Codex tmux Skill

Sends a prompt to a Codex CLI session running in a tmux pane and polls until the response is complete.

Procedure

1. Identify the Codex pane

tmux list-panes -a -F '#{session_name}:#{window_index}.#{pane_index}  #{pane_current_command}  #{pane_title}'

Find the pane whose command is codex (or node if Codex is running via Node). If the user didn't specify a pane, pick the first match.

If no matching pane is found, proceed to Step 1.5 to auto-launch Codex. Otherwise, set <pane> to the matched pane and skip to Step 2.

1.5. Auto-launch Codex (only if no pane found in Step 1)

# Create a new window named "codex" in the main session
tmux new-window -t main -n codex
# Launch codex (inherits env from shell, including API keys via ~/.zshrc)
tmux send-keys -t main:codex 'codex' C-m

Then poll every 2 s until the › idle prompt appears (max 30 s):

for i in $(seq 1 15); do
  sleep 2
  output=$(tmux capture-pane -p -t main:codex -S -50)
  echo "$output" | grep -qE '›\s*$' && break
done

Set <pane> to main:codex.

If the loop completes without seeing › , abort and report:

"Codex の起動を確認できませんでした。codex コマンドが PATH にあるか確認してください。"

2. Capture a baseline snapshot

tmux capture-pane -p -t <pane> -S -200

Save this as the baseline so you know where the new response starts.

3. Send the prompt

Detect whether $prompt contains newlines and use the appropriate method:

Single-line prompt:

tmux send-keys -t <pane> '<prompt>' C-m

Multi-line prompt (contains \n):

# 1. Write to temp file
printf '%s' "$prompt" > /tmp/tmux_send.txt
# 2. Load into tmux buffer
tmux load-buffer /tmp/tmux_send.txt
# 3. Paste into pane (preserves newlines)
tmux paste-buffer -t <pane>
sleep 0.5
# 4. Execute
tmux send-keys -t <pane> C-m

4. Poll until response is complete

Repeat every 3 seconds (max 120 s total):

tmux capture-pane -p -t <pane> -S -200

Stop polling when either condition is met:

• The last non-empty line of the captured output ends with › (Codex idle prompt)
• Two consecutive captures are identical (output has stabilised)

5. Extract and return the response

Take the diff between the baseline and the final capture. Strip the trailing › prompt line. Return the extracted text to the user.

6. Timeout handling

If 120 s elapse without detecting completion, return whatever was captured and note that the timeout was reached so the user can check manually.