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)

cmd/cyrene/wire.go

@@ -101,6 +101,16 @@ type app struct {
 	gw            *gateway.Server
 	channelBridge *channelbridge.Bridge
 	channels      gateway.ChannelConfigStore
+	// capacitySink broadcasts capacity_awaken transitions to the Web UI (design/14
+	// §5). It is satisfied by *gateway.Server (BroadcastCapacityAwaken); held as an
+	// interface so the composition root is unit-testable with a fake sink.
+	capacitySink capacityEmitter
+}
+
+// capacityEmitter is the gateway-facing sink for capacity_awaken broadcasts.
+// *gateway.Server implements it; tests substitute a fake to capture the args.
+type capacityEmitter interface {
+	BroadcastCapacityAwaken(sessionID, fromState, toState, tool string)
 }
 
 // newApp assembles the full Cyrene runtime from cfg and log. It opens the
@@ -330,6 +340,7 @@ func (a *app) wireRuntime() error {
 		Tools:         a.dispatcher,
 		Capacity:      capDecider,
 		OnCapacity:    a.onCapacity,
+		OnAudit:       a.onToolAudit,
 		Memory:        store,
 		Retriever:     a.retriever,
 		Extractor:     memory.NewManualExtractor(personaID, "owner"),
@@ -379,12 +390,33 @@ func (a *app) buildCapacity() *capacity.Decider {
 }
 
 // onCapacity is the loop's OnCapacity sink: it observes a capacity_awaken
-// transition for observability. A future frontend wire (design/14 §5) will render
-// the "能力点亮" progression from these events.
+// transition and pushes it to the Web UI (design/14 §5) as a `capacity_awaken`
+// frame. The loop only emits on a real state transition (never every round), and
+// the event carries the originating session id so it routes to that session, not
+// globally. The payload is the frontend AwakenPayload {fromState,toState,tool}.
 func (a *app) onCapacity(ev capacity.Event) {
-	a.log.Debug("capacity event",
+	logger.WithSource(a.log, "cmd/cyrene").Debug("capacity event",
 		"type", ev.Type, "from", ev.FromState, "to", ev.ToState,
-		"tool", ev.Tool, "at", ev.At)
+		"tool", ev.Tool, "at", ev.At, "session", ev.SessionID)
+	if a.capacitySink == nil {
+		return
+	}
+	a.capacitySink.BroadcastCapacityAwaken(ev.SessionID, string(ev.FromState), string(ev.ToState), ev.Tool)
+}
+
+// onToolAudit is the loop's OnAudit sink: it forwards a tool call denied by the
+// design/11 §3 security gate in a group/stranger environment into the gateway
+// audit trail, so the rejection is "audit 可查" and the frontend can render the
+// small "xx 工具在群聊被拒" note (design/11 §8). The gateway is built after the
+// loop (wireFrontend), so the sink reads a.gw at call time — by then the app is
+// fully assembled and any Run is served through the gateway.
+func (a *app) onToolAudit(ev loop.AuditEvent) {
+	if a.gw == nil {
+		logger.WithSource(a.log, "cmd/cyrene").Warn("loop audit dropped: gateway not ready",
+			"who", ev.Who, "what", ev.What, "result", ev.Result)
+		return
+	}
+	a.gw.AuditEvent(ev.Who, ev.What, ev.Result)
 }
 
 // parseCapacityToolSets decodes the CAPACITY_TOOL_SETS JSON map of
@@ -606,6 +638,7 @@ func (a *app) wireFrontend() error {
 		})
 	}
 	a.gw = gw
+	a.capacitySink = gw
 	return nil
 }
 

cmd/cyrene/wire_test.go

@@ -0,0 +1,63 @@
+package main
+
+import (
+	"testing"
+
+	"git.yeij.top/AskaEth/Cyrene_v2/internal/runtime/capacity"
+)
+
+// fakeCapacityEmitter captures the capacity_awaken broadcast args so the
+// composition-root test can assert the wire forwards the transition to the sink.
+type fakeCapacityEmitter struct {
+	callCount int
+	sessionID string
+	fromState string
+	toState   string
+	tool      string
+}
+
+func (f *fakeCapacityEmitter) BroadcastCapacityAwaken(sessionID, fromState, toState, tool string) {
+	f.callCount++
+	f.sessionID = sessionID
+	f.fromState = fromState
+	f.toState = toState
+	f.tool = tool
+}
+
+// TestOnCapacityBroadcastsAwaken exercises the audit-gap fix (#6): a capacity
+// state transition received by the loop's OnCapacity sink is forwarded to the
+// gateway as a capacity_awaken broadcast carrying the originating session and the
+// frontend AwakenPayload fields (fromState, toState, tool).
+func TestOnCapacityBroadcastsAwaken(t *testing.T) {
+	fake := &fakeCapacityEmitter{}
+	a := &app{log: testLog(), capacitySink: fake}
+
+	a.onCapacity(capacity.Event{
+		Type:      capacity.EventAwaken,
+		FromState: capacity.StateQuiet,
+		ToState:   capacity.StateAwake,
+		Tool:      "web_search",
+		SessionID: "web",
+	})
+
+	if fake.callCount != 1 {
+		t.Fatalf("broadcast call count = %d, want 1", fake.callCount)
+	}
+	if fake.sessionID != "web" || fake.fromState != "quiet" || fake.toState != "awake" || fake.tool != "web_search" {
+		t.Fatalf("broadcast args = {sessionID:%q from:%q to:%q tool:%q}, want {web,quiet,awake,web_search}",
+			fake.sessionID, fake.fromState, fake.toState, fake.tool)
+	}
+}
+
+// TestOnCapacityNilSinkNoPanic locks the guard: a boot/degraded path where the
+// gateway is not yet wired must not panic — the sink is optional.
+func TestOnCapacityNilSinkNoPanic(t *testing.T) {
+	a := &app{log: testLog()}
+	a.onCapacity(capacity.Event{
+		Type:      capacity.EventAwaken,
+		FromState: capacity.StateQuiet,
+		ToState:   capacity.StateFull,
+		Tool:      "code",
+		SessionID: "web",
+	})
+}

frontend/web/src/admin/AdminApp.tsx

@@ -42,16 +42,22 @@ function renderSection(section: AdminSectionId): React.ReactNode {
   }
 }
 
