Every debug surface

Each panel below is live in the dashboard whenever the app runs with the devtools tag. Panels fail soft: if a single endpoint is unreachable, the others keep rendering.

App cards

Six cards at the top of the dashboard: Health (polled against /health every 5s), Port, Swagger URL (if docs/swagger.json exists), GraphQL URL (if gqlgen.yml exists), and Queue URL (if an asynqmon compose service is running).

Metrics

Parses the Prometheus text output from /metrics (emitted by pkg/observability) and surfaces: total requests served, in-flight count, and average latency. Refreshes every 5s via SSE.

Profiles

One-click links to every profile net/http/pprof exposes: CPU (30s capture), heap, goroutine, mutex, block, allocs, threadcreate, execution trace (5s), and the pprof index. Each opens in a new tab — point go tool pprof at the URL, or use the built-in HTML UI.

Goroutines

Parses /debug/pprof/goroutine?debug=2, groups goroutines by top-of-stack function, and sorts descending by count. A spike in net/http.(*conn).serve is normal under load; a spike in your own code usually means a leak.

Routes

Scraped from docs/swagger.json. Each row shows method, path, operation summary, request type, and response type — extracted from the OpenAPI 2.0 parameters[in=body].schema or OpenAPI 3.0 requestBody.content.schema, with the primary 2xx response picked automatically.

Services

Runtime state of every compose service attached to this dev session — queried via docker compose ps --format json every 5s. Shows name + health/state as colored pills.

N+1 findings

Grouping by (trace_id, normalized_template) across the SQL ring. Templates are normalized by collapsing string / numeric literals to ? and compressing whitespace. Any group with three or more hits surfaces as a finding, sorted by count descending.

Recent requests

The last 200 requests captured by the devtools middleware (ring buffer — oldest request is overwritten when the cap is hit, see Ring sizes to change the cap). Each row:

Time — wall clock
Method — colored badge (GET = cyan, POST = green, PATCH = amber, DELETE = red)
Path — URL path
Status — pill (2xx green, 3xx cyan, 4xx amber, 5xx red)
Duration — wall-clock milliseconds
Trace — CTA button that opens the trace detail card below the Traces list
Replay — opens the inline editor

An Export HAR button in the header downloads the current ring as HAR 1.2 JSON for Chrome DevTools, Insomnia, or any HAR-aware viewer.

Recent SQL

Every statement captured by the GORM callback, with wall-clock duration, rows affected, error (if any), and the trace ID that triggered it. Each SELECT row also has an EXPLAIN button.

Traces & trace waterfall

The trace ring holds the most recent 50 traces — older ones are rotated out when new traces arrive (see Ring sizes if 50 is too tight). Clicking a row (or the trace CTA in a request row) pins that trace and opens a detail card below the list. The card fetches /debug/traces/{id} on demand (the summary list endpoint returns traces without spans to keep SSE payloads cheap) and renders:

A waterfall of nested spans — controller → service → repository → SQL — with bar widths proportional to duration and offsets relative to the root span’s start
The span’s kind (server, internal, client) next to its name
A click-to-expand stack beside each span name showing up to 20 call frames captured at span start via runtime.Callers
OTel events attached to the span (error markers, custom annotations)

Two tabs inside the card:

Waterfall (default) — the span tree
Logs — on-demand fetch of every slog record emitted during this request (pulled from /debug/logs?trace_id=…)

The detail card lives in its own DOM surface, so 5s SSE refreshes don’t collapse whatever waterfall you’re reading. A pinned-row banner above the Traces list keeps the inspected trace visible even if its row rotates out of view.

See Tracing & Waterfalls for how to get a multi-bar waterfall in your own code — and how to turn SQL into waterfall bars.

Exceptions

devtools.Recovery wraps pkg/middleware’s recovery middleware. Every panic is recorded with the recovered value, a 20-frame stack, method + path of the offending request, and the trace ID. The table shows all recoveries this session with their stacks rendered as expandable <pre> blocks.

Cache ops

devtools.WrapCache decorates cache.CacheService. Every Get/Set/Delete/Flush/Ping records an entry with op, key, hit/miss flag (for Get), duration, error, and trace ID. Per-trace aggregation answers “did this page actually use the cache?” at a glance.