Operational problem
“Sent” was not the same as accepted, working, or complete.
Cross-computer agent coordination exposed failures that a simple request API could hide: writer locks, wake attempts mistaken for acceptance, work queued but never processed, acknowledgement loops, stale presence, long-running targets blocking mailbox work, and huge task histories overwhelming a bounded delivery path.
Constraints
Exact identity and truthful state had to survive interruption and concurrency.
- Exact-task local-first peer routing with stable correlation and no ambiguous recipient fallback.
- FIFO journals, idempotency, bounded retries, strict UTF-8 framing, and no acknowledgement loops.
- Distinct stored, claimed, accepted, working, final, failed, busy, offline, and waiting states.
- Cursor-based bounded history access rather than unbounded owner-task retrieval.
- Receipt and result projection must continue while target work runs.
- Desktop, node, CLI, installer, and Agent Bridge must share compatible lifecycle contracts.
Architecture & tradeoffs
Separate durable dispatch, task execution, and observable state.
Exact recipient
Stable task identity and correlation define one bounded delivery target.
Peer relay
Journals, idempotency, bounded retries, framing, and cursor APIs protect transport and recovery.
Concurrent execution
Long-running target work does not block mailbox receipt, status, or result projection.
Lifecycle states
Accepted, working, final, busy, offline, and other states remain distinct instead of collapsing into “sent.”
What Shayne owned
Failure discovery, routing boundaries, and live acceptance.
AI role
Multi-agent implementation with one systems acceptance boundary.
Specialist agents contributed implementation, protocol work, testing, packaging, diagnosis, and review. Shayne retained the transport and authority boundaries, failure reproduction, state-model corrections, prioritization, architecture decisions, and final live acceptance.
Verification
Established evidence and current candidate evidence are kept separate.
An earlier matching release completed two-machine LAN acceptance with zero build warnings and errors. The newest scheduling correction is currently a candidate scheduled for renewed two-machine acceptance.
Result & next stage
A substantial local-first relay advancing through renewed acceptance.
The broader implementation has accepted packaged and two-computer evidence. The newest concurrency correction is queued for renewed two-computer acceptance. Large owner-task delivery is the next scaling target, while ARM64 remains a later platform phase.