Problem

The current UX loop is:

1. User opens GitHub to create an issue (or uses gh issue create manually)
2. User notes the issue number
3. User runs ao spawn <project> <number>
4. Repeat for every issue

The "orchestrator" tmux session that ao start creates is just a plain shell. There's no AI in it. It offers nothing beyond what the user could do in any terminal.

What it should be

The orchestrator session should be a Claude Code (or similar) session that:

• Understands the project — repo, open issues, recent PRs, running agents
• Creates issues from descriptions — "fix the config loading bug" → creates GitHub issue → spawns agent
• Monitors agents naturally — you ask "how's ao-1 doing?" → it peeks at the tmux session and summarizes
• Coordinates dependencies — "spawn issue 8 only after 5, 6, 7 are merged" → it watches and acts

Concretely, the loop should become:

# In the orchestrator session (Claude Code running there)
> we found a bug where defaults.agentConfig is ignored by spawn
→ Creates issue #74 with proper description
→ Spawns ao spawn ao 74
→ "Agent ao-1 is working on feat/74. I'll notify you when it opens a PR."

Why this matters

Without this, AO is infrastructure that requires manual orchestration. The viral moment — "I described work and 4 agents started in parallel" — doesn't happen because the human is still doing all the coordination manually.

The CLAUDE.orchestrator.md file hints at this vision but the implementation is just a blank tmux shell.

Suggested implementation

1. Instead of tmux new-session for the orchestrator, launch Claude Code in the project root with a system prompt that gives it:

   • ao CLI access
   • gh CLI access
   • Awareness of all running sessions and their status
   • Instructions to act as orchestrator (create issues, spawn agents, monitor, coordinate)

2. The user interacts with it conversationally and the AI handles all the ao spawn / gh issue create plumbing.