What problem does it solve? When orchestrating multiple AI agent workers (claude/codex) across cmux workspaces, there is no obvious push channel for workers to report results back. Polling terminal screens with read-screen is fragile, and cmux has no free-form emit RPC. This Skill shows how to use the cmux event bus for real worker-to-worker and worker-to-orchestrator messaging with replay and resume. ## Core Features & Use Cases - Pub/sub over cmux log: Publish base64-encoded messages via cmux log (which emits sidebar.log.appended events) and subscribe with cmux events filters, decoding payloads by topic tag. - Claim-check pattern for large payloads: Work around the 16 KiB frame cap by writing results to files and publishing only id/status/file references. - Race-condition handling: Avoid the idle-bus hang on cmux events --limit and prevent lost wakeups by capturing latest_seq before spawning workers and subscribing with --after <seq>. - Use Case: Fan out five codex workers across cmux workspaces, have each publish a completion message with its result file path, and have one orchestrator subscriber collect, dedupe, and read all results event-driven. ## Quick Start Ask the agent to set up cmux event bus messaging so your workers publish completion signals and the orchestrator subscribes with replay from a captured sequence number.
