@driftgard/node 1.5.1 → 1.7.0
This diff represents the content of publicly available package versions that have been released to one of the supported registries. The information contained in this diff is provided for informational purposes only and reflects changes between package versions as they appear in their respective public registries.
- package/README.md +46 -0
- package/dist/index.d.ts +18 -2
- package/dist/index.js +48 -0
- package/dist/types.d.ts +37 -0
- package/package.json +1 -1
package/README.md
CHANGED
|
@@ -89,6 +89,52 @@ const result = await dg.evaluate({
|
|
|
89
89
|
|
|
90
90
|
All fields in `usage` are optional. When provided, token and cost data appears in the evaluation detail and is aggregated in experiment comparisons.
|
|
91
91
|
|
|
92
|
+
## Cost alerts
|
|
93
|
+
|
|
94
|
+
When cost alerting is enabled on your project, the response includes a `cost_alert` field if a threshold is exceeded:
|
|
95
|
+
|
|
96
|
+
```typescript
|
|
97
|
+
const result = await dg.evaluate({ ... });
|
|
98
|
+
|
|
99
|
+
if (result.cost_alert) {
|
|
100
|
+
console.warn(`Cost alert: ${result.cost_alert.scope} spend $${result.cost_alert.actual_usd} exceeds $${result.cost_alert.threshold_usd}`);
|
|
101
|
+
// Throttle the agent, notify the user, etc.
|
|
102
|
+
}
|
|
103
|
+
```
|
|
104
|
+
|
|
105
|
+
Configure thresholds in Settings — per-project, per-model, or per-session. Session-scoped alerts catch runaway agent loops in real-time.
|
|
106
|
+
|
|
107
|
+
## Tool call validation
|
|
108
|
+
|
|
109
|
+
Validate AI agent tool/function calls against your control pack's tool rules:
|
|
110
|
+
|
|
111
|
+
```typescript
|
|
112
|
+
// Direct tool call evaluation
|
|
113
|
+
const result = await dg.evaluateToolCall({
|
|
114
|
+
project_id: "your-project-id",
|
|
115
|
+
model_id: "gpt-4o",
|
|
116
|
+
tool_name: "transfer_money",
|
|
117
|
+
parameters: { amount: 500, to_account: "account_123" },
|
|
118
|
+
session_id: "sess_abc123",
|
|
119
|
+
});
|
|
120
|
+
|
|
121
|
+
if (!result.evaluation.allowed) {
|
|
122
|
+
console.log("Tool blocked:", result.fallback?.message);
|
|
123
|
+
}
|
|
124
|
+
|
|
125
|
+
// Or wrap a tool function — blocks automatically
|
|
126
|
+
const safeTransfer = dg.guard(transferMoney, "transfer_money", "your-project-id");
|
|
127
|
+
await safeTransfer({ amount: 500, to_account: "account_123" }); // throws if blocked
|
|
128
|
+
|
|
129
|
+
// Report execution outcome (optional)
|
|
130
|
+
await dg.reportOutcome(result.evaluation_id, "your-project-id", {
|
|
131
|
+
execution_status: "success",
|
|
132
|
+
duration_ms: 230,
|
|
133
|
+
});
|
|
134
|
+
```
|
|
135
|
+
|
|
136
|
+
For Strands agents, use the `BeforeToolCallEvent` hook — see the integration guide.
|
|
137
|
+
|
|
92
138
|
## Features
|
|
93
139
|
|
|
94
140
|
- Single `evaluate()` method — send prompt/response, get verdict
|
package/dist/index.d.ts
CHANGED
|
@@ -1,5 +1,5 @@
|
|
|
1
|
-
import { DriftgardConfig, EvaluateRequest, EvaluateResponse } from "./types";
|
|
2
|
-
export { DriftgardConfig, EvaluateRequest, EvaluateResponse, CircuitBreakerConfig } from "./types";
|
|
1
|
+
import { DriftgardConfig, EvaluateRequest, EvaluateResponse, ToolCallRequest, OutcomeRequest } from "./types";
|
|
2
|
+
export { DriftgardConfig, EvaluateRequest, EvaluateResponse, CircuitBreakerConfig, ToolCallRequest, OutcomeRequest } from "./types";
|
|
3
3
|
export { Violation, EvaluationResult, FallbackResponse, HitlInfo } from "./types";
|
|
4
4
|
export { DriftgardError, AuthError, RateLimitError, FeatureNotAvailableError, ChainDepthExceededError } from "./errors";
|
|
5
5
|
type CircuitState = "closed" | "open" | "half-open";
|
|
@@ -22,6 +22,22 @@ export declare class Driftgard {
|
|
|
22
22
|
evaluate(req: EvaluateRequest): Promise<EvaluateResponse>;
|
|
23
23
|
/** Generate a synthetic response based on failure mode. */
|
|
24
24
|
private syntheticResponse;
|
|
25
|
+
/**
|
|
26
|
+
* Evaluate a tool/function call against your control pack's tool_rules.
