feat: 修复审计落差 #1/#2/#3/#6 + 管理面板移动端适配

#1 工具布尔标记生效(agent-106, 安全红线): loop toolSchemas/runToolCalls 加 allow-set 安全门 restrictSchemas/applyToolSecurityGate——群聊/陌生人只暴露允许工具(默认关), 被拒not_allowed软提示+AuditEvent, owner全工具不变; 身份靠 ActorEnv(非提示词)
#2 checkpoint 恢复(agent-109): run.go LoadLatest + resumableStopReason门控(只恢复in-flight工具循环,不出答案不续跑), 幂等不重放副作用, 新会话不误恢复; resume_test 4用例
#3 tool_result(agent-108, 协议层): Anthropic 原生tool_result块(tool_use_id/is_error)非user纯文本 + OpenAI tool消息补tool_call_id; Content加ToolCallID/IsError等; (loop接线层2处留待,见备注)
#6 capacity_awaken推送(agent-107): onCapacity从只Debug改推WS hub(BroadcastCapacityAwaken, sessionId定向+payload对齐前端AwakenPayload)
管理面板移动端(agent-110): AdminApp 汉堡菜单+抽屉侧栏+标题truncate防换行, RBAC守卫不受影响
全仓 build/vet/test绿 + 前端build绿; 前端无涉差异已对齐

备注: #3 loop接线层2处(assistant工具轮append历史 + Message.ToolCallID/IsError)留给后续; #4代审/#5子代理触发/#7拆文件后置(主人说后三不急)

internal/runtime/loop/capacity_l4_loop_test.go

@@ -86,12 +86,11 @@ func TestLoop_Capacity_L4_RestrictedEnv_NoOwnerHabitRead(t *testing.T) {
 		t.Fatal("no streaming request captured")
 	}
 	set := toolSet(ad.streamReqs[0].Tools)
-	// Restricted env → L4 silent → undecided turn stays quiet → only quiet tools, no
-	// web-search, no owner habit read.
-	if !hasAll(set, "terminal", "file") {
-		t.Fatalf("restricted tools = %v, want quiet set terminal+file", toolNames(ad.streamReqs[0].Tools))
-	}
-	if set["web-search"] {
-		t.Fatalf("restricted tools = %v, must NOT wake (owner habit not read)", toolNames(ad.streamReqs[0].Tools))
+	// Restricted env → L4 silent → the turn stays undecided, and the design/11 §3
+	// security gate hides every tool that is not marked AllowInGroup (terminal/file
+	// are unmarked). So the model sees nothing: web-search is absent (owner habit
+	// not read) and even the quiet-set tools are closed to a group session.
+	if len(set) != 0 {
+		t.Fatalf("restricted tools = %v, want none (security gate closes unmarked tools)", toolNames(ad.streamReqs[0].Tools))
 	}
 }

internal/runtime/loop/checkpoint.go

@@ -3,12 +3,26 @@ package loop
 import (
 	"context"
 	"encoding/json"
+	"errors"
 	"fmt"
 
+	"git.yeij.top/AskaEth/Cyrene_v2/internal/runtime/agent"
 	"git.yeij.top/AskaEth/Cyrene_v2/internal/runtime/model"
 	"git.yeij.top/AskaEth/Cyrene_v2/internal/runtime/persist"
 )
 
+// checkpointStore is the minimal surface the loop needs to persist a durable
+// snapshot and resume an interrupted run (design §1 durable execution).
+// *persist.Store satisfies it in production; tests substitute a fake to assert
+// the resume path without a real database.
+type checkpointStore interface {
+	// Save upserts a checkpoint.
+	Save(ctx context.Context, cp *persist.Checkpoint) error
+	// LoadLatest returns the most recently updated checkpoint for agentID, or
+	// ErrNotFound when the agent has none.
+	LoadLatest(ctx context.Context, agentID string) (*persist.Checkpoint, error)
+}
+
 // checkpointState is the durable snapshot persisted at every step. Carrying the
 // message window and the step index lets a long-running task be resumed from the
 // latest checkpoint without losing context (design §1 durable execution).
@@ -44,3 +58,59 @@ func (l *Loop) saveCheckpoint(ctx context.Context, sessionID string, step int, m
 		l.log.Warn("checkpoint save", "step", step, "error", err)
 	}
 }
+
+// loadCheckpoint resumes an interrupted run from the session's latest durable
+// checkpoint (design §1 durable execution). It returns the restored message
+// window, the next loop step to run (the checkpoint step + 1) and whether a
+// resumable checkpoint was found. When no checkpoint exists, or the latest one
+// reflects a completed/blocked run, it returns the caller's messages with step 1
+// so the loop starts fresh.
+//
+// Restoration is idempotent: the checkpoint captures the message window exactly as
+// it stood at the end of the last completed step (assistant tool calls plus their
+// results are already fused in), so resuming at the next step rebuilds the model
+// context from history and never re-executes a completed step's tool call nor
+// regenerates a finished answer. Loading is read-only — a resume never deletes the
+// checkpoint, so every step's snapshot stays queryable.
+func (l *Loop) loadCheckpoint(ctx context.Context, sessionID string, messages []model.Message) ([]model.Message, int, bool) {
+	if l.checkpoint == nil {
+		return messages, 1, false
+	}
+	cp, err := l.checkpoint.LoadLatest(ctx, sessionID)
+	if err != nil {
+		if !errors.Is(err, persist.ErrNotFound) {
+			l.log.Warn("checkpoint load", "session", sessionID, "error", err)
+		}
+		return messages, 1, false
+	}
+	st := checkpointState{}
+	if err := json.Unmarshal(cp.State, &st); err != nil {
+		l.log.Warn("checkpoint unmarshal", "session", sessionID, "step", cp.Step, "error", err)
+		return messages, 1, false
+	}
+	if !resumableStopReason(st.StopReason) {
+		l.log.Debug("checkpoint not resumable", "session", sessionID, "step", cp.Step, "stop_reason", st.StopReason)
+		return messages, 1, false
+	}
+	l.log.Info("checkpoint resumed", "session", sessionID, "step", cp.Step, "next_step", cp.Step+1, "messages", len(st.Messages))
+	return st.Messages, cp.Step + 1, true
+}
+
+// resumableStopReason reports whether a checkpoint whose last step ended with the
+// given stop reason represents an in-flight run worth resuming. Only a
+// mid-tool-loop step is genuinely "interrupted": the loop folded the tool results
+// into the window and was about to continue, so picking up at the next step
+// rebuilds context without duplicating work. A checkpoint carrying a final answer,
+// a budget exhaustion, a safety-net termination or a guardrail denial is a
+// completed/blocked run — resuming it would regenerate an already-produced answer
+// (a duplicate side effect) — so those fall through to a fresh start (a caller
+// opening a genuinely new conversation uses a distinct sessionID, which misses
+// LoadLatest anyway).
+func resumableStopReason(stopReason string) bool {
+	switch stopReason {
+	case "", agent.StopReasonToolCalls:
+		return true
+	default:
+		return false
+	}
+}

