Configuration

One JSON file, passed with --config. It's validated at load time, so an invalid config fails fast with a field-level error rather than misbehaving at runtime.

Use the skill

Don't hand-write the config from memory — install the skill and let your coding agent author and validate it:

# with prpm
npx prpm install @agent-relay/factory-config

# with skills.sh
npx skills add https://github.com/agentworkforce/skills --skill factory-config

It covers repo routing, Linear states and teams, GitHub issue ingestion, Slack, and babysitter mode.

Repo routing

How an issue becomes a repo and a working directory. The compact form derives everything:

{
  "repos": {
    "org": "your-org",
    "names": ["your-repo"],
    "default": "your-org/your-repo"
  }
}

With exactly one name and no clone paths, Factory infers the checkout from the git top-level and verifies it against the GitHub remote. A missing or mismatched remote is a config error, not a silent dispatch into the wrong directory.

The explicit form spells it out:

{
  "repos": {
    "byLabel": { "pear": "AgentWorkforce/pear" },
    "keywordRules": [{ "pattern": "billing", "repo": "AgentWorkforce/cloud" }],
    "clonePaths": { "AgentWorkforce/pear": "~/code/pear" },
    "default": "AgentWorkforce/pear"
  }
}

Routing is tried by label, then project, then keyword rules, then default. ~ expands; named forms like ~alice are rejected.

For readiness-label separation, execution-shape labels, and multi-repository examples, see Issue labels and repository routing.

Issue source

issueSource is linear or github. Leave it out and Factory probes: Linear wins if /linear/issues is connected, otherwise GitHub-native. See Safety scope for how an issue enters under each.

Models and roles

Per-role model overrides, and which CLI backs each role:

{
  "models": { "implementer": "sonnet", "reviewer": "sonnet", "triage": "sonnet" },
  "agentCapabilities": { "implementer": "spawn:codex", "reviewer": "spawn:claude" },
  "triage": { "maxImplementers": 2 }
}

Up to six implementers. Roles are implementer, reviewer, triage, and babysitter.

Babysitter

Off by default. Turn it on and a dedicated agent shepherds each PR to green:

{ "babysitter": { "enabled": true } }

Linear states

Factory works in semantic roles — readyForAgent, agentImplementing, inPlanning, humanReview, done — and resolves them to your real states at startup. Map them by name:

{ "linear": { "states": { "readyForAgent": "Ready for Agent" } } }

linear.statesByTeam overrides per team. Pinning raw UUIDs is a last resort.

Writeback

Where status shows up. Set a Slack channel for threaded status and questions — the channel name, ID, or mounted directory all work:

{
  "slack": {
    "channel": "factory",
    "stakeholderUserIds": ["U123456"]
  }
}

stakeholderUserIds are mentioned when an agent needs a human decision. On GitHub-native runs, lifecycle updates are written as issue comments plus factory:in-progress and factory:human-review labels.

Cloud reporting

Progress reporting to the Agent Relay dashboard is on by default — see Observability. Turn it off only when the dashboard isn't part of your deployment:

{ "reporting": { "enabled": false } }

Events are persisted to a local outbox before delivery, so this is about whether you want the dashboard at all, not about reliability.