|
|
27
|
+
*/
|
|
28
|
+
evaluateToolCall(req: ToolCallRequest): Promise<EvaluateResponse>;
|
|
29
|
+
/**
|
|
30
|
+
* Report the outcome of a tool execution after it was allowed.
|
|
31
|
+
*/
|
|
32
|
+
reportOutcome(evaluationId: string, projectId: string, outcome: OutcomeRequest): Promise<any>;
|
|
33
|
+
/**
|
|
34
|
+
* Wrap a tool function with Driftgard policy check.
|
|
35
|
+
* Returns a guarded version that evaluates before executing.
|
|
36
|
+
*/
|
|
37
|
+
guard<T extends (...args: any[]) => any>(toolFn: T, toolName: string, projectId: string, opts?: {
|
|
38
|
+
modelId?: string;
|
|
39
|
+
sessionId?: string;
|
|
40
|
+
}): (...args: Parameters<T>) => Promise<ReturnType<T>>;
|
|
25
41
|
/** Get current circuit breaker state (for observability). */
|
|
26
42
|
get circuitBreakerState(): {
|
|
27
43
|
state: CircuitState;
|
package/dist/index.js
CHANGED
|
@@ -58,6 +58,8 @@ class Driftgard {
|
|
|
58
58
|
...(req.control_pack_id ? { control_pack_id: req.control_pack_id } : {}),
|
|
59
59
|
...(req.control_pack_version ? { control_pack_version: req.control_pack_version } : {}),
|
|
60
60
|
...(req.dry_run ? { dry_run: req.dry_run } : {}),
|
|
61
|
+
...(req.eval_mode ? { eval_mode: req.eval_mode } : {}),
|
|
62
|
+
...(req.tool_call ? { tool_call: req.tool_call } : {}),
|
|
61
63
|
...(req.usage ? { usage: req.usage } : {}),
|
|
62
64
|
}, idempotencyKey);
|
|
63
65
|
// Success — reset circuit breaker
|
|
@@ -117,6 +119,52 @@ class Driftgard {
|
|
|
117
119
|
}),
|
|
118
120
|
};
|
|
119
121
|
}
|
|
122
|
+
/**
|
|
123
|
+
* Evaluate a tool/function call against your control pack's tool_rules.
|
|
124
|
+
*/
|
|
125
|
+
async evaluateToolCall(req) {
|
|
126
|
+
return this.evaluate({
|
|
127
|
+
project_id: req.project_id,
|
|
128
|
+
prompt: "",
|
|
129
|
+
response: "",
|
|
130
|
+
model_id: req.model_id,
|
|
131
|
+
eval_mode: "tool_call",
|
|
132
|
+
tool_call: { tool_name: req.tool_name, parameters: req.parameters || {} },
|
|
133
|
+
session_id: req.session_id,
|
|
134
|
+
parent_evaluation_id: req.parent_evaluation_id,
|
|
135
|
+
idempotency_key: req.idempotency_key,
|
|
136
|
+
});
|
|
137
|
+
}
|
|
138
|
+
/**
|
|
139
|
+
* Report the outcome of a tool execution after it was allowed.
|
|
140
|
+
*/
|
|
141
|
+
async reportOutcome(evaluationId, projectId, outcome) {
|
|
142
|
+
return this.post(`/audit/evaluations/${encodeURIComponent(evaluationId)}/outcome`, {
|
|
143
|
+
project_id: projectId,
|
|
144
|
+
...outcome,
|
|
145
|
+
});
|
|
146
|
+
}
|
|
147
|
+
/**
|
|
148
|
+
* Wrap a tool function with Driftgard policy check.