internal/runtime/loop/loop.go

@@ -24,7 +24,6 @@ import (
 	"git.yeij.top/AskaEth/Cyrene_v2/internal/runtime/guardrail"
 	"git.yeij.top/AskaEth/Cyrene_v2/internal/runtime/memory"
 	"git.yeij.top/AskaEth/Cyrene_v2/internal/runtime/model"
-	"git.yeij.top/AskaEth/Cyrene_v2/internal/runtime/persist"
 	"git.yeij.top/AskaEth/Cyrene_v2/internal/runtime/tool"
 	"git.yeij.top/AskaEth/Cyrene_v2/internal/runtime/trace"
 )
@@ -89,7 +88,7 @@ type Loop struct {
 	retriever       *memory.Retriever
 	working         *memory.WorkingMemory
 	extractor       memory.Extractor
-	checkpoint      *persist.Store
+	checkpoint      checkpointStore
 	tracer          *trace.Tracer
 	guards          *guardrail.Pipeline
 	personaID       string
@@ -100,6 +99,7 @@ type Loop struct {
 	toolParallelism int
 	hooks              ToolLoopHooks
 	onStream           func(model.StreamEvent)
+	onAudit            func(AuditEvent)
 	maxContextTokens   int
 	compactor          Compactor
 	transformContext   func(context.Context, []model.Message, int) []model.Message
@@ -150,6 +150,7 @@ func New(opts Options) *Loop {
 		toolParallelism:  toolParallelism,
 		hooks:            opts.Hooks,
 		onStream:         opts.OnStream,
+		onAudit:          opts.OnAudit,
 		maxContextTokens: opts.MaxContextTokens,
 		compactor:        compactor,
 		transformContext: opts.TransformContext,

internal/runtime/loop/options.go

@@ -9,7 +9,6 @@ import (
 	"git.yeij.top/AskaEth/Cyrene_v2/internal/runtime/guardrail"
 	"git.yeij.top/AskaEth/Cyrene_v2/internal/runtime/memory"
 	"git.yeij.top/AskaEth/Cyrene_v2/internal/runtime/model"
-	"git.yeij.top/AskaEth/Cyrene_v2/internal/runtime/persist"
 	"git.yeij.top/AskaEth/Cyrene_v2/internal/runtime/tool"
 	"git.yeij.top/AskaEth/Cyrene_v2/internal/runtime/trace"
 )
@@ -38,6 +37,12 @@ type Options struct {
 	// loop observes while deciding a round (design/14 §4), so a caller can
 	// surface the "能力点亮" progression to the frontend. Nil = no-op.
 	OnCapacity func(capacity.Event)
+	// OnAudit, when non-nil, receives an AuditEvent every time the loop denies a
+	// tool call in a restricted (group/stranger) environment (design/11 §3 §8),
+	// so the rejection lands in the audit trail and the frontend can render the
+	// small "xx 工具在群聊被拒" note. Nil = the rejection is still fed back to the
+	// model as a soft not_allowed result, just not recorded to an external sink.
+	OnAudit func(AuditEvent)
 	// Memory is the long-term memory store used to persist extracted items.
 	Memory *memory.Store
 	// Retriever injects relevant memories into the context, re-evaluated every
@@ -48,8 +53,11 @@ type Options struct {
 	WorkingMemory *memory.WorkingMemory
 	// Extractor derives memory items to persist after every step.
 	Extractor memory.Extractor
-	// Checkpoint persists a durable snapshot of the loop state at every step.
-	Checkpoint *persist.Store
+	// Checkpoint persists a durable snapshot of the loop state at every step and,
+	// on a later Run with the same sessionID, resumes an interrupted run from it
+	// (design §1 durable execution). *persist.Store satisfies the narrow
+	// checkpointStore interface; nil disables both persistence and resume.
+	Checkpoint checkpointStore
 	// Tracer emits a structured span for every step, LLM call and tool call.
 	Tracer *trace.Tracer
 	// Guards validates request/response payloads on the streaming path, where

