diff --git a/api/ops/intent_router.py b/api/ops/intent_router.py
new file mode 100644
index 00000000..2649d350
--- /dev/null
+++ b/api/ops/intent_router.py
@@ -0,0 +1,174 @@
+"""Ops Chat LLM intent router（P1-4）。
+
+基于轻量 LLM JSON 输出的混合意图分类器：
+- 环境变量 `OPS_CHAT_LLM_ROUTER=1` 时优先调用 LLM router。
+- LLM 输出必须含 `intent`、`slots`、`confidence`（0~1）。
+- 低置信度或非法 JSON / LLM 异常时降级为规则分类器，并记录
+  `intent_router.fallback` 事件到 `ops_run_events`。
+- 默认未开启时行为与原有规则分类器完全一致。
+"""
+
+from __future__ import annotations
+
+import json
+import logging
+import os
+import re
+import time
+from collections.abc import Callable
+from typing import Any
+
+from api.ops.llm import chat_completion
+
+logger = logging.getLogger(__name__)
+
+# 与 api.ops.orchestrator.core.Intent 保持一致（避免循环导入）
+_ALLOWED_INTENTS = {
+    "metrics_trend",
+    "issue_list",
+    "pr_list",
+    "issue_contribution",
+    "graph_module",
+    "scan_status",
+    "demo",
+    "fallback",
+}
+
+_MIN_CONFIDENCE = float(os.getenv("OPS_CHAT_LLM_ROUTER_MIN_CONFIDENCE") or "0.7")
+
+
+def _is_enabled() -> bool:
+    return os.getenv("OPS_CHAT_LLM_ROUTER", "") == "1"
+
+
+def _build_prompt(message: str) -> str:
+    return (
+        "你是 Ops Chat 意图分类助手。根据用户输入，输出严格 JSON（不要 markdown 代码块，不要解释）：\n"
+        "{\n"
+        '  "intent": "以下之一：metrics_trend, issue_list, pr_list, issue_contribution, '
+        "graph_module, scan_status, demo, fallback\",\n"
+        '  "slots": {},\n'
+        '  "confidence": 0.0\n'
+        "}\n"
+        "slots 可包含 issue_number、days、metric 等键；没有时填 {}。\n"
+        "confidence 为 0~1 的浮点数，表示分类置信度。\n\n"
+        f"用户输入：{message}"
+    )
+
+
+def _extract_json_obj(content: str) -> dict[str, Any]:
+    """从 LLM 输出中提取 JSON 对象；失败时抛出异常。"""
+    try:
+        data = json.loads(content)
+    except json.JSONDecodeError:
+        match = re.search(r"\{.*\}", content, re.S)
+        if not match:
+            raise ValueError("No JSON object found in LLM response") from None
+        data = json.loads(match.group(0))
+    if not isinstance(data, dict):
+        raise ValueError("LLM response JSON is not an object")
+    return data
+
+
+def _normalize_intent(value: Any) -> str:
+    """规范化 intent 为允许值之一，未知则 fallback。"""
+    if not isinstance(value, str):
+        return "fallback"
+    intent = value.strip().lower()
+    if intent in _ALLOWED_INTENTS:
+        return intent
+    return "fallback"
+
+
+def _clamp_confidence(value: Any) -> float:
+    """将 confidence 限制在 [0, 1]。"""
+    try:
+        confidence = float(value) if value is not None else 0.0
+    except (TypeError, ValueError):
+        confidence = 0.0
+    if not 0.0 <= confidence <= 1.0:
+        confidence = 0.0
+    return confidence
+
+
+def _record_fallback_event(
+    run_id: str | None,
+    store: Any,
+    reason: str,
+    detail: str,
+) -> None:
+    """记录 intent_router.fallback 事件；store 未提供时仅记录日志。"""
+    if not run_id:
+        return
+    try:
+        from api.ops.store.runs import append_event
+
+        append_event(
+            run_id,
+            "intent_router.fallback",
+            {"reason": reason, "detail": detail},
+            store=store,
+        )
+    except Exception:  # pragma: no cover - 防御性降级
+        logger.exception("intent_router.fallback event write failed")
+
+
+def llm_classify_intent(message: str) -> tuple[str, dict[str, Any], float]:
+    """直接调用 LLM 分类意图。
+
+    返回 (intent, slots, confidence)。失败时抛出异常，由调用方降级。
+    """
+    prompt = _build_prompt(message)
+    start = time.perf_counter()
+    try:
+        result = chat_completion(
+            [{"role": "user", "content": prompt}],
+            step="intent_router",
+            temperature=0.1,
+        )
+    except Exception as exc:
+        latency_ms = (time.perf_counter() - start) * 1000
+        logger.warning("router.latency: %.2f ms; LLM failed: %s", latency_ms, exc)
+        raise
+
+    latency_ms = (time.perf_counter() - start) * 1000
+    logger.info("router.latency: %.2f ms", latency_ms)
+
+    data = _extract_json_obj(result.content)
+    intent = _normalize_intent(data.get("intent"))
+    slots = data.get("slots") or {}
+    if not isinstance(slots, dict):
+        slots = {}
+    confidence = _clamp_confidence(data.get("confidence"))
+    return intent, slots, confidence
+
+
+def classify_intent_with_llm(
+    message: str,
+    fallback_fn: Callable[[str], tuple[str, dict[str, Any]]],
+    *,
+    run_id: str | None = None,
+    store: Any = None,
+) -> tuple[str, dict[str, Any]]:
+    """混合意图分类器。
+
+    - 当 `OPS_CHAT_LLM_ROUTER=1` 时优先调用 LLM router。
+    - LLM 低置信度或异常时调用 `fallback_fn` 并记录事件。
+    - 默认关闭时直接返回 `fallback_fn(message)`，保持向后兼容。
+    """
+    if not _is_enabled():
+        return fallback_fn(message)
+
+    try:
+        intent, slots, confidence = llm_classify_intent(message)
+    except Exception as exc:
+        result = fallback_fn(message)
+        _record_fallback_event(run_id, store, "llm_error", str(exc))
+        return result
+
+    if confidence < _MIN_CONFIDENCE:
