fix(autocoder): stop the manager dying on the bypass-permissions dialog

[laird]

Problem

The parallel-agent manager silently fails to start under --agent claude. Its tmux window is left at a bare shell in the project root, which reads as an unrelated login session rather than a failed manager.

Observed scrollback:

claude --dangerously-skip-permissions --model claude-opus-5

  WARNING: Claude Code running in Bypass Permissions mode
  ❯ No, exit
    Yes, I accept

kink-party main ? ✗

Cause

The manager is the only claude path that launches a REPL and then types its prompt into it. Workers use WORKER_LAUNCH_MODE=shell, where claude-worker-loop.sh starts Claude itself, so no send-keys ever races a TUI — which is why workers were unaffected and this looked intermittent.

The manager's prompt was sent after a fixed sleep 5 with no readiness check. --dangerously-skip-permissions opens a consent dialog defaulting to No, exit, so the trailing Enter answered the dialog instead of submitting the prompt.

Fix

New MANAGER_COMMAND_MODE (argv|agent-input|shell). In argv mode the prompt is appended to MANAGER_LAUNCH_CMD via printf %q and the post-launch send-keys is dropped entirely — no stray Enter, so the dialog waits for a real keypress and the prompt runs once accepted.

Agent	Mode	Behavior
claude	argv	changed
gemini	agent-input	unchanged
codex (goals branch)	agent-input	unchanged
codex (fallback), droid	shell	unchanged

Applied at both the tmux and cmux manager launch sites.

Verification

• claude -p "/help" returns "/help isn't available in this environment" — confirming a slash command passed as argv is parsed as a slash command, not literal prompt text.
• Dry-run of the composed command: claude --dangerously-skip-permissions --model claude-opus-5 /autocoder:monitor-loop
• bash -n clean on both files.

🤖 Generated with Claude Code

https://claude.ai/code/session_011JjWeRTcWvmuUPKRDvveEG