PROTOCOL

Frozen at v0.2. Implementations may add platforms. They may not add curiosity.

1. Mapping

A mapping is a markdown document: YAML front matter + one table. It is the only config.

Source: a filesystem path or an https:// URL. The agent does not fetch anything else.

Required front matter: workspace, platform (slack | discord), scope, agent. Optional: collect_seconds (default 45).

Required columns: id, name, role, direction.

name is documentation. Routing uses id only.

2. Roles

A map without a management in/both row, or without a reports out/both row, is invalid.

3. The wall

ScopeGuard.allow(id) is called before every read and every write.

• Unknown ids are dropped before the adapter sees the event.
• Logs and exceptions substitute <outside-scope>. The raw id is not interpolated.
• Reports may only name ids that allow returns true for.

A second workspace is a second process with a second file. Not a smarter prompt.

4. Relay, then report

1. Event arrives from a management channel.
2. Guard checks the source id.
3. Fan-out to every room in the map, excluding the source.
4. Collect window gathers replies from those rooms only.
5. One report, posted to every reports channel.

Report shape (v0.2):

# report from {agent}
scope: {scope}
broadcast: {id}
rooms: {n}
replies: {n}
window: {seconds}s
filed: {n} reports channel(s)

5. Reload

SIGHUP re-reads the mapping source and rebuilds the wall. No restart. No drain of in-flight collect windows — they finish against the map they started with.

6. Failure

Fail closed. A parse error, a missing token, an unknown platform — exit non-zero. Do not run half a map.