCall{{
 			ID: "c1", Name: "mcp__fake__echo", Arguments: `{"text":"via-backend"}`,
@@ -190,7 +225,15 @@ func TestBackendMCPToolsWiredViaOptions(t *testing.T) {
 		}}},
 		llm.Message{Role: llm.RoleAssistant, Content: "ok", StopReason: llm.StopEnd},
 	)
-	b := New(Options{Model: "fake", Stream: stream, MCP: mgr})
+	b := New(Options{
+		ModelOptions: ModelOptions{
+			Model:  "fake",
+			Stream: stream,
+		},
+		ToolOptions: ToolOptions{
+			MCP: mgr,
+		},
+	})
 	dir := t.TempDir()
 	res, err2 := b.Execute(context.Background(), exec.Request{
 		WorkDir: dir, Prompt: "call mcp", Targets: []string{"output.md"},
@@ -229,7 +272,15 @@ func TestBackendSkillsIndexAndLoader(t *testing.T) {
 		}}},
 		llm.Message{Role: llm.RoleAssistant, Content: "ok", StopReason: llm.StopEnd},
 	)
-	b := New(Options{Model: "fake", Stream: stream, SkillsRoots: []string{skillsRoot}})
+	b := New(Options{
+		ModelOptions: ModelOptions{
+			Model:  "fake",
+			Stream: stream,
+		},
+		ToolOptions: ToolOptions{
+			SkillsRoots: []string{skillsRoot},
+		},
+	})
 	dir := t.TempDir()
 	res, err := b.Execute(context.Background(), exec.Request{
 		WorkDir: dir, Prompt: "写", Targets: []string{"output.md"}, Skills: []string{"docx-author"},
@@ -257,7 +308,15 @@ func TestBackendSkillsIndexAndLoader(t *testing.T) {
 }

 func TestBackendDeclaredSkillMissingFailsClosed(t *testing.T) {
-	b := New(Options{Model: "fake", Stream: script(), SkillsRoots: []string{t.TempDir()}})
+	b := New(Options{
+		ModelOptions: ModelOptions{
+			Model:  "fake",
+			Stream: script(),
+		},
+		ToolOptions: ToolOptions{
+			SkillsRoots: []string{t.TempDir()},
+		},
+	})
 	res, err := b.Execute(context.Background(), exec.Request{
 		WorkDir: t.TempDir(), Prompt: "x", Targets: []string{"output.md"},
 		Skills: []string{"ghost-skill"},
@@ -277,7 +336,12 @@ func TestBackendWorkspaceSkillsRootAutoMerged(t *testing.T) {
 		}}},
 		llm.Message{Role: llm.RoleAssistant, Content: "ok", StopReason: llm.StopEnd},
 	)
-	b := New(Options{Model: "fake", Stream: stream})
+	b := New(Options{
+		ModelOptions: ModelOptions{
+			Model:  "fake",
+			Stream: stream,
+		},
+	})
 	dir := t.TempDir()
 	wsSkill := filepath.Join(dir, ".agent", "skills", "ws-skill")
 	if err := os.MkdirAll(wsSkill, 0o755); err != nil {
@@ -311,11 +375,15 @@ func TestBackendSubagentWiredViaOptions(t *testing.T) {
 		llm.Message{Role: llm.RoleAssistant, Content: "ok", StopReason: llm.StopEnd},
 	)
 	b := New(Options{
-		Model:  "fake",
-		Stream: stream,
-		Subagent: func(ctx context.Context, req subagent.Request) (subagent.Result, error) {
-			delegated = &req
-			return subagent.Result{Success: true, Text: "结论已查明"}, nil
+		ModelOptions: ModelOptions{
+			Model:  "fake",
+			Stream: stream,
+		},
+		ToolOptions: ToolOptions{
+			Subagent: func(ctx context.Context, req subagent.Request) (subagent.Result, error) {
+				delegated = &req
+				return subagent.Result{Success: true, Text: "结论已查明"}, nil
+			},
 		},
 	})
 	dir := t.TempDir()
@@ -0,0 +1,4 @@
+execution_id=cx session=
+agent_start
+turn_start
+message_start
@@ -25,9 +25,13 @@ func (h hostTool) Execute(ctx context.Context, args json.RawMessage) (tools.Resu
 func TestHostToolAndUsageEvent(t *testing.T) {
 	var events []exec.ProcessEvent
 	b := New(Options{
-		Model:     "fake",
-		HostTools: []tools.Tool{hostTool{name: "register_artifact"}},
-		Stream:    usageScript(),
+		ModelOptions: ModelOptions{
+			Model:  "fake",
+			Stream: usageScript(),
+		},
+		ToolOptions: ToolOptions{
+			HostTools: []tools.Tool{hostTool{name: "register_artifact"}},
+		},
 	})
 	res, err := b.Execute(context.Background(), exec.Request{
 		WorkDir: t.TempDir(), Prompt: "登记",
@@ -65,9 +69,13 @@ func TestHostToolAndUsageEvent(t *testing.T) {
 func TestHostToolSeesStepWorkDir(t *testing.T) {
 	var seen string
 	b := New(Options{
-		Model:     "fake",
-		Stream:    script(llm.Message{Role: llm.RoleAssistant, StopReason: llm.StopToolCalls, ToolCalls: []llm.ToolCall{{ID: "c1", Name: "peek", Arguments: `{}`}}}),
-		HostTools: []tools.Tool{peekTool{got: &seen}},
+		ModelOptions: ModelOptions{
+			Model:  "fake",
+			Stream: script(llm.Message{Role: llm.RoleAssistant, StopReason: llm.StopToolCalls, ToolCalls: []llm.ToolCall{{ID: "c1", Name: "peek", Arguments: `{}`}}}),
+		},
+		ToolOptions: ToolOptions{
+			HostTools: []tools.Tool{peekTool{got: &seen}},
+		},
 	})
 	dir := t.TempDir()
 	_, err := b.Execute(context.Background(), exec.Request{WorkDir: dir, Prompt: "x", Targets: []string{"output.md"}})
@@ -91,9 +99,13 @@ func (p peekTool) Execute(ctx context.Context, args json.RawMessage) (tools.Resu

 func TestHostToolNameConflictFailsClosed(t *testing.T) {
 	b := New(Options{
-		Model:     "fake",
-		Stream:    script(llm.Message{Role: llm.RoleAssistant, Content: "x", StopReason: llm.StopEnd}),
-		HostTools: []tools.Tool{hostTool{name: "read_file"}},
+		ModelOptions: ModelOptions{
+			Model:  "fake",
+			Stream: script(llm.Message{Role: llm.RoleAssistant, Content: "x", StopReason: llm.StopEnd}),
+		},
+		ToolOptions: ToolOptions{
+			HostTools: []tools.Tool{hostTool{name: "read_file"}},
+		},
 	})
 	res, err := b.Execute(context.Background(), exec.Request{
 		WorkDir: t.TempDir(), Prompt: "x", Targets: []string{"output.md"},
@@ -0,0 +1 @@
+hi
@@ -0,0 +1,65 @@
+package native
+
+import (
+	"context"
+	"encoding/json"
+	"os"
+	"os/exec"
+	"testing"
+
+	"agent-native/rpc"
+	"agent-native/runtime"
+	"agent-native/runtime/llm"
+	"agent-native/runtime/session"
+)
+
+func TestEventAndRPCFixtures(t *testing.T) {
+	var types []string
+	agent := runtime.NewAgentSession(runtime.Config{
+		Model: "fake",
+		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: "ok", StopReason: llm.StopEnd}}
+			close(ch)
+			return ch
+		},
+		OnEvent: func(ev runtime.Event) { types = append(types, ev.Type) },
+	})
+	if _, err := agent.Prompt(context.Background(), llm.Message{Role: llm.RoleUser, Content: "hi"}); err != nil {
+		t.Fatal(err)
+	}
+	raw, _ := json.Marshal(types)
+	for _, want := range []string{"message_start", "message_end", "agent_settled"} {
+		if !contains(types, want) {
+			t.Fatalf("fixture missing %s in %s", want, raw)
+		}
+	}
+	dir := t.TempDir()
+	store, err := session.Open(dir + "/s.jsonl")
+	if err != nil {
+		t.Fatal(err)
+	}
+	srv := &rpc.Server{Agent: agent, Store: store}
+	resp, err := srv.Handle(context.Background(), []byte(`{"type":"get_session_stats"}`))
+	if err != nil || !json.Valid(resp) {
+		t.Fatalf("rpc fixture: %s %v", resp, err)
+	}
+}
+
+func TestPiCrossCompareOptional(t *testing.T) {
+	if _, err := exec.LookPath("pi"); err != nil {
+		t.Skip("pi is not installed")
+	}
+	if os.Getenv("PI_CROSS") != "1" {
+		t.Skip("set PI_CROSS=1 to compare a live pi process")
+	}
+}
+
+func contains(list []string, want string) bool {
+	for _, item := range list {
+		if item == want {
+			return true
+		}
+	}
+	return false
+}
@@ -0,0 +1,40 @@
+package native
+
+import (
+	"os"
+	"path/filepath"
+	"strings"
+	"testing"
+)
+
+func TestR2PoisonMarkerSurvivesNewProcess(t *testing.T) {
+	dir := t.TempDir()
+	if err := os.MkdirAll(filepath.Join(dir, ".agentx"), 0o755); err != nil {
+		t.Fatal(err)
+	}
+	transcript := filepath.Join(dir, ".agentx", "transcript.jsonl")
+	if err := os.WriteFile(transcript, []byte("{}\n"), 0o644); err != nil {
+		t.Fatal(err)
+	}
+	if err := os.WriteFile(filepath.Join(dir, lastFailureFile), []byte("context length exceeded"), 0o644); err != nil {
+		t.Fatal(err)
+	}
+	b := New(Options{})
+	b.rotateIfPoisoned(dir)
+	if _, err := os.Stat(transcript); !os.IsNotExist(err) {
+		t.Fatal("poisoned transcript still at the live path")
+	}
+	entries, err := os.ReadDir(filepath.Join(dir, ".agentx"))
+	if err != nil {
+		t.Fatal(err)
+	}
+	var renamed bool
+	for _, e := range entries {
+		if strings.Contains(e.Name(), "transcript.poisoned.") {
+			renamed = true
+		}
+	}
+	if !renamed {
+		t.Fatal("expected archived transcript")
+	}
+}
@@ -58,7 +58,12 @@ func toolNames(decls []llm.ToolDecl) []string {

 func TestBackendRequestModelOverride(t *testing.T) {
 	stream, c := captureStream()
-	b := New(Options{Model: "default-model", Stream: stream})
+	b := New(Options{
+		ModelOptions: ModelOptions{
+			Model:  "default-model",
+			Stream: stream,
+		},
+	})
 	dir := t.TempDir()
 	res, err := b.Execute(context.Background(), exec.Request{
 		WorkDir: dir, Prompt: "hi", Targets: []string{"output.md"}, Model: "override-model",
@@ -82,7 +87,12 @@ func TestBackendRequestToolAllowlist(t *testing.T) {
 		}}},
 		llm.Message{Role: llm.RoleAssistant, Content: "gave up", StopReason: llm.StopEnd},
 	)
-	b := New(Options{Model: "fake", Stream: stream})
+	b := New(Options{
+		ModelOptions: ModelOptions{
+			Model:  "fake",
+			Stream: stream,
+		},
+	})
 	dir := t.TempDir()
 	// 目标 output2.md 不在 allowlist 内能写吗？——write_file 被整体收窄，
 	// 模型两次尝试都会得到 unknown tool，最终无交付物 → 失败。
@@ -109,7 +119,12 @@ func TestBackendRequestToolAllowlist(t *testing.T) {
 }

 func TestBackendRequestUnknownAllowlistFailsClosed(t *testing.T) {
-	b := New(Options{Model: "fake", Stream: script()})
+	b := New(Options{
+		ModelOptions: ModelOptions{
+			Model:  "fake",
+			Stream: script(),
+		},
+	})
 	res, err := b.Execute(context.Background(), exec.Request{
 		WorkDir: t.TempDir(), Prompt: "x", Targets: []string{"output.md"},
 		ToolAllowlist: []string{"read_file", "no_such_tool"},
@@ -139,7 +154,12 @@ func TestBackendRequestBudgetDeadlineBindsContext(t *testing.T) {
 		}()
 		return ch
 	}
-	b := New(Options{Model: "fake", Stream: stream})
+	b := New(Options{
+		ModelOptions: ModelOptions{
+			Model:  "fake",
+			Stream: stream,
+		},
+	})
 	start := time.Now()
 	res, err := b.Execute(context.Background(), exec.Request{
 		WorkDir: t.TempDir(), Prompt: "x", Targets: []string{"output.md"},
@@ -164,12 +184,17 @@ func TestBackendRequestBudgetDeadlineBindsContext(t *testing.T) {

 func TestBackendRequestExecutionIDInLogs(t *testing.T) {
 	stream, _ := captureStream()
-	b := New(Options{Model: "fake", Stream: stream})
+	b := New(Options{
+		ModelOptions: ModelOptions{
+			Model:  "fake",
+			Stream: stream,
+		},
+	})
 	dir := t.TempDir()
 	res, err := b.Execute(context.Background(), exec.Request{
 		WorkDir: dir, Prompt: "hi", Targets: []string{"output.md"},
 		Instructions: "你是角色",
-		ExecutionID: "exec-42", SessionScope: "run-9/step-3",
+		ExecutionID:  "exec-42", SessionScope: "run-9/step-3",
 		ToolProfile: exec.ToolProfileReasoning,
 	})
 	if err != nil || !res.Success {
@@ -38,7 +38,12 @@ func failThenSucceed(failMsg string) llm.StreamFunc {

 func TestSessionRotationAfterPoisonedFailure(t *testing.T) {
 	dir := t.TempDir()
-	b := New(Options{Model: "fake", Stream: failThenSucceed("maximum context length exceeded (context_overflow)")})
+	b := New(Options{
+		ModelOptions: ModelOptions{
+			Model:  "fake",
+			Stream: failThenSucceed("maximum context length exceeded (context_overflow)"),
+		},
+	})

 	// 第一次执行：失败（上下文溢出），留下毒化会话。
 	res1, err := b.Execute(context.Background(), execReq(dir, []string{"output.md"}))
@@ -87,7 +92,12 @@ func TestSessionRotationAfterPoisonedFailure(t *testing.T) {

 func TestNoRotationOnRetryableNetworkFailure(t *testing.T) {
 	dir := t.TempDir()
-	b := New(Options{Model: "fake", Stream: failThenSucceed("connection reset by peer (provider_network)")})
+	b := New(Options{
+		ModelOptions: ModelOptions{
+			Model:  "fake",
+			Stream: failThenSucceed("connection reset by peer (provider_network)"),
+		},
+	})
 	_, _ = b.Execute(context.Background(), execReq(dir, []string{"output.md"}))
 	transcript := filepath.Join(dir, ".agentx", "transcript.jsonl")
 	_ = os.MkdirAll(filepath.Dir(transcript), 0o755)
@@ -3,28 +3,28 @@ package catalogcheck
 import (
 	"testing"

-	"agent-native/contract"
+	"agent-native/exec"
 	"error-code-system/errcode"
 )

 // 常量取值必须与错误码库种子码相同。旧短字符串只作为错误码库里的别名存在。
 func TestContractCodesMatchCatalog(t *testing.T) {
 	pairs := []struct {
-		got  contract.ErrorCode
+		got  exec.ErrorCode
 		want errcode.Code
 	}{
-		{contract.ErrCanceled, errcode.AgentCanceled},
-		{contract.ErrDeadlineExceeded, errcode.AgentDeadlineExceeded},
-		{contract.ErrBudgetExceeded, errcode.AgentBudgetExceeded},
-		{contract.ErrProviderUnavailable, errcode.AgentProviderUnavailable},
-		{contract.ErrProviderRejected, errcode.AgentProviderRejected},
-		{contract.ErrToolRejected, errcode.AgentToolRejected},
-		{contract.ErrPathOutsideWorkspace, errcode.AgentPathOutsideWorkspace},
-		{contract.ErrMissingCapability, errcode.AgentMissingCapability},
-		{contract.ErrMissingDeliverable, errcode.AgentMissingDeliverable},
-		{contract.ErrEmptyOutput, errcode.AgentEmptyOutput},
-		{contract.ErrUnknownOutcome, errcode.AgentUnknownOutcome},
-		{contract.ErrInternal, errcode.AgentInternal},
+		{exec.ErrCanceled, errcode.AgentCanceled},
+		{exec.ErrDeadlineExceeded, errcode.AgentDeadlineExceeded},
+		{exec.ErrBudgetExceeded, errcode.AgentBudgetExceeded},
+		{exec.ErrProviderUnavailable, errcode.AgentProviderUnavailable},
+		{exec.ErrProviderRejected, errcode.AgentProviderRejected},
+		{exec.ErrToolRejected, errcode.AgentToolRejected},
+		{exec.ErrPathOutsideWorkspace, errcode.AgentPathOutsideWorkspace},
+		{exec.ErrMissingCapability, errcode.AgentMissingCapability},
+		{exec.ErrMissingDeliverable, errcode.AgentMissingDeliverable},
+		{exec.ErrEmptyOutput, errcode.AgentEmptyOutput},
+		{exec.ErrUnknownOutcome, errcode.AgentUnknownOutcome},
+		{exec.ErrInternal, errcode.AgentInternal},
 	}
 	for _, pair := range pairs {
 		if string(pair.got) != string(pair.want) {
@@ -0,0 +1,127 @@
+package main
+
+import (
+	"bufio"
+	"context"
+	"encoding/json"
+	"flag"
+	"fmt"
+	"io"
+	"os"
+	"path/filepath"
+
+	"agent-native/rpc"
+	"agent-native/runtime"
+	"agent-native/runtime/llm"
+	"agent-native/runtime/session"
+)
+
+func main() {
+	mode := flag.String("mode", "print", "print, json, or rpc")
+	model := flag.String("model", "", "provider/model[:thinking]")
+	prompt := flag.String("prompt", "", "user message for print and json modes")
+	sessionPath := flag.String("session", "", "transcript path")
+	noSession := flag.Bool("no-session", false, "do not read or write a transcript")
+	export := flag.String("export", "", "write the active session as HTML and exit")
+	flag.Parse()
+
+	if *sessionPath == "" && !*noSession {
+		home, _ := os.UserHomeDir()
+		*sessionPath = filepath.Join(home, ".agent-native", "session.jsonl")
+	}
+	var store *session.Store
+	if !*noSession {
+		st, err := session.Open(*sessionPath)
+		if err != nil {
+			fmt.Fprintln(os.Stderr, err)
+			os.Exit(1)
+		}
+		store = st
+	}
+	if *export != "" {
+		if store == nil {
+			fmt.Fprintln(os.Stderr, "export requires a session")
+			os.Exit(1)
+		}
+		if err := store.ExportHTML(*export); err != nil {
+			fmt.Fprintln(os.Stderr, err)
+			os.Exit(1)
+		}
+		return
+	}
+
+	ref, err := llm.ParseModelRef(*model)
+	if *model != "" && err != nil {
+		fmt.Fprintln(os.Stderr, err)
+		os.Exit(1)
+	}
+	agent := runtime.NewAgentSession(runtime.Config{
+		Model:         ref.Model,
+		ThinkingLevel: ref.Thinking,
+		Session:       store,
+		Stream:        liveStream(ref),
+		OnEvent: func(ev runtime.Event) {
+			if *mode != "json" {
+				return
+			}
+			raw, _ := json.Marshal(map[string]any{"type": ev.Type, "text": ev.Text, "tool": ev.Tool})
+			fmt.Println(string(raw))
+		},
+	})
+	srv := &rpc.Server{Agent: agent, Store: store, NoStore: *noSession}
+	ctx := context.Background()
+	switch *mode {
+	case "rpc":
+		in := bufio.NewScanner(os.Stdin)
+		in.Buffer(make([]byte, 0, 64*1024), 8*1024*1024)
+		for in.Scan() {
+			resp, err := srv.Handle(ctx, in.Bytes())
+			if err != nil {
+				fmt.Fprintln(os.Stderr, err)
+				os.Exit(1)
+			}
+			os.Stdout.Write(resp)
+		}
+	case "print", "json":
+		if *prompt == "" {
+			*prompt = readAll(os.Stdin)
+		}
+		res, err := agent.Prompt(ctx, llm.Message{Role: llm.RoleUser, Content: *prompt})
+		if err != nil {
+			fmt.Fprintln(os.Stderr, err)
+			os.Exit(1)
+		}
+		if res.Err != nil {
+			fmt.Fprintln(os.Stderr, res.Err)
+			os.Exit(1)
+		}
+		if *mode == "print" {
+			fmt.Println(res.Text)
+		}
+	default:
+		fmt.Fprintln(os.Stderr, "unknown mode")
+		os.Exit(1)
+	}
+}
+
+func liveStream(ref llm.ModelRef) llm.StreamFunc {
+	provider, ok := llm.Lookup(ref.Provider, nil)
+	if !ok {
+		provider = llm.Provider{Protocol: llm.ProtocolOpenAIChat, BaseURL: os.Getenv("OPENAI_BASE_URL"), APIKey: os.Getenv("OPENAI_API_KEY")}
+	}
+	client, err := llm.ClientFor(provider)
+	if err != nil {
+		return func(ctx context.Context, model string, messages []llm.Message, tools []llm.ToolDecl, opts llm.StreamOptions) <-chan llm.Event {
+			ch := make(chan llm.Event, 1)
+			ch <- llm.Event{Kind: llm.KindError, Err: err, Message: llm.Message{StopReason: llm.StopError}}
+			close(ch)
+			return ch
+		}
+	}
+	return llm.WithRetry(client.Stream, 2, nil)
+}
+
+func readAll(r io.Reader) string {
+	raw, _ := io.ReadAll(r)
+	return string(raw)
+}
@@ -11,27 +11,27 @@ import (
 	"testing"
 	"time"

-	"agent-native/contract"
+	"agent-native/exec"
 )

 // FakeBackend 是确定性假实现，仅用于契约套件自测与消费方单测。
 // 不模拟任何模型行为；按 Script 逐项返回。
 type FakeBackend struct {
-	Desc contract.BackendDescriptor
+	Desc exec.BackendDescriptor
 	// Result 为每次 Execute 返回的终态（浅拷贝后返回）。
-	Result contract.Result
+	Result exec.ContractResult
 	// BlockUntil 若非 nil，Execute 阻塞直到该 channel 关闭或 ctx 取消，
 	// 用于验证取消语义。
 	BlockUntil <-chan struct{}
 	// CanceledWithCtx 记录 Execute 返回时 ctx 是否已取消。
 	CanceledWithCtx bool
 	// LastRequest 记录最近一次请求。
-	LastRequest contract.Request
+	LastRequest exec.ContractRequest
 }

-func (f *FakeBackend) Descriptor() contract.BackendDescriptor { return f.Desc }
+func (f *FakeBackend) Descriptor() exec.BackendDescriptor { return f.Desc }

-func (f *FakeBackend) Execute(ctx context.Context, req contract.Request) (*contract.Result, error) {
+func (f *FakeBackend) Execute(ctx context.Context, req exec.ContractRequest) (*exec.ContractResult, error) {
 	f.LastRequest = req
 	if f.BlockUntil != nil {
 		select {
@@ -39,10 +39,10 @@ func (f *FakeBackend) Execute(ctx context.Context, req contract.Request) (*contr
 		case <-ctx.Done():
 			f.CanceledWithCtx = true
 			res := *(&f.Result)
-			res.Outcome = contract.OutcomeCanceled
-			res.Err = &contract.Error{
-				Code:        contract.ErrCanceled,
-				SideEffects: contract.SideEffectNone,
+			res.Outcome = exec.OutcomeCanceled
+			res.Err = &exec.Error{
+				Code:        exec.ErrCanceled,
+				SideEffects: exec.SideEffectNone,
 			}
 			return &res, nil
 		}
@@ -55,15 +55,15 @@ func (f *FakeBackend) Execute(ctx context.Context, req contract.Request) (*contr
 type Case struct {
 	Name string
 	// Backend 为被测实现。
-	Backend contract.Backend
+	Backend exec.ContractBackend
 	// Request 为请求模板。
-	Request contract.Request
+	Request exec.ContractRequest
 	// CancelAfter 若非零，在等待该时长后取消 ctx（验证取消路径）。
 	CancelAfter time.Duration
 	// WantOutcome 为期望终态。
-	WantOutcome contract.Outcome
+	WantOutcome exec.Outcome
 	// WantErrCode 为期望错误码（Failed/Canceled/Unknown 时）。
-	WantErrCode contract.ErrorCode
+	WantErrCode exec.ErrorCode
 }

 // Run 执行一条契约用例。
@@ -96,23 +96,23 @@ func Run(t *testing.T, c Case) {
 			t.Fatalf("[%s] Err.SideEffects must be explicit (none|possible), got empty", c.Name)
 		}
 	}
-	if res.Outcome == contract.OutcomeSucceeded && res.Err != nil {
+	if res.Outcome == exec.OutcomeSucceeded && res.Err != nil {
 		t.Fatalf("[%s] Succeeded result must not carry Err", c.Name)
 	}
 }

 // Suite 是最小契约套件：成功、失败带副作用标记、取消三态。
 // 每个真实 Backend 实现都必须通过；套件外按实现能力补充专项用例。
-func Suite(t *testing.T, mk func() contract.Backend) {
+func Suite(t *testing.T, mk func() exec.ContractBackend) {
 	t.Helper()

 	t.Run("succeeded_minimal", func(t *testing.T) {
 		b := mk()
 		Run(t, Case{
-			Name:     "succeeded",
-			Backend:  b,
-			Request:  contract.Request{ExecutionID: "e1", Prompt: "hi"},
-			WantOutcome: contract.OutcomeSucceeded,
+			Name:        "succeeded",
+			Backend:     b,
+			Request:     exec.ContractRequest{ExecutionID: "e1", Prompt: "hi"},
+			WantOutcome: exec.OutcomeSucceeded,
 		})
 	})

@@ -130,13 +130,13 @@ func Suite(t *testing.T, mk func() contract.Backend) {
 	t.Run("error_sideeffects_explicit", func(t *testing.T) {
 		b := mk()
 		// 通过描述能力判定是否有失败路径可测；无失败路径的实现跳过。
-		if !b.Descriptor().Has(contract.CapText) && !b.Descriptor().Has(contract.CapFiles) {
+		if !b.Descriptor().Has(exec.CapText) && !b.Descriptor().Has(exec.CapFiles) {
 			t.Skip("no text/files capability to force failure path")
 		}
 		// 空提示在多数实现下是确定性失败或成功空输出；仅验证当 Outcome=Failed
 		// 时错误字段完整。具体实现应补充自身失败注入用例。
-		res, _ := b.Execute(context.Background(), contract.Request{ExecutionID: "e2"})
-		if res != nil && res.Outcome == contract.OutcomeFailed {
+		res, _ := b.Execute(context.Background(), exec.ContractRequest{ExecutionID: "e2"})
+		if res != nil && res.Outcome == exec.OutcomeFailed {
 			if res.Err == nil || res.Err.Code == "" || res.Err.SideEffects == "" {
 				t.Fatalf("failed result must carry Code and SideEffects, got %+v", res.Err)
 			}
@@ -0,0 +1,41 @@
+package conformance
+
+import (
+	"testing"
+
+	"agent-native/exec"
+)
+
+func TestFakeBackendSuite(t *testing.T) {
+	Suite(t, func() exec.ContractBackend {
+		return &FakeBackend{
+			Desc: exec.BackendDescriptor{
+				Name:         "fake",
+				Capabilities: []exec.Capability{exec.CapText},
+			},
+			Result: exec.ContractResult{
+				Outcome: exec.OutcomeSucceeded,
+				Text:    "ok",
+			},
+		}
+	})
+}
+
+func TestFakeCancelSemantics(t *testing.T) {
+	block := make(chan struct{})
+	b := &FakeBackend{
+		Desc:       exec.BackendDescriptor{Name: "fake", Capabilities: []exec.Capability{exec.CapCancel}},
+		BlockUntil: block,
+	}
+	Run(t, Case{
+		Name:        "cancel",
+		Backend:     b,
+		Request:     exec.ContractRequest{ExecutionID: "e-cancel"},
+		CancelAfter: 50_000_000,
+		WantOutcome: exec.OutcomeCanceled,
+		WantErrCode: exec.ErrCanceled,
+	})
+	if !b.CanceledWithCtx {
+		t.Fatal("backend must observe ctx cancellation before returning")
+	}
+}
@@ -1,63 +0,0 @@
-package contract_test
-
-import (
-	"testing"
-
-	"agent-native/conformance"
-	"agent-native/contract"
-)
-
-// 契约自测：fake 后端通过最小套件，验证套件本身与类型语义一致。
-func TestFakeBackendSuite(t *testing.T) {
-	conformance.Suite(t, func() contract.Backend {
-		return &conformance.FakeBackend{
-			Desc: contract.BackendDescriptor{
-				Name:         "fake",
-				Capabilities: []contract.Capability{contract.CapText},
-			},
-			Result: contract.Result{
-				Outcome: contract.OutcomeSucceeded,
-				Text:    "ok",
-			},
-		}
-	})
-}
-
-func TestFakeCancelSemantics(t *testing.T) {
-	block := make(chan struct{})
-	b := &conformance.FakeBackend{
-		Desc:       contract.BackendDescriptor{Name: "fake", Capabilities: []contract.Capability{contract.CapCancel}},
-		BlockUntil: block,
-	}
-	conformance.Run(t, conformance.Case{
-		Name:        "cancel",
-		Backend:     b,
-		Request:     contract.Request{ExecutionID: "e-cancel"},
-		CancelAfter: 50_000_000, // 50ms in ns (avoid time import here)
-		WantOutcome: contract.OutcomeCanceled,
-		WantErrCode: contract.ErrCanceled,
-	})
-	if !b.CanceledWithCtx {
-		t.Fatal("backend must observe ctx cancellation before returning")
-	}
-}
-
-func TestErrorDuckTypeForLogs(t *testing.T) {
-	e := &contract.Error{
-		Code:        contract.ErrProviderUnavailable,
-		Reason:      "down",
-		Retryable:   true,
-		SideEffects: contract.SideEffectNone,
-	}
-	if e.ErrorCode() != "agent.provider_unavailable" {
-		t.Fatal(e.ErrorCode())
-	}
-	fields := e.ErrorFields()
-	if fields["reason"] != "down" || fields["side_effects"] != "none" || fields["retryable"] != true {
-		t.Fatalf("%v", fields)
-	}
-	var nilErr *contract.Error
-	if nilErr.ErrorCode() != "" || nilErr.ErrorFields() != nil {
-		t.Fatal("nil error")
-	}
-}
@@ -89,3 +89,19 @@ baseline_fixed: true

 安全声明（不冒充已解决）：run_command **不是 OS 沙箱**（宿主负责隔离）；技能正文与 MCP
 server 描述是提示注入面（只信受信根 + 长度上限）；fetch 的 DNS 解析存在 rebinding 窗口。
+扩展进程、安装的包和外部 CLI 都以调用者的完整权限运行，同样不是沙箱。
+
+## R16 无界面全量（pi-headless，2026-10-01）
+
+对标口径改为 pi 的无界面能力。不做 TUI、主题、快捷键，也不做订阅制 OAuth。
+
+| 域 | 现状 |
+|---|---|
+| 会话树 | available：id/parentId、fork/clone、叶子切换、continue/resume、追加式压缩、旧线性文件仍可 Load、export_html |
+| 供应商 | available：Anthropic、OpenAI Chat 与 Responses、Google、Azure、兼容预设、models.json、`provider/id:thinking`、图片、缓存 token、自动重试；上下文溢出不重试 |
+| AgentSession | available：steer/follow-up、one 与 all、abort、clear_queue、中途换模型；事件含 message_start、tool_execution_update、queue_update、agent_settled |
+| 资源 | available：pi 工具名与原名并存、上下文文件、提示词模板、`/skill:name`、项目信任 |
+| 扩展与包 | available：stdio JSONL 子进程；install/remove/update/list/config |
+| 模式 | available：`cmd/agent-native` 的 print/json/rpc，以及 Go `rpc.Client` |
+| exec.Backend | available：`adapters/native` 的 Execute 经 AgentSession，请求/结果字段不变 |
+
@@ -16,7 +16,7 @@ type Request struct {
 	// OnEvent 可选过程事件回调（nil = 调用方不需要过程，后端行为与现状完全一致）。
 	// 实现方必须保证：回调在 Execute 返回前完成（同步调用或等待收尾），无泄漏 goroutine；
 	// 回调内不得阻塞（消费方自行节流/异步入队）。
-	OnEvent func(ev ProcessEvent)
+	OnEvent func(ev ProcessEvent) `json:"-"`

 	// ---- v0.2 能力字段（全部可选；零值 = 既有行为，向后兼容） ----

@@ -1,7 +1,8 @@
-// Package contract 定义 Agent-Native 的公共执行契约。
+// 原 contract 包并入 exec。与本文件同名的执行门面仍是 Request、Result、Backend；
+// 设计契约里冲突的四个名字加 Contract 前缀。
 //
 // 本包是助手（Core）与小队（Squad）共用的唯一执行接口层：Core 注入具体
-// Backend 实现，Squad 只表达需求并使用已注入的后端。契约不依赖任何具体
+// ContractBackend 实现，Squad 只表达需求并使用已注入的后端。契约不依赖任何具体
 // 内核、产品类型、数据库或配置包。
 //
 // 语义不变量（实现方与消费方都必须遵守）：
@@ -20,7 +21,7 @@
 //     上下文，不授予跨会话可见性。
 //  7. Workspace 的路径约束不是 OS 沙箱；执行不可信第三方工具需要调用方
 //     提供的进程/容器隔离，不能由本契约推导。
-package contract
+package exec

 import (
 	"context"
@@ -50,12 +51,6 @@ const (
 	CapStructuredOutput Capability = "structured-output"
 )

-// ToolProfile 是预定义工具策略名；与现有产品配置兼容。
-const (
-	ToolProfileWorkspace = "workspace" // 默认：文档类工具受限集合
-	ToolProfileReasoning = "reasoning" // 无工具，仅推理
-)
-
 // ToolPolicy 描述执行允许的工具范围。实现不得自行扩展；未知工具一律拒绝。
 type ToolPolicy struct {
 	// Profile 为策略名（workspace/reasoning/或产品注册的自定义名）。
@@ -82,8 +77,8 @@ type SessionScope struct {
 	Session string `json:"session,omitempty"`
 }

-// Budget 约束一次执行（含共享 ExecutionID 的重试）。
-type Budget struct {
+// ContractBudget 约束一次执行（含共享 ExecutionID 的重试）。
+type ContractBudget struct {
 	// MaxPromptTokens/MaxCompletionTokens 为 0 时不限制。
 	MaxPromptTokens     int `json:"max_prompt_tokens,omitempty"`
 	MaxCompletionTokens int `json:"max_completion_tokens,omitempty"`
@@ -98,9 +93,9 @@ type Message struct {
 	Content string `json:"content"`
 }

-// Request 是一次授权执行的统一输入。字段为设计契约；现有 file-only 流程
+// ContractRequest 是一次授权执行的统一输入。字段为设计契约；现有 file-only 流程
 // （Prompt+Instructions+Targets→reply.md）是其子集 facade。
-type Request struct {
+type ContractRequest struct {
 	// ExecutionID 由调用方生成，重试与受控 fallback 复用同值。
 	ExecutionID string `json:"execution_id"`
 	// SessionScope 为隔离身份；可为零值（一次性任务）。
@@ -118,8 +113,8 @@ type Request struct {
 	Workspace WorkspaceHandle `json:"workspace"`
 	// ToolPolicy 为工具范围；空 Profile 等同 reasoning。
 	ToolPolicy ToolPolicy `json:"tool_policy"`
-	// Budget 为本次执行预算。
-	Budget Budget `json:"budget"`
+	// ContractBudget 为本次执行预算。
+	ContractBudget ContractBudget `json:"budget"`
 	// Required 为必须满足的能力；解析器据此选择实现，缺失即拒绝。
 	Required []Capability `json:"required,omitempty"`

@@ -235,14 +230,14 @@ type Usage struct {
 	CompletionTokens int `json:"completion_tokens,omitempty"`
 }

-// Result 是执行终态。Text 与 Files 依据任务可为空；Outcome 决定解释。
-type Result struct {
+// ContractResult 是执行终态。Text 与 Files 依据任务可为空；Outcome 决定解释。
+type ContractResult struct {
 	Outcome Outcome `json:"outcome"`
 	// Text 为文本结果（CapText 任务必填）。
 	Text string `json:"text,omitempty"`
 	// Files 为文件候选产物（内容字节），由调用方决定正式落盘与提交。
 	Files map[string][]byte `json:"files,omitempty"`
-	// Logs 为执行日志/过程摘要（兼容现有 Result.Logs）。
+	// Logs 为执行日志/过程摘要（兼容现有 ContractResult.Logs）。
 	Logs string `json:"logs,omitempty"`
 	// Err 为 Failed/Canceled/Unknown 时的标准化错误；Succeeded 时为 nil。
 	Err *Error `json:"err,omitempty"`
@@ -271,19 +266,19 @@ func (d BackendDescriptor) Has(c Capability) bool {
 	return false
 }

-// Backend 是统一执行接口。实现必须：
+// ContractBackend 是统一执行接口。实现必须：
 //   - 在 ctx 取消时尽快返回（Canceled 或 Unknown）；
 //   - 不持有跨执行的可变会话状态（连接池等无状态资源可复用）；
 //   - 不自行扩大工具范围或读取 Workspace 之外的路径。
-type Backend interface {
+type ContractBackend interface {
 	Descriptor() BackendDescriptor
-	// Execute 执行一次授权调用。返回的 Result 永不为 nil；ctx 错误映射为
+	// Execute 执行一次授权调用。返回的 ContractResult 永不为 nil；ctx 错误映射为
 	// Outcome=Canceled/Unknown，不返回 Go error 表达业务终态。
-	Execute(ctx context.Context, req Request) (*Result, error)
+	Execute(ctx context.Context, req ContractRequest) (*ContractResult, error)
 }

 // Resolver 按名称与必需能力解析后端。未知名必须返回错误，生产路径不得
 // 回落到 Stub。
 type Resolver interface {
-	Resolve(name string, required []Capability) (Backend, error)
+	Resolve(name string, required []Capability) (ContractBackend, error)
 }
@@ -0,0 +1,23 @@
+package exec
+
+import "testing"
+
+func TestErrorDuckTypeForLogs(t *testing.T) {
+	e := &Error{
+		Code:        ErrProviderUnavailable,
+		Reason:      "down",
+		Retryable:   true,
+		SideEffects: SideEffectNone,
+	}
+	if e.ErrorCode() != "agent.provider_unavailable" {
+		t.Fatal(e.ErrorCode())
+	}
+	fields := e.ErrorFields()
+	if fields["reason"] != "down" || fields["side_effects"] != "none" || fields["retryable"] != true {
+		t.Fatalf("%v", fields)
+	}
+	var nilErr *Error
+	if nilErr.ErrorCode() != "" || nilErr.ErrorFields() != nil {
+		t.Fatal("nil error")
+	}
+}
@@ -3,13 +3,10 @@ module agent-native
 go 1.27

 require (
-	agent-intent v0.0.0
 	error-code-system v0.0.0
 	logging-system v0.0.0
 )