-function AdminSidebar({ active }: { active: AdminSectionId }): React.ReactNode {
+/**
+ * Sidebar content shared by the desktop rail and the mobile drawer (design/10).
+ * Both render ADMIN_GROUPS with the same `onNavigate` hook — the drawer passes
+ * its close callback (tap an item → close), the desktop rail passes a no-op — so
+ * desktop and mobile never diverge.
+ */
+function adminNavContent({ active, onNavigate }: { active: AdminSectionId; onNavigate: () => void }): React.ReactNode {
   return (
-    <nav className="glass hidden h-full w-[var(--mu-sidebar-w)] shrink-0 flex-col gap-1 overflow-y-auto border-r border-[var(--mu-glass-border)] p-3 md:flex">
+    <>
       {ADMIN_GROUPS.map((group) => (
         <div key={group} className="mb-1">
           <div className="px-2 py-2 text-xs font-semibold uppercase tracking-wider text-[var(--mu-text-lo)]">{group}</div>
           {ADMIN_SECTIONS.filter((s) => s.group === group).map((s) => (
             <button
               key={s.id}
-              onClick={() => navigateAdmin(s.id)}
+              onClick={() => { navigateAdmin(s.id); onNavigate(); }}
               className={`flex w-full items-center gap-2.5 rounded-[var(--mu-radius)] px-3 py-2 text-sm text-left transition-all duration-200 ease-[var(--mu-spring)] active:scale-[0.98] ${
                 s.id === active
                   ? "bg-[var(--mu-accent)]/15 font-medium text-[var(--mu-accent-hi)]"
@@ -64,11 +70,50 @@ function AdminSidebar({ active }: { active: AdminSectionId }): React.ReactNode {
           ))}
         </div>
       ))}
+    </>
+  );
+}
+
+/** Desktop left rail — fixed, `hidden md:flex`. */
+function AdminSidebar({ active }: { active: AdminSectionId }): React.ReactNode {
+  return (
+    <nav className="glass hidden h-full w-[var(--mu-sidebar-w)] shrink-0 flex-col gap-1 overflow-y-auto border-r border-[var(--mu-glass-border)] p-3 md:flex">
+      {adminNavContent({ active, onNavigate: () => {} })}
     </nav>
   );
 }
 
-function AdminTopBar({ section }: { section: AdminSectionId }): React.ReactNode {
+/** Mobile drawer — scrim + spring slide-in panel (reuses mu-drawer-in). */
+function AdminMobileNav({ open, active, onClose }: { open: boolean; active: AdminSectionId; onClose: () => void }): React.ReactNode {
+  if (!open) return null;
+  return (
+    <div className="fixed inset-0 z-40 md:hidden" role="dialog" aria-modal="true">
+      <div
+        className="absolute inset-0 bg-black/50 backdrop-blur-[2px]"
+        onClick={onClose}
+        aria-label="关闭侧边栏"
+      />
+      <nav
+        className="glass absolute inset-y-0 left-0 flex w-[var(--mu-sidebar-w)] max-w-[80vw] flex-col gap-1 overflow-y-auto border-r border-[var(--mu-glass-border)] p-3"
+        style={{ animation: "mu-drawer-in 0.3s var(--mu-spring) both" }}
+      >
+        <div className="flex items-center justify-between px-2 pb-1">
+          <span className="text-xs font-medium text-[var(--mu-text-lo)]">菜单</span>
+          <button
+            onClick={onClose}
+            aria-label="关闭"
+            className="inline-flex h-8 w-8 items-center justify-center rounded-[var(--mu-radius)] text-[var(--mu-text-mid)] transition-all duration-200 hover:bg-[var(--mu-surface-4)] active:scale-95"
+          >
+            <Icon name="close" size={18} />
+          </button>
+        </div>
+        {adminNavContent({ active, onNavigate: onClose })}
+      </nav>
+    </div>
+  );
+}
+
+function AdminTopBar({ section, onMenu }: { section: AdminSectionId; onMenu?: () => void }): React.ReactNode {
   const role = getStoredRole();
   const [theme, setTheme] = useState<ThemeMode>(readTheme);
   const meta = ADMIN_SECTIONS.find((s) => s.id === section);
@@ -79,24 +124,38 @@ function AdminTopBar({ section }: { section: AdminSectionId }): React.ReactNode
   };
 
