agentengine-sdk-python 0.8.4__py3-none-any.whl → 0.8.5__py3-none-any.whl

ksadk/plugins/teams/domain.py

@@ -0,0 +1,1427 @@
+"""Persistent collaboration decisions. No models, network calls or Kernel loops.
+
+The plugin dispatcher submits this domain's outbox through Host Services.
+Provider tools invoke it with a verified Actor. Kernel facts return through
+the projector; a model's result candidate never completes a task by itself.
+"""
+
+from __future__ import annotations
+
+from copy import deepcopy
+from datetime import datetime
+from typing import Any, Callable
+
+from .contracts import (
+    API_VERSION,
+    TERMINAL,
+    Actor,
+    Budget,
+    GroupCreateInput,
+    MemberInput,
+    MessageInput,
+    TaskCreateInput,
+    plain_text,
+)
+from .errors import TeamsError
+from .store import TeamsStore, Transaction, digest, new_id, now
+
+
+from .domain_values import member_key, public, run_member_key
+from .task_decisions import TaskDecisions
+
+
+class TeamsDomain(TaskDecisions):
+    def __init__(self, store: TeamsStore, *, authority_ref: str) -> None:
+        self.store = store
+        self.authority_ref = authority_ref
+
+    def authorize(
+        self,
+        tx: Transaction,
+        actor: Actor,
+        group_id: str,
+        *,
+        owner: bool = False,
+        leader: bool = False,
+    ) -> dict[str, Any]:
+        group = tx.get("group", group_id)
+        if group["tenantId"] != actor.tenant_id:
+            raise TeamsError("not_found", "团队不存在或不可访问", status=404)
+        if actor.kind == "human":
+            if group["ownerSubject"] != actor.subject:
+                raise TeamsError("forbidden", "没有访问该团队的权限", status=403)
+        elif actor.kind == "member":
+            if owner or actor.group_id != group_id or not actor.member_id:
+                raise TeamsError("forbidden", "成员无权执行此操作", status=403)
+            if not actor.team_run_id:
+                raise TeamsError("invocation_required", "缺少受信执行上下文", status=403)
+            run = tx.get("team_run", actor.team_run_id)
+            if (
+                run["groupId"] != group_id
+                or run["status"] in TERMINAL
+                or run["status"] == "cancel_requested"
+            ):
+                raise TeamsError("team_run_closed", "本轮已停止，不能继续协作")
+            member = self.run_member(tx, actor.team_run_id, actor.member_id)
+            if member["status"] != "active" or (leader and member["role"] != "leader"):
+                raise TeamsError("forbidden", "成员授权已失效", status=403)
+        elif actor.kind != "host":
+            raise TeamsError("forbidden", "无效操作主体", status=403)
+        return group
+
+    @staticmethod
+    def run_member(tx: Transaction, run_id: str, member_id: str) -> dict[str, Any]:
+        """Resolve the frozen invocation identity, including pre-RunMember history."""
+        key = run_member_key(run_id, member_id)
+        stored = tx.get("run_member", key, required=False)
+        if stored:
+            return stored
+        run = tx.get("team_run", run_id)
+        frozen = next((m for m in run.get("_roster", []) if m["memberId"] == member_id), None)
+        if not frozen or frozen["status"] != "active":
+            raise TeamsError("member_unavailable", "成员不属于本轮冻结配置", status=409)
+        return {
+            **deepcopy(frozen),
+            "runMemberId": key,
+            "teamRunId": run_id,
+            "groupRevision": run["groupRevision"],
+        }
+
+    @staticmethod
+    def delivery_member(tx: Transaction, delivery: dict[str, Any]) -> dict[str, Any]:
+        """Historical execution identity must survive a Leader session takeover."""
+        if delivery.get("_memberSnapshot"):
+            return deepcopy(delivery["_memberSnapshot"])
+        run = tx.get("team_run", delivery["teamRunId"])
+        frozen = next(
+            (
+                m
+                for m in run.get("_roster", [])
+                if m["memberId"] == delivery["memberId"]
+                and m["sessionId"] == delivery["_sessionId"]
+            ),
+            None,
+        )
+        if frozen is None:
+            raise TeamsError(
+                "delivery_scope_unavailable", "原始执行身份不可用，需要核对历史来源", status=409
+            )
+        return {
+            **deepcopy(frozen),
+            "runMemberId": run_member_key(run["teamRunId"], frozen["memberId"]),
+            "teamRunId": run["teamRunId"],
+            "groupRevision": run["groupRevision"],
+        }
+
+    @staticmethod
+    def run_policy(tx: Transaction, run: dict[str, Any]) -> dict[str, Any]:
+        if "_policy" in run:
+            return run["_policy"]
+        revision = tx.get(
+            "group_revision", f"{run['groupId']}:{run['groupRevision']}", required=False
+        )
+        return (revision or tx.get("group", run["groupId"]))["_policy"]
+
+    def publish_run_member(self, tx: Transaction, member: dict[str, Any]) -> None:
+        self.publish(tx, "run_member", member["runMemberId"], member)
+        current = tx.get("member", member_key(member["groupId"], member["memberId"]))
+        active = [
+            m
+            for m in tx.list("run_member", member["groupId"])
+            if m["memberId"] == member["memberId"] and m["executionStatus"] != "idle"
+        ]
+        state = next(
+            (
+                state
+                for state in ("running", "waiting", "queued")
+                if any(m["executionStatus"] == state for m in active)
+            ),
+            "idle",
+        )
+        current.update(
+            executionStatus=state,
+            activeRunId=active[0].get("activeRunId") if len(active) == 1 else None,
+            activeTeamRunIds=[m["teamRunId"] for m in active],
+            revision=current["revision"] + 1,
+        )
+        self.publish(tx, "member", member_key(member["groupId"], member["memberId"]), current)
+
+    def update_active_clock(self, tx: Transaction, run: dict[str, Any]) -> None:
+        timestamp = now()
+        previous = run.get("activeDurationSeconds", 0)
+        if run.get("_activeSince"):
+            elapsed = (
+                datetime.fromisoformat(timestamp) - datetime.fromisoformat(run["_activeSince"])
+            ).total_seconds()
+            total = previous + run.get("_activeRemainderSeconds", 0) + max(0, elapsed)
+            run["activeDurationSeconds"] = int(total)
+            run["_activeRemainderSeconds"] = total - int(total)
+        executing = (
+            run["status"] not in TERMINAL | {"cancel_requested", "awaiting_acceptance"}
+            and not run["dispatchSuspended"]
+            and any(
+                d.get("_runStatus") == "running"
+                and not d.get("_terminalState")
+                and not d.get("_fenced")
+                for d in tx.list("delivery", run["groupId"], team_run_id=run["teamRunId"])
+            )
+        )
+        run["_activeSince"] = timestamp if executing else None
+        if run.get("activeDurationSeconds", 0) != previous:
+            run["revision"] += 1
+            self.publish(tx, "team_run", run["teamRunId"], run)
+
+    def mutate(
+        self,
+        actor: Actor,
+        group_id: str,
+        key: str,
+        payload: Any,
+        operation: Callable[[Transaction], Any],
+        *,
+        owner: bool = False,
+        leader: bool = False,
+    ) -> Any:
+        return self.store.mutate(
+            f"{actor.tenant_id}:{group_id}:{actor.kind}:{actor.subject}:{actor.member_id or ''}",
+            key,
+            payload,
+            operation,
+            authorize=lambda tx: self.authorize(tx, actor, group_id, owner=owner, leader=leader),
+        )
+
+    @staticmethod
+    def publish(
+        tx: Transaction,
+        kind: str,
+        key: str,
+        value: dict[str, Any],
+        *,
