Documentation

Use AI to need less AI.

These docs cover two parts of Coldworks: Reviews, the reviewer, and the audit, which reads the traces your agents already produce. Each part says what works today and what does not exist yet.

ReviewsRisk-routed code review for the agent era. It scores every pull request, clears the majority, and routes the risky few to a human.Read the section →The auditReads your trace exports on your machine and finds the judgments your agents keep re-buying. A design preview: not installable yet.Read the section →

Reviews

Coldworks reviews pull requests: risk-routed code review for the agent era. It scores every pull request, clears the majority, and routes the risky few to a human — with evidence attached. Every merge starts a clock against this repo’s reverts. It never blocks a merge.

Watch Coldworks score its own pull requests on the queue, read the scoreboard (adjudicated and pending counters tick there now; miss_rate stays null until the pre-registered interval fires, and a count is not a rate). The GitHub App is dogfooding on this repository and is not a self-serve product for other orgs yet. The self-serve measurement tool is still the backtest CLI: replay your repo’s merged history, label defect-inducing PRs from revert anchors in git, and measure which PRs Coldworks would have routed. The report is the demo — and the same numbers we publish for ourselves.

QuickstartBacktest any public repository in one command — no account, no server.Start here →What Coldworks gets wrongColdworks reviews its own pull requests, and the team logs what it got wrong.Honesty →CLI · doug-backtestEvery flag, with examples, for the CLI that replays a repo’s merged history.Reference →

Early preview. Coldworks is in active development. Everything marked Available works today; things marked Preview or Planned are described so you can see where this is going — not to pretend they exist.

The shape of Reviews

# 1 · replay history, honestly
doug-backtest your-org/your-repo

# 2 · route the live queue
62 open → 5 need you · rest cleared w/ receipts

# 3 · ask before you write             [planned]
coldworks.ask("backfill NOT NULL on a hot table")

The audit

Design preview · coldworks-audit has not shipped

Coldworks reads the traces your agents already produce, finds the judgments they keep re-buying, and compiles the repeats into deterministic code you own and you run. These pages cover the part you touch: getting your traces in, running the audit, and reading what it finds.

QuickstartOne command, no signup, no install step. See a full audit on sample data in under a minute.Start here →Connect your tracesLangfuse, LangSmith, or OpenTelemetry. A batch export is a file transfer, not an integration.Three steps →The audit CLIEvery command, flag, and file the audit reads or writes — and what never leaves your machine.Reference →

The shape of it

Everything below runs on your machine. For the audit, Coldworks hosts one thing: a read-only registry page for figures you choose to publish. That page cannot receive your traces, your documents, or your code — by architecture, not by policy.

The whole pipeline

your trace store  ──export──▶  a file on your disk  ──coldworks-audit──▶  a report on your disk

nothing crosses the boundary. the report is yours to forward, or not.

What is real right now

These pages hold themselves to the same rule as the audit: states are answers.

Claim	State
The connection flow in these pages is the front-door design under review.	DESIGNED
The recovery engine behind the audit exists and is deterministic.	BUILT · UNRELEASED
`coldworks-audit` is installable.	NOT YET
Validated against a real Langfuse / LangSmith export.	GATED · NEXT STEP
Anything in the audit uploads your data.	NO — BY DESIGN