Soften short-intent gates and add tool-result refs plus MCP wire invalidate.

Prefer report tools without hard-hiding CLI, keep soft CLI budgets, store compact+result_ref for the model, and clear frozen wire after MCP sync.

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

interfaces/admin/routes.py

@@ -3860,6 +3860,12 @@ def build_admin_router() -> APIRouter:
                 status="ok",
                 detail={"synced_tools": len(norm), "compat_mode": "bailian_webparser"},
             )
+            try:
+                from runtime.direct_loop import invalidate_tool_wire_cache
+
+                invalidate_tool_wire_cache(reason=f"mcp_admin_sync:{server_id}")
+            except Exception:
+                pass
             return {"ok": True, "server_id": server_id, "tools": norm, "compat_mode": "bailian_webparser"}
         rt = mcp_runtime_for_row(row, store=store)
         try:
@@ -3883,6 +3889,12 @@ def build_admin_router() -> APIRouter:
                 )
             store.replace_mcp_server_tools(server_id=server_id, tools=norm)
             store.set_mcp_server_health(server_id=server_id, status="ok", detail={"synced_tools": len(norm)})
+            try:
+                from runtime.direct_loop import invalidate_tool_wire_cache
+
+                invalidate_tool_wire_cache(reason=f"mcp_admin_sync:{server_id}")
+            except Exception:
+                pass
             return {"ok": True, "server_id": server_id, "tools": norm}
         finally:
             rt.stop()

runtime/application/gateway/ops_short_intent.py

@@ -8,42 +8,50 @@ from typing import Any
 _BOT_MENTION_RE = re.compile(r"@\S+")
 
 # intent -> (en hint, zh hint)
+# Prefer report path strongly; keep CLI/inventory available with soft budget (not hard-hidden).
 _HINTS: dict[str, tuple[str, str]] = {
     "excel_export": (
-        "[Ops short-intent: Excel export. Call ume_alarm_xlsx_report(..., deliverable=true) now. "
-        "write_xlsx / inventory / CLI / run_command are hidden this turn — do not build xlsx via shell.]",
-        "[短指令：导出 Excel。立即 ume_alarm_xlsx_report(..., deliverable=true)。"
-        "本轮已隐藏 write_xlsx/清单/CLI/run_command，禁止用 shell 搓表。]",
+        "[Ops short-intent: Excel export. Prefer ume_alarm_xlsx_report(..., deliverable=true) now. "
+        "write_xlsx / run_command are hidden this turn. "
+        "CLI/inventory stay available with a soft budget for device checks — prefer one hetero targets batch, not single-NE loops.]",
+        "[短指令：导出 Excel。优先 ume_alarm_xlsx_report(..., deliverable=true)。"
+        "本轮已隐藏 write_xlsx/run_command。"
+        "CLI/清单仍可用（soft budget）：核实设备优先一次 hetero targets batch，勿单台循环。]",
     ),
     "license": (
-        "[Ops short-intent: license/capacity. Call ume_alarm_xlsx_report(mode=list, keyword=license, deliverable=true) "
-        "or aggregateUmeAlarms/queryUmeAlarmsRaw; write_xlsx/CLI/inventory are hidden this turn.]",
-        "[短指令：License/容量。立即 ume_alarm_xlsx_report(mode=list, keyword=license, deliverable=true) "
-        "或 aggregate/queryUmeAlarmsRaw；本轮已隐藏 write_xlsx/CLI/清单。]",
+        "[Ops short-intent: license/capacity. Prefer ume_alarm_xlsx_report(mode=list, keyword=license, deliverable=true) "
+        "or aggregateUmeAlarms/queryUmeAlarmsRaw. write_xlsx/run_command hidden; "
+        "CLI available with soft budget for confirmation (one targets batch preferred).]",
+        "[短指令：License/容量。优先 ume_alarm_xlsx_report(mode=list, keyword=license, deliverable=true) "
+        "或 aggregate/queryUmeAlarmsRaw；write_xlsx/run_command 已隐藏；"
+        "CLI 可用（soft budget），确认阶段优先一次 targets batch。]",
     ),
     "congestion": (
-        "[Ops short-intent: bandwidth congestion. Call ume_alarm_xlsx_report(mode=list, deliverable=true) "
-        "or aggregateUmeAlarms/queryUmeAlarmsRaw; write_xlsx/CLI/inventory/sql are hidden this turn.]",
-        "[短指令：带宽拥塞。立即 ume_alarm_xlsx_report(mode=list, deliverable=true) 或 aggregate/query；"
-        "本轮已隐藏 write_xlsx/CLI/清单/sql。]",
+        "[Ops short-intent: bandwidth congestion. Prefer ume_alarm_xlsx_report(mode=list, deliverable=true) "
+        "or aggregateUmeAlarms/queryUmeAlarmsRaw. write_xlsx/run_command hidden; "
+        "CLI available with soft budget (one targets batch preferred).]",
+        "[短指令：带宽拥塞。优先 ume_alarm_xlsx_report(mode=list, deliverable=true) 或 aggregate/query；"
+        "write_xlsx/run_command 已隐藏；CLI 可用（soft budget），优先一次 targets batch。]",
     ),
     "fiber_cut": (
-        "[Ops short-intent: fiber/LOS. Call ume_alarm_xlsx_report(mode=fiber_cut, deliverable=true) now. "
-        "write_xlsx / inventory / CLI are hidden this turn — do not try listCliTargets/execManagedNe/write_xlsx.]",
-        "[短指令：断纤/LOS。立即 ume_alarm_xlsx_report(mode=fiber_cut, deliverable=true)。"
-        "本轮已隐藏 write_xlsx/清单/CLI，勿调用 listCliTargets/execManagedNe/write_xlsx。]",
+        "[Ops short-intent: fiber/LOS. Prefer ume_alarm_xlsx_report(mode=fiber_cut, deliverable=true) now. "
+        "write_xlsx/run_command hidden. CLI/inventory available with soft budget — "
+        "use at most one hetero targets batch for device confirmation, not listCliTargets→single-NE loops.]",
+        "[短指令：断纤/LOS。优先 ume_alarm_xlsx_report(mode=fiber_cut, deliverable=true)。"
+        "write_xlsx/run_command 已隐藏。CLI/清单可用（soft budget）："
+        "设备核实最多一次 hetero targets batch，勿 listCliTargets→单台循环。]",
     ),
     "offline": (
-        "[Ops short-intent: offline NE. Call ume_alarm_xlsx_report(mode=offline, deliverable=true) now. "
-        "write_xlsx / inventory / CLI are hidden this turn.]",
-        "[短指令：离线网元。立即 ume_alarm_xlsx_report(mode=offline, deliverable=true)。"
-        "本轮已隐藏 write_xlsx/清单/CLI。]",
+        "[Ops short-intent: offline NE. Prefer ume_alarm_xlsx_report(mode=offline, deliverable=true) now. "
+        "write_xlsx/run_command hidden; CLI/inventory with soft budget (one targets batch preferred).]",
+        "[短指令：离线网元。优先 ume_alarm_xlsx_report(mode=offline, deliverable=true)。"
+        "write_xlsx/run_command 已隐藏；CLI/清单 soft budget（优先一次 targets batch）。]",
     ),
     "alarm_tally": (
-        "[Ops short-intent: alarm tally/top. Call ume_alarm_xlsx_report(mode=aggregate_by_host, deliverable=true) "
-        "or aggregateUmeAlarms; write_xlsx / inventory / CLI are hidden this turn.]",
-        "[短指令：告警统计/Top。立即 ume_alarm_xlsx_report(mode=aggregate_by_host, deliverable=true) "
-        "或 aggregateUmeAlarms；本轮已隐藏 write_xlsx/清单/CLI。]",
+        "[Ops short-intent: alarm tally/top. Prefer ume_alarm_xlsx_report(mode=aggregate_by_host, deliverable=true) "
+        "or aggregateUmeAlarms. write_xlsx/run_command hidden; CLI soft budget for spot checks.]",
+        "[短指令：告警统计/Top。优先 ume_alarm_xlsx_report(mode=aggregate_by_host, deliverable=true) "
+        "或 aggregateUmeAlarms；write_xlsx/run_command 已隐藏；CLI soft budget 仅作抽检。]",
     ),
     "continue": (
         "[Ops short-intent: continue/confirm. Resume the unfinished prior task immediately; "
@@ -52,29 +60,21 @@ _HINTS: dict[str, tuple[str, str]] = {
     ),
 }
 
