Skip to content

docs: add the agmsgd RFC for comment - #1105

Open
fujibee wants to merge 2 commits into
integration/terminal-driver-v1from
docs/agmsgd-rfc
Open

docs: add the agmsgd RFC for comment#1105
fujibee wants to merge 2 commits into
integration/terminal-driver-v1from
docs/agmsgd-rfc

Conversation

@fujibee

@fujibee fujibee commented Sep 9, 2026

Copy link
Copy Markdown
Owner

Publishes the agmsgd design for comment, ahead of implementation, as docs/design/agmsgd-rfc.md (with a Japanese translation alongside).

What it proposes

  • One resident daemon replacing the per-session watcher processes, the per-team sync engines, and the Codex helper stack.
  • noticed split from read — a notification can never consume a message; only an actual inbox fetch marks anything read.
  • A one-line notification instead of pushed full text, so an agent mid-task spends context on a message only when it decides to, with search (Gmail-style operators over subject/summary/body) as the other half of that trade.
  • Subscription carries identity, which makes claiming a role explicitly unnecessary: a session subscribes at startup from what it already knows, and the subscribed name is locked by default.
  • Control messages executed by the daemon, mechanically and by kind, instead of delivered to the target model to comply with.
  • Codex without the launch wrapper — hooks plus a poke into a named pane, neither of which changes how Codex starts.

Review focus

This is prose, so the useful review is factual rather than structural:

  • Anything asserted about current behavior that is wrong. Every figure and mechanism in it was measured rather than recalled, but measurements can still target the wrong thing.
  • Anything promised that the design does not actually deliver, or that a later ruling already reversed.
  • Places where a reader who does not already know the internals would be lost.

The six open questions at the end are the ones intended for outside readers once this is published for discussion; they are not blockers for merging the document.

Not in scope here

It is an RFC, not a decision record — settled decisions continue to land as ADRs under docs/adr/. Nothing in it changes code.

… control

Publishes the design for comment before implementation, in English and
Japanese. It proposes folding the per-session watchers, the per-team sync
engines and the Codex helper stack into one resident daemon; splitting
noticed from read so a notification can never consume a message; replacing
pushed full text with a one-line notification plus search; letting a
subscription carry identity so claiming a role explicitly is no longer a
separate step; and executing control messages in the daemon rather than
asking the receiving model to comply.

It is an RFC, not a decision record: settled decisions continue to land as
ADRs. Six open questions at the end mark where outside experience would
change the design.
A semicolon inside a sequenceDiagram message is a statement separator, so
"full text; receipt: read" ended the message early and left the rest
unparseable. Replace it with a comma, and drop the quotes and brackets from
the notify line for the same reason: they are structural characters in
mermaid, and the parser stops at the first one it cannot place.
@fujibee
fujibee changed the base branch from main to integration/terminal-driver-v1 September 10, 2026 09:04
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

1 participant