internal/runtime/loop/request.go

@@ -58,6 +58,12 @@ func (l *Loop) temperature() *float64 {
 // (a view over the dispatcher, per design/14 §4): the model sees/invokes only
 // the allowed tools, while Invoke itself stays unblocked. A nil or disabled
 // decider keeps the pre-existing behavior (every tool advertised).
+//
+// The design/11 §3 security gate runs last over whatever the capacity view left:
+// in a group/stranger environment only tools whose editable marker allows that
+// environment are advertised (default closed), and the owner environment keeps
+// the full set. Security is a hard red line, not a convenience hint, so it can
+// only shrink the surface — it never re-admits a tool capacity excluded.
 func (l *Loop) toolSchemas(ctx context.Context, messages []model.Message) []model.ToolSpec {
 	d := l.dispatcher()
 	if d == nil {
@@ -65,7 +71,7 @@ func (l *Loop) toolSchemas(ctx context.Context, messages []model.Message) []mode
 	}
 	schemas := d.ToolSchemas()
 	if l.capacity == nil || !l.capacity.Enabled() {
-		return schemas
+		return l.restrictSchemas(ctx, d, schemas)
 	}
 
 	names := make([]string, 0, len(schemas))
@@ -81,7 +87,9 @@ func (l *Loop) toolSchemas(ctx context.Context, messages []model.Message) []mode
 	})
 	acc.state = dec.State
 	if l.onCapacity != nil && !isMutedBroadcast(ctx) {
+		sessionID := loopSession(ctx)
 		for _, ev := range dec.Events {
+			ev.SessionID = sessionID
 			l.onCapacity(ev)
 		}
 	}
@@ -96,7 +104,7 @@ func (l *Loop) toolSchemas(ctx context.Context, messages []model.Message) []mode
 			out = append(out, s)
 		}
 	}
-	return out
+	return l.restrictSchemas(ctx, d, out)
 }
 
 // capacityTexts extracts the recent conversation text for the capacity decider:

