permission-audit

permission-audit

Read-only scan → one severity-ranked report. No auto-fix, no auto-revoke, no notification.

Verbs (the installed lark-* skills): drive +search (--doc-types/--folder-tokens/--space-ids/--creator-ids — inventory files; NO permission flag), docs +fetch (content for PII scan), base +table-list. Raw CLI has no verb that returns per-file share/permission state — get permission metadata via the atomic lark-drive skill (or note the gap in the report). Atomic shapes + token flags: connectors/LARK-RECIPES.md.

Workflow

1. Scope. Default: all Drive/Wiki/Base the user admins. Narrow via --folder-tokens, --space-ids, --creator-ids. >1000 items → warn + suggest batching.
2. Inventory. drive +search --doc-types doc,sheet,bitable,file --folder-tokens <t> --jq '.data' (peek shape; + --space-ids for wiki; needs search:docs:read scope). Permission state → atomic lark-drive skill per item (batch). Cache for the session — no re-fetch.
3. Classify by memory rule: public-internet → HIGH (CRIT if strict policy) · anyone-with-link → MED · external_users >0 → MED (HIGH if folder is Confidential/HR/Finance) · inactive_owner → MED.
4. PII/DLP scan on HIGH+ items. docs +fetch --doc <url> --limit 1 (cap content), regex memory/regulated-data-types.md: VN CCCD \b\d{9}\b|\b\d{12}\b, Luhn-valid 13-19 digit cards, email+phone batch (>10/file), API keys (sk-, ghp_, AKIA, xoxb-). Each match: severity +1, mask 50% digits.
5. Dedupe vs incident-log.md — already-raised → mark RECURRENT.
6. Format report — SUMMARY (counts by severity) + CRIT/HIGH findings (token/url/owner/share/PII/policy ref/action/status) + OWNER EXPOSURE top 5 + RECOMMENDED ACTIONS. Ends "read-only output, no changes made."

Hard rules

1. READ-ONLY — never call +update/+apply-permission/+delete. "Fix luôn" → redirect to admin Lark UI or offer a task (confirm).
2. Mask sensitive data — 4111-****-****-1234, CCCD 3+3 digits only.
3. Don't dump file contents — report token + URL only.
4. Scan only where user has admin. 403 → skip silent, no brute force.
5. Severity from policy — every finding refs a memory/policies.md §.
6. Report to admin privately — 1:1 chat, never wide group.

Memory

Required: policies.md (sharing rules + severity map), regulated-data-types.md (PII patterns). Recommended: incident-log.md (dedupe recurrent).

Empty memory → conservative defaults (anyone-with-link = HIGH) + warning to set up policies.md.