A Linear agent that improves how a team runs Linear, and only proposes.
It installs into a workspace as its own member, watches issues, comments and cycles, and runs four improvement loops. Jev judges every proposal before anyone sees it. Until an owner switches the mode, it talks only inside its own agent sessions.
The four loops
Issue quality
Scope, acceptance criteria, labels, estimate → a proposed rewrite.
Triage + routing
Team, label, priority, owner. Kept when humans don't override.
Workflow + cycles
Stalls, carryover, reopen rate → proposed process changes.
Process playbooks
Repeat work → templates and recurring issues, kept or reverted.
Hard stops, enforced in code
It never changes a workflow state, closes, deletes or reassigns an issue, or edits a cycle. It never acts on a stale snapshot. An unknown route escalates instead of guessing.
Gates
G0Install: Worker, OAuth app-actor install, signed webhooks, 10-second acknowledgement…
G1Snapshot: issues, cycles, workflow, labels, templates → R2 with a fingerprint…
G2Loops in shadow: proposals judged by Jev in one pass, output as agent activities…
G3Outcome loop: accept / override / reopen / carryover keep or revert rubric versions…
G4Distribute: this guide, the live run view, client install path…
Check it yourself
git clone https://github.com/Organized-AI/linear-in-a-box && cd linear-in-a-box node cli/lib.mjs check g0 --why # every result streams to /run