Unify Rust court authority and make chat stop/end UX match mainstream agents.

Collapse extra_roots, mcp_*, and session grants into the host court; keep partial replies on stop, list only deliverable artifacts, and reuse a detached backend only when jwt_fp matches.

Co-authored-by: Cursor <cursoragent@cursor.com>

Author: wu1w

backend/agent/exit_reasons.py

@@ -11,60 +11,78 @@
 # code → (title, body, recovery_hint, severity)
 _EXIT_CATALOG: dict[str, tuple[str, str, str, str]] = {
     "budget_grace": (
-        "迭代预算已用尽（宽限终答）",
-        "本段迭代配额已满，系统已强制生成最终回复，未再调用工具。",
+        "本轮步数已用完",
+        "本轮步数已经用完，已根据现有结果给出答复。",
         "可提高 agent_max_iterations / 自动续段，或拆小任务后重试；"
         "进程若仍 suspended 可在控制台「恢复」。",
         "warn",
     ),
     "budget_exhausted": (
-        "迭代预算耗尽",
-        "运行在耗尽迭代配额后停止，未能进入宽限终答。",
+        "本轮步数已用完",
+        "本轮步数已经用完，没有再继续调用工具。",
         "提高迭代上限或缩小目标后重开会话；检查 Goal 自动续段设置。",
         "error",
     ),
     # Token process budget (charge_tokens) — MUST NOT reuse budget_exhausted title
     "kernel_token_budget_exhausted": (
-        "Token 预算耗尽",
-        "本进程 token 额度已用尽，运行中断（这不是迭代轮次上限）。",
+        "额度已用完",
+        "这一轮的用量额度已经用完，所以停在这里。",
         "主会话/CEO：系统会自动 chat_elastic top_up（上限 agent_chat_budget_hard_cap）；"
         "若仍中断多半已达天花板或 host 不可用。编制工单：crew_steward top_up / 提高 token_budget。",
         "error",
     ),
     "kernel_iteration_exhausted": (
-        "内核迭代预算耗尽",
-        "Rust policy 侧 iteration 预算已耗尽，已触发优雅收尾。",
+        "本轮步数已用完",
+        "本轮步数已经用完，已根据现有结果收束。",
         "在控制台查看 /kernel/policy/{process_id}；必要时 resume 或新开 run。",
         "warn",
     ),
     "kernel_budget_precheck": (
-        "Token 预算不足（事前中断）",
-        "进程 token 预算不足以支撑下一次 LLM 调用，已避免烧穿预算。",
+        "额度不足",
+        "剩余额度不够支撑下一轮，已提前停下以免空转。",
         "提高进程 token_budget / top_up，或收窄任务范围后重试。",
         "error",
     ),
     "kernel_gate_stop": (
-        "内核门闩停止",
-        "运行被 kernel 仲裁门（stop/挂起超时）中断。",
+        "运行已中断",
+        "这一轮被系统中断了。",
         "若进程为 suspended：POST /kernel/processes/{id}/resume 恢复；"
         "否则查看 decision_trail。",
         "info",
     ),
     "doom_loop": (
-        "工具空转熔断（doom loop）",
-        "同一工具与相近参数连续重复，系统已熔断并改为直接作答。",
+        "已停止重复操作",
+        "同样的步骤重复了多次，已改为直接作答。",
         "换参数/换工具，或人工指定下一步；可在权限中调整 doom_loop 策略。",
         "warn",
     ),
     "thrash": (
-        "工具空转熔断",
-        "检测到重复工具调用，已停止工具轮次。",
+        "已停止重复操作",
+        "检测到重复操作，已停止继续调用工具。",
         "根据已有结果作答，或换一种工具路径；避免同一命令连点。",
         "warn",
     ),
+    "empty_content_thrash": (
+        "没有生成回复",
+        "这一轮没有生成可见回复。直接发送下一条即可。",
+        "重试同一问题，或换一种说法。",
+        "warn",
+    ),
+    "max_tool_rounds": (
+        "本轮步数已用完",
+        "本轮步数已经用完，已根据现有结果给出答复。",
+        "发送「请继续」或缩小任务后重试。",
+        "warn",
+    ),
+    "max_segment_budget": (
+        "本轮步数已用完",
+        "本轮步数已经用完，已根据现有结果给出答复。",
+        "发送「请继续」可接着做。",
+        "warn",
+    ),
     "stopped_by_user": (
-        "用户停止",
-        "运行被用户手动停止。",
+        "已停止",
+        "已按你的操作停止生成。",
         "可重新发送消息继续；挂起进程可用 resume。",
         "info",
     ),
@@ -110,15 +128,44 @@
 }
 
