@attest-ai/core 1.0.0-rc.1 → 1.0.0-rc.2
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/dist/cjs/index.cjs +285 -110
- package/dist/cjs/index.cjs.map +1 -1
- package/dist/cjs/index.d.cts +129 -24
- package/dist/cjs/proto/index.cjs +14 -0
- package/dist/cjs/proto/index.cjs.map +1 -1
- package/dist/cjs/proto/index.d.cts +74 -1
- package/dist/esm/cli.js +2 -2
- package/dist/esm/cli.js.map +1 -1
- package/dist/esm/index.d.ts +129 -24
- package/dist/esm/index.js +275 -109
- package/dist/esm/index.js.map +1 -1
- package/dist/esm/proto/index.d.ts +74 -1
- package/dist/esm/proto/index.js +11 -0
- package/dist/esm/proto/index.js.map +1 -1
- package/package.json +1 -1
package/dist/cjs/index.d.cts
CHANGED
|
@@ -1,12 +1,12 @@
|
|
|
1
|
-
import { InitializeResult, ProtocolLogger, ProtocolDiagnostic, Trace, Assertion, EvaluateBatchResult, DriftReport, SimulatePersona, ConversationMessage, SimulateFaultConfig, Step, AssertionResult, Outcome } from './proto/index.cjs';
|
|
2
|
-
export { AgentId, AssertionId, AssertionSpec, BiasProbe, ConstraintSpec, ContentSpec, ERR_ASSERTION_ERROR, ERR_ENGINE_ERROR, ERR_INVALID_TRACE, ERR_PROVIDER_ERROR, ERR_SESSION_ERROR, ERR_TIMEOUT, EmbeddingSpec, EngineTimeoutError, ErrorData, EvaluateBatchParams, FailureClass, GenerateUserMessageParams, GenerateUserMessageResult, InitializeParams, JudgeAgreement, JudgeMetadata, LlmJudgeSpec, OUTCOME_FAIL, OUTCOME_PASS, OUTCOME_SKIP, PluginSpec, ProtocolDesyncError, ProtocolDiagnosticBuffer, ProtocolDiagnosticKind, ProtocolError, QueryDriftParams, QueryDriftResult, REASON_MISSING_EVIDENCE, REASON_PROVIDER_NOT_CONFIGURED, REASON_PROVIDER_UNAVAILABLE, REASON_SHORT_CIRCUITED, REASON_SIMULATION, RPCError, RPCRequest, RPCResponse, STATUS_HARD_FAIL, STATUS_NOT_EVALUATED, STATUS_PASS, STATUS_SOFT_FAIL, STEP_AGENT_CALL, STEP_LLM_CALL, STEP_RETRIEVAL, STEP_TOOL_CALL, SchemaSpec, ShutdownResult, SubmitPluginResultParams, TYPE_CONSTRAINT, TYPE_CONTENT, TYPE_EMBEDDING, TYPE_LLM_JUDGE, TYPE_PLUGIN, TYPE_SCHEMA, TYPE_TRACE, TYPE_TRACE_TREE, TraceId, TraceMetadata, TraceSpec, TraceTreeSpec, agentId, assertionId, decodeResponse, encodeRequest, extractId, extractResult, layerForType, previewLine, traceId } from './proto/index.cjs';
|
|
1
|
+
import { InitializeResult, ProtocolLogger, ProtocolDiagnostic, Trace, Assertion, BatchBudget, EvaluateBatchResult, DriftReport, SimulatePersona, ConversationMessage, SimulateFaultConfig, Step, AssertionResult, Usage, Outcome } from './proto/index.cjs';
|
|
2
|
+
export { AgentId, AssertionId, AssertionSpec, BiasProbe, ConstraintSpec, ContentSpec, ERR_ASSERTION_ERROR, ERR_ENGINE_ERROR, ERR_INVALID_TRACE, ERR_PROVIDER_ERROR, ERR_SESSION_ERROR, ERR_TIMEOUT, EmbeddingSpec, EngineTimeoutError, ErrorData, EvaluateBatchParams, Exchange, FailureClass, GenerateUserMessageParams, GenerateUserMessageResult, Handoff, InitializeParams, JudgeAgreement, JudgeMetadata, LlmJudgeSpec, OTLPConversionError, OUTCOME_FAIL, OUTCOME_PASS, OUTCOME_SKIP, PluginSpec, ProtocolDesyncError, ProtocolDiagnosticBuffer, ProtocolDiagnosticKind, ProtocolError, QueryDriftParams, QueryDriftResult, REASON_BUDGET_EXHAUSTED, REASON_JUDGE_ABSTAINED, REASON_MISSING_EVIDENCE, REASON_PROVIDER_NOT_CONFIGURED, REASON_PROVIDER_UNAVAILABLE, REASON_SHORT_CIRCUITED, REASON_SIMULATION, RPCError, RPCRequest, RPCResponse, STATUS_HARD_FAIL, STATUS_NOT_EVALUATED, STATUS_PASS, STATUS_SOFT_FAIL, STEP_AGENT_CALL, STEP_LLM_CALL, STEP_RETRIEVAL, STEP_TOOL_CALL, SchemaSpec, ShutdownResult, StepError, SubmitPluginResultParams, TYPE_CONSTRAINT, TYPE_CONTENT, TYPE_EMBEDDING, TYPE_LLM_JUDGE, TYPE_PLUGIN, TYPE_SCHEMA, TYPE_TRACE, TYPE_TRACE_TREE, TraceId, TraceMetadata, TraceSpec, TraceTreeSpec, Turn, agentId, assertionId, decodeResponse, encodeRequest, extractId, extractResult, layerForType, previewLine, traceId } from './proto/index.cjs';
|
|
3
3
|
import { ChildProcess } from 'node:child_process';
|
|
4
4
|
import { Interface } from 'node:readline';
|
|
5
5
|
import { AsyncLocalStorage } from 'node:async_hooks';
|
|
6
6
|
import { EventEmitter } from 'node:events';
|
|
7
7
|
|
|
8
|
-
declare const VERSION = "1.0.0-rc.
|
|
9
|
-
declare const ENGINE_VERSION = "1.0.0-rc.
|
|
8
|
+
declare const VERSION = "1.0.0-rc.2";
|
|
9
|
+
declare const ENGINE_VERSION = "1.0.0-rc.2";
|
|
10
10
|
|
|
11
11
|
declare class EngineManager {
|
|
12
12
|
private enginePath;
|
|
@@ -83,6 +83,7 @@ declare class AttestClient {
|
|
|
83
83
|
sendRequest(method: string, params: Record<string, unknown>, timeoutMs?: number): Promise<unknown>;
|
|
84
84
|
evaluateBatch(trace: Trace, assertions: readonly Assertion[], options?: {
|
|
85
85
|
timeout?: number;
|
|
86
|
+
budget?: BatchBudget;
|
|
86
87
|
}): Promise<EvaluateBatchResult>;
|
|
87
88
|
submitPluginResult(traceId: string, pluginName: string, assertionId: string, status: string, score: number, explanation: string, options?: {
|
|
88
89
|
timeout?: number;
|
|
@@ -146,7 +147,9 @@ declare class AgentResult {
|
|
|
146
147
|
readonly assertionResults: readonly AssertionResult[];
|
|
147
148
|
readonly totalCost: number;
|
|
148
149
|
readonly totalDurationMs: number;
|
|
149
|
-
|
|
150
|
+
/** Attest model usage of the evaluation that produced this result. */
|
|
151
|
+
readonly usage: Usage;
|
|
152
|
+
constructor(trace: Trace, assertionResults?: readonly AssertionResult[], totalCost?: number, totalDurationMs?: number, usage?: Usage);
|
|
150
153
|
/** True if every assertion has status "pass". */
|
|
151
154
|
get passed(): boolean;
|
|
152
155
|
/**
|
|
@@ -177,6 +180,29 @@ declare class AgentResult {
|
|
|
177
180
|
skipReason(): string;
|
|
178
181
|
}
|
|
179
182
|
|
|
183
|
+
/** Usage of an evaluation that made no model calls. */
|
|
184
|
+
declare function zeroUsage(): Usage;
|
|
185
|
+
/** Field-wise sum of two usages. */
|
|
186
|
+
declare function addUsage(a: Usage, b: Usage): Usage;
|
|
187
|
+
/** Field-wise sum of any number of usages; zero for none. */
|
|
188
|
+
declare function sumUsage(usages: Iterable<Usage>): Usage;
|
|
189
|
+
/**
|
|
190
|
+
* Validate the `usage` object of an evaluate_batch result: all nine fields
|
|
191
|
+
* present, non-negative, integers except the two `*_usd` amounts.
|
|
192
|
+
*/
|
|
193
|
+
declare function parseUsage(raw: unknown): Usage;
|
|
194
|
+
/**
|
|
195
|
+
* "{model_calls} model calls · {tokens} tokens ({cached} cached) · ${cost} · retry cost ${retry}",
|
|
196
|
+
* plus refused and unpriced calls when present. Same format as the engine's
|
|
197
|
+
* report.FormatUsage.
