ctxpipe: Skill for Cursor

Observability

Hosted stack: Railway project ctxpipe-observability. URLs, accounts, dashboards, MCP clients, and laptop OTLP: USING.md. Attribute history: ADR-011.

Find a signal

1. Take the id you have: response header x-request-id (request.id), a W3C traceId, enduser.id plus a time window, ctxpipe.org.slug / ctxpipe.org.id, or an OpenWorkflow run (SpanName workflow_run.create, workflow_run.execute, or step_attempt.execute on scope openworkflow). Correlate runs with openworkflow.run.id. Execution spans carry ctxpipe.actor.type=job. Each execution is its own trace, linked to workflow_run.create; step spans are children of that execution.
2. Name the environment: production, pr-<digits>, observability, or local-<name>. Dashboards and the shared filter: USING.md.
3. Read logs, traces, and metrics in HyperDX (MCP hyperdx). Read LLM prompt text in Langfuse (MCP langfuse). Read deploy status and stdout in Railway (MCP railway) when the process exited before export. URLs and keys: USING.md.
4. Done when the failing span or log line is named, or the window is empty and the environment, service, and time range are written down.

Stay inside a few hours. otel parts older than 3 days sit on the cold bucket; a cold read fails within 120s when the bucket is down. Langfuse traces, observations, and scores move after 30 days and are not deleted; read them with FINAL.

SELECT Timestamp, ServiceName, SpanName, StatusCode, Duration, TraceId
FROM otel.otel_traces
WHERE Timestamp >= now() - INTERVAL 2 HOUR
  AND DeploymentEnvironment = 'production'
  AND ServiceName = 'backend'
  AND SpanAttributes['request.id'] = '<id>'
LIMIT 50

Time range first, then DeploymentEnvironment, then ServiceName or TraceId. Map lookups come after those. Duration is nanoseconds.