refactor(github-action-diagnose): 更新troubleshooting及技能工具权限

.gitignore

@@ -13,6 +13,7 @@
 # Claude Code settings
 .claude/
 .omc/
+CLAUDE.local.md
 
 # Temporary files
 *.tmp

skills/infrastructure/github-action-diagnose/SKILL.md

@@ -1,7 +1,7 @@
 ---
 name: github-action-diagnose
 description: 诊断昇腾（Ascend）NPU 集群上 GitHub Actions 执行失败的原因，定位基础设施故障与根因分析
-allowed-tools: Bash(gh run view:*), Bash(gh run list:*), Bash(gh api:*), Bash(kubectl get:*), Bash(kubectl describe:*), Bash(kubectl logs:*), Bash(kubectl config:*), Read, Grep
+allowed-tools: Bash(gh run view:*), Bash(gh run list:*), Bash(gh pr view:*), Bash(gh api:*), Bash(kubectl get:*), Bash(kubectl describe:*), Bash(kubectl logs:*), Read, Grep
 ---
 
 ## Your task
@@ -49,18 +49,12 @@ allowed-tools: Bash(gh run view:*), Bash(gh run list:*), Bash(gh api:*), Bash(ku
    - Pod 已销毁：在日志中搜索 `Successfully assigned <runner_name> to <node_name>`。
 
 #### 2b. 环境故障深度定位
-检查以下故障模式：
 
-- **驱动/硬件失效**：`npu-smi info` 报错、`ERR99999`、`error code 507035`、`Device not found`。
-- **资源溢出 (OOM)**：系统级 `Killed` 信号、`Bus error`（SHM 不足）、`No space left on device`。
-- **多机连锁超时**：多机任务中某节点报 `Timeout`，优先识别并检查 Master 节点。Master 节点识别方法：
-  - 检查 `RANK_TABLE_FILE` 中 `rank_id=0` 对应的节点
-  - 或在日志中搜索 `master_addr` / `MASTER_ADDR` 环境变量指向的节点
-  - 确认 Master 节点是否有 `Unexpected Exit` 或驱动报错
+读取 `references/ascend-troubleshooting.md`，按其中的故障模式逐一比对日志，涵盖：驱动/硬件失效、资源饱和、OOM、设备插件异常、Runner 版本过期、多机连锁超时、节点出口网络故障
 
 ### Step 3: 非环境问题判定
 
-如果 Step 2 未发现环境/硬件故障，检查以下非环境因素：
+如果 Step 2 未发现环境/硬件/网络故障，检查以下非环境因素：
 
 - **YAML 语法错误**：如 `undefined variable "False"`（应为小写 `false`）。
 - **业务逻辑报错**：`AssertionError`、Python Traceback 指向业务源码、测试用例失败。

skills/infrastructure/github-action-diagnose/references/ascend-troubleshooting.md

@@ -22,6 +22,11 @@ Standard output of `npu-smi info`:
 | HBM Usage | If near 100%, memory fragmentation or leakage may occur. |
 | Temp | Should be < 80°C typically. High temp leads to frequency reduction. |
 
+Driver/hardware failure keywords in logs:
+- `ERR99999`: General NPU driver error.
+- `error code 507035`: NPU device initialization failure.
+- `Device not found`: NPU not recognized by driver.
+
 ## 3. Kernel & Driver Logs (dmesg)
 Key strings to search for:
 - `hiai: npu heartbeat loss`: Hardware/Firmware crash.
@@ -37,3 +42,43 @@ Look for:
 - `get npu device count failed`
 - `npu device is unhealthy`
 
+## 5. K8s Runner Logs
+Key strings to search for:
+- **Runner version v2.330.0 is deprecated and cannot receive messages**: github action runner exits for its version is deprecated
+
+## 6. OOM / Resource Overflow
+- `Killed`: system-level OOM kill signal.
+- `Bus error`: insufficient shared memory (SHM).
+- `No space left on device`: disk full.
+
+## 7. Multi-node Cascading Timeout
+When a multi-node job reports `Timeout`, identify the Master node first — it is the most likely root cause.
+
+Master node identification:
+- Check `RANK_TABLE_FILE` for the node with `rank_id=0`.
+- Search logs for `master_addr` / `MASTER_ADDR` environment variable.
+- Confirm whether the Master node has `Unexpected Exit` or driver errors.
+
+## 8. Node Network Failure (出口网络故障)
+
+When a runner node has degraded or broken external network connectivity, symptoms appear during dependency installation or artifact download steps — not in NPU-related steps.
+
+### Identification Signals
+
+| curl Error | Code | Description |
+|---|---|---|
+| `SSL connection timeout` | 28 | SSL handshake blocked; 0 bytes transferred over minutes |
+| `HTTP/2 stream 0 was not closed cleanly: PROTOCOL_ERROR (err 1)` | 92 | Connection reset by server/proxy after partial download |
+| `Could not resolve host` | 6 | DNS resolution failure |
+
+### Precursor Pattern
+- Download speed consistently < 50 KB/s for multi-MB files throughout the job (visible in wget/curl progress output)
+- Example: a 2.35 MB file taking 4+ minutes at 8–14 KB/s before the final curl failure
+
+### Differentiation from Code Issues
+- Failure is in a `curl`/`wget`/`pip` download command, not a Python traceback or test assertion
+- Other concurrent jobs on different runner nodes complete the same installation steps successfully in the same Run
+- Error codes are OS-level (curl exit codes), not Python exceptions
+
+### Node Tracing
+Apply the standard Step 2a node tracing procedure. If multiple failing jobs share the same runner name prefix (same node pool), the fault is pool-level rather than isolated to a single node.