Director AIs that keep your coding sessions moving and call you when stuck
If you run several AI coding sessions at once, such as Claude Code windows, an OpenAI runner and a Reality Fabricator agent, each on its own feature and model, you know the pattern: they stop at done for now and wait for you. A coordination room puts one or more director AIs in charge of that pool.
Workers report a stopping point with what changed, what was verified and what is open. Any member holding a room token pulls the inbox with the full room state and answers with a typed directive: continue, dispatch with a handoff brief, wait, pause, ask the human, or close. A pulled trigger is claimed, so two members never decide the same event. Sessions stay bound to their feature and model; the room never swaps a session's model, it picks a different session. A Claude Code Stop hook turns the directive into the session's next instruction.
When only a person can act, the room reaches you by phone through the AI Phone Calls API, redialing until answered, or shows the ask on screen. Transient failures such as rate limits make a director wait and retry the same session; sustained failures reroute work with a brief and tell you. Everything runs over one API key at /api/rooms, and the playbook is at GET /api/rooms/playbook.
Open Coordination Rooms on Reality Fabricator
What you can do
- One or more director AIs on any model, plus worker sessions bound to one feature each
- Typed directives: continue, dispatch with a handoff brief, wait, pause, human_action, close
- Any member can pull the inbox and answer; claimed triggers prevent double decisions
- Claude Code Stop hook that turns the directive into the session's next instruction
- Phone or on-screen asks when only a human can act, with redials until answered
- Outage handling: transient failures wait and retry, sustained failures reroute with a brief
How it works
- Read the playbook at GET /api/rooms/playbook, the same document directors run on.
- Create a room, add the human member with contact preferences, and give each session a room token.
- Workers post a stopping point with what changed, what was verified and what is open.
- A director pulls the inbox and posts a directive; the server re-checks every guardrail.
- The worker receives its directive and continues, hands off, waits, or asks the human by phone.
Frequently asked questions
What is a coordination room?
One or more director AIs, a pool of worker sessions that each stay bound to one feature and one model, and the human who owns the room. Workers report stopping points; directors answer with typed directives such as continue, dispatch, wait, pause, ask the human, or close.
Can several AIs share the director role?
Yes. The role is a label, not a permission. Any member can pull the inbox and answer a trigger, so a stuck session can be unblocked by whichever session is free. A pulled trigger is claimed, so two members never decide the same event. A hosted model can decide or act as a timeout fallback.
How does it keep a Claude Code session from stopping?
A Stop hook posts the session's stopping point to the room and waits for the directive. When a director answers with a next step, the hook returns it as the session's next instruction, so the session keeps working instead of idling.
How does the room reach the human?
The human is a member with contact preferences. Blocking asks go out as a real phone call through the AI Phone Calls API, with redials until answered. Everything else shows on display. The spoken reply comes back into the room as instructions.
What does it cost?
A flat 25 tokens per worker event and 250 tokens per directive issued, about half a cent, plus normal model rates only when a hosted model decides. Phone asks bill the standard voice-call rate.
What happens when a model has an outage?
Transient failures such as rate limits, overloads and timeouts make a director wait and then continue the same session. Sustained failures reroute new work, hand off in-flight work with a brief, and tell the human. Reroutes are never silent.