civitai XGuard Policy Manager
Use this skill to manage XGuard policy options against the orchestrator's /v1/manager/xguard/* admin endpoints.
Each label's policy is the natural-language prompt text + threshold + action that the orchestrator applies when evaluating an XGuard call. See docs/features/scanner-prompt-tuning.md for how this fits into the scanner refinement lifecycle. Phase 4 (refine policy) is the typical use of this skill: fetch the current options, edit a policy, PUT the new version, then track FP/FN rates against the new policyHash in the audit log.
Setup
The skill reads ORCHESTRATOR_ENDPOINT and ORCHESTRATOR_ACCESS_TOKEN from the project .env (or .claude/skills/xguard-manager/.env if you want skill-specific overrides). No additional setup beyond having those env vars set.
Commands
node .claude/skills/xguard-manager/manage.mjs <command> [args] [options]
| Command | Description |
|---|---|
get <mode> |
GET current options for the mode (text or prompt) |
defaults <mode> |
GET hardcoded defaults baked into the orchestrator |
put <mode> |
PUT new options for the mode (requires --writable + --file) |
reset <mode> |
POST reset back to defaults (requires --writable) |
export |
GET bulk export across all modes |
import |
PUT bulk import across all modes (requires --writable + --file) |
Options
| Flag | Description |
|---|---|
--writable |
Allow destructive operations (PUT / POST). Required for put, reset, import. |
--file <path>, -f <path> |
Read the request body (JSON) from a file. Required for put and import. |
--output <path>, -o <path> |
Save response to a file (instead of printing to stdout). |
--quiet, -q |
Only print the response body, no connection headers. |
--timeout <s>, -t <s> |
Request timeout in seconds (default: 30). |
Examples
# Inspect current prompt-mode options
node .claude/skills/xguard-manager/manage.mjs get prompt
# Save the current prompt-mode options to a file you can edit
node .claude/skills/xguard-manager/manage.mjs get prompt -o /tmp/prompt-policies.json
# See what the defaults look like (useful for scaffolding a new label)
node .claude/skills/xguard-manager/manage.mjs defaults prompt
# Ship an edited policy
node .claude/skills/xguard-manager/manage.mjs put prompt -f /tmp/prompt-policies.json --writable
# Wipe prompt-mode policies back to defaults (destructive)
node .claude/skills/xguard-manager/manage.mjs reset prompt --writable
# Back up the entire policy registry before a risky edit
node .claude/skills/xguard-manager/manage.mjs export -o /tmp/xguard-backup.json
# Restore from a backup
node .claude/skills/xguard-manager/manage.mjs import -f /tmp/xguard-backup.json --writable
Safety
- Read-only by default:
get,defaults, andexportalways work without--writable. - Destructive operations require explicit
--writable:put,reset, andimportchange orchestrator state and refuse to run without the flag. - Always ask the user before using
--writable: policy changes are global — every subsequent scan will use the new policy text. Confirm the intent before running. - Back up before editing: run
export -o backup.jsonbefore any non-trivial change so you canimportto restore if needed.
Typical workflow for refining a policy (Phase 4)
# 1. Snapshot current state for safe rollback
node .claude/skills/xguard-manager/manage.mjs export -o /tmp/xguard-backup-$(date +%Y%m%d).json
# 2. Pull the current options for the mode you're editing
node .claude/skills/xguard-manager/manage.mjs get prompt -o /tmp/prompt-current.json
# 3. Edit /tmp/prompt-current.json with the refined policy text
# 4. Ship it (after user confirms)
node .claude/skills/xguard-manager/manage.mjs put prompt -f /tmp/prompt-current.json --writable
# 5. Track FP/FN rate of new policyHash via the audit log + focused review
Why this is separate from the postgres-query / clickhouse-query skills
Those query the audit data (verdicts, scores, matched terms). This skill manages the policy itself — the prompt text the orchestrator hands to XGuard. A typical tuning loop uses both: query data to identify FP patterns, edit the policy here, then re-query after new scans land under the new policyHash.