   return (
-    <header className="glass sticky top-0 z-30 flex h-[var(--mu-topbar-h)] items-center gap-3 border-b border-[var(--mu-glass-border)] px-4">
-      <div className="flex items-center gap-2">
-        <img src="/icon.svg" alt="Cyrene" className="h-7 w-7" />
-        <div className="leading-tight">
-          <div className="text-sm font-semibold">Cyrene · 管理面板</div>
-          <div className="text-[11px] text-[var(--mu-text-lo)]">{meta?.label ?? "仪表盘"}</div>
+    <header className="glass sticky top-0 z-30 flex h-[var(--mu-topbar-h)] items-center gap-2 border-b border-[var(--mu-glass-border)] px-3 sm:px-4">
+      {onMenu ? (
+        <button
+          onClick={onMenu}
+          aria-label="打开菜单"
+          title="菜单"
+          className="inline-flex h-8 w-8 shrink-0 items-center justify-center rounded-[var(--mu-radius)] text-[var(--mu-text-mid)] transition-all duration-200 ease-[var(--mu-spring)] hover:bg-[var(--mu-surface-4)] active:scale-95 md:hidden"
+        >
+          <Icon name="menu" size={20} />
+        </button>
+      ) : null}
+      <div className="flex min-w-0 items-center gap-2">
+        <img src="/icon.svg" alt="Cyrene" className="h-7 w-7 shrink-0" />
+        <div className="min-w-0 leading-tight">
+          <div className="truncate whitespace-nowrap text-sm font-semibold">Cyrene · 管理面板</div>
+          <div className="truncate whitespace-nowrap text-[11px] text-[var(--mu-text-lo)]">{meta?.label ?? "仪表盘"}</div>
         </div>
       </div>
       <div className="flex-1" />
-      {role ? <Tag tone={role === "owner" ? "accent" : role === "admin" ? "ok" : "neutral"}>{role}</Tag> : null}
+      {role ? (
+        <Tag tone={role === "owner" ? "accent" : role === "admin" ? "ok" : "neutral"} className="hidden shrink-0 sm:inline-flex">
+          {role}
+        </Tag>
+      ) : null}
       <button
         onClick={toggleTheme}
         aria-label="切换主题"
-        className="inline-flex h-8 w-8 items-center justify-center rounded-[var(--mu-radius)] bg-[var(--mu-surface-4)] text-sm transition-all duration-200 ease-[var(--mu-spring)] hover:bg-[var(--mu-surface-5)] active:scale-95"
+        className="inline-flex h-8 w-8 shrink-0 items-center justify-center rounded-[var(--mu-radius)] bg-[var(--mu-surface-4)] text-sm transition-all duration-200 ease-[var(--mu-spring)] hover:bg-[var(--mu-surface-5)] active:scale-95"
       >
         <Icon name={theme === "dark" ? "sun" : "moon"} size={18} />
       </button>
-      <Button size="sm" variant="ghost" onClick={() => { window.location.hash = ""; }}>
+      <Button size="sm" variant="ghost" className="shrink-0" onClick={() => { window.location.hash = ""; }}>
         返回用户端
       </Button>
     </header>
@@ -137,6 +196,10 @@ function Forbidden403(): React.ReactNode {
 }
 
 export function AdminApp({ section }: { section: AdminSectionId }): React.ReactNode {
+  // Mobile nav drawer — only surfaced on the admin shell (md+ keeps the fixed
+  // sidebar); the login/403 guards never render it but still hold the hook.
+  const [adminMenuOpen, setAdminMenuOpen] = useState(false);
+
   // RBAC guard (design/10 §2 + design/11 §7). A logged-in viewer or an
   // unauthenticated user never sees the admin module; the gateway would also
   // reject them with 401/403.
@@ -159,13 +222,14 @@ export function AdminApp({ section }: { section: AdminSectionId }): React.ReactN
 
   return (
     <div className="flex h-full flex-col">
-      <AdminTopBar section={section} />
+      <AdminTopBar section={section} onMenu={() => setAdminMenuOpen(true)} />
       <div className="flex min-h-0 flex-1">
         <AdminSidebar active={section} />
         <main className="relative flex-1 overflow-y-auto p-4 md:p-6">
           <div className="mx-auto max-w-6xl">{renderSection(section)}</div>
         </main>
       </div>
+      <AdminMobileNav open={adminMenuOpen} active={section} onClose={() => setAdminMenuOpen(false)} />
     </div>
   );
 }

internal/runtime/capacity/capacity.go

@@ -49,7 +49,13 @@ type Event struct {
 	// Tool is the tool that this transition lit up (点亮工具). Empty when no
 	// single tool can be named.
 	Tool string
-	At   time.Time
+	// At is the wall-clock time the transition was observed.
+	At time.Time
+	// SessionID is set by the runtime layer (loop) to the conversation session the
+	// transition belongs to, so the frontend targets the correct session rather
+	// than treating the event as global. It is empty when the emitter never ran
+	// inside a session-scoped turn (e.g. a direct toolSchemas call in a unit test).
+	SessionID string
 }
 
 // IsAwaken reports whether the event is a capacity_awaken transition.

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 {

internal/runtime/model/anthropic.go

@@ -304,7 +304,7 @@ func buildAnthropicMessages(msgs []Message) (string, []anthropicMessage, error)
 			}
 			continue
 		}
-		blocks, err := encodeAnthropicContent(m.Content)
+		blocks, err := encodeAnthropicMessage(m)
 		if err != nil {
 			return "", nil, err
 		}
@@ -316,11 +316,13 @@ func buildAnthropicMessages(msgs []Message) (string, []anthropicMessage, error)
 	return sys.String(), out, nil
 }
 
-// messageText concatenates the text parts of a message (used for system prompts).
+// messageText concatenates the text-bearing parts of a message (system prompts
+// and tool-result bodies). Tool-result parts carry their body in Text, so a
+// RoleTool turn's content serializes correctly for OpenAI's tool message.
 func messageText(m Message) string {
 	var b strings.Builder
 	for _, part := range m.Content {
-		if part.Type == ContentTypeText {
+		if part.Type == ContentTypeText || part.Type == ContentTypeToolResult {
 			b.WriteString(part.Text)
 		}
 	}
@@ -340,8 +342,43 @@ func anthropicRole(r Role) string {
 	}
 }
 