+        result = fallback_fn(message)
+        _record_fallback_event(run_id, store, "low_confidence", f"confidence={confidence:.3f}")
+        return result
+
+    return intent, slots
diff --git a/api/ops/orchestrator/core.py b/api/ops/orchestrator/core.py
index 803c9ea9..b647f4af 100644
--- a/api/ops/orchestrator/core.py
+++ b/api/ops/orchestrator/core.py
@@ -5,6 +5,7 @@
 import re
 from typing import Any
 
+from api.ops import intent_router as _intent_router
 from api.ops.agents.graph_analyst import analyze_graph
 from api.ops.agents.issue_analyst import analyze_issue
 from api.ops.agents.scan_analyst import analyze_scan
@@ -31,7 +32,7 @@ class Intent:
     FALLBACK = "fallback"
 
 
-def classify_intent(message: str) -> tuple[str, dict[str, Any]]:
+def _rule_classify_intent(message: str) -> tuple[str, dict[str, Any]]:
     """基于规则快速分类；返回 (intent, slots)。"""
     msg = message.lower().strip()
     slots: dict[str, Any] = {}
@@ -114,6 +115,24 @@ def classify_intent(message: str) -> tuple[str, dict[str, Any]]:
     return Intent.FALLBACK, {}
 
 
+def classify_intent(
+    message: str,
+    run_id: str | None = None,
+    store: Any | None = None,
+) -> tuple[str, dict[str, Any]]:
+    """混合意图分类：默认规则；`OPS_CHAT_LLM_ROUTER=1` 时优先 LLM router。
+
+    当 LLM router 低置信度或异常时，自动降级为 `_rule_classify_intent`，
+    并通过 `append_event` 记录 `intent_router.fallback` 事件（提供 run_id/store 时）。
+    """
+    return _intent_router.classify_intent_with_llm(
+        message,
+        _rule_classify_intent,
+        run_id=run_id,
+        store=store,
+    )
+
+
 def is_fast_intent(intent: str) -> bool:
     return intent in (Intent.METRICS_TREND, Intent.ISSUE_LIST, Intent.PR_LIST, Intent.DEMO)
 
diff --git a/docs/_tech_graph/02_version.md b/docs/_tech_graph/02_version.md
index c9474185..75479a4e 100644
--- a/docs/_tech_graph/02_version.md
+++ b/docs/_tech_graph/02_version.md
@@ -56,5 +56,6 @@ timeline
   2026-07-07 : dd89b870 auto: api/agently_lab/__init__.py
   2026-07-08 : 4bf5782c auto: api/ops/orchestrator/__init__.py
   2026-07-09 : db09fd40 auto: api/ops/events_schema.py
+  2026-07-10 : 13553deb auto: api/ops/intent_router.py
 ```
 
diff --git a/docs/harness/invokes/by-task/ops-chat-session-sink-p0-p1/invoke_20260710_0929_30_ops_chat_session_sink_p0_p1_P1-4.md b/docs/harness/invokes/by-task/ops-chat-session-sink-p0-p1/invoke_20260710_0929_30_ops_chat_session_sink_p0_p1_P1-4.md
new file mode 100644
index 00000000..729e936f
--- /dev/null
+++ b/docs/harness/invokes/by-task/ops-chat-session-sink-p0-p1/invoke_20260710_0929_30_ops_chat_session_sink_p0_p1_P1-4.md
@@ -0,0 +1,79 @@
+# Invoke Snapshot · 30-execute-code · ops-chat-session-sink-p0-p1 · P1-4
+
+| 字段 | 值 |
+| --- | --- |
+| **hat** | 30-execute-code |
+| **task** | docs/harness/tasks/active/task_ops_chat_session_sink_p0_p1_v1.md |
+| **subproject** | ai-ink-brain-api-python |
+| **branch** | task/ops-chat-session-sink-p0-p1 |
+| **scope** | P1-4 LLM Router：`OPS_CHAT_LLM_ROUTER` · JSON intent · 规则 fallback |
+| **timestamp** | 2026-07-10 09:29 |
+| **verify_command** | `pytest tests/ops tests/ops_desk -m "not intent_eval and not intent_benchmark" -q && ruff check api/ops` |
+| **human_gate** | `HG-TASK-DRAFT`: approved (task §行为变更 · human_gate) · `HG-AUDIT-R1`: approved (task §行为变更 · human_gate) |
+
+## 用户消息快照
+
+```text
+你正在扮演工作区 Harness「30-execute-code · 执行编码帽」，严格遵循 docs/harness/prompts/30-execute-code.md。
+
+**输入（已替换占位符）**
+- 主 task 路径（相对 Projects/）：`docs/harness/tasks/active/task_ops_chat_session_sink_p0_p1_v1.md`
+- 逻辑子仓（相对 Projects/）：`ai-ink-brain-api-python`
+- Worktree 研发目录（所有 git/pytest/ruff 默认 cwd）：`ai-ink-brain-api-python`
+- 当前分支：`task/ops-chat-session-sink-p0-p1`（已基于 main fast-forward，包含 P1-3 merge）
+- 合并前须跑通的验证命令：
+  ```bash
+  pytest tests/ops tests/ops_desk -m "not intent_eval and not intent_benchmark" -q && ruff check api/ops
+  ```
+- 关联任务审核书面结论路径：`ai-ink-brain-api-python/docs/harness/reviews/task_ops_chat_session_sink_p0_p1_v1_audit_R2_20260708.md`
+- 关联 PLAN / 总规：`docs/harness/guides/PLAN_ops_chat_session_sink_p0_p1_v1_zh.md`
+- 关联结构化输出缺口矩阵：`docs/harness/guides/RUNTIME_structured_output_gap_matrix_v0_zh.md`
+
+**本棒目标：P1-4 LLM Router：`OPS_CHAT_LLM_ROUTER` · JSON intent · 规则 fallback**
+
+P1-4 具体要求（来自 PLAN §2、§3.1 D9、task §失败路径、§实现备忘）：
+- 新增 `api/ops/intent_router.py`：实现基于 LLM 的轻量 JSON intent 路由器。
+- 当环境变量 `OPS_CHAT_LLM_ROUTER=1` 时，`api/ops/orchestrator/core.py` 中的 `classify_intent` 优先调用 LLM router。
+- LLM router 输出 JSON：至少含 `intent`（字符串）、`slots`（对象）、`confidence`（float，0~1）。
+- 低置信度或非法 JSON 时降级为原有规则 `classify_intent`（即 task §失败路径的 `intent_router.fallback` event / `router.latency` 日志）。
+- 默认 `OPS_CHAT_LLM_ROUTER` 未开启或未设置时，行为与之前完全一致（向后兼容）。
+- 记录 `intent_router.fallback` event 到 `ops_run_events`（可复用 P0-2 `append_event`）。
+
+**范围限制**
+- 只做 P1-4；不改 P1-1 artifact、P1-2 checkpoint、P1-3 clarify 已交付行为
+- 不改 `harness_runtime` 生产图
+- 不改 Agently lab
+- 不改前端代码
+
+**test_strategy: required**
+- 先写/调整可失败的自动化测试，再改实现
+- 新增 `tests/ops/test_intent_router.py` 覆盖：
