Temporary Councils Guide

​

Mental Model

​

Source-Owned Capabilities

• The source owner stays the authority. Tool, auth, realm, and filesystem boundaries remain those of the source execution context. The council request has no credential, auth override, or mutable session state field, so it cannot widen them.
• The capability is held by reference. The council persists the capability reference so a crashed coordinator can still revoke it at the owning runtime — including a HOST-owned capability whose record lives in a remote host’s store.
• No bearer material ever reaches a caller. Results and record projections carry non-secret provenance only: owner route, fork session, source session with its selected prefix length and prefix digest, granted scope, reuse policy, expiry, and a non-secret correlation hint. There is no wire field that could hold a capability id, bearer token, revocation id, or cleanup id.

​

Real Temporary Mobs

​

Bounds

​

Merge-Back Modes

​

Durability And Crash Cleanup

• durable requires a council store that actually survives a restart. If the serving runtime’s store is process-bound, the request is refused with capability_unavailable rather than silently promising crash recovery.
• process_bound is the explicit opt-in for tests, embedders, and ephemeral surfaces. It says out loud that a process death loses the record and the source capability TTL is the only remaining backstop.

​

Host Bootstrap

​

Idempotency

• Same id + same request ⇒ join the running execution, or replay its sealed result.
• Same id + a materially different request ⇒ refused as a conflict (duplicate_input), with both fingerprints in the error payload.
• A second coordinator may take a council over only after observing the claim lease expired; the takeover advances the claim epoch and fences the previous executor (stale_fence).

​

Agent Tool

• Detached (hosts that outlive the call). A host with a runtime adapter (the JSON-RPC, REST, and MCP servers, Meta), rkat run --keep-alive, and rkat mob deploy --surface rpc can deliver a completion after the call returns, provided the convener can be made live again to receive it: it is a mob member, or the host installs a DetachedOwnerHost (the JSON-RPC host does). There the call returns at once with {"status": "running", "council_id": ..., "job_id": ..., "note": ...}. When the council ends, its sealed outcome (result, cleanup, replayed) is recorded once in the convener’s transcript as a durable BackgroundJob system notice (the header “Background council job <job_id> finished (completed or failed):”, with the outcome as the typed block’s detail and persisted: true). It is committed once on the council’s custody record and submitted as a continuation, durable from submission, then applied as a runtime input with the idempotency key council:<job_id>: a convener in the middle of a turn sees it at that turn’s next model call, saved once and with no second turn (one follow-up turn only if the turn ends before another model call), and an idle convener runs one wake turn that sees it. A convener that is no longer live (the runtime retired its idle executor) is revived when the continuation is applied: through its mob when it is a mob member, or through the host’s DetachedOwnerHost when it is a plain session. The status is failed when the exit reason is a failure (participant_seating_failed, participant_fork_instruction_refused, participant_source_replaced, wiring_incomplete, exchange_failed, coordinator_interrupted); the record still carries the full sealed outcome, and bound-limited ends (max_exchanges_reached, deadline_exceeded) complete normally with their partial result. The notice survives later model calls, turns, and reloads and is readable through the convener’s session history.
• Blocking. The call waits for the council and returns the sealed outcome as its ordinary tool result, with blocked_because: host_declared_unavailable on a host that may exit when the turn ends (the rkat CLI unless it stays alive, such as rkat run without --keep-alive); no_runtime_adapter for a host that declares delivery without a runtime; no_continuation_owner for a host that never bound its runtime delivery inbox; or no_owner_revival_host for a convener that is a plain session on a host without a DetachedOwnerHost (a top-level REST, MCP-server, or rkat run --keep-alive session). No detached job is owed, so no result can be lost.

​

Example Tool Call

{
  "topic": "Is the storage migration safe to ship this week?",
  "participants": [
    {
      "mob_id": "design-team",
      "member_id": "author",
      "role": "author"
    },
    {
      "mob_id": "platform-team",
      "member_id": "reviewer",
      "role": "reviewer",
      "prefix_message_count": 12
    }
  ],
  "max_rounds": 2,
  "max_exchanges": 8,
  "max_result_bytes": 8192,
  "timeout_seconds": 300,
  "merge": {
    "policy": "summary",
    "finalizer": 1,
    "max_bytes": 4096
  }
}

​

Errors

​

Related

• Mobs Guide
• Mob Architecture