-replace agent-intent => ./deps/agent-intent
-
 replace error-code-system => ./deps/error-code-system

 replace logging-system => ./deps/logging-system
@@ -0,0 +1,54 @@
+package rpc
+
+import (
+	"bufio"
+	"context"
+	"encoding/json"
+	"io"
+)
+
+// Client is the Go SDK for the JSONL RPC mode.
+type Client struct {
+	in  io.Writer
+	out *bufio.Reader
+}
+
+// NewClient wraps a bidirectional pipe. Commands are one JSON object per line.
+func NewClient(in io.Writer, out io.Reader) *Client {
+	return &Client{in: in, out: bufio.NewReader(out)}
+}
+
+// Call sends one command and reads the matching response.
+func (c *Client) Call(ctx context.Context, req Request) (Response, error) {
+	raw, err := json.Marshal(req)
+	if err != nil {
+		return Response{}, err
+	}
+	if _, err := c.in.Write(append(raw, '\n')); err != nil {
+		return Response{}, err
+	}
+	type result struct {
+		resp Response
+		err  error
+	}
+	ch := make(chan result, 1)
+	go func() {
+		line, err := c.out.ReadBytes('\n')
+		if err != nil {
+			ch <- result{err: err}
+			return
+		}
+		var resp Response
+		if err := json.Unmarshal(line, &resp); err != nil {
+			ch <- result{err: err}
+			return
+		}
+		ch <- result{resp: resp}
+	}()
+	select {
+	case <-ctx.Done():
+		return Response{}, ctx.Err()
+	case got := <-ch:
+		return got.resp, got.err
+	}
+}
@@ -0,0 +1,213 @@
+package rpc
+
+import (
+	"context"
+	"encoding/json"
+	"fmt"
+
+	"agent-native/runtime"
+	"agent-native/runtime/llm"
+	"agent-native/runtime/session"
+)
+
+// Request is one RPC command. Type matches the pi headless command set.
+type Request struct {
+	ID            string `json:"id,omitempty"`
+	Type          string `json:"type"`
+	Message       string `json:"message,omitempty"`
+	Model         string `json:"model,omitempty"`
+	ThinkingLevel string `json:"thinkingLevel,omitempty"`
+	Path          string `json:"path,omitempty"`
+	Name          string `json:"name,omitempty"`
+	Leaf          string `json:"leaf,omitempty"`
+	Keep          int    `json:"keep,omitempty"`
+}
+
+// Response is the synchronous result of one command.
+type Response struct {
+	ID     string `json:"id,omitempty"`
+	Type   string `json:"type"`
+	OK     bool   `json:"ok"`
+	Text   string `json:"text,omitempty"`
+	Error  string `json:"error,omitempty"`
+	Result any    `json:"result,omitempty"`
+	// Status 是命令回执：started 已接手，queued 先排队，handled 已处理完。
+	Status string `json:"status,omitempty"`
+}
+
+// Server serves RPC commands against one agent session and its transcript.
+type Server struct {
+	Agent   *runtime.AgentSession
+	Store   *session.Store
+	NoStore bool
+}
+
+// Handle runs one command and returns the JSON response line.
+func (s *Server) Handle(ctx context.Context, line []byte) ([]byte, error) {
+	var req Request
+	if err := json.Unmarshal(line, &req); err != nil {
+		return encode(Response{Type: "response", OK: false, Error: err.Error()})
+	}
+	resp := s.dispatch(ctx, req)
+	resp.ID = req.ID
+	if resp.Type == "" {
+		resp.Type = "response"
+	}
+	return encode(resp)
+}
+
+func (s *Server) dispatch(ctx context.Context, req Request) Response {
+	switch req.Type {
+	case "prompt":
+		if s.Agent == nil {
+			return fail("no agent")
+		}
+		msg := llm.Message{Role: llm.RoleUser, Content: req.Message}
+		if s.Agent.Busy() {
+			s.Agent.FollowUp(msg)
+			return Response{OK: true, Status: "queued"}
+		}
+		res, err := s.Agent.Prompt(ctx, msg)
+		if err != nil {
+			return fail(err.Error())
+		}
+		if res.Err != nil {
+			return fail(res.Err.Error())
+		}
+		return Response{OK: true, Text: res.Text, Status: "handled"}
+	case "steer":
+		if s.Agent == nil {
+			return fail("no agent")
+		}
+		s.Agent.Steer(llm.Message{Role: llm.RoleUser, Content: req.Message})
+		return Response{OK: true, Status: queueStatus(s.Agent.Busy())}
+	case "follow_up":
+		if s.Agent == nil {
+			return fail("no agent")
+		}
+		s.Agent.FollowUp(llm.Message{Role: llm.RoleUser, Content: req.Message})
+		return Response{OK: true, Status: queueStatus(s.Agent.Busy())}
+	case "abort":
+		s.Agent.Abort()
+		return Response{OK: true}
+	case "clear_queue":
+		s.Agent.ClearQueue()
+		return Response{OK: true}
+	case "set_model":
+		if err := s.Agent.SetModel(req.Model, req.ThinkingLevel); err != nil {
+			return fail(err.Error())
+		}
+		return Response{OK: true}
+	case "compact":
+		keep := req.Keep
+		if keep <= 0 {
+			keep = 4
+		}
+		done, err := s.Agent.Compact(keep, func(messages []llm.Message) (string, error) {
+			return fmt.Sprintf("%d messages", len(messages)), nil
+		})
+		if err != nil {
+			return fail(err.Error())
+		}
+		return Response{OK: true, Result: done}
+	case "get_state":
+		leaf, _ := s.Store.Leaf()
+		return Response{OK: true, Result: map[string]any{"leaf": leaf, "name": s.Store.Name(), "path": s.Store.Path()}}
+	case "get_tree":
+		entries, err := s.Store.Entries()
+		if err != nil {
+			return fail(err.Error())
+		}
+		return Response{OK: true, Result: entries}
+	case "get_entries":
+		entries, err := s.Store.Entries()
+		if err != nil {
+			return fail(err.Error())
+		}
+		return Response{OK: true, Result: entries}
+	case "get_last_assistant_text":
+		msgs, err := s.Store.Load()
+		if err != nil {
+			return fail(err.Error())
+		}
+		text := ""
+		for _, m := range msgs {
+			if m.Role == llm.RoleAssistant {
+				text = m.Content
+			}
+		}
+		return Response{OK: true, Text: text}
+	case "set_session_name":
+		if err := s.Store.SetName(req.Name); err != nil {
+			return fail(err.Error())
+		}
+		return Response{OK: true}
+	case "fork":
+		dest := req.Path
+		if dest == "" {
+			return fail("path is required")
+		}
+		if _, err := s.Store.Fork(dest, req.Leaf); err != nil {
+			return fail(err.Error())
+		}
+		return Response{OK: true, Text: dest}
+	case "clone":
+		if req.Path == "" {
+			return fail("path is required")
+		}
+		if _, err := s.Store.Clone(req.Path); err != nil {
+			return fail(err.Error())
+		}
+		return Response{OK: true, Text: req.Path}
+	case "switch_session":
+		if req.Path == "" {
+			return fail("path is required")
+		}
+		st, err := session.Open(req.Path)
+		if err != nil {
+			return fail(err.Error())
+		}
+		s.Store = st
+		return Response{OK: true}
+	case "new_session":
+		if req.Path == "" {
+			return fail("path is required")
+		}
+		st, err := session.Open(req.Path)
+		if err != nil {
+			return fail(err.Error())
+		}
+		s.Store = st
+		return Response{OK: true, Text: req.Path}
+	case "export_html":
+		if req.Path == "" {
+			return fail("path is required")
+		}
+		if err := s.Store.ExportHTML(req.Path); err != nil {
+			return fail(err.Error())
+		}
+		return Response{OK: true, Text: req.Path}
+	case "get_session_stats":
+		st := s.Store.Stats()
+		return Response{OK: true, Result: st}
+	default:
+		return fail("unknown command " + req.Type)
+	}
+}
+
+func fail(msg string) Response { return Response{OK: false, Error: msg} }
+
+func queueStatus(busy bool) string {
+	if busy {
+		return "queued"
+	}
+	return "started"
+}
+
+func encode(resp Response) ([]byte, error) {
+	raw, err := json.Marshal(resp)
+	if err != nil {
+		return nil, err
+	}
+	return append(raw, '\n'), nil
+}
@@ -0,0 +1,75 @@
+package rpc
+
+import (
+	"context"
+	"io"
+	"path/filepath"
+	"testing"
+
+	"agent-native/runtime"
+	"agent-native/runtime/llm"
+	"agent-native/runtime/session"
+)
+
+func TestRPCSessionCommandsAndClient(t *testing.T) {
+	dir := t.TempDir()
+	store, err := session.Open(filepath.Join(dir, "s.jsonl"))
+	if err != nil {
+		t.Fatal(err)
+	}
+	if err := store.Append(llm.Message{Role: llm.RoleAssistant, Content: "ready"}); err != nil {
+		t.Fatal(err)
+	}
+	agent := runtime.NewAgentSession(runtime.Config{
+		Model: "fake",
+		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: "echo", StopReason: llm.StopEnd}}
+			close(ch)
+			return ch
+		},
+	})
+	srv := &Server{Agent: agent, Store: store}
+	clientIn, serverOut := io.Pipe()
+	serverIn, clientOut := io.Pipe()
+	go func() {
+		defer serverOut.Close()
+		buf := make([]byte, 4096)
+		for {
+			n, err := serverIn.Read(buf)
+			if n > 0 {
+				resp, herr := srv.Handle(context.Background(), buf[:n])
+				if herr == nil {
+					_, _ = serverOut.Write(resp)
+				}
+			}
+			if err != nil {
+				return
+			}
+		}
+	}()
+	client := NewClient(clientOut, clientIn)
+	resp, err := client.Call(context.Background(), Request{ID: "1", Type: "get_last_assistant_text"})
+	if err != nil || !resp.OK || resp.Text != "ready" {
+		t.Fatalf("%+v %v", resp, err)
+	}
+	fork := filepath.Join(dir, "fork.jsonl")
+	resp, err = client.Call(context.Background(), Request{Type: "fork", Path: fork})
+	if err != nil || !resp.OK {
+		t.Fatalf("fork %+v %v", resp, err)
+	}
+	resp, err = client.Call(context.Background(), Request{Type: "prompt", Message: "hi"})
+	if err != nil || !resp.OK || resp.Text != "echo" || resp.Status != "handled" {
+		t.Fatalf("prompt %+v %v", resp, err)
+	}
+	resp, err = client.Call(context.Background(), Request{Type: "steer", Message: "later"})
+	if err != nil || !resp.OK || resp.Status != "started" {
+		t.Fatalf("steer %+v %v", resp, err)
+	}
+	htmlPath := filepath.Join(dir, "out.html")
+	resp, err = client.Call(context.Background(), Request{Type: "export_html", Path: htmlPath})
+	if err != nil || !resp.OK {
+		t.Fatalf("export %+v %v", resp, err)
+	}
+	_ = clientOut.Close()
+}
@@ -0,0 +1,95 @@
+package runtime
+
+import (
+	"context"
+	"errors"
+	"testing"
+
+	"agent-native/runtime/hooks"
+	"agent-native/runtime/llm"
+	"agent-native/runtime/tools"
+)
+
+func TestHookAdaptersMatchDirectCallbacks(t *testing.T) {
+	direct := runGate(t, false)
+	adapted := runGate(t, true)
+	if direct.stop != adapted.stop || direct.text != adapted.text {
+		t.Fatalf("stop/text direct=%s %q adapted=%s %q", direct.stop, direct.text, adapted.stop, adapted.text)
+	}
+	if direct.denies != adapted.denies || direct.reminders != adapted.reminders || direct.errors != adapted.errors {
+		t.Fatalf("counts direct=%+v adapted=%+v", direct, adapted)
+	}
+	if direct.denies != 1 || direct.reminders == 0 {
+		t.Fatalf("unexpected counts %+v", direct)
+	}
+}
+
+type gateStats struct {
+	stop, text                string
+	denies, reminders, errors int
+}
+
+func runGate(t *testing.T, viaHooks bool) gateStats {
+	t.Helper()
+	var stats gateStats
+	before := func(call llm.ToolCall) error {
+		if call.Name == "read_file" {
+			stats.denies++
+			return errors.New("denied")
+		}
+		return nil
+	}
+	reminders := func(int) []string {
+		stats.reminders++
+		return []string{"记一笔"}
+	}
+	stop := func(last llm.Message, info StopInfo) StopDecision {
+		if info.Continuations < 1 {
+			return StopDecision{Continue: true, Reason: "再看一眼"}
+		}
+		return StopDecision{}
+	}
+	cfg := Config{
+		Model: "fake",
+		Tools: tools.WorkspaceTools(t.TempDir()),
+		Stream: script(
+			llm.Message{Role: llm.RoleAssistant, StopReason: llm.StopToolCalls, ToolCalls: []llm.ToolCall{{
+				ID: "c1", Name: "read_file", Arguments: `{"path":"a.md"}`,
+			}}},
+			llm.Message{Role: llm.RoleAssistant, Content: "done", StopReason: llm.StopEnd},
+		),
+		Reminders: reminders,
+	}
+	if viaHooks {
+		bus := hooks.New()
+		bus.On(hooks.AdaptBefore(func(name, args string) error {
+			return before(llm.ToolCall{Name: name, Arguments: args})
+		}))
+		bus.On(func(c hooks.Call) hooks.Outcome {
+			if c.Event != hooks.Stop {
+				return hooks.Outcome{}
+			}
+			d := stop(llm.Message{Content: c.LastText}, StopInfo{Continuations: c.Continuations, ToolCalls: c.ToolCalls})
+			if d.Continue {
+				return hooks.Outcome{Decision: hooks.Block, Reason: d.Reason, Replacement: d.Replacement, Notice: d.Notice}
+			}
+			return hooks.Outcome{Decision: hooks.Allow, Notice: d.Notice}
+		})
+		cfg.Hooks = bus
+	} else {
+		cfg.BeforeTool = before
+		cfg.Stop = stop
+	}
+	res, err := Run(context.Background(), cfg, llm.Message{Role: llm.RoleUser, Content: "读文件"})
+	if err != nil {
+		t.Fatal(err)
+	}
+	stats.stop = string(res.StopReason)
+	stats.text = res.Text
+	for _, m := range res.Messages {
+		if m.Role == llm.RoleTool && m.IsError {
+			stats.errors++
+		}
+	}
+	return stats
+}
@@ -0,0 +1,216 @@
+// Package agentspec 定义子代理画像（AgentSpec）与 agent-*.md 文件形态。
+// 宿主与内置画像可声明全部字段；插件来源的画像只允许身份与工具面，
+// 不得声明钩子、MCP 与权限模式（fail-closed，错误定位到字段）。
+package agentspec
+
+import (
+	"fmt"
+	"os"
+	"path/filepath"
+	"strconv"
+	"strings"
+)
+
+// AgentSpec 是一个子代理画像。字段对齐公开子代理 frontmatter 约定，
+// 另加 outputSchema。空字段表示沿用宿主内置。
+type AgentSpec struct {
+	Name              string
+	Description       string
+	Prompt            string // 正文（frontmatter 之后的内容）
+	Tools             []string
+	DisallowedTools   []string
+	Model             string
+	PermissionMode    string
+	Skills            []string
+	MaxTurns          int
+	MemoryScope       string
+	Effort            string
+	Background        bool
+	WorktreeIsolation bool
+	OutputSchema      string
+}
+
+// pluginBannedKeys 是插件来源画像禁带的键：钩子、MCP 与权限模式归宿主与基座。
+var pluginBannedKeys = map[string]bool{
+	"hooks":          true,
+	"mcp":            true,
+	"permissionMode": true,
+}
+
+// Parse 解析宿主/内置画像文件：frontmatter（--- 包裹的 key: value 行，
+// 列表值用逗号分隔）+ 正文。未知键报错；name 必填且须小写连字符。
+func Parse(content string) (AgentSpec, error) {
+	return parse(content, nil)
+}
+
+// ParsePlugin 解析插件来源画像文件：在 Parse 之上禁带 hooks/mcp/permissionMode 键。
+func ParsePlugin(content string) (AgentSpec, error) {
+	return parse(content, pluginBannedKeys)
+}
+
+func parse(content string, banned map[string]bool) (AgentSpec, error) {
+	fm, body, err := splitFrontmatter(content)
+	if err != nil {
+		return AgentSpec{}, err
+	}
+	var spec AgentSpec
+	for key, value := range fm {
+		if banned[key] {
+			return AgentSpec{}, fmt.Errorf("agentspec: 插件画像不得声明 %s", key)
+		}
+		if err := assign(&spec, key, value); err != nil {
+			return AgentSpec{}, err
+		}
+	}
+	spec.Prompt = strings.TrimSpace(body)
+	if spec.Name == "" {
+		return AgentSpec{}, fmt.Errorf("agentspec: name required")
+	}
+	if !validName(spec.Name) {
+		return AgentSpec{}, fmt.Errorf("agentspec: name %q 须为小写连字符（如 doc-reviewer）", spec.Name)
+	}
+	return spec, nil
+}
+
+// LoadDir 整批加载目录下的 *.md 画像：任一文件解析失败则全批拒绝；
+// 重名拒绝。parser 按来源选 Parse 或 ParsePlugin。返回按 Name 索引。
+func LoadDir(dir string, parser func(string) (AgentSpec, error)) (map[string]AgentSpec, error) {
+	if parser == nil {
+		parser = Parse
+	}
+	names, err := filepath.Glob(filepath.Join(dir, "*.md"))
+	if err != nil {
+		return nil, fmt.Errorf("agentspec: glob %s: %w", dir, err)
+	}
+	out := map[string]AgentSpec{}
+	for _, name := range names {
+		raw, err := os.ReadFile(name)
+		if err != nil {
+			return nil, fmt.Errorf("agentspec: %s: %w", filepath.Base(name), err)
+		}
+		spec, err := parser(string(raw))
+		if err != nil {
+			return nil, fmt.Errorf("agentspec: %s: %w", filepath.Base(name), err)
+		}
+		if _, dup := out[spec.Name]; dup {
+			return nil, fmt.Errorf("agentspec: duplicate name %q in %s", spec.Name, dir)
+		}
+		out[spec.Name] = spec
+	}
+	return out, nil
+}
+
+func splitFrontmatter(content string) (map[string]string, string, error) {
+	text := strings.ReplaceAll(content, "\r\n", "\n")
+	if !strings.HasPrefix(text, "---\n") {
+		return nil, "", fmt.Errorf("agentspec: missing frontmatter（须以 --- 开头）")
+	}
+	rest := text[len("---\n"):]
+	end := strings.Index(rest, "\n---")
+	if end < 0 {
+		return nil, "", fmt.Errorf("agentspec: frontmatter 未闭合（缺 --- 结尾）")
+	}
+	fm := map[string]string{}
+	for _, line := range strings.Split(rest[:end], "\n") {
+		line = strings.TrimSpace(line)
+		if line == "" || strings.HasPrefix(line, "#") {
+			continue
+		}
+		i := strings.Index(line, ":")
+		if i <= 0 {
+			return nil, "", fmt.Errorf("agentspec: 无法解析的行 %q（须为 key: value）", line)
+		}
+		fm[strings.TrimSpace(line[:i])] = strings.TrimSpace(line[i+1:])
+	}
+	body := rest[end+len("\n---"):]
+	body = strings.TrimLeft(body, "\n")
+	return fm, body, nil
+}
+
+func assign(spec *AgentSpec, key, value string) error {
+	switch key {
+	case "name":
+		spec.Name = value
+	case "description":
+		spec.Description = value
+	case "tools":
+		spec.Tools = splitList(value)
+	case "disallowedTools":
+		spec.DisallowedTools = splitList(value)
+	case "model":
+		spec.Model = value
+	case "permissionMode":
+		spec.PermissionMode = value
+	case "skills":
+		spec.Skills = splitList(value)
+	case "maxTurns":
+		n, err := strconv.Atoi(value)
+		if err != nil || n < 0 {
+			return fmt.Errorf("agentspec: maxTurns %q 须为非负整数", value)
+		}
+		spec.MaxTurns = n
+	case "memoryScope":
+		spec.MemoryScope = value
+	case "effort":
+		spec.Effort = value
+	case "background":
+		b, err := parseBool(key, value)
+		if err != nil {
+			return err
+		}
+		spec.Background = b
+	case "worktreeIsolation":
+		b, err := parseBool(key, value)
+		if err != nil {
+			return err
+		}
+		spec.WorktreeIsolation = b
+	case "outputSchema":
+		spec.OutputSchema = value
+	default:
+		return fmt.Errorf("agentspec: 未知键 %q", key)
+	}
+	return nil
+}
+
+func parseBool(key, value string) (bool, error) {
+	switch value {
+	case "true":
+		return true, nil
+	case "false":
+		return false, nil
+	default:
+		return false, fmt.Errorf("agentspec: %s %q 须为 true/false", key, value)
+	}
+}
+
+func splitList(value string) []string {
+	if strings.TrimSpace(value) == "" {
+		return nil
+	}
+	parts := strings.Split(value, ",")
+	out := make([]string, 0, len(parts))
+	for _, p := range parts {
+		if s := strings.TrimSpace(p); s != "" {
+			out = append(out, s)
+		}
+	}
+	return out
+}
+
+func validName(name string) bool {
+	if name == "" || strings.HasPrefix(name, "-") || strings.HasSuffix(name, "-") {
+		return false
+	}
+	for _, seg := range strings.Split(name, "-") {
+		if seg == "" {
+			return false
+		}
+		for _, r := range seg {
+			if (r < 'a' || r > 'z') && (r < '0' || r > '9') {
+				return false
+			}
+		}
+	}
+	return true
+}
@@ -0,0 +1,131 @@
+package agentspec
+
+import (
+	"os"
+	"path/filepath"
+	"strings"
+	"testing"
+)
+
+const goodSpec = `---
+name: doc-reviewer
+description: 评审体系文档章节
+tools: read_file, grep, write_file
+disallowedTools: run_shell
+model: lightweight
+skills: house-style, rdm-review
+maxTurns: 8
+memoryScope: project
+effort: high
+background: true
+worktreeIsolation: false
+outputSchema: review-result
+---
+你是章节评审代理。逐条核对来源标注。
+`
+
+func TestParseFullFields(t *testing.T) {
+	spec, err := Parse(goodSpec)
+	if err != nil {
+		t.Fatal(err)
+	}
+	if spec.Name != "doc-reviewer" || spec.Description == "" {
+		t.Fatalf("identity = %+v", spec)
+	}
+	if len(spec.Tools) != 3 || spec.Tools[2] != "write_file" {
+		t.Fatalf("tools = %v", spec.Tools)
+	}
+	if len(spec.DisallowedTools) != 1 || spec.DisallowedTools[0] != "run_shell" {
+		t.Fatalf("disallowed = %v", spec.DisallowedTools)
+	}
+	if spec.Model != "lightweight" || spec.MaxTurns != 8 || spec.MemoryScope != "project" || spec.Effort != "high" {
+		t.Fatalf("scalars = %+v", spec)
+	}
+	if !spec.Background || spec.WorktreeIsolation {
+		t.Fatalf("bools = %+v", spec)
+	}
+	if len(spec.Skills) != 2 || spec.OutputSchema != "review-result" {
+		t.Fatalf("skills/schema = %+v", spec)
+	}
+	if spec.Prompt == "" {
+		t.Fatal("prompt body empty")
+	}
+}
+
+func TestParseDefaultsAndMinimal(t *testing.T) {
+	spec, err := Parse("---\nname: helper\n---\n做杂事。\n")
+	if err != nil {
+		t.Fatal(err)
+	}
+	if spec.MaxTurns != 0 || spec.Background || len(spec.Tools) != 0 || spec.PermissionMode != "" {
+		t.Fatalf("defaults = %+v", spec)
+	}
+	if spec.Prompt != "做杂事。" {
+		t.Fatalf("prompt = %q", spec.Prompt)
+	}
+}
+
+func TestParseRejectsBadInputs(t *testing.T) {
+	cases := map[string]string{
+		"missing frontmatter":   "name: x\n---\n正文\n",
+		"unclosed frontmatter":  "---\nname: x\n正文\n",
+		"unknown key":           "---\nname: x\nbogus: 1\n---\n正文\n",
+		"missing name":          "---\ndescription: x\n---\n正文\n",
+		"bad name uppercase":    "---\nname: Bad\n---\n正文\n",
+		"bad name underscore":   "---\nname: has_under\n---\n正文\n",
+		"bad maxTurns":          "---\nname: x\nmaxTurns: many\n---\n正文\n",
+		"bad bool":              "---\nname: x\nbackground: yes\n---\n正文\n",
+		"malformed line":        "---\nname: x\nno-colon-here\n---\n正文\n",
+	}
+	for label, content := range cases {
+		if _, err := Parse(content); err == nil {
+			t.Fatalf("%s: expected error", label)
+		}
+	}
+}
+
+func TestParsePluginBansHostOnlyKeys(t *testing.T) {
+	for _, key := range []string{"hooks", "mcp", "permissionMode"} {
+		content := "---\nname: plug-agent\n" + key + ": anything\n---\n正文\n"
+		if _, err := ParsePlugin(content); err == nil {
+			t.Fatalf("plugin key %s: expected rejection", key)
+		} else if got := err.Error(); !strings.Contains(got, "不得声明 "+key) {
+			t.Fatalf("plugin key %s: error %q 未定位到字段", key, got)
+		}
+	}
+	// 宿主来源不受此限：permissionMode 在 Parse 下可用。
+	if _, err := Parse("---\nname: host-agent\npermissionMode: plan\n---\n正文\n"); err != nil {
+		t.Fatalf("host parse with permissionMode: %v", err)
+	}
+	// 插件的合法字段照常解析。
+	spec, err := ParsePlugin("---\nname: plug-agent\ntools: read_file\n---\n正文\n")
+	if err != nil || len(spec.Tools) != 1 {
+		t.Fatalf("plugin good spec: %+v err=%v", spec, err)
+	}
+}
+
+func TestLoadDirAtomicAndDuplicate(t *testing.T) {
+	dir := t.TempDir()
+	write := func(name, content string) {
+		if err := os.WriteFile(filepath.Join(dir, name), []byte(content), 0o644); err != nil {
+			t.Fatal(err)
+		}
+	}
+	write("a.md", "---\nname: alpha\n---\nA\n")
+	write("b.md", "---\nname: beta\n---\nB\n")
+	got, err := LoadDir(dir, nil)
+	if err != nil || len(got) != 2 {
+		t.Fatalf("load = %d err=%v", len(got), err)
+	}
+
+	write("c.md", "---\nname: alpha\n---\nC\n")
+	if _, err := LoadDir(dir, nil); err == nil {
+		t.Fatal("duplicate name accepted")
+	}
+	os.Remove(filepath.Join(dir, "c.md"))
+
+	write("bad.md", "---\nname: x\nbogus: 1\n---\n")
+	if _, err := LoadDir(dir, nil); err == nil {
+		t.Fatal("bad file did not fail the batch")
+	}
+}
@@ -0,0 +1,242 @@
+// Package checkpoint 记录写文件之前的内容，并按用户消息把文件退回去。
+// 只追踪显式交给它的写入。shell 与 python 的改动不经过这里。
+package checkpoint
+
+import (
+	"bytes"
+	"encoding/json"
+	"os"
+	"path/filepath"
+	"sync"
+)
+
+// Record 是一次被追踪的写入。Before 是写入前的字节，After 是写入后的字节。
+type Record struct {
+	Seq     int    `json:"seq"`
+	Anchor  string `json:"anchor,omitempty"`
+	Root    string `json:"root"`
+	Path    string `json:"path"`
+	Existed bool   `json:"existed"`
+	Before  []byte `json:"before,omitempty"`
+	After   []byte `json:"after,omitempty"`
+}
+
+// Result 是一次回滚的分类。Skipped 是当前内容既不是写入后也不是写入前的文件。
+type Result struct {
+	Restored []string
+	Skipped  []string
+}
+
+// Log 是追加写的检查点文件。
+type Log struct {
+	path string
+	mu   sync.Mutex
+}
+
+func Open(path string) (*Log, error) {
+	if err := os.MkdirAll(filepath.Dir(path), 0o755); err != nil {
+		return nil, err
+	}
+	f, err := os.OpenFile(path, os.O_CREATE, 0o644)
+	if err != nil {
+		return nil, err
+	}
+	_ = f.Close()
+	return &Log{path: path}, nil
+}
+
+func (l *Log) Path() string { return l.path }
+
+// Record 追加一条检查点。rel 是相对 root 的路径。
+func (l *Log) Record(anchor, root, rel string, before []byte, existed bool, after []byte) error {
+	if l == nil {
+		return nil
+	}
+	l.mu.Lock()
+	defer l.mu.Unlock()
+	all, err := l.read()
+	if err != nil {
+		return err
+	}
+	rec := Record{
+		Seq: len(all) + 1, Anchor: anchor, Root: root, Path: filepath.ToSlash(rel),
+		Existed: existed, Before: append([]byte(nil), before...), After: append([]byte(nil), after...),
+	}
+	raw, err := json.Marshal(rec)
+	if err != nil {
+		return err
+	}
+	f, err := os.OpenFile(l.path, os.O_APPEND|os.O_WRONLY, 0o644)
+	if err != nil {
+		return err
+	}
+	defer f.Close()
+	_, err = f.Write(append(raw, '\n'))
+	return err
+}
+
+// Plan 列出从 anchor 起（含）会被回滚的文件，不改磁盘。
+// ordered 是用户消息 id 的时间顺序。
+func (l *Log) Plan(anchor string, ordered []string) (Result, error) {
+	if l == nil {
+		return Result{}, nil
+	}
+	l.mu.Lock()
+	defer l.mu.Unlock()
+	all, err := l.read()
+	if err != nil {
+		return Result{}, err
+	}
+	return classify(all, anchor, ordered), nil
+}
+
+// Rewind 把 anchor 及其后的被追踪写入退回写入前。
+// 当前内容与最后一次记录的 After 不一致时跳过并列入 Skipped。
+func (l *Log) Rewind(anchor string, ordered []string) (Result, error) {
+	if l == nil {
+		return Result{}, nil
+	}
+	l.mu.Lock()
+	defer l.mu.Unlock()
+	all, err := l.read()
+	if err != nil {
+		return Result{}, err
+	}
+	plan := classify(all, anchor, ordered)
+	for _, rel := range plan.Restored {
+		group := groupFor(selected(all, anchor, ordered), rel)
+		if err := restore(group); err != nil {
+			return Result{}, err
+		}
+	}
+	return plan, nil
+}
+
+func (l *Log) read() ([]Record, error) {
+	b, err := os.ReadFile(l.path)
+	if err != nil {
+		return nil, err
+	}
+	var out []Record
+	for _, line := range bytes.Split(b, []byte("\n")) {
+		line = bytes.TrimSpace(line)
+		if len(line) == 0 {
+			continue
+		}
+		var rec Record
+		if err := json.Unmarshal(line, &rec); err != nil {
+			return nil, err
+		}
+		out = append(out, rec)
+	}
+	return out, nil
+}
+
+func selected(all []Record, anchor string, ordered []string) []Record {
+	keep := map[string]bool{}
+	on := false
+	for _, id := range ordered {
+		if id == anchor {
+			on = true
+		}
+		if on {
+			keep[id] = true
+		}
+	}
+	if anchor != "" && len(ordered) == 0 {
+		keep[anchor] = true
+	}
+	var out []Record
+	for _, rec := range all {
+		if keep[rec.Anchor] {
+			out = append(out, rec)
+		}
+	}
+	return out
+}
+
+func classify(all []Record, anchor string, ordered []string) Result {
+	sel := selected(all, anchor, ordered)
+	seen := map[string]bool{}
+	var restored, skipped []string
+	for _, rec := range sel {
+		if seen[rec.Path] {
+			continue
+		}
+		seen[rec.Path] = true
+		group := groupFor(sel, rec.Path)
+		switch fileState(group) {
+		case stateRestore:
+			restored = append(restored, rec.Path)
+		case stateSkip:
+			skipped = append(skipped, rec.Path)
+		}
+	}
+	return Result{Restored: restored, Skipped: skipped}
+}
+
+func groupFor(sel []Record, rel string) []Record {
+	var group []Record
+	for _, rec := range sel {
+		if rec.Path == rel {
+			group = append(group, rec)
+		}
+	}
+	return group
+}
+
+const (
+	stateAlready = iota
+	stateRestore
+	stateSkip
+)
+
+func fileState(group []Record) int {
+	if len(group) == 0 {
+		return stateAlready
+	}
+	first, last := group[0], group[len(group)-1]
+	abs := filepath.Join(last.Root, filepath.FromSlash(last.Path))
+	current, err := os.ReadFile(abs)
+	missing := os.IsNotExist(err)
+	if err != nil && !missing {
+		return stateSkip
+	}
+	if missing {
+		if !last.Existed && len(last.After) == 0 {
+			return stateAlready
+		}
+		if !first.Existed && bytes.Equal(nil, first.Before) && len(last.After) > 0 {
+			return stateSkip
+		}
+		return stateSkip
+	}
+	if bytes.Equal(current, last.After) {
+		if !first.Existed {
+			return stateRestore
+		}
+		if bytes.Equal(current, first.Before) {
+			return stateAlready
+		}
+		return stateRestore
+	}
+	if first.Existed && bytes.Equal(current, first.Before) {
+		return stateAlready
+	}
+	return stateSkip
+}
+
+func restore(group []Record) error {
+	if len(group) == 0 {
+		return nil
+	}
+	first := group[0]
+	abs := filepath.Join(first.Root, filepath.FromSlash(first.Path))
+	if !first.Existed {
+		return os.Remove(abs)
+	}
+	if err := os.MkdirAll(filepath.Dir(abs), 0o755); err != nil {
+		return err
+	}
+	return os.WriteFile(abs, first.Before, 0o644)
+}
@@ -0,0 +1,80 @@
+package checkpoint
+
+import (
+	"os"
+	"path/filepath"
+	"testing"
+)
+
+func TestRewindRestoresTrackedWriteAndSkipsExternalChange(t *testing.T) {
+	root := t.TempDir()
+	log, err := Open(filepath.Join(t.TempDir(), "cp.jsonl"))
+	if err != nil {
+		t.Fatal(err)
+	}
+	if err := os.WriteFile(filepath.Join(root, "keep.txt"), []byte("old"), 0o644); err != nil {
+		t.Fatal(err)
+	}
+	if err := log.Record("m1", root, "keep.txt", []byte("old"), true, []byte("new")); err != nil {
+		t.Fatal(err)
+	}
+	if err := os.WriteFile(filepath.Join(root, "keep.txt"), []byte("new"), 0o644); err != nil {
+		t.Fatal(err)
+	}
+	if err := os.WriteFile(filepath.Join(root, "side.txt"), []byte("from-shell"), 0o644); err != nil {
+		t.Fatal(err)
+	}
+	if err := log.Record("m1", root, "moved.txt", []byte("a"), true, []byte("b")); err != nil {
+		t.Fatal(err)
+	}
+	if err := os.WriteFile(filepath.Join(root, "moved.txt"), []byte("someone-else"), 0o644); err != nil {
+		t.Fatal(err)
+	}
+
+	got, err := log.Rewind("m1", []string{"m1"})
+	if err != nil {
+		t.Fatal(err)
+	}
+	if len(got.Restored) != 1 || got.Restored[0] != "keep.txt" {
+		t.Fatalf("restored %#v", got.Restored)
+	}
+	if len(got.Skipped) != 1 || got.Skipped[0] != "moved.txt" {
+		t.Fatalf("skipped %#v", got.Skipped)
+	}
+	body, err := os.ReadFile(filepath.Join(root, "keep.txt"))
+	if err != nil || string(body) != "old" {
+		t.Fatalf("keep.txt %q %v", body, err)
+	}
+	side, err := os.ReadFile(filepath.Join(root, "side.txt"))
+	if err != nil || string(side) != "from-shell" {
+		t.Fatalf("untracked file changed: %q %v", side, err)
+	}
+	moved, _ := os.ReadFile(filepath.Join(root, "moved.txt"))
+	if string(moved) != "someone-else" {
+		t.Fatalf("external edit was reverted: %q", moved)
+	}
+}
+
+func TestRewindRemovesCreatedFile(t *testing.T) {
+	root := t.TempDir()
+	log, err := Open(filepath.Join(t.TempDir(), "cp.jsonl"))
+	if err != nil {
+		t.Fatal(err)
+	}
+	if err := os.WriteFile(filepath.Join(root, "new.txt"), []byte("hi"), 0o644); err != nil {
+		t.Fatal(err)
+	}
+	if err := log.Record("m2", root, "new.txt", nil, false, []byte("hi")); err != nil {
+		t.Fatal(err)
+	}
+	got, err := log.Rewind("m2", []string{"m1", "m2"})
+	if err != nil {
+		t.Fatal(err)
+	}
+	if len(got.Restored) != 1 {
+		t.Fatalf("restored %#v", got.Restored)
+	}
+	if _, err := os.Stat(filepath.Join(root, "new.txt")); !os.IsNotExist(err) {
+		t.Fatalf("created file still present: %v", err)
+	}
+}
@@ -0,0 +1,51 @@
+"""工作流脚本回调。用法：import agentwf as wf。
+
+请求写到 fd 3，响应从 fd 4 按行读回。宿主一次只处理一条。
+"""
+
+import json
+import os
+
+_seq = 0
+_out = os.fdopen(3, "w", encoding="utf-8", buffering=1)
+_inp = os.fdopen(4, "r", encoding="utf-8", buffering=1)
+
+
+def _rpc(op, **fields):
+    global _seq
+    _seq += 1
+    msg = {"id": str(_seq), "op": op}
+    msg.update(fields)
+    _out.write(json.dumps(msg, ensure_ascii=False) + "\n")
+    _out.flush()
+    line = _inp.readline()
+    if not line:
+        raise RuntimeError("workflow host closed")
+    resp = json.loads(line)
+    if not resp.get("ok"):
+        raise RuntimeError(resp.get("error") or "workflow call failed")
+    return resp.get("output", "")
+
+
+def tool(name, args=None):
+    return _rpc("tool.call", name=name, args=args or {})
+
+
+def ask(question, choices=None):
+    return _rpc("hitl.ask", question=question, choices=choices or [])
+
+
+def agent(name, task):
+    return _rpc("agent.spawn", agent=name, task=task)
+
+
+def artifact(name, content):
+    return _rpc("artifact.register", name=name, content=content)
+
+
+def card(title, kind="result"):
+    return _rpc("card.create", title=title, kind=kind)
+
+
+def step(step_id, title, status):
+    return _rpc("step", stepId=step_id, title=title, status=status)
@@ -0,0 +1,12 @@
+package codeorch
+
+// Forbidden 是脚本不能再调用的工具：解释器、自身、派发和工作流。
+func Forbidden(name string) bool {
+	switch name {
+	case "run_python", "run_shell", "run_command", "run_tool_script",
+		"delegate_task", "squad_orchestrate", "run_workflow", "review_findings":
+		return true
+	default:
+		return false
+	}
+}
@@ -0,0 +1,380 @@
+// Package codeorch 在会话目录里跑一段 Python，并按行把工具调用交回宿主。
+// 工作流脚本和 run_tool_script 共用这一套进程与限额。
+package codeorch
+
+import (
+	"bufio"
+	"bytes"
+	"context"
+	"encoding/json"
+	"fmt"
+	"io"
+	"os"
+	"os/exec"
+	"path/filepath"
+	"strings"
+	"sync"
+	"time"
+	"unicode/utf8"
+
+	_ "embed"
+)
+
+//go:embed agentwf.py
+var agentwfSource string
+
+const (
+	defaultTimeout = 120 * time.Second
+	defaultMaxWait = 600 * time.Second
+	defaultCalls   = 50
+	defaultResult  = 1 << 20
+	defaultModel   = 16000
+)
+
+var (
+	helperOnce sync.Once
+	helperDir  string
+	helperErr  error
+)
+
+// Call 是脚本发来的一行请求。Args 与 Choices 保持原始 JSON。
+type Call struct {
+	ID       string          `json:"id"`
+	Op       string          `json:"op"`
+	Name     string          `json:"name"`
+	Args     json.RawMessage `json:"args"`
+	Question string          `json:"question"`
+	Choices  json.RawMessage `json:"choices"`
+	Agent    string          `json:"agent"`
+	Task     string          `json:"task"`
+	Content  string          `json:"content"`
+	Title    string          `json:"title"`
+	Kind     string          `json:"kind"`
+	Status   string          `json:"status"`
+	StepID   string          `json:"stepId"`
+}
+
+// Limits 控制一次脚本运行。零值使用工具路径的缺省：120 秒、最长 600 秒、50 次、1 MiB、16k 字符。
+type Limits struct {
+	Timeout        time.Duration
+	MaxTimeout     time.Duration
+	MaxCalls       int
+	MaxResultBytes int
+	MaxModelChars  int
+	FullDir        string
+}
+
+// Outcome 是给模型的文本、已成功的调用，以及超限时的全文路径。
+type Outcome struct {
+	Output   string
+	FullPath string
+	Calls    []string
+	Err      error
+}
+
+// Starter 按工作目录和脚本路径准备命令。空则用 python3。
+type Starter func(ctx context.Context, workDir, path string) (*exec.Cmd, error)
+
+// Dispatch 处理一条回调。返回的错误会写回脚本并记入最终结果。
+type Dispatch func(ctx context.Context, call Call) (string, error)
+
+// ModelText 把截断后的输出、已执行调用和错误拼成给模型的文本。
+func (o Outcome) ModelText() string {
+	text := o.Output
+	if o.Err != nil {
+		msg := o.Err.Error()
+		if len(o.Calls) > 0 {
+			msg = "已执行：" + strings.Join(o.Calls, "、") + "\n" + msg
+		}
+		if text != "" {
+			text += "\n"
+		}
+		text += msg
+	}
+	if o.FullPath != "" {
+		if text != "" {
+			text += "\n"
+		}
+		text += "全文在 " + o.FullPath
+	}
+	return text
+}
+
+// RunFile 执行已经落盘的脚本。
+func RunFile(ctx context.Context, workDir, path string, lim Limits, start Starter, dispatch Dispatch) Outcome {
+	if ctx == nil {
+		ctx = context.Background()
+	}
+	timeout := lim.bounded()
+	runCtx, cancel := context.WithTimeout(ctx, timeout)
+	defer cancel()
+	helper, err := helperPath()
+	if err != nil {
+		return Outcome{Err: err}
+	}
+	if start == nil {
+		start = defaultStart
+	}
+	cmd, err := start(runCtx, workDir, path)
+	if err != nil {
+		return Outcome{Err: err}
+	}
+	cmd.Env = prependPythonPath(cmd.Env, helper)
+	reqR, reqW, err := os.Pipe()
+	if err != nil {
+		return Outcome{Err: err}
+	}
+	respR, respW, err := os.Pipe()
+	if err != nil {
+		reqR.Close()
+		reqW.Close()
+		return Outcome{Err: err}
+	}
+	cmd.ExtraFiles = []*os.File{reqW, respR}
+	var stdout, stderr bytes.Buffer
+	cmd.Stdout = &stdout
+	cmd.Stderr = &stderr
+	if err := cmd.Start(); err != nil {
+		reqR.Close()
+		reqW.Close()
+		respR.Close()
+		respW.Close()
+		return Outcome{Err: err}
+	}
+	reqW.Close()
+	respR.Close()
+	defer reqR.Close()
+	defer respW.Close()
+	calls, callErr := serve(runCtx, reqR, respW, lim, dispatch)
+	if callErr != nil {
+		cancel()
+	}
+	waitErr := cmd.Wait()
+	out := Outcome{Output: clipJoined(&stdout, &stderr), Calls: calls}
+	out = clipModel(out, lim)
+	switch {
+	case callErr != nil:
+		out.Err = callErr
+	case runCtx.Err() == context.DeadlineExceeded:
+		out.Err = fmt.Errorf("脚本超时")
+	case runCtx.Err() == context.Canceled:
+		out.Err = fmt.Errorf("脚本已取消")
+	case waitErr != nil:
+		out.Err = fmt.Errorf("脚本退出：%v", waitErr)
+	}
+	return out
+}
+
+// Run 把脚本体写入工作目录后再执行。
+func Run(ctx context.Context, workDir, body string, lim Limits, start Starter, dispatch Dispatch) Outcome {
+	dir := filepath.Join(workDir, ".agent")
+	if err := os.MkdirAll(dir, 0o755); err != nil {
+		return Outcome{Err: err}
+	}
+	path := filepath.Join(dir, fmt.Sprintf("orch-%d.py", time.Now().UnixNano()))
+	if err := os.WriteFile(path, []byte(body), 0o600); err != nil {
+		return Outcome{Err: err}
+	}
+	return RunFile(ctx, workDir, path, lim, start, dispatch)
+}
+
+func (l Limits) bounded() time.Duration {
+	d := l.Timeout
+	if d <= 0 {
+		d = defaultTimeout
+	}
+	max := l.MaxTimeout
+	if max <= 0 {
+		max = defaultMaxWait
+	}
+	if d > max {
+		return max
+	}
+	return d
+}
+
+func (l Limits) calls() int {
+	if l.MaxCalls <= 0 {
+		return defaultCalls
+	}
+	return l.MaxCalls
+}
+
+func (l Limits) resultBytes() int {
+	if l.MaxResultBytes <= 0 {
+		return defaultResult
+	}
+	return l.MaxResultBytes
+}
+
+func (l Limits) modelChars() int {
+	if l.MaxModelChars <= 0 {
+		return defaultModel
+	}
+	return l.MaxModelChars
+}
+
+func serve(ctx context.Context, reqR, respW *os.File, lim Limits, dispatch Dispatch) ([]string, error) {
+	sc := bufio.NewScanner(reqR)
+	sc.Buffer(make([]byte, 64*1024), lim.resultBytes())
+	var calls []string
+	n := 0
+	for sc.Scan() {
+		line := strings.TrimSpace(sc.Text())
+		if line == "" {
+			continue
+		}
+		n++
+		var call Call
+		if err := json.Unmarshal([]byte(line), &call); err != nil {
+			_ = writeResp(respW, call.ID, false, "", "请求不是合法 JSON")
+			continue
+		}
+		if n > lim.calls() {
+			_ = writeResp(respW, call.ID, false, "", "脚本回调超过 50 次")
+			return calls, fmt.Errorf("脚本回调超过 50 次")
+		}
+		if ctx.Err() != nil {
+			return calls, ctx.Err()
+		}
+		if dispatch == nil {
+			_ = writeResp(respW, call.ID, false, "", "没有回调处理")
+			continue
+		}
+		output, err := dispatch(ctx, call)
+		if err != nil {
+			_ = writeResp(respW, call.ID, false, "", err.Error())
+			if ctx.Err() != nil {
+				return calls, ctx.Err()
+			}
+			continue
+		}
+		if call.Op == "tool.call" && call.Name != "" {
+			calls = append(calls, call.Name)
+		}
+		if err := writeResp(respW, call.ID, true, clipBytes(output, lim.resultBytes()), ""); err != nil {
+			return calls, err
+		}
+	}
+	if err := sc.Err(); err != nil {
+		if err == bufio.ErrTooLong {
+			return calls, fmt.Errorf("单行请求超过 1MB")
+		}
+		if ctx.Err() != nil {
+			return calls, ctx.Err()
+		}
+		return calls, err
+	}
+	return calls, nil
+}
+
+func writeResp(w io.Writer, id string, ok bool, output, errText string) error {
+	raw, err := json.Marshal(struct {
+		ID     string `json:"id"`
+		OK     bool   `json:"ok"`
+		Output string `json:"output,omitempty"`
+		Error  string `json:"error,omitempty"`
+	}{ID: id, OK: ok, Output: output, Error: errText})
+	if err != nil {
+		return err
+	}
+	_, err = w.Write(append(raw, '\n'))
+	return err
+}
+
+func clipBytes(s string, n int) string {
+	if n <= 0 || len(s) <= n {
+		return s
+	}
+	cut := s[:n]
+	for !utf8.ValidString(cut) && len(cut) > 0 {
+		cut = cut[:len(cut)-1]
+	}
+	return cut
+}
+
+func clipJoined(stdout, stderr *bytes.Buffer) string {
+	text := strings.TrimSpace(stdout.String())
+	errText := strings.TrimSpace(stderr.String())
+	if errText != "" {
+		if text != "" {
+			text += "\n"
+		}
+		text += errText
+	}
+	return text
+}
+
+func clipModel(out Outcome, lim Limits) Outcome {
+	limit := lim.modelChars()
+	if utf8.RuneCountInString(out.Output) <= limit {
+		return out
+	}
+	dir := lim.FullDir
+	if dir == "" {
+		dir = os.TempDir()
+	}
+	if err := os.MkdirAll(dir, 0o755); err != nil {
+		out.Output = clipRunes(out.Output, limit)
+		return out
+	}
+	path := filepath.Join(dir, fmt.Sprintf("script-output-%d.txt", time.Now().UnixNano()))
+	if err := os.WriteFile(path, []byte(out.Output), 0o644); err != nil {
+		out.Output = clipRunes(out.Output, limit)
+		return out
+	}
+	out.FullPath = path
+	out.Output = clipRunes(out.Output, limit)
+	return out
+}
+
+func clipRunes(s string, n int) string {
+	if utf8.RuneCountInString(s) <= n {
+		return s
+	}
+	i := 0
+	for count := 0; count < n && i < len(s); count++ {
+		_, size := utf8.DecodeRuneInString(s[i:])
+		i += size
+	}
+	return s[:i] + "…（超长截断）"
+}
+