+  - `OPS_CHAT_LLM_ROUTER=1` 时 LLM router 返回合法 JSON intent
+  - 低置信度时降级规则 fallback
+  - LLM 超时/非法 JSON 时降级规则 fallback（对应 task §失败路径）
+  - 默认未开启时走原有规则
+  - 1 个集成测：通过 `classify_intent` 走 LLM router
+- 最终验证命令必须绿
+
+**失败路径硬性检查**
+- task §失败路径已列 `LLM router 超时/非法 JSON`：行为 = 降级 `classify_intent` 规则；可观测 = `intent_router.fallback` event / `router.latency` 日志；可重试 = 否；验证命令 = `pytest tests/ops/test_intent_router.py -k fallback`
+
+**你必须完成**
+0. **Invoke 快照（开帽起点）**：将本用户消息全文落盘到 `ai-ink-brain-api-python/docs/harness/invokes/by-task/ops-chat-session-sink-p0-p1/invoke_YYYYMMDD_HHMM_30_ops_chat_session_sink_p0_p1_P1-4.md`（含元数据表 + 快照 fenced code）。同一会话内追问 **不** 再新增快照文件。
+0b. **人工闸**：扫描 task / 关联 reviews 的 human_gate。若任一对本帽（30）为 pending → 仅输出须人改的 gate_id 与路径，拒开工；禁止代填 approved。
+1. 通读 task 全文：头部 gates_before_code、audit_profile、orchestration、chain_prompt、test_strategy / test_strategy_note、failure_paths、验收标准、必读列表、非范围。
+2. 阅读 PLAN §2、§3.1、`RUNTIME_structured_output_gap_matrix_v0_zh.md` 与关联 SNAPSHOT。
+3. 先读现有代码：`api/ops/orchestrator/core.py`（`classify_intent` 与 `Intent` 枚举）、`api/ops/events_schema.py`、`api/ops/chat_service.py`。
+4. 先写失败可复现的测试（`tests/ops/test_intent_router.py`），再实现 `api/ops/intent_router.py` 与 `classify_intent` 改造。
+5. 执行验证命令，保留可核对输出要点；修复直至通过。
+6. 按 40-self-check.md 将结论与命令摘要回填至 task 正文「### 自检结论（执行者）· P1-4」小节（不要覆盖 P0、P1-1、P1-2、P1-3 已有结论）。
+7. 对话回复：生成可以完整复制的 Prompt，用于直接交给下一棒 40 自检执行。
+8. **自动 commit**：在输出下一棒 Prompt 且本轮代码/测试/task 自检回填已落盘后，按 HANDOFF_AUTO_COMMIT.md 在 ai-ink-brain-api-python/ commit（仅本轮路径；禁止 git add -A；对话报 short-hash）。
+9. **禁止**自行 push；由 Lead 合并。
+
+**输出要求**
+- 若拒开工：仅 Markdown 阻塞清单
+- 若执行：diff 摘要、验证命令输出、commit short-hash、下一棒 40 Prompt
+
+**Judgment（本帽 · 对话末尾必填）**：experience_capture / gate/risk / hat_self
+```
diff --git a/docs/harness/invokes/by-task/ops-chat-session-sink-p0-p1/invoke_20260710_0930_40_ops_chat_session_sink_p0_p1_P1-4.md b/docs/harness/invokes/by-task/ops-chat-session-sink-p0-p1/invoke_20260710_0930_40_ops_chat_session_sink_p0_p1_P1-4.md
new file mode 100644
index 00000000..ecde18aa
--- /dev/null
+++ b/docs/harness/invokes/by-task/ops-chat-session-sink-p0-p1/invoke_20260710_0930_40_ops_chat_session_sink_p0_p1_P1-4.md
@@ -0,0 +1,75 @@
+---
+hat: "40-self-check"
+task: "ops-chat-session-sink-p0-p1"
+phase: "P1-4"
+subproject: "ai-ink-brain-api-python"
+branch: "task/ops-chat-session-sink-p0-p1"
+worktree_root: "ai-ink-brain-api-python/"
+date: "2026-07-10"
+time: "09:30"
+---
+
+| 字段 | 值 |
+| --- | --- |
+| **hat** | 40-self-check |
+| **task** | ops-chat-session-sink-p0-p1 |
+| **phase** | P1-4 LLM Router |
+| **subproject** | ai-ink-brain-api-python |
+| **branch** | task/ops-chat-session-sink-p0-p1 |
+| **worktree_root** | ai-ink-brain-api-python/ |
+| **date** | 2026-07-10 |
+| **time** | 09:30 |
+
+## 用户消息快照
+
+```text
+你正在扮演工作区 Harness「40-self-check · 执行者自检帽」，严格遵循 docs/harness/prompts/40-self-check.md。
+
+**输入（已替换占位符）**
+- 主 task 路径（相对 Projects/）：`docs/harness/tasks/active/task_ops_chat_session_sink_p0_p1_v1.md`
+- 逻辑子仓（相对 Projects/）：`ai-ink-brain-api-python`
+- Worktree 研发目录（所有 git/pytest/ruff 默认 cwd）：`ai-ink-brain-api-python`
+- 当前分支：`task/ops-chat-session-sink-p0-p1`（已基于 main fast-forward，包含 P1-3 merge）
+- 合并前须跑通的验证命令：
+  ```bash
+  pytest tests/ops tests/ops_desk -m "not intent_eval and not intent_benchmark" -q && ruff check api/ops
+  ```
+- 失败路径验证命令：
+  ```bash
+  pytest tests/ops/test_intent_router.py -k fallback -q
+  ```
+- 上一棒 30 commit：`ai-ink-brain-api-python@760179a5`
+- 关联任务审核书面结论路径：`ai-ink-brain-api-python/docs/harness/reviews/task_ops_chat_session_sink_p0_p1_v1_audit_R2_20260708.md`
+- 关联 PLAN / 总规：`docs/harness/guides/PLAN_ops_chat_session_sink_p0_p1_v1_zh.md`
+- 关联结构化输出缺口矩阵：`docs/harness/guides/RUNTIME_structured_output_gap_matrix_v0_zh.md`
+
+**本棒目标：P1-4 自检复核**
+
+你必须完成：
+0. **Invoke 快照（开帽起点）**：将本用户消息全文落盘到 `ai-ink-brain-api-python/docs/harness/invokes/by-task/ops-chat-session-sink-p0-p1/invoke_YYYYMMDD_HHMM_40_ops_chat_session_sink_p0_p1_P1-4.md`（含元数据表 + 快照 fenced code）。同一会话内追问 **不** 再新增快照文件。
+0b. **人工闸**：扫描 task / 关联 reviews 的 human_gate。若任一对本帽（40）为 pending → 仅输出须人改的 gate_id 与路径，拒开工；禁止代填 approved。
+1. 独立阅读 task 正文「### 自检结论（执行者）· P1-4」小节与上一棒 30 invoke 快照 `ai-ink-brain-api-python/docs/harness/invokes/by-task/ops-chat-session-sink-p0-p1/invoke_20260710_0929_30_ops_chat_session_sink_p0_p1_P1-4.md`。
+2. 独立阅读本轮 P1-4 改动代码：
+   - `api/ops/intent_router.py`
+   - `api/ops/orchestrator/core.py`（`classify_intent` / `_rule_classify_intent`）
+   - `tests/ops/test_intent_router.py`
+3. 在 `ai-ink-brain-api-python/` 内完整执行 30 声明的验证命令：
+   ```bash
+   pytest tests/ops tests/ops_desk -m "not intent_eval and not intent_benchmark" -q && ruff check api/ops
+   ```
+   并单独执行失败路径验证命令：
+   ```bash
+   pytest tests/ops/test_intent_router.py -k fallback -q
+   ```
+4. 通过 `git diff origin/main...HEAD --stat`（在 `ai-ink-brain-api-python` 内）核对全量变更路径，确认未扩 scope 到 P1-1 artifact、P1-2 checkpoint、P1-3 clarify、`harness_runtime` 生产图、Agently lab、前端代码。
+5. 按 40-self-check.md 将结论与命令摘要回填至 task 正文「### 自检结论（40 复核）· P1-4」小节（不要覆盖 P0、P1-1、P1-2、P1-3 或 30 已有结论）。
