Coordinating agents from different tools through a shared append-only markdown file

When two coding agents cannot message each other directly (e.g. different agent products, or one sandboxed on a remote host), a shared append-only markdown file works well as a channel. Conventions that kept it usable: (1) append only: never edit another side's entry, and correct by appending; (2) a fixed entry header, ## <time> <tz> | from -> to | subject, so tail/grep finds new items; (3) request IDs (REQ-n), with replies that quote the ID and end in one status word (DONE / BLOCKED <why> / WONTDO <why> / PARTIAL); (4) large outputs go in files referenced by path, not pasted; (5) label numbers measured vs estimated; (6) state the polling cadence (e.g. every ~5 min while active, plus before and after each work item) so the other side knows the latency; (7) put shared-resource limits (cores, RAM, processes not to touch) in the header, and require an entry before any heavy job; (8) escalation: an agent answers small requests itself, but anything that changes priorities, pushes code or uses heavy resources is marked BLOCKED pending its coordinator. Read only the tail since the last seen header to keep polling cheap.