-# Report-style short intents: hide inventory/CLI loops that dominate WA tool spam.
+# Report-style short intents: soft-nudge toward alarm/xlsx; hide only shell/xlsx DIY.
 _REPORT_TOOL_FILTER_INTENTS = frozenset(
     {"fiber_cut", "offline", "alarm_tally", "excel_export", "license", "congestion"}
 )
 
-# Match bare tool names and mcp__netx__* / legacy netx_* aliases.
+# Soft CLI budgets during report short-intents (prefer report; allow limited device checks).
+_SHORT_INTENT_CLI_SOFT_SINGLE = 2
+_SHORT_INTENT_CLI_SOFT_FAIL = 2
+_SHORT_INTENT_CLI_SOFT_BATCH = 1
+
+# Only hard-hide DIY spreadsheet / shell paths — keep inventory + execManagedNe visible.
 _SUPPRESSED_TOOL_NAMES = frozenset(
     {
-        "listclitargets",
-        "listmanagedne",
-        "getmanagedne",
-        "execmanagedne",
-        "queryumeneinventory",
-        "getumene",
-        "findtopologypaths",
-        "sqlqueryume",
         "run_command",
         "write_xlsx",
-        "netx_list_managed_ne",
-        "netx_get_managed_ne",
-        "netx_exec_managed_ne",
-        "netx_sql_query_ume",
-        "netx_list_cli_targets",
     }
 )
 
@@ -93,6 +93,17 @@ def ops_short_intent_should_filter_tools(intent: str | None) -> bool:
     return str(intent or "").strip() in _REPORT_TOOL_FILTER_INTENTS
 
 
+def ops_short_intent_cli_soft_budgets(intent: str | None) -> dict[str, int] | None:
+    """Tighter execManagedNe budgets for report short-intents; None = use global defaults."""
+    if not ops_short_intent_should_filter_tools(intent):
+        return None
+    return {
+        "single": int(_SHORT_INTENT_CLI_SOFT_SINGLE),
+        "fail": int(_SHORT_INTENT_CLI_SOFT_FAIL),
+        "batch": int(_SHORT_INTENT_CLI_SOFT_BATCH),
+    }
+
+
 def is_ops_short_intent_suppressed_tool(tool_name: str, *, intent: str | None) -> bool:
     if not ops_short_intent_should_filter_tools(intent):
         return False
@@ -101,7 +112,7 @@ def is_ops_short_intent_suppressed_tool(tool_name: str, *, intent: str | None) -
 
 
 def filter_tool_specs_for_ops_short_intent(tools: list[Any], *, intent: str | None) -> list[Any]:
-    """Drop inventory/CLI tools for report-style short intents (keep alarm/xlsx path)."""
+    """Drop DIY xlsx/shell tools for report-style short intents (keep alarm path + CLI)."""
     if not ops_short_intent_should_filter_tools(intent):
         return list(tools or [])
     out: list[Any] = []
@@ -188,5 +199,6 @@ __all__ = [
     "is_ops_short_intent_suppressed_tool",
     "maybe_ops_short_intent_system_hint",
     "normalize_ops_user_text",
+    "ops_short_intent_cli_soft_budgets",
     "ops_short_intent_should_filter_tools",
 ]

runtime/chat/exec_managed_ne_guard.py

@@ -128,6 +128,8 @@ def budget_block_payload(
     single_budget: int,
     fail_used: int = 0,
     fail_budget: int = 0,
+    batch_used: int = 0,
+    batch_budget: int = 0,
 ) -> dict[str, Any]:
     en = str(lang or "").strip().lower().startswith("en")
     if reason == "fail_budget":
@@ -136,14 +138,24 @@ def budget_block_payload(
             f"execManagedNe already failed {fail_used}/{fail_budget} times this turn. "
             "Stop one-NE loops; use ONE execManagedNe batch: "
             "ne_ids|ume_ne_ids + shared commands, or targets=[{ume_ne_id, commands},…] when commands differ — "
-            "or summarize reachable failures; do not keep probing."
+            "or summarize reachable failures / switch to ume_alarm_xlsx_report; do not keep probing."
             if en
             else f"本轮 execManagedNe 已失败 {fail_used}/{fail_budget} 次。"
             "停止单台循环；改用一次 batch："
             "同命令用 ne_ids/ume_ne_ids，每台命令不同用 targets=[{ume_ne_id, commands},…]；"
-            "或汇总可达性失败，勿继续盲探。"
+            "或汇总可达性失败 / 改走 ume_alarm_xlsx_report，勿继续盲探。"
         )
         err = "cli_fail_budget_exceeded"
+    elif reason == "batch_budget":
+        code = "cli_batch_budget_exceeded"
+        hint = (
+            f"execManagedNe batch soft budget exhausted ({batch_used}/{batch_budget} this turn). "
+            "Prefer ume_alarm_xlsx_report / alarm aggregate for the answer; do not start another CLI batch."
+            if en
+            else f"本轮 execManagedNe batch soft budget 已用尽（{batch_used}/{batch_budget}）。"
+            "优先 ume_alarm_xlsx_report / 告警聚合收口，勿再开新 CLI batch。"
+        )
+        err = "cli_batch_budget_exceeded"
     else:
         code = "cli_call_budget_exceeded"
         hint = (
@@ -181,6 +193,8 @@ def budget_block_payload(
         "single_budget": int(single_budget),
         "fail_used": int(fail_used),
         "fail_budget": int(fail_budget),
+        "batch_used": int(batch_used),
+        "batch_budget": int(batch_budget),
     }
 