+6. 对话回复：生成可以完整复制的 Prompt，用于直接交给下一棒 50 独立复检执行。
+7. 自动 commit：在输出下一棒 Prompt 且本轮 task 自检回填已落盘后，按 HANDOFF_AUTO_COMMIT.md 在 `ai-ink-brain-api-python/` commit（仅本轮路径；禁止 git add -A；对话报 short-hash）。
+8. **禁止**自行 push；由 Lead 合并。
+
+**输出要求**
+- 若拒开工：仅 Markdown 阻塞清单
+- 若执行：diff 摘要、验证命令输出、commit short-hash、下一棒 50 Prompt
+
+**Judgment（本帽 · 对话末尾必填）**：experience_capture / gate/risk / hat_self
+```
diff --git a/docs/harness/invokes/by-task/ops-chat-session-sink-p0-p1/invoke_20260710_0947_40_ops_chat_session_sink_p0_p1_P1-4.md b/docs/harness/invokes/by-task/ops-chat-session-sink-p0-p1/invoke_20260710_0947_40_ops_chat_session_sink_p0_p1_P1-4.md
new file mode 100644
index 00000000..61e3aa02
--- /dev/null
+++ b/docs/harness/invokes/by-task/ops-chat-session-sink-p0-p1/invoke_20260710_0947_40_ops_chat_session_sink_p0_p1_P1-4.md
@@ -0,0 +1,67 @@
+# Invoke Snapshot · 40-self-check · ops-chat-session-sink-p0-p1 · P1-4
+
+| 字段 | 值 |
+| --- | --- |
+| **hat** | 40-self-check |
+| **task** | docs/harness/tasks/active/task_ops_chat_session_sink_p0_p1_v1.md |
+| **subproject** | ai-ink-brain-api-python |
+| **branch** | task/ops-chat-session-sink-p0-p1 |
+| **scope** | P1-4 LLM Router：`OPS_CHAT_LLM_ROUTER` · JSON intent · 规则 fallback |
+| **timestamp** | 2026-07-10 09:47 |
+| **verify_command** | `pytest tests/ops tests/ops_desk -m "not intent_eval and not intent_benchmark" -q && ruff check api/ops` |
+| **failure_path_verify_command** | `pytest tests/ops/test_intent_router.py -k fallback -q` |
+| **human_gate** | `HG-TASK-DRAFT`: approved (task §行为变更 · human_gate) · `HG-AUDIT-R1`: approved (task §行为变更 · human_gate) |
+
+## 用户消息快照
+
+```text
+你正在扮演工作区 Harness「40-self-check · 执行者自检帽」，严格遵循 docs/harness/prompts/40-self-check.md。
+
+**输入（已替换占位符）**
+- 主 task 路径（相对 Projects/）：`docs/harness/tasks/active/task_ops_chat_session_sink_p0_p1_v1.md`
+- 逻辑子仓（相对 Projects/）：`ai-ink-brain-api-python`
+- Worktree 研发目录（所有 git/pytest/ruff 默认 cwd）：`ai-ink-brain-api-python`
+- 当前分支：`task/ops-chat-session-sink-p0-p1`（已基于 main fast-forward，包含 P1-3 merge）
+- 合并前须跑通的验证命令：
+  ```bash
+  pytest tests/ops tests/ops_desk -m "not intent_eval and not intent_benchmark" -q && ruff check api/ops
+  ```
+- 失败路径验证命令：
+  ```bash
+  pytest tests/ops/test_intent_router.py -k fallback -q
+  ```
+- 上一棒 30 commit：`ai-ink-brain-api-python@760179a5`
+- 关联任务审核书面结论路径：`ai-ink-brain-api-python/docs/harness/reviews/task_ops_chat_session_sink_p0_p1_v1_audit_R2_20260708.md`
+- 关联 PLAN / 总规：`docs/harness/guides/PLAN_ops_chat_session_sink_p0_p1_v1_zh.md`
+- 关联结构化输出缺口矩阵：`docs/harness/guides/RUNTIME_structured_output_gap_matrix_v0_zh.md`
+
+**本棒目标：P1-4 自检复核**
+
+你必须完成：
+0. **Invoke 快照（开帽起点）**：将本用户消息全文落盘到 `ai-ink-brain-api-python/docs/harness/invokes/by-task/ops-chat-session-sink-p0-p1/invoke_YYYYMMDD_HHMM_40_ops_chat_session_sink_p0_p1_P1-4.md`（含元数据表 + 快照 fenced code）。同一会话内追问 **不** 再新增快照文件。
+0b. **人工闸**：扫描 task / 关联 reviews 的 human_gate。若任一对本帽（40）为 pending → 仅输出须人改的 gate_id 与路径，拒开工；禁止代填 approved。
+1. 独立阅读 task 正文「### 自检结论（执行者）· P1-4」小节与上一棒 30 invoke 快照 `ai-ink-brain-api-python/docs/harness/invokes/by-task/ops-chat-session-sink-p0-p1/invoke_20260710_0929_30_ops_chat_session_sink_p0_p1_P1-4.md`。
+2. 独立阅读本轮 P1-4 改动代码：
+   - `api/ops/intent_router.py`
+   - `api/ops/orchestrator/core.py`（`classify_intent` / `_rule_classify_intent`）
+   - `tests/ops/test_intent_router.py`
+3. 在 `ai-ink-brain-api-python/` 内完整执行 30 声明的验证命令：
+   ```bash
+   pytest tests/ops tests/ops_desk -m "not intent_eval and not intent_benchmark" -q && ruff check api/ops
+   ```
+   并单独执行失败路径验证命令：
+   ```bash
+   pytest tests/ops/test_intent_router.py -k fallback -q
+   ```
+4. 通过 `git diff origin/main...HEAD --stat`（在 `ai-ink-brain-api-python` 内）核对全量变更路径，确认未扩 scope 到 P1-1 artifact、P1-2 checkpoint、P1-3 clarify、`harness_runtime` 生产图、Agently lab、前端代码。
+5. 按 40-self-check.md 将结论与命令摘要回填至 task 正文「### 自检结论（40 复核）· P1-4」小节（不要覆盖 P0、P1-1、P1-2、P1-3 或 30 已有结论）。
+6. 对话回复：生成可以完整复制的 Prompt，用于直接交给下一棒 50 独立复检执行。
+7. 自动 commit：在输出下一棒 Prompt 且本轮 task 自检回填已落盘后，按 HANDOFF_AUTO_COMMIT.md 在 `ai-ink-brain-api-python/` commit（仅本轮路径；禁止 git add -A；对话报 short-hash）。
+8. **禁止**自行 push；由 Lead 合并。
+
+**输出要求**
+- 若拒开工：仅 Markdown 阻塞清单
+- 若执行：diff 摘要、验证命令输出、commit short-hash、下一棒 50 Prompt
+
+**Judgment（本帽 · 对话末尾必填）**：experience_capture / gate/risk / hat_self
+```
diff --git a/docs/harness/invokes/by-task/ops-chat-session-sink-p0-p1/invoke_20260710_0947_50_ops_chat_session_sink_p0_p1_P1-4.md b/docs/harness/invokes/by-task/ops-chat-session-sink-p0-p1/invoke_20260710_0947_50_ops_chat_session_sink_p0_p1_P1-4.md
new file mode 100644
index 00000000..858d5dda
--- /dev/null
+++ b/docs/harness/invokes/by-task/ops-chat-session-sink-p0-p1/invoke_20260710_0947_50_ops_chat_session_sink_p0_p1_P1-4.md
@@ -0,0 +1,69 @@
+# 50-self-check Invoke Snapshot · P1-4
+
+| 项 | 内容 |
+| --- | --- |
+| **hat** | 50-self-check |
+| **task** | `docs/harness/tasks/active/task_ops_chat_session_sink_p0_p1_v1.md` |
+| **subproject** | `ai-ink-brain-api-python` |