+func defaultStart(ctx context.Context, workDir, path string) (*exec.Cmd, error) {
+	bin, err := exec.LookPath("python3")
+	if err != nil {
+		return nil, fmt.Errorf("找不到 python3: %w", err)
+	}
+	cmd := exec.CommandContext(ctx, bin, path)
+	cmd.Dir = workDir
+	return cmd, nil
+}
+
+func helperPath() (string, error) {
+	helperOnce.Do(func() {
+		dir, err := os.MkdirTemp("", "agentwf-*")
+		if err != nil {
+			helperErr = err
+			return
+		}
+		if err := os.WriteFile(filepath.Join(dir, "agentwf.py"), []byte(agentwfSource), 0o644); err != nil {
+			helperErr = err
+			return
+		}
+		helperDir = dir
+	})
+	return helperDir, helperErr
+}
+
+func prependPythonPath(env []string, dir string) []string {
+	if env == nil {
+		env = os.Environ()
+	}
+	for i, kv := range env {
+		if strings.HasPrefix(kv, "PYTHONPATH=") {
+			env[i] = "PYTHONPATH=" + dir + string(os.PathListSeparator) + strings.TrimPrefix(kv, "PYTHONPATH=")
+			return env
+		}
+	}
+	return append(env, "PYTHONPATH="+dir)
+}
@@ -0,0 +1,89 @@
+package codeorch
+
+import (
+	"context"
+	"errors"
+	"os"
+	"strings"
+	"testing"
+	"time"
+	"unicode/utf8"
+)
+
+func TestScriptCallsTwoToolsAndReportsTheOneThatRan(t *testing.T) {
+	body := "import agentwf as wf\nprint(wf.tool('alpha', {}))\nprint(wf.tool('beta', {}))\n"
+	out := Run(context.Background(), t.TempDir(), body, Limits{}, nil, func(_ context.Context, call Call) (string, error) {
+		if call.Name == "beta" {
+			return "", errors.New("beta 失败")
+		}
+		return "A", nil
+	})
+	text := out.ModelText()
+	if len(out.Calls) != 1 || out.Calls[0] != "alpha" {
+		t.Fatalf("calls %+v", out.Calls)
+	}
+	if !strings.Contains(text, "已执行：alpha") || !strings.Contains(text, "beta 失败") {
+		t.Fatalf("text %q err %v", text, out.Err)
+	}
+}
+
+func TestForbiddenToolIsRejectedBeforeDispatchRunsIt(t *testing.T) {
+	body := "import agentwf as wf\nprint(wf.tool('read_file', {}))\nprint(wf.tool('run_shell', {}))\n"
+	var ran []string
+	out := Run(context.Background(), t.TempDir(), body, Limits{}, nil, func(_ context.Context, call Call) (string, error) {
+		if Forbidden(call.Name) {
+			return "", errors.New("不能调用 " + call.Name)
+		}
+		ran = append(ran, call.Name)
+		return "ok", nil
+	})
+	if len(ran) != 1 || ran[0] != "read_file" {
+		t.Fatalf("ran %+v", ran)
+	}
+	if !strings.Contains(out.ModelText(), "不能调用 run_shell") {
+		t.Fatalf("text %q", out.ModelText())
+	}
+}
+
+func TestLongOutputIsClippedAndStored(t *testing.T) {
+	body := "import agentwf as wf\nprint(wf.tool('big', {}))\n"
+	dir := t.TempDir()
+	out := Run(context.Background(), dir, body, Limits{MaxModelChars: 8, FullDir: dir}, nil, func(_ context.Context, call Call) (string, error) {
+		return strings.Repeat("测", 20), nil
+	})
+	if out.FullPath == "" {
+		t.Fatal("expected a full-text file")
+	}
+	full, err := os.ReadFile(out.FullPath)
+	if err != nil {
+		t.Fatal(err)
+	}
+	if utf8.RuneCount(full) < 20 {
+		t.Fatalf("full text too short: %q", full)
+	}
+	if !strings.HasSuffix(out.Output, "…（超长截断）") || utf8.RuneCountInString(strings.TrimSuffix(out.Output, "…（超长截断）")) != 8 {
+		t.Fatalf("model output not clipped: %q", out.Output)
+	}
+	if !strings.Contains(out.ModelText(), out.FullPath) {
+		t.Fatalf("model text missing path: %q", out.ModelText())
+	}
+}
+
+func TestResultGivenToScriptIsCapped(t *testing.T) {
+	got := clipBytes(strings.Repeat("x", 20), 8)
+	if got != "xxxxxxxx" {
+		t.Fatalf("clip %q", got)
+	}
+	if !utf8.ValidString(clipBytes("测测测", 4)) {
+		t.Fatal("clip broke utf-8")
+	}
+}
+
+func TestTimeoutIsCapped(t *testing.T) {
+	if (Limits{Timeout: 900 * time.Second, MaxTimeout: 600 * time.Second}).bounded() != 600*time.Second {
+		t.Fatal("timeout must clamp to the max")
+	}
+	if (Limits{}).bounded() != defaultTimeout {
+		t.Fatal("empty limits must use the 120s default")
+	}
+}
@@ -0,0 +1,210 @@
+package runtime
+
+import (
+	"context"
+	"path/filepath"
+	"strings"
+	"testing"
+
+	"agent-native/runtime/llm"
+	"agent-native/runtime/session"
+	"agent-native/runtime/tools"
+)
+
+type replayTool struct {
+	countingTool
+	replay string
+}
+
+func (r *replayTool) Replay() string {
+	if r.replay == "" {
+		return tools.ReplayNever
+	}
+	return r.replay
+}
+
+func TestPendingUnsafeToolIsNotRerun(t *testing.T) {
+	dir := t.TempDir()
+	st, err := session.Open(filepath.Join(dir, "s.jsonl"))
+	if err != nil {
+		t.Fatal(err)
+	}
+	if err := st.Append(llm.Message{
+		Role: llm.RoleAssistant, StopReason: llm.StopToolCalls,
+		ToolCalls: []llm.ToolCall{{ID: "c1", Name: "write_file", Arguments: `{}`}},
+	}); err != nil {
+		t.Fatal(err)
+	}
+	if err := st.AppendToolIntent(session.ToolIntent{CallID: "c1", Name: "write_file", Arguments: `{}`, Replay: tools.ReplayNever}); err != nil {
+		t.Fatal(err)
+	}
+	tool := &countingTool{name: "write_file"}
+	reg := tools.NewRegistry(dir)
+	reg.Register(tool)
+	_, err = Run(context.Background(), Config{
+		Model: "fake", Session: st, Tools: reg, MaxTurns: 2,
+		Stream: script(llm.Message{Role: llm.RoleAssistant, Content: "later", StopReason: llm.StopEnd}),
+	}, llm.Message{Role: llm.RoleUser, Content: "next"})
+	if err != nil {
+		t.Fatal(err)
+	}
+	if tool.calls != 0 {
+		t.Fatalf("unsafe tool reran: %d", tool.calls)
+	}
+	msgs, err := st.Load()
+	if err != nil {
+		t.Fatal(err)
+	}
+	var paired bool
+	for _, m := range msgs {
+		if m.Role == llm.RoleTool && m.ToolCallID == "c1" && m.Content == unknownSideEffect {
+			paired = true
+		}
+	}
+	if !paired {
+		t.Fatalf("pair missing: %#v", msgs)
+	}
+}
+
+func TestPendingSafeToolRerunsOnce(t *testing.T) {
+	dir := t.TempDir()
+	st, err := session.Open(filepath.Join(dir, "s.jsonl"))
+	if err != nil {
+		t.Fatal(err)
+	}
+	if err := st.Append(llm.Message{
+		Role: llm.RoleAssistant, StopReason: llm.StopToolCalls,
+		ToolCalls: []llm.ToolCall{{ID: "c1", Name: "read_file", Arguments: `{}`}},
+	}); err != nil {
+		t.Fatal(err)
+	}
+	if err := st.AppendToolIntent(session.ToolIntent{CallID: "c1", Name: "read_file", Arguments: `{}`, Replay: tools.ReplaySafe}); err != nil {
+		t.Fatal(err)
+	}
+	tool := &replayTool{countingTool: countingTool{name: "read_file"}, replay: tools.ReplaySafe}
+	reg := tools.NewRegistry(dir)
+	reg.Register(tool)
+	_, err = Run(context.Background(), Config{
+		Model: "fake", Session: st, Tools: reg, MaxTurns: 2,
+		Stream: script(llm.Message{Role: llm.RoleAssistant, Content: "later", StopReason: llm.StopEnd}),
+	}, llm.Message{Role: llm.RoleUser, Content: "next"})
+	if err != nil {
+		t.Fatal(err)
+	}
+	if tool.calls != 1 {
+		t.Fatalf("calls %d", tool.calls)
+	}
+	msgs, err := st.Load()
+	if err != nil {
+		t.Fatal(err)
+	}
+	for _, m := range msgs {
+		if m.Role == llm.RoleTool && m.ToolCallID == "c1" {
+			if m.Content != "ok" || strings.Contains(m.Content, unknownSideEffect) {
+				t.Fatalf("result %#v", m)
+			}
+			return
+		}
+	}
+	t.Fatal("missing tool result")
+}
+
+func TestStoredSafeCurrentNeverDoesNotRerun(t *testing.T) {
+	dir := t.TempDir()
+	st, err := session.Open(filepath.Join(dir, "s.jsonl"))
+	if err != nil {
+		t.Fatal(err)
+	}
+	if err := st.AppendToolIntent(session.ToolIntent{CallID: "c1", Name: "read_file", Arguments: `{}`, Replay: tools.ReplaySafe}); err != nil {
+		t.Fatal(err)
+	}
+	tool := &replayTool{countingTool: countingTool{name: "read_file"}, replay: tools.ReplayNever}
+	reg := tools.NewRegistry(dir)
+	reg.Register(tool)
+	if err := recoverPending(context.Background(), Config{Session: st, Tools: reg}); err != nil {
+		t.Fatal(err)
+	}
+	if tool.calls != 0 {
+		t.Fatalf("calls %d", tool.calls)
+	}
+}
+
+func TestFollowUpIsPersistedBeforeNextRequest(t *testing.T) {
+	dir := t.TempDir()
+	st, err := session.Open(filepath.Join(dir, "s.jsonl"))
+	if err != nil {
+		t.Fatal(err)
+	}
+	sent := false
+	var saw bool
+	stream := func(ctx context.Context, model string, messages []llm.Message, tools []llm.ToolDecl, opts llm.StreamOptions) <-chan llm.Event {
+		ch := make(chan llm.Event, 1)
+		go func() {
+			defer close(ch)
+			for _, m := range messages {
+				if m.Content == "follow-please" {
+					loaded, err := st.Load()
+					if err != nil {
+						t.Errorf("load %v", err)
+					}
+					for _, got := range loaded {
+						if got.Content == "follow-please" {
+							saw = true
+						}
+					}
+				}
+			}
+			ch <- llm.Event{Kind: llm.KindMessageEnd, Message: llm.Message{Role: llm.RoleAssistant, Content: "ok", StopReason: llm.StopEnd}}
+		}()
+		return ch
+	}
+	_, err = Run(context.Background(), Config{
+		Model: "fake", Session: st, MaxTurns: 3, Stream: stream,
+		GetFollowUp: func() []llm.Message {
+			if sent {
+				return nil
+			}
+			sent = true
+			return []llm.Message{{Role: llm.RoleUser, Content: "follow-please"}}
+		},
+	}, llm.Message{Role: llm.RoleUser, Content: "go"})
+	if err != nil {
+		t.Fatal(err)
+	}
+	if !saw {
+		t.Fatal("follow-up entered the request before it was stored")
+	}
+}
+
+func TestSpansUseAgreedNamesWithoutPrompt(t *testing.T) {
+	var names []string
+	var leaked bool
+	reg := tools.NewRegistry(t.TempDir())
+	reg.Register(&countingTool{name: "read_file"})
+	_, err := Run(context.Background(), Config{
+		Model: "fake", Tools: reg, MaxTurns: 2,
+		Trace: func(name string, attrs map[string]string) func() {
+			names = append(names, name)
+			if _, ok := attrs["prompt"]; ok || attrs["content"] != "" {
+				leaked = true
+			}
+			return func() {}
+		},
+		Stream: script(
+			llm.Message{Role: llm.RoleAssistant, StopReason: llm.StopToolCalls, ToolCalls: []llm.ToolCall{{ID: "1", Name: "read_file", Arguments: `{}`}}},
+			llm.Message{Role: llm.RoleAssistant, Content: "done", StopReason: llm.StopEnd},
+		),
+	}, llm.Message{Role: llm.RoleUser, Content: "用户的秘密提示词"})
+	if err != nil {
+		t.Fatal(err)
+	}
+	joined := strings.Join(names, ",")
+	for _, want := range []string{"invoke_agent", "chat", "execute_tool"} {
+		if !strings.Contains(joined, want) {
+			t.Fatalf("names %s", joined)
+		}
+	}
+	if leaked {
+		t.Fatal("prompt leaked into span attrs")
+	}
+}
@@ -0,0 +1,181 @@
+package pkg
+
+import (
+	"encoding/json"
+	"fmt"
+	"io"
+	"os"
+	"os/exec"
+	"path/filepath"
+	"strings"
+)
+
+// Manifest is the agent-native.json discovered in a package.
+type Manifest struct {
+	Name       string            `json:"name"`
+	Extensions []string          `json:"extensions,omitempty"`
+	Config     map[string]string `json:"config,omitempty"`
+}
+
+// Manager installs packages under Dir.
+type Manager struct {
+	Dir string
+}
+
+func (m Manager) root() string {
+	if m.Dir != "" {
+		return m.Dir
+	}
+	home, _ := os.UserHomeDir()
+	return filepath.Join(home, ".agent-native", "packages")
+}
+
+// Install copies a local directory, clones a git URL, or runs npm pack
+// when src starts with "npm:". The package name comes from its manifest.
+func (m Manager) Install(src string) (Manifest, error) {
+	dir, err := os.MkdirTemp("", "agent-native-pkg-*")
+	if err != nil {
+		return Manifest{}, err
+	}
+	defer os.RemoveAll(dir)
+	switch {
+	case strings.HasPrefix(src, "npm:"):
+		if _, err := exec.LookPath("npm"); err != nil {
+			return Manifest{}, fmt.Errorf("pkg: npm is not installed")
+		}
+		cmd := exec.Command("npm", "pack", strings.TrimPrefix(src, "npm:"))
+		cmd.Dir = dir
+		if out, err := cmd.CombinedOutput(); err != nil {
+			return Manifest{}, fmt.Errorf("pkg: npm pack: %s", out)
+		}
+	case strings.HasPrefix(src, "git+") || strings.HasSuffix(src, ".git") || strings.Contains(src, "://"):
+		cmd := exec.Command("git", "clone", "--depth", "1", src, filepath.Join(dir, "src"))
+		if out, err := cmd.CombinedOutput(); err != nil {
+			return Manifest{}, fmt.Errorf("pkg: git clone: %s", out)
+		}
+		dir = filepath.Join(dir, "src")
+	default:
+		if err := copyDir(src, filepath.Join(dir, "src")); err != nil {
+			return Manifest{}, err
+		}
+		dir = filepath.Join(dir, "src")
+	}
+	manifest, err := ReadManifest(dir)
+	if err != nil {
+		return Manifest{}, err
+	}
+	dest := filepath.Join(m.root(), manifest.Name)
+	if err := os.RemoveAll(dest); err != nil {
+		return Manifest{}, err
+	}
+	if err := copyDir(dir, dest); err != nil {
+		return Manifest{}, err
+	}
+	return manifest, nil
+}
+
+// Remove deletes an installed package.
+func (m Manager) Remove(name string) error {
+	return os.RemoveAll(filepath.Join(m.root(), name))
+}
+
+// Update reinstalls from the recorded source when present, otherwise no-ops.
+func (m Manager) Update(name string) (Manifest, error) {
+	manifest, err := ReadManifest(filepath.Join(m.root(), name))
+	if err != nil {
+		return Manifest{}, err
+	}
+	src := manifest.Config["source"]
+	if src == "" {
+		return manifest, nil
+	}
+	return m.Install(src)
+}
+
+// List returns installed manifests.
+func (m Manager) List() ([]Manifest, error) {
+	entries, err := os.ReadDir(m.root())
+	if err != nil {
+		if os.IsNotExist(err) {
+			return nil, nil
+		}
+		return nil, err
+	}
+	var out []Manifest
+	for _, ent := range entries {
+		if !ent.IsDir() {
+			continue
+		}
+		manifest, err := ReadManifest(filepath.Join(m.root(), ent.Name()))
+		if err != nil {
+			continue
+		}
+		out = append(out, manifest)
+	}
+	return out, nil
+}
+
+// Config sets one package config key.
+func (m Manager) Config(name, key, value string) error {
+	path := filepath.Join(m.root(), name)
+	manifest, err := ReadManifest(path)
+	if err != nil {
+		return err
+	}
+	if manifest.Config == nil {
+		manifest.Config = map[string]string{}
+	}
+	manifest.Config[key] = value
+	raw, err := json.MarshalIndent(manifest, "", "  ")
+	if err != nil {
+		return err
+	}
+	return os.WriteFile(filepath.Join(path, "agent-native.json"), raw, 0o644)
+}
+
+// ReadManifest loads agent-native.json from a directory.
+func ReadManifest(dir string) (Manifest, error) {
+	raw, err := os.ReadFile(filepath.Join(dir, "agent-native.json"))
+	if err != nil {
+		return Manifest{}, err
+	}
+	var manifest Manifest
+	if err := json.Unmarshal(raw, &manifest); err != nil {
+		return Manifest{}, err
+	}
+	if manifest.Name == "" {
+		return Manifest{}, fmt.Errorf("pkg: manifest name is empty")
+	}
+	return manifest, nil
+}
+
+func copyDir(src, dst string) error {
+	return filepath.Walk(src, func(path string, info os.FileInfo, err error) error {
+		if err != nil {
+			return err
+		}
+		rel, err := filepath.Rel(src, path)
+		if err != nil {
+			return err
+		}
+		target := filepath.Join(dst, rel)
+		if info.IsDir() {
+			return os.MkdirAll(target, 0o755)
+		}
+		if err := os.MkdirAll(filepath.Dir(target), 0o755); err != nil {
+			return err
+		}
+		in, err := os.Open(path)
+		if err != nil {
+			return err
+		}
+		defer in.Close()
+		out, err := os.OpenFile(target, os.O_CREATE|os.O_TRUNC|os.O_WRONLY, info.Mode())
+		if err != nil {
+			return err
+		}
+		defer out.Close()
+		_, err = io.Copy(out, in)
+		return err
+	})
+}
@@ -0,0 +1,235 @@
+package ext
+
+import (
+	"bufio"
+	"context"
+	"encoding/json"
+	"fmt"
+	"io"
+	"os/exec"
+	"sync"
+
+	"agent-native/runtime/llm"
+	"agent-native/runtime/tools"
+)
+
+// Message is one JSONL frame on the extension pipe.
+type Message struct {
+	Type        string          `json:"type"`
+	ID          string          `json:"id,omitempty"`
+	Name        string          `json:"name,omitempty"`
+	Description string          `json:"description,omitempty"`
+	Parameters  json.RawMessage `json:"parameters,omitempty"`
+	Arguments   json.RawMessage `json:"arguments,omitempty"`
+	Content     string          `json:"content,omitempty"`
+	IsError     bool            `json:"isError,omitempty"`
+	Event       string          `json:"event,omitempty"`
+	Block       bool            `json:"block,omitempty"`
+	Reason      string          `json:"reason,omitempty"`
+	Message     string          `json:"message,omitempty"`
+	Protocol    string          `json:"protocol,omitempty"`
+	BaseURL     string          `json:"baseURL,omitempty"`
+	CustomType  string          `json:"customType,omitempty"`
+	Data        json.RawMessage `json:"data,omitempty"`
+	Command     string          `json:"command,omitempty"`
+}
+
+// UIRequest is a question an extension wants a human to answer.
+type UIRequest struct {
+	ID      string
+	Message string
+}
+
+// Process is one extension subprocess.
+type Process struct {
+	cmd      *exec.Cmd
+	stdin    io.WriteCloser
+	mu       sync.Mutex
+	tools    []registeredTool
+	commands []string
+	providers []llm.Provider
+	pending  map[string]chan Message
+	UI       chan UIRequest
+	OnCustom func(customType string, data json.RawMessage)
+	closed   bool
+}
+
+type registeredTool struct {
+	name        string
+	description string
+	parameters  json.RawMessage
+	proc        *Process
+}
+
+func (t registeredTool) Name() string                { return t.name }
+func (t registeredTool) Description() string         { return t.description }
+func (t registeredTool) Parameters() json.RawMessage { return t.parameters }
+func (t registeredTool) Execute(ctx context.Context, args json.RawMessage) (tools.Result, error) {
+	return t.proc.callTool(ctx, t.name, args)
+}
+
+// Start launches command and reads its registration frames until it goes idle
+// or the context ends. command is argv, not a shell line.
+func Start(ctx context.Context, command []string) (*Process, error) {
+	if len(command) == 0 {
+		return nil, fmt.Errorf("ext: empty command")
+	}
+	cmd := exec.CommandContext(ctx, command[0], command[1:]...)
+	stdin, err := cmd.StdinPipe()
+	if err != nil {
+		return nil, err
+	}
+	stdout, err := cmd.StdoutPipe()
+	if err != nil {
+		return nil, err
+	}
+	if err := cmd.Start(); err != nil {
+		return nil, err
+	}
+	p := &Process{
+		cmd: cmd, stdin: stdin, pending: map[string]chan Message{},
+		UI: make(chan UIRequest, 4),
+	}
+	go p.read(stdout)
+	if err := p.send(Message{Type: "init"}); err != nil {
+		_ = p.Close()
+		return nil, err
+	}
+	return p, nil
+}
+
+// Tools returns the tools the extension registered.
+func (p *Process) Tools() []tools.Tool {
+	p.mu.Lock()
+	defer p.mu.Unlock()
+	out := make([]tools.Tool, len(p.tools))
+	for i := range p.tools {
+		out[i] = p.tools[i]
+	}
+	return out
+}
+
+// Commands returns slash-command names.
+func (p *Process) Commands() []string {
+	p.mu.Lock()
+	defer p.mu.Unlock()
+	return append([]string(nil), p.commands...)
+}
+
+// Providers returns endpoints the extension registered.
+func (p *Process) Providers() []llm.Provider {
+	p.mu.Lock()
+	defer p.mu.Unlock()
+	return append([]llm.Provider(nil), p.providers...)
+}
+
+// Hook tells the extension about an event. A block result stops the host action.
+func (p *Process) Hook(ctx context.Context, event, name string, args json.RawMessage) (bool, string) {
+	id := name + ":" + event
+	ch := make(chan Message, 1)
+	p.mu.Lock()
+	p.pending[id] = ch
+	p.mu.Unlock()
+	if err := p.send(Message{Type: "event", ID: id, Event: event, Name: name, Arguments: args}); err != nil {
+		return false, ""
+	}
+	select {
+	case <-ctx.Done():
+		return false, ctx.Err().Error()
+	case msg := <-ch:
+		return msg.Block, msg.Reason
+	}
+}
+
+func (p *Process) callTool(ctx context.Context, name string, args json.RawMessage) (tools.Result, error) {
+	id := name
+	ch := make(chan Message, 1)
+	p.mu.Lock()
+	p.pending[id] = ch
+	p.mu.Unlock()
+	if err := p.send(Message{Type: "tool_call", ID: id, Name: name, Arguments: args}); err != nil {
+		return tools.Result{IsError: true, Content: err.Error()}, err
+	}
+	select {
+	case <-ctx.Done():
+		return tools.Result{IsError: true, Content: ctx.Err().Error()}, ctx.Err()
+	case msg := <-ch:
+		return tools.Result{Content: msg.Content, IsError: msg.IsError}, nil
+	}
+}
+
+func (p *Process) send(msg Message) error {
+	raw, err := json.Marshal(msg)
+	if err != nil {
+		return err
+	}
+	p.mu.Lock()
+	defer p.mu.Unlock()
+	if p.closed {
+		return fmt.Errorf("ext: closed")
+	}
+	_, err = p.stdin.Write(append(raw, '\n'))
+	return err
+}
+
+func (p *Process) read(r io.Reader) {
+	sc := bufio.NewScanner(r)
+	sc.Buffer(make([]byte, 0, 64*1024), 4*1024*1024)
+	for sc.Scan() {
+		var msg Message
+		if err := json.Unmarshal(sc.Bytes(), &msg); err != nil {
+			continue
+		}
+		p.dispatch(msg)
+	}
+}
+
+func (p *Process) dispatch(msg Message) {
+	switch msg.Type {
+	case "register_tool":
+		p.mu.Lock()
+		p.tools = append(p.tools, registeredTool{
+			name: msg.Name, description: msg.Description, parameters: msg.Parameters, proc: p,
+		})
+		p.mu.Unlock()
+	case "register_command":
+		p.mu.Lock()
+		p.commands = append(p.commands, msg.Name)
+		p.mu.Unlock()
+	case "register_provider":
+		p.mu.Lock()
+		p.providers = append(p.providers, llm.Provider{
+			Name: msg.Name, Protocol: llm.Protocol(msg.Protocol), BaseURL: msg.BaseURL,
+		})
+		p.mu.Unlock()
+	case "tool_result", "hook_result":
+		p.mu.Lock()
+		ch := p.pending[msg.ID]
+		delete(p.pending, msg.ID)
+		p.mu.Unlock()
+		if ch != nil {
+			ch <- msg
+		}
+	case "ui_request":
+		select {
+		case p.UI <- UIRequest{ID: msg.ID, Message: msg.Message}:
+		default:
+		}
+	case "custom_entry":
+		if p.OnCustom != nil {
+			p.OnCustom(msg.CustomType, msg.Data)
+		}
+	}
+}
+
+// Close stops the subprocess.
+func (p *Process) Close() error {
+	p.mu.Lock()
+	p.closed = true
+	_ = p.stdin.Close()
+	p.mu.Unlock()
+	if p.cmd.Process != nil {
+		_ = p.cmd.Process.Kill()
+	}
+	return p.cmd.Wait()
+}
@@ -0,0 +1,100 @@
+package ext
+
+import (
+	"context"
+	"os"
+	"path/filepath"
+	"testing"
+	"time"
+
+	"agent-native/runtime/ext/pkg"
+)
+
+func TestExtensionRegistersAndHooks(t *testing.T) {
+	dir := t.TempDir()
+	script := filepath.Join(dir, "ext.sh")
+	body := `#!/bin/sh
+while IFS= read -r line; do
+  case "$line" in
+    *'"type":"init"'*)
+      echo '{"type":"register_tool","name":"ping","description":"p","parameters":{"type":"object"}}'
+      echo '{"type":"register_command","name":"hi"}'
+      echo '{"type":"register_provider","name":"local","protocol":"openai-chat","baseURL":"http://127.0.0.1:9"}'
+      echo '{"type":"ui_request","id":"u1","message":"choose"}'
+      ;;
+    *'"type":"tool_call"'*)
+      echo '{"type":"tool_result","id":"ping","content":"pong"}'
+      ;;
+    *'"type":"event"'*)
+      echo '{"type":"hook_result","id":"ping:tool_call","block":true,"reason":"no"}'
+      ;;
+  esac
+done
+`
+	if err := os.WriteFile(script, []byte(body), 0o755); err != nil {
+		t.Fatal(err)
+	}
+	ctx, cancel := context.WithTimeout(context.Background(), 5*time.Second)
+	defer cancel()
+	proc, err := Start(ctx, []string{"/bin/sh", script})
+	if err != nil {
+		t.Fatal(err)
+	}
+	defer proc.Close()
+	deadline := time.Now().Add(2 * time.Second)
+	for len(proc.Tools()) == 0 && time.Now().Before(deadline) {
+		time.Sleep(20 * time.Millisecond)
+	}
+	tools := proc.Tools()
+	if len(tools) != 1 || tools[0].Name() != "ping" {
+		t.Fatalf("tools %+v", tools)
+	}
+	res, err := tools[0].Execute(ctx, []byte(`{}`))
+	if err != nil || res.Content != "pong" {
+		t.Fatalf("tool: %+v %v", res, err)
+	}
+	block, reason := proc.Hook(ctx, "tool_call", "ping", []byte(`{}`))
+	if !block || reason != "no" {
+		t.Fatalf("hook %v %s", block, reason)
+	}
+	if len(proc.Commands()) != 1 || len(proc.Providers()) != 1 {
+		t.Fatalf("commands %v providers %v", proc.Commands(), proc.Providers())
+	}
+	select {
+	case ui := <-proc.UI:
+		if ui.Message != "choose" {
+			t.Fatal(ui)
+		}
+	case <-time.After(time.Second):
+		t.Fatal("missing ui request")
+	}
+}
+
+func TestPackageInstallListConfigRemove(t *testing.T) {
+	src := t.TempDir()
+	if err := os.WriteFile(filepath.Join(src, "agent-native.json"), []byte(`{"name":"demo","extensions":["./ext.sh"]}`), 0o644); err != nil {
+		t.Fatal(err)
+	}
+	mgr := pkg.Manager{Dir: t.TempDir()}
+	manifest, err := mgr.Install(src)
+	if err != nil || manifest.Name != "demo" {
+		t.Fatalf("%+v %v", manifest, err)
+	}
+	if err := mgr.Config("demo", "source", src); err != nil {
+		t.Fatal(err)
+	}
+	list, err := mgr.List()
+	if err != nil || len(list) != 1 || list[0].Config["source"] != src {
+		t.Fatalf("list %+v %v", list, err)
+	}
+	if _, err := mgr.Update("demo"); err != nil {
+		t.Fatal(err)
+	}
+	if err := mgr.Remove("demo"); err != nil {
+		t.Fatal(err)
+	}
+	list, _ = mgr.List()
+	if len(list) != 0 {
+		t.Fatalf("still installed: %+v", list)
+	}
+}
@@ -0,0 +1,63 @@
+// Package filestate 记录本轮读过的文件，并在改写前核对修改时间与大小。
+package filestate
+
+import (
+	"fmt"
+	"os"
+	"sync"
+	"time"
+)
+
+type stamp struct {
+	mod  time.Time
+	size int64
+}
+
+// Tracker 是一次 run 内的读取记录。nil 接收者表示不检查。
+type Tracker struct {
+	mu   sync.Mutex
+	seen map[string]stamp
+}
+
+func New() *Tracker {
+	return &Tracker{seen: map[string]stamp{}}
+}
+
+// NoteRead 记下当前的修改时间与大小。
+func (t *Tracker) NoteRead(path string) error {
+	if t == nil {
+		return nil
+	}
+	info, err := os.Stat(path)
+	if err != nil {
+		return err
+	}
+	t.mu.Lock()
+	t.seen[path] = stamp{mod: info.ModTime(), size: info.Size()}
+	t.mu.Unlock()
+	return nil
+}
+
+// CheckWrite 允许新建。已存在的文件必须在本轮读过，且修改时间与大小未变。
+func (t *Tracker) CheckWrite(path string) error {
+	if t == nil {
+		return nil
+	}
+	info, err := os.Stat(path)
+	if err != nil {
+		if os.IsNotExist(err) {
+			return nil
+		}
+		return err
+	}
+	t.mu.Lock()
+	got, ok := t.seen[path]
+	t.mu.Unlock()
+	if !ok {
+		return fmt.Errorf("请先读取该文件再修改：%s", path)
+	}
+	if !got.mod.Equal(info.ModTime()) || got.size != info.Size() {
+		return fmt.Errorf("文件在读取后已变化，请重新读取：%s", path)
+	}
+	return nil
+}
@@ -0,0 +1,40 @@
+package filestate
+
+import (
+	"os"
+	"path/filepath"
+	"strings"
+	"testing"
+	"time"
+)
+
+func TestCheckWriteRejectsStaleFile(t *testing.T) {
+	dir := t.TempDir()
+	path := filepath.Join(dir, "note.txt")
+	if err := os.WriteFile(path, []byte("v1"), 0o644); err != nil {
+		t.Fatal(err)
+	}
+	tr := New()
+	if err := tr.CheckWrite(path); err == nil || !strings.Contains(err.Error(), "请先读取") {
+		t.Fatalf("unread write: %v", err)
+	}
+	if err := tr.NoteRead(path); err != nil {
+		t.Fatal(err)
+	}
+	if err := tr.CheckWrite(path); err != nil {
+		t.Fatal(err)
+	}
+	// 改大小，并尽量推开修改时间。
+	time.Sleep(20 * time.Millisecond)
+	if err := os.WriteFile(path, []byte("v1-changed"), 0o644); err != nil {
+		t.Fatal(err)
+	}
+	err := tr.CheckWrite(path)
+	if err == nil || !strings.Contains(err.Error(), "请重新读取") {
+		t.Fatalf("stale write: %v", err)
+	}
+	missing := filepath.Join(dir, "new.txt")
+	if err := tr.CheckWrite(missing); err != nil {
+		t.Fatalf("create: %v", err)
+	}
+}
@@ -0,0 +1,177 @@
+// Package gates 提供终轮门禁与进度提醒的组合机制（自产品下沉）。
+// 产品依赖——工件存取、质量评分、验证器模型、提示词、进度快照——全部端口化；
+// 判定逻辑（正则族）、续跑上限与节流策略由本包持有。
+// 端口为 nil 时语义与「未启用该检查」一致，不改变既有行为。
+package gates
+
+import (
+	"encoding/json"
+	"fmt"
+	"path/filepath"
+	"regexp"
+	"strings"
+)
+
+// 续跑与节流上限。与产品现网取值一致；消费方引用本包常量，不再各自硬编码。
+const (
+	// MaxDeliverableGuards 是「声称交付但无工件」的纠正轮次上限。
+	MaxDeliverableGuards = 2
+	// MaxVerifyContinuations 是终轮检查的续跑上限（与内核 Stop 的 3 次上限对齐）。
+	MaxVerifyContinuations = 3
+	// ReminderToolGap 是连续这么多个工具轮没有更新计划或待办就提醒一次。
+	ReminderToolGap = 3
+	// MaxChecklistReminders 是每个 run 的进度提醒次数上限。
+	MaxChecklistReminders = 3
+)
+
+// ArtifactLookup 按 ID 取工件：名称、内容与是否存在。
+type ArtifactLookup func(id string) (name string, content []byte, ok bool)
+
+// QualityScorer 对文本类工件评分；nil 表示跳过评分（视为通过）。
+// why 在未通过时给出可读原因。
+type QualityScorer func(name string, content []byte) (passed bool, why string)
+
+// ArtifactRef 是一次工件引用（库内最小形态；宿主负责投影自有类型）。
+type ArtifactRef struct {
+	ID   string
+	Name string
+	Kind string
+}
+
+var (
+	// ArtifactIDPattern 是工件编号的文本形态（门禁与宿主的引用提取共用）。
+	ArtifactIDPattern = regexp.MustCompile(`art_[0-9a-f]{12}`)
+	fileClaimRe       = regexp.MustCompile(`(?s)(?:已生成|已保存|已导出|已写入).{0,120}?\.(?:docx|xlsx|pptx|pdf|png|jpe?g|csv|md|txt|json)\b`)
+	softClaimRe       = regexp.MustCompile(`(?s)(?:文件|文档|工件).{0,40}?已(?:正常)?(?:生成|登记|保存|导出|写入)`)
+	forgedCallRe      = regexp.MustCompile(`\[(?:调用工具|工具结果|产出工件) `)
+)
+
+// ClaimGate 判断终轮正文是否在没有对应工件的情况下声称交付。
+// 引用本轮已产出、或确实存在的工件不算谎称；伪造调用标记始终拦截。
+// lookup 为 nil 时，任何引用未产出工件的 art_ 编号都视为不存在（拦截）。
+func ClaimGate(body string, produced []ArtifactRef, lookup ArtifactLookup) (bool, string) {
+	if strings.TrimSpace(body) == "" {
+		return false, ""
+	}
+	if forgedCallRe.MatchString(body) {
+		return true, "回复正文里出现了工具调用标记，但本轮没有真实的工具调用"
+	}
+	producedIDs := map[string]bool{}
+	for _, ref := range produced {
+		if ref.ID != "" {
+			producedIDs[ref.ID] = true
+		}
+	}
+	knownExisting := false
+	for _, id := range ArtifactIDPattern.FindAllString(body, -1) {
+		if producedIDs[id] {
+			continue
+		}
+		if lookup != nil {
+			if _, _, ok := lookup(id); ok {
+				knownExisting = true
+				continue
+			}
+		}
+		return true, "回复引用了不存在的工件"
+	}
+	if len(producedIDs) > 0 || knownExisting {
+		return false, ""
+	}
+	if fileClaimRe.MatchString(body) || softClaimRe.MatchString(body) {
+		return true, "回复声称已经生成或登记了文件，但这一轮没有产生对应的工件"
+	}
+	return false, ""
+}
+
+// CheckArtifacts 确认登记的工件存在且非空。
+// md、txt、docx 还要过 scorer；其他类型只检查存在与非空。
+func CheckArtifacts(refs []ArtifactRef, lookup ArtifactLookup, scorer QualityScorer) []string {
+	var reasons []string
+	for _, ref := range refs {
+		if ref.ID == "" {
+			reasons = append(reasons, "工件引用缺少 ID")
+			continue
+		}
+		name := ref.Name
+		if name == "" {
+			name = ref.ID
+		}
+		var (
+			content []byte
+			ok      bool
+			artName string
+		)
+		if lookup != nil {
+			artName, content, ok = lookup(ref.ID)
+		}
+		if !ok {
+			reasons = append(reasons, name+" 不存在")
+			continue
+		}
+		if artName != "" {
+			name = artName
+		}
+		if len(content) == 0 {
+			reasons = append(reasons, name+" 内容为空")
+			continue
+		}
+		switch strings.ToLower(filepath.Ext(name)) {
+		case ".md", ".txt", ".docx":
+			if scorer == nil {
+				continue
+			}
+			passed, why := scorer(name, content)
+			if !passed {
+				if why == "" {
+					why = "评分未通过"
+				}
+				reasons = append(reasons, name+"："+why)
+			}
+		}
+	}
+	return reasons
+}
+
+// Verdict 是独立验证器的裁决。
+type Verdict struct {
+	Passed     bool   `json:"passed"`
+	Reason     string `json:"reason"`
+	NextAction string `json:"nextAction"`
+}
+
+// ParseVerdict 解析验证器输出：剥围栏、取最外层 JSON、必须有 passed 键。
+func ParseVerdict(raw string) (Verdict, error) {
+	s := stripFence(raw)
+	start := strings.Index(s, "{")
+	end := strings.LastIndex(s, "}")
+	if start < 0 || end < start {
+		return Verdict{}, fmt.Errorf("验证器返回无法解析")
+	}
+	s = s[start : end+1]
+	var probe map[string]json.RawMessage
+	if err := json.Unmarshal([]byte(s), &probe); err != nil {
+		return Verdict{}, err
+	}
+	if _, ok := probe["passed"]; !ok {
+		return Verdict{}, fmt.Errorf("验证器返回无法解析")
+	}
+	var verdict Verdict
+	if err := json.Unmarshal([]byte(s), &verdict); err != nil {
+		return Verdict{}, err
+	}
+	return verdict, nil
+}
+
+func stripFence(s string) string {
+	s = strings.TrimSpace(s)
+	if !strings.HasPrefix(s, "```") {
+		return s
+	}
+	s = strings.TrimPrefix(s, "```")
+	if i := strings.IndexByte(s, '\n'); i >= 0 {
+		s = s[i+1:]
+	}
+	s = strings.TrimSpace(strings.TrimSuffix(strings.TrimSpace(s), "```"))
+	return s
+}
@@ -0,0 +1,133 @@
+package gates
+
+import (
+	"strings"
+	"testing"
+	"time"
+)
+
+func lookupWith(ids map[string][]byte) ArtifactLookup {
+	return func(id string) (string, []byte, bool) {
+		c, ok := ids[id]
+		return id + ".md", c, ok
+	}
+}
+
+func TestClaimGateForgedMarkerAlwaysBlocks(t *testing.T) {
+	block, why := ClaimGate("结果如下 [调用工具 write_file]", nil, nil)
+	if !block || why == "" {
+		t.Fatalf("block=%v why=%q", block, why)
+	}
+}
+
+func TestClaimGateProducedAndExistingDoNotBlock(t *testing.T) {
+	produced := []ArtifactRef{{ID: "art_aaaabbbbcccc"}}
+	if block, _ := ClaimGate("已生成 art_aaaabbbbcccc 请查收", produced, nil); block {
+		t.Fatal("produced reference blocked")
+	}
+	lookup := lookupWith(map[string][]byte{"art_dddd11112222": []byte("x")})
+	if block, _ := ClaimGate("之前的 art_dddd11112222 里有结论", nil, lookup); block {
+		t.Fatal("existing artifact reference blocked")
+	}
+}
+
+func TestClaimGateUnknownArtifactBlocks(t *testing.T) {
+	block, why := ClaimGate("见附件 art_000011112222", nil, lookupWith(nil))
+	if !block || !strings.Contains(why, "不存在") {
+		t.Fatalf("block=%v why=%q", block, why)
+	}
+}
+
+func TestClaimGateBareClaimWithoutArtifactsBlocks(t *testing.T) {
+	for _, body := range []string{
+		"已生成报告.md，请查收",
+		"文件已生成好了",
+	} {
+		if block, _ := ClaimGate(body, nil, nil); !block {
+			t.Fatalf("claim not blocked: %q", body)
+		}
+	}
+	// 如实说明没有生成则不拦。
+	if block, _ := ClaimGate("这一轮没有生成文件", nil, nil); block {
+		t.Fatal("honest note blocked")
+	}
+}
+
+func TestCheckArtifactsReasons(t *testing.T) {
+	lookup := func(id string) (string, []byte, bool) {
+		switch id {
+		case "ok":
+			return "ok.md", []byte("# 正常"), true
+		case "empty":
+			return "empty.md", nil, true
+		default:
+			return "", nil, false
+		}
+	}
+	scorer := func(name string, content []byte) (bool, string) { return false, "标题缺失" }
+	refs := []ArtifactRef{
+		{ID: ""},
+		{ID: "missing", Name: "缺失.md"},
+		{ID: "empty"},
+		{ID: "ok"},
+	}
+	reasons := CheckArtifacts(refs, lookup, scorer)
+	joined := strings.Join(reasons, "；")
+	for _, want := range []string{"缺少 ID", "缺失.md 不存在", "empty.md 内容为空", "ok.md：标题缺失"} {
+		if !strings.Contains(joined, want) {
+			t.Fatalf("reasons %q 缺 %q", joined, want)
+		}
+	}
+}
+
+func TestCheckArtifactsNilScorerSkipsQuality(t *testing.T) {
+	lookup := lookupWith(map[string][]byte{"a": []byte("x")})
+	if reasons := CheckArtifacts([]ArtifactRef{{ID: "a"}}, lookup, nil); len(reasons) != 0 {
+		t.Fatalf("reasons = %v", reasons)
+	}
+}
+
+func TestParseVerdictForms(t *testing.T) {
+	v, err := ParseVerdict("```json\n{\"passed\":true,\"reason\":\"ok\"}\n```")
+	if err != nil || !v.Passed {
+		t.Fatalf("v=%+v err=%v", v, err)
+	}
+	if _, err := ParseVerdict("没有 JSON"); err == nil {
+		t.Fatal("garbage accepted")
+	}
+	if _, err := ParseVerdict("{\"reason\":\"no passed key\"}"); err == nil {
+		t.Fatal("missing passed accepted")
+	}
+}
+
+func TestProgressReminderThrottleAndFormat(t *testing.T) {
+	r := &ProgressReminder{}
+	now := time.Now()
+	first := r.Remind([]string{"写第三章"}, []string{"评审"}, now)
+	if len(first) != 1 {
+		t.Fatalf("first = %v", first)
+	}
+	text := first[0]
+	for _, want := range []string{"已连续 3 个工具轮", "- 待办：写第三章", "- 计划：评审", "请更新计划或待办"} {
+		if !strings.Contains(text, want) {
+			t.Fatalf("text 缺 %q：\n%s", want, text)
+		}
+	}
+	// 内容相同且计数不同：key 不同，应继续发出直到封顶。
+	if got := r.Remind([]string{"写第三章"}, []string{"评审"}, now); len(got) != 1 {
+		t.Fatalf("second = %v", got)
+	}
+	if got := r.Remind([]string{"写第三章"}, []string{"评审"}, now); len(got) != 1 {
+		t.Fatalf("third = %v", got)
+	}
+	if got := r.Remind([]string{"写第三章"}, []string{"评审"}, now); got != nil {
+		t.Fatalf("fourth should be capped: %v", got)
+	}
+	if r.Sent() != 3 {
+		t.Fatalf("sent = %d", r.Sent())
+	}
+	// 无内容不发。
+	if got := (&ProgressReminder{}).Remind(nil, nil, now); got != nil {
+		t.Fatalf("empty = %v", got)
+	}
+}
@@ -0,0 +1,78 @@
+package gates
+
+import (
+	"fmt"
+	"strings"
+	"time"
+
+	"agent-native/runtime/reminders"
+)
+
+// ProgressReminder 是进度提醒的节流状态机与格式化机制。
+// 计数（第几个工具轮）由调用方按自己的事件流维护；本机制负责
+// 「达到阈值后的去重、封顶与文案」。
+type ProgressReminder struct {
+	Gap   int // 连续多少个工具轮未更新即提醒；0 = ReminderToolGap
+	Cap   int // 每 run 提醒上限；0 = MaxChecklistReminders
+	notes *reminders.Bus
+	sent  int
+}
+
+func (r *ProgressReminder) gap() int {
+	if r.Gap <= 0 {
+		return ReminderToolGap
+	}
+	return r.Gap
+}
+
+func (r *ProgressReminder) cap() int {
+	if r.Cap <= 0 {
+		return MaxChecklistReminders
+	}
+	return r.Cap
+}
+
+// Sent 报告已发出的提醒条数（调用方判定封顶用）。
+func (r *ProgressReminder) Sent() int { return r.sent }
+
+// Remind 在给定未完成项时发一条提醒；节流或没有内容时返回 nil。
+// 文案与产品现网一致：「计划和待办已连续 N 个工具轮没有更新」。
+func (r *ProgressReminder) Remind(todos, plans []string, now time.Time) []string {
+	if r.sent >= r.cap() {
+		return nil
+	}
+	text := FormatUnfinished(r.gap(), todos, plans)
+	if text == "" {
+		return nil
+	}
+	if r.notes == nil {
+		r.notes = reminders.New(0, r.cap())
+	}
+	key := fmt.Sprintf("%d:%s", r.sent, text)
+	if !r.notes.Admit(reminders.Note{Kind: reminders.KindTodo, Key: key, Text: text}, now) {
+		return nil
+	}
+	r.sent++
+	return []string{text}
+}
+
+// FormatUnfinished 渲染未完成项提醒；无内容时返回空串。
+func FormatUnfinished(gap int, todos, plans []string) string {
+	var lines []string
+	for _, item := range todos {
+		if text := strings.TrimSpace(item); text != "" {
+			lines = append(lines, "- 待办："+text)
+		}
+	}
+	for _, step := range plans {
+		if text := strings.TrimSpace(step); text != "" {
+			lines = append(lines, "- 计划："+text)
+		}
+	}
+	if len(lines) == 0 {
+		return ""
+	}
+	return fmt.Sprintf("计划和待办已连续 %d 个工具轮没有更新。未完成项：\n", gap) +
+		strings.Join(lines, "\n") +
+		"\n请更新计划或待办，或在回复里说明为什么先不做。"
+}
@@ -0,0 +1,94 @@
+// Package guardrails 提供三类语义化护栏端口：输入（只作用于入口智能体）、
+// 输出（只作用于给出最终结果的智能体）、工具（每次工具调用）。
+// 护栏以钩子总线注册：它是类型化语义层，钩子仍是通用事件层。
+// 当前形态护栏拿不到运行 ctx（事件载荷不含它）；需要取消语义时直接写钩子。
+package guardrails
+
+import (
+	"context"
+
+	"agent-native/runtime/hooks"
+)
+
+// Verdict 是一次护栏表态。空 Decision 表示不插手。
+// Deny 拒绝（Output 层表现为要求模型改进后继续）；Modify 改写载荷
+// （Input 改用户输入、Output 改最终文本、Tool 改参数）；Reason 附说明。
+type Verdict struct {
+	Decision    hooks.Decision
+	Reason      string
+	Replacement string
+}
+
+type InputGuardrail func(ctx context.Context, text string) Verdict
+type OutputGuardrail func(ctx context.Context, text string) Verdict
+type ToolGuardrail func(ctx context.Context, name, args string) Verdict
+
+// Set 是一组护栏；nil 字段表示该层不启用。
+type Set struct {
+	Input  InputGuardrail
+	Output OutputGuardrail
+	Tool   ToolGuardrail
+}
+
+// Scope 标注本 run 的角色：Entry=本 run 处理入口用户输入；
+// Final=本 run 给出最终答案。子代理 run 两者都为 false，
+// Input/Output 护栏不会作用于它；Tool 护栏作用于一切工具调用。
+type Scope struct {
+	Entry bool
+	Final bool
+}
+
+// Register 把护栏挂到总线。语义：
+//   - Input（仅 scope.Entry）：挂 UserPromptSubmit。deny → 循环不启动，
+//     以理由直接收尾；modify → 改写后的输入追加为新的用户消息（原文留痕）。
+//   - Output（仅 scope.Final）：挂 Stop。deny → 视同续跑，理由交给模型改进
+//     （计入内核续跑上限，达上限后收尾并附 Notice）；modify → 替换最终文本并续跑。
+//   - Tool：挂 PreToolUse，与其他钩子按既有优先级合并
+//     （deny/block > ask > modify > allow）；modify 的 Replacement 改写工具参数。
+func Register(bus *hooks.Bus, set Set, scope Scope) {
+	if bus == nil {
+		return
+	}
+	if set.Input != nil && scope.Entry {
+		g := set.Input
+		bus.On(func(c hooks.Call) hooks.Outcome {
+			if c.Event != hooks.UserPromptSubmit {
+				return hooks.Outcome{}
+			}
+			v := g(context.Background(), c.LastText)
+			return hooks.Outcome{Decision: v.Decision, Reason: v.Reason, Replacement: v.Replacement}
+		})
+	}
+	if set.Output != nil && scope.Final {
+		g := set.Output
+		bus.On(func(c hooks.Call) hooks.Outcome {
+			if c.Event != hooks.Stop {
+				return hooks.Outcome{}
+			}
+			v := g(context.Background(), c.LastText)
+			switch v.Decision {
+			case hooks.Deny, hooks.Block:
+				// 对 Stop 事件而言「拒绝最终答案」= 要求模型改进后再来一轮。
+				return hooks.Outcome{Decision: hooks.Block, Reason: v.Reason}
+			case hooks.Modify:
+				return hooks.Outcome{Decision: hooks.Block, Reason: v.Reason, Replacement: v.Replacement}
+			default:
+				return hooks.Outcome{Decision: v.Decision, Reason: v.Reason}
+			}
+		})
+	}
+	if set.Tool != nil {
+		g := set.Tool
+		bus.On(func(c hooks.Call) hooks.Outcome {
+			if c.Event != hooks.PreToolUse {
+				return hooks.Outcome{}
+			}
+			v := g(context.Background(), c.Tool, c.Args)
+			out := hooks.Outcome{Decision: v.Decision, Reason: v.Reason}
+			if v.Decision == hooks.Modify {
+				out.Args = v.Replacement
+			}
+			return out
+		})
+	}
+}
@@ -0,0 +1,89 @@
+package guardrails
+
+import (
+	"context"
+	"testing"
+
+	"agent-native/runtime/hooks"
+)
+
+func TestToolGuardrailMergesOnPreToolUse(t *testing.T) {