 

runtime/chat/history_tool_result_compact.py

@@ -46,6 +46,7 @@ def _guard_tool_result_text_for_history(
     cap_chars: int,
     image_cap_chars: int,
     video_cap_chars: int,
+    session_id: str = "",
 ) -> tuple[str, bool]:
     """Return (new_raw, changed) following the same strategy as context replay guard."""
     text = str(raw or "")
@@ -110,6 +111,33 @@ def _guard_tool_result_text_for_history(
     if len(text) <= int(cap_chars):
         return text, False
 
+    existing_ref = ""
+    if isinstance(obj, dict):
+        existing_ref = str(obj.get("result_ref") or "").strip()
+    if not existing_ref:
+        try:
+            from runtime.chat.tool_result_store import save_tool_result_blob
+
+            sid = str(session_id or getattr(store, "_history_compact_session_id", "") or "").strip() or "_history"
+            parsed_full = obj if isinstance(obj, dict) else None
+            if parsed_full is None:
+                try:
+                    p = json.loads(text)
+                    parsed_full = p if isinstance(p, dict) else {"ok": None, "raw": text}
+                except Exception:
+                    parsed_full = {"ok": None, "raw": text[: min(len(text), 200_000)]}
+            existing_ref = str(
+                save_tool_result_blob(
+                    session_id=sid,
+                    tool_call_id=f"history-{hash(text) & 0xFFFFFFFF:08x}",
+                    result=parsed_full,
+                    force=True,
+                )
+                or ""
+            )
+        except Exception:
+            existing_ref = ""
+
     preview = text[: max(1, min(4000, int(cap_chars) - 400))] + "\n...<tool_result_guard_truncated>"
     guarded_obj = {
         "ok": bool(ok) if ok is not None else None,
@@ -121,13 +149,15 @@ def _guard_tool_result_text_for_history(
         "preview": preview,
         "hint": (
             "Tool output was too large for safe context replay; it was truncated for history storage. "
-            "Use narrower queries (e.g., smaller glob/max_results) or adjust AIA_TOOL_LLM_MESSAGE_MAX_CHARS. / "
-            "工具输出过大，已压缩写回历史；请缩小范围或配置 AIA_TOOL_LLM_MESSAGE_MAX_CHARS。"
+            "Use fetch_tool_result(result_ref=...) when present, or narrower queries. / "
+            "工具输出过大，已压缩写回历史；有 result_ref 时用 fetch_tool_result 取全文。"
         ),
     }
+    if existing_ref:
+        guarded_obj["result_ref"] = existing_ref
+        guarded_obj["fetch_tool"] = "fetch_tool_result"
     return _json_dumps_safe(guarded_obj), True
 
-
 def compact_tool_results_in_session_history(
     *,
     store: Any,
@@ -150,6 +180,10 @@ def compact_tool_results_in_session_history(
     max_seen = 0
 
     # We only need to scan tool messages; fetching ids+content is enough.
+    try:
+        store._history_compact_session_id = sid  # noqa: SLF001
+    except Exception:
+        pass
     with store._connect() as conn:  # noqa: SLF001
         cur = conn.execute(
             "select id, content from chat_message where session_id=? and role='tool' "
@@ -168,6 +202,7 @@ def compact_tool_results_in_session_history(
                 cap_chars=cap,
                 image_cap_chars=image_cap,
                 video_cap_chars=video_cap,
+                session_id=sid,
             )
             # Defensive fallback: if content is still over cap but guard didn't report change,
             # force a minimal guard so polluted history can always be compacted.

