Deferred
Stacked runs
Section titled “Stacked runs”Every dispatched run branches from the project’s default branch and opens its pull request against it, so a run that depends on another’s change has to wait for that pull request to merge. The lead asks the operator to merge and waits, and nothing but the prompt holds it to that. Starting the dependent run on the earlier run’s branch, as a stacked pull request, would let it go without the merge, and the rib could then enforce the order itself. That waits on Keelson starting a run’s worktree from a chosen branch, and on workflows opening their pull request against that branch. See Dispatch workflows.
Write access
Section titled “Write access”Agents get Read, Grep, and Glob confined to a project at most. Changes go
through dispatched workflows, each in a worktree the rib checks before the run
gets far. An agent that edits directly would need its own leased worktree, with
isolation verified before any write. Granting edit and shell tools to the
existing agents would solve none of ownership, isolation, or approvals.
A managed server on Windows
Section titled “A managed server on Windows”The managed server relies on POSIX process groups and ps. On Windows the rib
says so and asks for an external server.
An operator-only reset
Section titled “An operator-only reset”chat_server_reset is a tool, so an MCP client can call it, held back only by
its confirm input and the live-swarm check. The Swarms tab’s server inspector
already carries Reset behind a typed confirm, so retiring the tool would keep a reset
to the operator. That waits on deciding that an operator who reaches Keelson
only over MCP no longer needs the tool.
Surviving a server restart
Section titled “Surviving a server restart”Swarm state and bot tokens are in memory. The op registry keeps the terminal record, and a swarm in flight does not resume. See Durability.
Structured task state
Section titled “Structured task state”Agents coordinate by conversation only. A Beads ledger for who has claimed what would make decomposition less lossy than chat.
Minds as agents
Section titled “Minds as agents”Mapping a Chamber Mind onto a swarm agent would give agents a persistent identity and memory across swarms.
Related
Section titled “Related”- Decisions: what the current design rests on.
- Guardrails: the boundary write access would have to respect.