1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
|
#+TITLE: Design: ai-vterm — in-Emacs Claude launcher
#+AUTHOR: Craig Jennings
#+DATE: 2026-05-07
#+OPTIONS: toc:nil num:nil
* Status
Draft.
* Problem
Claude Code currently launches outside Emacs via the =ai= shell script, which builds candidate projects from =~/.emacs.d=, =~/code/*=, =~/projects/*= (anything with =.ai/protocols.org=), opens each in a tmux window, and runs =claude "Read .ai/protocols.org and follow all instructions."= per project. The shell-out pulls focus to a terminal, and tmux's horizontal split is the wrong shape for a code-on-left, Claude-on-right reading layout.
The in-Emacs alternative today is =vterm-toggle= at F12, which uses a horizontal bottom split via =display-buffer-at-bottom=. No project picker, no per-project session model, no vertical split.
Building this in Emacs eliminates the context switch and gives the side-by-side layout that matches how the work is actually read.
* Non-Goals
- Replicating the =ai= script's git prep / auto-pull. Phase A.0 of the startup workflow already handles pulls at session start.
- A multi-project session switcher with its own UI. =consult-buffer= and the buffer list already navigate between =claude [...]= buffers.
- Replacing =vterm-toggle= at F12. The existing bottom-split flow stays for non-AI shells.
- Tab-bar or frame-per-project layouts.
- Auto-launching tmux inside the AI vterm. Claude under tmux adds a session-management layer for no benefit here.
* Approaches Considered
** Recommended: wrap =vterm= directly with per-project named buffers + a parallel display rule
A new module =modules/ai-vterm.el= adds a command that picks a Claude-template project, opens (or reuses) a vterm buffer named =claude [<basename>]=, and lets a new =display-buffer-alist= entry route any buffer matching that prefix to a right-side window. Multiple projects produce multiple coexisting buffers, all sharing the same right-side slot. Switching among them is a buffer-switch, not a kill-and-recreate.
Pros:
- Same package (=vterm=) as the existing config.
- Per-project buffers run simultaneously without conflict.
- Right-side placement is one =display-buffer-alist= entry.
- Existing windmove (Shift-arrows) handles code↔Claude focus toggling. =buffer-move= (C-M-arrows) handles side-swap. Neither needs new bindings.
Cons:
- Re-implements toggle/show-hide logic that =vterm-toggle= would handle for free. Acceptable because =vterm-toggle= is built around one toggle-able buffer, and the per-project model is what's wanted.
** Rejected: wrap =vterm-toggle=
=vterm-toggle='s contract is one buffer toggled visible/hidden. Per-project buffers running simultaneously is outside that contract. Wrapping it would mean fighting the abstraction.
** Rejected: project-per-tab via =tab-bar-mode=
Each project gets its own tab. Matches the =ai= / tmux model cleanly, but adds tab-bar UI that isn't in current use. Bigger lifestyle change for a one-window task.
** Rejected: frame-per-project
Each Claude session opens in a new Emacs frame. Hyprland-native, clean isolation, but frame creation under Wayland has historical jank, and it breaks the easy windmove flow between code and Claude.
** Rejected: window-configuration-per-project
Save and restore named window configs (code buffers + Claude vterm together). Preserves the surrounding thinking environment, but window configs go stale when buffers die, and it adds a parallel mechanism to project.el. Overkill for v1.
* Design
** Architecture
New module =modules/ai-vterm.el=. Required after =eshell-vterm-config= in =init.el= so =vterm= is loaded.
Components:
| Function | Kind | Responsibility |
|----------+------+----------------|
| =cj/--ai-vterm-candidates= | pure | Walks =~/.emacs.d=, =~/code/*=, =~/projects/*=; returns abs paths containing =.ai/protocols.org= |
| =cj/--ai-vterm-pick-project= | interactive helper | =completing-read= over candidates; returns picked path |
| =cj/--ai-vterm-buffer-name= | pure | =(format "claude [%s]" basename)= |
| =cj/--ai-vterm-show-or-create= | internal | Given dir + name: display existing buffer, or create vterm + send claude command |
| =cj/ai-vterm= | interactive entry | Composes picker + show-or-create |
The =display-buffer-alist= entry is added at module load:
#+begin_src emacs-lisp
(add-to-list 'display-buffer-alist
'("\\`claude \\["
(display-buffer-in-side-window)
(side . right)
(window-width . 0.5)
(dedicated . t)))
#+end_src
** Data Flow
On =M-x cj/ai-vterm=:
1. Pick a project via =completing-read=. Display in =~/relative= form. Return absolute path.
2. Compute buffer name: =claude [<basename-of-dir>]=.
3. Branch:
- *Buffer exists with live process* → =display-buffer= it. Side-window rule routes it to the right slot.
