[Bug]: Multiple runtime policy & extension issues: lspci bridge diagnosis blocked in View mode, phantom bcm_search_docs tool, broken test_bcm_docs.py import, and Grafana crash on empty prometheusUrl #69

AI Factory Operations Agent version

main branch (commit a5d07b6) / v0.0.1

Kubernetes version

1.30+

Installation method

Helm chart from source

Enabled modules and relevant configuration

modules:
kubernetes:
enabled: true
bcm:
enabled: true
observability:
enabled: true
grafana:
enabled: true
edit:
enabled: false # Default View mode
observability:
prometheusUrl: "" # When Prometheus is omitted or VictoriaMetrics/Loki used independently
loki:
url: "http://loki:3100"

Description

Several reproducible issues across extensions, diagnostic policies, and unit tests impact cluster investigations and developer workflows:

1. lspci -s <bridge> -vvv Blocked in View Mode:

   • skills/gpu-networking/references/gpu-network-path.md explicitly instructs agents to inspect PCIe ancestor bridges using lspci -s <bridge> -vvv or lspci -s "$bridge_bdf" -vvv.
   • However, diagnostic-policy.ts validates lspci with flags(/^(-[Dknmvx]+|-[a-zA-Z]{2,})$/). This rejects the -s flag (fails both the character set and min-length requirement) as well as the BDF address argument (e.g. 0000:01:00.0 does not begin with -).
   • Consequently, in View mode, automation-guard classifies the command as an unapproved Edit mutation and blocks it.

2. Phantom bcm_search_docs Tool Registration in bcm/index.ts:

   • In commit f138c26d, BCM manuals were updated to be extracted directly from the UI image into the workspace, and SKILL.md was updated with: "Never call bcm_search_docs; that tool is not available".
   • bcm_search_docs was removed from openclaw.plugin.json contracts and bcm-mcp-ssh-server.py.
   • However, bcm/index.ts (lines 909–933) still registers bcm_search_docs and dispatches to "search_bcm_docs". Because it was omitted from READ_ONLY_TOOLS in automation-context.ts, it is blocked in View mode and fails if executed.

3. Orphan test_bcm_docs.py Fails Python Test Discovery:

   • helm/mosaic-stack/tests/test_bcm_docs.py unconditionally imports files/docs/download_bcm_docs.py, which is not present in the repository. Running python3 -m unittest discover fails with an uncaught FileNotFoundError.

4. Grafana Extension Crashes Agent Startup on Empty prometheusUrl:

   • observability/grafana.ts unconditionally executes resolveString(api.pluginConfig, "prometheusUrl") without a fallback.
   • If Grafana is enabled for Loki logs alone or without a separate Prometheus endpoint (e.g. when configured with --set observability.prometheusUrl=""), OpenClaw crashes at startup with Error: observability.prometheusUrl is required.

5. Path Prefix Discarded in utils/mosaic-mcp.mjs:

   • new URL(path, baseUrl) strips any pathname in MOSAIC_URL when path begins with /, breaking MCP clients connecting through subpath ingress routes (e.g., https://cluster.internal/mosaic).

Reproduction steps

# 1. Verify lspci bridge diagnosis is rejected by diagnostic policy
node -e '
import("./helm/mosaic-stack/files/openclaw-seed/extensions/diagnostic-policy.ts")
  .then(m => console.log("lspci -s bridge allowed:", m.isAllowedDiagnosticArgv(["lspci", "-s", "0000:01:00.0", "-vvv"])));
'

# 2. Verify test discovery failure due to missing download_bcm_docs.py
python3 -m unittest discover -s helm/mosaic-stack/tests

# 3. Verify Grafana plugin startup crash on empty prometheusUrl
node -e '
import("./helm/mosaic-stack/files/openclaw-seed/extensions/observability/grafana.ts")
  .then(m => m.registerGrafanaTools({
    pluginConfig: {
      grafanaEnabled: true,
      grafanaUrl: "http://grafana:3000",
      prometheusUrl: "",
      datasourceUid: "loki-uid",
      datasourceType: "loki"
    },
    registerTool: () => {}
  }));
'

# 4. Verify subpath loss in mosaic-mcp URL construction
node -e 'console.log(new URL("/api/openclaw/tools", "https://cluster.internal/mosaic").href)'

Relevant sanitized logs

# lspci output:
lspci -s bridge allowed: false

# Test discovery output:
ERROR: test_bcm_docs (unittest.loader._FailedTest.test_bcm_docs)
ImportError: Failed to import test module: test_bcm_docs
FileNotFoundError: [Errno 2] No such file or directory: '.../helm/mosaic-stack/files/docs/download_bcm_docs.py'
FAILED (errors=1)

# Grafana plugin output:
Error: observability.prometheusUrl is required
    at resolveString (.../grafana.ts:37:9)
    at registerGrafanaTools (.../grafana.ts:363:25)

# URL resolution output:
https://cluster.internal/api/openclaw/tools # Note: /mosaic path prefix was dropped

Code of Conduct

• I agree to follow the AI Factory Operations Agent Code of Conduct.
• I searched open issues and found no duplicate.