[!IMPORTANT]
This repository is retired, as of 2026-10-19. On retirement it is renamed
coldworkshq/doug-archive and archived, so it is read-only. Work continues in
the new private repository coldworkshq/coldworks. In this repository, every
reference to coldworkshq/coldworks or coldworks#N means
coldworkshq/coldworks-archive, the repository that held that name before the
retirement.
Coldworks works out which ones do.
Every AI code reviewer on the market runs a language model over every diff. That makes their cost scale with the exact thing coding agents are inflating, and it still leaves a person reading bot comments on 100% of pull requests. Coldworks inverts the attention: most PRs clear, and only the small fraction that carries real risk gets a human. When the reader is on, that routing verdict still comes from an LLM reading the diff — the deterministic rules are the labeled fallback, not a claim that production is model-free.
This repository is the pull request reviewer in Coldworks, called Reviews on coldworks.dev. It keeps its codename, doug, in names you type: the repository, the Python package, the doug-backtest command, and the DOUG_* settings.
Three rules
Route, never block. The PR proceeds either way. Coldworks only decides who has to look. Tools that block get disabled.
Never write code, never open a PR. The moment it authors, it owns the authorship.
Publish the miss rate. Every quarter, including the incidents that came from PRs it cleared. A gate that never publishes its errors is a marketing claim; one that does can survive being wrong. The scoreboard is live; the miss-rate number is not yet — miss_rate stays null until enough adjudications exist.
What it looks at
Live scoring (production sets DOUG_READER=1) is an LLM diff-read, unless a repository turns its deep read off on /dashboard/settings — that setting narrows, never widens. The deterministic fallback looks at cheap structural signals, including:
- Sensitive paths (auth, billing, security) and schema migrations in the same diff
- Runtime dependency bumps with no test delta
- Fast single-reviewer approval on a large diff
- Sizeable changes with no test files touched
- Bot / agent authorship
- Static hotspot path segments (historically high-revert areas)
Diff size is deliberately de-weighted. It predicts poorly once the rest are controlled for. Rolling-window hotspot learning and several shape rules exist in the backtest CLI only, not in the live App path.
Repo layout
api/ — Python 3.14 + FastAPI, managed with uv. doug/reader.py (LLM scoring when enabled), doug/worker.py (drain + check run), doug/check_run.py, doug/features.py + doug/scoring.py (deterministic fallback). POST /webhooks/github verifies, records, and enqueues a review job, then 202s. GET /v1/queue is the token-gated ledger queue; GET /v1/showcase/queue and GET /v1/showcase/scoreboard are the public dogfood surfaces.web/ — Next.js 16 + Tailwind 4 + shadcn/ui. Landing, /queue, /docs, /scoreboard, WorkOS sign-in, and the signed-in /dashboard (ledger, repositories, and /dashboard/settings — flag line, PR comment, deep read, per repository). Public pages fetch /v1/showcase/* when the API is up, and fall back to a bundled fixture when it isn't.console/ — Operator console (Next.js). Shares the root npm workspaces lockfile with web/.
make dev # API on :8000, web on :3000
make test # API + Node workspace tests
make lint # ruff + eslint
npm ci # install web + console from the root lockfile
Services ship Dockerfiles built for Cloud Run (respect PORT, scale to zero). The API still uses an app-dir build context; Node apps must be built from the monorepo root because of npm workspaces:
gcloud run deploy doug-api --source api
docker build -f web/Dockerfile -t doug-web .
docker build -f console/Dockerfile -t doug-console .
# Prod path: api/deploy/gcp.sh web|console builds via Cloud Build, then
# gcloud run deploy --image …
#
# Before the first Node image deploy, create the Artifact Registry repo
# (or re-run setup): PROJECT=doug-prod0 ./api/deploy/gcp.sh setup
#
# console is IAM-gated and not in CI — after a root lockfile / Dockerfile
# change, rebuild it with: PROJECT=doug-prod0 ./api/deploy/gcp.sh console
Status
Early preview, dogfooding on this repository. The reviewer runs as a GitHub App, installed as dougs-review: webhook ingest, a durable worker, a neutral check run on every PR (ADR-0010). Merge to main deploys API + web (ADR-0009). It is not a self-serve product for other orgs yet.
The LLM reader is the scoring path when enabled (ADR-0004). The 2026-07 probe's AUC 0.69 / 0.67 is that probe, not a measurement of the shipped 100k-char reader (ADR-0012). Deterministic rule weights remain priors.
Not yet: outside installs, first published miss rate, Pattern Garden MCP server.
The self-serve measurement tool is still doug-backtest: replay a public repo's merged history, label defect-inducing changes via revert anchors, and plot capture against a size-only baseline.
License
FSL-1.1-ALv2 — Functional Source License. Read it, run it internally, modify it, learn from it. The one thing you can't do is ship it as a competing commercial product or service. It converts to Apache-2.0 automatically two years after each version is published.
Source-available rather than open source, and deliberately so: the point is freedom without free-riding, not enclosure.
