Would Herdr instead of tmux make sense? #16

New issue

Closed

opened 2026-07-29 16:52:14 +01:00 by pedromcaraujo · 1 comment

pedromcaraujo commented

2026-07-29 16:52:14 +01:00

Owner

No description provided.

pedromcaraujo commented

2026-07-29 17:09:36 +01:00

Author

Owner

Looked into this. Correcting the premise first: herdr has no browser UI. It's a
single Rust TUI binary — the author explicitly sells "no Electron, no hosted control
plane". Remote access is SSH, or herdr --remote streaming the TUI back to a local
terminal client. No nginx, no xterm.js, no PWA, no phone keybar, no upload dropzone.

So it's not a forge replacement. The real question is the one in the title: swap
herdr in for tmux, underneath forge's existing browser layer — ttyd would exec
herdr attach <channel> instead of tmux new-session -A.

Pros

Agent state detection comes free. herdr integration install claude gives
blocked/working/done/idle over a local socket API. That's exactly what abc8169
(#15) just built by scraping Claude's transcript JSONL — replaced by a supported
contract.
The socket API beats shelling out. web/tmux_ctl.py is 289 lines of tmux
subprocess calls for splits/zoom/kill/lazygit plus transcript parsing.
Already in nixpkgs (0.7.5), so packaging is a non-issue.
Agent-shaped primitives — pane metadata/titles, native agent resume, live
handoff. Closer to where forge would go if it ever wants N agents per repo.

Cons

It's a TUI, and we render TUIs through xterm.js. modules/shell.nix is 119
lines of tmux.conf tuned for exactly that (extended keys, OSC 52, mouse scroll,
Shift+Enter). All of it would need re-solving against herdr's own keybinding/theme
config, with no guarantee its sidebar renders sanely at phone width in an iframe.
herdr's sidebar duplicates forge's sidebar. Two channel lists in one viewport.
We'd run herdr with its sidebar off — throwing away its headline feature and using
it as a state API + multiplexer only.
Pre-1.0, ~5 months old. 0.7.x, no sandboxing, breaking changes likely. tmux is
the most boring dependency in this repo, which is a feature for something driven
from a phone on a train.
Not a small change. tmux touches all four nix modules plus tmux_ctl.py and
index.html — ~86 references across ~900 lines. It's a rewrite of the whole
session layer.

Decision — not now

The one genuine draw is the state API, and we just solved that ourselves in #15, and
it works. Rewriting the session layer to un-solve and re-solve our newest feature,
against a 0.7 dependency, on the layer that is hardest to debug from a phone, is a
bad trade.

There's no partial adoption either: herdr is the multiplexer, so tmux_ctl.py
can't just read its socket while tmux stays underneath. All-or-nothing.

Revisit when herdr reaches 1.0, or when we actually want multiple agents per
channel — that's the point where its model starts paying for the migration.

Refs: https://herdr.dev/docs/ · https://github.com/ogulcancelik/herdr

Looked into this. Correcting the premise first: **herdr has no browser UI.** It's a single Rust TUI binary — the author explicitly sells "no Electron, no hosted control plane". Remote access is SSH, or `herdr --remote` streaming the TUI back to a *local terminal client*. No nginx, no xterm.js, no PWA, no phone keybar, no upload dropzone. So it's not a forge replacement. The real question is the one in the title: swap herdr in for **tmux**, underneath forge's existing browser layer — ttyd would exec `herdr attach <channel>` instead of `tmux new-session -A`. ## Pros - **Agent state detection comes free.** `herdr integration install claude` gives blocked/working/done/idle over a local socket API. That's exactly what abc8169 (#15) just built by scraping Claude's transcript JSONL — replaced by a supported contract. - **The socket API beats shelling out.** `web/tmux_ctl.py` is 289 lines of `tmux` subprocess calls for splits/zoom/kill/lazygit plus transcript parsing. - **Already in nixpkgs** (0.7.5), so packaging is a non-issue. - **Agent-shaped primitives** — pane metadata/titles, native agent resume, live handoff. Closer to where forge would go if it ever wants N agents per repo. ## Cons - **It's a TUI, and we render TUIs through xterm.js.** `modules/shell.nix` is 119 lines of tmux.conf tuned for exactly that (extended keys, OSC 52, mouse scroll, Shift+Enter). All of it would need re-solving against herdr's own keybinding/theme config, with no guarantee its sidebar renders sanely at phone width in an iframe. - **herdr's sidebar duplicates forge's sidebar.** Two channel lists in one viewport. We'd run herdr with its sidebar off — throwing away its headline feature and using it as a state API + multiplexer only. - **Pre-1.0, ~5 months old.** 0.7.x, no sandboxing, breaking changes likely. tmux is the most boring dependency in this repo, which is a feature for something driven from a phone on a train. - **Not a small change.** tmux touches all four nix modules plus `tmux_ctl.py` and `index.html` — ~86 references across ~900 lines. It's a rewrite of the whole session layer. ## Decision — not now The one genuine draw is the state API, and we just solved that ourselves in #15, and it works. Rewriting the session layer to un-solve and re-solve our newest feature, against a 0.7 dependency, on the layer that is hardest to debug from a phone, is a bad trade. There's no partial adoption either: herdr *is* the multiplexer, so `tmux_ctl.py` can't just read its socket while tmux stays underneath. All-or-nothing. **Revisit when** herdr reaches 1.0, or when we actually want multiple agents per channel — that's the point where its model starts paying for the migration. Refs: https://herdr.dev/docs/ · https://github.com/ogulcancelik/herdr

pedromcaraujo closed this issue

2026-07-29 17:09:43 +01:00