+        created: bool = False,
+        **refs: Any,
+    ) -> None:
+        tx.put(kind, key, value)
+        event_type = {
+            "team_run": "team_run.updated",
+            "message": "message.created" if created else "message.updated",
+        }.get(kind, f"{kind}.updated")
+        payload_key = {"team_run": "teamRun", "run_member": "runMember"}.get(kind, kind)
+        event = tx.event(value["groupId"], event_type, {payload_key: public(value)}, **refs)
+        if kind == "message" and created:
+            value["_createdSeq"] = event["groupSeq"]
+            tx.put(kind, key, value)
+
+    def create_group(
+        self, actor: Actor, request: GroupCreateInput, bindings: dict[str, dict[str, Any]]
+    ) -> dict[str, Any]:
+        if actor.kind != "human" or not actor.subject or not actor.tenant_id:
+            raise TeamsError("human_owner_required", "需要真实用户创建团队", status=403)
+        for member in request.members:
+            binding = bindings.get(member.bindingRef)
+            if not binding or not binding.get("capabilities", {}).get("enqueue"):
+                raise TeamsError("binding_unavailable", "所选成员尚不支持团队执行", status=422)
+            if member.memberId == request.leaderMemberId and not binding["capabilities"].get(
+                "leader"
+            ):
+                raise TeamsError("leader_unavailable", "所选成员尚不支持协调团队", status=422)
+
+        def create(tx: Transaction) -> dict[str, Any]:
+            group_id, timestamp = new_id("grp"), now()
+            group = {
+                "groupId": group_id,
+                "name": request.name,
+                "authorityRef": self.authority_ref,
+                "tenantId": actor.tenant_id,
+                "ownerSubject": actor.subject,
+                "leaderMemberId": request.leaderMemberId,
+                "revision": 1,
+                "status": "active",
+                "createdAt": timestamp,
+                "updatedAt": timestamp,
+                "_policy": {"taskAcceptance": request.taskAcceptance, "peerWake": False},
+            }
+            if request.leaderStandbyBindingRef:
+                group["_leaderStandbyBindingRef"] = request.leaderStandbyBindingRef
+            self.publish(tx, "group", group_id, group)
+            for selected in request.members:
+                member = {
+                    "memberId": selected.memberId,
+                    "groupId": group_id,
+                    "name": selected.name,
+                    "role": "leader" if selected.memberId == request.leaderMemberId else "member",
+                    "bindingRef": selected.bindingRef,
+                    "binding": deepcopy(bindings[selected.bindingRef]),
+                    "responsibility": selected.responsibility,
+                    "sessionId": new_id("ses"),
+                    "revision": 1,
+                    "status": "active",
+                    "executionStatus": "idle",
+                }
+                self.publish(tx, "member", member_key(group_id, selected.memberId), member)
+            tx.put(
+                "group_revision", f"{group_id}:1", {**group, "members": tx.list("member", group_id)}
+            )
+            return self._snapshot(tx, group_id)
+
+        return self.store.mutate(
+            f"{actor.tenant_id}:{actor.subject}:create_group",
+            request.idempotencyKey,
+            request.model_dump(),
+            create,
+        )
+
+    def snapshot(
+        self, actor: Actor, group_id: str, team_run_id: str | None = None
+    ) -> dict[str, Any]:
+        with self.store.transaction() as tx:
+            self.authorize(tx, actor, group_id)
+            if actor.kind == "member" and team_run_id and team_run_id != actor.team_run_id:
+                raise TeamsError("forbidden", "不能读取其他轮次", status=403)
+            if team_run_id and tx.get("team_run", team_run_id)["groupId"] != group_id:
+                raise TeamsError("not_found", "轮次不存在", status=404)
+            return self._snapshot(tx, group_id, actor=actor, team_run_id=team_run_id)
+
+    def _snapshot(
+        self,
+        tx: Transaction,
+        group_id: str,
+        *,
+        actor: Actor | None = None,
+        team_run_id: str | None = None,
+    ) -> dict[str, Any]:
+        snapshot = {
+            "apiVersion": API_VERSION,
+            "group": tx.get("group", group_id),
+            "watermark": tx.watermark(group_id),
+        }
+        for key, kind in {
+            "members": "member",
+            "runMembers": "run_member",
+            "messages": "message",
+            "teamRuns": "team_run",
+            "tasks": "task",
+            "deliveries": "delivery",
+            "interactions": "interaction",
+            "artifacts": "artifact",
+        }.items():
+            values = tx.list(kind, group_id)
+            if kind == "run_member":
+                indexed = {m["runMemberId"]: m for m in values}
+                for run in tx.list("team_run", group_id):
+                    for frozen in run.get("_roster", []):
+                        if frozen["status"] == "active":
+                            identity = run_member_key(run["teamRunId"], frozen["memberId"])
+                            indexed.setdefault(
+                                identity, self.run_member(tx, run["teamRunId"], frozen["memberId"])
+                            )
+                values = list(indexed.values())
+            selected_run = actor.team_run_id if actor and actor.kind == "member" else team_run_id
+            if selected_run and kind != "member":
+                values = [v for v in values if v.get("teamRunId") == selected_run]
+            if key == "messages":
+                values = [value for value in values if value.get("visibility") != "internal"]
+            if actor and actor.kind == "member" and key == "interactions":
+                values = [value for value in values if value["ref"]["memberId"] == actor.member_id]
+            snapshot[key] = values
+        return public(snapshot)
+
+    def list_groups(
+        self, actor: Actor, *, cursor: str | None = None, limit: int = 50
+    ) -> dict[str, Any]:
+        if actor.kind != "human" or not 1 <= limit <= 100:
+            raise TeamsError("forbidden", "无法列出团队", status=403)
+        with self.store.transaction() as tx:
+            groups = sorted(
+                (
+                    group
+                    for group in tx.list("group")
+                    if group["tenantId"] == actor.tenant_id
+                    and group["ownerSubject"] == actor.subject
+                ),
+                key=lambda group: group["groupId"],
+            )
+            groups = [group for group in groups if not cursor or group["groupId"] > cursor]
+            items = []
+            for group in groups[:limit]:
+                gid = group["groupId"]
+                runs = tx.list("team_run", gid)
+                active = next(
+                    (run for run in reversed(runs) if run["status"] not in TERMINAL), None
+                )
+                messages = tx.list("message", gid)
+                read = tx.get("read_watermark", f"{gid}:{actor.subject}", required=False)
+                items.append(
+                    {
+                        **group,
+                        "memberCount": sum(
+                            member["status"] == "active" for member in tx.list("member", gid)
+                        ),
+                        "pendingCount": sum(
+                            item["status"] == "pending" for item in tx.list("interaction", gid)
+                        )
+                        + sum(
+                            task["status"] == "awaiting_acceptance" for task in tx.list("task", gid)
+                        )
+                        + sum(run["status"] == "awaiting_acceptance" for run in runs),
