cli-fleet 0.3.0

Launch multiple enforced Claude Code agent teams in parallel terminal windows, hardware-aware, coordinating via a shared filesystem mailbox.

CLI Fleet

Launch multiple enforced Claude Code agent teams in parallel — each team a separate claude session in its own window, coordinating via a shared filesystem mailbox. A thin, hardware-aware wrapper that completes the stack:

cli-wikia  →  cli-enforcement  →  cli-fleet
(knowledge)    (the brakes)        (the power)

cli-fleet bundles fleetcode's proven shell orchestration verbatim and adds:

• Hardware gating — refuses to launch more parallel agents than the box can take.
• Automatic enforcement — deploys the cli-enforcement engine into each team (hooks merged with fleetcode's mailbox hook), so every spawned team is governed.

Install

pip install cli-fleet        # pulls in cli-enforcement + cli-wikia

Usage

cli-fleet launch fleet.json            # hardware-gated, auto-enforced launch
cli-fleet launch fleet.json --background  # use `claude -p` instead of windows
cli-fleet status [meta-team]           # fleet status (auto-detects if one fleet)
cli-fleet dashboard [meta-team]        # scorecard: teams, mailbox, findings + per-team enforcement points
cli-fleet dashboard --watch 5          # redraw every 5s until Ctrl-C
cli-fleet send team-a team-b "message" --type finding   # cross-team mailbox message (--team to pick fleet)
cli-fleet cleanup [meta-team]          # tear down (auto-detects if one fleet)

Note: cli-fleet launch requires Linux with a terminal emulator (gnome-terminal/xterm family) unless --background is used.

Multi-CLI fleets (v0.3.0)

Teams are no longer claude-only: each team in the config may set an optional "model" (default claude). cli-fleet deploys the matching cli-enforcement engine into that team's workdir, merges a mailbox-check hook into the model's own hook file (without clobbering the enforcement hooks), and spawns the team via that CLI's headless command.

{
  "meta_team": "mixed-demo",
  "project_dir": "/path/to/project",
  "teams": [
    { "name": "team-claude",   "role": "backend",  "teammates": 2,
      "task": "Audit the API layer." },
    { "name": "team-gemini",   "role": "frontend", "teammates": 2,
      "task": "Audit the UI layer.",   "model": "gemini" },
    { "name": "team-deepseek", "role": "infra",    "teammates": 2,
      "task": "Audit the deploy scripts.", "model": "deepseek" }
  ]
}

model	binary	headless spawn	mailbox hook event	launch-verified?
claude	claude	claude --dangerously-skip-permissions -p <task>	UserPromptSubmit	✅ verified
gemini	gemini	gemini --approval-mode yolo -p <task>	BeforeAgent	✅ verified
deepseek	deepseek-code	deepseek-code --dangerously-skip-permissions -p <task>	UserPromptSubmit	✅ verified
copilot	copilot	copilot --allow-all-tools --allow-all-paths -p <task>	UserPromptSubmit	✅ verified
chatgpt	codex	codex exec --full-auto <task>	UserPromptSubmit	⚠ best-effort (from wiki docs; codex not verified locally — launch prints a warning)
antigravity	agy	agy --dangerously-skip-permissions -p <task>	PreInvocation	✅ verified

Shared state now lives at the model-neutral ~/.cli-fleet/meta-teams/ (META_TEAM_DIR env still wins). If a legacy fleetcode ~/.claude/meta-teams exists, the new root is created as a symlink to it, so old fleets keep working. The dashboard shows each team's model.

License

MIT — see LICENSE.