+	bus := hooks.New()
+	Register(bus, Set{
+		Tool: func(ctx context.Context, name, args string) Verdict {
+			if name == "run_shell" {
+				return Verdict{Decision: hooks.Deny, Reason: "shell 被护栏禁用"}
+			}
+			return Verdict{}
+		},
+	}, Scope{})
+	out := bus.Emit(hooks.Call{Event: hooks.PreToolUse, Tool: "run_shell"})
+	if out.Decision != hooks.Deny || out.Reason == "" {
+		t.Fatalf("out = %+v", out)
+	}
+	out = bus.Emit(hooks.Call{Event: hooks.PreToolUse, Tool: "read_file"})
+	if out.Decision != "" {
+		t.Fatalf("read_file out = %+v", out)
+	}
+}
+
+func TestToolGuardrailModifyRewritesArgs(t *testing.T) {
+	bus := hooks.New()
+	Register(bus, Set{
+		Tool: func(ctx context.Context, name, args string) Verdict {
+			return Verdict{Decision: hooks.Modify, Replacement: `{"path":"safe.txt"}`}
+		},
+	}, Scope{})
+	out := bus.Emit(hooks.Call{Event: hooks.PreToolUse, Tool: "write_file"})
+	if out.Decision != hooks.Modify || out.Args != `{"path":"safe.txt"}` {
+		t.Fatalf("out = %+v", out)
+	}
+}
+
+func TestOutputGuardrailDenyBecomesContinuation(t *testing.T) {
+	bus := hooks.New()
+	Register(bus, Set{
+		Output: func(ctx context.Context, text string) Verdict {
+			return Verdict{Decision: hooks.Deny, Reason: "缺少来源标注"}
+		},
+	}, Scope{Final: true})
+	out := bus.Emit(hooks.Call{Event: hooks.Stop, LastText: "结论"})
+	// Output deny 必须表达为续跑（Block），而不是把 run 掐死。
+	if out.Decision != hooks.Block || out.Reason != "缺少来源标注" {
+		t.Fatalf("out = %+v", out)
+	}
+}
+
+func TestScopeGatesInputAndOutput(t *testing.T) {
+	fired := map[string]bool{}
+	mk := func() Set {
+		return Set{
+			Input:  func(ctx context.Context, text string) Verdict { fired["input"] = true; return Verdict{} },
+			Output: func(ctx context.Context, text string) Verdict { fired["output"] = true; return Verdict{} },
+		}
+	}
+	// 子代理 run（Entry/Final 都为 false）：Input/Output 不注册。
+	bus := hooks.New()
+	Register(bus, mk(), Scope{})
+	bus.Emit(hooks.Call{Event: hooks.UserPromptSubmit, LastText: "x"})
+	bus.Emit(hooks.Call{Event: hooks.Stop, LastText: "y"})
+	if fired["input"] || fired["output"] {
+		t.Fatalf("sub-run scope fired: %v", fired)
+	}
+	// 入口且最终：两层都触发。
+	bus2 := hooks.New()
+	Register(bus2, mk(), Scope{Entry: true, Final: true})
+	bus2.Emit(hooks.Call{Event: hooks.UserPromptSubmit, LastText: "x"})
+	bus2.Emit(hooks.Call{Event: hooks.Stop, LastText: "y"})
+	if !fired["input"] || !fired["output"] {
+		t.Fatalf("entry/final scope did not fire: %v", fired)
+	}
+}
+
+func TestNilBusIsNoop(t *testing.T) {
+	Register(nil, Set{
+		Input:  func(ctx context.Context, text string) Verdict { return Verdict{Decision: hooks.Deny} },
+		Output: func(ctx context.Context, text string) Verdict { return Verdict{Decision: hooks.Deny} },
+		Tool:   func(ctx context.Context, name, args string) Verdict { return Verdict{Decision: hooks.Deny} },
+	}, Scope{Entry: true, Final: true})
+}
@@ -0,0 +1,136 @@
+// Package hooks 是进程内钩子总线。决策按 block/deny、ask、modify、allow 取更强的一条。
+// 没有表态的处理函数不改变结果。
+package hooks
+
+import "sync"
+
+type Event string
+
+const (
+	SessionStart       Event = "SessionStart"
+	UserPromptSubmit   Event = "UserPromptSubmit"
+	PreToolUse         Event = "PreToolUse"
+	PermissionRequest  Event = "PermissionRequest"
+	PostToolUse        Event = "PostToolUse"
+	PostToolUseFailure Event = "PostToolUseFailure"
+	PostToolBatch      Event = "PostToolBatch"
+	Stop               Event = "Stop"
+	SubagentStart      Event = "SubagentStart"
+	SubagentStop       Event = "SubagentStop"
+	PreCompact         Event = "PreCompact"
+	SessionEnd         Event = "SessionEnd"
+)
+
+type Decision string
+
+const (
+	Allow  Decision = "allow"
+	Deny   Decision = "deny"
+	Ask    Decision = "ask"
+	Modify Decision = "modify"
+	Block  Decision = "block"
+)
+
+// Call 是一次事件。工具字段只在工具事件上有意义。
+type Call struct {
+	Event         Event
+	Tool          string
+	Args          string
+	Result        string
+	IsError       bool
+	Turn          int
+	LastText      string
+	Continuations int
+	ToolCalls     int
+}
+
+// Outcome 是一个处理函数的表态。空 Decision 表示不插手。
+type Outcome struct {
+	Decision    Decision
+	Reason      string
+	Args        string
+	Notice      string
+	Replacement string
+}
+
+type Handler func(Call) Outcome
+
+type Bus struct {
+	mu sync.Mutex
+	hs []Handler
+}
+
+func New() *Bus { return &Bus{} }
+
+func (b *Bus) On(h Handler) {
+	if b == nil || h == nil {
+		return
+	}
+	b.mu.Lock()
+	b.hs = append(b.hs, h)
+	b.mu.Unlock()
+}
+
+// Emit 按优先级合并表态。deny 与 block 同级，先登记的保留。
+func (b *Bus) Emit(c Call) Outcome {
+	if b == nil {
+		return Outcome{}
+	}
+	b.mu.Lock()
+	hs := append([]Handler(nil), b.hs...)
+	b.mu.Unlock()
+	var merged Outcome
+	for _, h := range hs {
+		merged = prefer(merged, h(c))
+	}
+	return merged
+}
+
+func prefer(cur, next Outcome) Outcome {
+	if rank(next.Decision) > rank(cur.Decision) {
+		return next
+	}
+	return cur
+}
+
+func rank(d Decision) int {
+	switch d {
+	case Block, Deny:
+		return 4
+	case Ask:
+		return 3
+	case Modify:
+		return 2
+	case Allow:
+		return 1
+	default:
+		return 0
+	}
+}
+
+// AdaptBefore 把「返回错误即拒绝」收成 PreToolUse 处理函数。
+func AdaptBefore(fn func(name, args string) error) Handler {
+	return func(c Call) Outcome {
+		if c.Event != PreToolUse || fn == nil {
+			return Outcome{}
+		}
+		if err := fn(c.Tool, c.Args); err != nil {
+			return Outcome{Decision: Deny, Reason: err.Error()}
+		}
+		return Outcome{Decision: Allow}
+	}
+}
+
+// AdaptAfter 把工具结果观察收成 PostToolUse 或失败事件。它不改写结果。
+func AdaptAfter(fn func(name, args, result string, isError bool)) Handler {
+	return func(c Call) Outcome {
+		if fn == nil {
+			return Outcome{}
+		}
+		switch c.Event {
+		case PostToolUse, PostToolUseFailure:
+			fn(c.Tool, c.Args, c.Result, c.IsError)
+		}
+		return Outcome{}
+	}
+}
@@ -0,0 +1,40 @@
+package hooks
+
+import "testing"
+
+func TestDenyBeatsAllow(t *testing.T) {
+	b := New()
+	b.On(func(Call) Outcome { return Outcome{Decision: Allow} })
+	b.On(func(c Call) Outcome {
+		if c.Event == PreToolUse {
+			return Outcome{Decision: Deny, Reason: "no"}
+		}
+		return Outcome{}
+	})
+	out := b.Emit(Call{Event: PreToolUse, Tool: "write_file"})
+	if out.Decision != Deny || out.Reason != "no" {
+		t.Fatalf("%+v", out)
+	}
+}
+
+func TestAdaptBefore(t *testing.T) {
+	h := AdaptBefore(func(name, args string) error {
+		if name == "write_file" {
+			return errString("blocked")
+		}
+		return nil
+	})
+	if h(Call{Event: PreToolUse, Tool: "read_file"}).Decision != Allow {
+		t.Fatal("read")
+	}
+	if h(Call{Event: PreToolUse, Tool: "write_file"}).Decision != Deny {
+		t.Fatal("write")
+	}
+	if h(Call{Event: Stop}).Decision != "" {
+		t.Fatal("stop should be ignored")
+	}
+}
+
+type errString string
+
+func (e errString) Error() string { return string(e) }
@@ -42,6 +42,9 @@ func (c *Client) run(ctx context.Context, model string, messages []Message, tool
 		emitError(out, "", false, "agentx llm client is missing base url, api key, or model")
 		return
 	}
+	if opts.MessageShape != "" {
+		messages = PrepareMessages(messages, opts.MessageShape)
+	}
 	body := map[string]any{
 		"model":    model,
 		"messages": toProviderMessages(messages),
@@ -53,6 +56,9 @@ func (c *Client) run(ctx context.Context, model string, messages []Message, tool
 	if opts.MaxTokens > 0 {
 		body["max_tokens"] = opts.MaxTokens
 	}
+	if effort := OpenAIReasoningEffort(opts.ThinkingLevel); effort != "" {
+		body["reasoning_effort"] = effort
+	}
 	if len(tools) > 0 {
 		body["tools"] = toProviderTools(tools)
 		body["tool_choice"] = "auto"
@@ -86,8 +92,20 @@ func (c *Client) run(ctx context.Context, model string, messages []Message, tool
 	defer resp.Body.Close()
 	if resp.StatusCode >= 400 {
 		slurp, _ := io.ReadAll(io.LimitReader(resp.Body, 4096))
-		retryable := resp.StatusCode == http.StatusTooManyRequests || resp.StatusCode >= 500
-		emitError(out, "", retryable, fmt.Sprintf("provider status %d: %s", resp.StatusCode, strings.TrimSpace(string(slurp))))
+		msg := fmt.Sprintf("provider status %d: %s", resp.StatusCode, strings.TrimSpace(string(slurp)))
+		ev := Event{
+			Kind:         KindError,
+			HTTPStatus:   resp.StatusCode,
+			Retryable:    StatusRetryable(resp.StatusCode) && !ContextOverflow(msg),
+			ErrorMessage: msg,
+			Err:          &StreamError{Message: msg, Retryable: StatusRetryable(resp.StatusCode) && !ContextOverflow(msg)},
+			Message:      Message{Role: RoleAssistant, StopReason: StopError},
+		}
+		if d, ok := ParseRetryAfter(resp.Header.Get("Retry-After"), time.Now()); ok {
+			ev.HasRetryAfter = true
+			ev.RetryAfter = d
+		}
+		out <- ev
 		return
 	}
 	for ev := range ParseSSE(ctx, resp.Body) {
@@ -125,7 +143,7 @@ func toProviderMessages(messages []Message) []map[string]any {
 		default:
 			out = append(out, map[string]any{
 				"role":    string(m.Role),
-				"content": m.Content,
+				"content": messageContent(m),
 			})
 		}
 	}
@@ -0,0 +1,79 @@
+package llm
+
+import "context"
+
+// WithFallback 在尚未转发任何增量时，遇到 5xx 就改用链上下一个模型。
+// fallbacks 为空时返回原来的 StreamFunc，不改变行为。
+func WithFallback(base StreamFunc, fallbacks []string, onSwitch func(from, to string)) StreamFunc {
+	chain := nonEmpty(fallbacks)
+	if base == nil || len(chain) == 0 {
+		return base
+	}
+	return func(ctx context.Context, model string, messages []Message, tools []ToolDecl, opts StreamOptions) <-chan Event {
+		out := make(chan Event)
+		go func() {
+			defer close(out)
+			models := append([]string{model}, chain...)
+			for i, current := range models {
+				forwarded := false
+				switchNext := false
+				for ev := range base(ctx, current, messages, tools, opts) {
+					if isDelta(ev) {
+						forwarded = true
+					}
+					if ev.Kind == KindError && !forwarded && serverError(ev) && i+1 < len(models) {
+						switchNext = true
+						if onSwitch != nil {
+							onSwitch(current, models[i+1])
+						}
+						break
+					}
+					select {
+					case out <- ev:
+					case <-ctx.Done():
+						return
+					}
+					if ev.Kind == KindMessageEnd || ev.Kind == KindError {
+						return
+					}
+				}
+				if !switchNext {
+					return
+				}
+			}
+		}()
+		return out
+	}
+}
+
+func nonEmpty(in []string) []string {
+	var out []string
+	seen := map[string]bool{}
+	for _, s := range in {
+		s = trimModel(s)
+		if s == "" || seen[s] {
+			continue
+		}
+		seen[s] = true
+		out = append(out, s)
+	}
+	return out
+}
+
+func trimModel(s string) string {
+	for len(s) > 0 && (s[0] == ' ' || s[0] == '\t') {
+		s = s[1:]
+	}
+	for len(s) > 0 && (s[len(s)-1] == ' ' || s[len(s)-1] == '\t') {
+		s = s[:len(s)-1]
+	}
+	return s
+}
+
+func isDelta(ev Event) bool {
+	return ev.Kind == KindTextDelta || ev.Kind == KindReasoningDelta || ev.Kind == KindToolCallDelta
+}
+
+func serverError(ev Event) bool {
+	return ev.HTTPStatus >= 500 && ev.HTTPStatus <= 599
+}
@@ -0,0 +1,74 @@
+package llm
+
+import (
+	"context"
+	"testing"
+)
+
+func TestFallbackSwitchesOn5xxBeforeOutput(t *testing.T) {
+	var switched string
+	base := func(ctx context.Context, model string, messages []Message, tools []ToolDecl, opts StreamOptions) <-chan Event {
+		ch := make(chan Event, 2)
+		go func() {
+			defer close(ch)
+			if model == "primary" {
+				ch <- Event{Kind: KindError, HTTPStatus: 503, ErrorMessage: "provider status 503", Message: Message{Role: RoleAssistant, StopReason: StopError}}
+				return
+			}
+			ch <- Event{Kind: KindTextDelta, Text: "ok"}
+			ch <- Event{Kind: KindMessageEnd, Message: Message{Role: RoleAssistant, Content: "ok", StopReason: StopEnd}}
+		}()
+		return ch
+	}
+	stream := WithFallback(base, []string{"backup"}, func(from, to string) {
+		switched = from + "->" + to
+	})
+	var text string
+	var sawError bool
+	for ev := range stream(context.Background(), "primary", nil, nil, StreamOptions{}) {
+		if ev.Kind == KindTextDelta {
+			text += ev.Text
+		}
+		if ev.Kind == KindError {
+			sawError = true
+		}
+	}
+	if switched != "primary->backup" || text != "ok" || sawError {
+		t.Fatalf("switched %q text %q error %v", switched, text, sawError)
+	}
+}
+
+func TestFallbackStaysAfterFirstDelta(t *testing.T) {
+	calls := 0
+	base := func(ctx context.Context, model string, messages []Message, tools []ToolDecl, opts StreamOptions) <-chan Event {
+		calls++
+		ch := make(chan Event, 2)
+		go func() {
+			defer close(ch)
+			ch <- Event{Kind: KindTextDelta, Text: "partial"}
+			ch <- Event{Kind: KindError, HTTPStatus: 500, Message: Message{Role: RoleAssistant, StopReason: StopError}}
+		}()
+		return ch
+	}
+	stream := WithFallback(base, []string{"backup"}, func(from, to string) {
+		t.Fatal("switched after output")
+	})
+	n := 0
+	for range stream(context.Background(), "primary", nil, nil, StreamOptions{}) {
+		n++
+	}
+	if calls != 1 || n != 2 {
+		t.Fatalf("calls %d events %d", calls, n)
+	}
+}
+
+func TestEmptyFallbackIsIdentity(t *testing.T) {
+	base := func(ctx context.Context, model string, messages []Message, tools []ToolDecl, opts StreamOptions) <-chan Event {
+		ch := make(chan Event)
+		close(ch)
+		return ch
+	}
+	if WithFallback(base, nil, nil) == nil {
+		t.Fatal("nil stream")
+	}
+}
@@ -0,0 +1,60 @@
+package llm
+
+// PrepareMessages 按 endpoint 整形消息。
+// merge 合并相邻、且不携带工具调用的同角色消息。
+// strict 在此之后把第一条之后的 system 收成 user，再合并，满足严格交替。
+// 工具消息和带 tool_calls 的助手消息保持原样，避免拆开调用与结果。
+func PrepareMessages(msgs []Message, shape string) []Message {
+	switch shape {
+	case "", "merge", "strict":
+	default:
+		return msgs
+	}
+	out := mergeSameRole(msgs)
+	if shape == "strict" {
+		out = foldExtraSystems(out)
+		out = mergeSameRole(out)
+	}
+	return out
+}
+
+func mergeable(m Message) bool {
+	return m.Role != RoleTool && len(m.ToolCalls) == 0 && m.ToolCallID == ""
+}
+
+func mergeSameRole(msgs []Message) []Message {
+	if len(msgs) == 0 {
+		return msgs
+	}
+	out := make([]Message, 0, len(msgs))
+	for _, m := range msgs {
+		if n := len(out); n > 0 && out[n-1].Role == m.Role && mergeable(out[n-1]) && mergeable(m) {
+			if out[n-1].Content != "" && m.Content != "" {
+				out[n-1].Content += "\n\n" + m.Content
+			} else {
+				out[n-1].Content += m.Content
+			}
+			continue
+		}
+		out = append(out, m)
+	}
+	return out
+}
+
+func foldExtraSystems(msgs []Message) []Message {
+	seenSystem := false
+	out := make([]Message, 0, len(msgs))
+	for _, m := range msgs {
+		if m.Role == RoleSystem {
+			if seenSystem {
+				m.Role = RoleUser
+				if m.Content != "" {
+					m.Content = "[system]\n" + m.Content
+				}
+			}
+			seenSystem = true
+		}
+		out = append(out, m)
+	}
+	return out
+}
@@ -0,0 +1,38 @@
+package llm
+
+import "testing"
+
+func TestPrepareMergesAdjacentSystems(t *testing.T) {
+	out := PrepareMessages([]Message{
+		{Role: RoleSystem, Content: "a"},
+		{Role: RoleSystem, Content: "b"},
+		{Role: RoleUser, Content: "hi"},
+	}, "merge")
+	if len(out) != 2 || out[0].Content != "a\n\nb" {
+		t.Fatalf("merged = %#v", out)
+	}
+}
+
+func TestPrepareStrictFoldsLaterSystem(t *testing.T) {
+	out := PrepareMessages([]Message{
+		{Role: RoleSystem, Content: "base"},
+		{Role: RoleUser, Content: "hi"},
+		{Role: RoleSystem, Content: "note"},
+	}, "strict")
+	if len(out) != 2 {
+		t.Fatalf("len = %d %#v", len(out), out)
+	}
+	if out[1].Role != RoleUser || out[1].Content != "hi\n\n[system]\nnote" {
+		t.Fatalf("folded = %#v", out[1])
+	}
+}
+
+func TestPrepareKeepsToolPairs(t *testing.T) {
+	out := PrepareMessages([]Message{
+		{Role: RoleAssistant, Content: "call", ToolCalls: []ToolCall{{ID: "1", Name: "t"}}},
+		{Role: RoleTool, ToolCallID: "1", Content: "ok"},
+	}, "merge")
+	if len(out) != 2 || out[1].Role != RoleTool {
+		t.Fatalf("tools = %#v", out)
+	}
+}
@@ -0,0 +1,575 @@
+package llm
+
+import (
+	"context"
+	"encoding/json"
+	"fmt"
+	"math/rand"
+	"net/http"
+	"os"
+	"strconv"
+	"strings"
+	"time"
+)
+
+// Thinking levels, from off to max. Unknown values are rejected by NormalizeThinking.
+var ThinkingLevels = []string{"off", "minimal", "low", "medium", "high", "xhigh", "max"}
+
+// NormalizeThinking returns the canonical level. Empty stays empty.
+func NormalizeThinking(level string) (string, error) {
+	if level == "" {
+		return "", nil
+	}
+	level = strings.ToLower(strings.TrimSpace(level))
+	for _, known := range ThinkingLevels {
+		if level == known {
+			return level, nil
+		}
+	}
+	return "", fmt.Errorf("llm: unknown thinking level %q", level)
+}
+
+// OpenAIReasoningEffort maps a thinking level onto OpenAI reasoning_effort.
+// off and empty produce no parameter.
+func OpenAIReasoningEffort(level string) string {
+	switch level {
+	case "minimal":
+		return "minimal"
+	case "low":
+		return "low"
+	case "medium":
+		return "medium"
+	case "high", "xhigh", "max":
+		return "high"
+	default:
+		return ""
+	}
+}
+
+// AnthropicThinkingBudget maps a thinking level onto a token budget.
+// off and empty disable thinking.
+func AnthropicThinkingBudget(level string) int {
+	switch level {
+	case "minimal":
+		return 1024
+	case "low":
+		return 2048
+	case "medium":
+		return 4096
+	case "high":
+		return 8192
+	case "xhigh":
+		return 16384
+	case "max":
+		return 32768
+	default:
+		return 0
+	}
+}
+
+// GoogleThinkingBudget maps a thinking level onto Gemini thinkingBudget.
+func GoogleThinkingBudget(level string) int {
+	switch level {
+	case "minimal", "low":
+		return 1024
+	case "medium":
+		return 4096
+	case "high":
+		return 8192
+	case "xhigh", "max":
+		return 24576
+	default:
+		return 0
+	}
+}
+
+// ModelRef is provider/model with an optional :thinking suffix.
+type ModelRef struct {
+	Provider string
+	Model    string
+	Thinking string
+}
+
+// ParseModelRef accepts "model", "provider/model", and "provider/model:level".
+func ParseModelRef(spec string) (ModelRef, error) {
+	spec = strings.TrimSpace(spec)
+	if spec == "" {
+		return ModelRef{}, fmt.Errorf("llm: empty model ref")
+	}
+	thinking := ""
+	if i := strings.LastIndex(spec, ":"); i >= 0 {
+		thinking = spec[i+1:]
+		spec = spec[:i]
+		norm, err := NormalizeThinking(thinking)
+		if err != nil {
+			return ModelRef{}, err
+		}
+		thinking = norm
+	}
+	provider, model := "", spec
+	if i := strings.Index(spec, "/"); i >= 0 {
+		provider = spec[:i]
+		model = spec[i+1:]
+	}
+	if model == "" {
+		return ModelRef{}, fmt.Errorf("llm: model id missing in %q", spec)
+	}
+	return ModelRef{Provider: provider, Model: model, Thinking: thinking}, nil
+}
+
+// Protocol names the wire format a provider speaks.
+type Protocol string
+
+const (
+	ProtocolOpenAIChat      Protocol = "openai-chat"
+	ProtocolOpenAIResponses Protocol = "openai-responses"
+	ProtocolAnthropic       Protocol = "anthropic"
+	ProtocolGoogle          Protocol = "google"
+	ProtocolAzure           Protocol = "azure"
+)
+
+// Provider is one configured endpoint.
+type Provider struct {
+	Name      string   `json:"name"`
+	Protocol  Protocol `json:"protocol"`
+	BaseURL   string   `json:"baseURL"`
+	APIKeyEnv string   `json:"apiKeyEnv,omitempty"`
+	APIKey    string   `json:"-"`
+}
+
+// Presets are OpenAI-compatible or native endpoints that only need an API key.
+func Presets() []Provider {
+	return []Provider{
+		{Name: "anthropic", Protocol: ProtocolOpenAIChat, BaseURL: "https://api.anthropic.com/v1", APIKeyEnv: "ANTHROPIC_API_KEY"},
+		{Name: "openai", Protocol: ProtocolOpenAIChat, BaseURL: "https://api.openai.com/v1", APIKeyEnv: "OPENAI_API_KEY"},
+		{Name: "openai-responses", Protocol: ProtocolOpenAIResponses, BaseURL: "https://api.openai.com/v1", APIKeyEnv: "OPENAI_API_KEY"},
+		{Name: "google", Protocol: ProtocolOpenAIChat, BaseURL: "https://generativelanguage.googleapis.com/v1beta/openai", APIKeyEnv: "GEMINI_API_KEY"},
+		{Name: "azure", Protocol: ProtocolAzure, BaseURL: "", APIKeyEnv: "AZURE_OPENAI_API_KEY"},
+		{Name: "deepseek", Protocol: ProtocolOpenAIChat, BaseURL: "https://api.deepseek.com/v1", APIKeyEnv: "DEEPSEEK_API_KEY"},
+		{Name: "minimax", Protocol: ProtocolOpenAIChat, BaseURL: "https://api.minimaxi.com/v1", APIKeyEnv: "MINIMAX_API_KEY"},
+		{Name: "openrouter", Protocol: ProtocolOpenAIChat, BaseURL: "https://openrouter.ai/api/v1", APIKeyEnv: "OPENROUTER_API_KEY"},
+		{Name: "groq", Protocol: ProtocolOpenAIChat, BaseURL: "https://api.groq.com/openai/v1", APIKeyEnv: "GROQ_API_KEY"},
+	}
+}
+
+// modelsFile is the on-disk shape of models.json.
+type modelsFile struct {
+	Providers []Provider `json:"providers"`
+}
+
+// LoadModelsJSON reads custom providers. Missing file yields an empty list.
+func LoadModelsJSON(path string) ([]Provider, error) {
+	raw, err := os.ReadFile(path)
+	if err != nil {
+		if os.IsNotExist(err) {
+			return nil, nil
+		}
+		return nil, err
+	}
+	var file modelsFile
+	if err := json.Unmarshal(raw, &file); err != nil {
+		return nil, err
+	}
+	for i := range file.Providers {
+		if file.Providers[i].APIKey == "" && file.Providers[i].APIKeyEnv != "" {
+			file.Providers[i].APIKey = os.Getenv(file.Providers[i].APIKeyEnv)
+		}
+	}
+	return file.Providers, nil
+}
+
+// Lookup finds a preset or a custom provider by name.
+func Lookup(name string, custom []Provider) (Provider, bool) {
+	for _, p := range custom {
+		if p.Name == name {
+			return p, true
+		}
+	}
+	for _, p := range Presets() {
+		if p.Name == name {
+			if p.APIKey == "" {
+				p.APIKey = os.Getenv(p.APIKeyEnv)
+			}
+			return p, true
+		}
+	}
+	return Provider{}, false
+}
+
+// ClientFor 返回会说该协议的客户端。responses、azure 以及原生 Anthropic/Google
+// 协议当前没有接到客户端上，返回明确错误而不是发到错误的路径。
+func ClientFor(p Provider) (*Client, error) {
+	switch p.Protocol {
+	case ProtocolOpenAIChat:
+		return &Client{BaseURL: p.BaseURL, APIKey: p.APIKey}, nil
+	default:
+		return nil, fmt.Errorf("llm: protocol %q is not implemented", p.Protocol)
+	}
+}
+
+func messageContent(m Message) any {
+	if len(m.Images) == 0 {
+		return m.Content
+	}
+	parts := []map[string]any{}
+	if m.Content != "" {
+		parts = append(parts, map[string]any{"type": "text", "text": m.Content})
+	}
+	for _, img := range m.Images {
+		url := img.URL
+		if url == "" && img.Data != "" {
+			mime := img.MIME
+			if mime == "" {
+				mime = "image/png"
+			}
+			url = "data:" + mime + ";base64," + img.Data
+		}
+		parts = append(parts, map[string]any{
+			"type":      "image_url",
+			"image_url": map[string]any{"url": url},
+		})
+	}
+	return parts
+}
+
+// BuildBody returns the JSON body for one native protocol.
+func BuildBody(protocol Protocol, model string, messages []Message, tools []ToolDecl, opts StreamOptions) ([]byte, error) {
+	switch protocol {
+	case ProtocolOpenAIChat, ProtocolAzure:
+		body := map[string]any{
+			"model":    model,
+			"messages": toProviderMessages(messages),
+			"stream":   true,
+		}
+		if opts.MaxTokens > 0 {
+			body["max_tokens"] = opts.MaxTokens
+		}
+		if effort := OpenAIReasoningEffort(opts.ThinkingLevel); effort != "" {
+			body["reasoning_effort"] = effort
+		}
+		if len(tools) > 0 {
+			body["tools"] = toProviderTools(tools)
+		}
+		return json.Marshal(body)
+	case ProtocolOpenAIResponses:
+		body := map[string]any{
+			"model":  model,
+			"input":  responsesInput(messages),
+			"stream": true,
+		}
+		if opts.MaxTokens > 0 {
+			body["max_output_tokens"] = opts.MaxTokens
+		}
+		if effort := OpenAIReasoningEffort(opts.ThinkingLevel); effort != "" {
+			body["reasoning"] = map[string]any{"effort": effort}
+		}
+		return json.Marshal(body)
+	// ProtocolAnthropic 与 ProtocolGoogle 只供离线组包。
+	// 当前 Client 只发 OpenAI chat completions，这两支没有接到客户端上。
+	// 原生协议是后续阶段的事；预设表里的 anthropic / google 已改走兼容端点。
+	case ProtocolAnthropic:
+		sys, msgs := anthropicMessages(messages)
+		body := map[string]any{
+			"model":      model,
+			"messages":   msgs,
+			"max_tokens": opts.MaxTokens,
+			"stream":     true,
+		}
+		// Anthropic 要求 max_tokens。0 表示调用方不限制：用足够大的默认值，
+		// 避免 4096 把长回复截断。真正的上限由调用方 StreamOptions.MaxTokens 决定。
+		if body["max_tokens"] == 0 {
+			body["max_tokens"] = 128000
+		}
+		if sys != "" {
+			body["system"] = sys
+		}
+		if budget := AnthropicThinkingBudget(opts.ThinkingLevel); budget > 0 {
+			body["thinking"] = map[string]any{"type": "enabled", "budget_tokens": budget}
+		}
+		return json.Marshal(body)
+	case ProtocolGoogle:
+		body := map[string]any{
+			"contents": googleContents(messages),
+		}
+		if budget := GoogleThinkingBudget(opts.ThinkingLevel); budget > 0 {
+			body["generationConfig"] = map[string]any{
+				"thinkingConfig": map[string]any{"thinkingBudget": budget},
+			}
+		}
+		return json.Marshal(body)
+	default:
+		return nil, fmt.Errorf("llm: unknown protocol %q", protocol)
+	}
+}
+
+func responsesInput(messages []Message) []map[string]any {
+	out := make([]map[string]any, 0, len(messages))
+	for _, m := range messages {
+		role := string(m.Role)
+		if role == "" {
+			role = "user"
+		}
+		out = append(out, map[string]any{
+			"role":    role,
+			"content": messageContent(m),
+		})
+	}
+	return out
+}
+
+func anthropicMessages(messages []Message) (string, []map[string]any) {
+	var system strings.Builder
+	var out []map[string]any
+	for _, m := range messages {
+		if m.Role == RoleSystem {
+			if system.Len() > 0 {
+				system.WriteString("\n")
+			}
+			system.WriteString(m.Content)
+			continue
+		}
+		role := "user"
+		if m.Role == RoleAssistant {
+			role = "assistant"
+		}
+		content := []map[string]any{}
+		if m.Content != "" {
+			content = append(content, map[string]any{"type": "text", "text": m.Content})
+		}
+		for _, img := range m.Images {
+			content = append(content, map[string]any{
+				"type": "image",
+				"source": map[string]any{
+					"type":       "base64",
+					"media_type": firstNonEmpty(img.MIME, "image/png"),
+					"data":       img.Data,
+				},
+			})
+		}
+		out = append(out, map[string]any{"role": role, "content": content})
+	}
+	return system.String(), out
+}
+
+func googleContents(messages []Message) []map[string]any {
+	var out []map[string]any
+	for _, m := range messages {
+		if m.Role == RoleSystem {
+			continue
+		}
+		role := "user"
+		if m.Role == RoleAssistant {
+			role = "model"
+		}
+		parts := []map[string]any{}
+		if m.Content != "" {
+			parts = append(parts, map[string]any{"text": m.Content})
+		}
+		for _, img := range m.Images {
+			parts = append(parts, map[string]any{
+				"inlineData": map[string]any{
+					"mimeType": firstNonEmpty(img.MIME, "image/png"),
+					"data":     img.Data,
+				},
+			})
+		}
+		out = append(out, map[string]any{"role": role, "parts": parts})
+	}
+	return out
+}
+
+func firstNonEmpty(a, b string) string {
+	if a != "" {
+		return a
+	}
+	return b
+}
+
+const (
+	retryBaseDelay = 200 * time.Millisecond
+	maxRetryAfter  = 30 * time.Second
+)
+
+// overflowNeedles are substrings that mean the prompt does not fit.
+// Matching is case-insensitive. A bare "context" is not enough.
+var overflowNeedles = []string{
+	"status 413",
+	"payload too large",
+	"context_length_exceeded",
+	"context length",
+	"maximum context",
+	"context window",
+	"prompt is too long",
+	"exceeds the maximum",
+	"exceeds max",
+	"too many tokens",
+	"input is too long",
+	"input length",
+	"tokens exceed",
+	"上下文",
+	"超出最大",
+	"输入过长",
+	"提示过长",
+	"请求过长",
+	"请求体过大",
+}
+
+// ContextOverflow reports provider errors that ask the caller to compact.
+func ContextOverflow(message string) bool {
+	m := strings.ToLower(message)
+	for _, needle := range overflowNeedles {
+		if strings.Contains(m, strings.ToLower(needle)) {
+			return true
+		}
+	}
+	return false
+}
+
+// CapacityLimited reports an in-stream overload that is worth retrying
+// before any delta has been forwarded.
+func CapacityLimited(message string) bool {
+	m := strings.ToLower(message)
+	return strings.Contains(m, "overloaded") || strings.Contains(m, "at capacity") || strings.Contains(m, "too many requests") || strings.Contains(m, "rate limit")
+}
+
+// StatusRetryable is true for timeouts, rate limits, server errors, and 529.
+func StatusRetryable(code int) bool {
+	return code == http.StatusRequestTimeout || code == http.StatusTooManyRequests || code == 529 || code >= 500
+}
+
+// ParseRetryAfter reads a delay in seconds or an HTTP date.
+// The returned duration is capped. ok is false when the header is missing or garbage.
+func ParseRetryAfter(raw string, now time.Time) (time.Duration, bool) {
+	raw = strings.TrimSpace(raw)
+	if raw == "" {
+		return 0, false
+	}
+	if secs, err := strconv.Atoi(raw); err == nil {
+		if secs < 0 {
+			return 0, false
+		}
+		return capRetryAfter(time.Duration(secs) * time.Second), true
+	}
+	when, err := http.ParseTime(raw)
+	if err != nil {
+		return 0, false
+	}
+	d := when.Sub(now)
+	if d < 0 {
+		d = 0
+	}
+	return capRetryAfter(d), true
+}
+
+func capRetryAfter(d time.Duration) time.Duration {
+	if d > maxRetryAfter {
+		return maxRetryAfter
+	}
+	return d
+}
+
+func retryBackoff(extraTry int) time.Duration {
+	if extraTry < 1 {
+		extraTry = 1
+	}
+	shift := extraTry - 1
+	if shift > 4 {
+		shift = 4
+	}
+	delay := retryBaseDelay * time.Duration(1<<uint(shift))
+	jitter := time.Duration(rand.Int63n(int64(delay/2) + 1))
+	return delay + jitter
+}
+
+func forwardedDelta(ev Event) bool {
+	switch ev.Kind {
+	case KindTextDelta, KindReasoningDelta, KindToolCallDelta:
+		return true
+	default:
+		return false
+	}
+}
+
+// WithRetry re-invokes base on a retryable terminal error, up to attempts
+// extra tries. Events are forwarded as they arrive. A retry happens only when
+// no delta has been forwarded yet. onEvent receives auto_retry_start and
+// auto_retry_end. Context-overflow errors are returned immediately.
+// attempts <= 0 disables extra tries.
+func WithRetry(base StreamFunc, attempts int, onEvent func(kind, detail string)) StreamFunc {
+	if attempts < 0 {
+		attempts = 0
+	}
+	return func(ctx context.Context, model string, messages []Message, tools []ToolDecl, opts StreamOptions) <-chan Event {
+		out := make(chan Event)
+		go func() {
+			defer close(out)
+			headerWait := time.Duration(0)
+			useHeader := false
+			for try := 0; try <= attempts; try++ {
+				if try > 0 {
+					delay := retryBackoff(try)
+					if useHeader {
+						delay = headerWait
+					}
+					detail := fmt.Sprintf("attempt %d", try+1)
+					if onEvent != nil {
+						onEvent("auto_retry_start", detail)
+					}
+					select {
+					case <-ctx.Done():
+						emitAbort(out, "", ctx.Err())
+						return
+					case <-time.After(delay):
+					}
+				}
+				useHeader = false
+				forwarded := false
+				var terminal *Event
+				for ev := range base(ctx, model, messages, tools, opts) {
+					if terminal != nil {
+						continue
+					}
+					if ev.Kind == KindError {
+						cp := ev
+						terminal = &cp
+						continue
+					}
+					if forwardedDelta(ev) {
+						forwarded = true
+					}
+					select {
+					case <-ctx.Done():
+						return
+					case out <- ev:
+					}
+					if ev.Kind == KindMessageEnd {
+						if try > 0 && onEvent != nil {
+							onEvent("auto_retry_end", fmt.Sprintf("attempt %d", try+1))
+						}
+						return
+					}
+				}
+				if terminal == nil {
+					return
+				}
+				overflow := ContextOverflow(terminal.ErrorMessage)
+				if forwarded || overflow || !terminal.Retryable || try == attempts {
+					if try > 0 && onEvent != nil {
+						onEvent("auto_retry_end", fmt.Sprintf("attempt %d", try+1))
+					}
+					select {
+					case <-ctx.Done():
+						return
+					case out <- *terminal:
+					}
+					return
+				}
+				if terminal.HasRetryAfter {
+					useHeader = true
+					headerWait = terminal.RetryAfter
+				}
+			}
+		}()
+		return out
+	}
+}
@@ -0,0 +1,231 @@
+package llm
+
+import (
+	"context"
+	"encoding/json"
+	"net/http"
+	"net/http/httptest"
+	"os"
+	"path/filepath"
+	"strings"
+	"testing"
+	"time"
+)
+
+func TestParseModelRefAndThinking(t *testing.T) {
+	ref, err := ParseModelRef("anthropic/claude-sonnet:high")
+	if err != nil {
+		t.Fatal(err)
+	}
+	if ref.Provider != "anthropic" || ref.Model != "claude-sonnet" || ref.Thinking != "high" {
+		t.Fatalf("%+v", ref)
+	}
+	if _, err := ParseModelRef("openai/gpt:nope"); err == nil {
+		t.Fatal("unknown thinking level must fail")
+	}
+	if AnthropicThinkingBudget("high") != 8192 || OpenAIReasoningEffort("off") != "" {
+		t.Fatal("thinking map")
+	}
+	p, ok := Lookup("groq", nil)
+	if !ok || !strings.Contains(p.BaseURL, "groq") || p.Protocol != ProtocolOpenAIChat {
+		t.Fatalf("preset: %+v", p)
+	}
+}
+
+func TestModelsJSONAndBodies(t *testing.T) {
+	dir := t.TempDir()
+	path := filepath.Join(dir, "models.json")
+	raw := []byte(`{"providers":[{"name":"local","protocol":"openai-chat","baseURL":"http://127.0.0.1:9","apiKeyEnv":"UNUSED_KEY"}]}`)
+	if err := os.WriteFile(path, raw, 0o644); err != nil {
+		t.Fatal(err)
+	}
+	got, err := LoadModelsJSON(path)
+	if err != nil || len(got) != 1 || got[0].Name != "local" {
+		t.Fatalf("%+v %v", got, err)
+	}
+	msgs := []Message{{
+		Role: RoleUser, Content: "see",
+		Images: []Image{{MIME: "image/png", Data: "aaaa"}},
+	}}
+	body, err := BuildBody(ProtocolAnthropic, "claude", append([]Message{{Role: RoleSystem, Content: "sys"}}, msgs...), nil, StreamOptions{ThinkingLevel: "low"})
+	if err != nil {
+		t.Fatal(err)
+	}
+	if !strings.Contains(string(body), "budget_tokens") || !strings.Contains(string(body), "base64") {
+		t.Fatalf("anthropic body: %s", body)
+	}
+	gbody, err := BuildBody(ProtocolGoogle, "gemini", msgs, nil, StreamOptions{ThinkingLevel: "max"})
+	if err != nil || !strings.Contains(string(gbody), "inlineData") || !strings.Contains(string(gbody), "thinkingBudget") {
+		t.Fatalf("google: %s %v", gbody, err)
+	}
+	rbody, err := BuildBody(ProtocolOpenAIResponses, "gpt", msgs, nil, StreamOptions{ThinkingLevel: "medium", MaxTokens: 20})
+	if err != nil || !strings.Contains(string(rbody), "max_output_tokens") || !strings.Contains(string(rbody), "image_url") {
+		t.Fatalf("responses: %s %v", rbody, err)
+	}
+	var decoded map[string]any
+	if err := json.Unmarshal(rbody, &decoded); err != nil {
+		t.Fatal(err)
+	}
+}
+
+func TestPresetsAndClientForChatCompletions(t *testing.T) {
+	want := map[string]string{
+		"anthropic": "https://api.anthropic.com/v1",
+		"google":    "https://generativelanguage.googleapis.com/v1beta/openai",
+	}
+	for _, p := range Presets() {
+		base, ok := want[p.Name]
+		if !ok {
+			continue
+		}
+		if p.Protocol != ProtocolOpenAIChat || p.BaseURL != base {
+			t.Fatalf("%s preset: %+v", p.Name, p)
+		}
+	}
+	for _, name := range []string{"openai-responses", "azure"} {
+		p, ok := Lookup(name, nil)
+		if !ok {
+			t.Fatal(name)
+		}
+		if _, err := ClientFor(p); err == nil {
+			t.Fatalf("%s should fail closed", name)
+		}
+	}
+	var path string
+	srv := httptest.NewServer(http.HandlerFunc(func(w http.ResponseWriter, r *http.Request) {
+		path = r.URL.Path
+		w.Header().Set("Content-Type", "text/event-stream")
+		_, _ = w.Write([]byte("data: [DONE]\n\n"))
+	}))
+	defer srv.Close()
+	client, err := ClientFor(Provider{Protocol: ProtocolOpenAIChat, BaseURL: srv.URL, APIKey: "k"})
+	if err != nil {
+		t.Fatal(err)
+	}
+	for ev := range client.Stream(context.Background(), "m", []Message{{Role: RoleUser, Content: "hi"}}, nil, StreamOptions{}) {
+		if ev.Kind == KindError {
+			t.Fatal(ev.Err)
+		}
+	}
+	if path != "/chat/completions" {
+		t.Fatalf("path %q", path)
+	}
+}
+
+func TestRetryThenSuccess(t *testing.T) {
+	n := 0
+	base := StreamFunc(func(ctx context.Context, model string, messages []Message, tools []ToolDecl, opts StreamOptions) <-chan Event {
+		out := make(chan Event, 2)
+		n++
+		if n == 1 {
+			out <- Event{Kind: KindError, Retryable: true, ErrorMessage: "429"}
+		} else {
+			out <- Event{Kind: KindTextDelta, Text: "ok"}
+			out <- Event{Kind: KindMessageEnd, Message: Message{Role: RoleAssistant, Content: "ok"}}
+		}
+		close(out)
+		return out
+	})
+	var events []string
+	wrapped := WithRetry(base, 2, func(kind, detail string) { events = append(events, kind) })
+	var texts []string
+	for ev := range wrapped(context.Background(), "m", nil, nil, StreamOptions{}) {
+		if ev.Kind == KindTextDelta {
+			texts = append(texts, ev.Text)
+		}
+	}
+	if len(texts) != 1 || texts[0] != "ok" {
+		t.Fatalf("texts %v calls %d", texts, n)
+	}
+	if len(events) == 0 || events[0] != "auto_retry_start" {
+		t.Fatalf("events %v", events)
+	}
+}
+
+func TestContextOverflowPatterns(t *testing.T) {
+	hits := []string{
+		"provider status 413: payload too large",
+		"context_length_exceeded",
+		"prompt is too long",
+		"input length exceeds the maximum",
+		"maximum context length exceeded",
+		"context length overflow",
+		"too many tokens",
+		"输入过长，请缩短上下文",
+		"请求过长",
+		"超出最大长度",
+	}
+	for _, msg := range hits {
+		if !ContextOverflow(msg) {
+			t.Fatalf("want overflow: %s", msg)
+		}
+	}
+	if ContextOverflow("context cancelled") || ContextOverflow("temporary network error") {
+		t.Fatal("unrelated context text must not count as overflow")
+	}
+}
+
+func TestParseRetryAfter(t *testing.T) {
+	now := time.Date(2026, 10, 3, 12, 0, 0, 0, time.UTC)
+	if d, ok := ParseRetryAfter("2", now); !ok || d != 2*time.Second {
+		t.Fatalf("seconds: %v %v", d, ok)
+	}
+	if d, ok := ParseRetryAfter("not-a-date", now); ok || d != 0 {
+		t.Fatalf("garbage should fail: %v %v", d, ok)
+	}
+	if _, ok := ParseRetryAfter("", now); ok {
+		t.Fatal("empty header")
+	}
+	far := now.Add(2 * time.Hour).Format(http.TimeFormat)
+	if d, ok := ParseRetryAfter(far, now); !ok || d != maxRetryAfter {
+		t.Fatalf("date cap: %v %v", d, ok)
+	}
+	past := now.Add(-time.Minute).Format(http.TimeFormat)
+	if d, ok := ParseRetryAfter(past, now); !ok || d != 0 {
+		t.Fatalf("past date: %v %v", d, ok)
+	}
+}
+
+func TestRetryDoesNotRepeatDeltas(t *testing.T) {
+	n := 0
+	base := StreamFunc(func(ctx context.Context, model string, messages []Message, tools []ToolDecl, opts StreamOptions) <-chan Event {
+		out := make(chan Event, 2)
+		n++
+		out <- Event{Kind: KindTextDelta, Text: "a"}
+		out <- Event{Kind: KindError, Retryable: true, ErrorMessage: "overloaded"}
+		close(out)
+		return out
+	})
+	var texts []string
+	var errors int
+	for ev := range WithRetry(base, 2, nil)(context.Background(), "m", nil, nil, StreamOptions{}) {
+		if ev.Kind == KindTextDelta {
+			texts = append(texts, ev.Text)
+		}
+		if ev.Kind == KindError {
+			errors++
+		}
+	}
+	if n != 1 || len(texts) != 1 || texts[0] != "a" || errors != 1 {
+		t.Fatalf("calls=%d texts=%v errors=%d", n, texts, errors)
+	}
+}
+
+func TestContextOverflowIsNotRetried(t *testing.T) {
+	if !ContextOverflow("maximum context length exceeded") {
+		t.Fatal("expected overflow")
+	}
+	n := 0
+	base := StreamFunc(func(ctx context.Context, model string, messages []Message, tools []ToolDecl, opts StreamOptions) <-chan Event {
+		out := make(chan Event, 1)
+		n++
+		out <- Event{Kind: KindError, Retryable: true, ErrorMessage: "context length overflow"}
+		close(out)
+		return out
+	})
+	for range WithRetry(base, 3, nil)(context.Background(), "m", nil, nil, StreamOptions{}) {
+	}
+	if n != 1 {
+		t.Fatalf("overflow must not retry, calls=%d", n)
+	}
+}
@@ -0,0 +1,29 @@
+package llm
+
+import (
+	"context"
+	"strings"
+	"testing"
+)
+
+func TestR2SSEJoinsMultilineData(t *testing.T) {
+	body := strings.Join([]string{
+		`data: {"choices":[{"delta":{"content":"hi"}`,
+		`data: }]}`,
+		``,
+		`data: [DONE]`,
+		``,
+	}, "\n")
+	var got string
+	for ev := range ParseSSE(context.Background(), strings.NewReader(body)) {
+		if ev.Kind == KindTextDelta {
+			got += ev.Text
+		}
+		if ev.Kind == KindError {
+			t.Fatal(ev.Err)
+		}
+	}
+	if got != "hi" {
+		t.Fatalf("joined content %q", got)
+	}
+}
@@ -26,12 +26,19 @@ type sseChoice struct {