|
|
198
|
+
*/
|
|
199
|
+
declare function formatUsage(u: Usage): string;
|
|
200
|
+
/**
|
|
201
|
+
* The run-summary line "attest usage: …" when the session made a model call
|
|
202
|
+
* or refused one by budget; undefined otherwise.
|
|
203
|
+
*/
|
|
204
|
+
declare function usageSummaryLine(u: Usage): string | undefined;
|
|
205
|
+
|
|
180
206
|
/**
|
|
181
207
|
* Terminal-grade rendering of Attest assertion failures.
|
|
182
208
|
*
|
|
@@ -220,6 +246,8 @@ declare class TraceTree {
|
|
|
220
246
|
get aggregateLatency(): number;
|
|
221
247
|
}
|
|
222
248
|
|
|
249
|
+
type TrajectoryMode = "strict" | "unordered" | "subset" | "superset";
|
|
250
|
+
type ArgOperator = "eq" | "ne" | "lt" | "lte" | "gt" | "gte" | "in" | "not_in";
|
|
223
251
|
declare class ExpectChain {
|
|
224
252
|
private readonly _result;
|
|
225
253
|
private readonly _assertions;
|
|
@@ -484,6 +512,59 @@ declare class ExpectChain {
|
|
|
484
512
|
forbiddenTools(tools: string[], opts?: {
|
|
485
513
|
soft?: boolean;
|
|
486
514
|
}): this;
|
|
515
|
+
/**
|
|
516
|
+
* Assert how many times a tool is called, per value of argument `per` when set.
|
|
517
|
+
*
|
|
518
|
+
* @example
|
|
519
|
+
* ```ts
|
|
520
|
+
* // fails with "issue_refund called 2× for order_id=O-1042 (max 1)"
|
|
521
|
+
* expect(result).toolCardinality("issue_refund", { per: "order_id", atMost: 1 });
|
|
522
|
+
* ```
|
|
523
|
+
*/
|
|
524
|
+
toolCardinality(tool: string, opts: {
|
|
525
|
+
per?: string;
|
|
526
|
+
atMost?: number;
|
|
527
|
+
atLeast?: number;
|
|
528
|
+
soft?: boolean;
|
|
529
|
+
}): this;
|
|
530
|
+
/**
|
|
531
|
+
* Assert the tool-call trajectory matches `tools`: `strict` (same sequence,
|
|
532
|
+
* the default), `unordered` (same calls, any order), `subset` (no call
|
|
533
|
+
* outside `tools`) or `superset` (every call in `tools` made). Use
|
|
534
|
+
* {@link trajectoryPartialOrder} for ordering constraints.
|
|
535
|
+
*/
|
|
536
|
+
trajectory(tools: string[], opts?: {
|
|
537
|
+
mode?: TrajectoryMode;
|
|
538
|
+
soft?: boolean;
|
|
539
|
+
}): this;
|
|
540
|
+
/**
|
|
541
|
+
* Assert every tool in `tools` and `order` is called and each
|
|
542
|
+
* `[before, after]` pair holds as in {@link toolCalledBefore}.
|
|
543
|
+
*/
|
|
544
|
+
trajectoryPartialOrder(order: Array<[string, string]>, opts?: {
|
|
545
|
+
tools?: string[];
|
|
546
|
+
soft?: boolean;
|
|
547
|
+
}): this;
|
|
548
|
+
/** Assert every call to `after` has an earlier call to `before`; holds when `after` is never called. */
|
|
549
|
+
toolCalledBefore(before: string, after: string, opts?: {
|
|
550
|
+
soft?: boolean;
|
|
551
|
+
}): this;
|
|
552
|
+
/** Assert no block of `period` (1–4) tool calls repeats `minRepeats` (default 2) times in a row; A→B→A→B fails period 2. */
|
|
553
|
+
noToolCycles(period: number, opts?: {
|
|
554
|
+
minRepeats?: number;
|
|
555
|
+
soft?: boolean;
|
|
556
|
+
}): this;
|
|
557
|
+
/** Assert at most `maxStreak` consecutive tool calls end in error (trace v2). */
|
|
558
|
+
maxToolFailureStreak(maxStreak: number, opts?: {
|
|
559
|
+
soft?: boolean;
|
|
560
|
+
}): this;
|
|
561
|
+
/**
|
|
562
|
+
* Assert every call to `tool` has argument `arg` (a dot path) satisfying
|
|
563
|
+
* `op value`; holds when `tool` is never called.
|
|
564
|
+
*/
|
|
565
|
+
toolArgSatisfies(tool: string, arg: string, op: ArgOperator, value: unknown, opts?: {
|
|
566
|
+
soft?: boolean;
|
|
567
|
+
}): this;
|
|
487
568
|
/**
|
|
488
569
|
* Assert the output message contains the given string.
|
|
489
570
|
*
|
|
@@ -1163,27 +1244,51 @@ declare class OllamaAdapter {
|
|
|
1163
1244
|
}): Trace;
|
|
1164
1245
|
}
|
|
1165
1246
|
|
|
1166
|
-
|
|
1167
|
-
|
|
1168
|
-
}
|
|
1247
|
+
type AttributeValue = string | number | boolean | readonly (string | number | boolean | null | undefined)[];
|
|
1248
|
+
/** The parts of an `@opentelemetry/sdk-trace-base` `ReadableSpan` the adapter reads. */
|
|
1169
1249
|
interface OTelSpan {
|
|
1170
|
-
name: string;
|
|
1171
|
-
|
|
1172
|
-
|
|
1173
|
-
|
|
1174
|
-
|
|
1175
|
-
|
|
1250
|
+
readonly name: string;
|
|
1251
|
+
/** `[seconds, nanoseconds]` */
|
|
1252
|
+
readonly startTime: readonly [number, number];
|
|
1253
|
+
readonly endTime: readonly [number, number];
|
|
1254
|
+
readonly attributes?: Readonly<Record<string, AttributeValue | undefined>>;
|
|
1255
|
+
/** Parent span id (sdk-trace-base 1.x). */
|
|
1256
|
+
readonly parentSpanId?: string;
|
|
1257
|
+
/** Parent span context (sdk-trace-base 2.x). */
|
|
1258
|
+
readonly parentSpanContext?: {
|
|
1259
|
+
readonly spanId: string;
|
|
1260
|
+
};
|
|
1261
|
+
readonly status?: {
|
|
1262
|
+
readonly code: number;
|
|
1263
|
+
readonly message?: string;
|
|
1264
|
+
};
|
|
1265
|
+
spanContext(): {
|
|
1266
|
+
readonly traceId: string;
|
|
1267
|
+
readonly spanId: string;
|
|
1268
|
+
};
|
|
1176
1269
|
}
|
|
1270
|
+
/** An OTLP/JSON export: a parsed `ExportTraceServiceRequest`, its JSON text, or its bytes. */
|
|
1271
|
+
type OTLPInput = Record<string, unknown> | string | Uint8Array;
|
|
1177
1272
|
declare class OTelAdapter {
|
|
1178
|
-
|
|
1179
|
-
|
|
1180
|
-
|
|
1181
|
-
|
|
1182
|
-
|
|
1183
|
-
|
|
1184
|
-
|
|
1185
|
-
|
|
1186
|
-
|
|
1273
|
+
/**
|
|
1274
|
+
* Build a trace from finished OpenTelemetry spans of one trace.
|
|
1275
|
+
*
|
|
1276
|
+
* @param agentId Trace `agent_id`; defaults to the first turn's `gen_ai.agent.name`.
|
|
1277
|
+
* @throws OTLPConversionError when the engine rejects the spans.
|
|
1278
|
+
*/
|
|
1279
|
+
static fromSpans(spans: readonly OTelSpan[], agentId?: string): Promise<Trace>;
|
|
1280
|
+
/**
|
|
1281
|
+
* Convert an OTLP/JSON export (`resourceSpans` → `scopeSpans` → `spans`).
|
|
1282
|
+
*
|
|
1283
|
+
* @param agentId Trace `agent_id`; defaults to the first turn's `gen_ai.agent.name`.
|
|
1284
|
+
* @throws OTLPConversionError when the engine rejects the export (malformed JSON, invalid
|
|
1285
|
+
* ids or timestamps, more than one trace, non-numeric token counts, ...).
|
|
1286
|
+
* @throws RangeError when a parsed export holds a non-finite number.
|
|
1287
|
+
* @throws OTLPConversionError when the engine times out, overflows the output buffer or
|
|
1288
|
+
* prints no trace.
|
|
1289
|
+
* @throws Error from `spawnSync` when the engine cannot start.