|
|
149
|
+
* Returns a guarded version that evaluates before executing.
|
|
150
|
+
*/
|
|
151
|
+
guard(toolFn, toolName, projectId, opts) {
|
|
152
|
+
const dg = this;
|
|
153
|
+
return async (...args) => {
|
|
154
|
+
const parameters = args.length === 1 && typeof args[0] === "object" ? args[0] : { args };
|
|
155
|
+
const result = await dg.evaluateToolCall({
|
|
156
|
+
project_id: projectId,
|
|
157
|
+
model_id: opts?.modelId || "agent",
|
|
158
|
+
tool_name: toolName,
|
|
159
|
+
parameters,
|
|
160
|
+
session_id: opts?.sessionId,
|
|
161
|
+
});
|
|
162
|
+
if (!result.evaluation.allowed) {
|
|
163
|
+
throw new errors_1.DriftgardError(result.fallback?.message || `Tool "${toolName}" blocked by policy`, 403, "tool_blocked");
|
|
164
|
+
}
|
|
165
|
+
return toolFn(...args);
|
|
166
|
+
};
|
|
167
|
+
}
|
|
120
168
|
/** Get current circuit breaker state (for observability). */
|
|
121
169
|
get circuitBreakerState() {
|
|
122
170
|
return { state: this.cbState, failures: this.cbFailures, openedAt: this.cbOpenedAt };
|
package/dist/types.d.ts
CHANGED
|
@@ -26,6 +26,11 @@ export interface EvaluateRequest {
|
|
|
26
26
|
control_pack_id?: string;
|
|
27
27
|
control_pack_version?: number;
|
|
28
28
|
dry_run?: boolean;
|
|
29
|
+
eval_mode?: "text" | "tool_call";
|
|
30
|
+
tool_call?: {
|
|
31
|
+
tool_name: string;
|
|
32
|
+
parameters?: Record<string, any>;
|
|
33
|
+
};
|
|
29
34
|
/** Caller-provided idempotency key. If omitted, the SDK generates one. */
|
|
30
35
|
idempotency_key?: string;
|
|
31
36
|
usage?: {
|
|
@@ -103,4 +108,36 @@ export interface EvaluateResponse {
|
|
|
103
108
|
hitl?: HitlInfo;
|
|
104
109
|
evaluation: EvaluationResult;
|
|
105
110
|
fallback?: FallbackResponse;
|
|
111
|
+
cost_alert?: {
|
|
112
|
+
type: string;
|
|
113
|
+
scope: string;
|
|
114
|
+
session_id?: string;
|
|
115
|
+
model_id?: string;
|
|
116
|
+
window: string;
|
|
117
|
+
threshold_usd: number;
|
|
118
|
+
actual_usd: number;
|
|
119
|
+
};
|
|
120
|
+
eval_mode?: string;
|
|
121
|
+
tool_call?: {
|
|
122
|
+
tool_name: string;
|
|
123
|
+
parameters?: Record<string, any>;
|
|
124
|
+
};
|
|
125
|
+
}
|
|
126
|
+
export interface ToolCallRequest {
|
|
127
|
+
project_id: string;
|
|
128
|
+
model_id: string;
|
|
129
|
+
tool_name: string;
|
|
130
|
+
parameters?: Record<string, any>;
|
|
131
|
+
session_id?: string;
|
|
132
|
+
parent_evaluation_id?: string;
|
|
133
|
+
idempotency_key?: string;
|
|
134
|
+
}
|
|
135
|
+
export interface OutcomeRequest {
|
|
136
|
+
execution_status: "success" | "failed" | "rolled_back";
|
|
137
|
+
duration_ms?: number;
|
|
138
|
+
error_code?: string;
|
|
139
|
+
error_message?: string;
|
|
140
|
+
affected_resources?: string[];
|
|
141
|
+
compensation_action?: string;
|
|
142
|
+
external_ref?: string;
|
|
106
143
|
}
|
package/package.json
CHANGED