Tailnet policy review
Draft: tested on synthetic policies only, never a real one. The redactor passed one real machine run. Inspect the whole response before sharing it.
Say first: "This is read-only. I will read the policy file you provide and show a redacted report without changing your policy, tailnet or business data."
Read references/safety-contract.md and references/commands.md.
Get the local policy path and the operator's agent-tag and admin-tag lists. Accept a file already exported or checked out from their policy repo. Never fetch a policy or ask for credentials. If tags are unknown, say that agent-to-admin reach remains incomplete.
Resolve scripts from the runtime-supplied skill directory. From that directory, run the following with the provided path and repeated tag flags:
python3 scripts/review.py "$POLICY_FILE" --agent-tag tag:agent --admin-tag tag:admin | python3 scripts/redact.pyThe tag names above are examples. Substitute only operator-provided tags, quoting arguments as data. Python 3.10+ is required; never install it.
Show the redacted table. Each row names the rule location, why to inspect it, a safer shape, and official docs. Explain that static selectors do not resolve groups, hosts, CIDRs, posture, or application capabilities. Zero rows cannot prove isolation. Invalid HuJSON means the review is incomplete. Tailscale's policy reference says to prefer grants;
aclsstill work.Stop after a short "What to look at" list. Offer shapes for the operator to consider, without editing, applying or publishing a policy.
What this skill never does
- Change files, network access, tags, ownership, or SSH rules.
- Call an API, retrieve credentials, or open an admin console.
- Read business data, print policy text, save raw output or publish the report.
- Claim to calculate effective access or replace policy tests.