darkroom

Evidence-based autonomous software delivery: agents build, sealed exams judge, and every claim ships with proof.

darkroom runs a convergence loop in which a builder agent works toward criteria it is never shown. A non-LLM harness drives the system under test as a black box and captures typed evidence (HTTP transcripts, command output, screenshots, screencasts); a judge scores that evidence against rubrics sealed in a vault the builder cannot address; gates ratchet so no scenario ships below its peak. The exam lives outside the repository, the scores live outside the builder's reach, and the manifest, not the code, is what you review, diff, and gate on.

This is proven live, not aspirational. A greenfield example app was delivered by agents under darkroom: its API, its security behaviors, its full UI and design language, even its Makefile, every behavior converging to a gated 100 on captured evidence, for single-digit dollars of metered spend per campaign, with the exams' screencasts as the receipts.

The name references the dark factory pattern (lights-off autonomous production) and the clean room pattern (independent implementation from specification). A darkroom is a controlled, light-sealed environment where evidence is developed and evaluated without contamination from the implementation side.

New here? Start with the Quickstart: four stages, ~15 minutes, covering first evidence, the exam, the convergence loop with zero spend, then real agents. Its runnable stages are executed by CI, so it cannot rot.

Install

pip install darkroom-ai

The core is dependency-free. Extras add capabilities:

The distribution is named darkroom-ai (the bare darkroom name is squatted on PyPI); the import name is darkroom throughout. The CLI installs as darkroom (alias: darkrm).

How it works

A tenant repository declares only run-me facts in darkroom.toml:

[project]
name = "quicknotes"

[commands]
serve = "make serve PORT={port}"
test = "darkroom drive"

The exam is data, held operator-side as one drive script per scenario. The engine boots the app fresh, executes the steps against it as a black box, and captures every exchange as evidence:

scenario = "note_saved"
record = true          # screencast the whole scenario

[[step]]
name = "save"
kind = "http"
method = "POST"
url = "{base_url}/notes"
json = { text = "first light" }
expect = { status = 201 }
save = { note_id = "$.id" }

[[step]]
name = "read_back"
url = "{base_url}/notes/{note_id}"
expect = { status = 200, body_contains = "first light" }

Step kinds cover HTTP, commands, Ed25519 keygen/signing, assertions, waits, container failure injection, and real browser interaction (goto/click/fill/screenshot, with viewport control and headless passkey ceremonies via a virtual authenticator). A failing scenario keeps its evidence: a failing scenario is still judgeable, which is the point.

darkroom verify checks each run against an evidence contract (per-scenario required kinds, counts, steps, trials), so "this build produced its proof" is a CI gate before any judging happens.

The loop, darkroom auto, runs assess → judge → build to convergence. Judge and builder can be shell hooks (see the Quickstart's zero-spend demo) or full agents configured operator-side:

# ~/.darkroom/projects/quicknotes/operator.toml (never in the tenant)
[judge]
model = "claude-opus-5"

[builder]
model = "claude-sonnet-5"
escalated_model = "claude-opus-5"

[loop]
max_iterations = 6

darkroom auto --scenario note_saved     # operator config discovered from the home

Authority lives in the per-project darkroom home (~/.darkroom/projects/<name>/, mode 700): operator config, drive scripts, loop state, and the vault of sealed rubrics, which the judge reads and the builder never can. darkroom vault seal moves rubrics out of the tenant; derive-contract regenerates the builder-safe contract from them; the OpenBao backend adds token-gated reads and server-side audit. The loop stagnation-escalates (diagnostic access, model escalation, sharper feedback), rolls back regressions to the best checkpoint, and ratchets evidence-gates.json on convergence. A red gate always means something real: rubric changes re-baseline; they never masquerade as regressions.

Every agent invocation is metered (model, tokens, cost) into loop state, and darkroom dossier assembles the cross-run record into one operator-facing bundle: score trajectories, checkpoints, escalations, gates, spend.

Evidence capture without the loop

The capture layer stands alone. Installing the package registers a pytest plugin; tests request the evidence fixture and runs become manifest-backed:

def test_login_flow(page, evidence):
    evidence.screenshot(page, "after_login", full_page=True)
    evidence.log("api_response", {"status": 200})

EVIDENCE_MODE=1 EVIDENCE_DIR=./evidence pytest

The same API is importable directly (EvidenceCapture, start_run/end_run), and darkroom gallery renders any run as a static contact sheet.

The formats are a spec

Every artifact (manifest, contract, evaluation, gates, tickets, drive scripts, role hooks, dossier) is a written, versioned format with conformance rules and a trust-boundary map: see docs/spec/. Any harness in any language can emit a darkroom manifest; any judge infrastructure can consume one.

Claude skills

Four skills ship inside the package, version-locked to the engine:

• darkroom-interview: the intent interview, covering charter, scenario enumeration, thresholds, rubric drafting under the capturable-evidence rule (with structured witness citations), gaming self-audit, preflight-validated artifacts.
• darkroom-converge: the operator's run-a-scenario loop — author the drive, audit, dry-run before spending, commit both repos, launch auto, triage blockers, regress, deploy.
• darkroom-status: the standing report — where every scenario's campaign stands now, blockers with their triage, what is shipped versus merely gated, the queue, spend — in one fixed shape, every number from darkroom status.
• darkroom-chronicle: narrates a project's delivery from its dossier, covering progression, what the judge witnessed, sticking points, novelties, spend.

Install them either way:

# as a Claude Code plugin, pinned to the release you are running
claude plugin marketplace add withtwoemms/darkroom
claude plugin install darkroom@darkroom          # -> darkroom:interview, :converge, :status, :chronicle

# or from the installed package
darkroom skills install                          # symlinks into ~/.claude/skills (--copy to copy)
darkroom skills check                            # exit 1 if any installed copy drifts from this engine

darkroom skills check compares every installed copy against the packaged one by content hash, so a stale or hand-edited skill is reported rather than quietly steering an operator wrong.

Documentation

• Quickstart -- first evidence to real agents in four stages
• docs/spec/ -- the format specifications (the handoff contracts, frozen at 1.0)
• ROADMAP.md -- milestones from foundation through v1 and beyond
• CHANGELOG.md -- the release-by-release record
• The Judge & Builder: implementing a dark factory -- the pattern essay: why evaluation must be structurally separated from implementation
• docs/vision.md -- design fiction: building a web app in the dark
• docs/rubric-lifecycle.md -- how a rubric is made, hardened, and revised
• docs/generalization-plan.md -- how darkroom absorbed the Judge-Builder framework

Development

make venv       # create virtualenv and install deps
make test-unit  # run unit tests
make test       # run all tests with coverage
make help       # see all targets

Licensed under Apache-2.0.