|
|
1290
|
+
*/
|
|
1291
|
+
static fromOtlp(otlp: OTLPInput, agentId?: string): Promise<Trace>;
|
|
1187
1292
|
}
|
|
1188
1293
|
|
|
1189
1294
|
/**
|
|
@@ -1250,4 +1355,4 @@ interface TraceAdapter {
|
|
|
1250
1355
|
traceFromResponse(response: unknown, options?: Record<string, unknown>): Trace;
|
|
1251
1356
|
}
|
|
1252
1357
|
|
|
1253
|
-
export { ADVERSARIAL_USER, Agent, AgentResult, type AgreementResult, AlertDispatcher, type AlertPayload, AnthropicAdapter, Assertion, AssertionResult, AttestClient, type AttestClientOptions, type AttestPlugin, CONFUSED_USER, COOPERATIVE_USER, ContinuousEvalRunner, ConversationMessage, DriftReport, ENGINE_VERSION, EngineManager, EvaluateBatchResult, ExpectChain, FRIENDLY_USER, GeminiAdapter, InitializeResult, type LabelPair, type LabeledRecord, LangChainAdapter, ManualAdapter, OTelAdapter, OllamaAdapter, OpenAIAdapter, Outcome, PluginRegistry, type PluginResult, ProtocolDiagnostic, ProtocolLogger, type RenderOptions, Sampler, SimulateFaultConfig, SimulatePersona, Step, TIER_1, TIER_2, TIER_3, Trace, type TraceAdapter, TraceBuilder, TraceTree, VERSION, activeBuilder, agent, attestExpect, computeAgreement, config, delegate, executePluginAssertion, getSampleRate, isSimulationMode, loadLabelsCSV, loadLabelsJSONL, promptHash, renderAssertionFailure, renderDiagnostics, renderSummary, requireSimulationCapability, resetConfig, simulationEnvEnabled, tier };
|
|
1358
|
+
export { ADVERSARIAL_USER, Agent, AgentResult, type AgreementResult, AlertDispatcher, type AlertPayload, AnthropicAdapter, Assertion, AssertionResult, AttestClient, type AttestClientOptions, type AttestPlugin, BatchBudget, CONFUSED_USER, COOPERATIVE_USER, ContinuousEvalRunner, ConversationMessage, DriftReport, ENGINE_VERSION, EngineManager, EvaluateBatchResult, ExpectChain, FRIENDLY_USER, GeminiAdapter, InitializeResult, type LabelPair, type LabeledRecord, LangChainAdapter, ManualAdapter, OTelAdapter, OllamaAdapter, OpenAIAdapter, Outcome, PluginRegistry, type PluginResult, ProtocolDiagnostic, ProtocolLogger, type RenderOptions, Sampler, SimulateFaultConfig, SimulatePersona, Step, TIER_1, TIER_2, TIER_3, Trace, type TraceAdapter, TraceBuilder, TraceTree, Usage, VERSION, activeBuilder, addUsage, agent, attestExpect, computeAgreement, config, delegate, executePluginAssertion, formatUsage, getSampleRate, isSimulationMode, loadLabelsCSV, loadLabelsJSONL, parseUsage, promptHash, renderAssertionFailure, renderDiagnostics, renderSummary, requireSimulationCapability, resetConfig, simulationEnvEnabled, sumUsage, tier, usageSummaryLine, zeroUsage };
|
package/dist/cjs/proto/index.cjs
CHANGED
|
@@ -28,12 +28,15 @@ __export(proto_exports, {
|
|
|
28
28
|
ERR_TIMEOUT: () => ERR_TIMEOUT,
|
|
29
29
|
EngineTimeoutError: () => EngineTimeoutError,
|
|
30
30
|
FailureClass: () => FailureClass,
|
|
31
|
+
OTLPConversionError: () => OTLPConversionError,
|
|
31
32
|
OUTCOME_FAIL: () => OUTCOME_FAIL,
|
|
32
33
|
OUTCOME_PASS: () => OUTCOME_PASS,
|
|
33
34
|
OUTCOME_SKIP: () => OUTCOME_SKIP,
|
|
34
35
|
ProtocolDesyncError: () => ProtocolDesyncError,
|
|
35
36
|
ProtocolDiagnosticBuffer: () => ProtocolDiagnosticBuffer,
|
|
36
37
|
ProtocolError: () => ProtocolError,
|
|
38
|
+
REASON_BUDGET_EXHAUSTED: () => REASON_BUDGET_EXHAUSTED,
|
|
39
|
+
REASON_JUDGE_ABSTAINED: () => REASON_JUDGE_ABSTAINED,
|
|
37
40
|
REASON_MISSING_EVIDENCE: () => REASON_MISSING_EVIDENCE,
|
|
38
41
|
REASON_PROVIDER_NOT_CONFIGURED: () => REASON_PROVIDER_NOT_CONFIGURED,
|
|
39
42
|
REASON_PROVIDER_UNAVAILABLE: () => REASON_PROVIDER_UNAVAILABLE,
|
|
@@ -81,6 +84,8 @@ var REASON_PROVIDER_NOT_CONFIGURED = "provider_not_configured";
|
|
|
81
84
|
var REASON_SIMULATION = "simulation";
|
|
82
85
|
var REASON_MISSING_EVIDENCE = "missing_evidence";
|
|
83
86
|
var REASON_SHORT_CIRCUITED = "short_circuited";
|
|
87
|
+
var REASON_BUDGET_EXHAUSTED = "budget_exhausted";
|
|
88
|
+
var REASON_JUDGE_ABSTAINED = "judge_abstained";
|
|
84
89
|
var OUTCOME_PASS = "pass";
|
|
85
90
|
var OUTCOME_FAIL = "fail";
|
|
86
91
|
var OUTCOME_SKIP = "skip";
|
|
@@ -171,6 +176,12 @@ var ProtocolDesyncError = class extends Error {
|
|
|
171
176
|
this.diagnostics = diagnostics;
|
|
172
177
|
}
|
|
173
178
|
};
|
|
179
|
+
var OTLPConversionError = class extends Error {
|
|
180
|
+
constructor(message) {
|
|
181
|
+
super(message);
|
|
182
|
+
this.name = "OTLPConversionError";
|
|
183
|
+
}
|
|
184
|
+
};
|
|
174
185
|
|
|
175
186
|
// src/proto/diagnostics.ts
|
|
176
187
|
var DEFAULT_BUFFER_SIZE = 32;
|
|
@@ -285,12 +296,15 @@ function extractId(response) {
|
|
|
285
296
|
ERR_TIMEOUT,
|
|
286
297
|
EngineTimeoutError,
|
|
287
298
|
FailureClass,
|
|
299
|
+
OTLPConversionError,
|
|
288
300
|
OUTCOME_FAIL,
|
|
289
301
|
OUTCOME_PASS,
|
|
290
302
|
OUTCOME_SKIP,
|
|
291
303
|
ProtocolDesyncError,
|
|
292
304
|
ProtocolDiagnosticBuffer,
|
|
293
305
|
ProtocolError,
|
|
306
|
+
REASON_BUDGET_EXHAUSTED,
|
|
307
|
+
REASON_JUDGE_ABSTAINED,
|
|
294
308
|
REASON_MISSING_EVIDENCE,
|
|
295
309
|
REASON_PROVIDER_NOT_CONFIGURED,
|
|
296
310
|
REASON_PROVIDER_UNAVAILABLE,
|
|
@@ -1 +1 @@
|
|
|
1
|
-
{"version":3,"sources":["../../../src/proto/index.ts","../../../src/proto/constants.ts","../../../src/proto/types.ts","../../../src/proto/errors.ts","../../../src/proto/diagnostics.ts","../../../src/proto/codec.ts"],"sourcesContent":["export * from \"./constants.js\";\nexport * from \"./types.js\";\nexport * from \"./errors.js\";\nexport {\n ProtocolDiagnosticBuffer,\n previewLine,\n} from \"./diagnostics.js\";\nexport type {\n ProtocolDiagnostic,\n ProtocolDiagnosticKind,\n ProtocolLogger,\n} from \"./diagnostics.js\";\nexport { encodeRequest, decodeResponse, extractResult, extractId } from \"./codec.js\";\n","// Step type constants\nexport const STEP_LLM_CALL = \"llm_call\" as const;\nexport const STEP_TOOL_CALL = \"tool_call\" as const;\nexport const STEP_RETRIEVAL = \"retrieval\" as const;\nexport const STEP_AGENT_CALL = \"agent_call\" as const;\n\n// Status constants\nexport const STATUS_PASS = \"pass\" as const;\nexport const STATUS_SOFT_FAIL = \"soft_fail\" as const;\nexport const STATUS_HARD_FAIL = \"hard_fail\" as const;\nexport const STATUS_NOT_EVALUATED = \"not_evaluated\" as const;\n\n// Reasons carried by a not_evaluated result (AssertionResult.reason)\nexport const REASON_PROVIDER_UNAVAILABLE = \"provider_unavailable\" as const;\nexport const REASON_PROVIDER_NOT_CONFIGURED = \"provider_not_configured\" as const;\nexport const REASON_SIMULATION = \"simulation\" as const;\nexport const REASON_MISSING_EVIDENCE = \"missing_evidence\" as const;\nexport const REASON_SHORT_CIRCUITED = \"short_circuited\" as const;\n\n// Test outcomes derived from a batch of assertion results (AgentResult.outcome)\nexport const OUTCOME_PASS = \"pass\" as const;\nexport const OUTCOME_FAIL = \"fail\" as const;\nexport const OUTCOME_SKIP = \"skip\" as const;\nexport type Outcome = typeof OUTCOME_PASS | typeof OUTCOME_FAIL | typeof OUTCOME_SKIP;\n\n// Assertion type constants\nexport const TYPE_SCHEMA = \"schema\" as const;\nexport const TYPE_CONSTRAINT = \"constraint\" as const;\nexport const TYPE_TRACE = \"trace\" as const;\nexport const TYPE_CONTENT = \"content\" as const;\nexport const TYPE_EMBEDDING = \"embedding\" as const;\nexport const TYPE_LLM_JUDGE = \"llm_judge\" as const;\nexport const TYPE_TRACE_TREE = \"trace_tree\" as const;\nexport const TYPE_PLUGIN = \"plugin\" as const;\n\n// Error code constants\nexport const ERR_INVALID_TRACE = 1001 as const;\nexport const ERR_ASSERTION_ERROR = 1002 as const;\nexport const ERR_PROVIDER_ERROR = 2001 as const;\nexport const ERR_ENGINE_ERROR = 3001 as