agents

What each agent tracks

8 agents, one wire format. They write byte-identical sketches and normalize a statement to the same shape, so a service split across two languages is one history rather than two. What differs is what each runtime allows.

Why the gaps are where they are

• CPU and waiting are Python only. Splitting an operation’s own time into the half that was computing and the half that was queueing needs a per-thread clock, and a span that owns its thread for its whole duration. time.thread_time() gives both. Node’s process.cpuUsage() is process-wide with many async contexts interleaved on one thread; Go’s goroutines migrate between threads. Neither could report a number that was true.
• Query plans are Postgres, on Node, Ruby and PHP. A generic-plan EXPLAIN, cached per statement and issued outside the caller’s span so it is never measured as their work. MySQL and MongoDB can only explain a query that still has its values in it, so capturing a plan there would mean retaining somebody’s data to compose a command out of it. The same refusal in every agent.
• Round trips are MongoDB only. The driver fetches the first 101 documents and then as much as fits in 16MB per trip, so the waiting grows with bytes rather than with documents. No SQL driver exposes anything equivalent.
• Everything else is universal by construction. The measurement core is shared and the collector-side detections — repeated queries, over-fetching, unbounded reads, recursion — read data every server agent already sends. An agent does not opt into them.

A client agent is an addition rather than an alternative. It sees the half of a request that happens after the server span ends — and, in an app with no server of its own, it is the only thing that sees anything at all.

Found by reading the code

These cannot be measured. A handler that was never attached emits nothing, and an effect that re-runs forever is a property of a dependency array. So they come from the build transform reading the source, and they are ranked by how much production traffic the file they live in actually serves — which is the reason they are here rather than in a linter. ESLint cannot tell a dead button in your checkout from one in a page nobody loads.

• dead_interaction — does nothing: users click this and nothing observable happens
• effect_loop — effect loop: an effect writes state it also depends on — found by reading the code, not by measuring
• missing_cleanup — never cleaned up: a subscription, interval or listener is created without a teardown
• exposed_secret — exposed key: A password-like key is written into code that gets downloaded by everyone who visits the site. Anyone who looks can copy it and use it as you.
• insecure_transport — unencrypted: This talks to another service over a plain, unencrypted connection. Anything travelling on it can be read or changed along the way.

Every agent