+| **phase** | P1-4 LLM Router |
+| **branch** | `task/ops-chat-session-sink-p0-p1` |
+| **timestamp** | 2026-07-10 09:47 |
+| **verify_command** | `pytest tests/ops tests/ops_desk -m "not intent_eval and not intent_benchmark" -q && ruff check api/ops` |
+| **failure_path_verify_command** | `pytest tests/ops/test_intent_router.py -k fallback -q` |
+
+## Input Snapshot
+
+```text
+你正在扮演工作区 Harness「50-self-check · 独立复检 + 全局验收帽」，严格遵循 docs/harness/prompts/50-self-check.md。
+
+**输入（已替换占位符）**
+- 主 task 路径（相对 Projects/）：`docs/harness/tasks/active/task_ops_chat_session_sink_p0_p1_v1.md`
+- 逻辑子仓（相对 Projects/）：`ai-ink-brain-api-python`
+- Worktree 研发目录（所有 git/pytest/ruff 默认 cwd）：`ai-ink-brain-api-python`
+- 当前分支：`task/ops-chat-session-sink-p0-p1`（已基于 main fast-forward，包含 P1-3 merge）
+- 合并前须跑通的验证命令：
+  ```bash
+  pytest tests/ops tests/ops_desk -m "not intent_eval and not intent_benchmark" -q && ruff check api/ops
+  ```
+- 失败路径验证命令：
+  ```bash
+  pytest tests/ops/test_intent_router.py -k fallback -q
+  ```
+- 上一棒 30 commit：`ai-ink-brain-api-python @ 760179a5`
+- 上一棒 40 commit：`ai-ink-brain-api-python @ 5c4f05c6`
+- 关联任务审核书面结论路径：`ai-ink-brain-api-python/docs/harness/reviews/task_ops_chat_session_sink_p0_p1_v1_audit_R2_20260708.md`
+- 关联 PLAN / 总规：`docs/harness/guides/PLAN_ops_chat_session_sink_p0_p1_v1_zh.md`
+- 关联结构化输出缺口矩阵：`docs/harness/guides/RUNTIME_structured_output_gap_matrix_v0_zh.md`
+
+**本棒目标：P1-4 独立复检 + 全局验收**
+
+你必须完成：
+0. **Invoke 快照（开帽起点）**：将本用户消息全文落盘到 `ai-ink-brain-api-python/docs/harness/invokes/by-task/ops-chat-session-sink-p0-p1/invoke_YYYYMMDD_HHMM_50_ops_chat_session_sink_p0_p1_P1-4.md`（含元数据表 + 快照 fenced code）。同一会话内追问 **不** 再新增快照文件。
+0b. **人工闸**：扫描 task / 关联 reviews 的 human_gate。若任一对本帽（50）为 pending → 仅输出须人改的 gate_id 与路径，拒开工；禁止代填 approved。
+1. 独立阅读 task 正文「### 自检结论（执行者）· P1-4」与「### 自检结论（40 复核）· P1-4」小节。
+2. 独立阅读本轮 P1-4 改动代码：
+   - `api/ops/intent_router.py`
+   - `api/ops/orchestrator/core.py`（`classify_intent` / `_rule_classify_intent`）
+   - `tests/ops/test_intent_router.py`
+3. 在 `ai-ink-brain-api-python/` 内完整执行验证命令：
+   ```bash
+   pytest tests/ops tests/ops_desk -m "not intent_eval and not intent_benchmark" -q && ruff check api/ops
+   ```
+   并单独执行失败路径验证命令：
+   ```bash
+   pytest tests/ops/test_intent_router.py -k fallback -q
+   ```
+4. 执行 `git diff origin/main...HEAD --stat`（在 `ai-ink-brain-api-python` 内）核对全量变更路径，确认未扩 scope 到 P1-1 artifact、P1-2 checkpoint、P1-3 clarify、`harness_runtime` 生产图、Agently lab、前端代码。
+5. 与 30 commit `760179a5`、40 commit（待回填）、R2 任务审核书面结论逐条核对。
+6. 50 复检书面结论落盘：`ai-ink-brain-api-python/docs/harness/reviews/task_ops_chat_session_sink_p0_p1_v1_audit_R1_20260710_50_P1-4.md`。
+7. 按 50-self-check.md 将结论与命令摘要回填至 task 正文「### 自检结论（50 复检）· P1-4」小节（不要覆盖 P0、P1-1、P1-2、P1-3、30、40 已有结论）。
+8. 对话回复：给出是否建议合并、阻塞项清单、全局验收结论。
+9. 自动 commit：在输出最终结论且本轮 reviews/task 自检回填已落盘后，按 HANDOFF_AUTO_COMMIT.md 在 `ai-ink-brain-api-python/` commit（仅本轮路径；禁止 git add -A；对话报 short-hash）。
+10. **禁止**自行 push；由 Lead 合并。
+
+**输出要求**
+- 若拒开工：仅 Markdown 阻塞清单
+- 若执行：diff 摘要、验证命令输出、commit short-hash、合并建议、全局验收表
+
+**Judgment（本帽 · 对话末尾必填）**：experience_capture / gate/risk / hat_self
+```
diff --git a/docs/harness/reviews/task_ops_chat_session_sink_p0_p1_v1_audit_R1_20260710_50_P1-4.md b/docs/harness/reviews/task_ops_chat_session_sink_p0_p1_v1_audit_R1_20260710_50_P1-4.md
new file mode 100644
index 00000000..06c25c51
--- /dev/null
+++ b/docs/harness/reviews/task_ops_chat_session_sink_p0_p1_v1_audit_R1_20260710_50_P1-4.md
@@ -0,0 +1,164 @@
+# 50 独立复检 + 全局验收报告 · Ops Chat Session Sink P0+P1 · P1-4 LLM Router
+
+| 项 | 内容 |
+| --- | --- |
+| **task** | `docs/harness/tasks/active/task_ops_chat_session_sink_p0_p1_v1.md` |
+| **subproject** | `ai-ink-brain-api-python` |
+| **scope** | P1-4 LLM Router：`OPS_CHAT_LLM_ROUTER` · JSON intent · 规则 fallback |
+| **hat** | 50-independent-reinspect |
+| **date** | 2026-07-10 |
+| **30 commit** | `ai-ink-brain-api-python@760179a5` |
+| **40 commit** | `ai-ink-brain-api-python@5c4f05c6` · `Projects@9fef12f` |
+
+---
+
+## 复核方法
+
+1. 独立阅读 task 正文「### 自检结论（执行者）· P1-4」与「### 自检结论（40 复核）· P1-4」小节。
+2. 独立阅读本轮 P1-4 改动代码：
+   - `api/ops/intent_router.py`
+   - `api/ops/orchestrator/core.py`（`classify_intent` / `_rule_classify_intent`）
+   - `tests/ops/test_intent_router.py`
+3. 在 `ai-ink-brain-api-python/` 内完整执行 30 声明的验证命令：
+   ```bash
+   pytest tests/ops tests/ops_desk -m "not intent_eval and not intent_benchmark" -q && ruff check api/ops
+   ```
+4. 单独执行失败路径验证命令：
+   ```bash
+   pytest tests/ops/test_intent_router.py -k fallback -q
+   ```
+5. 执行 `git diff origin/main...HEAD --stat` 核对全量变更路径，确认未扩 scope 到 P1-1 artifact、P1-2 checkpoint、P1-3 clarify、`harness_runtime` 生产图、Agently lab、前端代码。
+6. 与 30 commit `760179a5`、40 commit `5c4f05c6` / `Projects@9fef12f`、R2 任务审核书面结论逐条核对。
+
+---
+
+## 命令输出
+
+### 完整验证命令
+
+```text
+pytest tests/ops tests/ops_desk -m "not intent_eval and not intent_benchmark" -q
+............................................s........................... [ 22%]