internal/runtime/loop/resume_test.go

@@ -0,0 +1,194 @@
+package loop
+
+import (
+	"context"
+	"encoding/json"
+	"testing"
+
+	"git.yeij.top/AskaEth/Cyrene_v2/internal/runtime/agent"
+	"git.yeij.top/AskaEth/Cyrene_v2/internal/runtime/model"
+	"git.yeij.top/AskaEth/Cyrene_v2/internal/runtime/persist"
+)
+
+// fakeCheckpointStore is a seam for asserting the resume path without a real
+// database. It records every LoadLatest call and returns a scripted checkpoint or
+// error, so a test can prove the loop asks the store for the session's checkpoint
+// and resumes (or stays fresh) based on what comes back.
+type fakeCheckpointStore struct {
+	latest    *persist.Checkpoint
+	latestErr error
+	loadCalls []string
+}
+
+func (f *fakeCheckpointStore) Save(context.Context, *persist.Checkpoint) error { return nil }
+
+func (f *fakeCheckpointStore) LoadLatest(_ context.Context, agentID string) (*persist.Checkpoint, error) {
+	f.loadCalls = append(f.loadCalls, agentID)
+	return f.latest, f.latestErr
+}
+
+// TestRun_ResumeFromCheckpoint_ContinuesNoRepeat simulates a long task that was
+// interrupted mid-tool-loop: the loop left a tool_calls checkpoint at step 2 with
+// the message window as it stood after folding two tool batches in, then the
+// process died before step 3 could run. A fresh Run on the same sessionID must
+// resume at step 3 — rebuilding the context from the checkpoint rather than
+// restarting — and must not re-execute the already-completed steps.
+func TestRun_ResumeFromCheckpoint_ContinuesNoRepeat(t *testing.T) {
+	t.Parallel()
+	ctx := context.Background()
+	ad := &mockAdapter{chat: []mgResp{{resp: model.Response{Content: "final"}}}}
+	a := newTestAgent(t, ad, nil)
+	cp := newTestCheckpointStore(t)
+	l := New(Options{Agent: a, Checkpoint: cp, MaxRounds: 8})
+
+	const sess = "sess-resume"
+	interrupted := []model.Message{
+		model.TextMessage(model.RoleUser, "do the task"),
+		model.TextMessage(model.RoleTool, "tool aa result: va"),
+		model.TextMessage(model.RoleTool, "tool bb result: vb"),
+	}
+	// Seed exactly what a crashed run leaves behind: a tool_calls checkpoint at
+	// step 2 with the window computed after step 2's tool batch.
+	l.saveCheckpoint(ctx, sess, 2, interrupted, agent.StopReasonToolCalls)
+
+	res, err := l.Run(ctx, sess, []model.Message{model.TextMessage(model.RoleUser, "do the task")})
+	if err != nil {
+		t.Fatalf("Run: %v", err)
+	}
+
+	// Resume continues at step 3 (the interrupted step), not step 1.
+	if res.Steps != 3 {
+		t.Fatalf("Steps = %d, want 3 (resumed at step 3)", res.Steps)
+	}
+	// No completed step's tool call is re-executed.
+	if res.ToolCallsExecuted != 0 {
+		t.Fatalf("ToolCallsExecuted = %d, want 0 (completed steps not re-run)", res.ToolCallsExecuted)
+	}
+	if res.Text != "final" {
+		t.Fatalf("Text = %q, want final", res.Text)
+	}
+
+	// The resumed run makes exactly one model call (step 3), fed the restored
+	// window — not a restart with a fresh conversation.
+	if len(ad.reqs) != 1 {
+		t.Fatalf("resumed run made %d model calls, want 1", len(ad.reqs))
+	}
+	got := ad.reqs[0].Messages
+	if len(got) != len(interrupted)+1 {
+		t.Fatalf("resumed request messages = %d, want %d (system + restored window)", len(got), len(interrupted)+1)
+	}
+	if got[0].Role != model.RoleSystem {
+		t.Fatalf("leading message role = %q, want system (instruction)", got[0].Role)
+	}
+	for i, m := range interrupted {
+		if got[i+1].Role != m.Role || messageText(got[i+1]) != messageText(m) {
+			t.Fatalf("restored message %d mismatch: got %+v, want %+v", i, got[i+1], m)
+		}
+	}
+}
+
+// TestRun_NewSession_NotResumed proves a different sessionID never resurrects an
+// old checkpoint: the caller opens a brand-new session, LoadLatest misses, and the
+// loop runs fresh from step 1 with the caller's own messages.
+func TestRun_NewSession_NotResumed(t *testing.T) {
+	t.Parallel()
+	ctx := context.Background()
+	call := `{"tool_calls":[{"name":"echo","args":{"q":"x"}}]}`
+	ad := &mockAdapter{chat: []mgResp{
+		{resp: model.Response{Content: call}},
+		{resp: model.Response{Content: "done"}},
+	}}
+	td := dispatcherWith(t, &mockTool{name: "echo", result: "hi"})
+	a := newTestAgent(t, ad, td)
+	cp := newTestCheckpointStore(t)
+	l := New(Options{Agent: a, Checkpoint: cp, MaxRounds: 8})
+
+	// sess-a holds an interrupted run's checkpoint.
+	l.saveCheckpoint(ctx, "sess-a", 2, []model.Message{
+		model.TextMessage(model.RoleUser, "a task"),
+		model.TextMessage(model.RoleTool, "tool echo result: hi"),
+	}, agent.StopReasonToolCalls)
+
+	res, err := l.Run(ctx, "sess-b", []model.Message{model.TextMessage(model.RoleUser, "go")})
+	if err != nil {
+		t.Fatalf("Run: %v", err)
+	}
+	if res.Steps != 2 {
+		t.Fatalf("Steps = %d, want 2 (fresh two-step run)", res.Steps)
+	}
+	if res.ToolCallsExecuted != 1 {
+		t.Fatalf("ToolCallsExecuted = %d, want 1", res.ToolCallsExecuted)
+	}
+	// The first model call must start from the caller's fresh message, never from
+	// sess-a's restored window.
+	first := ad.reqs[0].Messages
+	if len(first) != 2 {
+		t.Fatalf("fresh run first request messages = %d, want 2 (system + user)", len(first))
+	}
+	if messageText(first[1]) != "go" {
+		t.Fatalf("fresh run first user message = %q, want 'go'", messageText(first[1]))
+	}
+}
+
+// TestRun_Resume_LoadLatestCalled asserts Run asks the store for the session's
+// latest checkpoint and resumes from it (LoadLatest → tool_calls checkpoint →
+// continue at the next step).
+func TestRun_Resume_LoadLatestCalled(t *testing.T) {
+	t.Parallel()
+	ctx := context.Background()
+	ad := &mockAdapter{chat: []mgResp{{resp: model.Response{Content: "final"}}}}
+	a := newTestAgent(t, ad, nil)
+
+	buf, err := json.Marshal(checkpointState{
+		Step:       2,
+		Messages:   []model.Message{model.TextMessage(model.RoleUser, "keep going")},
+		StopReason: agent.StopReasonToolCalls,
+	})
+	if err != nil {
+		t.Fatalf("marshal checkpoint: %v", err)
+	}
+	fake := &fakeCheckpointStore{latest: &persist.Checkpoint{ID: "sess:2", AgentID: "sess", Step: 2, State: buf}}
+	l := New(Options{Agent: a, Checkpoint: fake, MaxRounds: 8})
+
+	res, err := l.Run(ctx, "sess", []model.Message{model.TextMessage(model.RoleUser, "new")})
+	if err != nil {
+		t.Fatalf("Run: %v", err)
+	}
+	if len(fake.loadCalls) != 1 || fake.loadCalls[0] != "sess" {
+		t.Fatalf("LoadLatest calls = %v, want exactly ['sess']", fake.loadCalls)
+	}
+	if res.Steps != 3 {
+		t.Fatalf("Steps = %d, want 3 (resumed at step 3)", res.Steps)
+	}
+	got := ad.reqs[0].Messages
+	if len(got) != 2 || messageText(got[1]) != "keep going" {
+		t.Fatalf("resumed request = %+v, want [system, 'keep going']", got)
+	}
+}
+
+// TestRun_Resume_NotFound_Fresh asserts LoadLatest is still called on the fresh
+// path and that ErrNotFound yields a full restart from step 1 with the caller's
+// messages (no checkpoint to resurrect).
+func TestRun_Resume_NotFound_Fresh(t *testing.T) {
+	t.Parallel()
+	ctx := context.Background()
+	ad := &mockAdapter{chat: []mgResp{{resp: model.Response{Content: "hello"}}}}
+	a := newTestAgent(t, ad, nil)
+	fake := &fakeCheckpointStore{latestErr: persist.ErrNotFound}
+	l := New(Options{Agent: a, Checkpoint: fake, MaxRounds: 8})
+
+	res, err := l.Run(ctx, "sess", []model.Message{model.TextMessage(model.RoleUser, "hi")})
+	if err != nil {
+		t.Fatalf("Run: %v", err)
+	}
+	if len(fake.loadCalls) != 1 || fake.loadCalls[0] != "sess" {
+		t.Fatalf("LoadLatest calls = %v, want exactly ['sess']", fake.loadCalls)
+	}
+	if res.Steps != 1 {
+		t.Fatalf("Steps = %d, want 1 (fresh run)", res.Steps)
+	}
+	got := ad.reqs[0].Messages
+	if len(got) != 2 || messageText(got[1]) != "hi" {
+		t.Fatalf("fresh request = %+v, want [system, 'hi']", got)
+	}
+}

