Fleetmind

Command a fleet of AI agents — with a review gate that keeps quality high.

Fleetmind is a thin, portable installer that stands up a whole crew of agentic engineering tools and wires them together, so in one command you go from a bare machine to a working setup where you're the captain and a crew of agents does the work: isolated parallel worktrees, long-running autonomous loops, and a review gate that reviews every change and opens a real PR/MR with a risk level.

It works with GitHub or GitLab (auto-detected per repo), on macOS, Linux, and Windows (native — no WSL required — or WSL2). The agentic tools are the excellent MIT-licensed work of Kun Chen (@kunchenguid) — Fleetmind is the portable wrapper that installs and configures them. See CREDITS.

📖 Start here: USER_GUIDE.md — the daily rhythm, copy-paste recipes (ship a change, run tasks in parallel, overnight loops), the review gate explained, testing, and troubleshooting.

---

Install (one command)

Windows (native — no WSL, full parity):

.\bootstrap.ps1 -Path lite     # installs the crew, quality gate, and hooks natively
.\doctor.ps1                   # expect CAPTAIN READY

Windows + WSL (only for tmux/firstmate, or corporate/VPN GitLab networking):

.\gitlab-setup.ps1 -GitlabHost gitlab.yourcompany.com

macOS / Linux / WSL:

bash gitlab-setup.sh --host gitlab.yourcompany.com

Installs the whole crew — Node (nvm), Go, treehouse, gnhf, no-mistakes, glab and gh, Claude Code, and the skills — no sudo, idempotent, non-clobbering.

Then the two logins (only you can — they're your credentials):

glab auth login --hostname gitlab.yourcompany.com   # PAT: api + write_repository
gh auth login                                        # only if you also use GitHub
claude                                               # sign in on first launch
source ~/.bashrc

Prefer the modular installer or a GitHub-only setup? bootstrap.sh / bootstrap.ps1 offer that (Windows: native full parity via bootstrap.ps1, or WSL2 if you want tmux/firstmate).

---

Prove it works (one command)

bash e2e-test.sh --host gitlab.yourcompany.com

Runs every feature end-to-end and writes a report + evidence to ~/.captain/e2e/:

• Tier A (no login): tools present · glab reaches your GitLab · treehouse worktree pool · no-mistakes daemon health.
• Tier B (needs a logged-in agent): gnhf makes a real commit · no-mistakes reviews a change and opens a real MR with a risk level.

A green run:

### E2E result: 6 passed, 0 failed, 0 skipped
  ⚓ FULL PASS — every feature proven end-to-end.

--quick runs Tier A only. Reproducible — anyone can run it.

---

The workflow

   PLAN            ISOLATE          BUILD            GATE                 SHIP
 talk to an   →  treehouse    →   the agent    →   git push         →   approve by
 agent (plans     (isolated       does the         no-mistakes           risk, merge
  visually)       worktree)        work             (review + MR)

You spend your time at PLAN and SHIP; the crew runs the middle.

The review gate (risk-tiered)

Don't hand-review every diff — ship through the gate and sign off only on risk-worthy changes:

git push no-mistakes <branch>

no-mistakes runs Review → Test → Docs → Lint → Push → PR/MR → CI in an isolated worktree, auto-fixes safe things, and assigns a risk level:

Risk	You do
low	glance at the summary; merge
medium	read the diff + recorded evidence; approve or send back
high	review carefully — this is where your attention belongs

It auto-detects the forge (glab for GitLab, gh for GitHub). A push-guard hook blocks direct git push origin so nothing bypasses the gate.

The pre-MR quality gate (no false greens)

Before a PR/MR is opened, pre-mr-check refuses it unless the branch/commit are named per convention and fresh validation evidence exists that covers this change — so a green that measured nothing can't ship. Each repo declares its own specifics in one committed .fleetmind/project.json (target, reviewer, ticket pattern, validate command + evidence); the gate itself stays generic and opens the PR/MR via forge. The review-gate hook also blocks raw gh pr create / glab mr create, so the gate is the only path. Every crew agent carries the matching completion contract (done-with-evidence or blocked-with-proof). See docs/quality-gate.md.

Everyday commands

treehouse                                   # isolated worktree (auto-reused pool)
gnhf "raise coverage to 90%" --agent claude # long-running autonomous loop
git push no-mistakes my-branch              # ship through the review gate
no-mistakes runs                            # runs + MR links

See USER_GUIDE.md for full recipes.

---

What gets installed

Layer	Components
Crew & pipeline	treehouse · gnhf · no-mistakes · firstmate* · axi/gh-axi + lavish skills
Forge CLIs	glab (GitLab) and gh (GitHub) — no-mistakes picks per repo
Quality gate	pre-mr-check (naming + validation-evidence gate) · per-repo .fleetmind/project.json · review-gate hook (blocks bypass) · forge (gh|glab)
Harnesses	Claude Code, Codex CLI, opencode
Runtimes	Node (via nvm), Go
Ship (optional)	WezTerm, tmux*, Neovim, ripgrep/fd/fzf, starship
Brains	global memory (CLAUDE.md ↔ AGENTS.md), the senior engineering-standards baseline (with the no-false-green completion contract), review-gate + e2e-testing skills

* tmux + firstmate are Unix-only. On Windows they need WSL2; everything else — crew, quality gate, hooks, forge — runs natively.

---

What can't be automated (true for everyone)

• The two logins (glab / claude) — an AI review gate can't review without an authenticated agent, and no software can log in as you.
• WSL2 provisioning on Windows — needs admin + a reboot (the installer sets it up).
• Your judgment — deciding what to build. That's the captain's job.

Everything else is one command.

---

Docs

• USER_GUIDE.md — day-to-day usage, recipes, troubleshooting
• CONTRIBUTING.md — dev setup + how to verify changes
• SECURITY.md — what it runs, and reporting issues
• CREDITS.md — the upstream tools this builds on
• CHANGELOG.md · CODE_OF_CONDUCT.md

License

MIT © 2026 vinaygiri. The bundled crew tools are MIT © their authors — see CREDITS.