+........................................................................ [ 44%]
+........................................................................ [ 66%]
+.........................................ss......................sssssss [ 88%]
+......................................                                   [100%]
+=============================== warnings summary ================================
+../../../miniconda3/lib/python3.13/site-packages/fastapi/testclient.py:1
+  /Users/cyning/miniconda3/lib/python3.13/site-packages/fastapi/testclient.py:1: StarletteDeprecationWarning: Using `httpx` with `starlette.testclient` is deprecated; install `httpx2` instead.
+    from starlette.testclient import TestClient as TestClient  # noqa
+
+-- Docs: https://docs.pytest.org/en/stable/howto/capture-warnings.html
+=========================== short test summary info ============================
+SKIPPED [1] tests/ops/test_events_schema.py:181: 需要真实 Supabase 连接；本地/CI 环境缺失时跳过
+SKIPPED [2] tests/ops_desk/test_run_schema_p1.py:102: public 中表已存在，跳过写测试以避免破坏数据
+SKIPPED [7] tests/ops_desk/test_schema_p0.py:102: public 中表已存在，跳过写测试以避免破坏数据
+316 passed, 10 skipped, 1 warning in 57.58s
+
+ruff check api/ops
+All checks passed!
+```
+
+- pytest 退出码：`0`
+- ruff 退出码：`0`
+- 10 skipped 中：1 个为 `tests/ops/test_events_schema.py::test_append_event_integration_with_real_store`（显式 skip，需真实 Supabase 连接）；其余 9 个为 `tests/ops_desk/test_run_schema_p1.py` / `tests/ops_desk/test_schema_p0.py` 中环境感知跳过（表已存在），与 P1-4 改动无关。
+
+### 失败路径额外验证
+
+```text
+pytest tests/ops/test_intent_router.py -k fallback -q
+...                                                                      [100%]
+3 passed, 3 deselected in 0.38s
+```
+
+- pytest 退出码：`0`
+- 覆盖：低置信度 fallback、非法 JSON fallback、LLM 超时 fallback；均记录 `intent_router.fallback` event。
+
+---
+
+## 全量变更路径核对
+
+```text
+ api/ops/intent_router.py                           | 174 ++++++++++++++++
+ api/ops/orchestrator/core.py                       |  21 +-
+ docs/_tech_graph/02_version.md                     |   1 +
+ ...710_0929_30_ops_chat_session_sink_p0_p1_P1-4.md |  79 +++++++
+ ...710_0930_40_ops_chat_session_sink_p0_p1_P1-4.md |  75 +++++++
+ ...710_0947_40_ops_chat_session_sink_p0_p1_P1-4.md |  67 ++++++
+ ...710_0947_50_ops_chat_session_sink_p0_p1_P1-4.md |  69 ++++++
+ tests/ops/test_intent_router.py                    | 231 +++++++++++++++++++++
+ 8 files changed, 716 insertions(+), 1 deletion(-)
+```
+
+- 代码变更仅涉及 `api/ops/intent_router.py`、`api/ops/orchestrator/core.py`、`tests/ops/test_intent_router.py`。
+- `docs/_tech_graph/02_version.md` 追加版本时间点，属技术图谱例行更新，未改生产代码。
+- 文档变更为本轮 invoke 快照、task 自检结论与本 50 复检报告，属 harness 落盘工件。
+- 未涉及 P1-1 artifact、P1-2 checkpoint、P1-3 clarify、`harness_runtime` 生产图、Agently lab、前端代码。
+
+---
+
+## 与 30 / 40 结论差异核对
+
+| 30 / 40 声称项 | 50 独立复核 | 结果 |
+| --- | --- | --- |
+| 新增 `api/ops/intent_router.py` 轻量 JSON LLM router | 文件存在；`_build_prompt` / `llm_classify_intent` / `classify_intent_with_llm` 实现完整；输出含 `intent` / `slots` / `confidence` | 一致 |
+| `OPS_CHAT_LLM_ROUTER=1` 时 `classify_intent` 优先调用 LLM router | `core.py:118-133` 委托 `_intent_router.classify_intent_with_llm`；集成测 `test_classify_intent_uses_llm_router` 通过 | 一致 |
+| 低置信度降级规则 fallback | `intent_router.py:169-172` 在 `confidence < _MIN_CONFIDENCE` 时调用 fallback；`test_llm_router_low_confidence_fallback` 通过 | 一致 |
+| 非法 JSON / LLM 异常降级规则 fallback | `_extract_json_obj` 抛异常与 `chat_completion` 异常均在 `classify_intent_with_llm` 捕获并 fallback；对应测例通过 | 一致 |
+| 默认未开启时行为与之前一致 | `_is_enabled()` 仅在 `OPS_CHAT_LLM_ROUTER=1` 时启用；`test_llm_router_disabled_uses_rule` 通过 | 一致 |
+| 记录 `intent_router.fallback` event | `_record_fallback_event` 复用 P0-2 `append_event`；fallback 测例断言 store 中存在 `event_type=intent_router.fallback` | 一致 |
+| `router.latency` 日志 | `llm_classify_intent` 在成功/失败路径均记录 `router.latency` | 一致 |
+| 最终验证命令绿 | 本轮独立跑通 `316 passed, 10 skipped` + ruff 全绿 | 一致 |
+| 未扩 scope | 全量 diff 仅 P1-4 router 相关文件 + docs 版本时间点 + invoke；未涉及 P1-1/2/3、Session 生产图、Agently lab、前端 | 一致 |
+| P0 ~ P1-3 已有实现未被破坏 | `review/rules.py`、`events_schema.py`、`store/runs.py`、`chat_context.py`、`chat_service.py`、`orchestrator/clarify.py` 未改动；相关测例全绿 | 一致 |
+
+**差异项**：无。
+
+---
+
+## 验收项复核表
+
+| 验收项 | 状态 | 证据 | 备注 |
+| --- | --- | --- | --- |
+| `api/ops/intent_router.py` 存在且实现 LLM JSON intent router | pass | 文件新增；`_build_prompt` / `llm_classify_intent` / `classify_intent_with_llm` 完整；返回含 `intent` / `slots` / `confidence` | — |
+| `OPS_CHAT_LLM_ROUTER=1` 时 `classify_intent` 优先调用 LLM router | pass | `core.py:118-133`；集成测 `test_classify_intent_uses_llm_router` 通过 | — |
+| 低置信度降级规则 fallback | pass | `intent_router.py:169-172`；`test_llm_router_low_confidence_fallback` 通过 | — |
+| 非法 JSON / LLM 异常降级规则 fallback | pass | `classify_intent_with_llm` 异常捕获分支；`test_llm_router_invalid_json_fallback` / `test_llm_router_timeout_fallback` 通过 | — |
+| 默认未开启时行为与之前一致 | pass | `_is_enabled()` 仅在 `OPS_CHAT_LLM_ROUTER=1` 时启用；`test_llm_router_disabled_uses_rule` 通过 | — |
