Loop Engineering Team is a prototype ticket/spec-to-job adapter on top of AI Dev Team. It is disabled by default and intentionally narrower than the longer-term design recorded in the types and configuration.
- AI Team remains the Slack/chat/team-context product.
- AI Dev Team remains the supervised coding-agent harness.
- Loop Engineering Team can bind supplied GitHub issue context to a normal AI Dev Team job. It does not currently discover or update GitHub Project fields itself.
GitHub Issue / Project
-> AI Team triage and spec drafting
-> Ready for Dev project status
-> AI Dev Team /internal/loop-runs
-> loop.source / loop.spec / loop.executor events
-> normal AI Dev Team job pipeline
-> verification and policy-gated PR publishing
-> normal AI Dev Team result and memory events
AI Team owns conversation, stakeholder updates, triage and spec drafting. In this
release AI Dev Team accepts the resulting ticket context through
/internal/loop-runs, records Loop metadata, and runs only the existing internal
planner/orchestrator/reviewer pipeline.
Use GitHub Issues as the durable ticket and, if desired, Projects as an
operator-managed board. Configure loopEngineering.github.allowedRepos as a
fail-closed repository allow list; an empty list permits no repositories. Status names and Project fields are retained as
integration metadata; this release does not call the Projects API to discover or
change them.
Inbox, Triage, Needs Spec, Spec Review, Ready for Dev, Claimed by Agent, In Progress, PR Open, CI/Review Repair, Human Review, Ready to Merge, Done, Blocked, Won't Do.
An upstream service may supply fields such as Risk, Spec Level, Agent Profile, Human Gate, and acceptance criteria. They are not discovered by AI Dev
Team itself.
none: the issue body is enough for a small low-risk change.mini: the issue or a comment includes Goal, Acceptance Criteria, Verification and Risk.full: the target repo containsspecs/0123-short-slug/withbrief.md,requirements.md,design.md,tasks.md,acceptance.mdanddecisions.md.
High-risk or dangerous work requires an approved spec before development.
Only internal is integrated: it delegates to the normal AI Dev Team job
pipeline. The mock, external-command, codex-sdk, and claude-agent-sdk
values are reserved adapter boundaries; Loop run creation fails closed when one
is selected. Their core classes remain design scaffolding, not a supported
execution path.
The only implemented Loop human gate is before-dev. before-pr and
before-merge are reserved values and are rejected rather than silently ignored.
PR publication is independently governed by the global fail-closed
externalActions policy. Provider credentials must come from the runner
environment, not specs or issue bodies.
Internal Loop runs consume repository AGENTS.md and CLAUDE.md through the
existing project-memory path. Repo-instruction stewardship settings are reserved
for integration by an upstream service. The Loop-specific context provider
settings are also reserved; they do not add a second context service in this
release. Default repository posture:
AGENTS.mdis a normal file with install/test/build/lint/typecheck commands.CLAUDE.mdimports@AGENTS.md.- Dangerous areas are named: auth, billing, secrets, migrations, deploy, Kubernetes, Terraform, Ansible and privacy/data deletion.
For an evaluation repository, start with loopEngineering.enabled: true,
allowedRepos restricted to that repository, executor: internal, and
humanGate: before-dev. Submit a ticket only after the issue body, acceptance
criteria, verification steps, and any required approved spec are present. Keep
the global externalActions policy approval-gated until the whole flow has been
tested in a disposable repository.