runtime/chat/tool_result_store.py

@@ -0,0 +1,167 @@
+"""Persist full tool results for later fetch_tool_result while LLM sees compact payloads."""
+
+from __future__ import annotations
+
+import hashlib
+import json
+import re
+import time
+from pathlib import Path
+from typing import Any
+
+from svc.config.paths import attachments_dir
+
+_REF_RE = re.compile(r"^tr:[0-9a-f]{32}$")
+_SAVE_MIN_CHARS = 4_000
+_FETCH_DEFAULT_MAX_CHARS = 120_000
+
+
+def _root() -> Path:
+    p = (attachments_dir() / "tool_results").resolve()
+    p.mkdir(parents=True, exist_ok=True)
+    return p
+
+
+def _session_dir(session_id: str) -> Path:
+    sid = str(session_id or "").strip() or "_anon"
+    digest = hashlib.sha256(sid.encode("utf-8", errors="ignore")).hexdigest()[:24]
+    d = _root() / digest
+    d.mkdir(parents=True, exist_ok=True)
+    return d
+
+
+def _json_dumps(obj: Any) -> str:
+    return json.dumps(obj, ensure_ascii=False, default=str)
+
+
+def _json_size(obj: Any) -> int:
+    try:
+        return len(_json_dumps(obj))
+    except Exception:
+        return 0
+
+
+def make_result_ref(*, session_id: str, tool_call_id: str, payload: Any) -> str:
+    raw = f"{session_id}|{tool_call_id}|{_json_size(payload)}|{time.time_ns()}"
+    return "tr:" + hashlib.sha256(raw.encode("utf-8", errors="ignore")).hexdigest()[:32]
+
+
+def is_tool_result_ref(value: str) -> bool:
+    return bool(_REF_RE.fullmatch(str(value or "").strip().lower()))
+
+
+def save_tool_result_blob(
+    *,
+    session_id: str,
+    tool_call_id: str,
+    result: Any,
+    force: bool = False,
+) -> str | None:
+    """Store full tool JSON when large enough (or force=True). Returns result_ref or None."""
+    sid = str(session_id or "").strip()
+    if not sid:
+        return None
+    if not isinstance(result, dict):
+        return None
+    size = _json_size(result)
+    if (not force) and size < int(_SAVE_MIN_CHARS):
+        return None
+    ref = make_result_ref(session_id=sid, tool_call_id=str(tool_call_id or ""), payload=result)
+    path = _session_dir(sid) / f"{ref[3:]}.json"
+    meta = {
+        "result_ref": ref,
+        "session_id": sid,
+        "tool_call_id": str(tool_call_id or ""),
+        "chars": int(size),
+        "saved_at_ms": int(time.time() * 1000),
+    }
+    path.write_text(_json_dumps({"meta": meta, "result": result}), encoding="utf-8")
+    return ref
+
+
+def load_tool_result_blob(
+    result_ref: str,
+    *,
+    session_id: str,
+    max_chars: int | None = None,
+) -> dict[str, Any]:
+    ref = str(result_ref or "").strip().lower()
+    sid = str(session_id or "").strip()
+    if not is_tool_result_ref(ref):
+        return {"ok": False, "error_code": "invalid_result_ref", "error": "invalid_result_ref"}
+    if not sid:
+        return {"ok": False, "error_code": "session_required", "error": "session_required"}
+    path = _session_dir(sid) / f"{ref[3:]}.json"
+    if not path.is_file():
+        # Fallback: scan root for orphaned refs (session hash mismatch / migrate).
+        found = None
+        for child in _root().glob(f"*/{ref[3:]}.json"):
+            found = child
+            break
+        if found is None:
+            return {"ok": False, "error_code": "result_ref_not_found", "error": "result_ref_not_found", "result_ref": ref}
+        path = found
+    try:
+        data = json.loads(path.read_text(encoding="utf-8"))
+    except Exception as exc:
+        return {"ok": False, "error_code": "result_ref_read_failed", "error": f"{type(exc).__name__}: {exc}"}
+    meta = data.get("meta") if isinstance(data, dict) else None
+    if isinstance(meta, dict):
+        owner = str(meta.get("session_id") or "").strip()
+        if owner and owner != sid:
+            return {"ok": False, "error_code": "result_ref_session_mismatch", "error": "result_ref_session_mismatch"}
+    result = data.get("result") if isinstance(data, dict) else None
+    if not isinstance(result, dict):
+        return {"ok": False, "error_code": "result_ref_invalid_payload", "error": "result_ref_invalid_payload"}
+    cap = int(_FETCH_DEFAULT_MAX_CHARS if max_chars is None else max_chars)
+    cap = max(4_000, min(cap, 500_000))
+    body = _json_dumps(result)
+    if len(body) <= cap:
+        out = dict(result)
+        out["result_ref"] = ref
+        out["_fetched_full"] = True
+        return {"ok": True, "result_ref": ref, "result": out, "chars": len(body), "truncated": False}
+    from runtime.chat.tool_runtime import truncate_tool_result_for_llm_messages
+
+    slim = truncate_tool_result_for_llm_messages(result, max_chars=cap)
+    if isinstance(slim, dict):
+        slim = dict(slim)
+        slim["result_ref"] = ref
+        slim["_fetched_truncated"] = True
+        slim["hint"] = (
+            str(slim.get("hint") or "")
+            + " Full blob still on disk; narrow the tool query or raise max_chars on fetch_tool_result."
+        ).strip()
+    return {
+        "ok": True,
+        "result_ref": ref,
+        "result": slim,
+        "chars": len(body),
+        "truncated": True,
+        "fetch_cap_chars": cap,
+    }
+
+
+def attach_result_ref(payload: dict[str, Any], *, result_ref: str | None) -> dict[str, Any]:
+    if not result_ref or not isinstance(payload, dict):
+        return payload
+    out = dict(payload)
+    out["result_ref"] = str(result_ref)
+    if out.get("_truncated_for_llm") or out.get("_tool_result_guarded") or out.get("_history_compacted"):
+        out["fetch_tool"] = "fetch_tool_result"
+        hint = str(out.get("hint") or "").strip()
+        extra = (
+            f"Full tool output stored as result_ref={result_ref}. "
+            f"Call fetch_tool_result(result_ref=\"{result_ref}\") when you need details."
+        )
+        out["hint"] = f"{hint} {extra}".strip() if hint else extra
+    return out
+
+
+__all__ = [
+    "attach_result_ref",
+    "is_tool_result_ref",
+    "load_tool_result_blob",
+    "make_result_ref",
+    "save_tool_result_blob",
+]

runtime/chat/tool_runtime.py

@@ -1095,15 +1095,27 @@ class ToolExecutor:
             is_exec_managed_ne_tool,
             load_turn_exec_managed_ne_stats,
         )
+        from runtime.application.gateway.ops_short_intent import ops_short_intent_cli_soft_budgets
 
-        prior_single_exec, _prior_batch_exec, prior_exec_fails = load_turn_exec_managed_ne_stats(
+        prior_single_exec, prior_batch_exec, prior_exec_fails = load_turn_exec_managed_ne_stats(
             ctx.store,
             session_id=ctx.session_id,
             turn_uuid=str(ctx.turn_uuid or ""),
         )
         single_exec_budget = exec_managed_ne_single_budget()
         fail_exec_budget = exec_managed_ne_fail_budget()
+        batch_exec_budget = 0  # 0 = unlimited (global default)
+        try:
+            md_budget = ctx.inbound_metadata if isinstance(ctx.inbound_metadata, dict) else {}
+            soft = ops_short_intent_cli_soft_budgets(str(md_budget.get("ops_short_intent") or "").strip() or None)
+            if soft:
+                single_exec_budget = int(soft.get("single") or single_exec_budget)
+                fail_exec_budget = int(soft.get("fail") or fail_exec_budget)
+                batch_exec_budget = int(soft.get("batch") or 0)
+        except Exception:
+            pass
         local_single_exec = 0
+        local_batch_exec = 0
         local_exec_fails = 0
 