+| 记录 `intent_router.fallback` event | pass | `_record_fallback_event` 复用 P0-2 `append_event`；fallback 测例断言 event 存在 | — |
+| `router.latency` 日志 | pass | `llm_classify_intent` 成功/失败路径均记录 `router.latency` | — |
+| `tests/ops/test_intent_router.py` 覆盖目标场景 | pass | 6 测例全绿：启用返回合法 intent、低置信度 fallback、非法 JSON fallback、超时 fallback、默认规则、集成路径 | — |
+| task §失败路径验证命令绿 | pass | `pytest tests/ops/test_intent_router.py -k fallback -q` → `3 passed, 3 deselected` | — |
+| 最终验证命令绿 | pass | pytest `316 passed, 10 skipped` + ruff `All checks passed!`；退出码均为 `0` | — |
+| 未静默扩大 scope | pass | 全量 diff 路径清单见上文 | 未改 P1-1/2/3、Session 生产图、Agently lab、前端 |
+| P0 ~ P1-3 已有实现未被破坏 | pass | `review/rules.py`、`events_schema.py`、`store/runs.py`、`chat_context.py`、`chat_service.py`、`orchestrator/clarify.py` 未改动；相关测例全绿 | — |
+
+---
+
+## 阻塞项清单
+
+无。
+
+---
+
+## 是否建议合并
+
+**建议合并**。P1-4 LLM Router 实现、测试、30 执行、40 自检、50 独立复检均通过，无 scope creep，人工闸 `HG-TASK-DRAFT` / `HG-AUDIT-R1` 已 approved。
+
+---
+
+## 执行路线与 Commit 回溯
+
+| 阶段 | 帽子 | 关键动作 | 落盘工件 | 对应 commit |
+|------|------|----------|----------|-------------|
+| P1-4 | 30 execute | LLM intent router + `classify_intent` 混合分类器 + 测试 | `api/ops/intent_router.py`, `api/ops/orchestrator/core.py`, `tests/ops/test_intent_router.py` | `ai-ink-brain-api-python@760179a5` |
+| P1-4 | 40 self-check | 复核 P1-4 验收 | task 内 P1-4 30/40 自检结论 | `ai-ink-brain-api-python@5c4f05c6` · `Projects@9fef12f` |
+| P1-4 | 50 reinspect R1 | 独立复检 + 全局验收 | `ai-ink-brain-api-python/docs/harness/reviews/task_ops_chat_session_sink_p0_p1_v1_audit_R1_20260710_50_P1-4.md` | 待本审查落盘后 commit |
+
+---
+
+## Judgment（50）
+
+- **experience_capture**: `required` — LLM router JSON intent、confidence 阈值、规则 fallback、事件记录模式可复用到后续 ChatBI / Session 路由。
+- **gate/risk**: 无 — `HG-TASK-DRAFT` / `HG-AUDIT-R1` 均为 `approved`；50 未遇 pending 人工闸。
+- **hat_self**: `pass` — 独立复检与 30/40 结论一致，验证命令绿，失败路径额外验证绿，输出 pass/fail 表、阻塞项清单、合并建议与执行路线回溯。
diff --git a/tests/ops/test_intent_router.py b/tests/ops/test_intent_router.py
new file mode 100644
index 00000000..0349466f
--- /dev/null
+++ b/tests/ops/test_intent_router.py
@@ -0,0 +1,231 @@
+"""P1-4: Ops Chat LLM intent router 单测。
+
+覆盖：
+- OPS_CHAT_LLM_ROUTER=1 时 LLM router 返回合法 JSON intent。
+- 低置信度时降级为规则 fallback。
+- LLM 超时 / 非法 JSON 时降级为规则 fallback。
+- 默认未开启时走原有规则。
+- 通过 classify_intent 走 LLM router 的集成路径。
+"""
+
+from __future__ import annotations
+
+import json
+from typing import Any
+
+import pytest
+
+from api.ops.llm.types import LlmCompletionResult, LlmUsage
+
+
+def _make_llm_result(
+    content: str,
+    step: str = "intent_router",
+    latency_ms: float = 80.0,
+) -> LlmCompletionResult:
+    return LlmCompletionResult(
+        content=content,
+        usage=LlmUsage(
+            provider="siliconflow",
+            model="Qwen/Qwen2.5-72B-Instruct",
+            prompt_tokens=20,
+            completion_tokens=10,
+            total_tokens=30,
+            latency_ms=latency_ms,
+            step=step,
+        ),
+    )
+
+
+class FakeStore:
+    """内存版 OpsRunStore，仅记录 append_event。"""
+
+    def __init__(self) -> None:
+        self.events: list[dict[str, Any]] = []
+
+    def append_event(
+        self,
+        run_id: str,
+        agent_role: str,
+        event_type: str,
+        payload: dict[str, Any] | None = None,
+        node_id: str | None = None,
+        seq: int | None = None,
+    ) -> dict[str, Any]:
+        evt: dict[str, Any] = {
+            "run_id": run_id,
+            "agent_role": agent_role,
+            "event_type": event_type,
+            "payload": payload or {},
+            "node_id": node_id,
+            "seq": seq or len(self.events) + 1,
+        }
+        self.events.append(evt)
+        return evt
+
+
+def test_llm_router_enabled_returns_valid_intent(monkeypatch: pytest.MonkeyPatch) -> None:
+    """OPS_CHAT_LLM_ROUTER=1 且 LLM 返回合法 JSON 时，优先采用 LLM 结果。"""
+    monkeypatch.setenv("OPS_CHAT_LLM_ROUTER", "1")
+    from api.ops import intent_router
+
+    def fake_chat_completion(messages: list[dict[str, str]], **kwargs: Any) -> LlmCompletionResult:
+        return _make_llm_result(
+            json.dumps(
+                {"intent": "issue_contribution", "slots": {"issue_number": 123}, "confidence": 0.95},
+                ensure_ascii=False,
+            )
+        )
+
+    monkeypatch.setattr(intent_router, "chat_completion", fake_chat_completion)
+
+    intent, slots = intent_router.classify_intent_with_llm("分析 #123", lambda _msg: ("fallback", {}))
+
+    assert intent == "issue_contribution"
+    assert slots == {"issue_number": 123}
+
+
+def test_llm_router_low_confidence_fallback(monkeypatch: pytest.MonkeyPatch) -> None:
+    """置信度低于阈值时降级为规则 fallback，并记录 intent_router.fallback 事件。"""
+    monkeypatch.setenv("OPS_CHAT_LLM_ROUTER", "1")
+    from api.ops import intent_router
+
+    def fake_chat_completion(messages: list[dict[str, str]], **kwargs: Any) -> LlmCompletionResult:
+        return _make_llm_result(
+            json.dumps(
+                {"intent": "issue_contribution", "slots": {"issue_number": 123}, "confidence": 0.2},