internal/runtime/loop/run.go

@@ -35,12 +35,21 @@ func (l *Loop) Run(ctx context.Context, sessionID string, messages []model.Messa
 	// Seed the per-Run capacity accumulator at quiet so the state can progress
 	// (quiet→awake→full) across a multi-step tool loop; a nil/disabled decider
 	// leaves it unused. It is scoped to this Run, so concurrent runs never share.
+	// The session id rides the same context so a capacity_awaken emission can be
+	// targeted to the session that produced it (design/14 §5).
 	ctx = withCapState(ctx, &capStateAccum{state: capacity.StateQuiet})
+	ctx = withLoopSession(ctx, sessionID)
 
-	msgs := l.resolveImages(ctx, messages)
+	// Resume an interrupted run (design §1 durable execution): when the session
+	// has a resumable checkpoint, rebuild its context and continue at the next
+	// step rather than restarting from a fresh conversation. A different
+	// sessionID — or a checkpoint left by a completed run — falls through to a
+	// fresh start (LoadLatest misses, or the stop reason is not resumable).
+	msgs, startStep, _ := l.loadCheckpoint(ctx, sessionID, messages)
+	msgs = l.resolveImages(ctx, msgs)
 	toolState := &toolLoopState{prevFP: make(map[string]bool)}
 	res := &Result{}
-	for step := 1; step <= l.maxRounds; step++ {
+	for step := startStep; step <= l.maxRounds; step++ {
 		if err := ctx.Err(); err != nil {
 			return nil, fmt.Errorf("loop: run: step %d: %w", step, err)
 		}

internal/runtime/loop/session_ctx.go

@@ -0,0 +1,21 @@
+package loop
+
+import "context"
+
+// loopSessionIDKey is the context key for the session id the current Run is
+// driving. It is seeded at the top of Run and read where the loop emits a
+// session-scoped UI event (capacity_awaken), so the event carries the right
+// session rather than being treated as global.
+type loopSessionIDKey struct{}
+
+// withLoopSession returns a context carrying the Run's session id.
+func withLoopSession(ctx context.Context, sessionID string) context.Context {
+	return context.WithValue(ctx, loopSessionIDKey{}, sessionID)
+}
+
+// loopSession returns the session id on ctx, or "" when none was seeded (a
+// defensive fallback for a caller that drives the loop outside Run).
+func loopSession(ctx context.Context) string {
+	s, _ := ctx.Value(loopSessionIDKey{}).(string)
+	return s
+}