         results_by_id: dict[str, tuple[dict[str, Any], int]] = {}
@@ -1237,7 +1249,31 @@ class ToolExecutor:
             if is_exec_managed_ne_tool(tool_name):
                 batchish = is_batch_exec_args(dict(tc.arguments or {}))
                 single_used = int(prior_single_exec) + int(local_single_exec)
+                batch_used = int(prior_batch_exec) + int(local_batch_exec)
                 fail_used = int(prior_exec_fails) + int(local_exec_fails)
+                if batchish and int(batch_exec_budget) > 0 and batch_used >= int(batch_exec_budget):
+                    results_by_id[tc.id] = (
+                        budget_block_payload(
+                            reason="batch_budget",
+                            lang=str(ctx.lang or "en"),
+                            single_used=single_used,
+                            single_budget=single_exec_budget,
+                            fail_used=fail_used,
+                            fail_budget=fail_exec_budget,
+                            batch_used=batch_used,
+                            batch_budget=batch_exec_budget,
+                        ),
+                        0,
+                    )
+                    _trace(
+                        "cli_batch_budget_exceeded",
+                        {
+                            "tool_name": tc.name,
+                            "batch_used": batch_used,
+                            "batch_budget": batch_exec_budget,
+                        },
+                    )
+                    continue
                 if (not batchish) and fail_used >= int(fail_exec_budget):
                     results_by_id[tc.id] = (
                         budget_block_payload(
@@ -1247,6 +1283,8 @@ class ToolExecutor:
                             single_budget=single_exec_budget,
                             fail_used=fail_used,
                             fail_budget=fail_exec_budget,
+                            batch_used=batch_used,
+                            batch_budget=batch_exec_budget,
                         ),
                         0,
                     )
@@ -1264,6 +1302,8 @@ class ToolExecutor:
                             single_budget=single_exec_budget,
                             fail_used=fail_used,
                             fail_budget=fail_exec_budget,
+                            batch_used=batch_used,
+                            batch_budget=batch_exec_budget,
                         ),
                         0,
                     )
@@ -1330,8 +1370,11 @@ class ToolExecutor:
                 )
                 continue
             first_tool_call_id_by_signature[sig] = str(tc.id or "")
-            if is_exec_managed_ne_tool(tool_name) and not is_batch_exec_args(dict(tc.arguments or {})):
-                local_single_exec += 1
+            if is_exec_managed_ne_tool(tool_name):
+                if is_batch_exec_args(dict(tc.arguments or {})):
+                    local_batch_exec += 1
+                else:
+                    local_single_exec += 1
             runnable_tool_uses.append(tc)
 
         for batch in partition_tool_use_batches(runnable_tool_uses, ctx.tools):
@@ -1399,6 +1442,7 @@ class ToolExecutor:
                     if str(result.get("error_code") or "") not in {
                         "cli_call_budget_exceeded",
                         "cli_fail_budget_exceeded",
+                        "cli_batch_budget_exceeded",
                         "identical_retry_blocked",
                         "retry_forbidden_blocked",
                         "tool_loop_guard",
@@ -1427,11 +1471,21 @@ class ToolExecutor:
                 duration_ms=duration_ms,
             )
             tool_log_write_ms = int((time.perf_counter() - t_db1) * 1000)
-            # Keep full payload during the active turn. History compaction is deferred
-            # until the turn finishes, so current-round model context remains lossless.
+            # Persist full blob when large; model context gets compact + result_ref.
             t_trunc = time.perf_counter()
             observed_rows_this_call = int(_estimate_observed_rows(result))
             result_for_llm = dict(persisted_result or {})
+            result_ref: str | None = None
+            try:
+                from runtime.chat.tool_result_store import attach_result_ref, save_tool_result_blob
+
+                result_ref = save_tool_result_blob(
+                    session_id=str(ctx.session_id or ""),
+                    tool_call_id=str(tc.id or ""),
+                    result=persisted_result if isinstance(persisted_result, dict) else result_for_llm,
+                )
+            except Exception:
+                result_ref = None
             if tc.name in _SQL_REPLAY_COMPACT_TOOL_NAMES:
                 current = int(local_turn_tool_name_counts.get(tc.name, 0))
                 current_rows = int(local_turn_tool_observed_rows.get(tc.name, 0))
@@ -1444,6 +1498,27 @@ class ToolExecutor:
                     result_for_llm["_tool_observed_rows_this_call"] = int(observed_rows_this_call)
                     result_for_llm["_tool_observed_rows_cumulative_in_turn"] = int(cumulative_rows)
                     result_for_llm["audit_note"] = "Result compacted for history replay safety."
+            # Compact oversized payloads for the model wire; full body remains in blob/tool_log.
+            try:
+                capped = truncate_tool_result_for_llm_messages(result_for_llm)
+                if isinstance(capped, dict) and capped.get("_truncated_for_llm"):
+                    result_for_llm = capped
+                    if not result_ref:
+                        from runtime.chat.tool_result_store import save_tool_result_blob
+
+                        result_ref = save_tool_result_blob(
+                            session_id=str(ctx.session_id or ""),
+                            tool_call_id=str(tc.id or ""),
+                            result=persisted_result if isinstance(persisted_result, dict) else dict(result_for_llm),
+                            force=True,
+                        )
+                if result_ref:
+                    from runtime.chat.tool_result_store import attach_result_ref
+
+                    result_for_llm = attach_result_ref(result_for_llm, result_ref=result_ref)
+            except Exception:
+                if result_ref and isinstance(result_for_llm, dict):
+                    result_for_llm["result_ref"] = result_ref
             trunc_ms = int((time.perf_counter() - t_trunc) * 1000)
             tool_content = self._json_dumps_safe(result_for_llm)
             t_db2 = time.perf_counter()
@@ -1469,6 +1544,7 @@ class ToolExecutor:
                         isinstance(result, dict) and _result_is_retry_forbidden(result)
                     ),
                     **({"exec_ne_mode": exec_ne_mode} if exec_ne_mode else {}),
+                    **({"result_ref": result_ref} if result_ref else {}),
                 },
             )
             try:

runtime/chat/turn_idle_guard.py

@@ -100,6 +100,7 @@ class TurnIdleTracker:
                 "tool_loop_guard",
                 "cli_call_budget_exceeded",
                 "cli_fail_budget_exceeded",
+                "cli_batch_budget_exceeded",
             }:
                 guard += 1
         stats = RoundStats(

runtime/direct_loop.py

@@ -155,6 +155,41 @@ def _tool_wire_freeze_enabled(store: Any) -> bool:
     return raw in {"1", "true", "yes", "on"}
 
