COLDWORKS/ docs / audit / quickstart
The audit · Getting startedDesign preview · coldworks-audit has not shipped

Quickstart

First audit in under a minute. There is nothing to install and no account to make — the only prerequisite is uv, which most Python developers already carry.

What to download

Nothing, in the usual sense. uvx fetches and runs the audit in one step, in an isolated environment it cleans up itself. If you prefer a persistent install:

Either works
$ uvx coldworks-audit --help          # run without installing
$ uv tool install coldworks-audit     # or keep it on your PATH
No API key, no signup, no Docker, no database. The audit is a program that reads a file on your disk and writes a report next to it.

See it work before you export anything

Empty-handed? The package writes a sample span export itself, so the first report costs you nothing but the command:

Sample run
$ uvx coldworks-audit run --sample --open

This renders the full report — headline, repeat groups, nondeterminism findings, loss populations — on data we wrote, labeled as such on every screen. The release-1 sample is span-only; documents arrive with --docs in release 2.

Your first real audit

  1. Export traces from wherever they live — takes about 30 seconds. See Connect your traces for the Langfuse, LangSmith, and OpenTelemetry paths.
  2. Run the audit on the file:
First real run
$ uvx coldworks-audit run ~/Downloads/traces.jsonl --open
  1. Read the headline. The share of your agent's context that sits in exact repeats of a judgment it already made — in your own tool vocabulary. Repeats count as cache candidates only when their results agree; disagreeing repeats are reported separately as nondeterminism findings. The report opens with one sentence: “This report is the first step of an audit, not a dashboard. Where your platform already prints this number, start from that number; what this adds is the groups that disagree and the populations it could not read.”

If the audit can't read your export

It doesn't report a silent zero. It prints a field census — "your spans carry tool names but not structured arguments" — so even a failed audit hands you a concrete, forwardable finding about your own telemetry.

Where to go next

  • Connect your traces — the per-source paths, and how to make the audit standing instead of one-off.
  • The audit CLI — every flag, and the exact list of what never leaves your machine.