internal/runtime/loop/tool_policy.go

@@ -0,0 +1,86 @@
+package loop
+
+import (
+	"context"
+
+	"git.yeij.top/AskaEth/Cyrene_v2/internal/runtime/agent"
+	"git.yeij.top/AskaEth/Cyrene_v2/internal/runtime/model"
+	"git.yeij.top/AskaEth/Cyrene_v2/internal/runtime/tool"
+)
+
+// auditToolRejectNotAllowed is the audit "what" for a tool call denied by the
+// design/11 §3 security gate in a group/stranger environment.
+const auditToolRejectNotAllowed = "tool.reject.not_allowed"
+
+// AuditEvent reports a security-gated tool rejection to the audit sink
+// (design/11 §3 §8). The loop emits one when it denies a tool call in a
+// restricted (group/stranger) environment, so the rejection is auditable and the
+// frontend can render the small "xx 工具在群聊被拒" note. It is wired to the
+// gateway audit trail via Options.OnAudit.
+type AuditEvent struct {
+	// Who is the acting actor environment label ("group_member"/"stranger").
+	Who string
+	// What is the operation, e.g. "tool.reject.not_allowed".
+	What string
+	// Result is the outcome detail — the not_allowed hint naming the denied tool.
+	Result string
+}
+
+// restrictSchemas returns only the tool schemas the session env on ctx may
+// invoke, per the design/11 §3 security gate: a group/stranger session may
+// invoke only tools whose editable marker allows that environment (default
+// closed), while the owner environment keeps the full set. It is the final word
+// over the capacity allowed-set view — a security red line, not a convenience
+// hint — so it runs last and can only shrink the surface.
+func (l *Loop) restrictSchemas(ctx context.Context, d *tool.Dispatcher, schemas []model.ToolSpec) []model.ToolSpec {
+	policy := d.PolicyEngine(sessionEnvFrom(ctx).ActorEnv)
+	kept := make([]model.ToolSpec, 0, len(schemas))
+	for _, s := range schemas {
+		if policy.Allow(s.Name) {
+			kept = append(kept, s)
+		}
+	}
+	return kept
+}
+
+// applyToolSecurityGate blocks every call whose tool is not allowed in the
+// session env on ctx (design/11 §3). It mutates actions in place, overriding
+// each runnable call to a skip carrying the not_allowed hint and emitting an
+// AuditEvent for the rejection. A nil receiver or nil dispatcher is a no-op.
+func (l *Loop) applyToolSecurityGate(ctx context.Context, d *tool.Dispatcher, calls []agent.ToolCall, actions []toolCallAction) {
+	env := sessionEnvFrom(ctx)
+	policy := d.PolicyEngine(env.ActorEnv)
+	for i := range calls {
+		if !actions[i].run {
+			continue
+		}
+		if policy.Allow(calls[i].Name) {
+			continue
+		}
+		reason := notAllowedReason(calls[i].Name)
+		actions[i] = toolCallAction{run: false, reason: reason}
+		l.emitAudit(AuditEvent{
+			Who:    string(env.ActorEnv),
+			What:   auditToolRejectNotAllowed,
+			Result: reason,
+		})
+	}
+}
+
+// notAllowedReason is the soft hint fed back to the model when a restricted
+// (group/stranger) session invokes a tool whose editable marker denies that
+// environment (design/11 §3). It is a soft result — injected into the
+// conversation, never a panic or hard crash — so the model can reassess.
+func notAllowedReason(name string) string {
+	return "not_allowed: tool " + name + " is not allowed in a group/stranger conversation"
+}
+
+// emitAudit forwards an AuditEvent to the configured OnAudit sink. It is a no-op
+// when no sink is wired, so the security gate still works (soft not_allowed
+// result) in leaf tests that do not attach an audit sink.
+func (l *Loop) emitAudit(ev AuditEvent) {
+	if l.onAudit == nil {
+		return
+	}
+	l.onAudit(ev)
+}