 
+def _mcp_tools_fingerprint(store: Any) -> str:
+    """Stable hash of enabled MCP server tool catalogs (invalidates freeze after sync)."""
+    try:
+        import hashlib
+
+        rows = store.list_mcp_servers(enabled_only=True) if store else []
+        parts: list[str] = []
+        for row in rows or []:
+            if not isinstance(row, dict):
+                continue
+            sid = str(row.get("server_id") or "").strip()
+            if not sid:
+                continue
+            tools = []
+            try:
+                tools = store.list_mcp_server_tools(server_id=sid) or []
+            except Exception:
+                tools = []
+            names: list[str] = []
+            for t in tools:
+                if isinstance(t, dict):
+                    n = str(t.get("tool_name") or t.get("name") or "").strip()
+                    if n:
+                        names.append(n)
+            names.sort()
+            parts.append(f"{sid}:{','.join(names)}")
+        parts.sort()
+        raw = "|".join(parts)
+        if not raw:
+            return "0"
+        return hashlib.sha256(raw.encode("utf-8", errors="ignore")).hexdigest()[:16]
+    except Exception:
+        return "x"
+
+
 def _tool_wire_settings_signature(store: Any) -> tuple[bool, str]:
     runtime_enabled = True
     try:
@@ -163,6 +198,7 @@ def _tool_wire_settings_signature(store: Any) -> tuple[bool, str]:
             runtime_enabled = raw_flag in {"1", "true", "yes", "on"}
     except Exception:
         runtime_enabled = True
+    mcp_fp = _mcp_tools_fingerprint(store)
     sig = "|".join(
         [
             f"rt={int(bool(runtime_enabled))}",
@@ -171,11 +207,21 @@ def _tool_wire_settings_signature(store: Any) -> tuple[bool, str]:
             f"skill_rt={str(store.get_setting('AIA_SKILL_RUNTIME_ENABLED') or '')}",
             f"skill_disabled={str(store.get_setting('AIA_SKILL_DISABLED_NAMES') or '')}",
             f"bind_en={str(store.get_setting('AIA_SKILL_ROLE_BINDING_ENABLED') or '')}",
+            f"mcp_tools={mcp_fp}",
         ]
     )
     return runtime_enabled, sig
 
 
+def invalidate_tool_wire_cache(*, reason: str = "") -> dict[str, Any]:
+    """Clear frozen tool-wire cache so the next turn rebuilds from current catalogs."""
+    global _TOOL_WIRE_FROZEN_SIGNATURE
+    with _TOOL_WIRE_CACHE_LOCK:
+        _TOOL_WIRE_FROZEN_SIGNATURE = None
+        n = len(_TOOL_WIRE_CACHE)
+        _TOOL_WIRE_CACHE.clear()
+    return {"ok": True, "cleared_entries": int(n), "reason": str(reason or "")}
+
 def _tool_wire_cache_key(
     *,
     store: Any,
@@ -416,6 +462,9 @@ def _guard_tool_results_for_llm_context(
         if len(raw) <= cap:
             out.append(_tool_message_with_content(m, raw, sid=session_id))
             continue
+        existing_ref = ""
+        if isinstance(obj, dict):
+            existing_ref = str(obj.get("result_ref") or "").strip()
         preview = raw[: max(1, min(4000, cap - 400))] + "\n...<tool_result_guard_truncated>"
         guarded_obj = {
             "ok": bool(ok) if ok is not None else None,
@@ -427,10 +476,29 @@ def _guard_tool_results_for_llm_context(
             "preview": preview,
             "hint": (
                 "Tool output was too large for safe context replay; it was truncated for the model context. "
-                "Use narrower queries (e.g., smaller glob/max_results) or adjust AIA_TOOL_LLM_MESSAGE_MAX_CHARS. / "
-                "工具输出过大，已在发给模型的上下文中强制截断；请缩小范围或配置 AIA_TOOL_LLM_MESSAGE_MAX_CHARS。"
+                "Call fetch_tool_result(result_ref=...) when result_ref is present, or use narrower queries. / "
+                "工具输出过大，已在上下文中截断；有 result_ref 时用 fetch_tool_result 取全文。"
             ),
         }
+        if existing_ref:
+            guarded_obj["result_ref"] = existing_ref
+            guarded_obj["fetch_tool"] = "fetch_tool_result"
+        else:
+            try:
+                from runtime.chat.tool_result_store import save_tool_result_blob
+
+                full_obj = obj if isinstance(obj, dict) else {"ok": ok, "raw": raw}
+                ref = save_tool_result_blob(
+                    session_id=str(session_id or ""),
+                    tool_call_id=f"msg-{int(getattr(m, 'id', 0) or 0)}",
+                    result=full_obj if isinstance(full_obj, dict) else {"payload": full_obj},
+                    force=True,
+                )
+                if ref:
+                    guarded_obj["result_ref"] = ref
+                    guarded_obj["fetch_tool"] = "fetch_tool_result"
+            except Exception:
+                pass
         guarded = _json_dumps_safe(guarded_obj)
         out.append(_tool_message_with_content(m, guarded, sid=session_id))
         if trace_id:
@@ -445,6 +513,7 @@ def _guard_tool_results_for_llm_context(
                     "original_chars": int(len(raw)),
                     "guarded_chars": int(len(guarded)),
                     "guard_cap_chars": int(cap),
+                    "result_ref": str(guarded_obj.get("result_ref") or ""),
                 },
                 run_id=run_id,
                 attempt_no=attempt_no,
@@ -728,13 +797,23 @@ def _prepare_llm_tools(
 ) -> list[dict[str, Any]]:
     global _TOOL_WIRE_FROZEN_SIGNATURE
     now = time.time()
-    runtime_enabled, sig = _tool_wire_settings_signature(store)
+    runtime_enabled, live_sig = _tool_wire_settings_signature(store)
     freeze_enabled = _tool_wire_freeze_enabled(store)
-    frozen_sig = _TOOL_WIRE_FROZEN_SIGNATURE if freeze_enabled else None
+    expected_frozen = f"rt={int(bool(runtime_enabled))}|{live_sig}"
+    with _TOOL_WIRE_CACHE_LOCK:
+        frozen_sig = _TOOL_WIRE_FROZEN_SIGNATURE if freeze_enabled else None
+        if isinstance(frozen_sig, str) and frozen_sig.strip():
+            # Reuse freeze only while settings + MCP catalog fingerprint still match.
+            if frozen_sig == expected_frozen:
+                sig = frozen_sig
+            else:
+                _TOOL_WIRE_FROZEN_SIGNATURE = None
+                _TOOL_WIRE_CACHE.clear()
+                sig = live_sig
+                frozen_sig = None
+        else:
+            sig = live_sig
     if isinstance(frozen_sig, str) and frozen_sig.strip():
-        # Startup-prewarmed frozen mode: execution path reuses precomputed tool wiring
-        # and does not perform per-turn policy revalidation.
-        sig = frozen_sig
         try:
             rt_head = str(frozen_sig).split("|", 1)[0].strip().lower()
             runtime_enabled = rt_head == "rt=1"
@@ -1670,5 +1749,11 @@ def run_direct_loop(**kwargs: Any) -> TurnRunOutcome:
     return run_oclaw_direct_loop(**kwargs)
 