+                ensure_ascii=False,
+            )
+        )
+
+    monkeypatch.setattr(intent_router, "chat_completion", fake_chat_completion)
+
+    fallback_calls: list[str] = []
+
+    def rule_fallback(msg: str) -> tuple[str, dict[str, Any]]:
+        fallback_calls.append(msg)
+        return "metrics_trend", {"days": 7}
+
+    store = FakeStore()
+    intent, slots = intent_router.classify_intent_with_llm(
+        "比较含糊的问题",
+        rule_fallback,
+        run_id="run-low-conf",
+        store=store,
+    )
+
+    assert intent == "metrics_trend"
+    assert slots == {"days": 7}
+    assert fallback_calls == ["比较含糊的问题"]
+
+    fallback_events = [e for e in store.events if e["event_type"] == "intent_router.fallback"]
+    assert len(fallback_events) == 1
+    assert fallback_events[0]["payload"]["reason"] == "low_confidence"
+    assert "0.2" in fallback_events[0]["payload"]["detail"]
+
+
+def test_llm_router_invalid_json_fallback(monkeypatch: pytest.MonkeyPatch) -> None:
+    """LLM 返回非法 JSON 时降级为规则 fallback，并记录 intent_router.fallback 事件。"""
+    monkeypatch.setenv("OPS_CHAT_LLM_ROUTER", "1")
+    from api.ops import intent_router
+
+    def fake_chat_completion(messages: list[dict[str, str]], **kwargs: Any) -> LlmCompletionResult:
+        return _make_llm_result("这不是 JSON")
+
+    monkeypatch.setattr(intent_router, "chat_completion", fake_chat_completion)
+
+    def rule_fallback(_msg: str) -> tuple[str, dict[str, Any]]:
+        return "fallback", {}
+
+    store = FakeStore()
+    intent, slots = intent_router.classify_intent_with_llm(
+        "模糊问题",
+        rule_fallback,
+        run_id="run-bad-json",
+        store=store,
+    )
+
+    assert intent == "fallback"
+    assert slots == {}
+
+    fallback_events = [e for e in store.events if e["event_type"] == "intent_router.fallback"]
+    assert len(fallback_events) == 1
+    assert fallback_events[0]["payload"]["reason"] == "llm_error"
+
+
+def test_llm_router_timeout_fallback(monkeypatch: pytest.MonkeyPatch) -> None:
+    """LLM 调用超时/抛异常时降级为规则 fallback，并记录 intent_router.fallback 事件。"""
+    monkeypatch.setenv("OPS_CHAT_LLM_ROUTER", "1")
+    from api.ops import intent_router
+
+    def fake_chat_completion(messages: list[dict[str, str]], **kwargs: Any) -> LlmCompletionResult:
+        raise TimeoutError("LLM timeout")
+
+    monkeypatch.setattr(intent_router, "chat_completion", fake_chat_completion)
+
+    def rule_fallback(_msg: str) -> tuple[str, dict[str, Any]]:
+        return "metrics_trend", {"metric": "cycle-time"}
+
+    store = FakeStore()
+    intent, slots = intent_router.classify_intent_with_llm(
+        "最近 PR 周期多少",
+        rule_fallback,
+        run_id="run-timeout",
+        store=store,
+    )
+
+    assert intent == "metrics_trend"
+    assert slots == {"metric": "cycle-time"}
+
+    fallback_events = [e for e in store.events if e["event_type"] == "intent_router.fallback"]
+    assert len(fallback_events) == 1
+    assert fallback_events[0]["payload"]["reason"] == "llm_error"
+    assert "LLM timeout" in fallback_events[0]["payload"]["detail"]
+
+
+def test_llm_router_disabled_uses_rule(monkeypatch: pytest.MonkeyPatch) -> None:
+    """默认未开启 OPS_CHAT_LLM_ROUTER 时，走原有规则，不调用 LLM。"""
+    monkeypatch.delenv("OPS_CHAT_LLM_ROUTER", raising=False)
+    from api.ops import intent_router
+
+    llm_calls: list[Any] = []
+
+    def fake_chat_completion(messages: list[dict[str, str]], **kwargs: Any) -> LlmCompletionResult:
+        llm_calls.append(messages)
+        return _make_llm_result(json.dumps({"intent": "issue_list", "slots": {}, "confidence": 0.9}))
+
+    monkeypatch.setattr(intent_router, "chat_completion", fake_chat_completion)
+
+    def rule_fallback(_msg: str) -> tuple[str, dict[str, Any]]:
+        return "pr_list", {}
+
+    intent, slots = intent_router.classify_intent_with_llm("列出 PR", rule_fallback)
+
+    assert intent == "pr_list"
+    assert slots == {}
+    assert llm_calls == []
+
+
+def test_classify_intent_uses_llm_router(monkeypatch: pytest.MonkeyPatch) -> None:
+    """集成测：通过 api.ops.orchestrator.classify_intent 走 LLM router。"""
+    monkeypatch.setenv("OPS_CHAT_LLM_ROUTER", "1")
+    from api.ops import intent_router
+    from api.ops.orchestrator import classify_intent
+
+    def fake_chat_completion(messages: list[dict[str, str]], **kwargs: Any) -> LlmCompletionResult:
+        return _make_llm_result(
+            json.dumps(
+                {"intent": "issue_list", "slots": {}, "confidence": 0.88},
+                ensure_ascii=False,
+            )
+        )
+
+    monkeypatch.setattr(intent_router, "chat_completion", fake_chat_completion)
+
+    store = FakeStore()
+    intent, slots = classify_intent("列出所有 issues", run_id="run-integration", store=store)
+
+    assert intent == "issue_list"
+    assert slots == {}
+    fallback_events = [e for e in store.events if e["event_type"] == "intent_router.fallback"]
+    assert fallback_events == []