verModal

AVER (Agent Versioning) — a version scheme for AI agents and robots instead of software releases. Semver tells you how much code changed, which isn't what you need to know when you're handing a task to an agent. AVER tracks four things separately: which model's driving, what it can touch, how much it's allowed to do without a human, and which guardrails are active.

Full writeup is in docs/aver. This repo is that spec plus a small CLI for generating and checking version strings.

v anthropic:sonnet-5.9f3c21a0.2-a3f9c1
  └vendor┘ └mind┘ └capability┘│└policy┘
                               autonomy

Install

Needs Bun, 1.3+. Runs the TS directly, no build step.

bun install

Use

bun run src/cli.ts generate \
  --vendor anthropic --mind sonnet-5 \
  --manifest examples/claude-code-manifest.json \
  --autonomy 2 \
  --policy-source examples/claude-code-policy.txt \
  --out examples/version.json

bun run src/cli.ts parse "vanthropic:sonnet-5.9f3c21a0.2-a3f9c1"

bun run src/cli.ts verify "vanthropic:sonnet-5.9f3c21a0.2-a3f9c1" \
  --manifest examples/claude-code-manifest.json

generate builds a version string from a tool/actuator manifest, an autonomy level (0–3), and a policy file, and writes {version, capability_manifest} if you pass --out. parse just splits a version string into its fields. verify checks that a version's capability hash actually matches a given manifest.

There's a sample manifest and policy in examples/, modeled on a Claude Code session's tool set.

Docs

Everything past the pitch above lives in docs/aver, that's where the actual field definitions, comparison rules, and worked examples are.

Status

Pre-alpha, v0.1. First-pass proto schema — the four fields and the hashing approach for capability feel right but haven't been used outside this repo yet. Vendor namespacing and how robotics extends autonomy are still open — see the changelog. Expect breaking changes.

License

Apache-2.0, see LICENSE.