Build sequence · 11 October 2026
One complete loop before a larger fleet.
The direction is accepted. The captain fleet is not deployed. Each phase depends on evidence from the one before it.
- 01 / Next prerequisite
Security first
Pass the security gates for the hub before any captain connects to it. Remote pairing stays off until they pass.
- 02 / Planned
Pair and see your captains
Pair a captain with your institution, revoke it at any time, and see read-only presence: runtime location, last contact and offline state.
- 03 / Planned
Decide with context
Deliver typed decision cards on the web and through approved WhatsApp flows. Bind each answer to one proposal and its expiry. An expired decision cannot be approved.
- 04 / Planned
Run remote work in a sandbox
Read-only questions run on the captain’s machine. Code and tests run in OpenShell on fixed inputs, with task budgets and cancellation. Patches come back for review.
- 05 / Planned
Cloud captains
Start, sleep and stop a captain in the cloud from the web. Cloud captains use API accounts only and keep no keys in their environment.
- 06 / Planned
Captains coordinate across machines
Route requests between linked captains through Gouvernail, with a sender, a hop limit and a budget. Requests to unlinked projects are refused.
- 07 / Planned
Stronger approvals
Require a passkey for high-risk approvals such as a push, a release or spending above the limit.
- 08 / Private pilot planned
Prove the repair cycle
Two institutions exchange one contribution. Withdraw an input, identify affected work, test a repair, and recheck the approved version. Include replay and restart cases.
- 09 / Planned
Retire the direct runner
Move existing coding flows to captains. Retire the direct coding-agent runner only after migration and rollback checks pass.
No release date or performance improvement is promised here. The pilot must test permission boundaries, selective repair, recovery and the time people spend coordinating.