internal/runtime/loop/tool_policy_gating_test.go

@@ -0,0 +1,244 @@
+package loop
+
+import (
+	"context"
+	"strings"
+	"sync"
+	"testing"
+
+	"git.yeij.top/AskaEth/Cyrene_v2/internal/capability/registry"
+	"git.yeij.top/AskaEth/Cyrene_v2/internal/capability/restrict"
+	"git.yeij.top/AskaEth/Cyrene_v2/internal/runtime/model"
+	"git.yeij.top/AskaEth/Cyrene_v2/internal/runtime/tool"
+)
+
+// newPolicyDispatcher builds a fresh registry + dispatcher and registers one
+// mock tool per name, returning both so a test can set the group/stranger
+// markers on the registry before exercising the design/11 §3 security gate.
+func newPolicyDispatcher(t *testing.T, names []string) (*registry.Registry, *tool.Dispatcher) {
+	t.Helper()
+	reg := registry.New()
+	d, err := tool.New(reg)
+	if err != nil {
+		t.Fatalf("tool.New: %v", err)
+	}
+	for _, name := range names {
+		if err := d.Register(&mockTool{name: name}); err != nil {
+			t.Fatalf("Register %q: %v", name, err)
+		}
+	}
+	return reg, d
+}
+
+// expectToolNames asserts got carries exactly the tool names in want, any order.
+// The registry List iterates a map, so registration order is not preserved;
+// a set comparison keeps these assertions order-independent.
+func expectToolNames(t *testing.T, got []model.ToolSpec, want ...string) {
+	t.Helper()
+	if len(got) != len(want) {
+		t.Fatalf("tool schema count = %d, want %d (%v)", len(got), len(want), want)
+	}
+	wantSet := make(map[string]bool, len(want))
+	for _, n := range want {
+		wantSet[n] = true
+	}
+	for _, s := range got {
+		if !wantSet[s.Name] {
+			t.Fatalf("unexpected tool %q in schemas %v", s.Name, toolNames(got))
+		}
+	}
+}
+
+// TestToolSchemas_RestrictedEnv_FiltersByMarkers proves a group/stranger session
+// is only shown the tools whose editable marker allows that environment, while
+// the owner session keeps the full set (design/11 §3).
+func TestToolSchemas_RestrictedEnv_FiltersByMarkers(t *testing.T) {
+	t.Parallel()
+	reg, d := newPolicyDispatcher(t, []string{"groupOnly", "strangerOnly", "both", "ownerOnly"})
+	mustSetPolicy(t, reg, "groupOnly", registry.ToolPolicy{AllowInGroup: true})
+	mustSetPolicy(t, reg, "strangerOnly", registry.ToolPolicy{AllowInStranger: true})
+	mustSetPolicy(t, reg, "both", registry.ToolPolicy{AllowInGroup: true, AllowInStranger: true})
+	// ownerOnly stays default-closed (zero marker).
+
+	l := New(Options{Tools: d})
+
+	group := l.toolSchemas(WithSessionEnv(context.Background(),
+		SessionEnv{ActorEnv: restrict.ActorGroupMember, MemWritable: false}), nil)
+	expectToolNames(t, group, "groupOnly", "both")
+
+	stranger := l.toolSchemas(WithSessionEnv(context.Background(),
+		SessionEnv{ActorEnv: restrict.ActorStranger, MemWritable: false}), nil)
+	expectToolNames(t, stranger, "strangerOnly", "both")
+
+	owner := l.toolSchemas(WithSessionEnv(context.Background(), OwnerEnv()), nil)
+	expectToolNames(t, owner, "groupOnly", "strangerOnly", "both", "ownerOnly")
+}
+
+// TestToolSchemas_DefaultClosed_UnmarkedToolHidden proves the defaulted-closed
+// marker: an unmarked tool is not advertised to a group/stranger session, so a
+// session cannot even see (let alone call) an owner-only capability.
+func TestToolSchemas_DefaultClosed_UnmarkedToolHidden(t *testing.T) {
+	t.Parallel()
+	reg, d := newPolicyDispatcher(t, []string{"secret", "shared"})
+	// Only shared is opened to groups; secret is default-closed.
+	mustSetPolicy(t, reg, "shared", registry.ToolPolicy{AllowInGroup: true})
+	l := New(Options{Tools: d})
+
+	ctx := WithSessionEnv(context.Background(),
+		SessionEnv{ActorEnv: restrict.ActorGroupMember, MemWritable: false})
+	expectToolNames(t, l.toolSchemas(ctx, nil), "shared")
+}
+
+// TestRun_Restricted_NotAllowed_BlockedAndAudited proves a group session calling
+// an owner-only tool never executes it: the call is fed back as a soft
+// not_allowed result (no panic/crash) and an AuditEvent is recorded for the
+// rejection.
+func TestRun_Restricted_NotAllowed_BlockedAndAudited(t *testing.T) {
+	t.Parallel()
+	call := `{"tool_calls":[{"name":"ownerOnly","args":{}}]}`
+	ad := &mockAdapter{chat: []mgResp{
+		{resp: model.Response{Content: call}},
+		{resp: model.Response{Content: "done"}},
+	}}
+	reg, d := newPolicyDispatcher(t, []string{"ownerOnly", "shared"})
+	mustSetPolicy(t, reg, "shared", registry.ToolPolicy{AllowInGroup: true})
+
+	a := newTestAgent(t, ad, d)
+	var mu sync.Mutex
+	var audited []AuditEvent
+	l := New(Options{Agent: a, MaxRounds: 8, OnAudit: func(ev AuditEvent) {
+		mu.Lock()
+		audited = append(audited, ev)
+		mu.Unlock()
+	}})
+
+	ctx := WithSessionEnv(context.Background(),
+		SessionEnv{ActorEnv: restrict.ActorGroupMember, MemWritable: false})
+	res, err := l.Run(ctx, "sess", []model.Message{model.TextMessage(model.RoleUser, "go")})
+	if err != nil {
+		t.Fatalf("Run: %v", err)
+	}
+	if res.Text != "done" {
+		t.Fatalf("Text = %q, want done", res.Text)
+	}
+	if res.ToolCallsExecuted != 0 {
+		t.Fatalf("ToolCallsExecuted = %d, want 0 (ownerOnly blocked in group)", res.ToolCallsExecuted)
+	}
+
+	last := ad.reqs[len(ad.reqs)-1].Messages
+	lastMsg := last[len(last)-1]
+	if !strings.Contains(lastMsg.Content[0].Text, "not_allowed") ||
+		!strings.Contains(lastMsg.Content[0].Text, "ownerOnly") {
