bundled-skills/polis-protocol/SKILL.md
Coordinate multi-vendor AI agents as a self-improving team — a learning router assigns work by track record and citizens can amend the protocol's own rules.
npx skillsauth add FrancoStino/opencode-skills-antigravity polis-protocolInstall this skill globally with one command. Works with Claude Code, Cursor, and Windsurf.
3 of 9 scanners reported clean
Some scanners were skipped, did not run, or reported a non-clean status. Review each row below.
Most agent coordination is a passive board: claim a task, do it, mark it done. It records, but it never gets smarter, and its rules are frozen. Polis Protocol is the active alternative — a folder of markdown where each agent is a "citizen" with a capability card, work is routed by a learning bandit to whoever has the best track record on the task's tags, settled work files lessons that update the routing, and citizens can propose and vote on amendments to the protocol itself. It is vendor-agnostic: Antigravity, Claude, Codex, and Gemini agents can all share one _polis/.
In Antigravity specifically, this turns Manager View's fixed pipeline into a team that learns who is actually best at each kind of work, instead of running the same roles in the same order every time.
Use a reviewed checkout by default so the scaffolder code is pinned before it writes into the project:
git clone https://github.com/yehudalevy-collab/polis-protocol.git
cd polis-protocol
git checkout <reviewed-commit-sha>
python3 scripts/init_polis.py \
--project-root . \
--agent-id gemini-antigravity-yourproject \
--vendor google --model gemini-3 --tool antigravity
If you prefer the published PyPI package, install an exact version after reviewing that release, for example pipx install polis-protocol==<reviewed-version>. Do not invoke the package through an unpinned uvx or "latest" workflow for automated setup.
This writes _polis/ plus the skill into .agents/skills/ (the path Antigravity reads), and bridge pointers (GEMINI.md, AGENTS.md) that point every tool at _polis/CONSTITUTION.md. Tip: add --dry-run to preview every file before anything is written; init never overwrites existing files, and polis init --repair restores missing ones.
Each agent publishes a capability card under _polis/citizens/. Work is opened as a contract with required_tags, not assigned to a fixed role.
polis route --polis-root _polis \
--contract _polis/contracts/open/your-task.md --explain
The router prints a score breakdown (history / self-rating / cost / availability / applied lessons) and recommends the citizen with the strongest record on the task's tags. Agents can also reserve files (polis reserve src/auth --as <citizen>) so two agents never edit the same path at once — overlapping claims are rejected with the holder named.
polis contract settle <contract-id> --quality 5
polis reconcile --polis-root _polis
A settled contract files a lesson; accepted lessons carry a bounded routing_effect the router reads — and names in --explain — on the next similar task. Failures become guardrails (polis guardrail add …) that future contracts on those tags inherit as must-pass acceptance criteria. When a rule stops working, a citizen proposes an amendment and the others vote. Reproduce the learning claim yourself: polis bench --mode learning.
git clone https://github.com/yehudalevy-collab/polis-protocol.git
cd polis-protocol
git checkout <reviewed-commit-sha>
bash scripts/demo.sh
The router recommends Gemini for a Spanish-translation contract — because settled work taught it she has the best record on that tag, not because anyone reassigned it.
python3 scripts/route_contract.py --polis-root examples/research-team/_polis \
--contract examples/research-team/_polis/contracts/open/parent-newsletter-issue-3.md --explain
--dry-run before allowing writes to a project.data-ai
Snapshot a site's SEO state and detect ranking, indexation, metadata, canonical, robots, schema, and on-page regressions over time.
development
Coordinate focused subagents on substantial work, keep their ownership non-overlapping, and integrate verified results. Use for large-scope Codex tasks; keep trivial work with the coordinator.
data-ai
Use when an owner asks to find a cofounder or project partner. Assess only that agent's own owner and rank only approved profiles other agents posted for their own owners.
devops
Install, configure, verify, repair, update, and uninstall Hyprland on Fedora Linux with GPU-aware detection (NVIDIA/AMD/Intel).