Tool Health Smoke

Overview

Use this skill for broad tool-by-tool smoke checks. It is narrower than the full qa skill: it only verifies that each live tool CLI is installed and that its health command succeeds through the normal sandbox credential path.

Do not invent one-off probes for each tool. The canonical smoke surface is:

<tool> health

Workflow

1. Run the bundled runner from a sandbox or environment where centaur-tools and tool shims are installed:

   uv run .agents/skills/tool-health-smoke/scripts/run_tool_health_smoke.py
   

2. The runner discovers tools with centaur-tools json, then executes each <tool> health command with a bounded timeout.

3. Review the generated Slack-friendly report. The first line is the overall result:

   Overall: PASS|FAIL|PARTIAL - <reason>
   

4. Return the report in the Slack thread as the assistant response. Do not call the Slack posting API unless the user explicitly asks you to post through the Slack tool.

Report Rules

• Treat returncode != 0, invalid JSON, missing ok, or ok: false as a failed tool health check.
• Treat a missing health command as a failure. Tools are expected to expose health.
• Report the runner output directly. Health commands are responsible for returning compact, safe JSON.
• Keep evidence compact: include the tool name, status, and one short detail or error.
• Use Slack mrkdwn bullets, not Markdown tables.
• Include all failed rows. If all tools pass, include a compact pass list or pass count.

Useful Options

uv run .agents/skills/tool-health-smoke/scripts/run_tool_health_smoke.py --timeout 45
uv run .agents/skills/tool-health-smoke/scripts/run_tool_health_smoke.py --concurrency 4
uv run .agents/skills/tool-health-smoke/scripts/run_tool_health_smoke.py --only slack,websearch,vlogs
uv run .agents/skills/tool-health-smoke/scripts/run_tool_health_smoke.py --json

Use --json only when you need machine-readable results for follow-up analysis. Use the default text output for Slack.

Interpreting Results

• PASS: every discovered tool health command returned valid JSON with ok: true.
• FAIL: one or more tools failed, timed out, returned malformed output, or omitted the required health result.
• PARTIAL: discovery succeeded but no tools were found, or the run was intentionally filtered.

When a tool fails because a credential is missing, check whether the client incorrectly requires an environment variable instead of using secret() or proxy-injected auth.