 type sseChunk struct {
 	Choices []sseChoice `json:"choices"`
-	Usage   *Usage      `json:"usage"`
+	Usage   *usageWire  `json:"usage"`
 	Error   *struct {
 		Message string `json:"message"`
 	} `json:"error"`
 }

+type usageWire struct {
+	Usage
+	Details struct {
+		Cached int `json:"cached_tokens"`
+	} `json:"prompt_tokens_details"`
+}
+
 // ParseSSE reads an OpenAI-compatible server-sent event body and emits
 // deltas plus one terminal message_end or error event.
 func ParseSSE(ctx context.Context, r io.Reader) <-chan Event {
@@ -49,22 +56,21 @@ func emitParsed(ctx context.Context, r io.Reader, out chan<- Event) {
 	order := []int{}
 	var usage Usage
 	finished := ""
+	sawDone := false
 	sc := bufio.NewScanner(r)
 	sc.Buffer(make([]byte, 0, 64*1024), 4*1024*1024)
-	for sc.Scan() {
+	reader := &sseEventReader{sc: sc}
+	for {
+		data, ok := reader.next()
+		if !ok {
+			break
+		}
 		if ctx.Err() != nil {
 			emitAbort(out, content.String(), ctx.Err())
 			return
 		}
-		line := strings.TrimSpace(sc.Text())
-		if line == "" || strings.HasPrefix(line, ":") {
-			continue
-		}
-		if !strings.HasPrefix(line, "data:") {
-			continue
-		}
-		data := strings.TrimSpace(strings.TrimPrefix(line, "data:"))
 		if data == "[DONE]" {
+			sawDone = true
 			break
 		}
 		var chunk sseChunk
@@ -73,11 +79,16 @@ func emitParsed(ctx context.Context, r io.Reader, out chan<- Event) {
 			return
 		}
 		if chunk.Error != nil && chunk.Error.Message != "" {
-			emitError(out, content.String(), false, chunk.Error.Message)
+			msg := chunk.Error.Message
+			retryable := CapacityLimited(msg) && !ContextOverflow(msg)
+			emitError(out, content.String(), retryable, msg)
 			return
 		}
 		if chunk.Usage != nil {
-			usage = *chunk.Usage
+			usage = chunk.Usage.Usage
+			if chunk.Usage.Details.Cached > 0 {
+				usage.CacheReadTokens = chunk.Usage.Details.Cached
+			}
 		}
 		if len(chunk.Choices) == 0 {
 			continue
@@ -144,6 +155,16 @@ func emitParsed(ctx context.Context, r io.Reader, out chan<- Event) {
 		emitError(out, content.String(), true, err.Error())
 		return
 	}
+	// 取消必须压过正常结束。select 在两个分支都就绪时随机挑选，
+	// 竞态下会把 message_end 发给已经取消的调用方。
+	if ctx.Err() != nil {
+		emitAbort(out, content.String(), ctx.Err())
+		return
+	}
+	// length、空结束原因，或没有 [DONE] 的中断，都不把半截工具调用交给执行层。
+	if !toolFinishExecutable(finished, sawDone) {
+		order = nil
+	}
 	msg := Message{Role: RoleAssistant, Content: content.String(), StopReason: mapFinish(finished, len(order) > 0)}
 	for _, idx := range order {
 		msg.ToolCalls = append(msg.ToolCalls, *calls[idx])
@@ -154,7 +175,75 @@ 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}:
+	}
+}
+
+// sseEventReader 按空行切事件，并把同一事件里的多行 data: 用换行拼回。
+// 已经是完整 JSON 的 data 行单独成事件，这样单行流不用空行也能解析。
+type sseEventReader struct {
+	sc    *bufio.Scanner
+	carry string
+}
+
+func (r *sseEventReader) next() (string, bool) {
+	var parts []string
+	nextPiece := func() (string, bool) {
+		if r.carry != "" {
+			piece := r.carry
+			r.carry = ""
+			return piece, true
+		}
+		if !r.sc.Scan() {
+			return "", false
+		}
+		return r.sc.Text(), true
+	}
+	for {
+		line, ok := nextPiece()
+		if !ok {
+			break
+		}
+		if strings.TrimSpace(line) == "" {
+			if len(parts) == 0 {
+				continue
+			}
+			return strings.Join(parts, "\n"), true
+		}
+		trimmed := strings.TrimSpace(line)
+		if strings.HasPrefix(trimmed, ":") {
+			continue
+		}
+		if !strings.HasPrefix(trimmed, "data:") {
+			continue
+		}
+		piece := strings.TrimSpace(strings.TrimPrefix(trimmed, "data:"))
+		if len(parts) > 0 && json.Valid([]byte(strings.Join(parts, "\n"))) {
+			r.carry = line
+			return strings.Join(parts, "\n"), true
+		}
+		parts = append(parts, piece)
+	}
+	if len(parts) > 0 {
+		return strings.Join(parts, "\n"), true
+	}
+	return "", false
+}
+
+// toolFinishExecutable 只在供应商明确以工具调用或正常 stop 结束时执行。
+// length、缺失的结束原因，以及没有 [DONE] 的断开，都视为未完成。
+func toolFinishExecutable(reason string, sawDone bool) bool {
+	switch reason {
+	case "tool_calls", "function_call":
+		return true
+	case "stop":
+		return true
+	case "length":
+		return false
+	case "":
+		return false
+	default:
+		return sawDone && reason != "error" && reason != "content_filter"
 	}
 }

@@ -164,7 +253,9 @@ func mapFinish(reason string, hasTools bool) StopReason {
 		return StopToolCalls
 	case "content_filter", "error":
 		return StopError
-	case "stop", "length", "":
+	case "length":
+		return StopEnd
+	case "stop", "":
 		if hasTools {
 			return StopToolCalls
 		}
@@ -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) {
@@ -70,6 +93,51 @@ func TestParseSSEToolCallIncrements(t *testing.T) {
 	}
 }

+func TestParseSSELengthDropsUnfinishedToolCall(t *testing.T) {
+	body := strings.Join([]string{
+		`data: {"choices":[{"delta":{"tool_calls":[{"index":0,"id":"call_1","function":{"name":"read_file","arguments":"{\"path\":"}}]}}]}`,
+		`data: {"choices":[{"delta":{},"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.Message.StopReason != StopEnd || len(end.Message.ToolCalls) != 0 {
+		t.Fatalf("truncated tool call must not be executable: %#v", end.Message)
+	}
+	if end.FinishReason != "length" {
+		t.Fatalf("finish %q", end.FinishReason)
+	}
+}
+
+func TestParseSSEUncleanEndDropsToolCall(t *testing.T) {
+	body := `data: {"choices":[{"delta":{"tool_calls":[{"index":0,"id":"call_1","function":{"name":"read_file","arguments":"{\"path\":"}}]}}]}`
+	var end Event
+	for ev := range ParseSSE(context.Background(), strings.NewReader(body)) {
+		if ev.Kind == KindMessageEnd {
+			end = ev
+		}
+	}
+	if end.Kind != KindMessageEnd || end.Message.StopReason != StopEnd || len(end.Message.ToolCalls) != 0 {
+		t.Fatalf("unclean stream must not execute tools: %#v", end)
+	}
+}
+
+func TestParseSSECapacityErrorIsRetryable(t *testing.T) {
+	body := "data: {\"error\":{\"message\":\"server is at capacity\"}}\n\n"
+	var got Event
+	for ev := range ParseSSE(context.Background(), strings.NewReader(body)) {
+		got = ev
+	}
+	if got.Kind != KindError || !got.Retryable || !CapacityLimited(got.ErrorMessage) {
+		t.Fatalf("event %#v", got)
+	}
+}
+
 func TestParseSSEMalformedIsError(t *testing.T) {
 	body := "data: {not-json}\n\n"
 	var got Event
@@ -3,6 +3,7 @@ package llm
 import (
 	"context"
 	"encoding/json"
+	"time"
 )

 // Role is a transcript role carried into the provider request.
@@ -48,6 +49,7 @@ type Message struct {
 	Role       Role       `json:"role"`
 	Content    string     `json:"content,omitempty"`
 	Name       string     `json:"name,omitempty"`
+	Images     []Image    `json:"images,omitempty"`
 	ToolCallID string     `json:"tool_call_id,omitempty"`
 	ToolCalls  []ToolCall `json:"tool_calls,omitempty"`
 	StopReason StopReason `json:"stop_reason,omitempty"`
@@ -61,11 +63,26 @@ type ToolDecl struct {
 	Parameters  json.RawMessage `json:"parameters"`
 }

-// Usage counts tokens reported by the provider. Zero means the provider omitted usage.
+// Image is one picture attached to a user message.
+type Image struct {
+	MIME string `json:"mime,omitempty"`
+	URL  string `json:"url,omitempty"`
+	Data string `json:"data,omitempty"`
+}
+
+// Usage counts tokens. On one provider response, PromptTokens, CompletionTokens
+// and TotalTokens are that request. On a loop Result they are the billing sum
+// across requests; Last* is only the latest request, for context observation.
+// Zero means the provider omitted usage.
 type Usage struct {
-	PromptTokens     int `json:"prompt_tokens"`
-	CompletionTokens int `json:"completion_tokens"`
-	TotalTokens      int `json:"total_tokens"`
+	PromptTokens         int `json:"prompt_tokens"`
+	CompletionTokens     int `json:"completion_tokens"`
+	TotalTokens          int `json:"total_tokens"`
+	CacheReadTokens      int `json:"cache_read_tokens,omitempty"`
+	CacheWriteTokens     int `json:"cache_write_tokens,omitempty"`
+	LastPromptTokens     int `json:"last_prompt_tokens,omitempty"`
+	LastCompletionTokens int `json:"last_completion_tokens,omitempty"`
+	LastTotalTokens      int `json:"last_total_tokens,omitempty"`
 }

 // Event is one streamed update. Failures are KindError plus a final assistant
@@ -80,9 +97,18 @@ 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
+	// HTTPStatus 是提供方的 HTTP 状态。0 表示这条错误不是 HTTP 响应。
+	HTTPStatus int
+	// RetryAfter is the provider's requested wait. HasRetryAfter distinguishes
+	// a real zero from a missing or unparseable header.
+	RetryAfter    time.Duration
+	HasRetryAfter bool
 }

 // StreamOptions tunes one provider request.
@@ -90,6 +116,11 @@ type StreamOptions struct {
 	MaxTokens int
 	APIBase   string
 	APIKey    string
+	// ThinkingLevel is off, minimal, low, medium, high, xhigh, or max.
+	// Empty leaves the provider default unchanged.
+	ThinkingLevel string
+	// MessageShape is merge, strict, or empty. Empty leaves the transcript unchanged.
+	MessageShape string
 }

 // StreamFunc is the provider boundary. It must return a channel that is closed
@@ -8,7 +8,9 @@ import (
 	"sync"
 	"time"

+	"agent-native/runtime/hooks"
 	"agent-native/runtime/llm"
+	"agent-native/runtime/reminders"
 	"agent-native/runtime/session"
 	"agent-native/runtime/tools"