 
-__all__ = ["run_oclaw_direct_loop", "run_direct_loop", "warm_tool_wire_cache", "tool_wire_freeze_status"]
+__all__ = [
+    "run_oclaw_direct_loop",
+    "run_direct_loop",
+    "warm_tool_wire_cache",
+    "invalidate_tool_wire_cache",
+    "tool_wire_freeze_status",
+]
 

runtime/tools/mcp/sync_tools.py

@@ -85,6 +85,12 @@ def sync_mcp_server_tools(store: Any, row: dict[str, Any]) -> dict[str, Any] | N
         store.replace_mcp_server_tools(server_id=sid, tools=tools)
         detail = {"synced_tools": len(tools), "compat_mode": "bailian_webparser"}
         store.set_mcp_server_health(server_id=sid, status="ok", detail=detail)
+        try:
+            from runtime.direct_loop import invalidate_tool_wire_cache
+
+            invalidate_tool_wire_cache(reason=f"mcp_sync:{sid}")
+        except Exception:
+            pass
         return {"server_id": sid, "ok": True, "health": detail, "tools_synced": len(tools)}
 
     rt = mcp_runtime_for_row(row, store=store)
@@ -128,6 +134,12 @@ def sync_mcp_server_tools(store: Any, row: dict[str, Any]) -> dict[str, Any] | N
                 "health": health,
             },
         )
+        try:
+            from runtime.direct_loop import invalidate_tool_wire_cache
+
+            invalidate_tool_wire_cache(reason=f"mcp_sync:{sid}")
+        except Exception:
+            pass
         return {
             "server_id": sid,
             "ok": True,

runtime/tools/public/fetch_tool_result_tool.py

@@ -0,0 +1,56 @@
+from __future__ import annotations
+
+from typing import Any
+
+from runtime.chat.tool_invocation_context import current_tool_lane_sessions
+from runtime.chat.tool_result_store import load_tool_result_blob
+from runtime.tools.base import ToolSpec
+
+
+def fetch_tool_result_tool() -> ToolSpec:
+    def handler(args: dict[str, Any]) -> dict[str, Any]:
+        ref = str(args.get("result_ref") or "").strip()
+        if not ref:
+            return {"ok": False, "error_code": "result_ref_required", "error": "result_ref_required"}
+        owner, sid = current_tool_lane_sessions()
+        session_id = str(sid or owner or "").strip()
+        max_chars = args.get("max_chars")
+        try:
+            max_i = int(max_chars) if max_chars is not None else None
+        except Exception:
+            max_i = None
+        return load_tool_result_blob(ref, session_id=session_id, max_chars=max_i)
+
+    return ToolSpec(
+        name="fetch_tool_result",
+        description=(
+            "Fetch a previously truncated/guarded tool result by result_ref. "
+            "Use when a tool payload includes result_ref / _truncated_for_llm / _tool_result_guarded "
+            "and you need more detail than the compact preview."
+        ),
+        parameters={
+            "type": "object",
+            "properties": {
+                "result_ref": {
+                    "type": "string",
+                    "description": "Opaque ref from a prior tool message (e.g. tr:…).",
+                },
+                "max_chars": {
+                    "type": "integer",
+                    "minimum": 4000,
+                    "maximum": 500000,
+                    "description": "Optional cap for returned JSON size (default ~120k).",
+                },
+            },
+            "required": ["result_ref"],
+            "additionalProperties": False,
+        },
+        handler=handler,
+        tags=frozenset({"system", "read", "tool_result"}),
+        read_only=True,
+        risk_level="low",
+        timeout_s=8.0,
+    )
+
+
+__all__ = ["fetch_tool_result_tool"]

tests/test_ops_short_intent_and_exec_hints.py

@@ -5,9 +5,9 @@ from runtime.application.gateway.ops_short_intent import (
     filter_tool_specs_for_ops_short_intent,
     is_ops_short_intent_suppressed_tool,
     maybe_ops_short_intent_system_hint,
+    ops_short_intent_cli_soft_budgets,
 )
 from runtime.tools.base import ToolSpec
-from runtime.tools.tool_error_hints import enrich_exec_managed_ne_error, enrich_get_managed_ne_error
 
 
 def test_detect_ops_short_intent_english_field() -> None:
@@ -20,18 +20,19 @@ def test_detect_ops_short_intent_english_field() -> None:
     assert detect_ops_short_intent("hello there how are you doing today with something else") is None
 
 
-def test_ops_short_intent_suppresses_inventory_cli_tools() -> None:
-    assert is_ops_short_intent_suppressed_tool("mcp__netx__listCliTargets", intent="fiber_cut")
-    assert is_ops_short_intent_suppressed_tool("mcp__netx__execManagedNe", intent="offline")
+def test_ops_short_intent_suppresses_diy_xlsx_shell_only() -> None:
     assert is_ops_short_intent_suppressed_tool("run_command", intent="excel_export")
     assert is_ops_short_intent_suppressed_tool("write_xlsx", intent="fiber_cut")
     assert is_ops_short_intent_suppressed_tool("write_xlsx", intent="excel_export")
+    # Soft preference: CLI/inventory stay available for device confirmation.
+    assert not is_ops_short_intent_suppressed_tool("mcp__netx__listCliTargets", intent="fiber_cut")
+    assert not is_ops_short_intent_suppressed_tool("mcp__netx__execManagedNe", intent="offline")
     assert not is_ops_short_intent_suppressed_tool("ume_alarm_xlsx_report", intent="fiber_cut")
     assert not is_ops_short_intent_suppressed_tool("mcp__netx__queryUmeAlarmsRaw", intent="alarm_tally")
     assert not is_ops_short_intent_suppressed_tool("mcp__netx__execManagedNe", intent="continue")
 
