openfabriq
Factory

factory.yaml

The repository, the people the factory trusts, its limits, its harness and the services it talks to.

factory.yaml holds the settings of the whole factory. Only name, repository and maintainers are required. This page is the key reference; the harness explains setup, commands, checks, env, secrets and network in detail.

Example

openfabriq/factory.yaml
name: Task app factory
description: From issue to review-ready pull request; a human merges.   # optional
repository: acme/task-app
maintainers: [alice, bob]       # only their actions start the factory
model: sonnet                   # default model of the agents
timezone: Europe/Paris          # optional: the time zone of schedules (the machine's by default)
signature: false                # optional: leave out the line that says openfabriq wrote it
answer_mentions: false          # optional: the engine no longer answers @openfabriq on its pull requests and issues

limits:
  usd_per_run: 5                # maximum cost of one run (API-equivalent)
  usd_per_day: 30               # maximum per day, all automations together
  minutes_per_run: 45
  review_rounds: 3              # rounds of review before a person takes over

setup: pnpm install --frozen-lockfile            # prepare each checkout before the agent (optional)
commands:                       # optional: the project's commands, by name
  test: pnpm test
  lint: pnpm lint
  size: scripts/bundle-size.sh 500   # runs openfabriq/scripts/bundle-size.sh, which must exist
checks: [test, lint]            # optional: run by the engine before code leaves the machine

env:                            # optional: values for setup, commands and agents; never a secret
  NODE_ENV: test
secrets: [DATABASE_URL]         # optional: names; values stay on the engine's machine
network: [packages, api.stripe.com] # optional: the sites agents and commands reach; packages by default

integrations:                   # optional: services the engine knows by name
  slack: { channel: factory }
  linear: { team: ENG }
  sentry: { organization: acme, project: web-app }

webhooks:                       # optional: webhooks the relay receives for the factory
  deploy-failed: { signature: standard }

mcp:                            # optional: MCP servers agents may use
  sentry:
    url: https://mcp.sentry.dev/mcp
    secret: MCP_SENTRY_TOKEN

wiki:                           # optional: the wiki the engine keeps; wiki: true for the defaults (off without the key)
  folder: docs/wiki
  refresh: after_merge          # or manual (the default)

Keys

KeyPurposeDefault
nameHow the studio names the factoryRequired
descriptionHow the studio describes the factory
repositoryThe GitHub repository the factory works on, as owner/nameRequired
maintainersGitHub logins whose actions start automations, without @: validate refuses @aliceRequired
modelDefault model of the agents: sonnet, opus, haiku or a full model idsonnet
timezoneIANA time zone for schedule, such as Europe/Paris or UTCThe machine's
signaturefalse leaves out the line that says openfabriq wrote a comment, a review, an issue or a pull requesttrue
limitsCost and time limits
setupOne command, or a list, run in each checkout before the agent (setup)
commands, checksThe project's commands, which can run a script, and the checks among them
env, secretsEnvironment variables and secrets for commands
networkThe sites agents and commands reach: none, packages, any, or a list of sitespackages
integrationsSlack, Linear, Sentry
webhooksWebhooks received by the relay
mcpMCP servers
wikiThe wiki: true turns it on, in wiki/, on demand; or set its folder and refresh (manual or after_merge); false or no key keeps it offOff
answer_mentionsfalse turns off the answer to mentions: the agent that opened a pull request or an issue no longer answers a maintainer's @openfabriq on ittrue

Limits

LimitEffectDefault
usd_per_runA run stops when its API-equivalent cost reaches this amount5
usd_per_dayOnce the day's runs reach this amount, further runs are skipped until the next day. A run stopped before Claude Code reports its cost (by minutes_per_run or an error) counts its whole usd_per_run30
minutes_per_runA run stops after this duration, setup included45
review_roundsAfter this many rounds of review, the engine hands the pull request to a person with needs-human: rounds of the internal review (reviewed_by), or review and rework rounds on GitHub3

plan reports a raised limit as a security change.

How cost is counted

With a subscription, nothing is billed per token: the cost shown is what the same tokens would cost on the API, a measure of how much of your quota a run uses. With an API key, it estimates what you pay, at Anthropic's list prices.

GitHub's own limits

GitHub limits how much an account writes: about 80 pull requests, comments, labels, reviews or merges a minute, and 500 an hour. The engine writes one thing at a time, a second apart, whatever the number of runs. When GitHub still refuses, it waits a minute, then two, then four, and tries again; after that, the run fails with "GitHub's rate limit". This waiting does not count in minutes_per_run, so a factory that writes a lot, such as a site migration, slows down instead of failing.

On this page