@@ -18,8 +20,15 @@ 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
@@ -29,23 +38,34 @@ 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.
 type Config struct {
-	Model         string
-	SystemPrompt  string
-	Tools         *tools.Registry
-	Stream        llm.StreamFunc
-	Session       *session.Store
+	Model        string
+	SystemPrompt string
+	Tools        *tools.Registry
+	Stream       llm.StreamFunc
+	Session      *session.Store
+	// Trace 开始一个 span，返回的函数结束它。nil 表示不采集。
+	// 属性里不要放提示词正文。
+	Trace         func(name string, attrs map[string]string) func()
 	MaxTurns      int
 	ToolExecution string
 	GetSteering   func() []llm.Message
 	GetFollowUp   func() []llm.Message
-	OnEvent       func(Event)
+	// OnEvent 观察循环事件。同一次 Run 内不会并发调用
+	// （与 BeforeTool、AfterTool 共用一把锁串行）。
+	OnEvent func(Event)
+	// ThinkingLevel is forwarded on each provider request.
+	ThinkingLevel string
+	// Resolve, when set, overrides Model and ThinkingLevel at the start of each turn.
+	Resolve func() (model, thinking string)
 	// Log 可选。nil 时丢弃日志，不改变循环结果。
 	Log *logsys.Logger
 	// ExecutionID 只用于日志字段，不参与控制流。
@@ -53,6 +73,68 @@ 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。
+	// 同一次 Run 内与 OnEvent、AfterTool 串行，不会并发调用。
+	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。
+	// 同一次 Run 内与 OnEvent、BeforeTool 串行，不会并发调用。
+	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，
+	// 并入最后一条工具结果或用户消息，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。
+	// 续写提示不写进 Session。次数由 MaxLengthContinuations 决定。
+	ContinueOnLength bool
+	// MaxLengthContinuations 是 length 续写的次数上限。
+	// 0 表示默认 3 次；负数表示不限制。
+	MaxLengthContinuations int
+	// ContextWindow 是模型窗口的字符估算。0 表示用保守默认值。
+	// 超过窗口减去 OutputReserve 与 ContextBuffer 时，清理最近若干条之前的旧工具结果。
+	ContextWindow int
+	// OutputReserve 是留给模型输出的字符数。0 表示默认预留。
+	OutputReserve int
+	// ContextBuffer 是窗口估算的余量。0 表示默认余量。
+	ContextBuffer int
+	// KeepToolResults 是清理时保留的最近工具结果条数。0 表示 5。
+	KeepToolResults int
+	// Hooks 是进程内钩子。空总线不改变停止、提醒和拒绝次数。
+	Hooks *hooks.Bus
+	// PermissionAsk 在钩子对一次工具调用表态 ask 时向宿主请求裁决。
+	// nil 时保持旧行为：ask 与 deny 一样按拒绝处理。
+	// 返回 Allow 时工具照常执行；宿主需要「本会话始终允许」时自行记录，
+	// 内核不替宿主保存放行状态。
+	PermissionAsk func(ctx context.Context, call llm.ToolCall) AskVerdict
+}
+
+// AskVerdict 是宿主对一次 ask 决策的裁决。
+type AskVerdict struct {
+	Allow  bool
+	Reason string // 拒绝时写进配对 tool 结果的理由；放行时忽略
+}
+
+// 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。
@@ -69,6 +151,7 @@ type Result struct {
 	Messages   []llm.Message
 	StopReason llm.StopReason
 	Usage      llm.Usage
+	Notice     string
 	Err        error
 	Retryable  bool
 }
@@ -84,6 +167,9 @@ func Run(ctx context.Context, cfg Config, user llm.Message) (Result, error) {
 	if cfg.ToolExecution == "" {
 		cfg.ToolExecution = ExecSequential
 	}
+	serializeHostCallbacks(&cfg)
+	endRun := startSpan(cfg, "invoke_agent", map[string]string{"model": cfg.Model})
+	defer endRun()
 	var messages []llm.Message
 	if cfg.Session != nil {
 		loaded, err := cfg.Session.Load()
@@ -92,6 +178,19 @@ func Run(ctx context.Context, cfg Config, user llm.Message) (Result, error) {
 		}
 		messages = loaded
 	}
+	if err := recoverPending(ctx, cfg); err != nil {
+		return Result{}, err
+	}
+	if cfg.Session != nil {
+		loaded, err := cfg.Session.Load()
+		if err != nil {
+			return Result{}, err
+		}
+		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...)
@@ -104,10 +203,39 @@ func Run(ctx context.Context, cfg Config, user llm.Message) (Result, error) {
 	}
 	messages = append(messages, user)
 	emit(cfg, Event{Type: "agent_start"})
+	cfg.Hooks.Emit(hooks.Call{Event: hooks.SessionStart})
+	if out := cfg.Hooks.Emit(hooks.Call{Event: hooks.UserPromptSubmit, LastText: user.Content}); out.Decision != "" {
+		switch out.Decision {
+		case hooks.Deny, hooks.Block:
+			// 输入护栏拒绝：不进入循环，以理由直接收尾（原文已持久化留痕）。
+			why := out.Reason
+			if why == "" {
+				why = "输入被护栏拒绝"
+			}
+			return Result{Text: why, Messages: messages, StopReason: llm.StopEnd, Notice: why}, nil
+		case hooks.Modify:
+			// 改写后的输入追加为新的用户消息（原文保留，可审计）。
+			if out.Replacement != "" {
+				note := llm.Message{Role: llm.RoleUser, Content: out.Replacement}
+				if err := persist(cfg, note); err != nil {
+					return Result{}, err
+				}
+				messages = append(messages, note)
+				user = note
+			}
+		}
+	}

 	var last llm.Message
 	var usage llm.Usage
 	toolCalls := 0
+	stopContinues := 0
+	lengthContinues := 0
+	lengthNudge := false
+	var repeats repeatState
+	defer func() {
+		cfg.Hooks.Emit(hooks.Call{Event: hooks.SessionEnd, LastText: last.Content, ToolCalls: toolCalls})
+	}()
 	for turn := 0; turn < cfg.MaxTurns; turn++ {
 		if ctx.Err() != nil {
 			cfg.logger().Error("runtime turn aborted", ctx.Err())
@@ -117,7 +245,9 @@ func Run(ctx context.Context, cfg Config, user llm.Message) (Result, error) {
 			return budgetStop(cfg, last, messages, usage, reason), nil
 		}
 		emit(cfg, Event{Type: "turn_start"})
-		end, errEv := consume(ctx, cfg, messages)
+		emit(cfg, Event{Type: "message_start"})
+		request := prepareRequest(cfg, messages, 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
@@ -126,18 +256,37 @@ func Run(ctx context.Context, cfg Config, user llm.Message) (Result, error) {
 		usage.PromptTokens += end.Usage.PromptTokens
 		usage.CompletionTokens += end.Usage.CompletionTokens
 		usage.TotalTokens += end.Usage.TotalTokens
+		usage.LastPromptTokens = end.Usage.PromptTokens
+		usage.LastCompletionTokens = end.Usage.CompletionTokens
+		usage.LastTotalTokens = end.Usage.TotalTokens
 		if err := persist(cfg, last); err != nil {
 			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 reason := budgetExceeded(cfg.Budget, usage, toolCalls, time.Now()); reason != "" {
+				var err error
+				messages, err = appendSkippedToolResults(cfg, last.ToolCalls, messages)
+				if err != nil {
+					return Result{}, err
+				}
+				return budgetStop(cfg, last, messages, usage, reason), nil
+			}
 			if cfg.Budget != nil && cfg.Budget.MaxToolCalls > 0 && toolCalls+len(last.ToolCalls) > cfg.Budget.MaxToolCalls {
 				reason := fmt.Sprintf("budget_exceeded: tool calls %d+%d > max %d", toolCalls, len(last.ToolCalls), cfg.Budget.MaxToolCalls)
+				var err error
+				messages, err = appendSkippedToolResults(cfg, last.ToolCalls, messages)
+				if err != nil {
+					return Result{}, err
+				}
 				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 {
@@ -153,10 +302,25 @@ func Run(ctx context.Context, cfg Config, user llm.Message) (Result, error) {
 					messages = append(messages, m)
 				}
 			}
+			cfg.Hooks.Emit(hooks.Call{Event: hooks.PostToolBatch, Turn: turn, ToolCalls: len(last.ToolCalls)})
 			emit(cfg, Event{Type: "turn_end", Message: last})
 			continue
 		}
 		emit(cfg, Event{Type: "turn_end", Message: last})
+		if cfg.ContinueOnLength && end.FinishReason == "length" && len(last.ToolCalls) == 0 {
+			lengthCap := cfg.MaxLengthContinuations
+			if lengthCap == 0 {
+				lengthCap = maxLengthContinuations
+			}
+			if lengthCap > 0 && lengthContinues >= lengthCap {
+				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 {
@@ -166,9 +330,30 @@ func Run(ctx context.Context, cfg Config, user llm.Message) (Result, error) {
 			}
 			continue
 		}
-		emit(cfg, Event{Type: "agent_end", Message: last})
-		emitUsage(cfg, usage)
-		return Result{Text: last.Content, Messages: messages, StopReason: last.StopReason, Usage: usage}, nil
+		if decision, hooked := stopDecision(cfg, last, stopContinues, toolCalls); hooked {
+			if decision.Continue && stopContinues < maxStopContinuations {
+				stopContinues++
+				if decision.Replacement != "" && len(messages) > 0 {
+					messages[len(messages)-1].Content = decision.Replacement
+					last.Content = decision.Replacement
+					note := llm.Message{Role: llm.RoleUser, Content: decision.Replacement}
+					if err := persist(cfg, note); err != nil {
+						return Result{}, err
+					}
+					messages = append(messages, note)
+				}
+				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)
@@ -216,15 +401,29 @@ func systemText(cfg Config) string {
 }

 func consume(ctx context.Context, cfg Config, messages []llm.Message) (llm.Event, error) {
+	endSpan := startSpan(cfg, "chat", map[string]string{"model": cfg.Model})
+	defer endSpan()
 	var decls []llm.ToolDecl
 	if cfg.Tools != nil {
 		for _, d := range cfg.Tools.Decls() {
 			decls = append(decls, llm.ToolDecl{Name: d.Name, Description: d.Description, Parameters: d.Parameters})
 		}
 	}
+	model := cfg.Model
+	opts := llm.StreamOptions{ThinkingLevel: cfg.ThinkingLevel}
+	if cfg.Resolve != nil {
+		if m, th := cfg.Resolve(); m != "" || th != "" {
+			if m != "" {
+				model = m
+			}
+			if th != "" {
+				opts.ThinkingLevel = th
+			}
+		}
+	}
 	var end llm.Event
 	sawEnd := false
-	for ev := range cfg.Stream(ctx, cfg.Model, messages, decls, llm.StreamOptions{}) {
+	for ev := range cfg.Stream(ctx, model, messages, decls, opts) {
 		switch ev.Kind {
 		case llm.KindTextDelta:
 			emit(cfg, Event{Type: "text_delta", Text: ev.Text})
@@ -245,12 +444,66 @@ func consume(ctx context.Context, cfg Config, messages []llm.Message) (llm.Event
 	return end, nil
 }

+const toolBudgetSkipContent = "not executed: tool budget exceeded"
+
+// appendSkippedToolResults 为未执行的 tool_call 各写一条错误结果，保持调用与结果成对。
+// 不发工具事件：这些调用没有进入执行。
+func appendSkippedToolResults(cfg Config, calls []llm.ToolCall, messages []llm.Message) ([]llm.Message, error) {
+	for _, call := range calls {
+		msg := llm.Message{
+			Role:       llm.RoleTool,
+			Name:       call.Name,
+			ToolCallID: call.ID,
+			Content:    toolBudgetSkipContent,
+			IsError:    true,
+		}
+		if err := persist(cfg, msg); err != nil {
+			return nil, err
+		}
+		messages = append(messages, msg)
+	}
+	return messages, nil
+}
+
+// serializeHostCallbacks 让同一次 Run 的宿主回调不会并发进入。
+func serializeHostCallbacks(cfg *Config) {
+	var mu sync.Mutex
+	if cfg.OnEvent != nil {
+		inner := cfg.OnEvent
+		cfg.OnEvent = func(ev Event) {
+			mu.Lock()
+			defer mu.Unlock()
+			inner(ev)
+		}
+	}
+	if cfg.BeforeTool != nil {
+		inner := cfg.BeforeTool
+		cfg.BeforeTool = func(call llm.ToolCall) error {
+			mu.Lock()
+			defer mu.Unlock()
+			return inner(call)
+		}
+	}
+	if cfg.AfterTool != nil {
+		inner := cfg.AfterTool
+		cfg.AfterTool = func(call llm.ToolCall, res llm.Message) {
+			mu.Lock()
+			defer mu.Unlock()
+			inner(call, res)
+		}
+	}
+}
+
 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)
 	}
 	out := make([]llm.Message, 0, len(assistant.ToolCalls))
 	for _, call := range assistant.ToolCalls {
+		if err := ctx.Err(); err != nil {
+			out = append(out, cancelledToolMessage(call, err))
+			continue
+		}
 		out = append(out, runOne(ctx, cfg, call))
 	}
 	return out
@@ -270,20 +523,314 @@ func executeParallel(ctx context.Context, cfg Config, assistant llm.Message) []l
 	return out
 }

+func cancelledToolMessage(call llm.ToolCall, err error) llm.Message {
+	return llm.Message{
+		Role:       llm.RoleTool,
+		Name:       call.Name,
+		ToolCallID: call.ID,
+		Content:    "not executed: " + err.Error(),
+		IsError:    true,
+	}
+}
+
+func rejectIncompleteToolCall(call llm.ToolCall) string {
+	if strings.TrimSpace(call.Name) == "" {
+		return "工具调用未执行：名称不完整"
+	}
+	args := strings.TrimSpace(call.Arguments)
+	if !json.Valid([]byte(args)) || !strings.HasPrefix(args, "{") {
+		return "工具调用未执行：参数不是完整的 JSON 对象"
+	}
+	return ""
+}
+
+func stopDecision(cfg Config, last llm.Message, continuations, toolCalls int) (StopDecision, bool) {
+	if cfg.Hooks != nil {
+		out := cfg.Hooks.Emit(hooks.Call{
+			Event: hooks.Stop, LastText: last.Content, Continuations: continuations, ToolCalls: toolCalls,
+		})
+		if out.Decision != "" {
+			return StopDecision{
+				Continue: out.Decision == hooks.Block, Reason: out.Reason,
+				Replacement: out.Replacement, Notice: out.Notice,
+			}, true
+		}
+	}
+	if cfg.Stop != nil {
+		return cfg.Stop(last, StopInfo{Continuations: continuations, ToolCalls: toolCalls}), true
+	}
+	return StopDecision{}, false
+}
+
+func gateTool(ctx context.Context, cfg Config, call llm.ToolCall) (llm.ToolCall, llm.Message, bool) {
+	if cfg.Hooks == nil {
+		return call, llm.Message{}, false
+	}
+	out := cfg.Hooks.Emit(hooks.Call{Event: hooks.PreToolUse, Tool: call.Name, Args: call.Arguments})
+	switch out.Decision {
+	case hooks.Ask:
+		// ask 先问宿主；没有裁决端口（或宿主拒绝）时保持旧行为：按拒绝处理。
+		cfg.Hooks.Emit(hooks.Call{Event: hooks.PermissionRequest, Tool: call.Name, Args: call.Arguments})
+		why := out.Reason
+		if cfg.PermissionAsk != nil {
+			if verdict := cfg.PermissionAsk(ctx, call); verdict.Allow {
+				return call, llm.Message{}, false
+			} else if verdict.Reason != "" {
+				why = verdict.Reason
+			}
+		}
+		return call, rejectToolMessage(call, why), true
+	case hooks.Deny, hooks.Block:
+		return call, rejectToolMessage(call, out.Reason), true
+	case hooks.Modify:
+		if out.Args != "" {
+			call.Arguments = out.Args
+		}
+	}
+	return call, llm.Message{}, false
+}
+
+func rejectToolMessage(call llm.ToolCall, why string) llm.Message {
+	if why == "" {
+		why = "工具调用被拒绝"
+	}
+	return llm.Message{Role: llm.RoleTool, Name: call.Name, ToolCallID: call.ID, Content: why, IsError: true}
+}
+
+func noteToolHook(cfg Config, call llm.ToolCall, msg llm.Message) {
+	if cfg.Hooks == nil {
+		return
+	}
+	ev := hooks.PostToolUse
+	if msg.IsError {
+		ev = hooks.PostToolUseFailure
+	}
+	cfg.Hooks.Emit(hooks.Call{Event: ev, Tool: call.Name, Args: call.Arguments, Result: msg.Content, IsError: msg.IsError})
+}
+
 func runOne(ctx context.Context, cfg Config, call llm.ToolCall) llm.Message {
-	emit(cfg, Event{Type: "tool_execution_start", Tool: call.Name})
+	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})
+		if cfg.AfterTool != nil {
+			cfg.AfterTool(call, msg)
+		}
+		return msg
+	}
+	emit(cfg, Event{Type: "tool_execution_start", Tool: call.Name, 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})
+		noteToolHook(cfg, call, msg)
+		if cfg.AfterTool != nil {
+			cfg.AfterTool(call, msg)
+		}
+		return msg
+	} else {
+		call = gated
+	}
+	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})
+			noteToolHook(cfg, call, msg)
+			if cfg.AfterTool != nil {
+				cfg.AfterTool(call, msg)
+			}
+			return msg
+		}
+	}
+	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})
+		return msg
+	}
+	endSpan := startSpan(cfg, "execute_tool", map[string]string{"tool": call.Name})
+	defer endSpan()
 	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})
+	noteToolHook(cfg, call, msg)
+	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 prepareRequest(cfg Config, messages []llm.Message, turn int, repeats *repeatState, lengthNudge *bool) []llm.Message {
+	request := clearOldResults(cfg, persistedProjection(cfg, messages))
+	return reminders.Attach(request, turnReminderTexts(cfg, turn, repeats, lengthNudge))
+}
+
+// persistedProjection 在有会话存储时只采用已经落盘的消息。
+// steer 与 follow-up 必须先写入存储，下一轮请求才会看见它们。
+func persistedProjection(cfg Config, messages []llm.Message) []llm.Message {
+	if cfg.Session == nil {
+		return messages
+	}
+	loaded, err := cfg.Session.Load()
+	if err != nil {
+		return messages
+	}
+	return loaded
+}
+
+const unknownSideEffect = "副作用未知，未重试"
+
+func recoverPending(ctx context.Context, cfg Config) error {
+	if cfg.Session == nil {
+		return nil
+	}
+	pending, err := cfg.Session.PendingIntents()
+	if err != nil {
+		return err
+	}
+	for _, in := range pending {
+		current := tools.ReplayNever
+		if cfg.Tools != nil {
+			current = cfg.Tools.Replay(in.Name)
+		}
+		msg := llm.Message{
+			Role: llm.RoleTool, Name: in.Name, ToolCallID: in.CallID,
+			Content: unknownSideEffect, IsError: true,
+		}
+		if in.Replay == tools.ReplaySafe && current == tools.ReplaySafe && cfg.Tools != nil {
+			res := cfg.Tools.Execute(ctx, in.Name, json.RawMessage(in.Arguments))
+			msg.Content = res.Content
+			msg.IsError = res.IsError
+		}
+		if err := persist(cfg, msg); err != nil {
+			return err
+		}
+	}
+	return nil
+}
+
+func commitToolIntent(cfg Config, call llm.ToolCall) error {
+	if cfg.Session == nil {
+		return nil
+	}
+	replay := tools.ReplayNever
+	if cfg.Tools != nil {
+		replay = cfg.Tools.Replay(call.Name)
+	}
+	return cfg.Session.AppendToolIntent(session.ToolIntent{
+		CallID: call.ID, Name: call.Name, Arguments: call.Arguments, Replay: replay,
+	})
+}
+
+func startSpan(cfg Config, name string, attrs map[string]string) func() {
+	if cfg.Trace == nil {
+		return func() {}
+	}
+	end := cfg.Trace(name, attrs)
+	if end == nil {
+		return func() {}
+	}
+	return end
+}
+
+func clearOldResults(cfg Config, messages []llm.Message) []llm.Message {
+	window := cfg.ContextWindow
+	if window <= 0 {
+		window = session.DefaultContextWindow
+	}
+	reserve := cfg.OutputReserve
+	if reserve <= 0 {
+		reserve = session.DefaultOutputReserve
+	}
+	buffer := cfg.ContextBuffer
+	if buffer <= 0 {
+		buffer = session.DefaultContextBuffer
+	}
+	if !session.NeedsClear(session.EstimateChars(messages), window, reserve, buffer) {
+		return messages
+	}
+	return session.ClearOldToolResults(messages, cfg.KeepToolResults)
+}
+
+func turnReminderTexts(cfg Config, turn int, repeats *repeatState, lengthNudge *bool) []string {
+	var extra []string
+	if cfg.Reminders != nil {
+		for _, text := range cfg.Reminders(turn) {
+			if strings.TrimSpace(text) == "" {
+				continue
+			}
+			extra = append(extra, text)
+		}
+	}
+	if lengthNudge != nil && *lengthNudge {
+		extra = append(extra, 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, repeatReminderText)
+			repeats.reminders++
+		}
+	}
+	return extra
+}
+
+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 {
@@ -301,7 +848,7 @@ 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) {
@@ -319,8 +866,8 @@ func budgetExceeded(b *Budget, usage llm.Usage, toolCalls int, now time.Time) st
 	if !b.Deadline.IsZero() && now.After(b.Deadline) {
 		return fmt.Sprintf("budget_exceeded: deadline %s passed", b.Deadline.Format(time.RFC3339))
 	}
-	if b.MaxPromptTokens > 0 && usage.PromptTokens > b.MaxPromptTokens {
-		return fmt.Sprintf("budget_exceeded: prompt tokens %d > max %d", usage.PromptTokens, b.MaxPromptTokens)
+	if b.MaxPromptTokens > 0 && usage.LastPromptTokens > b.MaxPromptTokens {
+		return fmt.Sprintf("budget_exceeded: prompt tokens %d > max %d", usage.LastPromptTokens, b.MaxPromptTokens)
 	}
 	if b.MaxCompletionTokens > 0 && usage.CompletionTokens > b.MaxCompletionTokens {
 		return fmt.Sprintf("budget_exceeded: completion tokens %d > max %d", usage.CompletionTokens, b.MaxCompletionTokens)
@@ -4,6 +4,7 @@ import (
 	"context"
 	"encoding/json"
 	"strings"
+	"sync"
 	"testing"
 	"time"

@@ -66,15 +67,108 @@ func TestLoopBudgetToolCallCap(t *testing.T) {
 	if !sawBudgetEvent {
 		t.Fatal("expected budget_exceeded event")
 	}
-	// 第二批工具调用不该被执行（预算在执行前拦截）：transcript 只含一次 tool 结果。
+	// 第一批执行 1 次；第二批超限，两条合成错误结果，不执行。
 	toolResults := 0
+	skipped := 0
 	for _, m := range res.Messages {
-		if m.Role == llm.RoleTool {
-			toolResults++
+		if m.Role != llm.RoleTool {
+			continue
+		}
+		toolResults++
+		if m.IsError && m.Content == toolBudgetSkipContent {
+			skipped++
 		}
 	}
-	if toolResults != 1 {
-		t.Fatalf("expected exactly 1 tool result, got %d", toolResults)
+	if toolResults != 3 || skipped != 2 {
+		t.Fatalf("expected 3 tool results including 2 skipped, got results=%d skipped=%d", toolResults, skipped)
+	}
+	assertToolCallsPaired(t, res.Messages)
+}
+
+func assertToolCallsPaired(t *testing.T, messages []llm.Message) {
+	t.Helper()
+	pending := map[string]string{}
+	for _, m := range messages {
+		if m.Role == llm.RoleAssistant {
+			for _, call := range m.ToolCalls {
+				if _, ok := pending[call.ID]; ok {
+					t.Fatalf("duplicate tool call id %s", call.ID)
+				}
+				pending[call.ID] = call.Name
+			}
+		}
+		if m.Role == llm.RoleTool {
+			name, ok := pending[m.ToolCallID]
+			if !ok {
+				t.Fatalf("tool result %s has no matching call", m.ToolCallID)
+			}
+			if name != m.Name {
+				t.Fatalf("tool result %s name %q != call %q", m.ToolCallID, m.Name, name)
+			}
+			delete(pending, m.ToolCallID)
+		}
+	}
+	if len(pending) != 0 {
+		t.Fatalf("unpaired tool calls: %v", pending)
+	}
+}
+
+type pauseTool struct{}
+
+func (pauseTool) Name() string        { return "pause" }
+func (pauseTool) Description() string { return "slow tool" }
+func (pauseTool) Parameters() json.RawMessage {
+	return json.RawMessage(`{"type":"object","properties":{}}`)
+}
+func (pauseTool) Execute(context.Context, json.RawMessage) (tools.Result, error) {
+	time.Sleep(40 * time.Millisecond)
+	return tools.Result{Content: "ok"}, nil
+}
+
+func TestParallelHostCallbacksAreSerial(t *testing.T) {
+	reg := tools.NewRegistry(t.TempDir())
+	reg.Register(pauseTool{})
+	var mu sync.Mutex
+	inFlight, maxInFlight := 0, 0
+	enter := func() {
+		mu.Lock()
+		inFlight++
+		if inFlight > maxInFlight {
+			maxInFlight = inFlight
+		}
+		mu.Unlock()
+		time.Sleep(20 * time.Millisecond)
+		mu.Lock()
+		inFlight--
+		mu.Unlock()
+	}
+	_, err := Run(context.Background(), Config{
+		Model:         "fake",
+		Tools:         reg,
+		ToolExecution: ExecParallel,
+		OnEvent: func(Event) {
+			enter()
+		},
+		BeforeTool: func(llm.ToolCall) error {
+			enter()
+			return nil
+		},
+		AfterTool: func(llm.ToolCall, llm.Message) {
+			enter()
+		},
+		Stream: usageScript(llm.Usage{},
+			llm.Message{Role: llm.RoleAssistant, StopReason: llm.StopToolCalls, ToolCalls: []llm.ToolCall{
+				{ID: "a", Name: "pause", Arguments: `{}`},
+				{ID: "b", Name: "pause", Arguments: `{}`},
+			}},
+			llm.Message{Role: llm.RoleAssistant, Content: "done", StopReason: llm.StopEnd},
+		),
+	}, llm.Message{Role: llm.RoleUser, Content: "go"})
+	if err != nil {
+		t.Fatal(err)
+	}
+	if maxInFlight != 1 {
+		t.Fatalf("host callbacks overlapped: max in flight %d", maxInFlight)
 	}
 }

@@ -100,6 +194,57 @@ func TestLoopBudgetPromptTokens(t *testing.T) {
 	}
 }

+func TestLoopUsageSplitsLastRequestFromCumulative(t *testing.T) {
+	res, err := Run(context.Background(), Config{
+		Model:  "fake",
+		Tools:  tools.WorkspaceTools(t.TempDir()),
+		Budget: &Budget{MaxPromptTokens: 10},
+		Stream: usageScript(llm.Usage{PromptTokens: 6, CompletionTokens: 1, TotalTokens: 7},
+			llm.Message{Role: llm.RoleAssistant, StopReason: llm.StopToolCalls, ToolCalls: []llm.ToolCall{{
+				ID: "c1", Name: "list_dir", Arguments: `{}`,
+			}}},
+			llm.Message{Role: llm.RoleAssistant, Content: "ok", StopReason: llm.StopEnd},
+		),
+	}, llm.Message{Role: llm.RoleUser, Content: "go"})
+	if err != nil {
+		t.Fatal(err)
+	}
+	if res.Err != nil || res.Text != "ok" {
+		t.Fatalf("last request stays under the window: err=%v text=%q", res.Err, res.Text)
+	}
+	if res.Usage.PromptTokens != 12 || res.Usage.LastPromptTokens != 6 {
+		t.Fatalf("billing sum %d last %d", res.Usage.PromptTokens, res.Usage.LastPromptTokens)
+	}
+}
+
+func TestLoopRejectsIncompleteToolCall(t *testing.T) {
+	tool := &countingTool{name: "list_dir"}
+	reg := tools.NewRegistry(t.TempDir())
+	reg.Register(tool)
+	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: "list_dir", Arguments: `{"path":`,
+		}}}, llm.Message{Role: llm.RoleAssistant, Content: "stopped", StopReason: llm.StopEnd}),
+	}, llm.Message{Role: llm.RoleUser, Content: "go"})
+	if err != nil {
+		t.Fatal(err)
+	}
+	if tool.calls != 0 {
+		t.Fatalf("incomplete call executed %d times", tool.calls)
+	}
+	var paired bool
+	for _, m := range res.Messages {
+		if m.Role == llm.RoleTool && m.ToolCallID == "c1" && m.IsError {
+			paired = true
+		}
+	}
+	if !paired {
+		t.Fatal("expected a paired error result")
+	}
+}
+
 func TestLoopBudgetDeadline(t *testing.T) {
 	res, err := Run(context.Background(), Config{
 		Model:  "fake",
@@ -0,0 +1,168 @@
+package runtime
+
+import (
+	"context"
+	"strings"
+	"testing"
+
+	"agent-native/runtime/guardrails"
+	"agent-native/runtime/hooks"
+	"agent-native/runtime/llm"
+	"agent-native/runtime/tools"
+)
+
+// TestInputGuardrailDenySkipsLoop：输入护栏 deny 时不进循环、不调模型，
+// 以理由直接收尾；原文已进转录（留痕）。
+func TestInputGuardrailDenySkipsLoop(t *testing.T) {
+	calls := 0
+	bus := hooks.New()
+	guardrails.Register(bus, guardrails.Set{
+		Input: func(ctx context.Context, text string) guardrails.Verdict {
+			return guardrails.Verdict{Decision: hooks.Deny, Reason: "含敏感指令"}
+		},
+	}, guardrails.Scope{Entry: true})
+	res, err := Run(context.Background(), Config{
+		Model: "fake",
+		Hooks: bus,
+		Stream: func(ctx context.Context, model string, messages []llm.Message, decls []llm.ToolDecl, opts llm.StreamOptions) <-chan llm.Event {
+			calls++
+			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: "bad input"})
+	if err != nil {
+		t.Fatal(err)
+	}
+	if calls != 0 {
+		t.Fatalf("model called %d times after input deny", calls)
+	}
+	if res.Text != "含敏感指令" {
+		t.Fatalf("result text = %q", res.Text)
+	}
+	if !strings.Contains(transcriptText(res.Messages), "bad input") {
+		t.Fatal("original input missing from transcript")
+	}
+}
+
+// TestInputGuardrailModifyAppendsRewrite：modify 时改写内容追加为新的用户消息，
+// 模型最终看到的是改写后的输入；原文保留在转录里。
+func TestInputGuardrailModifyAppendsRewrite(t *testing.T) {
+	var saw string
+	bus := hooks.New()
+	guardrails.Register(bus, guardrails.Set{
+		Input: func(ctx context.Context, text string) guardrails.Verdict {
+			return guardrails.Verdict{Decision: hooks.Modify, Replacement: "改写后的输入"}
+		},
+	}, guardrails.Scope{Entry: true})
+	_, err := Run(context.Background(), Config{
+		Model: "fake",
+		Hooks: bus,
+		Stream: func(ctx context.Context, model string, messages []llm.Message, decls []llm.ToolDecl, opts llm.StreamOptions) <-chan llm.Event {
+			saw = transcriptText(messages)
+			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: "原始输入"})
+	if err != nil {
+		t.Fatal(err)
+	}
+	if !strings.Contains(saw, "原始输入") || !strings.Contains(saw, "改写后的输入") {
+		t.Fatalf("request messages = %q", saw)
+	}
+}
+
+// TestOutputGuardrailDenyContinuesThenCaps：输出护栏 deny 视同续跑，
+// 理由进转录；达内核续跑上限后收尾。
+func TestOutputGuardrailDenyContinuesThenCaps(t *testing.T) {
+	calls := 0
+	bus := hooks.New()
+	guardrails.Register(bus, guardrails.Set{
+		Output: func(ctx context.Context, text string) guardrails.Verdict {
+			return guardrails.Verdict{Decision: hooks.Deny, Reason: "缺少来源标注"}
+		},
+	}, guardrails.Scope{Final: true})
+	res, err := Run(context.Background(), Config{
+		Model:    "fake",
+		MaxTurns: 8,
+		Hooks:    bus,
+		Stream: script(
+			llm.Message{Role: llm.RoleAssistant, Content: "v1", StopReason: llm.StopEnd},
+			llm.Message{Role: llm.RoleAssistant, Content: "v2", StopReason: llm.StopEnd},
+			llm.Message{Role: llm.RoleAssistant, Content: "v3", StopReason: llm.StopEnd},
+			llm.Message{Role: llm.RoleAssistant, Content: "v4", StopReason: llm.StopEnd},
+		),
+	}, llm.Message{Role: llm.RoleUser, Content: "go"})
+	if err != nil {
+		t.Fatal(err)
+	}
+	_ = calls
+	if strings.Count(transcriptText(res.Messages), "缺少来源标注") != 3 {
+		t.Fatalf("reason count = %s", transcriptText(res.Messages))
+	}
+	if res.Text != "v4" {
+		t.Fatalf("final text = %q", res.Text)
+	}
+}
+
+// TestToolGuardrailDenyPairsResult：工具护栏 deny 写配对错误结果且不执行工具。
+func TestToolGuardrailDenyPairsResult(t *testing.T) {
+	hits := 0
+	reg := tools.NewRegistry(t.TempDir())
+	reg.Register(hitTool{hits: &hits})
+	bus := hooks.New()
+	guardrails.Register(bus, guardrails.Set{
+		Tool: func(ctx context.Context, name, args string) guardrails.Verdict {
+			return guardrails.Verdict{Decision: hooks.Deny, Reason: "护栏禁用"}
+		},
+	}, guardrails.Scope{})
+	res, err := Run(context.Background(), Config{
+		Model: "fake",
+		Tools: reg,
+		Hooks: bus,
+		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},
+		),
+	}, llm.Message{Role: llm.RoleUser, Content: "go"})
+	if err != nil {
+		t.Fatal(err)
+	}
+	if hits != 0 {
+		t.Fatalf("tool executed %d times", hits)
+	}
+	paired := false
+	for _, m := range res.Messages {
+		if m.Role == llm.RoleTool && m.ToolCallID == "c1" && m.IsError && m.Content == "护栏禁用" {
+			paired = true
+		}
+	}
+	if !paired {
+		t.Fatalf("missing paired result: %s", transcriptText(res.Messages))
+	}
+}
+
+// TestSubRunScopeSkipsInputOutput：子代理 run（Entry/Final=false）注册同名护栏也不触发。
+func TestSubRunScopeSkipsInputOutput(t *testing.T) {
+	bus := hooks.New()
+	guardrails.Register(bus, guardrails.Set{
+		Input:  func(ctx context.Context, text string) guardrails.Verdict { return guardrails.Verdict{Decision: hooks.Deny, Reason: "x"} },
+		Output: func(ctx context.Context, text string) guardrails.Verdict { return guardrails.Verdict{Decision: hooks.Deny, Reason: "y"} },
+	}, guardrails.Scope{})
+	res, err := Run(context.Background(), Config{
+		Model: "fake",
+		Hooks: bus,
+		Stream: script(llm.Message{Role: llm.RoleAssistant, Content: "ok", StopReason: llm.StopEnd}),
+	}, llm.Message{Role: llm.RoleUser, Content: "go"})
+	if err != nil {
+		t.Fatal(err)
+	}
+	if res.Text != "ok" {
+		t.Fatalf("sub-scope run text = %q", res.Text)
+	}
+}
@@ -0,0 +1,372 @@
+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 TestLoopContinueOnLengthUnlimited(t *testing.T) {
+	var n int
+	stream := func(ctx context.Context, model string, messages []llm.Message, decls []llm.ToolDecl, opts llm.StreamOptions) <-chan llm.Event {
+		n++
+		ch := make(chan llm.Event, 1)
+		finish := "length"
+		text := "part"
+		if n >= 5 {
+			finish = "stop"
+			text = "done"
+		}
+		ch <- llm.Event{
+			Kind:         llm.KindMessageEnd,
+			FinishReason: finish,
+			Message:      llm.Message{Role: llm.RoleAssistant, Content: text, StopReason: llm.StopEnd},
+		}
+		close(ch)
+		return ch
+	}
+	res, err := Run(context.Background(), Config{
+		Model:                  "fake",
+		MaxTurns:               8,
+		ContinueOnLength:       true,
+		MaxLengthContinuations: -1,
+		Stream:                 stream,
+	}, llm.Message{Role: llm.RoleUser, Content: "go"})
+	if err != nil {
+		t.Fatal(err)
+	}
+	if res.Err != nil {
+		t.Fatalf("err = %v", res.Err)
+	}
+	if n != 5 {
+		t.Fatalf("model calls = %d, want 5", n)
+	}
+	if res.Text != "done" {
+		t.Fatalf("text = %q", res.Text)
+	}
+}
+
+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
+}
@@ -0,0 +1,253 @@
+package runtime
+
+import (
+	"context"
+	"strings"
+	"testing"
+
+	"agent-native/runtime/hooks"
+	"agent-native/runtime/llm"
+	"agent-native/runtime/tools"
+)
+
+type eventRecorder struct {
+	seen []hooks.Event
+}
+
+func (r *eventRecorder) handler(c hooks.Call) hooks.Outcome {
+	r.seen = append(r.seen, c.Event)
+	return hooks.Outcome{}
+}
+
+func (r *eventRecorder) has(ev hooks.Event) bool {
+	for _, e := range r.seen {
+		if e == ev {
+			return true
+		}
+	}
+	return false
+}
+
+// TestHookLifecycleOrder 断言一次带工具调用的 run 的事件顺序为
+// SessionStart → UserPromptSubmit → PreToolUse → PostToolUse →
+// PostToolBatch → Stop → SessionEnd。
+func TestHookLifecycleOrder(t *testing.T) {
+	rec := &eventRecorder{}
+	bus := hooks.New()
+	bus.On(rec.handler)
+	hits := 0
+	reg := tools.NewRegistry(t.TempDir())
+	reg.Register(hitTool{hits: &hits})
+	_, err := Run(context.Background(), Config{
+		Model: "fake",
+		Tools: reg,
+		Hooks: bus,
+		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},
+		),
+	}, llm.Message{Role: llm.RoleUser, Content: "go"})
+	if err != nil {
+		t.Fatal(err)
+	}
+	want := []hooks.Event{hooks.SessionStart, hooks.UserPromptSubmit, hooks.PreToolUse,
+		hooks.PostToolUse, hooks.PostToolBatch, hooks.Stop, hooks.SessionEnd}
+	if len(rec.seen) != len(want) {
+		t.Fatalf("events = %v, want %v", rec.seen, want)
+	}
+	for i := range want {
+		if rec.seen[i] != want[i] {
+			t.Fatalf("events = %v, want %v", rec.seen, want)
+		}
+	}
+}
+
+// TestHookUserPromptSubmitCarriesText 断言 UserPromptSubmit 事件带用户输入正文。
+func TestHookUserPromptSubmitCarriesText(t *testing.T) {
+	var got string
+	bus := hooks.New()
+	bus.On(func(c hooks.Call) hooks.Outcome {
+		if c.Event == hooks.UserPromptSubmit {
+			got = c.LastText
+		}
+		return hooks.Outcome{}
+	})
+	_, err := Run(context.Background(), Config{
+		Model: "fake",
+		Hooks: bus,
+		Stream: script(llm.Message{Role: llm.RoleAssistant, Content: "ok", StopReason: llm.StopEnd}),
+	}, llm.Message{Role: llm.RoleUser, Content: "ping"})
+	if err != nil {
+		t.Fatal(err)
+	}
+	if got != "ping" {
+		t.Fatalf("UserPromptSubmit text = %q", got)
+	}
+}
+
+// TestAskPortAllowProceeds：ask 决策经宿主裁决放行后工具照常执行。
+func TestAskPortAllowProceeds(t *testing.T) {
+	hits := 0
+	reg := tools.NewRegistry(t.TempDir())
+	reg.Register(hitTool{hits: &hits})
+	bus := hooks.New()
+	bus.On(func(c hooks.Call) hooks.Outcome {
+		if c.Event == hooks.PreToolUse {
+			return hooks.Outcome{Decision: hooks.Ask, Reason: "需要确认"}
+		}
+		return hooks.Outcome{}
+	})
+	sawRequest := false
+	bus.On(func(c hooks.Call) hooks.Outcome {
+		if c.Event == hooks.PermissionRequest {
+			sawRequest = true
+		}
+		return hooks.Outcome{}
+	})
+	res, err := Run(context.Background(), Config{
+		Model: "fake",
+		Tools: reg,
+		Hooks: bus,
+		PermissionAsk: func(ctx context.Context, call llm.ToolCall) AskVerdict {
+			return AskVerdict{Allow: true}
+		},
+		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},
+		),
+	}, llm.Message{Role: llm.RoleUser, Content: "go"})
+	if err != nil {
+		t.Fatal(err)
+	}
+	if hits != 1 {
+		t.Fatalf("tool executed %d times, want 1", hits)
+	}
+	if !sawRequest {
+		t.Fatal("PermissionRequest event missing")
+	}
+	if !strings.Contains(transcriptText(res.Messages), "ran") {
+		t.Fatalf("tool result missing: %s", transcriptText(res.Messages))
+	}
+}
+
+// TestAskPortDenyPairsResult：宿主拒绝时写配对错误结果，理由取宿主理由。
+func TestAskPortDenyPairsResult(t *testing.T) {
+	hits := 0
+	reg := tools.NewRegistry(t.TempDir())
+	reg.Register(hitTool{hits: &hits})
+	bus := hooks.New()
+	bus.On(func(c hooks.Call) hooks.Outcome {
+		if c.Event == hooks.PreToolUse {
+			return hooks.Outcome{Decision: hooks.Ask, Reason: "需要确认"}
+		}
+		return hooks.Outcome{}
+	})
+	res, err := Run(context.Background(), Config{
+		Model: "fake",
+		Tools: reg,
+		Hooks: bus,
+		PermissionAsk: func(ctx context.Context, call llm.ToolCall) AskVerdict {
+			return AskVerdict{Allow: false, Reason: "用户拒绝"}
+		},
+		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},
+		),
+	}, llm.Message{Role: llm.RoleUser, Content: "go"})
+	if err != nil {
+		t.Fatal(err)
+	}
+	if hits != 0 {
+		t.Fatalf("tool executed %d times, want 0", hits)
+	}
+	paired := false
+	for _, m := range res.Messages {
+		if m.Role == llm.RoleTool && m.ToolCallID == "c1" && m.IsError && m.Content == "用户拒绝" {
+			paired = true
+		}
+	}
+	if !paired {
+		t.Fatalf("missing paired deny result: %s", transcriptText(res.Messages))
+	}
+}
+
+// TestAskWithoutPortKeepsLegacyReject：没有裁决端口时 ask 仍按拒绝处理（旧行为）。
+func TestAskWithoutPortKeepsLegacyReject(t *testing.T) {
+	hits := 0
+	reg := tools.NewRegistry(t.TempDir())
+	reg.Register(hitTool{hits: &hits})
+	bus := hooks.New()
+	bus.On(func(c hooks.Call) hooks.Outcome {
+		if c.Event == hooks.PreToolUse {
+			return hooks.Outcome{Decision: hooks.Ask, Reason: "需要确认"}
+		}
+		return hooks.Outcome{}
+	})
+	res, err := Run(context.Background(), Config{
+		Model: "fake",
+		Tools: reg,
+		Hooks: bus,
+		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},
+		),
+	}, llm.Message{Role: llm.RoleUser, Content: "go"})
+	if err != nil {
+		t.Fatal(err)
+	}
+	if hits != 0 {
+		t.Fatalf("tool executed %d times, want 0", hits)
+	}
+	paired := false
+	for _, m := range res.Messages {
+		if m.Role == llm.RoleTool && m.ToolCallID == "c1" && m.IsError && m.Content == "需要确认" {
+			paired = true
+		}
+	}
+	if !paired {
+		t.Fatalf("missing paired ask-reject result: %s", transcriptText(res.Messages))
+	}
+}
+
+// TestStopContinuationFlagVisible：续跑后 Stop 事件的 Continuations 计数对下游可见。
+func TestStopContinuationFlagVisible(t *testing.T) {
+	var counts []int
+	bus := hooks.New()
+	bus.On(func(c hooks.Call) hooks.Outcome {
+		if c.Event == hooks.Stop {
+			counts = append(counts, c.Continuations)
+			return hooks.Outcome{Decision: hooks.Block, Reason: "再试"}
+		}
+		return hooks.Outcome{}
+	})
+	_, err := Run(context.Background(), Config{
+		Model:    "fake",
+		MaxTurns: 8,
+		Hooks:    bus,
+		Stream: script(
+			llm.Message{Role: llm.RoleAssistant, Content: "a", StopReason: llm.StopEnd},
+			llm.Message{Role: llm.RoleAssistant, Content: "a", StopReason: llm.StopEnd},
+			llm.Message{Role: llm.RoleAssistant, Content: "a", StopReason: llm.StopEnd},
+			llm.Message{Role: llm.RoleAssistant, Content: "a", StopReason: llm.StopEnd},
+		),
+	}, llm.Message{Role: llm.RoleUser, Content: "go"})
+	if err != nil {
+		t.Fatal(err)
+	}
+	if len(counts) != 4 {
+		t.Fatalf("stop events = %v", counts)
+	}
+	for i, c := range counts {
+		if c != i {
+			t.Fatalf("stop continuations = %v, want 0,1,2,3", counts)
+		}
+	}
+}
@@ -0,0 +1,27 @@
+package mcp
+
+import "context"
+
+type askKey struct{}
+
+// WithAsk 把当前回合的 AskUser 放进上下文。MCP 返回 input_required 时使用它。
+func WithAsk(ctx context.Context, fn func(context.Context, string) (string, error)) context.Context {
+	return context.WithValue(ctx, askKey{}, fn)
+}
+
+func askFrom(ctx context.Context) func(context.Context, string) (string, error) {
+	if ctx == nil {
+		return nil
+	}
+	fn, _ := ctx.Value(askKey{}).(func(context.Context, string) (string, error))
+	return fn
+}
+
+func inputQuestion(res callToolResult) (string, bool) {
+	for _, blk := range res.Content {
+		if blk.Type == "input_required" {
+			return blk.Text, true
+		}
+	}
+	return "", false
+}
@@ -8,6 +8,7 @@ import (
 	"bufio"
 	"encoding/json"
 	"os"
+	"time"
 )

