/fleet

Fleet orchestration — manage coding agents (Claude Code, Copilot, Amplifier) running across multiple Azure VMs via azlin.

Invoke as /fleet <command> or just describe what you want and Claude will pick the right command.

All Commands

Status & Monitoring

Admiral Control

Session Management

Task Management

Projects & Knowledge

Co-Pilot

Quick Reference

Key Options

fleet scout   [--session vm:session] [--vm VM] [--skip-adopt] [--incremental] [--save PATH]
fleet advance [--session vm:session] [--vm VM] [--force] [--save PATH]
fleet dry-run [--vm VM ...] [--backend auto|anthropic|copilot]
fleet adopt   <vm> [--sessions s1 s2]
fleet watch   <vm> <session> [--lines 30]
fleet auth    <vm> [--services github azure claude]
fleet add-task "prompt" [--priority high] [--repo URL]
fleet start   [--interval 300] [--max-cycles 10]

Environment

Admiral Configuration

• Model: Claude Opus 4.6 (claude-opus-4-6)
• Max output tokens: 128,000 (reasoning JSON)
• Context gathered per session: full tmux scrollback + first 50 & last 200 transcript messages
• Safety: 57 dangerous-input patterns blocked (with safe allow-list), confidence thresholds (60% send_input, 80% restart), --confirm default on advance
• Docs: docs/fleet-orchestration/ADMIRAL_REASONING.md

Status Icons

Performance & Architecture

• Sequential VM polling: Session discovery from azlin list (no SSH). Pane capture via sequential SSH
• Cached SSH output: Scout caches Phase 1 tmux captures for Phase 3 reasoning (no double-poll)
• Incremental scout: --incremental flag skips unchanged sessions using ~/.amplihack/fleet/last_scout.json
• Bastion tunnel reuse: Reuses existing SSH tunnels via get_existing_tunnels() instead of creating new ones
• PR URL detection: Uses gh pr list on remote VM for reliable PR detection from git state
• Health metrics in reasoning: fleet_health.py wired into SessionContext for admiral decisions
• Unified status classifier: Single canonical classifier in _status.py (no dual TUI/CLI divergence)
• Modular CLI: Commands split across _cli_session_ops.py, _cli_scout_advance.py, _cli_formatters.py (each under 400 LOC)
• Project grouping: Scout report groups sessions by registered project with open objectives displayed

How to Run

Execute via Bash:

fleet <command> [options]

Presenting Results

After running fleet scout or fleet advance, present the output to the user as:

1. Summary table — reformat the CLI output into a clean markdown table with columns: VM | Session | Status | Action | Conf | Summary

2. Proposed inputs — for any send_input decisions, show what the admiral wants to send

3. Follow-up commands — always end with the actionable next steps from the report:

   • fleet advance to send next command to all sessions
   • fleet advance --confirm to review each before executing
   • fleet advance --session vm:session to advance one specific session
   • fleet watch <vm> <session> to inspect specific sessions