+                        "unreadCount": sum(
+                            message.get("_createdSeq", 0) > (read or {}).get("watermark", 0)
+                            and message.get("senderPrincipal") != actor.subject
+                            for message in messages
+                        ),
+                        "activeTeamRunId": active["teamRunId"] if active else None,
+                        "activeTeamRunIds": [
+                            r["teamRunId"] for r in runs if r["status"] not in TERMINAL
+                        ],
+                        "activeStatus": active["status"] if active else None,
+                        "lastMessage": plain_text(messages[-1]["parts"])[:160] if messages else "",
+                    }
+                )
+            return {
+                "items": public(items),
+                "nextCursor": items[-1]["groupId"] if len(groups) > limit else None,
+            }
+
+    @staticmethod
+    def active_run(tx: Transaction, group_id: str) -> dict[str, Any] | None:
+        return next(
+            (
+                run
+                for run in reversed(tx.list("team_run", group_id))
+                if run["status"] not in TERMINAL
+            ),
+            None,
+        )
+
+    def _start(
+        self,
+        tx: Transaction,
+        group: dict[str, Any],
+        message: dict[str, Any],
+        budget: Budget | None = None,
+    ) -> dict[str, Any]:
+        existing = next(
+            (
+                run
+                for run in tx.list("team_run", group["groupId"])
+                if run["goalMessageId"] == message["messageId"]
+            ),
+            None,
+        )
+        if existing:
+            return existing
+        if group.get("legacySource"):
+            raise TeamsError(
+                "legacy_history_read_only", "导入历史只读，请创建新团队执行任务", status=409
+            )
+        if group["status"] != "active":
+            raise TeamsError("group_archived", "团队已归档", status=409)
+        timestamp = now()
+        run_id = new_id("tr")
+        roster = []
+        for current in tx.list("member", group["groupId"]):
+            if current["status"] != "active":
+                continue
+            member = {
+                **deepcopy(current),
+                "runMemberId": run_member_key(run_id, current["memberId"]),
+                "teamRunId": run_id,
+                "groupRevision": group["revision"],
+                "sessionId": new_id("ses"),
+                "revision": 1,
+                "executionStatus": "idle",
+                "activeRunId": None,
+            }
+            member.pop("_bindingHistory", None)
+            member.pop("activeTeamRunIds", None)
+            roster.append(member)
+        run = {
+            "teamRunId": run_id,
+            "groupId": group["groupId"],
+            "groupRevision": group["revision"],
+            "revision": 1,
+            "goalMessageId": message["messageId"],
+            "goal": plain_text(message["parts"]),
+            "leaderMemberId": group["leaderMemberId"],
+            "status": "planning",
+            "dispatchSuspended": False,
+            "dispatchEpoch": 1,
+            "budget": (budget or Budget()).model_dump(),
+            "createdAt": timestamp,
+            "updatedAt": timestamp,
+            "_roster": roster,
+            "_leaderStandbyBindingRef": group.get("_leaderStandbyBindingRef"),
+            "_policy": deepcopy(group["_policy"]),
+            "activeDurationSeconds": 0,
+            "workspace": deepcopy(message.get("workspace")),
+        }
+        self.publish(tx, "team_run", run["teamRunId"], run)
+        for member in roster:
+            self.publish(tx, "run_member", member["runMemberId"], member)
+        return run
+
+    def message_run(self, tx: Transaction, group_id: str, request: MessageInput):
+        if request.teamRunId:
+            run = tx.get("team_run", request.teamRunId)
+            if run["groupId"] != group_id:
+                raise TeamsError("not_found", "轮次不属于当前团队", status=404)
+            if request.intent != "note":
+                self._task_run(tx, group_id, request.teamRunId)
+            return run
+        if request.intent in {"start_goal", "note"}:
+            return None
+        active = [r for r in tx.list("team_run", group_id) if r["status"] not in TERMINAL]
+        if len(active) > 1:
+            raise TeamsError(
+                "team_run_ambiguous", "团队有多个进行中的任务，请选择本条消息所属任务", status=422
+            )
+        return active[0] if active else None
+
+    def send(self, actor: Actor, group_id: str, request: MessageInput) -> dict[str, Any]:
+        def send(tx: Transaction) -> dict[str, Any]:
+            group = self.authorize(tx, actor, group_id)
+            if group["status"] != "active":
+                raise TeamsError("group_archived", "团队已归档")
+            if request.replyTo:
+                parent = tx.get("message", request.replyTo)
+                if parent["groupId"] != group_id:
+                    raise TeamsError("invalid_reply", "不能引用其他团队的消息", status=422)
+            if actor.kind != "human" and request.intent == "start_goal":
+                raise TeamsError("human_goal_required", "只有群主可以发起新目标", status=403)
+            run = None if actor.kind == "member" else self.message_run(tx, group_id, request)
+            if actor.kind == "member":
+                if request.teamRunId and request.teamRunId != actor.team_run_id:
+                    raise TeamsError("forbidden", "不能发送到其他轮次", status=403)
+                run = self._task_run(tx, group_id, actor.team_run_id)
+            member = (
+                self.run_member(tx, actor.team_run_id, actor.member_id) if actor.member_id else None
+            )
+            message = {
+                "messageId": new_id("gm"),
+                "groupId": group_id,
+                "revision": 1,
+                "createdAt": now(),
+                "senderPrincipal": actor.subject,
+                "senderName": member["name"] if member else "我",
+                "groupRole": member["role"] if member else "owner",
+                "memberId": actor.member_id,
+                "parts": [part.model_dump(exclude_none=True) for part in request.parts],
+                "mentions": request.mentions,
+                "intent": request.intent,
+                "replyTo": request.replyTo,
+                "visibility": "public",
+                "workspace": request.workspace.model_dump() if request.workspace else None,
+            }
+            starting = request.intent == "start_goal" or (
+                request.intent == "directed" and run is None and actor.kind == "human"
+            )
+            if starting:
+                run = self._start(tx, group, message)
+            elif request.intent != "note" and run is None:
+                raise TeamsError("team_run_required", "请先发起本轮目标，或选择仅留言", status=422)
+            if run:
+                message["teamRunId"] = run["teamRunId"]
+            if request.replyTo and parent.get("teamRunId") != message.get("teamRunId"):
+                raise TeamsError("invalid_reply", "回复必须属于同一轮次", status=422)
+            for member_id in request.mentions:
+                self.run_member(tx, run["teamRunId"], member_id) if run else self._member(
+                    tx, group_id, member_id
+                )
+            for part in request.parts:
+                if part.kind == "attachment":
+                    artifact = tx.get("artifact", part.attachmentRef)
+                    if artifact["groupId"] != group_id:
+                        raise TeamsError("attachment_forbidden", "附件不属于该团队", status=403)
+                    # Explicit attachment sharing grants only this destination round access.