const;\nexport const ERR_TIMEOUT = 3002 as const;\nexport const ERR_SESSION_ERROR = 3003 as const;\n","// ---------------------------------------------------------------------------\n// Branded ID types (T6)\n// ---------------------------------------------------------------------------\n\ndeclare const __brand: unique symbol;\ntype Brand<T, B extends string> = T & { readonly [__brand]: B };\n\nexport type TraceId = Brand<string, \"TraceId\">;\nexport type AssertionId = Brand<string, \"AssertionId\">;\nexport type AgentId = Brand<string, \"AgentId\">;\n\nexport function traceId(id: string): TraceId {\n return id as TraceId;\n}\n\nexport function assertionId(id: string): AssertionId {\n return id as AssertionId;\n}\n\nexport function agentId(id: string): AgentId {\n return id as AgentId;\n}\n\n// ---------------------------------------------------------------------------\n// Assertion spec discriminated unions (T5)\n// ---------------------------------------------------------------------------\n\nexport interface SchemaSpec {\n readonly type: \"schema\";\n readonly target: string;\n readonly schema: Record<string, unknown>;\n}\n\nexport interface ConstraintSpec {\n readonly type: \"constraint\";\n readonly field: string;\n readonly operator: string;\n readonly value?: number;\n readonly min?: number;\n readonly max?: number;\n readonly soft: boolean;\n}\n\nexport interface TraceSpec {\n readonly type: \"trace\";\n readonly check: string;\n readonly tools?: string[];\n readonly tool?: string;\n readonly max_repetitions?: number;\n readonly soft: boolean;\n}\n\nexport interface ContentSpec {\n readonly type: \"content\";\n readonly target: string;\n readonly check: string;\n readonly value?: string;\n readonly values?: string[];\n readonly case_sensitive?: boolean;\n readonly soft?: boolean;\n}\n\nexport interface EmbeddingSpec {\n readonly type: \"embedding\";\n readonly target: string;\n readonly reference: string;\n readonly threshold: number | \"dynamic\";\n readonly model?: string;\n readonly soft: boolean;\n}\n\nexport interface LlmJudgeSpec {\n readonly type: \"llm_judge\";\n readonly target: string;\n readonly criteria: string;\n readonly rubric: string;\n readonly threshold: number | \"dynamic\";\n readonly model?: string;\n readonly soft: boolean;\n}\n\nexport interface TraceTreeSpec {\n readonly type: \"trace_tree\";\n readonly check: string;\n readonly [key: string]: unknown;\n}\n\nexport interface PluginSpec {\n readonly type: \"plugin\";\n readonly plugin_id: string;\n readonly config?: Record<string, unknown>;\n readonly soft: boolean;\n}\n\nexport type AssertionSpec =\n | SchemaSpec\n | ConstraintSpec\n | TraceSpec\n | ContentSpec\n | EmbeddingSpec\n | LlmJudgeSpec\n | TraceTreeSpec\n | PluginSpec;\n\n// ---------------------------------------------------------------------------\n// Core protocol types\n// ---------------------------------------------------------------------------\n\nexport interface TraceMetadata {\n readonly total_tokens?: number;\n readonly cost_usd?: number;\n readonly latency_ms?: number;\n readonly model?: string;\n readonly timestamp?: string;\n}\n\nexport interface Step {\n readonly type: string;\n readonly name: string;\n readonly args?: Record<string, unknown>;\n readonly result?: Record<string, unknown>;\n readonly sub_trace?: Trace;\n readonly metadata?: Record<string, unknown>;\n readonly started_at_ms?: number;\n readonly ended_at_ms?: number;\n readonly agent_id?: string;\n readonly agent_role?: string;\n}\n\nexport interface Trace {\n readonly trace_id: string;\n readonly output: Record<string, unknown>;\n readonly schema_version?: number;\n readonly agent_id?: string;\n readonly input?: Record<string, unknown>;\n readonly steps: readonly Step[];\n readonly metadata?: TraceMetadata;\n readonly parent_trace_id?: string;\n}\n\nexport interface Assertion {\n readonly assertion_id: string;\n readonly type: string;\n readonly spec: Record<string, unknown>;\n readonly request_id?: string;\n}\n\n/**\n * Audit trail for an LLM-judge evaluation. Mirrors\n * engine/pkg/types/assertion.go::JudgeMetadata so report v2 payloads\n * round-trip without losing the model / rubric / variance evidence\n * reviewers need to triage a judge failure.\n */\nexport interface JudgeMetadata {\n readonly model?: string;\n readonly rubric_name?: string;\n readonly rubric_version?: string;\n readonly prompt_hash?: string;\n readonly sample_scores?: readonly number[];\n readonly score_mean?: number;\n readonly score_stddev?: number;\n readonly high_variance?: boolean;\n readonly bias_probes?: readonly BiasProbe[];\n readonly calibration?: JudgeAgreement;\n}\n\n/**\n * One judge-bias probe result. Score is the score the judge produced on\n * the mutated input; delta is `score - baseline`. A well-calibrated\n * judge has |delta| close to zero. Mirrors\n * engine/pkg/types/assertion.go::BiasProbe.\n */\nexport interface BiasProbe {\n readonly name: string;\n readonly score: number;\n readonly delta: number;\n}\n\n/**\n * Calibration agreement summary keyed by\n * (rubric_name, rubric_version, prompt_hash). Mirrors\n * engine/pkg/types/assertion.go::JudgeAgreement.\n */\nexport interface JudgeAgreement {\n readonly label_count: number;\n readonly agreement: number;\n readonly cohen_kappa: number;\n readonly roc_auc?: number;\n}\n\nexport interface AssertionResult {\n readonly assertion_id: string;\n readonly status: string;\n readonly score: number;\n readonly explanation: string;\n readonly cost?: number;\n readonly duration_ms?: number;\n readonly request_id?: string;\n /** Set only when status is \"not_evaluated\" (REASON_* constants). */\n readonly reason?: string;\n readonly threshold_source?: string;\n readonly layer?: number;\n readonly type?: string;\n readonly trace_node_path?: string;\n readonly expected?: string;\n readonly actual?: string;\n readonly suggested_action?: string;\n readonly judge_metadata?: JudgeMetadata;\n /**\n * Heuristic root-cause label set by the engine\n * (broken_code / flaky_judge / bad_rubric / missing_trace_data /\n * stochastic_variance). Empty/undefined for passes and any result the\n * engine declined to label. Mirrors\n * engine/pkg/types/failure_class.go.\n */\n readonly failure_class?: string;\n}\n\n/**\n * FailureClass values mirror engine/pkg/types/failure_class.go. Use\n * these constants when matching on AssertionResult.failure_class so a\n * typo in either codebase fails fast.\n */\nexport const FailureClass = {\n BrokenCode: \"broken_code\",\n FlakyJudge: \"flaky_judge\",\n BadRubric: \"bad_rubric\",\n MissingTraceData: \"missing_trace_data\",\n StochasticVariance: \"stochastic_variance\",\n} as const;\nexport type FailureClass = (typeof FailureClass)[keyof typeof FailureClass];\n\n/**\n * Map an assertion-type string to its pipeline layer (1–8). Returns 0\n * when the type is unknown so callers can route uncategorised results\n * into their own bucket.