 
+# Chat-channel next step (no API paths / process ids — matches Cursor/ChatGPT tone)
+_CHAT_NEXT: dict[str, str] = {
+    "budget_grace": "需要的话，发送「请继续」。",
+    "budget_exhausted": "发送「请继续」可以接着做。",
+    "kernel_iteration_exhausted": "发送「请继续」可以接着做。",
+    "kernel_token_budget_exhausted": "缩小范围后再试，或在设置里提高上限。",
+    "kernel_budget_precheck": "缩小任务后再试。",
+    "kernel_gate_stop": "发送新消息即可继续。",
+    "doom_loop": "若还要改，直接说下一步即可。",
+    "thrash": "若还要改，直接说下一步即可。",
+    "stopped_by_user": "",
+    "completed": "",
+    "empty_content_thrash": "",
+    "max_tool_rounds": "发送「请继续」可以接着做。",
+    "max_segment_budget": "发送「请继续」可以接着做。",
+    "goal_stalled": "目标一段时间没有推进。直接说下一步，或发送「请继续」。",
+}
+
+# Overlay when catalog body still names kernel/host internals
+_CHAT_BODY: dict[str, str] = {
+    "host_down": "系统暂时连不上，请稍后再试。",
+    "host_abi_mismatch": "系统组件需要更新后才能继续。",
+    "sandbox_missing": "当前环境无法安全执行这项操作。",
+    "intent_denied": "这次操作没有获得所需权限。",
+    "resource_denied": "用量达到上限，这一轮先停在这里。",
+    "empty_content_thrash": "这一轮没有生成可见回复。直接发送下一条即可。",
+}
+
+
 def describe_exit_reason(code: str | None) -> dict[str, Any]:
-    """Return structured UX payload for a loop exit code."""
+    """Return structured UX payload for a loop exit code (API / console)."""
     c = (code or "").strip() or "completed"
     if c in _EXIT_CATALOG:
         title, body, hint, sev = _EXIT_CATALOG[c]
     else:
         title, body, hint, sev = (
-            f"结束原因：{c}",
-            f"运行以代码 {c} 结束。",
+            "这一轮已结束",
+            "这一轮已经结束。",
             "查看 decision_trail 与 process meta；必要时 resume 或新开 run。",
             "info",
         )
@@ -135,18 +182,29 @@ def describe_exit_reason(code: str | None) -> dict[str, Any]:
     }
 
 
-def format_exit_user_message(code: str | None, *, process_id: str | None = None) -> str:
+def format_exit_user_message(
+    code: str | None,
+    *,
+    process_id: str | None = None,
+    for_operator: bool = False,
+) -> str:
+    """User-chat copy by default. Operator/console keeps API hints when requested."""
     d = describe_exit_reason(code)
-    pid = (process_id or "").strip()
-    lines = [
-        f"[{d['title']}]",
-        d["message"],
-    ]
-    if d.get("recovery_hint"):
-        lines.append(f"恢复建议：{d['recovery_hint']}")
-    if pid:
-        lines.append(f"process_id={pid} · 控制台可 resume / 查看策略与决策轨迹。")
-    return "\n".join(lines)
+    if for_operator:
+        pid = (process_id or "").strip()
+        lines = [f"[{d['title']}]", d["message"]]
+        if d.get("recovery_hint"):
+            lines.append(f"恢复建议：{d['recovery_hint']}")
+        if pid:
+            lines.append(f"process_id={pid} · 控制台可 resume / 查看策略与决策轨迹。")
+        return "\n".join(lines)
+    c = str(d.get("code") or "")
+    body = _CHAT_BODY.get(c) or str(d.get("message") or "").strip()
+    lines = [body]
+    nxt = _CHAT_NEXT.get(c, "")
+    if nxt:
+        lines.append(nxt)
+    return "\n".join(x for x in lines if x)
 
 
 __all__ = ["describe_exit_reason", "format_exit_user_message"]

backend/agent/grant_store.py

@@ -196,6 +196,7 @@ def _prune_expired_locked() -> int:
     # MCP 管理与运行时（动态 mcp_* 走 tool_matches 前缀规则）
     "manage_mcp": "manage_mcp",
     "manage_skill": "manage_skill",
+    "generate_ppt": "file_rw",
 }
 
 
@@ -229,6 +230,9 @@ def sync_catalog_from_kernel(kernel: Any | None = None) -> bool:
         TOOL_TO_CREW_CAP.setdefault("clarify", "crew_steward")
         TOOL_TO_CREW_CAP.setdefault("manage_mcp", "manage_mcp")
         TOOL_TO_CREW_CAP.setdefault("manage_skill", "manage_skill")
+        TOOL_TO_CREW_CAP.setdefault("generate_ppt", "file_rw")
+        TOOL_TO_CREW_CAP.setdefault("current_time", "current_time")
+        TOOL_TO_CREW_CAP.setdefault("result_load", "file_read")
         logger.info("grant_store catalog synced from rust (%s entries)", len(TOOL_TO_CREW_CAP))
         return True
     except Exception as e:
@@ -270,11 +274,65 @@ def allow_signature(tool: str, arguments: dict[str, Any] | None = None) -> str:
     return tool
 