+                    if run and artifact.get("teamRunId") != run["teamRunId"]:
+                        if actor.kind != "human":
+                            raise TeamsError(
+                                "attachment_forbidden", "跨任务交付物只能由群主显式分享", status=403
+                            )
+                        grants = run.setdefault("_sharedArtifactIds", [])
+                        if part.attachmentRef not in grants:
+                            grants.append(part.attachmentRef)
+                            tx.put("team_run", run["teamRunId"], run)
+            self.publish(tx, "message", message["messageId"], message, created=True)
+            if request.intent != "note":
+                if actor.kind == "member":
+                    # Member text is never a trigger. Explicit tool wake is separate.
+                    pass
+                elif run["status"] == "cancel_requested":
+                    raise TeamsError("stop_in_progress", "本轮正在停止，请等待执行结束")
+                else:
+                    targets = (
+                        [run["leaderMemberId"]]
+                        if starting
+                        else request.mentions or [run["leaderMemberId"]]
+                    )
+                    for target in targets:
+                        self.dispatch(
+                            tx,
+                            run,
+                            message,
+                            target,
+                            source=message["messageId"],
+                            purpose="goal" if starting else "message",
+                        )
+            return {
+                "status": "accepted",
+                "groupId": group_id,
+                "messageId": message["messageId"],
+                "teamRunId": run["teamRunId"] if run else None,
+                "groupSeq": tx.watermark(group_id),
+            }
+
+        return self.mutate(
+            actor,
+            group_id,
+            request.idempotencyKey,
+            {"operation": "send", **request.model_dump()},
+            send,
+        )
+
+    def start_goal(self, actor: Actor, group_id: str, message_id: str, key: str) -> dict[str, Any]:
+        def start(tx: Transaction) -> dict[str, Any]:
+            group = tx.get("group", group_id)
+            message = tx.get("message", message_id)
+            if message["groupId"] != group_id or message["groupRole"] != "owner":
+                raise TeamsError("invalid_goal", "目标必须引用本群的用户消息", status=422)
+            run = self._start(tx, group, message)
+            self.dispatch(
+                tx, run, message, run["leaderMemberId"], source=message_id, purpose="goal"
+            )
+            return public(run)
+
+        return self.mutate(
+            actor, group_id, key, {"operation": "start", "messageId": message_id}, start, owner=True
+        )
+
+    @staticmethod
+    def _member(tx: Transaction, group_id: str, member_id: str) -> dict[str, Any]:
+        member = tx.get("member", member_key(group_id, member_id))
+        if member["status"] != "active":
+            raise TeamsError("member_unavailable", "目标成员已移除或不可用")
+        return member
+
+    def dispatch(
+        self,
+        tx: Transaction,
+        run: dict[str, Any],
+        message: dict[str, Any],
+        target: str,
+        *,
+        source: str,
+        purpose: str,
+        hop: int = 0,
+        root: str | None = None,
+        task_id: str | None = None,
+        attempt_id: str | None = None,
+    ) -> dict[str, Any] | None:
+        # Always re-read the budget: several targets can be dispatched in one tx.
+        run = tx.get("team_run", run["teamRunId"])
+        decision_id = "td_" + digest([run["teamRunId"], source, target, purpose])[7:]
+        previous = tx.get("trigger", decision_id, required=False)
+        if previous:
+            return (
+                tx.get("delivery", previous["deliveryId"], required=False)
+                if previous.get("deliveryId")
+                else None
+            )
+        member = self.run_member(tx, run["teamRunId"], target)
+        pending = sum(
+            delivery["memberId"] == target and delivery["status"] in {"pending", "uncertain"}
+            for delivery in tx.list("delivery", run["groupId"], team_run_id=run["teamRunId"])
+        )
+        reason = None
+        if run["status"] in TERMINAL or run["status"] == "cancel_requested":
+            reason = "team_run_closed"
+        elif hop > run["budget"]["maxHops"]:
+            reason = "hop_limit"
+        elif run["budget"]["startsUsed"] >= run["budget"]["maxStarts"]:
+            reason = "start_budget_exhausted"
+        elif pending >= 64:
+            reason = "member_queue_full"
+        decision = {
+            "decisionId": decision_id,
+            "groupId": run["groupId"],
+            "teamRunId": run["teamRunId"],
+            "sourceEventRef": source,
+            "rootTriggerId": root or source,
+            "hop": hop,
+            "targetMemberId": target,
+            "purpose": purpose,
+            "blockedReason": reason,
+        }
+        if reason:
+            tx.put("trigger", decision_id, decision)
+            if run["status"] not in TERMINAL and run["status"] != "cancel_requested":
+                run.update(
+                    status="needs_attention",
+                    reason=reason,
+                    revision=run["revision"] + 1,
+                    updatedAt=now(),
+                )
+                self.publish(tx, "team_run", run["teamRunId"], run)
+            tx.event(run["groupId"], "trigger.blocked", {"decision": decision})
+            return None
+        delivery = {
+            "deliveryId": new_id("gd"),
+            "groupId": run["groupId"],
+            "teamRunId": run["teamRunId"],
+            "messageId": message["messageId"],
+            "memberId": target,
+            "revision": 1,
+            "status": "pending",
+            "_decisionId": decision_id,
+            "_bindingRef": member["bindingRef"],
+            "_sessionId": member["sessionId"],
+            "_memberSnapshot": deepcopy(member),
+            "_dispatchEpoch": run["dispatchEpoch"],
+            "_payload": plain_text(message["parts"]),
+            "_parts": message["parts"],
+            "_taskId": task_id,
+            "_attemptId": attempt_id,
+            "_nextRetryAt": 0,
+            "_createdAt": now(),
+        }
+        delivery["_payloadDigest"] = digest(
+            {
+                "content": delivery["_payload"],
+                "parts": delivery["_parts"],
+                "taskId": task_id,
+                "attemptId": attempt_id,
+            }
+        )
+        decision["deliveryId"] = delivery["deliveryId"]
+        tx.put("trigger", decision_id, decision)
+        self.publish(tx, "delivery", delivery["deliveryId"], delivery)
+        run["budget"]["startsUsed"] += 1
+        run.update(revision=run["revision"] + 1, updatedAt=now())
+        self.publish(tx, "team_run", run["teamRunId"], run)
+        if member["executionStatus"] == "idle":
+            member.update(executionStatus="queued", revision=member["revision"] + 1)
+            self.publish_run_member(tx, member)
+        return delivery
+
+    def control(
+        self,
+        actor: Actor,
+        group_id: str,
+        run_id: str,
+        action: str,
+        expected_revision: int,
+        key: str,
+    ) -> dict[str, Any]:
+        if action not in {"suspend_dispatch", "resume_dispatch", "stop"}:
+            raise TeamsError("unsupported_control", "不支持该团队控制", status=422)
+
+        def control(tx: Transaction) -> dict[str, Any]:
+            run = tx.get("team_run", run_id)
+            if run["groupId"] != group_id:
+                raise TeamsError("not_found", "本轮不存在", status=404)
+            if run["revision"] != expected_revision:
+                raise TeamsError("revision_conflict", "本轮状态已更新，请刷新后重试")
+            if run["status"] in TERMINAL:
+                raise TeamsError("team_run_closed", "本轮已结束")
+            if run["status"] == "cancel_requested" and action != "stop":
+                raise TeamsError("stop_in_progress", "已请求停止，不能恢复旧命令")
+            run["dispatchEpoch"] += 1
+            # Resume is an asynchronous barrier operation too. Do not admit
+            # dispatch until all frozen member scopes are confirmed active.