+// encodeAnthropicMessage turns a normalized message into Anthropic content
+// blocks. A RoleTool turn is ALWAYS rendered as native tool_result blocks (never
+// a user plain-text turn), because Anthropic requires a tool result to arrive as
+// a tool_result block carrying the tool_use_id and the is_error flag.
+func encodeAnthropicMessage(m Message) ([]any, error) {
+	if m.Role == RoleTool {
+		return encodeAnthropicToolResult(m)
+	}
+	return encodeAnthropicContent(m.Content)
+}
+
+// encodeAnthropicToolResult renders a RoleTool turn as tool_result blocks. Each
+// ContentTypeToolResult part becomes one block; legacy text parts are wrapped as
+// tool_result blocks too so the turn never degrades into a plain user message.
+// The tool_use_id falls back to the message-level ToolCallID when a part does not
+// carry its own.
+func encodeAnthropicToolResult(m Message) ([]any, error) {
+	blocks := make([]any, 0, len(m.Content))
+	for _, p := range m.Content {
+		switch p.Type {
+		case ContentTypeToolResult, ContentTypeText:
+			id := p.ToolCallID
+			if id == "" {
+				id = m.ToolCallID
+			}
+			blocks = append(blocks, AnthropicToolResult(id, p.Text, p.IsError))
+		}
+	}
+	if len(blocks) == 0 {
+		blocks = append(blocks, AnthropicToolResult(m.ToolCallID, "", false))
+	}
+	return blocks, nil
+}
+
 // encodeAnthropicContent turns normalized content parts into Anthropic content
-// blocks.
+// blocks. Text / image / tool_use (assistant) parts are handled here; tool_result
+// parts are rendered at the message level by encodeAnthropicToolResult.
 func encodeAnthropicContent(parts []Content) ([]any, error) {
 	blocks := make([]any, 0, len(parts))
 	for _, p := range parts {
@@ -361,6 +398,19 @@ func encodeAnthropicContent(parts []Content) ([]any, error) {
 					"data":       data,
 				},
 			})
+		case ContentTypeToolCall:
+			args := p.ToolArgs
+			if args == nil {
+				args = map[string]any{}
+			}
+			blocks = append(blocks, map[string]any{
+				"type":  "tool_use",
+				"id":    p.ToolCallID,
+				"name":  p.ToolCallName,
+				"input": args,
+			})
+		case ContentTypeToolResult:
+			blocks = append(blocks, AnthropicToolResult(p.ToolCallID, p.Text, p.IsError))
 		default:
 			return nil, fmt.Errorf("model: %w: unsupported content type %q", ErrBadRequest, p.Type)
 		}
@@ -421,7 +471,7 @@ func responseFromAnthropicWire(wire *anthropicResponse) Response {
 		case "text":
 			b.WriteString(blk.Text)
 		case "tool_use":
-			toolCalls = append(toolCalls, ToolCall{Name: blk.Name, Args: blk.Input})
+			toolCalls = append(toolCalls, ToolCall{Name: blk.Name, ID: blk.ID, Args: blk.Input})
 		}
 	}
 	out.Content = b.String()

internal/runtime/model/anthropic_stream.go

@@ -232,6 +232,6 @@ func finalizeAnthropicStream(agg *Response, blocks map[int]*streamBlock, order [
 				args = map[string]any{}
 			}
 		}
-		agg.ToolCalls = append(agg.ToolCalls, ToolCall{Name: b.name, Args: args})
+		agg.ToolCalls = append(agg.ToolCalls, ToolCall{Name: b.name, ID: b.id, Args: args})
 	}
 }

