FluxyChatReference
Agent tool policy
JSON policy-as-code for agent tool allow, deny, and require_approval, aligned with EU AI Act HITL.
FluxyChat evaluates a per-project JSON policy before every agent tool call. Policies merge with existing HITL gates and EU AI Act high-risk overrides.
Effects
| Effect | Behavior |
|---|---|
allow | Tool runs normally (subject to HITL gate / EU AI Act) |
require_approval | Forces human approval even if tool not in alwaysRequire |
deny | Tool blocked with policy_denied; audit row written |
Policy shape
{
"version": 1,
"defaultEffect": "allow",
"rules": [
{
"id": "approve-send",
"tools": ["sendMessage", "postMessage"],
"effect": "require_approval",
"priority": 10,
"reason": "Outbound messages require HITL"
},
{
"id": "hold-search",
"tools": ["web_search", "search_*"],
"effect": "allow",
"onHoldPhrase": "Searching the knowledge base…",
"priority": 5
},
{
"id": "deny-delete",
"tools": ["delete_*"],
"effect": "deny",
"priority": 20
}
]
}Tool patterns support exact names, prefix* wildcards, or regex.
On-hold narration
Add onHoldPhrase to a rule to emit an agent_on_hold room event during tool execution (voice duplex or console preview). Barge-in cancels the filler on the client.
{
"id": "hold-search",
"tools": ["web_search"],
"effect": "allow",
"onHoldPhrase": "Searching the knowledge base…"
}API
| Method | Path | Description |
|---|---|---|
| GET | /agents/tool-policy | Load policy |
| PUT | /agents/tool-policy | Upsert policy |
| POST | /agents/tool-policy/evaluate | Dry-run evaluator |
Dashboard: Settings → Agent tool policy (/settings/agent-tools).
Audit
Every evaluation inserts into agent_tool_policy_audit (tool, effect, rule_id, run_id, room_id).