+            run["dispatchSuspended"] = True
+            if action == "stop":
+                run["status"] = "cancel_requested"
+            run.update(revision=run["revision"] + 1, updatedAt=now())
+            self.update_active_clock(tx, run)
+            self.publish(tx, "team_run", run_id, run)
+            for delivery in tx.list("delivery", group_id, team_run_id=run_id):
+                if delivery["status"] == "pending":
+                    if action == "stop":
+                        delivery["status"] = "cancelled"
+                    delivery.update(
+                        _dispatchEpoch=run["dispatchEpoch"], revision=delivery["revision"] + 1
+                    )
+                    self.publish(tx, "delivery", delivery["deliveryId"], delivery)
+            # Durable control outbox: the runtime must obtain EVERY member's
+            # Kernel barrier before acknowledging a completed stop/pause.
+            control_id = new_id("gc")
+            tx.put(
+                "control",
+                control_id,
+                {
+                    "controlId": control_id,
+                    "groupId": group_id,
+                    "teamRunId": run_id,
+                    "action": action,
+                    "epoch": run["dispatchEpoch"],
+                    "status": "pending",
+                    "barriers": {},
+                },
+            )
+            return public(run)
+
+        return self.mutate(
+            actor,
+            group_id,
+            key,
+            {
+                "operation": "control",
+                "runId": run_id,
+                "action": action,
+                "expectedRevision": expected_revision,
+            },
+            control,
+            owner=True,
+        )
+
+    def events(
+        self, actor: Actor, group_id: str, after: int = 0, limit: int = 200
+    ) -> list[dict[str, Any]]:
+        with self.store.transaction() as tx:
+            self.authorize(tx, actor, group_id)
+            if actor.kind == "member":
+                raise TeamsError("owner_stream_only", "成员通过受限上下文读取群信息", status=403)
+        return self.store.events(group_id, after, limit)
+
+    def project_run(
+        self,
+        *,
+        delivery_id: str,
+        event_id: str,
+        run_id: str,
+        status: str,
+        output: str = "",
+        tokens: int = 0,
+    ) -> dict[str, Any]:
+        """Trusted host projection after consulting the canonical Run record.
+
+        The plugin transport does not expose this operation to models/browsers.
+        Run state, output and usage must be obtained through Host Services.
+        """
+        if status not in {
+            "running",
+            "awaiting_approval",
+            "waiting_for_node",
+            "succeeded",
+            "failed",
+            "cancelled",
+            "interrupted",
+        }:
+            raise TeamsError("invalid_run_state", "无法识别执行状态", status=422)
+
+        def project(tx: Transaction) -> dict[str, Any]:
+            delivery = tx.get("delivery", delivery_id)
+            run = tx.get("team_run", delivery["teamRunId"])
+            member = self.delivery_member(tx, delivery)
+            if delivery.get("runId") and delivery["runId"] != run_id:
+                raise TeamsError("run_identity_conflict", "投递与执行引用不一致")
+            if delivery.get("_terminalState"):
+                return {"status": delivery["_terminalState"]}
+            delivery.update(runId=run_id, status="accepted", revision=delivery["revision"] + 1)
+            delivery["_runStatus"] = status
+            terminal = status in {"succeeded", "failed", "cancelled", "interrupted"}
+            if terminal:
+                delivery["_terminalState"] = status
+            self.publish(tx, "delivery", delivery_id, delivery)
+            self.update_active_clock(tx, run)
+            tx.put("team_run", run["teamRunId"], run)
+            source = {
+                "authorityRef": self.authority_ref,
+                "groupId": run["groupId"],
+                "memberId": member["memberId"],
+                "bindingRef": member["bindingRef"],
+                "providerRef": member["binding"]["providerRef"],
+                "sessionId": member["sessionId"],
+                "runId": run_id,
+            }
+            member.update(
+                activeRunId=None if terminal else run_id,
+                executionStatus="idle"
+                if terminal
+                else "waiting"
+                if status in {"awaiting_approval", "waiting_for_node"}
+                else "running",
+                revision=member["revision"] + 1,
+            )
+            current_member = self.run_member(tx, run["teamRunId"], delivery["memberId"])
+            if current_member["sessionId"] == member["sessionId"]:
+                member["revision"] = current_member["revision"] + 1
+                self.publish_run_member(tx, member)
+            if status == "running" and run["status"] == "planning":
+                run.update(status="running", revision=run["revision"] + 1, updatedAt=now())
+                self.publish(tx, "team_run", run["teamRunId"], run)
+            if delivery.get("_taskId"):
+                task = tx.get("task", delivery["_taskId"])
+                attempt = task["attempts"][-1]
+                if attempt["attemptId"] == delivery.get("_attemptId"):
+                    attempt["source"] = source
+                    if run_id not in attempt["_runIds"]:
+                        attempt["_runIds"].append(run_id)
+                    attempt.setdefault("startedAt", now())
+                    if terminal:
+                        attempt["endedAt"] = now()
+                        candidate = attempt.get("_candidate")
+                        if (
+                            status == "succeeded"
+                            and candidate
+                            and candidate["runId"] == run_id
+                            and not delivery.get("_fenced")
+                        ):
+                            attempt.update(
+                                result=candidate["result"], artifacts=candidate["artifacts"]
+                            )
+                            task["status"] = (
+                                "succeeded"
+                                if task["_acceptancePolicy"] == "result"
+                                else "awaiting_acceptance"
+                            )
+                        else:
+                            task["status"] = "cancelled" if status == "cancelled" else "failed"
+                            task["reason"] = (
+                                "执行结束但未提交可核验结果" if status == "succeeded" else status
+                            )
+                        attempt["status"] = task["status"]
+                    else:
+                        task["status"] = attempt["status"] = "running"
+                        if status == "awaiting_approval":
+                            task["reason"] = "等待人工审批"
+                        elif status == "waiting_for_node":
+                            task["reason"] = "等待设备上线"
+                        elif task.get("reason") in {"等待人工审批", "等待设备上线"}:
+                            task.pop("reason", None)
+                    task["revision"] += 1
+                    self.publish(tx, "task", task["taskId"], task)
+            if terminal:
+                run = tx.get("team_run", run["teamRunId"])
+                run["budget"]["tokensUsed"] += max(0, tokens)
+                if (
+                    member["memberId"] == run["leaderMemberId"]
+                    and not delivery.get("_fenced")
+                    and status in {"failed", "cancelled", "interrupted"}
+                    and run["status"] not in TERMINAL | {"cancel_requested"}
+                ):
+                    run.update(
+                        status="needs_attention", dispatchSuspended=True, reason=f"leader_{status}"