\n */\nexport function layerForType(assertionType: string | undefined): number {\n switch (assertionType) {\n case \"schema\":\n return 1;\n case \"constraint\":\n return 2;\n case \"trace\":\n case \"trace_tree\":\n return 3;\n case \"content\":\n return 4;\n case \"embedding\":\n return 5;\n case \"llm_judge\":\n return 6;\n case \"plugin\":\n return 8;\n default:\n return 0;\n }\n}\n\nexport interface ErrorData {\n readonly error_type: string;\n readonly retryable: boolean;\n readonly detail: string;\n}\n\nexport interface RPCError {\n readonly code: number;\n readonly message: string;\n readonly data?: ErrorData;\n}\n\nexport interface RPCRequest {\n readonly jsonrpc: \"2.0\";\n readonly id: number;\n readonly method: string;\n readonly params: Record<string, unknown>;\n}\n\nexport interface RPCResponse {\n readonly jsonrpc: \"2.0\";\n readonly id: number;\n readonly result?: unknown;\n readonly error?: RPCError;\n}\n\nexport interface InitializeParams {\n readonly sdk_name: string;\n readonly sdk_version: string;\n readonly protocol_version: number;\n readonly required_capabilities: readonly string[];\n readonly preferred_encoding?: string;\n}\n\nexport interface InitializeResult {\n readonly engine_version: string;\n readonly protocol_version: number;\n readonly capabilities: readonly string[];\n readonly missing: readonly string[];\n readonly compatible: boolean;\n readonly encoding?: string;\n readonly max_concurrent_requests?: number;\n readonly max_trace_size_bytes?: number;\n readonly max_steps_per_trace?: number;\n}\n\nexport interface EvaluateBatchParams {\n readonly trace: Trace;\n readonly assertions: readonly Assertion[];\n}\n\nexport interface EvaluateBatchResult {\n readonly results: readonly AssertionResult[];\n readonly total_cost?: number;\n readonly total_duration_ms?: number;\n /**\n * True when this result was produced by an SDK simulation shim rather\n * than a real engine evaluation. Engine-produced results always carry\n * simulated=false. CI gates can fail builds that accidentally evaluate\n * against the simulator by checking this flag.\n */\n readonly simulated?: boolean;\n}\n\nexport interface ShutdownResult {\n readonly sessions_completed: number;\n readonly assertions_evaluated: number;\n}\n\nexport interface SubmitPluginResultParams {\n readonly trace_id: string;\n readonly plugin_name: string;\n readonly assertion_id: string;\n readonly result: AssertionResult;\n}\n\n// ---------------------------------------------------------------------------\n// Drift detection types\n// ---------------------------------------------------------------------------\n\nexport interface DriftReport {\n readonly assertion_id: string;\n readonly mean: number;\n readonly stddev: number;\n readonly count: number;\n readonly latest_score: number;\n readonly deviation: number;\n readonly status: string;\n}\n\nexport interface QueryDriftParams {\n readonly assertion_id: string;\n readonly window_size: number;\n}\n\nexport interface QueryDriftResult {\n readonly report: DriftReport;\n}\n\n// ---------------------------------------------------------------------------\n// Simulation types\n// ---------------------------------------------------------------------------\n\nexport interface SimulatePersona {\n readonly name: string;\n readonly system_prompt: string;\n readonly style: string;\n readonly temperature: number;\n readonly max_tokens?: number;\n}\n\nexport interface SimulateFaultConfig {\n readonly error_rate: number;\n readonly latency_jitter_ms: number;\n readonly content_corruption: boolean;\n readonly timeout_after_ms: number;\n}\n\nexport interface ConversationMessage {\n readonly role: string;\n readonly content: string;\n}\n\nexport interface GenerateUserMessageParams {\n readonly persona: SimulatePersona;\n readonly conversation_history: readonly ConversationMessage[];\n readonly fault_config?: SimulateFaultConfig;\n}\n\nexport interface GenerateUserMessageResult {\n readonly message: string;\n}\n","import type { ErrorData } from \"./types.js\";\nimport type { ProtocolDiagnostic } from \"./diagnostics.js\";\n\nexport class ProtocolError extends Error {\n readonly code: number;\n readonly errorMessage: string;\n readonly data: ErrorData | undefined;\n\n constructor(code: number, message: string, data?: ErrorData) {\n super(message);\n this.name = \"ProtocolError\";\n this.code = code;\n this.errorMessage = message;\n this.data = data;\n }\n}\n\nexport class EngineTimeoutError extends Error {\n readonly method: string;\n readonly timeoutMs: number;\n\n constructor(method: string, timeoutMs: number) {\n super(\n `Engine request '${method}' timed out after ${timeoutMs}ms. ` +\n \"Set ATTEST_ENGINE_TIMEOUT to increase the limit.\",\n );\n this.name = \"EngineTimeoutError\";\n this.method = method;\n this.timeoutMs = timeoutMs;\n }\n}\n\n/**\n * Raised when the reader loop detects that engine output has lost\n * framing or contains repeated unparseable lines. All in-flight requests\n * are rejected with this error so callers see a visible failure rather\n * than a hung promise.\n */\nexport class ProtocolDesyncError extends Error {\n readonly diagnostics: readonly ProtocolDiagnostic[];\n\n constructor(message: string, diagnostics: readonly ProtocolDiagnostic[]) {\n super(message);\n this.name = \"ProtocolDesyncError\";\n this.diagnostics = diagnostics;\n }\n}\n","/**\n * Protocol diagnostic types for client-side observability.\n *\n * The reader loop emits diagnostics whenever it cannot route a response\n * line: malformed JSON, missing id, unknown id, or non-routable protocol\n * errors. Diagnostics are buffered in a bounded ring (32 entries) and\n * trigger a desync error when the rate exceeds the configured threshold.\n */\n\nexport type ProtocolDiagnosticKind =\n | \"malformed_json\"\n | \"non_object_response\"\n | \"invalid_jsonrpc_version\"\n | \"missing_id\"\n | \"unknown_id\"\n | \"non_routable_error\";\n\nexport interface ProtocolDiagnostic {\n readonly kind: ProtocolDiagnosticKind;\n readonly message: string;\n readonly rawLine: string;\n readonly timestampMs: number;\n}\n\nexport interface ProtocolLogger {\n warn(message: string, ...args: unknown[]): void;\n error(message: string, ...args: unknown[]): void;\n}\n\nconst DEFAULT_BUFFER_SIZE = 32;\nconst MAX_LINE_PREVIEW = 512;\n\nexport class ProtocolDiagnosticBuffer {\n private readonly capacity: number;\n private readonly entries: ProtocolDiagnostic[] = [];\n\n constructor(capacity: number = DEFAULT_BUFFER_SIZE) {\n if (capacity <= 0) {\n throw new Error(\n `ProtocolDiagnosticBuffer capacity must be positive, got ${capacity}`,\n );\n }\n this.capacity = capacity;\n }\n\n push(diagnostic: ProtocolDiagnostic): void {\n if (this.entries.length >= this.capacity) {\n this.entries.shift();\n }\n this.entries.push(diagnostic);\n }\n\n snapshot(): readonly ProtocolDiagnostic[] {\n return this.entries.slice();\n }\n\n countWithin(windowMs: number, nowMs: number): number {\n const cutoff = nowMs - windowMs;\n let count = 0;\n for (let i = this.entries.length - 1; i >= 0; i -= 1) {\n if (this.entries[i]!.timestampMs >= cutoff) {\n count += 1;\n } else {\n break;\n }\n }\n return count;\n }\n\n clear(): void {\n this.entries.length = 0;\n }\n}\n\nexport function previewLine(line: string): string {\n if (line.length <= MAX_LINE_PREVIEW) return line;\n return line.slice(0, MAX_LINE_PREVIEW) + \"…\";\n}\n","import type { ErrorData, RPCResponse } from \"./types.js\";\nimport { ProtocolError } from \"./errors.js\";\n\n/**\n * Encode a JSON-RPC 2.0 request as an NDJSON line.\n */\nexport function encodeRequest(\n id: number,\n method: string,\n params: Record<string, unknown>,\n): string {\n const msg = {\n jsonrpc: \"2.0\" as const,\n id,\n method,\n params,\n };\n return JSON.stringify(msg) + \"\\n\";\n}\n\n/**\n * Decode a JSON-RPC 2.0 response from an NDJSON line.\n * Throws ProtocolError if the response contains an error.\n * Throws Error for malformed JSON or invalid responses.\n */\nexport function decodeResponse(line: string): RPCResponse {\n const text = line.trim();\n if (!text) {\n throw new Error(\"empty response line\");\n }\n\n let data: unknown;\n try {\n data = JSON.parse(text);\n } catch (e) {\n throw new Error(`malformed JSON response: ${e instanceof Error ? e.message : String(e)}`);\n }\n\n if (typeof data !