 // RunFakeServer 在 stdin/stdout 上运行最小 MCP server（echo / boom / 未知方法），
@@ -67,6 +68,9 @@ func fakeServerHandle(req rpcRequest) rpcResponse {
 				Content: []contentBlock{{Type: "text", Text: "boom failed"}},
 				IsError: true,
 			})
+		case "slow":
+			time.Sleep(2 * time.Second)
+			return respond(callToolResult{Content: []contentBlock{{Type: "text", Text: "slow done"}}})
 		default:
 			return fail(-32602, "unknown tool "+p.Name)
 		}
@@ -87,6 +91,11 @@ func fakeServerTools() []toolInfo {
 			Description: "Always fails",
 			InputSchema: json.RawMessage(`{"type":"object","properties":{}}`),
 		},
+		{
+			Name:        "slow",
+			Description: "Responds after a delay",
+			InputSchema: json.RawMessage(`{"type":"object","properties":{}}`),
+		},
 	}
 }

@@ -92,7 +92,31 @@ func (c *httpClient) CallTool(ctx context.Context, name string, args json.RawMes
 	return res, err
 }

-func (c *httpClient) Close() error { return nil }
+func (c *httpClient) Close() error {
+	c.mu.Lock()
+	sid := c.sessionID
+	url := c.cfg.URL
+	client := c.client
+	c.mu.Unlock()
+	if sid == "" || url == "" || client == nil {
+		return nil
+	}
+	req, err := http.NewRequest(http.MethodDelete, url, nil)
+	if err != nil {
+		return err
+	}
+	req.Header.Set("Mcp-Session-Id", sid)
+	resp, err := client.Do(req)
+	if err != nil {
+		return err
+	}
+	defer resp.Body.Close()
+	_, _ = io.Copy(io.Discard, io.LimitReader(resp.Body, 2048))
+	if resp.StatusCode >= 400 {
+		return fmt.Errorf("mcp session terminate: http %d", resp.StatusCode)
+	}
+	return nil
+}

 func (c *httpClient) roundTrip(ctx context.Context, method string, params any, out any) error {
 	c.mu.Lock()
@@ -19,6 +19,8 @@ const (
 	maxToolNameBytes     = 128
 	callTimeoutDefault   = 30 * time.Second
 	registerTimeoutFixed = 30 * time.Second
+	// 工具数超过这个值且没有显式暴露级别时，整台服务器改为延迟暴露。
+	deferredToolThreshold = 8
 )

 // Manager 维护一组 MCP server 连接，并把它们的工具注册进 tools.Registry。
@@ -72,49 +74,138 @@ func (m *Manager) client(ctx context.Context, cfg ServerConfig) (Client, error)
 	return c, nil
 }

-// RegisterInto 连接全部 server、枚举工具并以 mcp__<server>__<tool> 命名注册。
-// 任一 server 不可达即返回错误（fail-closed，不静默跳过）。
-func (m *Manager) RegisterInto(ctx context.Context, reg *tools.Registry) error {
-	for _, cfg := range m.configs {
-		c, err := m.client(ctx, cfg)
-		if err != nil {
+// ApplyOverrides 用项目配置覆盖启用状态和暴露级别。空字符串不改。
+func (m *Manager) ApplyOverrides(raw string) error {
+	if m == nil || strings.TrimSpace(raw) == "" {
+		return nil
+	}
+	var over map[string]struct {
+		Enabled  *bool  `json:"enabled"`
+		Exposure string `json:"exposure"`
+	}
+	if err := json.Unmarshal([]byte(raw), &over); err != nil {
+		return err
+	}
+	for i, cfg := range m.configs {
+		o, ok := over[cfg.Name]
+		if !ok {
+			continue
+		}
+		if o.Enabled != nil {
+			m.configs[i].Enabled = o.Enabled
+		}
+		if o.Exposure != "" {
+			m.configs[i].Exposure = o.Exposure
+		}
+		if err := m.configs[i].Validate(); err != nil {
 			return err
 		}
-		listCtx, cancel := context.WithTimeout(ctx, registerTimeoutFixed)
-		list, err := c.ListTools(listCtx)
-		cancel()
-		if err != nil {
-			return fmt.Errorf("mcp server %q tools/list failed: %w", cfg.Name, err)
-		}
-		if len(list) > maxToolsPerServer {
-			return fmt.Errorf("mcp server %q exposes %d tools (max %d)", cfg.Name, len(list), maxToolsPerServer)
-		}
-		for _, ti := range list {
-			name := ToolName(cfg.Name, ti.Name)
-			if len(name) > maxToolNameBytes {
-				return fmt.Errorf("mcp tool name too long: %s", name)
-			}
-			desc := ti.Description
-			if len(desc) > maxDescriptionBytes {
-				desc = desc[:maxDescriptionBytes] + "…[truncated]"
-			}
-			schema := ti.InputSchema
-			if len(schema) == 0 {
-				schema = json.RawMessage(`{"type":"object","properties":{}}`)
-			}
-			reg.Register(&bridgeTool{
-				client:      c,
-				serverCfg:   cfg,
-				toolName:    ti.Name,
-				declName:    name,
-				description: desc,
-				schema:      schema,
-			})
-		}
 	}
 	return nil
 }

+// Ready 报告已启用的服务器是否都已经连上。已连上时注册只是枚举工具，不再握手。
+func (m *Manager) Ready() bool {
+	if m == nil {
+		return true
+	}
+	m.mu.Lock()
+	defer m.mu.Unlock()
+	for _, cfg := range m.configs {
+		if !cfg.IsEnabled() {
+			continue
+		}
+		if _, ok := m.clients[cfg.Name]; !ok {
+			return false
+		}
+	}
+	return true
+}
+func (m *Manager) StartBackground(ctx context.Context, reg *tools.Registry) {
+	if m == nil || reg == nil {
+		return
+	}
+	for _, cfg := range m.configs {
+		cfg := cfg
+		if !cfg.IsEnabled() {
+			continue
+		}
+		go func() { _ = m.registerServer(ctx, reg, cfg) }()
+	}
+}
+
+// RegisterInto 连接全部已启用的 server 并注册工具。未启用的跳过。
+func (m *Manager) RegisterInto(ctx context.Context, reg *tools.Registry) error {
+	for _, cfg := range m.configs {
+		if !cfg.IsEnabled() {
+			continue
+		}
+		if err := m.registerServer(ctx, reg, cfg); err != nil {
+			return err
+		}
+	}
+	return nil
+}
+
+func (m *Manager) registerServer(ctx context.Context, reg *tools.Registry, cfg ServerConfig) error {
+	c, err := m.client(ctx, cfg)
+	if err != nil {
+		return err
+	}
+	listCtx, cancel := context.WithTimeout(ctx, registerTimeoutFixed)
+	list, err := c.ListTools(listCtx)
+	cancel()
+	if err != nil {
+		return fmt.Errorf("mcp server %q tools/list failed: %w", cfg.Name, err)
+	}
+	if len(list) > maxToolsPerServer {
+		return fmt.Errorf("mcp server %q exposes %d tools (max %d)", cfg.Name, len(list), maxToolsPerServer)
+	}
+	exposure := chooseExposure(cfg.Exposure, len(list))
+	for _, ti := range list {
+		name := ToolName(cfg.Name, ti.Name)
+		if len(name) > maxToolNameBytes {
+			return fmt.Errorf("mcp tool name too long: %s", name)
+		}
+		desc := ti.Description
+		if len(desc) > maxDescriptionBytes {
+			desc = tools.CutUTF8(desc, maxDescriptionBytes) + "…[truncated]"
+		}
+		schema := ti.InputSchema
+		if len(schema) == 0 {
+			schema = json.RawMessage(`{"type":"object","properties":{}}`)
+		}
+		reg.Register(&bridgeTool{
+			client:      c,
+			serverCfg:   cfg,
+			toolName:    ti.Name,
+			declName:    name,
+			description: desc,
+			schema:      schema,
+			exposure:    exposure,
+			redial: func(ctx context.Context) (Client, error) {
+				m.mu.Lock()
+				delete(m.clients, cfg.Name)
+				m.mu.Unlock()
+				return m.client(ctx, cfg)
+			},
+		})
+	}
+	return nil
+}
+
+func chooseExposure(explicit string, n int) string {
+	switch explicit {
+	case tools.ExposureDirect, tools.ExposureDeferred:
+		return explicit
+	default:
+		if n > deferredToolThreshold {
+			return tools.ExposureDeferred
+		}
+		return tools.ExposureDirect
+	}
+}
+
 // ToolName 生成桥接工具名（命名空间隔离）。
 func ToolName(server, tool string) string {
 	return "mcp__" + server + "__" + tool
@@ -130,7 +221,7 @@ func (m *Manager) Close() {
 	m.clients = map[string]Client{}
 }

-// bridgeTool 把一个 MCP 工具适配为 tools.Tool；调用失败重连一次后重试。
+// bridgeTool 把一个 MCP 工具适配为 tools.Tool。CallTool 失败时重连一次再试。
 type bridgeTool struct {
 	client      Client
 	serverCfg   ServerConfig
@@ -138,21 +229,39 @@ type bridgeTool struct {
 	declName    string
 	description string
 	schema      json.RawMessage
+	exposure    string
+	redial      func(ctx context.Context) (Client, error)
 }

-func (t *bridgeTool) Name() string        { return t.declName }
-func (t *bridgeTool) Description() string { return t.description }
-func (t *bridgeTool) Parameters() json.RawMessage {
-	return t.schema
-}
+func (t *bridgeTool) Name() string                { return t.declName }
+func (t *bridgeTool) Description() string         { return t.description }
+func (t *bridgeTool) Parameters() json.RawMessage { return t.schema }
+func (t *bridgeTool) Exposure() string            { return t.exposure }

 func (t *bridgeTool) Execute(ctx context.Context, args json.RawMessage) (tools.Result, error) {
-	callCtx, cancel := context.WithTimeout(ctx, requestTimeout(t.serverCfg))
-	defer cancel()
-	res, err := t.client.CallTool(callCtx, t.toolName, args)
+	res, err := t.call(ctx, args)
+	if err != nil && t.redial != nil {
+		redial := t.redial
+		t.redial = nil
+		if c, rerr := redial(ctx); rerr == nil {
+			t.client = c
+			res, err = t.call(ctx, args)
+		}
+	}
 	if err != nil {
 		return tools.Result{Content: fmt.Sprintf("mcp tool %s failed: %v", t.declName, err), IsError: true}, nil
 	}
+	if question, ok := inputQuestion(res); ok {
+		ask := askFrom(ctx)
+		if ask == nil {
+			return tools.Result{Content: "需要用户输入：" + question, IsError: true}, nil
+		}
+		answer, aerr := ask(ctx, question)
+		if aerr != nil {
+			return tools.Result{Content: aerr.Error(), IsError: true}, nil
+		}
+		return tools.Result{Content: answer}, nil
+	}
 	var b strings.Builder
 	for _, blk := range res.Content {
 		if blk.Type == "text" && blk.Text != "" {
@@ -162,10 +271,16 @@ func (t *bridgeTool) Execute(ctx context.Context, args json.RawMessage) (tools.R
 	}
 	out := b.String()
 	if len(out) > maxTextResultBytes {
-		out = out[:maxTextResultBytes] + "\n[truncated]"
+		out = tools.CutUTF8(out, maxTextResultBytes) + "\n[truncated]"
 	}
 	if res.IsError {
 		return tools.Result{Content: out, IsError: true}, nil
 	}
 	return tools.Result{Content: out}, nil
 }
+
+func (t *bridgeTool) call(ctx context.Context, args json.RawMessage) (callToolResult, error) {
+	callCtx, cancel := context.WithTimeout(ctx, requestTimeout(t.serverCfg))
+	defer cancel()
+	return t.client.CallTool(callCtx, t.toolName, args)
+}
@@ -45,7 +45,7 @@ func TestStdioClientHandshakeListAndCall(t *testing.T) {
 	defer c.Close()

 	list, err := c.ListTools(ctx)
-	if err != nil || len(list) != 2 || list[0].Name != "echo" {
+	if err != nil || len(list) != 3 || list[0].Name != "echo" {
 		t.Fatalf("list failed: %v %+v", err, list)
 	}
 	res, err := c.CallTool(ctx, "echo", json.RawMessage(`{"text":"hi"}`))
@@ -75,6 +75,35 @@ func TestStdioClientServerRejectsBadSchema(t *testing.T) {
 	}
 }

+func TestStdioSlowCancelThenEcho(t *testing.T) {
+	ctx := context.Background()
+	c, err := DialStdio(ctx, fakeStdioConfig(t))
+	if err != nil {
+		t.Fatal(err)
+	}
+	defer c.Close()
+
+	slowCtx, cancel := context.WithTimeout(ctx, 80*time.Millisecond)
+	defer cancel()
+	_, err = c.CallTool(slowCtx, "slow", nil)
+	if err == nil {
+		t.Fatal("slow call should time out")
+	}
+
+	echoCtx, echoCancel := context.WithTimeout(ctx, 3*time.Second)
+	defer echoCancel()
+	res, err := c.CallTool(echoCtx, "echo", json.RawMessage(`{"text":"after-slow"}`))
+	if err != nil {
+		t.Fatal(err)
+	}
+	if len(res.Content) != 1 || res.Content[0].Text != "echo: after-slow" {
+		t.Fatalf("echo must receive its own result, got %+v", res)
+	}
+	if c.Readers() != 1 {
+		t.Fatalf("cancel must leave a single reader, got %d", c.Readers())
+	}
+}
+
 func fakeHTTPHandler(w http.ResponseWriter, r *http.Request) {
 	var req rpcRequest
 	if err := json.NewDecoder(r.Body).Decode(&req); err != nil {
@@ -104,7 +133,7 @@ func TestHTTPClientJSONMode(t *testing.T) {
 		t.Fatalf("session id must be captured, got %q", c.sessionID)
 	}
 	list, err := c.ListTools(ctx)
-	if err != nil || len(list) != 2 {
+	if err != nil || len(list) != 3 {
 		t.Fatalf("list failed: %v %+v", err, list)
 	}
 	res, err := c.CallTool(ctx, "echo", json.RawMessage(`{"text":"net"}`))
@@ -130,7 +159,7 @@ func TestHTTPClientSSEMode(t *testing.T) {
 	}
 	defer c.Close()
 	list, err := c.ListTools(ctx)
-	if err != nil || len(list) != 2 {
+	if err != nil || len(list) != 3 {
 		t.Fatalf("sse list failed: %v %+v", err, list)
 	}
 }
@@ -239,3 +268,113 @@ func TestManagerRegisterTimeoutBounded(t *testing.T) {
 		t.Fatalf("timeout must bound the dial, took %s", elapsed)
 	}
 }
+
+func TestBackgroundConnectDoesNotBlockTheCaller(t *testing.T) {
+	started := make(chan struct{})
+	srv := httptest.NewServer(http.HandlerFunc(func(w http.ResponseWriter, r *http.Request) {
+		select {
+		case <-started:
+		default:
+			close(started)
+		}
+		time.Sleep(400 * time.Millisecond)
+		fakeHTTPHandler(w, r)
+	}))
+	defer srv.Close()
+	mgr, err := NewManager([]ServerConfig{{Name: "late", URL: srv.URL}})
+	if err != nil {
+		t.Fatal(err)
+	}
+	defer mgr.Close()
+	reg := tools.NewRegistry(t.TempDir())
+	t0 := time.Now()
+	mgr.StartBackground(context.Background(), reg)
+	if time.Since(t0) > 150*time.Millisecond {
+		t.Fatalf("background connect blocked the caller for %s", time.Since(t0))
+	}
+	deadline := time.Now().Add(3 * time.Second)
+	for time.Now().Before(deadline) {
+		if reg.Has("mcp__late__echo") {
+			return
+		}
+		time.Sleep(20 * time.Millisecond)
+	}
+	t.Fatal("tools did not appear after the server became ready")
+}
+
+func TestDeferredMCPToolsUseSearchAndCall(t *testing.T) {
+	mgr, err := NewManager([]ServerConfig{{Name: "fake", Command: fakeStdioConfig(t).Command, Args: fakeStdioConfig(t).Args, Env: fakeStdioConfig(t).Env, Exposure: "deferred"}})
+	if err != nil {
+		t.Fatal(err)
+	}
+	defer mgr.Close()
+	reg := tools.NewRegistry(t.TempDir())
+	if err := mgr.RegisterInto(context.Background(), reg); err != nil {
+		t.Fatal(err)
+	}
+	for _, d := range reg.Decls() {
+		if strings.HasPrefix(d.Name, "mcp__fake__") {
+			t.Fatalf("deferred mcp tool listed: %s", d.Name)
+		}
+	}
+	if !strings.Contains(reg.Search("echo"), "mcp__fake__echo") {
+		t.Fatalf("search %q", reg.Search("echo"))
+	}
+	res := reg.Execute(context.Background(), "call_tool", json.RawMessage(`{"name":"mcp__fake__echo","arguments":{"text":"via"}}`))
+	if res.IsError || !strings.Contains(res.Content, "echo: via") {
+		t.Fatalf("call %+v", res)
+	}
+}
+
+func TestDisabledServerIsSkipped(t *testing.T) {
+	off := false
+	mgr, err := NewManager([]ServerConfig{{Name: "fake", Command: "x", Enabled: &off}})
+	if err != nil {
+		t.Fatal(err)
+	}
+	reg := tools.NewRegistry(t.TempDir())
+	if err := mgr.RegisterInto(context.Background(), reg); err != nil {
+		t.Fatal(err)
+	}
+	if reg.Has("mcp__fake__echo") {
+		t.Fatal("disabled server registered a tool")
+	}
+	if err := mgr.ApplyOverrides(`{"fake":{"enabled":true,"exposure":"direct"}}`); err != nil {
+		t.Fatal(err)
+	}
+	if !mgr.Configs()[0].IsEnabled() || mgr.Configs()[0].Exposure != "direct" {
+		t.Fatalf("override %+v", mgr.Configs()[0])
+	}
+}
+
+func TestManyToolsDefaultToDeferred(t *testing.T) {
+	if chooseExposure("", 9) != tools.ExposureDeferred {
+		t.Fatal("many tools should default to deferred")
+	}
+	if chooseExposure("direct", 9) != tools.ExposureDirect {
+		t.Fatal("explicit direct must win")
+	}
+}
+
+type askClient struct{}
+
+func (askClient) Initialize(context.Context) (initializeResult, error) {
+	return initializeResult{}, nil
+}
+func (askClient) ListTools(context.Context) ([]toolInfo, error) { return nil, nil }
+func (askClient) Close() error                                  { return nil }
+func (askClient) CallTool(context.Context, string, json.RawMessage) (callToolResult, error) {
+	return callToolResult{Content: []contentBlock{{Type: "input_required", Text: "选一个颜色"}}}, nil
+}
+
+func TestInputRequiredUsesAsk(t *testing.T) {
+	tool := &bridgeTool{
+		client: askClient{}, declName: "mcp__x__ask", toolName: "ask",
+		schema: json.RawMessage(`{"type":"object","properties":{}}`),
+	}
+	ctx := WithAsk(context.Background(), func(context.Context, string) (string, error) { return "蓝色", nil })
+	res, err := tool.Execute(ctx, json.RawMessage(`{}`))
+	if err != nil || res.IsError || res.Content != "蓝色" {
+		t.Fatalf("ask result %+v %v", res, err)
+	}
+}
@@ -25,6 +25,10 @@ type ServerConfig struct {
 	Headers map[string]string `json:"headers,omitempty"`
 	// Timeout 单次请求上限（默认 30s）。
 	Timeout time.Duration `json:"timeout,omitempty"`
+	// Enabled 为 false 时跳过这台服务器。nil 表示启用。
+	Enabled *bool `json:"enabled,omitempty"`
+	// Exposure 为 direct 或 deferred。空则工具多时延迟暴露。
+	Exposure string `json:"exposure,omitempty"`
 }

 // Validate 校验配置完整性（fail-closed：命名不规范或双传输/无传输都拒绝）。
@@ -37,9 +41,19 @@ func (c ServerConfig) Validate() error {
 	if hasCmd == hasURL {
 		return fmt.Errorf("mcp server %q: exactly one of command or url is required", c.Name)
 	}
+	switch c.Exposure {
+	case "", "direct", "deferred":
+	default:
+		return fmt.Errorf("mcp server %q: exposure must be direct or deferred", c.Name)
+	}
 	return nil
 }

+// IsEnabled 报告这台服务器是否要连接。
+func (c ServerConfig) IsEnabled() bool {
+	return c.Enabled == nil || *c.Enabled
+}
+
 func validName(s string) bool {
 	if s == "" {
 		return false
@@ -0,0 +1,66 @@
+package mcp
+
+import (
+	"context"
+	"encoding/json"
+	"fmt"
+	"net/http"
+	"net/http/httptest"
+	"strings"
+	"testing"
+)
+
+type flakyClient struct {
+	fail  bool
+	calls int
+}
+
+func (f *flakyClient) Initialize(context.Context) (initializeResult, error) {
+	return initializeResult{}, nil
+}
+func (f *flakyClient) ListTools(context.Context) ([]toolInfo, error) { return nil, nil }
+func (f *flakyClient) Close() error                                  { return nil }
+func (f *flakyClient) CallTool(context.Context, string, json.RawMessage) (callToolResult, error) {
+	f.calls++
+	if f.fail {
+		return callToolResult{}, fmt.Errorf("down")
+	}
+	return callToolResult{Content: []contentBlock{{Type: "text", Text: "back"}}}, nil
+}
+
+func TestR2MCPRetriesOnceAfterReconnect(t *testing.T) {
+	first := &flakyClient{fail: true}
+	second := &flakyClient{}
+	tool := &bridgeTool{
+		client:   first,
+		toolName: "echo",
+		declName: "mcp__s__echo",
+		redial: func(context.Context) (Client, error) {
+			return second, nil
+		},
+	}
+	res, err := tool.Execute(context.Background(), []byte(`{}`))
+	if err != nil || res.IsError {
+		t.Fatalf("retry: %+v %v", res, err)
+	}
+	if first.calls != 1 || second.calls != 1 || !strings.Contains(res.Content, "back") {
+		t.Fatalf("calls first=%d second=%d content=%q", first.calls, second.calls, res.Content)
+	}
+}
+
+func TestR2HTTPCloseTerminatesSession(t *testing.T) {
+	var method, sid string
+	srv := httptest.NewServer(http.HandlerFunc(func(w http.ResponseWriter, r *http.Request) {
+		method = r.Method
+		sid = r.Header.Get("Mcp-Session-Id")
+		w.WriteHeader(http.StatusOK)
+	}))
+	defer srv.Close()
+	c := &httpClient{cfg: ServerConfig{URL: srv.URL}, client: &http.Client{}, sessionID: "sess-1"}
+	if err := c.Close(); err != nil {
+		t.Fatal(err)
+	}
+	if method != http.MethodDelete || sid != "sess-1" {
+		t.Fatalf("terminate %s session %q", method, sid)
+	}
+}
@@ -12,14 +12,24 @@ import (
 )

 // stdioClient 通过「子进程 stdin/stdout 按行分隔的 JSON-RPC」与 server 通信。
-// 同一时刻只允许一个在途请求（串行锁），匹配简单且足以支撑智能体工具调用。
+// 一个长驻读循环按 id 把响应分发给各请求；写 stdin 仍串行。
+// 取消会摘掉等待项，迟到的回复直接丢弃。读循环退出时，在等的请求全部失败，客户端标记为已死。
 type stdioClient struct {
-	cmd    *exec.Cmd
-	stdin  io.WriteCloser
-	stdout *bufio.Reader
-	mu     sync.Mutex
-	nextID int64
-	closed bool
+	cmd        *exec.Cmd
+	stdin      io.WriteCloser
+	stdout     *bufio.Reader
+	writeMu    sync.Mutex
+	mu         sync.Mutex
+	pending    map[int64]chan stdioResult
+	nextID     int64
+	readers    atomic.Int32
+	closed     bool
+	readerOnce sync.Once
+}
+
+type stdioResult struct {
+	resp rpcResponse
+	err  error
 }

 // DialStdio 启动 server 子进程并完成 initialize 握手。
@@ -43,7 +53,13 @@ func DialStdio(ctx context.Context, cfg ServerConfig) (*stdioClient, error) {
 	if err := cmd.Start(); err != nil {
 		return nil, fmt.Errorf("mcp server %q: start %s: %w", cfg.Name, cfg.Command, err)
 	}
-	c := &stdioClient{cmd: cmd, stdin: stdin, stdout: bufio.NewReaderSize(stdout, 1<<20)}
+	c := &stdioClient{
+		cmd:     cmd,
+		stdin:   stdin,
+		stdout:  bufio.NewReaderSize(stdout, 1<<20),
+		pending: map[int64]chan stdioResult{},
+	}
+	c.startReader()
 	if err := c.handshake(ctx); err != nil {
 		_ = c.Close()
 		return nil, err
@@ -104,11 +120,12 @@ func (c *stdioClient) CallTool(ctx context.Context, name string, args json.RawMe

 func (c *stdioClient) Close() error {
 	c.mu.Lock()
-	defer c.mu.Unlock()
 	if c.closed {
+		c.mu.Unlock()
 		return nil
 	}
 	c.closed = true
+	c.mu.Unlock()
 	_ = c.stdin.Close()
 	if c.cmd.Process != nil {
 		_ = c.cmd.Process.Kill()
@@ -117,29 +134,43 @@ func (c *stdioClient) Close() error {
 }

 func (c *stdioClient) roundTrip(ctx context.Context, method string, params any, out any) error {
+	c.startReader()
+	id := atomic.AddInt64(&c.nextID, 1)
+	ch := make(chan stdioResult, 1)
 	c.mu.Lock()
-	defer c.mu.Unlock()
 	if c.closed {
+		c.mu.Unlock()
 		return fmt.Errorf("mcp client closed")
 	}
-	id := atomic.AddInt64(&c.nextID, 1)
+	c.pending[id] = ch
+	c.mu.Unlock()
+
 	req := rpcRequest{JSONRPC: "2.0", ID: id, Method: method, Params: params}
 	body, err := json.Marshal(req)
 	if err != nil {
+		c.dropPending(id)
 		return err
 	}
 	body = append(body, '\n')
-	if _, err := c.stdin.Write(body); err != nil {
+	c.writeMu.Lock()
+	_, err = c.stdin.Write(body)
+	c.writeMu.Unlock()
+	if err != nil {
+		c.dropPending(id)
 		return fmt.Errorf("mcp write %s: %w", method, err)
 	}
-	line, err := c.readResponseLine(ctx)
-	if err != nil {
-		return fmt.Errorf("mcp read %s: %w", method, err)
+
+	var got stdioResult
+	select {
+	case got = <-ch:
+	case <-ctx.Done():
+		c.dropPending(id)
+		return fmt.Errorf("mcp read %s: %w", method, ctx.Err())
 	}
-	var resp rpcResponse
-	if err := json.Unmarshal(line, &resp); err != nil {
-		return fmt.Errorf("mcp %s: bad response frame: %w", method, err)
+	if got.err != nil {
+		return fmt.Errorf("mcp read %s: %w", method, got.err)
 	}
+	resp := got.resp
 	if resp.ID == nil || *resp.ID != id {
 		return fmt.Errorf("mcp %s: response id mismatch", method)
 	}
@@ -152,40 +183,60 @@ func (c *stdioClient) roundTrip(ctx context.Context, method string, params any,
 	return nil
 }

-// readResponseLine 逐行读取，跳过通知帧（无 id），带 ctx 取消感知。
-func (c *stdioClient) readResponseLine(ctx context.Context) ([]byte, error) {
-	type result struct {
-		line []byte
-		err  error
-	}
-	done := make(chan result, 1)
-	go func() {
-		for {
-			line, err := c.stdout.ReadBytes('\n')
-			if err != nil {
-				done <- result{nil, err}
-				return
-			}
-			var probe rpcResponse
-			if json.Unmarshal(line, &probe) == nil && probe.ID != nil {
-				done <- result{line, nil}
-				return
-			}
-			// 通知或日志行：忽略。
+func (c *stdioClient) dropPending(id int64) {
+	c.mu.Lock()
+	delete(c.pending, id)
+	c.mu.Unlock()
+}
+
+func (c *stdioClient) startReader() {
+	c.readerOnce.Do(func() {
+		go c.readLoop()
+	})
+}
+
+func (c *stdioClient) readLoop() {
+	c.readers.Add(1)
+	defer c.readers.Add(-1)
+	var fatal error
+	for {
+		line, err := c.stdout.ReadBytes('\n')
+		if err != nil {
+			fatal = err
+			break
 		}
-	}()
-	select {
-	case r := <-done:
-		return r.line, r.err
-	case <-ctx.Done():
-		return nil, ctx.Err()
+		var resp rpcResponse
+		if json.Unmarshal(line, &resp) != nil || resp.ID == nil {
+			continue
+		}
+		c.mu.Lock()
+		ch := c.pending[*resp.ID]
+		if ch != nil {
+			delete(c.pending, *resp.ID)
+		}
+		c.mu.Unlock()
+		if ch == nil {
+			continue
+		}
+		ch <- stdioResult{resp: resp}
+	}
+	c.mu.Lock()
+	c.closed = true
+	pending := c.pending
+	c.pending = map[int64]chan stdioResult{}
+	c.mu.Unlock()
+	for _, ch := range pending {
+		ch <- stdioResult{err: fatal}
 	}
 }