+                    )
+                if (
+                    run["budget"]["tokensUsed"] >= run["budget"]["maxTokens"]
+                    and run["status"] not in TERMINAL
+                    and run["status"] != "cancel_requested"
+                ):
+                    run.update(
+                        status="needs_attention",
+                        dispatchSuspended=True,
+                        reason="token_budget_exhausted",
+                    )
+                run.update(revision=run["revision"] + 1, updatedAt=now())
+                self.publish(tx, "team_run", run["teamRunId"], run)
+                if output.strip():
+                    message = {
+                        "messageId": "result_" + digest([delivery_id, run_id])[7:],
+                        "groupId": run["groupId"],
+                        "teamRunId": run["teamRunId"],
+                        "revision": 1,
+                        "createdAt": now(),
+                        "senderPrincipal": f"member:{member['memberId']}",
+                        "senderName": member["name"],
+                        "groupRole": member["role"],
+                        "memberId": member["memberId"],
+                        "parts": [{"kind": "text", "text": output}],
+                        "mentions": [],
+                        "intent": "result",
+                        "visibility": "public",
+                        "sourceRefs": [source],
+                    }
+                    self.publish(tx, "message", message["messageId"], message, created=True)
+                self._advance(tx, run)
+                if not delivery.get("_fenced"):
+                    self.finalize_if_ready(tx, run["teamRunId"])
+            return {"status": status}
+
+        return self.store.mutate(
+            f"projection:{delivery_id}",
+            event_id,
+            {"runId": run_id, "status": status, "output": output, "tokens": tokens},
+            project,
+        )
+
+    def finalize_if_ready(self, tx, run_id):
+        run = tx.get("team_run", run_id)
+        candidate = run.get("_finalCandidate")
+        if not candidate or run["status"] in TERMINAL | {"cancel_requested", "awaiting_acceptance"}:
+            return
+        deliveries = tx.list("delivery", run["groupId"], team_run_id=run_id)
+        if not any(
+            d.get("runId") == candidate["runId"]
+            and d.get("_terminalState") == "succeeded"
+            and not d.get("_fenced")
+            for d in deliveries
+        ):
+            return
+        if any(
+            d["status"] in {"pending", "uncertain"}
+            or (d["status"] == "accepted" and not d.get("_terminalState"))
+            for d in deliveries
+        ):
+            return
+        if any(
+            task["status"] != "succeeded"
+            for task in tx.list("task", run["groupId"], team_run_id=run_id)
+        ):
+            return
+        run.update(
+            status="awaiting_acceptance",
+            result=candidate["result"],
+            revision=run["revision"] + 1,
+            updatedAt=now(),
+        )
+        self.publish(tx, "team_run", run_id, run)
+
+    def finish_candidate(self, actor: Actor, result: str, key: str) -> dict[str, Any]:
+        if (
+            actor.kind != "member"
+            or not actor.group_id
+            or not actor.team_run_id
+            or not actor.run_id
+        ):
+            raise TeamsError("invocation_required", "需要 Leader 的受信执行上下文", status=403)
+        if not result.strip() or len(result) > 100_000:
+            raise TeamsError("invalid_result", "请提供有效的交付结果", status=422)
+
+        def finish(tx: Transaction) -> dict[str, Any]:
+            run = self._task_run(tx, actor.group_id, actor.team_run_id)
+            tasks = tx.list("task", actor.group_id, team_run_id=actor.team_run_id)
+            if any(task["status"] != "succeeded" for task in tasks):
+                raise TeamsError("tasks_unsettled", "仍有任务未通过验收，不能提交最终成果")
+            run["_finalCandidate"] = {"runId": actor.run_id, "result": result}
+            run["revision"] += 1
+            tx.put("team_run", run["teamRunId"], run)
+            return {"status": "candidate_received", "teamRunId": run["teamRunId"]}
+
+        return self.mutate(
+            actor,
+            actor.group_id,
+            key,
+            {"operation": "finish", "result": result},
+            finish,
+            leader=True,
+        )
+
+    def member_message(
+        self,
+        actor: Actor,
+        *,
+        content: str,
+        target_member_id: str | None,
+        wake: bool,
+        source_delivery_id: str,
+        key: str,
+    ) -> dict[str, Any]:
+        """Explicit member communication; prose mentions never trigger work."""
+        if actor.kind != "member" or not actor.group_id or not actor.team_run_id:
+            raise TeamsError("invocation_required", "需要受信的成员执行上下文", status=403)
+        if not content.strip() or len(content) > 100_000:
+            raise TeamsError("invalid_message", "消息不能为空或超过大小限制", status=422)
+
+        def send(tx: Transaction) -> dict[str, Any]:
+            run = self._task_run(tx, actor.group_id, actor.team_run_id)
+            member = self.run_member(tx, actor.team_run_id, actor.member_id)
+            source_delivery = tx.get("delivery", source_delivery_id)
+            if (
+                source_delivery["memberId"] != actor.member_id
+                or source_delivery["teamRunId"] != actor.team_run_id
+                or source_delivery.get("runId") not in {None, actor.run_id}
+            ):
+                raise TeamsError("invocation_scope_mismatch", "消息来源执行不一致", status=403)
+            if target_member_id:
+                self.run_member(tx, actor.team_run_id, target_member_id)
+            if wake:
+                if not target_member_id:
+                    raise TeamsError("member_required", "唤醒需要明确接收成员", status=422)
+                if (
+                    member["role"] != "leader"
+                    and target_member_id != run["leaderMemberId"]
+                    and not self.run_policy(tx, run)["peerWake"]
+                ):
+                    raise TeamsError(
+                        "peer_wake_forbidden", "当前策略只允许向 Leader 请求后续处理", status=403
+                    )
+            message = {
+                "messageId": new_id("gm"),
+                "groupId": actor.group_id,
+                "teamRunId": actor.team_run_id,
+                "revision": 1,
+                "createdAt": now(),
+                "senderPrincipal": f"member:{actor.member_id}",
+                "senderName": member["name"],
+                "groupRole": member["role"],
+                "memberId": actor.member_id,
+                "parts": [{"kind": "text", "text": content.strip()}],
+                "mentions": [target_member_id] if target_member_id else [],
+                "intent": "directed" if wake else "progress",
+                "visibility": "public",
+            }
+            self.publish(tx, "message", message["messageId"], message, created=True)
+            delivery = None
+            if wake:
+                previous = tx.get("trigger", source_delivery["_decisionId"])
+                delivery = self.dispatch(
+                    tx,
+                    run,
+                    message,
+                    target_member_id,
+                    source=f"{actor.run_id}:{key}",
+                    purpose="member_request",
+                    hop=previous["hop"] + 1,
+                    root=previous["rootTriggerId"],
+                )
+            return {
+                "status": "accepted" if not wake or delivery else "blocked",
+                "messageId": message["messageId"],
+                "groupId": actor.group_id,
+                "deliveryId": delivery["deliveryId"] if delivery else None,