 
+def _rust_grant_rpc(method: str, params: dict[str, Any], kernel: Any | None = None) -> dict[str, Any] | None:
+    """Best-effort live court map. Old hosts without these RPCs are ignored."""
+    try:
+        k = kernel
+        if k is None:
+            from backend.kernel_rust.client import is_rust_host_available
+
+            if not is_rust_host_available():
+                return None
+            from backend.kernel import get_kernel
+
+            k = get_kernel()
+        if not hasattr(k, "_call"):
+            return None
+        r = k._call(method, params)
+        return r if isinstance(r, dict) else None
+    except Exception as e:
+        logger.debug("rust grant rpc %s skip: %s", method, e)
+        return None
+
+
+def rehydrate_session_grants_to_kernel(kernel: Any | None = None) -> int:
+    """Push persisted grants into the live Rust SessionGrantStore."""
+    _ensure_loaded()
+    n = 0
+    with _grants_lock:
+        items = [(sid, sorted(sigs)) for sid, sigs in _session_grants.items() if sigs]
+    for sid, sigs in items:
+        r = _rust_grant_rpc(
+            "session_grant_add",
+            {"session_id": sid, "sigs": sigs},
+            kernel=kernel,
+        )
+        if r is not None:
+            n += 1
+    if n:
+        logger.info("session grants rehydrated to rust sessions=%s", n)
+    return n
+
+
 def has_session_grant(session_id: str | None, tool: str, arguments: dict[str, Any] | None = None) -> bool:
     if not session_id:
         return False
     _ensure_loaded()
     sid = str(session_id)
+    sig = allow_signature(tool, arguments)
+    rust = _rust_grant_rpc(
+        "session_grant_has",
+        {"session_id": sid, "sig": sig},
+    )
+    if isinstance(rust, dict) and rust.get("has") is True:
+        return True
+    if rust is not None and tool != sig:
+        whole = _rust_grant_rpc(
+            "session_grant_has",
+            {"session_id": sid, "sig": tool},
+        )
+        if isinstance(whole, dict) and whole.get("has") is True:
+            return True
     with _grants_lock:
         # Lazy TTL check
         ttl = _grant_ttl_seconds()
@@ -289,7 +347,7 @@ def has_session_grant(session_id: str | None, tool: str, arguments: dict[str, An
         if not grants:
             return False
         # 细粒度签名（command:rm）或整工具名（本员工/本会话放行 command）
-        if allow_signature(tool, arguments) in grants:
+        if sig in grants:
             return True
         if tool in grants:
             return True
@@ -321,6 +379,10 @@ def add_session_grant(
             bucket.add(tool)
         _session_grant_ts[sid] = time.time()
         _persist()
+    sigs = [sig]
+    if whole_tool and tool != sig:
+        sigs.append(tool)
+    _rust_grant_rpc("session_grant_add", {"session_id": sid, "sigs": sigs})
     logger.info(
         "grant session allow session=%s sig=%s whole=%s",
         sid[:8],
@@ -447,17 +509,26 @@ def clear_session_grants(session_id: str | None) -> None:
         _session_grant_ts.pop(sid, None)
         if had:
             _persist()
+    _rust_grant_rpc("session_grant_clear", {"session_id": str(session_id)})
 
 
 def prune_expired_session_grants() -> int:
     """Drop grants past TTL. Returns number removed."""
     _ensure_loaded()
     with _grants_lock:
+        dead = [
+            sid
+            for sid, ts in list(_session_grant_ts.items())
+            if _grant_ttl_seconds() > 0
+            and (time.time() - float(ts or 0)) > _grant_ttl_seconds()
+        ]
         n = _prune_expired_locked()
         if n:
             _persist()
             logger.info("session grants TTL pruned count=%s", n)
-        return n
+    for sid in dead:
+        _rust_grant_rpc("session_grant_clear", {"session_id": sid})
+    return n
 
 
 def prune_orphan_session_grants(live_session_ids: set[str] | list[str] | None) -> int:
@@ -478,7 +549,9 @@ def prune_orphan_session_grants(live_session_ids: set[str] | list[str] | None) -
                 len(dead),
                 len(_session_grants),
             )
-        return len(dead)
+    for sid in dead:
+        _rust_grant_rpc("session_grant_clear", {"session_id": sid})
+    return len(dead)
 
 
 async def prune_session_grants_startup() -> dict[str, int]:
@@ -560,3 +633,5 @@ def reset_for_tests() -> None:
                 p.unlink()
         except OSError:
             pass
+    # Do not session_grant_clear the live host: pytest may share a desktop
+    # tevarn-kernel-host. Court tests clear their own session ids explicitly.