== \"object\" || data === null || Array.isArray(data)) {\n throw new Error(`expected JSON object, got ${Array.isArray(data) ? \"array\" : typeof data}`);\n }\n\n const obj = data as Record<string, unknown>;\n\n if (obj.jsonrpc !== \"2.0\") {\n throw new Error(`invalid jsonrpc version: ${String(obj.jsonrpc)}`);\n }\n\n if (obj.error != null) {\n const err = obj.error as Record<string, unknown>;\n let errorData: ErrorData | undefined;\n\n if (err.data != null) {\n const raw = err.data as Record<string, unknown>;\n errorData = {\n error_type: (raw.error_type as string) ?? \"\",\n retryable: (raw.retryable as boolean) ?? false,\n detail: (raw.detail as string) ?? \"\",\n };\n }\n\n throw new ProtocolError(\n (err.code as number) ?? -1,\n (err.message as string) ?? \"unknown error\",\n errorData,\n );\n }\n\n return obj as unknown as RPCResponse;\n}\n\n/**\n * Extract the result field from a decoded response.\n */\nexport function extractResult(response: RPCResponse): unknown {\n if (response.result === undefined) {\n throw new Error(\"response missing 'result' field\");\n }\n return response.result;\n}\n\n/**\n * Extract the request ID from a decoded response.\n */\nexport function extractId(response: RPCResponse): number {\n if (response.id === undefined) {\n throw new Error(\"response missing 'id' field\");\n }\n return Number(response.id);\n}\n"],"mappings":";;;;;;;;;;;;;;;;;;;;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;;;ACCO,IAAM,gBAAgB;AACtB,IAAM,iBAAiB;AACvB,IAAM,iBAAiB;AACvB,IAAM,kBAAkB;AAGxB,IAAM,cAAc;AACpB,IAAM,mBAAmB;AACzB,IAAM,mBAAmB;AACzB,IAAM,uBAAuB;AAG7B,IAAM,8BAA8B;AACpC,IAAM,iCAAiC;AACvC,IAAM,oBAAoB;AAC1B,IAAM,0BAA0B;AAChC,IAAM,yBAAyB;AAG/B,IAAM,eAAe;AACrB,IAAM,eAAe;AACrB,IAAM,eAAe;AAIrB,IAAM,cAAc;AACpB,IAAM,kBAAkB;AACxB,IAAM,aAAa;AACnB,IAAM,eAAe;AACrB,IAAM,iBAAiB;AACvB,IAAM,iBAAiB;AACvB,IAAM,kBAAkB;AACxB,IAAM,cAAc;AAGpB,IAAM,oBAAoB;AAC1B,IAAM,sBAAsB;AAC5B,IAAM,qBAAqB;AAC3B,IAAM,mBAAmB;AACzB,IAAM,cAAc;AACpB,IAAM,oBAAoB;;;AC9B1B,SAAS,QAAQ,IAAqB;AAC3C,SAAO;AACT;AAEO,SAAS,YAAY,IAAyB;AACnD,SAAO;AACT;AAEO,SAAS,QAAQ,IAAqB;AAC3C,SAAO;AACT;AA0MO,IAAM,eAAe;AAAA,EAC1B,YAAY;AAAA,EACZ,YAAY;AAAA,EACZ,WAAW;AAAA,EACX,kBAAkB;AAAA,EAClB,oBAAoB;AACtB;AAQO,SAAS,aAAa,eAA2C;AACtE,UAAQ,eAAe;AAAA,IACrB,KAAK;AACH,aAAO;AAAA,IACT,KAAK;AACH,aAAO;AAAA,IACT,KAAK;AAAA,IACL,KAAK;AACH,aAAO;AAAA,IACT,KAAK;AACH,aAAO;AAAA,IACT,KAAK;AACH,aAAO;AAAA,IACT,KAAK;AACH,aAAO;AAAA,IACT,KAAK;AACH,aAAO;AAAA,IACT;AACE,aAAO;AAAA,EACX;AACF;;;AC9PO,IAAM,gBAAN,cAA4B,MAAM;AAAA,EAC9B;AAAA,EACA;AAAA,EACA;AAAA,EAET,YAAY,MAAc,SAAiB,MAAkB;AAC3D,UAAM,OAAO;AACb,SAAK,OAAO;AACZ,SAAK,OAAO;AACZ,SAAK,eAAe;AACpB,SAAK,OAAO;AAAA,EACd;AACF;AAEO,IAAM,qBAAN,cAAiC,MAAM;AAAA,EACnC;AAAA,EACA;AAAA,EAET,YAAY,QAAgB,WAAmB;AAC7C;AAAA,MACE,mBAAmB,MAAM,qBAAqB,SAAS;AAAA,IAEzD;AACA,SAAK,OAAO;AACZ,SAAK,SAAS;AACd,SAAK,YAAY;AAAA,EACnB;AACF;AAQO,IAAM,sBAAN,cAAkC,MAAM;AAAA,EACpC;AAAA,EAET,YAAY,SAAiB,aAA4C;AACvE,UAAM,OAAO;AACb,SAAK,OAAO;AACZ,SAAK,cAAc;AAAA,EACrB;AACF;;;ACjBA,IAAM,sBAAsB;AAC5B,IAAM,mBAAmB;AAElB,IAAM,2BAAN,MAA+B;AAAA,EACnB;AAAA,EACA,UAAgC,CAAC;AAAA,EAElD,YAAY,WAAmB,qBAAqB;AAClD,QAAI,YAAY,GAAG;AACjB,YAAM,IAAI;AAAA,QACR,2DAA2D,QAAQ;AAAA,MACrE;AAAA,IACF;AACA,SAAK,WAAW;AAAA,EAClB;AAAA,EAEA,KAAK,YAAsC;AACzC,QAAI,KAAK,QAAQ,UAAU,KAAK,UAAU;AACxC,WAAK,QAAQ,MAAM;AAAA,IACrB;AACA,SAAK,QAAQ,KAAK,UAAU;AAAA,EAC9B;AAAA,EAEA,WAA0C;AACxC,WAAO,KAAK,QAAQ,MAAM;AAAA,EAC5B;AAAA,EAEA,YAAY,UAAkB,OAAuB;AACnD,UAAM,SAAS,QAAQ;AACvB,QAAI,QAAQ;AACZ,aAAS,IAAI,KAAK,QAAQ,SAAS,GAAG,KAAK,GAAG,KAAK,GAAG;AACpD,UAAI,KAAK,QAAQ,CAAC,EAAG,eAAe,QAAQ;AAC1C,iBAAS;AAAA,MACX,OAAO;AACL;AAAA,MACF;AAAA,IACF;AACA,WAAO;AAAA,EACT;AAAA,EAEA,QAAc;AACZ,SAAK,QAAQ,SAAS;AAAA,EACxB;AACF;AAEO,SAAS,YAAY,MAAsB;AAChD,MAAI,KAAK,UAAU,iBAAkB,QAAO;AAC5C,SAAO,KAAK,MAAM,GAAG,gBAAgB,IAAI;AAC3C;;;ACvEO,SAAS,cACd,IACA,QACA,QACQ;AACR,QAAM,MAAM;AAAA,IACV,SAAS;AAAA,IACT;AAAA,IACA;AAAA,IACA;AAAA,EACF;AACA,SAAO,KAAK,UAAU,GAAG,IAAI;AAC/B;AAOO,SAAS,eAAe,MAA2B;AACxD,QAAM,OAAO,KAAK,KAAK;AACvB,MAAI,CAAC,MAAM;AACT,UAAM,IAAI,MAAM,qBAAqB;AAAA,EACvC;AAEA,MAAI;AACJ,MAAI;AACF,WAAO,KAAK,MAAM,IAAI;AAAA,EACxB,SAAS,GAAG;AACV,UAAM,IAAI,MAAM,4BAA4B,aAAa,QAAQ,EAAE,UAAU,OAAO,CAAC,CAAC,EAAE;AAAA,EAC1F;AAEA,MAAI,OAAO,SAAS,YAAY,SAAS,QAAQ,MAAM,QAAQ,IAAI,GAAG;AACpE,UAAM,IAAI,MAAM,6BAA6B,MAAM,QAAQ,IAAI,IAAI,UAAU,OAAO,IAAI,EAAE;AAAA,EAC5F;AAEA,QAAM,MAAM;AAEZ,MAAI,IAAI,YAAY,OAAO;AACzB,UAAM,IAAI,MAAM,4BAA4B,OAAO,IAAI,OAAO,CAAC,EAAE;AAAA,EACnE;AAEA,MAAI,IAAI,SAAS,MAAM;AACrB,UAAM,MAAM,IAAI;AAChB,QAAI;AAEJ,QAAI,IAAI,QAAQ,MAAM;AACpB,YAAM,MAAM,IAAI;AAChB,kBAAY;AAAA,QACV,YAAa,IAAI,cAAyB;AAAA,QAC1C,WAAY,IAAI,aAAyB;AAAA,QACzC,QAAS,IAAI,UAAqB;AAAA,MACpC;AAAA,IACF;AAEA,UAAM,IAAI;AAAA,MACP,IAAI,QAAmB;AAAA,MACvB,IAAI,WAAsB;AAAA,MAC3B;AAAA,IACF;AAAA,EACF;AAEA,SAAO;AACT;AAKO,SAAS,cAAc,UAAgC;AAC5D,MAAI,SAAS,WAAW,QAAW;AACjC,UAAM,IAAI,MAAM,iCAAiC;AAAA,EACnD;AACA,SAAO,SAAS;AAClB;AAKO,SAAS,UAAU,UAA+B;AACvD,MAAI,SAAS,OAAO,QAAW;AAC7B,UAAM,IAAI,MAAM,6BAA6B;AAAA,EAC/C;AACA,SAAO,OAAO,SAAS,EAAE;AAC3B;","names":[]}
|
|
1
|
+
{"version":3,"sources":["../../../src/proto/index.ts","../../../src/proto/constants.ts","../../../src/proto/types.ts","../../../src/proto/errors.ts","../../../src/proto/diagnostics.ts","../../../src/proto/codec.ts"],"sourcesContent":["export * from \"./constants.js\";\nexport * from \"./types.js\";\nexport * from \"./errors.js\";\nexport {\n ProtocolDiagnosticBuffer,\n previewLine,\n} from \"./diagnostics.js\";\nexport type {\n ProtocolDiagnostic,\n ProtocolDiagnosticKind,\n ProtocolLogger,\n} from \"./diagnostics.js\";\nexport { encodeRequest, decodeResponse, extractResult, extractId } from \"./codec.js\";\n","// Step type constants\nexport const STEP_LLM_CALL = \"llm_call\" as const;\nexport const STEP_TOOL_CALL = \"tool_call\" as const;\nexport const STEP_RETRIEVAL = \"retrieval\" as const;\nexport const STEP_AGENT_CALL = \"agent_call\" as const;\n\n// Status constants\nexport const STATUS_PASS = \"pass\" as const;\nexport const STATUS_SOFT_FAIL = \"soft_fail\" as const;\nexport const STATUS_HARD_FAIL = \"hard_fail\" as const;\nexport const STATUS_NOT_EVALUATED = \"not_evaluated\" as const;\n\n// Reasons carried by a not_evaluated result (AssertionResult.reason)\nexport const REASON_PROVIDER_UNAVAILABLE = \"provider_unavailable\" as const;\nexport const REASON_PROVIDER_NOT_CONFIGURED = \"provider_not_configured\" as const;\nexport const REASON_SIMULATION = \"simulation\" as const;\nexport const REASON_MISSING_EVIDENCE = \"missing_evidence\" as const;\nexport const REASON_SHORT_CIRCUITED = \"short_circuited\" as const;\nexport const REASON_BUDGET_EXHAUSTED = \"budget_exhausted\" as const;\nexport const REASON_JUDGE_ABSTAINED = \"judge_abstained\" as const;\n\n// Test outcomes derived from a batch of assertion results (AgentResult.outcome)\nexport const OUTCOME_PASS = \"pass\" as const;\nexport const OUTCOME_FAIL = \"fail\" as const;\nexport const OUTCOME_SKIP = \"skip\" as const;\nexport type Outcome = typeof OUTCOME_PASS | typeof OUTCOME_FAIL | typeof OUTCOME_SKIP;\n\n// Assertion type constants\nexport const TYPE_SCHEMA = \"schema\" as const;\nexport const TYPE_CONSTRAINT = \"constraint\" as const;\nexport const TYPE_TRACE = \"trace\" as const;\nexport const TYPE_CONTENT = \"content\" as const;\nexport const TYPE_EMBEDDING = \"embedding\" as const;\nexport const TYPE_LLM_JUDGE = \"llm_judge\" as const;\nexport const TYPE_TRACE_TREE = \"trace_tree\" as const;\nexport const TYPE_PLUGIN = \"plugin\" as const;\n\n// Error code constants\nexport const ERR_INVALID_TRACE = 1001 as const;\nexport const ERR_ASSERTION_ERROR = 1002 as const;\nexport const ERR_PROVIDER_ERROR = 2001 as const;\nexport const ERR_ENGINE_ERROR = 3001 as