#4008 access can-i: evaluate a local grants file without --server

What

Let swamp access can-i evaluate a local grants file without --server. For example:

swamp access can-i --action run --on model:esqa --method provisionDevices \
  --as user:agent --grants-file .swamp/agent-grants.yaml --json

It would return allow or deny, plus the grant that decided it. A local run would still not be enforced. This is only an evaluation.

Why

The grant model already expresses what an AI coding agent may do: deny first, per-method methods: lists, CEL conditions on model name, tags, type and method, workflow vault caps, and approve for workflow approval steps. But a local swamp run without --server isn't subject to grants, and that's where agents run swamp.

Today the only way to apply grants to an agent is to route its work through a swamp serve. A command that evaluates grants offline would let a Claude Code PreToolUse hook ask swamp for each swamp model method run and swamp workflow run. The hook would map deny to a block and allow to a pass, and could ask the human for anything with no matching grant. The agent's policy would then live in swamp's own grant format, reviewed in the repo, not in a separate allow-list in each tool's settings.

Notes

• The hook would need the same subject resolution and condition evaluation that serve uses, so the answers match what serve would decide.
• A JSON result with the grant that decided it would let the hook explain its decision to the user.