+		t.Fatalf("expected not_allowed soft result, got %q", lastMsg.Content[0].Text)
+	}
+
+	mu.Lock()
+	defer mu.Unlock()
+	if len(audited) != 1 {
+		t.Fatalf("audit events = %d, want 1", len(audited))
+	}
+	if audited[0].What != auditToolRejectNotAllowed {
+		t.Fatalf("audit what = %q, want %q", audited[0].What, auditToolRejectNotAllowed)
+	}
+	if audited[0].Who != string(restrict.ActorGroupMember) {
+		t.Fatalf("audit who = %q, want %q", audited[0].Who, restrict.ActorGroupMember)
+	}
+	if !strings.Contains(audited[0].Result, "ownerOnly") {
+		t.Fatalf("audit result = %q, want to name ownerOnly", audited[0].Result)
+	}
+}
+
+// TestRun_Restricted_AllowedTool_Executes proves a group session may invoke a
+// tool whose AllowInGroup marker is set, and that it runs normally.
+func TestRun_Restricted_AllowedTool_Executes(t *testing.T) {
+	t.Parallel()
+	call := `{"tool_calls":[{"name":"shared","args":{"q":"hi"}}]}`
+	ad := &mockAdapter{chat: []mgResp{
+		{resp: model.Response{Content: call}},
+		{resp: model.Response{Content: "done"}},
+	}}
+	reg, d := newPolicyDispatcher(t, []string{"shared", "ownerOnly"})
+	mustSetPolicy(t, reg, "shared", registry.ToolPolicy{AllowInGroup: true})
+
+	a := newTestAgent(t, ad, d)
+	l := New(Options{Agent: a, MaxRounds: 8})
+
+	ctx := WithSessionEnv(context.Background(),
+		SessionEnv{ActorEnv: restrict.ActorGroupMember, MemWritable: false})
+	res, err := l.Run(ctx, "sess", []model.Message{model.TextMessage(model.RoleUser, "go")})
+	if err != nil {
+		t.Fatalf("Run: %v", err)
+	}
+	if res.ToolCallsExecuted != 1 {
+		t.Fatalf("ToolCallsExecuted = %d, want 1 (shared allowed in group)", res.ToolCallsExecuted)
+	}
+	if res.Text != "done" {
+		t.Fatalf("Text = %q, want done", res.Text)
+	}
+}
+
+// TestRun_Owner_AllToolsExecute proves the owner environment is unaffected by the
+// default-closed marker: the owner may invoke an owner-only (unmarked) tool.
+func TestRun_Owner_AllToolsExecute(t *testing.T) {
+	t.Parallel()
+	call := `{"tool_calls":[{"name":"ownerOnly","args":{}}]}`
+	ad := &mockAdapter{chat: []mgResp{
+		{resp: model.Response{Content: call}},
+		{resp: model.Response{Content: "done"}},
+	}}
+	_, d := newPolicyDispatcher(t, []string{"ownerOnly"})
+	// Default-closed marker (never opened): owner must still be able to call it.
+	a := newTestAgent(t, ad, d)
+	l := New(Options{Agent: a, MaxRounds: 8})
+
+	res, err := l.Run(WithSessionEnv(context.Background(), OwnerEnv()), "sess",
+		[]model.Message{model.TextMessage(model.RoleUser, "go")})
+	if err != nil {
+		t.Fatalf("Run: %v", err)
+	}
+	if res.ToolCallsExecuted != 1 {
+		t.Fatalf("ToolCallsExecuted = %d, want 1 (owner unaffected by default-closed)", res.ToolCallsExecuted)
+	}
+	if res.Text != "done" {
+		t.Fatalf("Text = %q, want done", res.Text)
+	}
+}
+
+// TestRun_Stranger_NotAllowed_Blocked proves the gate also applies to the
+// stranger private-chat environment, and that a group-only tool is denied there.
+func TestRun_Stranger_NotAllowed_Blocked(t *testing.T) {
+	t.Parallel()
+	call := `{"tool_calls":[{"name":"groupOnly","args":{}}]}`
+	ad := &mockAdapter{chat: []mgResp{
+		{resp: model.Response{Content: call}},
+		{resp: model.Response{Content: "done"}},
+	}}
+	reg, d := newPolicyDispatcher(t, []string{"groupOnly", "strangerOk"})
+	mustSetPolicy(t, reg, "groupOnly", registry.ToolPolicy{AllowInGroup: true})
+	mustSetPolicy(t, reg, "strangerOk", registry.ToolPolicy{AllowInStranger: true})
+
+	a := newTestAgent(t, ad, d)
+	l := New(Options{Agent: a, MaxRounds: 8})
+
+	ctx := WithSessionEnv(context.Background(),
+		SessionEnv{ActorEnv: restrict.ActorStranger, MemWritable: false})
+	res, err := l.Run(ctx, "sess", []model.Message{model.TextMessage(model.RoleUser, "go")})
+	if err != nil {
+		t.Fatalf("Run: %v", err)
+	}
+	if res.ToolCallsExecuted != 0 {
+		t.Fatalf("ToolCallsExecuted = %d, want 0 (groupOnly denied to stranger)", res.ToolCallsExecuted)
+	}
+	last := ad.reqs[len(ad.reqs)-1].Messages
+	if !strings.Contains(last[len(last)-1].Content[0].Text, "not_allowed") {
+		t.Fatalf("expected not_allowed soft result, got %q", last[len(last)-1].Content[0].Text)
+	}
+}
+
+// mustSetPolicy applies a ToolPolicy to the named tool, failing the test on error.
+func mustSetPolicy(t *testing.T, reg *registry.Registry, name string, pol registry.ToolPolicy) {
+	t.Helper()
+	if err := reg.SetPolicy(name, pol); err != nil {
+		t.Fatalf("SetPolicy %q: %v", name, err)
+	}
+}

internal/runtime/loop/tools.go

@@ -114,6 +114,15 @@ func (l *Loop) runToolCalls(ctx context.Context, messages []model.Message, calls
 		res.terminate = true
 	}
 
+	// Security gate (design/11 §3): a group/stranger session may invoke only
+	// tools whose editable marker allows that environment (default closed). A
+	// disallowed call never executes; it is fed back as a not_allowed soft result
+	// and recorded to the audit sink. It is the final authority over the batch —
+	// a later audit hook may tighten but never re-admit a denied tool.
+	if d != nil {
+		l.applyToolSecurityGate(ctx, d, calls, actions)
+	}
+
 	// BatchAdmit hook (audit / political review): block the whole batch.
 	if l.hooks.BatchAdmit != nil {
 		if ok, reason := l.hooks.BatchAdmit(ctx, calls); !ok {