- *Buffer exists, dead process* → kill it (log last 200 chars to =*Messages*=), then fall through to create.
- *No buffer* → =let=-bind =default-directory= to picked dir and =vterm-buffer-name= to computed name; call =(vterm)=. After process is live, send =claude "Read .ai/protocols.org and follow all instructions."= via =vterm-send-string= + =vterm-send-return=.
4. =select-window= on the displayed window so point lands in Claude. =C-u= prefix shows without selecting.
After this, all navigation is handled by existing global bindings: Shift-arrows (windmove) for focus, C-M-arrows (=buffer-move=) for directional side-swap.
** Error Handling
| Case | Response |
|------+----------|
| Picker cancelled (=quit=) | Silent no-op |
| No candidates found | =user-error= naming the search roots |
| Picked dir disappeared between scan and launch | =user-error= naming the path |
| Existing buffer with dead process | Kill + recreate; log last 200 chars |
| Side-window already showing a different =claude [...]= | =display-buffer= swaps which buffer occupies the slot; hidden one keeps running |
| =vterm= not installed | Module fails to load loudly (no graceful degradation) |
** Per-project tmux sessions
The launch command sent to a fresh AI-vterm shell is
#+begin_example
tmux new-session -A -s <basename> -c <dir> '<claude-cmd>; exec bash'
#+end_example
- =-A= reattaches to an existing session of the same name instead of creating a new one. So a second F9 on the same project after an Emacs crash brings the running Claude back without spawning a duplicate.
- =-s <basename>= names the session after the project's directory basename. =tmux ls= shows the active sessions by project name.
- =-c <dir>= sets the start directory for new sessions (ignored on attach).
- =exec bash= tails the shell command so the tmux window survives Claude exiting -- the session stays alive with a bare prompt for recovery, and reattach is unaffected.
** Tmux Auto-Launch Suppression
The existing =cj/vterm-launch-tmux= on =vterm-mode-hook= types =tmux\n= unconditionally for any vterm buffer. AI-vterm =let='s a dynamic =cj/--ai-vterm-suppress-tmux= flag around =(vterm)= so the hook skips its bare =tmux\n= and the project-named launch command (above) runs instead.
** Testing
Pure helpers tested against real inputs:
- =cj/--ai-vterm-buffer-name= — Normal, Boundary (trailing slash, dot-prefix dirs, spaces in basenames), Error (degenerate paths).
- =cj/--ai-vterm-candidates= — temp directory tree built with =make-temp-file= + =make-directory=, fake =.ai/protocols.org= markers. Assert returned paths, ignored entries.
Internal with mocked boundary:
- =cj/--ai-vterm-show-or-create= — =cl-letf= on =vterm= to skip process spawn; assert buffer name, =default-directory=, claude argv via captured =vterm-send-string= calls. Two branches (exists vs creates) tested with mocked =process-live-p=.
Display rule:
- After =add-to-list=, =display-buffer= on a buffer named =claude [test]= lands in a window with =(window-parameter w 'window-side) = 'right=.
Test files:
- =tests/test-ai-vterm--candidates.el=
- =tests/test-ai-vterm--buffer-name.el=
- =tests/test-ai-vterm--show-or-create.el=
- =tests/test-ai-vterm--display-rule.el=
Smoke test (=:slow= tag, excluded from default suite): launch against a fixture, verify live process.
* Open Questions
- [ ] Default split width — 50/50 vs 60/40 weighted to code. Starting with 50/50.
- [X] Keybinding — F9. Replaces the prior =cj/toggle-gptel= binding on F9; gptel moves to C-F9.
* Next Steps
- TDD implementation in this order: =buffer-name= → =candidates= → =show-or-create= → display rule → interactive entry.
- Wire into =init.el= after =eshell-vterm-config=.
- Pick a keybinding once the command is shipped.
|