+            }
+
+        return self.mutate(
+            actor,
+            actor.group_id,
+            key,
+            {
+                "operation": "member_message",
+                "content": content,
+                "target": target_member_id,
+                "wake": wake,
+                "sourceDeliveryId": source_delivery_id,
+            },
+            send,
+        )
+
+    def accept_run(
+        self,
+        actor: Actor,
+        group_id: str,
+        run_id: str,
+        expected_revision: int,
+        key: str,
+        *,
+        accepted: bool | None = None,
+        action: str | None = None,
+        reason: str | None = None,
+    ) -> dict[str, Any]:
+        chosen = action or ("accept" if accepted else "reject")
+        if chosen not in {"accept", "reject", "request_changes"}:
+            raise TeamsError("invalid_acceptance_action", "不支持该验收操作", status=422)
+        if action is None and accepted is None:
+            raise TeamsError("acceptance_action_required", "请选择验收操作", status=422)
+        if action and accepted is not None and (action != ("accept" if accepted else "reject")):
+            raise TeamsError("acceptance_action_conflict", "验收操作不能互相冲突", status=422)
+        if chosen == "request_changes" and not (reason or "").strip():
+            raise TeamsError("change_reason_required", "请说明需要修改的内容", status=422)
+        if reason and len(reason) > 2000:
+            raise TeamsError("invalid_reason", "修改理由不能超过2000字", status=422)
+
+        def accept(tx: Transaction) -> dict[str, Any]:
+            run = self._task_run(tx, group_id, run_id)
+            if run["revision"] != expected_revision:
+                raise TeamsError("revision_conflict", "本轮状态已更新，请刷新后重试")
+            if run["status"] != "awaiting_acceptance":
+                raise TeamsError("not_awaiting_acceptance", "本轮尚未提交最终成果")
+            if any(
+                delivery["status"] in {"pending", "uncertain"}
+                or (delivery["status"] == "accepted" and not delivery.get("_terminalState"))
+                for delivery in tx.list("delivery", group_id, team_run_id=run_id)
+            ):
+                raise TeamsError("execution_in_flight", "仍有执行未结束，请等待真实终态")
+            if chosen == "request_changes":
+                run.setdefault("resultHistory", []).append(
+                    {
+                        "result": run.get("result", ""),
+                        "reason": reason.strip(),
+                        "revision": run["revision"],
+                        "requestedAt": now(),
+                    }
+                )
+                run.pop("_finalCandidate", None)
+                run.pop("result", None)
+            run.update(
+                status="running"
+                if chosen == "request_changes"
+                else "succeeded"
+                if chosen == "accept"
+                else "failed",
+                revision=run["revision"] + 1,
+                updatedAt=now(),
+                _acceptedBy=actor.subject,
+                _acceptedAt=now(),
+            )
+            if reason:
+                run["reason"] = reason.strip()
+            self.publish(tx, "team_run", run_id, run)
+            if chosen == "request_changes":
+                message = {
+                    "messageId": new_id("gm"),
+                    "groupId": group_id,
+                    "teamRunId": run_id,
+                    "revision": 1,
+                    "createdAt": now(),
+                    "senderPrincipal": actor.subject,
+                    "senderName": "我",
+                    "groupRole": "owner",
+                    "parts": [
+                        {
+                            "kind": "text",
+                            "text": f"最终成果需要修改：{reason.strip()}\n请在本轮继续处理，"
+                            "保留已验收的有效结果，必要时创建修订子任务，完成后重新提交最终成果。",
+                        }
+                    ],
+                    "mentions": [run["leaderMemberId"]],
+                    "intent": "followup",
+                    "visibility": "public",
+                }
+                self.publish(tx, "message", message["messageId"], message, created=True)
+                self.dispatch(
+                    tx,
+                    run,
+                    message,
+                    run["leaderMemberId"],
+                    source=message["messageId"],
+                    purpose="revision",
+                )
+            return public(tx.get("team_run", run_id))
+
+        return self.mutate(
+            actor,
+            group_id,
+            key,
+            {
+                "operation": "accept_run",
+                "runId": run_id,
+                "expectedRevision": expected_revision,
+                "accepted": accepted,
+                "action": action,
+                "reason": reason,
+            },
+            accept,
+            owner=True,
+        )
+
+    def update_group(
+        self,
+        actor: Actor,
+        group_id: str,
+        expected_revision: int,
+        key: str,
+        *,
+        name: str | None = None,
+        leader_member_id: str | None = None,
+        remove_member_id: str | None = None,
+        task_acceptance: str | None = None,
+        peer_wake: bool | None = None,
+        archived: bool | None = None,
+        add_member: MemberInput | None = None,
+        rebind_member: MemberInput | None = None,
+        bindings: dict[str, dict[str, Any]] | None = None,
+    ) -> dict[str, Any]:
+        if name is not None and (not name.strip() or len(name.strip()) > 120):
+            raise TeamsError("invalid_name", "群名不能为空且不能超过120字", status=422)
+        if task_acceptance is not None and task_acceptance not in {"leader", "human", "result"}:
+            raise TeamsError("invalid_policy", "验收策略无效", status=422)
+        payload = {
+            "operation": "update_group",
+            "expectedRevision": expected_revision,
+            "name": name,
+            "leaderMemberId": leader_member_id,
+            "removeMemberId": remove_member_id,
+            "taskAcceptance": task_acceptance,
+            "peerWake": peer_wake,
+            "archived": archived,
+            "addMember": add_member.model_dump() if add_member else None,
+            "rebindMember": rebind_member.model_dump() if rebind_member else None,
+        }
+
+        def update(tx: Transaction) -> dict[str, Any]:
+            group = tx.get("group", group_id)
+            if group.get("legacySource"):
+                raise TeamsError(
+                    "legacy_history_read_only", "导入历史只读，请创建新团队", status=409
+                )
+            if group["revision"] != expected_revision:
+                raise TeamsError("revision_conflict", "团队配置已更新，请刷新后重试")
+            if name is not None:
+                group["name"] = name.strip()
+            for selected, adding in ((add_member, True), (rebind_member, False)):
+                if selected is None:
+                    continue
+                binding = (bindings or {}).get(selected.bindingRef)
+                if not binding or not binding.get("capabilities", {}).get("enqueue"):
+                    raise TeamsError("binding_unavailable", "所选Build不支持团队执行", status=422)
+                if adding:
+                    if (
+                        len([m for m in tx.list("member", group_id) if m["status"] == "active"])
+                        >= 8
+                    ):
+                        raise TeamsError("member_limit", "一个团队最多8位成员", status=422)
+                    if tx.get("member", member_key(group_id, selected.memberId), required=False):
+                        raise TeamsError(
+                            "member_id_exists", "成员标识已使用，请使用新标识", status=409
+                        )
+                    member = {
+                        "memberId": selected.memberId,
+                        "groupId": group_id,
+                        "role": "member",