internal/runtime/model/anthropic_test.go

@@ -509,3 +509,103 @@ func TestAnthropicChat_ToolResult(t *testing.T) {
 		t.Errorf("is_error = %v, want true", blk["is_error"])
 	}
 }
+
+// anthropicBlockContent asserts a message renders a single Anthropic content block
+// and returns it as a map for inspection.
+func anthropicBlockContent(t *testing.T, msg anthropicMessage) (map[string]any, bool) {
+	t.Helper()
+	blocks, ok := msg.Content.([]any)
+	if !ok || len(blocks) != 1 {
+		t.Fatalf("content = %#v, want exactly 1 content block", msg.Content)
+	}
+	block, ok := blocks[0].(map[string]any)
+	return block, ok
+}
+
+func TestAnthropicChat_Build_ToolResultBlock(t *testing.T) {
+	t.Parallel()
+	msgs := []Message{
+		TextMessage(RoleUser, "what's the weather in Paris?"),
+		AssistantToolCallMessage([]ToolCall{{ID: "toolu_1", Name: "weather", Args: map[string]any{"city": "Paris"}}}),
+		ToolResultMessage("toolu_1", `{"temp":22}`, true),
+	}
+	sys, out, err := buildAnthropicMessages(msgs)
+	if err != nil {
+		t.Fatalf("buildAnthropicMessages error: %v", err)
+	}
+	if sys != "" {
+		t.Errorf("system = %q, want empty", sys)
+	}
+	if len(out) != 3 {
+		t.Fatalf("messages = %d, want 3", len(out))
+	}
+
+	// The assistant turn must carry the native tool_use block, not plain text.
+	if out[1].Role != "assistant" {
+		t.Errorf("assistant role = %q, want assistant", out[1].Role)
+	}
+	use, ok := anthropicBlockContent(t, out[1])
+	if !ok {
+		t.Fatal("assistant content is not a single block")
+	}
+	if use["type"] != "tool_use" || use["id"] != "toolu_1" || use["name"] != "weather" {
+		t.Errorf("tool_use block = %#v, want type=tool_use id=toolu_1 name=weather", use)
+	}
+	if inp, _ := use["input"].(map[string]any); inp["city"] != "Paris" {
+		t.Errorf("tool_use input = %#v, want {city:Paris}", use["input"])
+	}
+
+	// The tool-result turn must be a user turn with a native tool_result block
+	// (the audit's core fix: never a user plain-text turn).
+	if out[2].Role != "user" {
+		t.Errorf("tool-result role = %q, want user", out[2].Role)
+	}
+	tr, ok := anthropicBlockContent(t, out[2])
+	if !ok {
+		t.Fatal("tool-result content is not a single block")
+	}
+	if tr["type"] != "tool_result" {
+		t.Errorf("type = %v, want tool_result (not plain user text)", tr["type"])
+	}
+	if tr["tool_use_id"] != "toolu_1" {
+		t.Errorf("tool_use_id = %v, want toolu_1", tr["tool_use_id"])
+	}
+	if tr["content"] != `{"temp":22}` {
+		t.Errorf("content = %v, want the raw result", tr["content"])
+	}
+	if tr["is_error"] != true {
+		t.Errorf("is_error = %v, want true", tr["is_error"])
+	}
+}
+
+func TestAnthropicChat_Build_ToolResultLegacyText(t *testing.T) {
+	t.Parallel()
+	// The loop's current form: a RoleTool message carrying a plain text part and
+	// Message.ToolCallID. The protocol layer must still render a native tool_result
+	// block (not a user plain-text turn).
+	m := Message{Role: RoleTool, ToolCallID: "toolu_9",
+		Content: []Content{{Type: ContentTypeText, Text: "tool weather result: 22"}}}
+	_, out, err := buildAnthropicMessages([]Message{m})
+	if err != nil {
+		t.Fatalf("buildAnthropicMessages error: %v", err)
+	}
+	if len(out) != 1 {
+		t.Fatalf("messages = %d, want 1", len(out))
+	}
+	if out[0].Role != "user" {
+		t.Errorf("role = %q, want user", out[0].Role)
+	}
+	tr, ok := anthropicBlockContent(t, out[0])
+	if !ok {
+		t.Fatal("content is not a single block")
+	}
+	if tr["type"] != "tool_result" {
+		t.Errorf("type = %v, want tool_result (not user text)", tr["type"])
+	}
+	if tr["tool_use_id"] != "toolu_9" {
+		t.Errorf("tool_use_id = %v, want toolu_9", tr["tool_use_id"])
+	}
+	if tr["content"] != "tool weather result: 22" {
+		t.Errorf("content = %v, want the text", tr["content"])
+	}
+}

internal/runtime/model/model.go

@@ -36,17 +36,36 @@ const (
 	ContentTypeText ContentType = "text"
 	// ContentTypeImage is an image content part.
 	ContentTypeImage ContentType = "image_url"
+	// ContentTypeToolCall is an assistant-originated tool invocation part
+	// (Anthropic tool_use / OpenAI tool_calls), carried on a RoleAssistant turn.
+	ContentTypeToolCall ContentType = "tool_call"
+	// ContentTypeToolResult is a tool execution result part, carried on a RoleTool
+	// turn and encoded natively (Anthropic tool_result / OpenAI tool message).
+	ContentTypeToolResult ContentType = "tool_result"
 )
 
 // Content is a single, possibly multimodal, part of a message.
 type Content struct {
-	// Type distinguishes a text part from an image part.
+	// Type distinguishes a text part from an image / tool part.
 	Type ContentType
-	// Text is the text payload when Type == ContentTypeText.
+	// Text is the text payload when Type == ContentTypeText, or the result body
+	// when Type == ContentTypeToolResult.
 	Text string
 	// ImageURL is the image source when Type == ContentTypeImage. It may be a
 	// data: URI or an http(s):// URL.
 	ImageURL string
+	// ToolCallID correlates the part to a provider tool_use/tool_call id: the call
+	// id for ContentTypeToolCall, or the tool_use_id being answered for
+	// ContentTypeToolResult.
+	ToolCallID string
+	// ToolCallName is the invoked tool name when Type == ContentTypeToolCall.
+	ToolCallName string
+	// ToolArgs is the structured argument object when Type == ContentTypeToolCall.
+	// It is never a JSON-encoded string.
+	ToolArgs map[string]any
+	// IsError marks a ContentTypeToolResult part as an execution failure so the
+	// provider's tool_result.is_error / error flag is set natively.
+	IsError bool
 }
 
 // Message is a single conversation turn, normalized across providers.
@@ -62,8 +81,11 @@ type Message struct {
 	// channel layer. Empty means the sender is unlabeled (the owner, or an
 	// unknown/unresolved sender), and callers may leave it unset.
 	SenderName string
-	// ToolCallID and ToolCallName are reserved as interface-level placeholders
-	// for the M2 tool loop; they are ignored by current adapters.
+	// ToolCallID correlates a RoleTool turn to the provider tool_use/tool_call id
+	// it answers, and ToolCallName is the invoked tool name carried as a
+	// convenience for callers that render the result. Both are set by the tool
+	// loop when it feeds a result back, and consumed by the adapters to emit
+	// native tool_result / tool-message blocks.
 	ToolCallID   string
 	ToolCallName string
 }
@@ -84,6 +106,25 @@ func ImageMessage(role Role, text, imageURL string) Message {
 	return Message{Role: role, Content: parts}
 }
 
