myFleet: Fleet Toolkit for Agentic Work

myFleet is a toolkit of agent-fleet machinery: agent skills, charters, and behavior policies from a working multi-agent setup, plus the harness-sync engine that projects them into eleven coding-agent checkouts — Claude, Codex, Gemini, Cursor, Factory droids, OpenCode, Kilo, Kimi, Antigravity, Vix, and ZCode. It is generalized from a production multi-agent workspace, where every piece is in daily use.

The problem it addresses is familiar to anyone running more than one agent harness: the same instructions, skills, and settings get maintained N times in N tools' private formats, and drift quietly. myFleet keeps one canonical copy of each artifact and treats every harness copy as generated. Policies are written as citable rule IDs (- **ID** — proposition) so a skill, charter, or review can reference a rule instead of restating it. Work lands through a review gate, not direct pushes. The content directories are data; the fleet package is the engine that moves the data.

Repository tour

Quickstart

Python 3.13+ is the only runtime requirement. The engine is checkout-first by design: the sync engine projects this repository's skills/, policies/, and agents/ trees into your harnesses, so it runs from a clone rather than a site-packages install. (A PyPI release of the wheel is planned; the memory and analytics CLIs are the install-independent subset.)

git clone https://gitlab.com/doering-ai/libs/fleet.git && cd fleet
uv sync --all-extras
uv run sync-agents -n     # read-only preview of every projection
uv run sync-agents --check  # report drift between sources and generated consumers
uv run sync-policies -n   # preview the permissions projection

Run the same commands without the uv run prefix once the environment is active. Applying is idempotent: symlinks already pointing at the right target are left alone, hook registrations de-duplicate, and unrelated keys in harness settings survive untouched.

The fleet package

Six console scripts cover the engine's jobs; two background pollers keep shared state fresh.

The pollers are plain scripts meant to run under a systemd user timer: the bulletin poller snapshots and cross-checks the shared agent bulletin (docs/bulletin.md), and the memory poller commits new agent-memory files to the default branch (docs/memory.md).

Design notes

• Data is not code. skills/, agents/, policies/, and settings/ are plain files a human can read and diff; the package only moves and renders them. Generated consumers are never sources.
• Dependency discipline. The ported modules stay stdlib-plus-pydantic by design, so they run inside externally-managed environments. Shared utilities come from my-basis where it clearly helps, and nothing depends on a private package.
• Preview, then idempotent apply. Every projection has a dry-run; applying twice changes nothing.
• Hooks are boring. Hook scripts are pure-stdlib programs that speak JSON over stdio; only the per-harness registration format differs.
• Rules are citable. Policies use stable ID anchors with nested scope, so agents can be pointed at a rule instead of a wall of prose.

Roadmap

• agent-worktree — per-agent git worktree isolation with a review-gated merge path. The policies describe the workflow today; the CLI has not been generalized yet.
• agentviz — interactive dashboards over the agentlog data.
• Tracker sync — reconcile a local task mirror with a hosted tracker; the use-plane skill documents the working method.

Provenance and sanitization

The contents are generalized from a production multi-agent workspace, and publication means the private provenance stays out. That is enforced, not promised: scripts/check_sanitized.py scans every tracked file in CI for known private patterns — names, hosts, home paths, vault references, ID fragments — and fails the pipeline on any genuine hit. The carve-outs are few and documented in the script: forge URLs in this repository's own namespace, secret-store product documentation in the auth skill, and the engine's own harness identifiers.

Contributing and license

GitLab is the development forge; report issues and propose changes at the GitLab project. Changes pass the same sanitizer and pipeline as everything else.

Licensed under MPL-2.0.