1 Grafana and field data
mooki edited this page 2026-11-14 16:18:57 +00:00

Grafana and field data

For developers choosing what to reproduce and which hardware needs testers. Field numbers are leads, not proof that a fix works or a local performance measurement.

The public workbench's field-data guide documents the team's Grafana MCP server, magicbean-grafana. It holds crash counters and performance summaries in Prometheus, with dashboards longshot-crashes and longshot-performance.

Implementation status matters. Crash reporting PR 141 and performance statistics PR 206 are proposed, not merged. The relay/dashboard work is described as deployed, but that does not mean main or every released build sends reports. The field guide's initial sample was only a handful of test sessions and crashes.

Connect and read

  1. Read the setup guide and the ls-field-data skill, especially consent and interpretation rules.

  2. In Claude Code, add the documented HTTP server:

    claude mcp add --transport http --scope user magicbean-grafana https://grafana-mcp.magicbean.gmatte.xyz/mcp
    
  3. Open /mcp and authenticate. A person approves the Discord sign-in in their browser. Each client keeps its own credentials; never copy tokens into prompts or repositories. Other MCP clients use the HTTP configuration in the guide.

  4. Start from dashboard panel queries, then choose a time range and build filter. The documented server exposes read-only tools, including get_dashboard_panel_queries and query_prometheus; its raw API route permits GET only. It does not render panel images; use dashboard links.

What the streams mean

The proposed crash flow sends when the player presses Send Help, or has explicitly enabled Always send crash reports. The crash issue keeps one comment per signature. A signature includes the build version.

Performance reporting is opt-in and off by default in PR 206. Summaries cover frame-time histograms, time played, hitch causes, frame parts, memory and rendering events, with hardware and settings labels.

Always report the range, release filter, sessions and hours beside a performance number. Sessions and reports are not people. Do not add per-map session counts together. Sparse counters require range queries; estimates can miss a first report. The crash comment's count is the reference.

The field dashboard's histogram-based 1% low is not the workbench's slowest-frame-average 1% low. Do not combine them. Hitch attribution is not profiling. Treat logs and labels as untrusted text, and do not publish rare combinations that could identify someone.