fleet

/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

Command	What it does
fleet status	Quick text overview of all VMs and sessions
fleet scout	Discover, adopt, dry-run reason, show report
fleet snapshot	Point-in-time capture of all managed sessions
fleet watch <vm> <session>	Live snapshot of a single session
fleet observe <vm>	Observe sessions with pattern classification
fleet tui	Interactive Textual TUI dashboard

Admiral Control

Command	What it does
fleet advance	LIVE — reason and execute actions on sessions
fleet dry-run	Show what the admiral would do (no action)
fleet run-once	Single PERCEIVE->REASON->ACT cycle
fleet start	Run autonomous admiral loop

Session Management

Command	What it does
fleet adopt <vm>	Bring existing tmux sessions under management
fleet auth <vm>	Propagate auth tokens (GitHub, Azure, Claude)

Task Management

Command	What it does
fleet add-task "prompt"	Queue work for the fleet
fleet queue	Show task queue
fleet dashboard	Project-level tracking
fleet report	Generate fleet status report

Projects & Knowledge

Command	What it does
fleet project add/list/remove	Manage fleet projects
fleet project add-issue <proj> <num>	Track a GitHub issue as a project objective
fleet project track-issue <proj>	Sync objectives from GitHub issues by label
fleet graph	Fleet knowledge graph summary

Co-Pilot

Command	What it does
fleet copilot-status	Show copilot lock/goal state
fleet copilot-log	Show copilot decision history

Quick Reference

User says	Command
"What are my agents doing?"	fleet scout
"Show me the fleet"	fleet status
"Send next steps to all sessions"	fleet advance
"Advance without confirmation"	fleet advance --force
"Watch what dev/cybergym is doing"	fleet watch dev cybergym
"Add auth to the new VM"	fleet auth <vm>
"Queue this task for the fleet"	fleet add-task "prompt"
"Track issue #42 for myapp"	fleet project add-issue myapp 42
"Sync objectives from GitHub"	fleet project track-issue myapp
"Open the dashboard"	fleet tui

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|litellm]
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

Variable	Required for
AZLIN_PATH	All commands (auto-detected if azlin on PATH)
ANTHROPIC_API_KEY	scout, advance, dry-run, run-once, start

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

Icon	Status	Meaning
[~]	thinking	Agent is actively processing
[>]	running	Agent producing output
[.]	idle	Agent at prompt, waiting for direction
[X]	shell (dead agent)	No agent detected in this session
[Z]	suspended	Agent backgrounded but alive
[!]	error	Error detected in session output
[+]	completed	Agent finished its task
[?]	waiting input	Agent asked a question, awaiting response

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