+// ToolResultMessage builds a RoleTool turn carrying the native result of one tool
+// call. toolUseID must equal the provider id of the tool_use block this result
+// answers; isError marks an execution failure; text is the result body.
+func ToolResultMessage(toolUseID, text string, isError bool) Message {
+	return Message{Role: RoleTool, ToolCallID: toolUseID,
+		Content: []Content{{Type: ContentTypeToolResult, Text: text, ToolCallID: toolUseID, IsError: isError}}}
+}
+
+// AssistantToolCallMessage builds a RoleAssistant turn advertising the tool calls
+// the model requested, so a subsequent RoleTool turn can be correlated to them by
+// the provider (OpenAI tool_call_id / Anthropic tool_use_id).
+func AssistantToolCallMessage(calls []ToolCall) Message {
+	parts := make([]Content, 0, len(calls))
+	for _, c := range calls {
+		parts = append(parts, Content{Type: ContentTypeToolCall, ToolCallID: c.ID, ToolCallName: c.Name, ToolArgs: c.Args})
+	}
+	return Message{Role: RoleAssistant, Content: parts}
+}
+
 // ToolSpec is an interface-level placeholder for a tool/function schema. The
 // full tool loop (registration, argument binding, result re-feed) is M2; here
 // it only lets a request advertise tools to the provider in the wire contract.
@@ -99,6 +140,10 @@ type ToolSpec struct {
 // (never a serialized JSON string), so a tool loop can bind it directly without
 // re-parsing.
 type ToolCall struct {
+	// ID is the provider tool_use/tool_call id, so a tool loop can correlate the
+	// resulting RoleTool message to the exact call (Anthropic tool_use_id / OpenAI
+	// tool_call_id). Empty when the provider emitted no id.
+	ID string
 	// Name is the tool being invoked.
 	Name string
 	// Args are the tool arguments, decoded from the provider's structured

internal/runtime/model/openai.go

@@ -48,10 +48,23 @@ type openAIStreamOpts struct {
 }
 
 // openAIMessage is one request message in the wire body. Content is either a
-// plain string (text-only) or an array of {type,text|image_url} parts.
+// plain string (text-only) or an array of {type,text|image_url} parts. Tool
+// messages carry ToolCallID (required by OpenAI); assistant messages that
+// requested functions carry ToolCalls so the following tool messages can be
+// correlated.
 type openAIMessage struct {
-	Role    string `json:"role"`
-	Content any    `json:"content"`
+	Role       string           `json:"role"`
+	Content    any              `json:"content"`
+	ToolCallID string           `json:"tool_call_id,omitempty"`
+	ToolCalls  []openAIToolCall `json:"tool_calls,omitempty"`
+}
+
+// openAIToolCall is the wire form of one assistant tool_call in a request. OpenAI
+// requires each subsequent tool message to reference the id back to these calls.
+type openAIToolCall struct {
+	ID       string           `json:"id"`
+	Type     string           `json:"type"`
+	Function openAIToolCallFn `json:"function"`
 }
 
 // openAITool is the wire form of a function tool schema (M2 placeholder).
@@ -86,8 +99,9 @@ type openAIChoice struct {
 
 // openAIResponseMessage is the assistant message inside a non-streaming choice.
 type openAIResponseMessage struct {
-	Role    string `json:"role"`
-	Content string `json:"content"`
+	Role      string           `json:"role"`
+	Content   string           `json:"content"`
+	ToolCalls []openAIToolCall `json:"tool_calls,omitempty"`
 }
 
 // openAIDelta is the incremental payload inside a stream chunk. Reasoning content
@@ -320,22 +334,75 @@ func (a *OpenAIChatAdapter) buildRequest(req Request, stream bool) ([]byte, erro
 func buildMessages(msgs []Message) ([]openAIMessage, error) {
 	out := make([]openAIMessage, 0, len(msgs))
 	for _, m := range msgs {
-		content, err := encodeContent(m.Content)
+		wire, err := encodeOpenAIMessage(m)
 		if err != nil {
 			return nil, err
 		}
-		out = append(out, openAIMessage{Role: string(m.Role), Content: content})
+		out = append(out, wire)
 	}
 	return out, nil
 }
 