const;\nexport const ERR_TIMEOUT = 3002 as const;\nexport const ERR_SESSION_ERROR = 3003 as const;\n","// ---------------------------------------------------------------------------\n// Branded ID types (T6)\n// ---------------------------------------------------------------------------\n\ndeclare const __brand: unique symbol;\ntype Brand<T, B extends string> = T & { readonly [__brand]: B };\n\nexport type TraceId = Brand<string, \"TraceId\">;\nexport type AssertionId = Brand<string, \"AssertionId\">;\nexport type AgentId = Brand<string, \"AgentId\">;\n\nexport function traceId(id: string): TraceId {\n return id as TraceId;\n}\n\nexport function assertionId(id: string): AssertionId {\n return id as AssertionId;\n}\n\nexport function agentId(id: string): AgentId {\n return id as AgentId;\n}\n\n// ---------------------------------------------------------------------------\n// Assertion spec discriminated unions (T5)\n// ---------------------------------------------------------------------------\n\nexport interface SchemaSpec {\n readonly type: \"schema\";\n readonly target: string;\n readonly schema: Record<string, unknown>;\n}\n\nexport interface ConstraintSpec {\n readonly type: \"constraint\";\n readonly field: string;\n readonly operator: string;\n readonly value?: number;\n readonly min?: number;\n readonly max?: number;\n readonly soft: boolean;\n}\n\nexport interface TraceSpec {\n readonly type: \"trace\";\n readonly check: string;\n readonly tools?: string[];\n readonly tool?: string;\n readonly max_repetitions?: number;\n readonly soft: boolean;\n}\n\nexport interface ContentSpec {\n readonly type: \"content\";\n readonly target: string;\n readonly check: string;\n readonly value?: string;\n readonly values?: string[];\n readonly case_sensitive?: boolean;\n readonly soft?: boolean;\n}\n\nexport interface EmbeddingSpec {\n readonly type: \"embedding\";\n readonly target: string;\n readonly reference: string;\n readonly threshold: number | \"dynamic\";\n readonly model?: string;\n readonly soft: boolean;\n}\n\nexport interface LlmJudgeSpec {\n readonly type: \"llm_judge\";\n readonly target: string;\n readonly criteria: string;\n readonly rubric: string;\n readonly threshold: number | \"dynamic\";\n readonly model?: string;\n readonly soft: boolean;\n}\n\nexport interface TraceTreeSpec {\n readonly type: \"trace_tree\";\n readonly check: string;\n readonly [key: string]: unknown;\n}\n\nexport interface PluginSpec {\n readonly type: \"plugin\";\n readonly plugin_id: string;\n readonly config?: Record<string, unknown>;\n readonly soft: boolean;\n}\n\nexport type AssertionSpec =\n | SchemaSpec\n | ConstraintSpec\n | TraceSpec\n | ContentSpec\n | EmbeddingSpec\n | LlmJudgeSpec\n | TraceTreeSpec\n | PluginSpec;\n\n// ---------------------------------------------------------------------------\n// Core protocol types\n// ---------------------------------------------------------------------------\n\nexport interface TraceMetadata {\n readonly total_tokens?: number;\n readonly cost_usd?: number;\n readonly latency_ms?: number;\n readonly model?: string;\n readonly timestamp?: string;\n}\n\n/** Trace v2: a step whose span ended in error. */\nexport interface StepError {\n readonly type: string;\n readonly message?: string;\n}\n\n/** Trace v2: a recorded model request/response pair. */\nexport interface Exchange {\n readonly request: Record<string, unknown>;\n readonly response: Record<string, unknown>;\n}\n\n/** Trace v2: one conversation turn and the indices of its steps. */\nexport interface Turn {\n readonly index: number;\n readonly span_id?: string;\n readonly input?: unknown;\n readonly output?: unknown;\n readonly step_indices: readonly number[];\n}\n\n/** Trace v2: control passing from one agent to another at an agent_call step. */\nexport interface Handoff {\n readonly from_agent: string;\n readonly to_agent: string;\n readonly step_index: number;\n}\n\nexport interface Step {\n readonly type: string;\n readonly name: string;\n readonly args?: Record<string, unknown>;\n readonly result?: Record<string, unknown>;\n readonly sub_trace?: Trace;\n readonly metadata?: Record<string, unknown>;\n readonly started_at_ms?: number;\n readonly ended_at_ms?: number;\n readonly agent_id?: string;\n readonly agent_role?: string;\n // Trace v2 (schema_version 2) fields.\n readonly span_id?: string;\n readonly parent_span_id?: string;\n readonly tool_call_id?: string;\n readonly retry_of?: number;\n readonly error?: StepError;\n readonly exchange?: Exchange;\n}\n\nexport interface Trace {\n readonly trace_id: string;\n readonly output: Record<string, unknown>;\n readonly schema_version?: number;\n readonly agent_id?: string;\n readonly input?: Record<string, unknown>;\n readonly steps: readonly Step[];\n readonly metadata?: TraceMetadata;\n readonly parent_trace_id?: string;\n // Trace v2 (schema_version 2) fields.\n readonly conversation_id?: string;\n readonly turns?: readonly Turn[];\n readonly handoffs?: readonly Handoff[];\n}\n\nexport interface Assertion {\n readonly assertion_id: string;\n readonly type: string;\n readonly spec: Record<string, unknown>;\n readonly request_id?: string;\n}\n\n/**\n * Audit trail for an LLM-judge evaluation. Mirrors\n * engine/pkg/types/assertion.go::JudgeMetadata so report v2 payloads\n * round-trip without losing the model / rubric / variance evidence\n * reviewers need to triage a judge failure.\n */\nexport interface JudgeMetadata {\n readonly provider?: string;\n readonly model?: string;\n readonly rubric_name?: string;\n readonly rubric_version?: string;\n readonly prompt_hash?: string;\n readonly sample_scores?: readonly number[];\n readonly score_mean?: number;\n readonly score_stddev?: number;\n readonly high_variance?: boolean;\n readonly bias_probes?: readonly BiasProbe[];\n readonly calibration?: JudgeAgreement;\n}\n\n/**\n * One judge-bias probe result. Score is the score the judge produced on\n * the mutated input; delta is `score - baseline`. A well-calibrated\n * judge has |delta| close to zero. Mirrors\n * engine/pkg/types/assertion.go::BiasProbe.\n */\nexport interface BiasProbe {\n readonly name: string;\n readonly score: number;\n readonly delta: number;\n}\n\n/**\n * Calibration agreement summary keyed by\n * (rubric_name, rubric_version, prompt_hash). Mirrors\n * engine/pkg/types/assertion.go::JudgeAgreement.\n */\nexport interface JudgeAgreement {\n readonly label_count: number;\n readonly agreement: number;\n readonly cohen_kappa: number;\n readonly roc_auc?: number;\n}\n\nexport interface AssertionResult {\n readonly assertion_id: string;\n readonly status: string;\n readonly score: number;\n readonly explanation: string;\n readonly cost?: number;\n readonly duration_ms?: number;\n readonly request_id?: string;\n /** Set only when status is \"not_evaluated\" (REASON_* constants). */\n readonly reason?: string;\n readonly threshold_source?: string;\n readonly layer?: number;\n readonly type?: string;\n readonly trace_node_path?: string;\n readonly expected?: string;\n readonly actual?: string;\n readonly suggested_action?: string;\n readonly judge_metadata?: JudgeMetadata;\n /**\n * Heuristic root-cause label set by the engine\n * (broken_code / flaky_judge / bad_rubric / missing_trace_data /\n * stochastic_variance). Empty/undefined for passes and any result the\n * engine declined to label. Mirrors\n * engine/pkg/types/failure_class.go.\n */\n readonly failure_class?: string;\n}\n\n/**\n * FailureClass values mirror engine/pkg/types/failure_class.go. Use\n * these constants when matching on AssertionResult.failure_class so a\n * typo in either codebase fails fast.