+                        "revision": 1,
+                        "status": "active",
+                        "executionStatus": "idle",
+                    }
+                else:
+                    member = self._member(tx, group_id, selected.memberId)
+                    if member["role"] == "leader" and not binding["capabilities"].get("leader"):
+                        raise TeamsError(
+                            "leader_unavailable", "新Build不支持Leader策略", status=422
+                        )
+                    member.setdefault("_bindingHistory", []).append(
+                        {
+                            key: deepcopy(member[key])
+                            for key in ("bindingRef", "binding", "sessionId")
+                        }
+                    )
+                    member["revision"] += 1
+                member.update(
+                    name=selected.name,
+                    bindingRef=selected.bindingRef,
+                    binding=deepcopy(binding),
+                    sessionId=new_id("ses"),
+                    responsibility=selected.responsibility,
+                )
+                self.publish(tx, "member", member_key(group_id, selected.memberId), member)
+            if leader_member_id:
+                new_leader = self._member(tx, group_id, leader_member_id)
+                if not new_leader["binding"]["capabilities"].get("leader"):
+                    raise TeamsError("leader_unavailable", "该成员不支持协调团队", status=422)
+                for member in tx.list("member", group_id):
+                    role = "leader" if member["memberId"] == leader_member_id else "member"
+                    if member["role"] != role:
+                        member.update(role=role, revision=member["revision"] + 1)
+                        self.publish(tx, "member", member_key(group_id, member["memberId"]), member)
+                group["leaderMemberId"] = leader_member_id
+            if remove_member_id:
+                if group["leaderMemberId"] == remove_member_id:
+                    raise TeamsError(
+                        "leader_required", "移除 Leader 前请指定新的 Leader", status=422
+                    )
+                member = self._member(tx, group_id, remove_member_id)
+                member.update(status="removed", revision=member["revision"] + 1)
+                self.publish(tx, "member", member_key(group_id, remove_member_id), member)
+            if task_acceptance is not None:
+                group["_policy"]["taskAcceptance"] = task_acceptance
+            if peer_wake is not None:
+                group["_policy"]["peerWake"] = peer_wake
+            if archived is not None:
+                group["status"] = "archived" if archived else "active"
+            group.update(revision=group["revision"] + 1, updatedAt=now())
+            self.publish(tx, "group", group_id, group)
+            tx.put(
+                "group_revision",
+                f"{group_id}:{group['revision']}",
+                {**group, "members": tx.list("member", group_id)},
+            )
+            return self._snapshot(tx, group_id)
+
+        return self.mutate(actor, group_id, key, payload, update, owner=True)
+
+    def mark_read(self, actor: Actor, group_id: str, watermark: int) -> dict[str, int]:
+        with self.store.transaction() as tx:
+            self.authorize(tx, actor, group_id, owner=True)
+            if watermark < 0 or watermark > tx.watermark(group_id):
+                raise TeamsError("invalid_cursor", "已读水位无效", status=422)
+            key = f"{group_id}:{actor.subject}"
+            previous = tx.get("read_watermark", key, required=False)
+            value = max(watermark, (previous or {}).get("watermark", 0))
+            tx.put("read_watermark", key, {"groupId": group_id, "watermark": value})
+            return {"watermark": value}
+
+    def execution(self, actor: Actor, group_id: str, run_id: str) -> dict[str, Any]:
+        with self.store.transaction() as tx:
+            self.authorize(tx, actor, group_id)
+            if actor.kind == "member" and run_id != actor.team_run_id:
+                raise TeamsError("forbidden", "不能读取其他轮次执行", status=403)
+            snapshot = self._snapshot(tx, group_id)
+            deliveries = tx.list("delivery", group_id, team_run_id=run_id)
+            children = tx.list("child_invocation", group_id, team_run_id=run_id)
+            members = {d["deliveryId"]: self.delivery_member(tx, d) for d in deliveries}
+            watermark = tx.watermark(group_id)
+        if not any(run["teamRunId"] == run_id for run in snapshot["teamRuns"]):
+            raise TeamsError("not_found", "执行轮次不存在", status=404)
+        nodes, edges = [], []
+        for task in snapshot["tasks"]:
+            if task["teamRunId"] != run_id:
+                continue
+            attempt = task["attempts"][-1] if task["attempts"] else None
+            node = {
+                "nodeId": task["taskId"],
+                "kind": "task",
+                "title": task["title"],
+                "status": task["status"],
+                "taskId": task["taskId"],
+                "memberId": task["ownerMemberId"],
+                "reason": task.get("reason"),
+            }
+            if attempt:
+                node.update(attemptId=attempt["attemptId"], source=attempt.get("source"))
+            nodes.append(node)
+            edges.extend(
+                {"source": dependency, "target": task["taskId"], "kind": "dependency"}
+                for dependency in task["dependencies"]
+            )
+        run_nodes = {}
+        for delivery in deliveries:
+            if not delivery.get("runId"):
+                continue
+            member = members[delivery["deliveryId"]]
+            binding = member
+            node_id = "run_" + digest([delivery["_sessionId"], delivery["runId"]])[7:]
+            run_nodes[(delivery["_sessionId"], delivery["runId"])] = node_id
+            node = {
+                "nodeId": node_id,
+                "kind": "run",
+                "title": member["name"],
+                "status": delivery.get("_terminalState") or delivery.get("_runStatus") or "queued",
+                "reason": "execution_uncertain_fenced"
+                if delivery.get("_fenced")
+                else delivery.get("reason"),
+                "memberId": member["memberId"],
+                "source": {
+                    "authorityRef": self.authority_ref,
+                    "groupId": group_id,
+                    "memberId": member["memberId"],
+                    "bindingRef": binding["bindingRef"],
+                    "providerRef": binding["binding"]["providerRef"],
+                    "sessionId": binding["sessionId"],
+                    "runId": delivery["runId"],
+                },
+            }
+            if delivery.get("_taskId"):
+                node.update(
+                    taskId=delivery["_taskId"],
+                    attemptId=delivery["_attemptId"],
+                    parentNodeId=delivery["_taskId"],
+                )
+                edges.append(
+                    {"source": delivery["_taskId"], "target": node_id, "kind": "invocation"}
+                )
+            nodes.append(node)
+        for child in children:
+            run_nodes[(child["source"]["sessionId"], child["_nativeRunId"])] = child["nodeId"]
+        for child in children:
+            node = public(child)
+            parent = run_nodes.get((child["source"]["sessionId"], child["_parentRunId"]))
+            if parent:
+                node["parentNodeId"] = parent
+                edges.append({"source": parent, "target": child["nodeId"], "kind": "invocation"})
+            nodes.append(node)
+        return {
+            "groupId": group_id,
+            "teamRunId": run_id,
+            "watermark": watermark,
+            "nodes": nodes,
+            "edges": edges,
+        }