+// encodeOpenAIMessage renders one normalized message in OpenAI wire form. A
+// RoleTool turn becomes a tool message carrying tool_call_id (required by OpenAI,
+// and the audit's gap); an assistant turn that advertised tool calls carries the
+// corresponding tool_calls array so the provider can correlate the tool messages.
+func encodeOpenAIMessage(m Message) (openAIMessage, error) {
+	if m.Role == RoleTool {
+		return openAIMessage{
+			Role:       string(RoleTool),
+			ToolCallID: toolMessageCallID(m),
+			Content:    messageText(m),
+		}, nil
+	}
+	content, err := encodeContent(m.Content)
+	if err != nil {
+		return openAIMessage{}, err
+	}
+	return openAIMessage{Role: string(m.Role), Content: content, ToolCalls: encodeToolCalls(m.Content)}, nil
+}
+
+// toolMessageCallID resolves the tool_call_id for a RoleTool turn: the
+// message-level id, else the id carried on the first tool-result part.
+func toolMessageCallID(m Message) string {
+	if m.ToolCallID != "" {
+		return m.ToolCallID
+	}
+	for _, p := range m.Content {
+		if p.ToolCallID != "" {
+			return p.ToolCallID
+		}
+	}
+	return ""
+}
+
+// encodeToolCalls builds the assistant tool_calls array from a turn's tool-call
+// parts. Arguments are serialized to JSON because OpenAI expects a string.
+func encodeToolCalls(parts []Content) []openAIToolCall {
+	calls := make([]openAIToolCall, 0, len(parts))
+	for _, p := range parts {
+		if p.Type != ContentTypeToolCall {
+			continue
+		}
+		arguments := "{}"
+		if len(p.ToolArgs) > 0 {
+			if b, err := json.Marshal(p.ToolArgs); err == nil {
+				arguments = string(b)
+			}
+		}
+		calls = append(calls, openAIToolCall{
+			ID:       p.ToolCallID,
+			Type:     "function",
+			Function: openAIToolCallFn{Name: p.ToolCallName, Arguments: arguments},
+		})
+	}
+	return calls
+}
+
 func encodeContent(parts []Content) (any, error) {
 	if len(parts) == 0 {
 		return "", nil
 	}
-	if len(parts) == 1 && parts[0].Type == ContentTypeText {
-		return parts[0].Text, nil
-	}
 	arr := make([]any, 0, len(parts))
 	for _, p := range parts {
 		switch p.Type {
@@ -346,10 +413,23 @@ func encodeContent(parts []Content) (any, error) {
 				"type":      "image_url",
 				"image_url": map[string]any{"url": p.ImageURL},
 			})
+		case ContentTypeToolCall, ContentTypeToolResult:
+			// Tool-call and tool-result parts are encoded at the message level
+			// (tool_calls array / tool_call_id), never as text-image content.
+			continue
 		default:
 			return nil, fmt.Errorf("model: %w: unsupported content type %q", ErrBadRequest, p.Type)
 		}
 	}
+	if len(arr) == 0 {
+		return "", nil
+	}
+	if len(arr) == 1 {
+		if blk, ok := arr[0].(map[string]any); ok && blk["type"] == "image_url" {
+			return arr, nil
+		}
+		return arr[0].(map[string]any)["text"], nil
+	}
 	return arr, nil
 }
 
@@ -408,6 +488,13 @@ func responseFromWire(wire *openAIChatResponse) Response {
 		text := wire.Choices[0].Message.Content
 		out.Content = text
 		out.Message = Message{Role: RoleAssistant, Content: []Content{{Type: ContentTypeText, Text: text}}}
+		for _, tc := range wire.Choices[0].Message.ToolCalls {
+			args := map[string]any{}
+			if tc.Function.Arguments != "" {
+				_ = json.Unmarshal([]byte(tc.Function.Arguments), &args)
+			}
+			out.ToolCalls = append(out.ToolCalls, ToolCall{Name: tc.Function.Name, ID: tc.ID, Args: args})
+		}
 	}
 	if u := wire.Usage; u != nil {
 		out.PromptTokens = u.PromptTokens
@@ -553,7 +640,7 @@ func scanSSE(r io.Reader, onEvent func(StreamEvent)) (Response, error) {
 				args = map[string]any{}
 			}
 		}
-		agg.ToolCalls = append(agg.ToolCalls, ToolCall{Name: acc.name, Args: args})
+		agg.ToolCalls = append(agg.ToolCalls, ToolCall{Name: acc.name, ID: acc.id, Args: args})
 	}
 	agg.Content = text.String()
 	agg.Message = Message{Role: RoleAssistant, Content: []Content{{Type: ContentTypeText, Text: agg.Content}}}

internal/runtime/model/openai_test.go

@@ -352,3 +352,80 @@ func TestOpenAIChat_ChatStream_ToolCall(t *testing.T) {
 		t.Errorf("events = %+v, want %+v", events, wantEvents)
 	}
 }
+
+func TestOpenAIChat_Chat_ToolRoundTrip(t *testing.T) {
+	t.Parallel()
+	ch := make(chan capturedRequest, 1)
+	srv := newCompletionServer(t, http.StatusOK, "application/json", okCompletion, ch)
+	ad := NewOpenAIChat(config.LLMConfig{BaseURL: srv.URL, Model: "gpt-test"})
+
+	req := Request{Messages: []Message{
+		TextMessage(RoleUser, "what's the weather in Paris?"),
+		AssistantToolCallMessage([]ToolCall{{ID: "call_1", Name: "weather", Args: map[string]any{"city": "Paris"}}}),
+		ToolResultMessage("call_1", `{"temp":22}`, true),
+	}, Model: "gpt-test"}
+	if _, err := ad.Chat(context.Background(), req, Options{}); err != nil {
+		t.Fatalf("Chat error: %v", err)
+	}
+
+	got := <-ch
+	msgs, _ := got.body["messages"].([]any)
+	if len(msgs) != 3 {
+		t.Fatalf("body.messages = %v, want 3 messages", got.body["messages"])
+	}
+
+	// The assistant turn must carry tool_calls (so the tool message can be
+	// correlated by the provider) — this was previously dropped entirely.
+	assistant, _ := msgs[1].(map[string]any)
+	if assistant["role"] != "assistant" {
+		t.Errorf("assistant role = %v, want assistant", assistant["role"])
+	}
+	tcs, _ := assistant["tool_calls"].([]any)
+	if len(tcs) != 1 {
+		t.Fatalf("assistant tool_calls = %v, want 1 entry", assistant["tool_calls"])
+	}
+	tc := tcs[0].(map[string]any)
+	if tc["id"] != "call_1" || tc["type"] != "function" {
+		t.Errorf("tool_call = %v, want id=call_1 type=function", tc)
+	}
+	fn, _ := tc["function"].(map[string]any)
+	if fn["name"] != "weather" || fn["arguments"] != `{"city":"Paris"}` {
+		t.Errorf("function = %#v, want name=weather arguments={\"city\":\"Paris\"}", fn)
+	}
+
+	// The tool message must carry tool_call_id (the audit's gap) and the result.
+	tool, _ := msgs[2].(map[string]any)
+	if tool["role"] != "tool" {
+		t.Errorf("tool role = %v, want tool", tool["role"])
+	}
+	if tool["tool_call_id"] != "call_1" {
+		t.Errorf("tool_call_id = %v, want call_1", tool["tool_call_id"])
+	}
+	if tool["content"] != `{"temp":22}` {
+		t.Errorf("tool content = %v, want the raw result", tool["content"])
+	}
+}
+
+func TestOpenAIChat_Build_ToolMessageLegacy(t *testing.T) {
+	t.Parallel()
+	// The loop's current form: a RoleTool message carrying a plain text part and
+	// Message.ToolCallID. buildMessages must emit a tool message with tool_call_id.
+	m := Message{Role: RoleTool, ToolCallID: "call_9",
+		Content: []Content{{Type: ContentTypeText, Text: "tool weather result: 22"}}}
+	msgs, err := buildMessages([]Message{m})
+	if err != nil {
+		t.Fatalf("buildMessages error: %v", err)
+	}
+	if len(msgs) != 1 {
+		t.Fatalf("messages = %d, want 1", len(msgs))
+	}
+	if msgs[0].Role != "tool" {
+		t.Errorf("role = %q, want tool", msgs[0].Role)
+	}
+	if msgs[0].ToolCallID != "call_9" {
+		t.Errorf("tool_call_id = %q, want call_9", msgs[0].ToolCallID)
+	}
+	if msgs[0].Content != "tool weather result: 22" {
+		t.Errorf("content = %v, want the text", msgs[0].Content)
+	}
+}

