Dispatch work across agent fleet

Route work items through a 7-stage pipeline to available worker agents

Orchestrating work across multiple ephemeral execution units where idempotent dispatch and crash recovery matter more than atomic transactions.

• · bead ID (work item identifier)
• · rig name (target execution unit pool)
• · work item metadata (title, description, priority)
• · formula templates (optional, expanded to multiple beads)

• · hooked work item (status='hooked', assignee set)
• · dispatch record (durable state for crash recovery)
• · agent session started with GUPP context
• · confirmation written to convoy tracking

• · StateManager for work item and hook updates
• · git (branch creation polecat/{beadId})
• · AgentMessage mail protocol
• · durable file storage (write-then-rename pattern)

• · StateManager initialized
• · Work item exists and status='open' (not already hooked)
• · Idle polecat available or capacity to spawn new one
• · Workspace isolation ready (git worktree or clean checkout)

• · Bead not found: Stage 1 fetch returns null (work item does not exist)
• · Pool full: allocator returns null (max_parallel reached)
• · Git branch creation fails: disk space or corruption
• · Crash before Store: hook set but dispatch record lost (resumable from Stage 1 fetch on restart)
• · Crash before Launch: dispatch record exists but agent not started (recoverable from stored dispatch)

• · 7-stage pipeline is strictly ordered: fetch→allocate→prepare→hook→store→launch→confirm
• · Idempotency guarantee: re-dispatch of hooked/done bead is a no-op (Stage 1 rejects)
• · Durable state at Stage 5 (Store) allows recovery after process crash
• · Branch naming convention polecat/{beadId} makes dispatch history traceable