 
-def test_filter_tool_specs_for_ops_short_intent_keeps_report_path() -> None:
+def test_filter_tool_specs_for_ops_short_intent_keeps_cli_and_report() -> None:
     def _spec(name: str) -> ToolSpec:
         return ToolSpec(
             name=name,
@@ -52,20 +53,33 @@ def test_filter_tool_specs_for_ops_short_intent_keeps_report_path() -> None:
     names = {t.name for t in kept}
     assert "ume_alarm_xlsx_report" in names
     assert "mcp__netx__aggregateUmeAlarms" in names
+    assert "mcp__netx__listCliTargets" in names
+    assert "mcp__netx__execManagedNe" in names
     assert "write_xlsx" not in names
-    assert "mcp__netx__listCliTargets" not in names
-    assert "mcp__netx__execManagedNe" not in names
     assert "run_command" not in names
     assert len(filter_tool_specs_for_ops_short_intent(tools, intent="continue")) == len(tools)
 
 
+def test_ops_short_intent_cli_soft_budgets() -> None:
+    soft = ops_short_intent_cli_soft_budgets("fiber_cut")
+    assert soft is not None
+    assert int(soft["single"]) <= 2
+    assert int(soft["batch"]) == 1
+    assert ops_short_intent_cli_soft_budgets("continue") is None
+    assert ops_short_intent_cli_soft_budgets(None) is None
+
+
 def test_ops_short_intent_hint_english_default() -> None:
     hint = maybe_ops_short_intent_system_hint(text="LOS on these sites", lang="en")
     assert "fiber" in hint.lower() or "LOS" in hint
     assert "ume_alarm_xlsx_report" in hint
+    assert "soft budget" in hint.lower() or "targets" in hint.lower()
     assert "断纤" not in hint
 
 
+from runtime.tools.tool_error_hints import enrich_exec_managed_ne_error, enrich_get_managed_ne_error
+
+
 def test_enrich_exec_timeout() -> None:
     out = enrich_exec_managed_ne_error(
         {"ok": False, "error_code": "tool_timeout_or_failed", "error": "timeout"}

tests/test_tool_result_ref.py

@@ -0,0 +1,47 @@
+from __future__ import annotations
+
+from pathlib import Path
+
+from runtime.chat.tool_result_store import (
+    attach_result_ref,
+    load_tool_result_blob,
+    save_tool_result_blob,
+)
+from runtime.tools.public.fetch_tool_result_tool import fetch_tool_result_tool
+
+
+def test_save_and_fetch_tool_result_blob(monkeypatch, tmp_path: Path) -> None:
+    monkeypatch.setattr("runtime.chat.tool_result_store.attachments_dir", lambda: tmp_path)
+    big = {"ok": True, "rows": ["x" * 100 for _ in range(80)]}
+    ref = save_tool_result_blob(session_id="sess-1", tool_call_id="tc1", result=big)
+    assert ref and ref.startswith("tr:")
+    loaded = load_tool_result_blob(ref, session_id="sess-1")
+    assert loaded["ok"] is True
+    assert loaded["result"]["ok"] is True
+    assert loaded["truncated"] is False
+
+
+def test_attach_result_ref_adds_fetch_hint() -> None:
+    out = attach_result_ref({"ok": True, "_truncated_for_llm": True, "preview": "p"}, result_ref="tr:abcd")
+    assert out["result_ref"] == "tr:abcd"
+    assert out["fetch_tool"] == "fetch_tool_result"
+    assert "fetch_tool_result" in str(out.get("hint") or "")
+
+
+def test_fetch_tool_result_public_tool(monkeypatch, tmp_path: Path) -> None:
+    monkeypatch.setattr("runtime.chat.tool_result_store.attachments_dir", lambda: tmp_path)
+    ref = save_tool_result_blob(
+        session_id="sess-2",
+        tool_call_id="tc2",
+        result={"ok": True, "data": "hello" * 2000},
+        force=True,
+    )
+    assert ref
+    monkeypatch.setattr(
+        "runtime.tools.public.fetch_tool_result_tool.current_tool_lane_sessions",
+        lambda: (None, "sess-2"),
+    )
+    tool = fetch_tool_result_tool()
+    out = tool.handler({"result_ref": ref})
+    assert out["ok"] is True
+    assert out["result_ref"] == ref

tests/test_tool_wire_freeze_switch.py

@@ -10,6 +10,14 @@ class _DummyStore:
     def get_setting(self, key: str) -> str:
         return str(self.settings.get(key, ""))
 
+    def list_mcp_servers(self, *, enabled_only: bool = False):
+        del enabled_only
+        return []
+
+    def list_mcp_server_tools(self, *, server_id: str):
+        del server_id
+        return []
+
 
 def test_tool_wire_freeze_default_on(monkeypatch) -> None:
     monkeypatch.setattr(dl, "_prepare_llm_tools", lambda **kwargs: [])
@@ -52,3 +60,35 @@ def test_warm_tool_wire_cache_clears_frozen_stale_entries(monkeypatch) -> None:
     st = dl.tool_wire_freeze_status(store=store)
     assert st["frozen"] is True
     assert st["last_warm_count"] == 2
+
+
+def test_invalidate_tool_wire_cache_clears_freeze(monkeypatch) -> None:
+    monkeypatch.setattr(dl, "_prepare_llm_tools", lambda **kwargs: [])
+    store = _DummyStore({"AIA_TOOL_WIRE_FROZEN_ON_STARTUP": "1"})
+    _ = dl.warm_tool_wire_cache(store=store, tools=object(), base_url="", roles=["generalist"])
+    assert dl.tool_wire_freeze_status(store=store)["frozen"] is True
+    out = dl.invalidate_tool_wire_cache(reason="test")
+    assert out["ok"] is True
+    assert dl.tool_wire_freeze_status(store=store)["frozen"] is False
+
+
+def test_mcp_tools_fingerprint_changes_with_catalog() -> None:
+    class _Store(_DummyStore):
+        def __init__(self):
+            super().__init__()
+            self.tools = [{"tool_name": "a"}]
+
+        def list_mcp_servers(self, *, enabled_only: bool = False):
+            del enabled_only
+            return [{"server_id": "netx"}]
+
+        def list_mcp_server_tools(self, *, server_id: str):
+            del server_id
+            return list(self.tools)
+
+    store = _Store()
+    fp1 = dl._mcp_tools_fingerprint(store)
+    store.tools = [{"tool_name": "a"}, {"tool_name": "b"}]
+    fp2 = dl._mcp_tools_fingerprint(store)
+    assert fp1 != fp2
+    assert fp1 != "x"