internal/runtime/tool/tool.go

@@ -17,6 +17,7 @@ import (
 	"time"
 
 	"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/pkg/plugins/sdk"
 )
@@ -187,6 +188,19 @@ func (d *Dispatcher) IsSequential(name string) bool {
 	return d.modes[name] == ExecutionModeSequential
 }
 
+// PolicyEngine returns a restrict.Policy scoped to env, reading the editable
+// reachability markers from the underlying capability registry (design/11 §3).
+// It is the loop's authoritative source for the group/stranger tool gate and is
+// fail-closed: a nil receiver or nil registry denies every tool in a non-owner
+// environment, and the owner environment always allows. The returned policy is
+// immutable and safe for concurrent use.
+func (d *Dispatcher) PolicyEngine(env restrict.ActorEnv) *restrict.Policy {
+	if d == nil || d.reg == nil {
+		return restrict.SetContext(nil, env)
+	}
+	return restrict.SetContext(d.reg, env)
+}
+
 // Invoke executes the tool registered as name with the given args. It resolves
 // the tool from the registry, validates args against the tool's Parameters JSON
 // Schema, and runs Execute under the tool's deadline (per-tool override, else

internal/services/gateway/extra_test.go

@@ -859,6 +859,26 @@ func TestHubBroadcastStream(t *testing.T) {
 	}
 }
 
+func TestServerBroadcastCapacityAwaken(t *testing.T) {
+	t.Parallel()
+	env := newEnv(t, nil, false)
+	c := newFakeConn(env.srv.hub, 32)
+	env.srv.hub.register(c)
+
+	env.srv.BroadcastCapacityAwaken("web", "quiet", "awake", "web_search")
+	ev := readFakeFrame(t, c)
+	if ev.Type != "capacity_awaken" {
+		t.Fatalf("frame type = %q, want capacity_awaken (frame=%+v)", ev.Type, ev)
+	}
+	if ev.SessionID != "web" {
+		t.Fatalf("frame sessionId = %q, want web", ev.SessionID)
+	}
+	data, ok := ev.Data.(map[string]any)
+	if !ok || data["fromState"] != "quiet" || data["toState"] != "awake" || data["tool"] != "web_search" {
+		t.Fatalf("frame data = %+v, want {fromState:quiet,toState:awake,tool:web_search}", ev.Data)
+	}
+}
+
 func TestHubStreamChatIdle(t *testing.T) {
 	t.Parallel()
 	hb := newTestHub(nil, nil) // nil runner

internal/services/gateway/server.go

@@ -508,6 +508,28 @@ func (s *Server) BroadcastPersisted(msg history.Message) {
 	s.hub.broadcast(wsEvent{Type: "message_persisted", SessionID: msg.SessionID, Data: map[string]any{"message": wire}})
 }
 
+// wsCapacityAwaken is the WS event type for a capacity state transition
+// (design/14 §5). Its data field is the frontend AwakenPayload {fromState,toState,tool}.
+const wsCapacityAwaken = "capacity_awaken"
+
+// BroadcastCapacityAwaken fans a capacity state transition to the Web UI as a
+// `capacity_awaken` frame (design/14 §5). sessionID is the conversation that
+// transitioned, so the client targets the correct session rather than treating the
+// wake event as global. data matches the frontend AwakenPayload {fromState, toState,
+// tool}; tool may be empty (no single tool was lit, the client then keeps the prior
+// tool badge).
+func (s *Server) BroadcastCapacityAwaken(sessionID, fromState, toState, tool string) {
+	s.hub.broadcast(wsEvent{
+		Type:      wsCapacityAwaken,
+		SessionID: sessionID,
+		Data: map[string]any{
+			"fromState": fromState,
+			"toState":   toState,
+			"tool":      tool,
+		},
+	})
+}
+
 // sourcedLog returns a logger tagged with source "gateway", substituting a
 // discard logger for nil so Server can always log safely.
 func sourcedLog(l *slog.Logger) *slog.Logger {