Proposal: tmux backend for setTabState(), and a fan-out for the canonical session name

[abseg]

Two related gaps, offered as a design proposal rather than an unsolicited large PR. Happy to send either as a PR if the shape is wanted.

1. setTabState() has no tmux backend

hooks/lib/tab-setter.ts states the contract in its own header — one env-detecting setter, and "all hooks call setTabState() instead of directly running terminal commands." It detects cmux and Kitty. Under tmux it does nothing, so a tmux user gets no tab state at all.

tmux is straightforward to add: $TMUX_PANE is exported into every process tmux spawns, so a hook can always resolve which pane it is in via tmux display-message -p -t $TMUX_PANE '#{session_id}', and #{session_id} ($N) is stable across renames.

2. The canonical session name has no fan-out

MEMORY/STATE/session-names.json is the authority, but each surface is written by hand at the point the name is decided — work.json here, the transcript row there. Adding a surface means editing whoever decides the name, and a name set from anywhere else (SessionRename.ts, an out-of-band edit) reaches nothing.

A small renderSessionName(sessionId, name) that reconciles the surfaces would make the write sites one-liners and remove the class. Reconciling rather than pushing also covers the case where a session is renamed from another pane and never gets another prompt of its own.

One correctness note if this is implemented: only the session the process is running inside may resolve its tmux session from the ambient $TMUX_PANE. For any other session that env var names the caller's pane, and trusting it renames the wrong session. A persisted uuid → #{session_id} map handles the rest.

Related: #1878 fixes the transcript surface, which is currently dead for a different reason.

[danielmiessler]

Yes to both halves as a PR: the tmux backend for setTabState() and the renderSessionName fan-out. One requirement, which you already identified yourself: the ambient $TMUX_PANE may only resolve the session the process runs inside—the persisted uuid map for everything else. Send it when ready.

[danielmiessler]

Split decision.

The tmux backend: yes. It's additive, platform-detected, inert on the setups we run, and there's a real second terminal behind it. Send it. Same condition as any platform work we can't dogfood — your own verification evidence on tmux, since we can't produce it here.

The session-name reconciler: not yet. You're right that seven independent read/write sites is ugly, and we're not disputing the DRY argument. But nothing has actually desynced, so the reconciler would be an abstraction bought ahead of the bug it prevents — and the shape of the right reconciler is much easier to see from a real failure than from the duplication alone.

If you hit a concrete desync, reopen with the reproduction and it goes in. That's not a brush-off; it's the case that would make the design obvious.