+func (c *stdioClient) Readers() int { return int(c.readers.Load()) }
+
 func (c *stdioClient) notify(method string) error {
 	c.mu.Lock()
-	defer c.mu.Unlock()
-	if c.closed {
+	closed := c.closed
+	c.mu.Unlock()
+	if closed {
 		return fmt.Errorf("mcp client closed")
 	}
 	body, err := json.Marshal(rpcNotification{JSONRPC: "2.0", Method: method})
@@ -193,6 +244,8 @@ func (c *stdioClient) notify(method string) error {
 		return err
 	}
 	body = append(body, '\n')
+	c.writeMu.Lock()
+	defer c.writeMu.Unlock()
 	_, err = c.stdin.Write(body)
 	return err
 }
@@ -0,0 +1,275 @@
+// Package policy 决定一次工具调用在当前权限模式下是拒绝、询问还是放行。
+// 默认模式不额外表态，调用方继续用原有画像策略。
+package policy
+
+import (
+	"strings"
+	"sync"
+)
+
+type Mode string
+
+const (
+	ModeDefault  Mode = "default"
+	ModePlan     Mode = "plan"
+	ModeAutoEdit Mode = "auto_edit"
+)
+
+type Effect string
+
+const (
+	EffectNone  Effect = ""
+	EffectDeny  Effect = "deny"
+	EffectAsk   Effect = "ask"
+	EffectAllow Effect = "allow"
+)
+
+type Traits struct {
+	ReadOnly bool
+	Edits    bool
+	Risk     string
+}
+
+type Rule struct {
+	Effect  Effect
+	Tool    string
+	ArgGlob string
+}
+
+type Request struct {
+	Mode         Mode
+	Name         string
+	Args         string
+	Traits       Traits
+	SessionAllow bool
+}
+
+func Normalize(mode string) Mode {
+	switch Mode(mode) {
+	case ModePlan, ModeAutoEdit, ModeDefault:
+		return Mode(mode)
+	default:
+		return ModeDefault
+	}
+}
+
+// Decide 的优先级是拒绝规则、计划模式禁写、本会话放行、询问规则、自动改文件、高风险命令。
+// 都没有命中时返回空，表示沿用画像的拒绝、确认和放行。
+func Decide(req Request, rules []Rule) Effect {
+	var sawAsk, sawAllow bool
+	for _, rule := range rules {
+		if !matchTool(rule.Tool, req.Name) || !glob(rule.ArgGlob, req.Args) {
+			continue
+		}
+		switch rule.Effect {
+		case EffectDeny:
+			return EffectDeny
+		case EffectAsk:
+			sawAsk = true
+		case EffectAllow:
+			sawAllow = true
+		}
+	}
+	traits := mergeTraits(req.Name, req.Traits)
+	if commandRisk(req.Name, req.Args) == "high" {
+		traits.Risk = "high"
+	}
+	if Normalize(string(req.Mode)) == ModePlan && !planVisible(req.Name, traits.ReadOnly) {
+		return EffectDeny
+	}
+	if req.SessionAllow {
+		return EffectAllow
+	}
+	if sawAsk {
+		return EffectAsk
+	}
+	if Normalize(string(req.Mode)) == ModeAutoEdit && traits.Edits && traits.Risk != "high" {
+		return EffectAllow
+	}
+	if sawAllow {
+		return EffectAllow
+	}
+	if traits.Risk == "high" && (req.Name == "run_shell" || req.Name == "run_command") {
+		return EffectAsk
+	}
+	return EffectNone
+}
+
+// SessionAllows 记录并查询「本会话始终允许」的决定。键的口径由实现方定义；
+// 宿主可以用自己的模式匹配（如参数通配）实现同一接口。
+type SessionAllows interface {
+	Remember(name, args string)
+	Allowed(name, args string) bool
+}
+
+// MemoryAllows 是进程内的会话放行存储，按「工具名 + 原样参数串」精确匹配。
+type MemoryAllows struct {
+	mu      sync.Mutex
+	allowed map[string]bool
+}
+
+func NewMemoryAllows() *MemoryAllows {
+	return &MemoryAllows{allowed: make(map[string]bool)}
+}
+
+func (m *MemoryAllows) Remember(name, args string) {
+	m.mu.Lock()
+	m.allowed[name+"\n"+args] = true
+	m.mu.Unlock()
+}
+
+func (m *MemoryAllows) Allowed(name, args string) bool {
+	m.mu.Lock()
+	defer m.mu.Unlock()
+	return m.allowed[name+"\n"+args]
+}
+
+// Policy 是一次会话的权限配置：模式、规则与会话放行存在同一个对象上，
+// Decide/Visible 语义与同名函数完全一致。
+type Policy struct {
+	Mode          Mode
+	Rules         []Rule
+	SessionAllows SessionAllows // 可为 nil：等价于 SessionAllow 恒为 false
+}
+
+// Decide 按 Policy 的模式、规则与会话放行判定一次工具调用。
+func (p Policy) Decide(name, args string, traits Traits) Effect {
+	var allowed bool
+	if p.SessionAllows != nil {
+		allowed = p.SessionAllows.Allowed(name, args)
+	}
+	return Decide(Request{Mode: p.Mode, Name: name, Args: args, Traits: traits, SessionAllow: allowed}, p.Rules)
+}
+
+// Visible 报告该模式下这个工具能否出现在清单里。
+func (p Policy) Visible(name string) bool {
+	return Visible(string(p.Mode), name, mergeTraits(name, Traits{}).ReadOnly)
+}
+
+// Visible 报告计划模式下这个工具能否出现在清单里。其他模式一律可见。
+func Visible(mode string, name string, readOnly bool) bool {
+	if Normalize(mode) != ModePlan {
+		return true
+	}
+	return planVisible(name, readOnly)
+}
+
+func planVisible(name string, readOnly bool) bool {
+	switch name {
+	case "update_plan", "update_todos", "ask_user", "exit_plan_mode":
+		return true
+	default:
+		return readOnly
+	}
+}
+
+// TraitsFor 是内置工具的只读、改文件和风险声明。未知工具为零值。
+func TraitsFor(name string) Traits {
+	switch name {
+	case "read_file", "list_dir", "grep", "glob", "load_skill",
+		"fs_read_file", "fs_list_dir", "time_now", "knowledge_search",
+		"list_conversations", "read_conversation", "quality_review",
+		"web_search", "tavily_search", "fetch_url", "ask_user", "exit_plan_mode",
+		"update_plan", "update_todos", "tool_search":
+		return Traits{ReadOnly: true, Risk: "low"}
+	case "write_file", "edit_file", "multi_edit", "fs_write_file", "write_artifact", "register_artifact":
+		return Traits{Edits: true, Risk: "medium"}
+	case "run_shell", "run_command", "run_python", "run_tool_script", "call_tool":
+		return Traits{Risk: "medium"}
+	default:
+		return Traits{}
+	}
+}
+
+func mergeTraits(name string, given Traits) Traits {
+	base := TraitsFor(name)
+	if given.ReadOnly {
+		base.ReadOnly = true
+	}
+	if given.Edits {
+		base.Edits = true
+	}
+	if given.Risk != "" {
+		base.Risk = given.Risk
+	}
+	return base
+}
+
+// ClassifyCommand 按命令文本分级。高风险会在默认模式下改为询问。
+func ClassifyCommand(cmd string) string {
+	c := strings.ToLower(cmd)
+	for _, n := range []string{"rm -rf", "rm -fr", "sudo ", "mkfs", "shutdown", "reboot", "> /dev/", "| sh", "|sh", "chmod -r 777"} {
+		if strings.Contains(c, n) {
+			return "high"
+		}
+	}
+	for _, n := range []string{"rm ", "mv ", "chmod ", "chown ", "kill "} {
+		if strings.Contains(c, n) {
+			return "medium"
+		}
+	}
+	return "low"
+}
+
+func commandRisk(name, args string) string {
+	if name != "run_shell" && name != "run_command" {
+		return ""
+	}
+	return ClassifyCommand(extractCommand(args))
+}
+
+func extractCommand(args string) string {
+	for _, key := range []string{`"command"`, `"cmd"`} {
+		i := strings.Index(args, key)
+		if i < 0 {
+			continue
+		}
+		rest := args[i+len(key):]
+		rest = strings.TrimLeft(rest, " \t\n:")
+		if !strings.HasPrefix(rest, `"`) {
+			continue
+		}
+		rest = rest[1:]
+		if j := strings.Index(rest, `"`); j >= 0 {
+			return rest[:j]
+		}
+	}
+	return args
+}
+
+func matchTool(pattern, name string) bool {
+	if pattern == "" || pattern == "*" {
+		return true
+	}
+	return glob(pattern, name)
+}
+
+func glob(pattern, s string) bool {
+	if pattern == "" || pattern == "*" {
+		return true
+	}
+	parts := strings.Split(pattern, "*")
+	if !strings.HasPrefix(pattern, "*") {
+		if !strings.HasPrefix(s, parts[0]) {
+			return false
+		}
+	}
+	if !strings.HasSuffix(pattern, "*") {
+		last := parts[len(parts)-1]
+		if !strings.HasSuffix(s, last) {
+			return false
+		}
+	}
+	idx := 0
+	for _, part := range parts {
+		if part == "" {
+			continue
+		}
+		j := strings.Index(s[idx:], part)
+		if j < 0 {
+			return false
+		}
+		idx += j + len(part)
+	}
+	return true
+}
@@ -0,0 +1,90 @@
+package policy
+
+import "testing"
+
+func TestPlanModeDeniesWrites(t *testing.T) {
+	effect := Decide(Request{Mode: ModePlan, Name: "write_file", Args: `{"path":"a.md"}`}, nil)
+	if effect != EffectDeny {
+		t.Fatalf("effect %q", effect)
+	}
+	if !Visible(string(ModePlan), "update_plan", false) {
+		t.Fatal("plan tool hidden")
+	}
+	if Visible(string(ModePlan), "write_file", false) {
+		t.Fatal("write visible in plan mode")
+	}
+}
+
+func TestDefaultModeStaysQuiet(t *testing.T) {
+	if got := Decide(Request{Mode: ModeDefault, Name: "write_file"}, nil); got != EffectNone {
+		t.Fatalf("default write %q", got)
+	}
+	if got := Decide(Request{Mode: ModeDefault, Name: "run_shell", Args: `{"command":"rm -rf /"}`}, nil); got != EffectAsk {
+		t.Fatalf("high shell %q", got)
+	}
+}
+
+func TestRulePrecedenceAndSessionAllow(t *testing.T) {
+	rules := []Rule{{Effect: EffectAllow, Tool: "write_file"}, {Effect: EffectDeny, Tool: "write_*", ArgGlob: "*secret*"}}
+	if got := Decide(Request{Name: "write_file", Args: `{"path":"secret.txt"}`}, rules); got != EffectDeny {
+		t.Fatalf("deny lost: %q", got)
+	}
+	if got := Decide(Request{Name: "http.fetch", SessionAllow: true}, []Rule{{Effect: EffectAsk, Tool: "http.*"}}); got != EffectAllow {
+		t.Fatalf("session allow %q", got)
+	}
+}
+
+func TestAutoEditAllowsWorkspaceWrites(t *testing.T) {
+	if got := Decide(Request{Mode: ModeAutoEdit, Name: "fs_write_file"}, nil); got != EffectAllow {
+		t.Fatalf("auto edit %q", got)
+	}
+}
+
+// TestPolicyDecideMatchesFunctionForm：Policy 结构体形态与函数形态判定一致。
+func TestPolicyDecideMatchesFunctionForm(t *testing.T) {
+	rules := []Rule{{Effect: EffectAllow, Tool: "write_file"}, {Effect: EffectDeny, Tool: "write_*", ArgGlob: "*secret*"}}
+	cases := []struct {
+		p    Policy
+		name string
+		args string
+		want Effect
+	}{
+		{Policy{Mode: ModePlan}, "write_file", `{"path":"a.md"}`, EffectDeny},
+		{Policy{Mode: ModeDefault, Rules: rules}, "write_file", `{"path":"a.md"}`, EffectAllow},
+		{Policy{Mode: ModeDefault, Rules: rules}, "write_file", `{"path":"secret.txt"}`, EffectDeny},
+		{Policy{Mode: ModeDefault}, "run_shell", `{"command":"rm -rf /"}`, EffectAsk},
+	}
+	for i, c := range cases {
+		if got := c.p.Decide(c.name, c.args, Traits{}); got != c.want {
+			t.Fatalf("case %d: effect %q, want %q", i, got, c.want)
+		}
+	}
+}
+
+// TestPolicySessionAllowsStore：会话放行经存储短路 ask，且跨调用存活。
+func TestPolicySessionAllowsStore(t *testing.T) {
+	store := NewMemoryAllows()
+	p := Policy{Mode: ModeDefault, Rules: []Rule{{Effect: EffectAsk, Tool: "run_shell"}}, SessionAllows: store}
+	if got := p.Decide("run_shell", `{"command":"ls"}`, Traits{}); got != EffectAsk {
+		t.Fatalf("before remember: %q, want ask", got)
+	}
+	store.Remember("run_shell", `{"command":"ls"}`)
+	if got := p.Decide("run_shell", `{"command":"ls"}`, Traits{}); got != EffectAllow {
+		t.Fatalf("after remember: %q, want allow", got)
+	}
+	if got := p.Decide("run_shell", `{"command":"rm -rf /"}`, Traits{}); got != EffectAsk {
+		t.Fatalf("other args: %q, want ask（精确匹配不放行）", got)
+	}
+}
+
+// TestPolicyNilStoreAndVisible：nil 存储等价于 SessionAllow 恒 false；Visible 方法可用。
+func TestPolicyNilStoreAndVisible(t *testing.T) {
+	p := Policy{Mode: ModePlan}
+	if got := p.Decide("write_file", `{}`, Traits{}); got != EffectDeny {
+		t.Fatalf("plan deny: %q", got)
+	}
+	if !p.Visible("update_plan") || p.Visible("write_file") {
+		t.Fatal("policy visible mismatch")
+	}
+}
+
@@ -0,0 +1,130 @@
+package runtime
+
+import (
+	"context"
+	"encoding/json"
+	"path/filepath"
+	"strings"
+	"testing"
+
+	"agent-native/runtime/llm"
+	"agent-native/runtime/session"
+	"agent-native/runtime/tools"
+)
+
+type countingTool struct {
+	name  string
+	calls int
+	hook  func()
+}
+
+func (c *countingTool) Name() string        { return c.name }
+func (c *countingTool) Description() string { return c.name }
+func (c *countingTool) Parameters() json.RawMessage {
+	return json.RawMessage(`{"type":"object","properties":{}}`)
+}
+func (c *countingTool) Execute(ctx context.Context, args json.RawMessage) (tools.Result, error) {
+	c.calls++
+	if c.hook != nil {
+		c.hook()
+	}
+	return tools.Result{Content: "ok"}, nil
+}
+
+func TestR2TokenBudgetSkipsTools(t *testing.T) {
+	tool := &countingTool{name: "list_dir"}
+	reg := tools.NewRegistry(t.TempDir())
+	reg.Register(tool)
+	res, err := Run(context.Background(), Config{
+		Model:  "fake",
+		Tools:  reg,
+		Budget: &Budget{MaxPromptTokens: 10},
+		Stream: usageScript(llm.Usage{PromptTokens: 500},
+			llm.Message{Role: llm.RoleAssistant, StopReason: llm.StopToolCalls, ToolCalls: []llm.ToolCall{{
+				ID: "c1", Name: "list_dir", Arguments: `{}`,
+			}}},
+		),
+	}, llm.Message{Role: llm.RoleUser, Content: "go"})
+	if err != nil {
+		t.Fatal(err)
+	}
+	if tool.calls != 0 {
+		t.Fatalf("tool ran after token budget: %d", tool.calls)
+	}
+	if res.Err == nil || !strings.Contains(res.Err.Error(), "prompt tokens") {
+		t.Fatalf("budget error: %v", res.Err)
+	}
+}
+
+func TestR2StopReplacementIsPersisted(t *testing.T) {
+	store, err := session.Open(filepath.Join(t.TempDir(), "t.jsonl"))
+	if err != nil {
+		t.Fatal(err)
+	}
+	_, err = Run(context.Background(), Config{
+		Model:    "fake",
+		MaxTurns: 3,
+		Session:  store,
+		Stream: script(
+			llm.Message{Role: llm.RoleAssistant, Content: "draft", StopReason: llm.StopEnd},
+			llm.Message{Role: llm.RoleAssistant, Content: "final", StopReason: llm.StopEnd},
+		),
+		Stop: func(last llm.Message, info StopInfo) StopDecision {
+			if info.Continuations == 0 {
+				return StopDecision{Continue: true, Replacement: "corrected-claim"}
+			}
+			return StopDecision{}
+		},
+	}, llm.Message{Role: llm.RoleUser, Content: "go"})
+	if err != nil {
+		t.Fatal(err)
+	}
+	msgs, err := store.Load()
+	if err != nil {
+		t.Fatal(err)
+	}
+	var saw bool
+	for _, m := range msgs {
+		if m.Content == "corrected-claim" {
+			saw = true
+		}
+	}
+	if !saw {
+		t.Fatalf("replacement missing from session: %+v", msgs)
+	}
+}
+
+func TestR2SequentialToolsStopAfterCancel(t *testing.T) {
+	ctx, cancel := context.WithCancel(context.Background())
+	first := &countingTool{name: "first", hook: cancel}
+	second := &countingTool{name: "second"}
+	reg := tools.NewRegistry(t.TempDir())
+	reg.Register(first)
+	reg.Register(second)
+	res, err := Run(ctx, Config{
+		Model: "fake",
+		Tools: reg,
+		Stream: script(
+			llm.Message{Role: llm.RoleAssistant, StopReason: llm.StopToolCalls, ToolCalls: []llm.ToolCall{
+				{ID: "1", Name: "first", Arguments: `{}`},
+				{ID: "2", Name: "second", Arguments: `{}`},
+			}},
+			llm.Message{Role: llm.RoleAssistant, Content: "after", StopReason: llm.StopEnd},
+		),
+	}, llm.Message{Role: llm.RoleUser, Content: "go"})
+	if err != nil {
+		t.Fatal(err)
+	}
+	if first.calls != 1 || second.calls != 0 {
+		t.Fatalf("calls first=%d second=%d", first.calls, second.calls)
+	}
+	var paired int
+	for _, m := range res.Messages {
+		if m.Role == llm.RoleTool {
+			paired++
+		}
+	}
+	if paired != 2 {
+		t.Fatalf("tool results = %d, want a pair for each call", paired)
+	}
+}
@@ -0,0 +1,97 @@
+// Package reminders 收集一轮里要带给模型的提醒，并做节流。
+// 提醒并入最后一条工具结果或用户消息，不另起一条 user 消息。
+package reminders
+
+import (
+	"strings"
+	"sync"
+	"time"
+
+	"agent-native/runtime/llm"
+)
+
+type Kind string
+
+const (
+	KindTodo        Kind = "todo"
+	KindPlan        Kind = "plan"
+	KindBackground  Kind = "background"
+	KindDeliverable Kind = "deliverable"
+	KindBudget      Kind = "budget"
+)
+
+// Note 是一条待投递的提醒。Key 相同的提醒在本轮只投递一次。
+type Note struct {
+	Kind Kind
+	Key  string
+	Text string
+}
+
+// Bus 按最小间隔、每 run 上限和去重决定是否投递。
+type Bus struct {
+	mu     sync.Mutex
+	minGap time.Duration
+	cap    int
+	sent   int
+	last   time.Time
+	seen   map[string]struct{}
+}
+
+func New(minGap time.Duration, cap int) *Bus {
+	if cap <= 0 {
+		cap = 3
+	}
+	return &Bus{minGap: minGap, cap: cap, seen: map[string]struct{}{}}
+}
+
+// Admit 在通过节流时返回真。同一 Kind+Key 只通过一次。
+func (b *Bus) Admit(n Note, now time.Time) bool {
+	if b == nil {
+		return false
+	}
+	b.mu.Lock()
+	defer b.mu.Unlock()
+	if b.sent >= b.cap {
+		return false
+	}
+	if b.minGap > 0 && !b.last.IsZero() && now.Sub(b.last) < b.minGap {
+		return false
+	}
+	key := string(n.Kind) + "\n" + n.Key
+	if _, ok := b.seen[key]; ok {
+		return false
+	}
+	b.seen[key] = struct{}{}
+	b.last = now
+	b.sent++
+	return true
+}
+
+// Attach 把提醒正文并入最后一条工具结果或用户消息。
+// 找不到这两种消息时原样返回，不追加 user 消息。
+func Attach(messages []llm.Message, texts []string) []llm.Message {
+	var lines []string
+	for _, text := range texts {
+		text = strings.TrimSpace(text)
+		if text != "" {
+			lines = append(lines, text)
+		}
+	}
+	if len(lines) == 0 || len(messages) == 0 {
+		return messages
+	}
+	block := "<system-reminder>\n" + strings.Join(lines, "\n") + "\n</system-reminder>"
+	out := append([]llm.Message{}, messages...)
+	for i := len(out) - 1; i >= 0; i-- {
+		if out[i].Role != llm.RoleTool && out[i].Role != llm.RoleUser {
+			continue
+		}
+		if out[i].Content != "" {
+			out[i].Content += "\n" + block
+		} else {
+			out[i].Content = block
+		}
+		return out
+	}
+	return messages
+}
@@ -0,0 +1,53 @@
+package reminders
+
+import (
+	"strings"
+	"testing"
+	"time"
+
+	"agent-native/runtime/llm"
+)
+
+func TestAttachDoesNotAddUserAfterTool(t *testing.T) {
+	msgs := []llm.Message{
+		{Role: llm.RoleUser, Content: "请改文件"},
+		{Role: llm.RoleAssistant, ToolCalls: []llm.ToolCall{{ID: "c1", Name: "read_file"}}},
+		{Role: llm.RoleTool, Name: "read_file", ToolCallID: "c1", Content: "原文"},
+	}
+	out := Attach(msgs, []string{"计划和待办已过期"})
+	if len(out) != len(msgs) {
+		t.Fatalf("len %d, want %d", len(out), len(msgs))
+	}
+	for i := range msgs {
+		if out[i].Role != msgs[i].Role {
+			t.Fatalf("role changed at %d: %s", i, out[i].Role)
+		}
+	}
+	last := out[len(out)-1]
+	if last.Role != llm.RoleTool || !strings.Contains(last.Content, "原文") || !strings.Contains(last.Content, "system-reminder") {
+		t.Fatalf("reminder not folded into tool result: %+v", last)
+	}
+	if strings.Contains(out[0].Content, "system-reminder") {
+		t.Fatal("reminder leaked onto the earlier user message")
+	}
+}
+
+func TestAdmitThrottles(t *testing.T) {
+	bus := New(time.Minute, 2)
+	now := time.Now()
+	if !bus.Admit(Note{Kind: KindBudget, Key: "near", Text: "预算将尽"}, now) {
+		t.Fatal("first admit")
+	}
+	if bus.Admit(Note{Kind: KindBudget, Key: "near", Text: "预算将尽"}, now.Add(time.Second)) {
+		t.Fatal("duplicate key admitted")
+	}
+	if bus.Admit(Note{Kind: KindBackground, Key: "job", Text: "后台结束"}, now.Add(time.Second)) {
+		t.Fatal("inside min gap admitted")
+	}
+	if !bus.Admit(Note{Kind: KindBackground, Key: "job", Text: "后台结束"}, now.Add(time.Minute)) {
+		t.Fatal("after gap")
+	}
+	if bus.Admit(Note{Kind: KindDeliverable, Key: "doc", Text: "交付物待确认"}, now.Add(2*time.Minute)) {
+		t.Fatal("over cap")
+	}
+}
@@ -0,0 +1,37 @@
+package resources
+
+import (
+	"os"
+	"path/filepath"
+	"strings"
+)
+
+// ContextFiles walks from dir toward the filesystem root and concatenates
+// AGENTS.md and CLAUDE.md (nearest first). SYSTEM.md and APPEND_SYSTEM.md
+// are read from dir only.
+func ContextFiles(dir string) (string, error) {
+	var parts []string
+	cur := dir
+	for {
+		for _, name := range []string{"AGENTS.md", "CLAUDE.md"} {
+			raw, err := os.ReadFile(filepath.Join(cur, name))
+			if err != nil {
+				continue
+			}
+			parts = append(parts, "# "+name+" ("+cur+")\n"+string(raw))
+		}
+		parent := filepath.Dir(cur)
+		if parent == cur {
+			break
+		}
+		cur = parent
+	}
+	for _, name := range []string{"SYSTEM.md", "APPEND_SYSTEM.md"} {
+		raw, err := os.ReadFile(filepath.Join(dir, name))
+		if err != nil {
+			continue
+		}
+		parts = append(parts, "# "+name+"\n"+string(raw))
+	}
+	return strings.Join(parts, "\n\n"), nil
+}
@@ -0,0 +1,44 @@
+package resources
+
+import (
+	"os"
+	"path/filepath"
+	"strings"
+)
+
+// Expand replaces {{var}} placeholders and /name includes.
+// An include reads <dir>/<name>.md. Unknown variables are left intact.
+func Expand(text, dir string, vars map[string]string) (string, error) {
+	out := text
+	for key, value := range vars {
+		out = strings.ReplaceAll(out, "{{"+key+"}}", value)
+	}
+	for {
+		i := strings.Index(out, "/")
+		if i < 0 {
+			break
+		}
+		rest := out[i+1:]
+		j := 0
+		for j < len(rest) && isName(rest[j]) {
+			j++
+		}
+		if j == 0 || (i > 0 && out[i-1] != ' ' && out[i-1] != '\n') {
+			break
+		}
+		name := rest[:j]
+		raw, err := os.ReadFile(filepath.Join(dir, name+".md"))
+		if err != nil {
+			break
+		}
+		out = out[:i] + string(raw) + out[i+1+j:]
+	}
+	for key, value := range vars {
+		out = strings.ReplaceAll(out, "{{"+key+"}}", value)
+	}
+	return out, nil
+}
+
+func isName(b byte) bool {
+	return b == '-' || b == '_' || (b >= 'a' && b <= 'z') || (b >= 'A' && b <= 'Z') || (b >= '0' && b <= '9')
+}
@@ -0,0 +1,72 @@
+package resources
+
+import (
+	"os"
+	"path/filepath"
+	"strings"
+	"testing"
+
+	"agent-native/runtime/skills"
+	"agent-native/runtime/tools"
+)
+
+func TestContextAndPromptAndTrust(t *testing.T) {
+	root := t.TempDir()
+	proj := filepath.Join(root, "proj")
+	if err := os.MkdirAll(proj, 0o755); err != nil {
+		t.Fatal(err)
+	}
+	if err := os.WriteFile(filepath.Join(root, "AGENTS.md"), []byte("root-agents"), 0o644); err != nil {
+		t.Fatal(err)
+	}
+	if err := os.WriteFile(filepath.Join(proj, "SYSTEM.md"), []byte("sys"), 0o644); err != nil {
+		t.Fatal(err)
+	}
+	text, err := ContextFiles(proj)
+	if err != nil || !strings.Contains(text, "root-agents") || !strings.Contains(text, "sys") {
+		t.Fatalf("context: %s %v", text, err)
+	}
+	if err := os.WriteFile(filepath.Join(proj, "greet.md"), []byte("hello {{name}}"), 0o644); err != nil {
+		t.Fatal(err)
+	}
+	got, err := Expand("say /greet", proj, map[string]string{"name": "Ada"})
+	if err != nil || got != "say hello Ada" {
+		t.Fatalf("expand: %q %v", got, err)
+	}
+	if ProjectTrusted(TrustAsk, false, false) {
+		t.Fatal("ask is not trusted without approve")
+	}
+	if !ProjectTrusted(TrustNever, true, false) {
+		t.Fatal("approve overrides never")
+	}
+	if ProjectTrusted(TrustAlways, false, true) {
+		t.Fatal("no-approve wins")
+	}
+}
+
+func TestPiAliasesFilterAndSkillCommand(t *testing.T) {
+	reg := tools.WorkspaceTools(t.TempDir())
+	reg.RegisterPiAliases()
+	if !reg.Has("read") || !reg.Has("ls") || !reg.Has("read_file") {
+		t.Fatal("aliases missing")
+	}
+	if err := reg.Filter(tools.FilterOptions{Exclude: []string{"read"}, NoBuiltin: false}); err != nil {
+		t.Fatal(err)
+	}
+	if reg.Has("read") || !reg.Has("read_file") {
+		t.Fatal("exclude should drop only the alias")
+	}
+	dir := t.TempDir()
+	skillDir := filepath.Join(dir, "review")
+	if err := os.MkdirAll(skillDir, 0o755); err != nil {
+		t.Fatal(err)
+	}
+	if err := os.WriteFile(filepath.Join(skillDir, "SKILL.md"), []byte("be thorough"), 0o644); err != nil {
+		t.Fatal(err)
+	}
+	store := skills.NewStore(dir)
+	out, err := store.ExpandCommands("use /skill:review now")
+	if err != nil || !strings.Contains(out, "be thorough") {
+		t.Fatalf("skill expand: %q %v", out, err)
+	}
+}
@@ -0,0 +1,102 @@
+package resources
+
+import (
+	"encoding/json"
+	"os"
+	"path/filepath"
+)
+
+// Trust is how a project directory is treated in non-interactive runs.
+type Trust string
+
+const (
+	TrustAsk    Trust = "ask"
+	TrustAlways Trust = "always"
+	TrustNever  Trust = "never"
+)
+
+// Settings is the global plus project configuration.
+type Settings struct {
+	Trust    Trust             `json:"trust,omitempty"`
+	Model    string            `json:"model,omitempty"`
+	Thinking string            `json:"thinking,omitempty"`
+	Tools    []string          `json:"tools,omitempty"`
+	Exclude  []string          `json:"exclude,omitempty"`
+	Vars     map[string]string `json:"vars,omitempty"`
+}
+
+// LoadSettings reads global (optional) then project settings.
+// Project keys override global keys. A missing file is an empty settings.
+func LoadSettings(globalPath, projectPath string) (Settings, error) {
+	global, err := readSettings(globalPath)
+	if err != nil {
+		return Settings{}, err
+	}
+	project, err := readSettings(projectPath)
+	if err != nil {
+		return Settings{}, err
+	}
+	return mergeSettings(global, project), nil
+}
+
+// ProjectTrusted reports whether project resources may be loaded.
+// ask and never refuse in non-interactive mode. approve forces trust.
+func ProjectTrusted(trust Trust, approve, noApprove bool) bool {
+	if noApprove {
+		return false
+	}
+	if approve {
+		return true
+	}
+	return trust == TrustAlways
+}
+
+func readSettings(path string) (Settings, error) {
+	if path == "" {
+		return Settings{}, nil
+	}
+	raw, err := os.ReadFile(path)
+	if err != nil {
+		if os.IsNotExist(err) {
+			return Settings{}, nil
+		}
+		return Settings{}, err
+	}
+	var s Settings
+	if err := json.Unmarshal(raw, &s); err != nil {
+		return Settings{}, err
+	}
+	return s, nil
+}
+
+func mergeSettings(base, over Settings) Settings {
+	if over.Trust != "" {
+		base.Trust = over.Trust
+	}
+	if over.Model != "" {
+		base.Model = over.Model
+	}
+	if over.Thinking != "" {
+		base.Thinking = over.Thinking
+	}
+	if over.Tools != nil {
+		base.Tools = over.Tools
+	}
+	if over.Exclude != nil {
+		base.Exclude = over.Exclude
+	}
+	if len(over.Vars) > 0 {
+		if base.Vars == nil {
+			base.Vars = map[string]string{}
+		}
+		for k, v := range over.Vars {
+			base.Vars[k] = v
+		}
+	}
+	return base
+}
+
+// DefaultPaths returns the global and project settings paths for a cwd.
+func DefaultPaths(home, cwd string) (string, string) {
+	return filepath.Join(home, ".agent-native", "settings.json"), filepath.Join(cwd, ".agent-native", "settings.json")
+}
@@ -0,0 +1,105 @@
+package session
+
+import (
+	"os"
+	"path/filepath"
+	"sort"
+	"strings"
+)
+
+// Info describes one session file in a directory.
+type Info struct {
+	ID      string
+	Path    string
+	Name    string
+	CWD     string
+	Leaf    string
+	Updated string
+}
+
+// List returns session files in dir, newest first.
+func List(dir string) ([]Info, error) {
+	ents, err := os.ReadDir(dir)
+	if err != nil {
+		if os.IsNotExist(err) {
+			return nil, nil
+		}
+		return nil, err
+	}
+	var out []Info
+	for _, ent := range ents {
+		name := ent.Name()
+		if ent.IsDir() || !strings.HasSuffix(name, ".jsonl") {
+			continue
+		}
+		path := filepath.Join(dir, name)
+		st, err := Open(path)
+		if err != nil {
+			continue
+		}
+		info, err := st.Info()
+		if err != nil {
+			continue
+		}
+		out = append(out, info)
+	}
+	sort.Slice(out, func(i, j int) bool { return out[i].Updated > out[j].Updated })
+	return out, nil
+}
+
+// Continue opens the newest session whose cwd matches. A second return
+// value of false means there is nothing to continue.
+func Continue(dir, cwd string) (*Store, bool, error) {
+	list, err := List(dir)
+	if err != nil {
+		return nil, false, err
+	}
+	for _, info := range list {
+		if cwd == "" || info.CWD == cwd {
+			st, err := Open(info.Path)
+			return st, err == nil, err
+		}
+	}
+	return nil, false, nil
+}
+
+// Resume opens a session by id (filename stem) or by name.
+func Resume(dir, idOrName string) (*Store, error) {
+	list, err := List(dir)
+	if err != nil {
+		return nil, err
+	}
+	for _, info := range list {
+		if info.ID == idOrName || info.Name == idOrName {
+			return Open(info.Path)
+		}
+	}
+	return nil, os.ErrNotExist
+}
+
+// SetCWD records the working directory used by Continue.
+func (s *Store) SetCWD(cwd string) error {
+	s.mu.Lock()
+	defer s.mu.Unlock()
+	return s.writeMetaKeyLocked("cwd", cwd)
+}
+
+// Info reads the catalog fields for this file.
+func (s *Store) Info() (Info, error) {
+	s.mu.Lock()
+	defer s.mu.Unlock()
+	entries, err := s.readEntriesLocked()
+	if err != nil {
+		return Info{}, err
+	}
+	meta := s.readMetaLocked()
+	updated := ""
+	if len(entries) > 0 {
+		updated = entries[len(entries)-1].Timestamp
+	}
+	id := strings.TrimSuffix(filepath.Base(s.path), ".jsonl")
+	return Info{
+		ID: id, Path: s.path, Name: meta["name"], CWD: meta["cwd"],
+		Leaf: s.leafLocked(entries), Updated: updated,
+	}, nil
+}
@@ -3,6 +3,7 @@ package session
 import (
 	"bufio"
 	"encoding/json"
+	"fmt"
 	"os"

 	"agent-native/runtime/llm"
@@ -66,73 +67,88 @@ func (s *Store) ReadMeta() map[string]string {
 	return meta
 }

-// Compact 把转录改写为「系统消息 + 摘要消息 + 最近 keepLastN 条」。
-// 摘要由 summarize 生成（由适配层注入，通常是一次无工具模型调用）；
-// 改写原子落盘（临时文件 + rename）。
-// 返回是否发生了压缩（消息数不足/已是紧凑形态时返回 false 且不动文件）。
+const maxCompactAttempts = 3
+
+// Compact appends a compaction entry. The file keeps every earlier record.
+// Load hides the summarized prefix behind the summary plus the last
+// keepLastN non-system messages. Returns false when there is nothing to fold.
+// A session change during summary is retried from the new transcript.
+// Three consecutive summary failures or conflicts trip the breaker and
+// leave the file unchanged.
 func (s *Store) Compact(keepLastN int, summarize func(messages []llm.Message) (string, error)) (bool, error) {
-	messages, err := s.Load()
+	var last error
+	for attempt := 0; attempt < maxCompactAttempts; attempt++ {
+		done, err, retry := s.compactOnce(keepLastN, summarize)
+		if !retry {
+			return done, err
+		}
+		last = err
+	}
+	if last == nil {
+		last = fmt.Errorf("compaction conflict: session changed during summary")
+	}
+	return false, fmt.Errorf("compaction circuit open: %w", last)
+}
+
+func (s *Store) compactOnce(keepLastN int, summarize func(messages []llm.Message) (string, error)) (bool, error, bool) {
+	s.mu.Lock()
+	entries, err := s.readEntriesLocked()
 	if err != nil {
-		return false, err
+		s.mu.Unlock()
+		return false, err, false
 	}
-	if len(messages) <= keepLastN+1 { // 系统消息 + 尾部，无可压缩头部
-		return false, nil
+	leaf := s.leafLocked(entries)
+	generation := len(entries)
+	s.mu.Unlock()
+	path := pathTo(entries, leaf)
+	var messages []llm.Message
+	var msgEntries []Entry
+	for _, e := range path {
+		if e.Type == TypeMessage && e.Message != nil {
+			messages = append(messages, *e.Message)
+			msgEntries = append(msgEntries, e)
+		}
+	}
+	if len(messages) <= keepLastN+1 {
+		return false, nil, false
 	}
-	var system *llm.Message
 	rest := messages
+	restEntries := msgEntries
 	if len(rest) > 0 && rest[0].Role == llm.RoleSystem {
-		copyOf := rest[0]
-		system = &copyOf
 		rest = rest[1:]
+		restEntries = restEntries[1:]
 	}
 	tailStart := len(rest) - keepLastN
 	if tailStart <= 0 {
-		return false, nil
+		return false, nil, false
 	}
+	tailStart = alignToolTail(rest, tailStart)
 	head := rest[:tailStart]
-	tail := rest[tailStart:]
 	summary, err := summarize(head)
 	if err != nil {
-		return false, err
+		return false, err, true
 	}
-	summaryMsg := llm.Message{
-		Role:    llm.RoleUser,
-		Content: "[conversation summary]\n" + summary + "\n[recent messages follow]",
+	var tailIDs []string
+	for _, e := range restEntries[tailStart:] {
+		tailIDs = append(tailIDs, e.ID)
 	}
-	var rebuilt []llm.Message
-	if system != nil {
-		rebuilt = append(rebuilt, *system)
-	}
-	rebuilt = append(rebuilt, summaryMsg)
-	rebuilt = append(rebuilt, tail...)
-
 	s.mu.Lock()
 	defer s.mu.Unlock()
-	tmp := s.path + ".compact.tmp"
-	f, err := os.OpenFile(tmp, os.O_CREATE|os.O_TRUNC|os.O_WRONLY, 0o644)
+	now, err := s.readEntriesLocked()
 	if err != nil {
-		return false, err
+		return false, err, false
 	}
-	w := bufio.NewWriter(f)
-	for _, m := range rebuilt {
-		raw, err := json.Marshal(m)
-		if err != nil {
-			_ = f.Close()
-			_ = os.Remove(tmp)
-			return false, err
-		}
-		w.Write(raw)
-		w.WriteByte('\n')
+	if len(now) != generation {
+		return false, fmt.Errorf("compaction conflict: session changed during summary"), true
 	}
-	if err := w.Flush(); err != nil {
-		_ = f.Close()
-		_ = os.Remove(tmp)
-		return false, err
+	_, err = s.appendEntryLocked(Entry{Type: TypeCompaction, Summary: summary, TailIDs: tailIDs})
+	return err == nil, err, false
+}
+
+// alignToolTail 让保留尾巴从一条非 tool 消息开始，避免把 tool_calls 和结果拆开。
+func alignToolTail(messages []llm.Message, tailStart int) int {
+	for tailStart > 0 && messages[tailStart].Role == llm.RoleTool {
+		tailStart--
 	}
-	_ = f.Close()
-	if err := os.Rename(tmp, s.path); err != nil {
-		_ = os.Remove(tmp)
-		return false, err
-	}
-	return true, nil
+	return tailStart
 }
@@ -114,9 +114,12 @@ func TestCompactRewritesTranscript(t *testing.T) {
 	if msgs[2].Content != "reply" || msgs[3].Content != "earlyj" || msgs[4].Content != "reply" {
 		t.Fatalf("tail order wrong: %+v", msgs[2:])
 	}
-	// 无 tmp 残留。
-	if _, err := os.Stat(s.Path() + ".compact.tmp"); !os.IsNotExist(err) {
-		t.Fatal("compact tmp file must not remain")
+	entries, err := s.Entries()
+	if err != nil {
+		t.Fatal(err)
+	}
+	if len(entries) <= 5 {
+		t.Fatalf("compaction must append, not rewrite; got %d entries", len(entries))
 	}
 }

@@ -0,0 +1,32 @@
+package session
+
+import (
+	"fmt"
+	"html"
+	"os"
+	"strings"
+)
+
+// ExportHTML writes the active branch as a standalone HTML document.
+func (s *Store) ExportHTML(path string) error {
+	msgs, err := s.Load()
+	if err != nil {
+		return err
+	}
+	var b strings.Builder
+	b.WriteString("<!DOCTYPE html><html><head><meta charset=\"utf-8\"><title>")
+	title := s.Name()
+	if title == "" {
+		title = "session"
+	}
+	b.WriteString(html.EscapeString(title))
+	b.WriteString("</title></head><body>\n")
+	for _, m := range msgs {
+		fmt.Fprintf(&b, "<article data-role=\"%s\"><h2>%s</h2><pre>%s</pre></article>\n",
+			html.EscapeString(string(m.Role)),
+			html.EscapeString(string(m.Role)),
+			html.EscapeString(m.Content))
+	}
+	b.WriteString("</body></html>\n")
+	return os.WriteFile(path, []byte(b.String()), 0o644)
+}
@@ -0,0 +1,57 @@
+package session
+
+import (
+	"encoding/json"
+
+	"agent-native/runtime/llm"
+)
+
+// ToolIntent 是一次尚未配对结果的工具调用声明。
+type ToolIntent struct {
+	CallID    string `json:"callId"`
+	Name      string `json:"name"`
+	Arguments string `json:"arguments,omitempty"`
+	Replay    string `json:"replay"`
+}
+
+// AppendToolIntent 先落下「即将执行」。结果另写一条 tool 消息。
+func (s *Store) AppendToolIntent(in ToolIntent) error {
+	raw, err := json.Marshal(in)
+	if err != nil {
+		return err
+	}
+	_, err = s.AppendEntry(Entry{Type: TypeToolIntent, Data: raw})
+	return err
+}
+
+// PendingIntents 返回活动分支上还没有对应 tool 结果的意图。
+func (s *Store) PendingIntents() ([]ToolIntent, error) {
+	s.mu.Lock()
+	defer s.mu.Unlock()
+	entries, err := s.readEntriesLocked()
+	if err != nil {
+		return nil, err
+	}
+	answered := map[string]bool{}
+	var intents []ToolIntent
+	for _, e := range pathTo(entries, s.leafLocked(entries)) {
+		if e.Type == TypeMessage && e.Message != nil && e.Message.Role == llm.RoleTool && e.Message.ToolCallID != "" {
+			answered[e.Message.ToolCallID] = true
+		}
+		if e.Type != TypeToolIntent {
+			continue
+		}
+		var in ToolIntent
+		if err := json.Unmarshal(e.Data, &in); err != nil {
+			return nil, err
+		}
+		intents = append(intents, in)
+	}
+	var pending []ToolIntent
+	for _, in := range intents {
+		if !answered[in.CallID] {
+			pending = append(pending, in)
+		}
+	}
+	return pending, nil
+}
@@ -1,17 +1,16 @@
 package session

 import (
-	"bufio"
-	"encoding/json"
 	"os"
 	"path/filepath"
-	"strings"
 	"sync"

 	"agent-native/runtime/llm"
 )

-// Store is an append-only JSONL transcript. Each line is one llm.Message.
+// Store is an append-only JSONL transcript. New sessions are a tree of
+// entries (id / parentId). Load still returns the model-visible message
+// path, and files written as one llm.Message per line remain readable.
 type Store struct {
 	path string
 	mu   sync.Mutex
@@ -31,56 +30,25 @@ func Open(path string) (*Store, error) {

 func (s *Store) Path() string { return s.path }

+// Append records one message on the active branch.
 func (s *Store) Append(msg llm.Message) error {
 	s.mu.Lock()
 	defer s.mu.Unlock()
-	raw, err := json.Marshal(msg)
-	if err != nil {
-		return err
-	}
-	f, err := os.OpenFile(s.path, os.O_APPEND|os.O_WRONLY, 0o644)
-	if err != nil {
-		return err
-	}
-	defer f.Close()
-	raw = append(raw, '\n')
-	_, err = f.Write(raw)
+	_, err := s.appendEntryLocked(Entry{Type: TypeMessage, Message: &msg})
 	return err
 }

-// Load 全量读取转录。崩溃残留容错：Append 恒以换行收尾，因此无换行结尾的
-// 尾部若不是完整 JSON 即为崩溃半行，丢弃而不是让整个会话不可用；
-// 行中间的坏行仍报错（那是文件损坏，不是崩溃残留）。
+// Load returns the model-visible messages on the active branch.
+// A compaction entry replaces earlier turns with its summary plus the
+// retained tail; the file itself is never rewritten.
+// Crash residue: a trailing line without a newline that is not valid JSON
+// is dropped. A bad line in the middle is still an error.
 func (s *Store) Load() ([]llm.Message, error) {
 	s.mu.Lock()
 	defer s.mu.Unlock()
-	raw, err := os.ReadFile(s.path)
+	entries, err := s.readEntriesLocked()
 	if err != nil {
 		return nil, err
 	}
-	text := string(raw)
-	if text != "" && !strings.HasSuffix(text, "\n") {
-		if idx := strings.LastIndexByte(text, '\n'); idx >= 0 {
-			if !json.Valid([]byte(text[idx+1:])) {
-				text = text[:idx+1]
-			}
-		} else if !json.Valid([]byte(text)) {
-			text = ""
-		}
-	}
-	var out []llm.Message
-	sc := bufio.NewScanner(strings.NewReader(text))
-	sc.Buffer(make([]byte, 0, 64*1024), 8*1024*1024)
-	for sc.Scan() {
-		line := sc.Bytes()
-		if len(line) == 0 {
-			continue
-		}
-		var msg llm.Message
-		if err := json.Unmarshal(line, &msg); err != nil {
-			return nil, err
-		}
-		out = append(out, msg)
-	}
-	return out, sc.Err()
+	return visibleMessages(entries, s.leafLocked(entries)), nil
 }
@@ -0,0 +1,70 @@
+package session
+
+import "agent-native/runtime/llm"
+
+const (
+	// DefaultKeepToolResults 是运行内清理时保留的最近工具结果条数。
+	DefaultKeepToolResults = 5
+	// DefaultContextWindow 是未配置模型窗口时的保守字符数。
+	DefaultContextWindow = 32000
+	// DefaultOutputReserve 是留给模型输出的字符预留。
+	DefaultOutputReserve = 4096
+	// DefaultContextBuffer 是窗口估算的余量，避免贴着上限才清理。
+	DefaultContextBuffer = 1024
+)
+
+// EstimateChars 用字符数近似这次请求的体积。
+func EstimateChars(msgs []llm.Message) int {
+	n := 0
+	for _, m := range msgs {
+		n += len([]rune(m.Content))
+		for _, c := range m.ToolCalls {
+			n += len([]rune(c.Name)) + len([]rune(c.Arguments))
+		}
+	}
+	return n
+}
+
+// NeedsClear 在估算超过「窗口 − 输出预留 − 缓冲」时返回真。
+// window ≤ 0 时用保守默认值。
+func NeedsClear(estimate, window, outputReserve, buffer int) bool {
+	if window <= 0 {
+		window = DefaultContextWindow
+	}
+	room := window - outputReserve - buffer
+	if room < 1 {
+		room = 1
+	}
+	return estimate > room
+}
+
+// ClearOldToolResults 把最近 keep 条之前的工具结果换成占位。
+// 消息条数、顺序和 tool_call 配对都不变；占位写明完整结果仍在会话记录里。
+// keep ≤ 0 时保留默认的 5 条。返回的切片不与输入共享被改写的消息。
+func ClearOldToolResults(msgs []llm.Message, keep int) []llm.Message {
+	if keep <= 0 {
+		keep = DefaultKeepToolResults
+	}
+	out := append([]llm.Message{}, msgs...)
+	var toolIdx []int
+	for i, m := range out {
+		if m.Role == llm.RoleTool {
+			toolIdx = append(toolIdx, i)
+		}
+	}
+	if len(toolIdx) <= keep {
+		return out
+	}
+	for _, i := range toolIdx[:len(toolIdx)-keep] {
+		name := out[i].Name
+		if name == "" {
+			name = "工具"
+		}
+		loc := "会话记录"
+		if out[i].ToolCallID != "" {
+			loc = "会话记录中的调用 " + out[i].ToolCallID
+		}
+		out[i].Content = "旧工具结果已清理。完整内容见" + loc + "（" + name + "）。"
+	}
+	return out
+}
@@ -0,0 +1,60 @@
+package session
+
+import (
+	"strings"
+	"testing"
+
+	"agent-native/runtime/llm"
+)
+
+func TestClearOldToolResultsKeepsLastFive(t *testing.T) {
+	var msgs []llm.Message
+	for i := 0; i < 8; i++ {
+		id := string(rune('a' + i))
+		msgs = append(msgs,
+			llm.Message{Role: llm.RoleAssistant, ToolCalls: []llm.ToolCall{{ID: id, Name: "read_file"}}},
+			llm.Message{Role: llm.RoleTool, Name: "read_file", ToolCallID: id, Content: "正文-" + id},
+		)
+	}
+	out := ClearOldToolResults(msgs, 5)
+	if len(out) != len(msgs) {
+		t.Fatalf("len %d, want %d", len(out), len(msgs))
+	}
+	cleared := 0
+	for i, m := range out {
+		if m.Role != llm.RoleTool {
+			if m.Role != msgs[i].Role || len(m.ToolCalls) != len(msgs[i].ToolCalls) {
+				t.Fatalf("pairing changed at %d", i)
+			}
+			continue
+		}
+		if m.ToolCallID != msgs[i].ToolCallID {
+			t.Fatalf("tool id changed at %d", i)
+		}
+		if strings.Contains(m.Content, "旧工具结果已清理") {
+			cleared++
+			if !strings.Contains(m.Content, m.ToolCallID) {
+				t.Fatalf("placeholder missing call id: %s", m.Content)
+			}
+		} else if m.Content != msgs[i].Content {
+			t.Fatalf("recent result changed: %s", m.Content)
+		}
+	}
+	if cleared != 3 {
+		t.Fatalf("cleared %d, want 3", cleared)
+	}
+	for _, m := range out[len(out)-10:] {
+		if m.Role == llm.RoleTool && strings.Contains(m.Content, "旧工具结果已清理") {
+			t.Fatalf("last five were cleared: %s", m.ToolCallID)
+		}
+	}
+}
+
+func TestNeedsClearUsesConservativeWindow(t *testing.T) {
+	if NeedsClear(10, 0, 0, 0) {
+		t.Fatal("small estimate must fit the default window")
+	}
+	if !NeedsClear(DefaultContextWindow, 0, DefaultOutputReserve, DefaultContextBuffer) {
+		t.Fatal("estimate at the default window must clear")
+	}
+}
--- a/Show More
+++ b/Show More