datadog-auto-detector

name	datadog-auto-detector
description	Auto-detection: when user mentions Datadog resources (app.datadoghq.com URLs, 'error rate of service X', 'check logs for Y', monitor/incident/service-health questions) and needs data, automatically fetches condensed summary via datadog-analyzer subagent. Skips auto-fetch for past tense mentions, already-fetched data, informational discussions, or vague references.
user-invocable	false

Datadog Auto-Detector Skill

Seamlessly integrates Datadog observability context into conversations without polluting the main context window. You decide WHEN to fetch; the datadog-analyzer subagent decides HOW to fetch and what to condense.

For explicit debugging requests (/schovi:debug, "find the root cause"), the debug skill owns the whole flow, including any Datadog fetching. Don't activate on top of it.

Codex Compatibility

If a Claude-style Agent tool or custom subagent_type is unavailable, use the Datadog MCP tools directly and return the same condensed summary shape (max ~1200 tokens). Never paste raw Datadog payloads. Use plugins/schovi/agents/datadog-analyzer/AGENT.md as reference instructions.

Pattern Recognition

URLs (https://app.datadoghq.com/...):

• Logs: /logs?query=...
• APM/Traces: /apm/traces?..., /apm/trace/[trace-id]
• Metrics: /metric/explorer?...
• Dashboards: /dashboard/[id]
• Monitors: /monitors/[id]
• Incidents: /incidents/[id]
• Services: /services/[name]
• RUM: /rum/...

Natural language:

• Metrics: "error rate of [service]", "latency of [service]", "CPU usage", "throughput"
• Logs: "logs for [service]", "error logs", "check [service] logs"
• Traces: "traces for [service]", "slow requests in [service]", "APM data"
• Incidents: "active incidents", "SEV-1 incidents", "current incidents"
• Monitors: "alerting monitors", "triggered monitors", "check monitors for [service]"
• Service health: "is [service] healthy?", "status of [service]"

When to Fetch

• Datadog URL shared and the user needs its content
• Observability question: "What's the error rate?", "Show me logs for service X"
• Status check: "Is pb-backend-web healthy?", "Check monitors"
• Investigation: "Users report 500 errors, can you check Datadog?"

When to Skip

• Past tense: "I checked the error rate yesterday", "Datadog showed high latency"
• Already fetched this conversation (re-fetch only on explicit request for fresh data)
• Informational: "Datadog is our monitoring tool", "We use Datadog for observability"
• Too vague: "something in Datadog" (ask for clarification instead)

Workflow

Step 1: Detect & Evaluate

Scan the message for URLs and observability keywords. Apply the fetch/skip rules above.

Step 2: Classify Intent

Full context | specific query | quick status | investigation | comparison. The intent shapes what the analyzer should focus on.

Step 3: Fetch

Tool: Agent
Parameters:
  subagent_type: "schovi:datadog-analyzer:datadog-analyzer"
  prompt: |
    Fetch and summarize [resource type] for [context].
    [If URL]: Datadog URL: [url]
    [If query]: Service: [name], Query Type: [type], Time Range: [range]
    Intent: [classified intent]
    Focus on: [specific aspects user cares about]
  description: "Fetching Datadog observability data"

Step 4: Integrate Naturally

Answer the user's question using the relevant parts of the summary. Don't regurgitate the full summary.

Session Memory

Track what you've fetched this conversation (check transcript for previous datadog-analyzer calls). Reuse existing context instead of re-fetching.

Limits & Error Handling

• Max 3 resources per response; for longer lists ask which ones matter
• Fetch failed: report in one line, suggest checking the Datadog MCP server config, continue with whatever the user can provide
• Ambiguous service name: ask which service before fetching