\n */\nexport const FailureClass = {\n BrokenCode: \"broken_code\",\n FlakyJudge: \"flaky_judge\",\n BadRubric: \"bad_rubric\",\n MissingTraceData: \"missing_trace_data\",\n StochasticVariance: \"stochastic_variance\",\n} as const;\nexport type FailureClass = (typeof FailureClass)[keyof typeof FailureClass];\n\n/**\n * Map an assertion-type string to its pipeline layer (1–8). Returns 0\n * when the type is unknown so callers can route uncategorised results\n * into their own bucket.\n */\nexport function layerForType(assertionType: string | undefined): number {\n switch (assertionType) {\n case \"schema\":\n return 1;\n case \"constraint\":\n return 2;\n case \"trace\":\n case \"trace_tree\":\n return 3;\n case \"content\":\n return 4;\n case \"embedding\":\n return 5;\n case \"llm_judge\":\n return 6;\n case \"plugin\":\n return 8;\n default:\n return 0;\n }\n}\n\nexport interface ErrorData {\n readonly error_type: string;\n readonly retryable: boolean;\n readonly detail: string;\n}\n\nexport interface RPCError {\n readonly code: number;\n readonly message: string;\n readonly data?: ErrorData;\n}\n\nexport interface RPCRequest {\n readonly jsonrpc: \"2.0\";\n readonly id: number;\n readonly method: string;\n readonly params: Record<string, unknown>;\n}\n\nexport interface RPCResponse {\n readonly jsonrpc: \"2.0\";\n readonly id: number;\n readonly result?: unknown;\n readonly error?: RPCError;\n}\n\nexport interface InitializeParams {\n readonly sdk_name: string;\n readonly sdk_version: string;\n readonly protocol_version: number;\n readonly required_capabilities: readonly string[];\n readonly preferred_encoding?: string;\n}\n\nexport interface InitializeResult {\n readonly engine_version: string;\n readonly protocol_version: number;\n readonly capabilities: readonly string[];\n readonly missing: readonly string[];\n readonly compatible: boolean;\n readonly encoding?: string;\n readonly max_concurrent_requests?: number;\n readonly max_trace_size_bytes?: number;\n readonly max_steps_per_trace?: number;\n}\n\n/**\n * Limits for one evaluate_batch call on top of the session budget. A judge or\n * embedding call that cannot reserve against them is not sent and its\n * assertion is not_evaluated with reason \"budget_exhausted\".\n */\nexport interface BatchBudget {\n readonly max_judge_calls?: number;\n readonly max_cost_usd?: number;\n}\n\nexport interface EvaluateBatchParams {\n readonly trace: Trace;\n readonly assertions: readonly Assertion[];\n /** Omitted when no per-batch budget applies. */\n readonly budget?: BatchBudget;\n}\n\n/**\n * Attest-initiated model usage (embeddings, judges, judge repeats, bias\n * probes, simulated-user messages). Agent calls not routed through attest\n * are not counted.\n */\nexport interface Usage {\n readonly model_calls: number;\n /** Judge-provider requests, the part of model_calls that max_judge_calls limits. */\n readonly judge_calls: number;\n readonly input_tokens: number;\n readonly output_tokens: number;\n readonly cached_tokens: number;\n readonly cost_usd: number;\n readonly retry_cost_usd: number;\n readonly unpriced_calls: number;\n readonly budget_refused: number;\n}\n\nexport interface EvaluateBatchResult {\n readonly results: readonly AssertionResult[];\n readonly total_cost?: number;\n readonly total_duration_ms?: number;\n /** Always sent by the engine; validated by AttestClient.evaluateBatch. */\n readonly usage: Usage;\n /**\n * True when this result was produced by an SDK simulation shim rather\n * than a real engine evaluation. Engine-produced results always carry\n * simulated=false. CI gates can fail builds that accidentally evaluate\n * against the simulator by checking this flag.\n */\n readonly simulated?: boolean;\n}\n\nexport interface ShutdownResult {\n readonly sessions_completed: number;\n readonly assertions_evaluated: number;\n}\n\nexport interface SubmitPluginResultParams {\n readonly trace_id: string;\n readonly plugin_name: string;\n readonly assertion_id: string;\n readonly result: AssertionResult;\n}\n\n// ---------------------------------------------------------------------------\n// Drift detection types\n// ---------------------------------------------------------------------------\n\nexport interface DriftReport {\n readonly assertion_id: string;\n readonly mean: number;\n readonly stddev: number;\n readonly count: number;\n readonly latest_score: number;\n readonly deviation: number;\n readonly status: string;\n}\n\nexport interface QueryDriftParams {\n readonly assertion_id: string;\n readonly window_size: number;\n}\n\nexport interface QueryDriftResult {\n readonly report: DriftReport;\n}\n\n// ---------------------------------------------------------------------------\n// Simulation types\n// ---------------------------------------------------------------------------\n\nexport interface SimulatePersona {\n readonly name: string;\n readonly system_prompt: string;\n readonly style: string;\n readonly temperature: number;\n readonly max_tokens?: number;\n}\n\nexport interface SimulateFaultConfig {\n readonly error_rate: number;\n readonly latency_jitter_ms: number;\n readonly content_corruption: boolean;\n readonly timeout_after_ms: number;\n}\n\nexport interface ConversationMessage {\n readonly role: string;\n readonly content: string;\n}\n\nexport interface GenerateUserMessageParams {\n readonly persona: SimulatePersona;\n readonly conversation_history: readonly ConversationMessage[];\n readonly fault_config?: SimulateFaultConfig;\n}\n\nexport interface GenerateUserMessageResult {\n readonly message: string;\n}\n","import type { ErrorData } from \"./types.js\";\nimport type { ProtocolDiagnostic } from \"./diagnostics.js\";\n\nexport class ProtocolError extends Error {\n readonly code: number;\n readonly errorMessage: string;\n readonly data: ErrorData | undefined;\n\n constructor(code: number, message: string, data?: ErrorData) {\n super(message);\n this.name = \"ProtocolError\";\n this.code = code;\n this.errorMessage = message;\n this.data = data;\n }\n}\n\nexport class EngineTimeoutError extends Error {\n readonly method: string;\n readonly timeoutMs: number;\n\n constructor(method: string, timeoutMs: number) {\n super(\n `Engine request '${method}' timed out after ${timeoutMs}ms. ` +\n \"Set ATTEST_ENGINE_TIMEOUT to increase the limit.\",\n );\n this.name = \"EngineTimeoutError\";\n this.method = method;\n this.timeoutMs = timeoutMs;\n }\n}\n\n/**\n * Raised when the reader loop detects that engine output has lost\n * framing or contains repeated unparseable lines. All in-flight requests\n * are rejected with this error so callers see a visible failure rather\n * than a hung promise.\n */\nexport class ProtocolDesyncError extends Error {\n readonly diagnostics: readonly ProtocolDiagnostic[];\n\n constructor(message: string, diagnostics: readonly ProtocolDiagnostic[]) {\n super(message);\n this.name = \"ProtocolDesyncError\";\n this.diagnostics = diagnostics;\n }\n}\n\n/**\n * The engine rejected an OTLP/JSON export (`attest-engine trace convert`). The message is\n * the engine's error, which names the offending span or field.\n */\nexport class OTLPConversionError extends Error {\n constructor(message: string) {\n super(message);\n this.name = \"OTLPConversionError\";\n }\n}\n","/**\n * Protocol diagnostic types for client-side observability.\n *\n * The reader loop emits diagnostics whenever it cannot route a response\n * line: malformed JSON, missing id, unknown id, or non-routable protocol\n * errors. Diagnostics are buffered in a bounded ring (32 entries) and\n * trigger a desync error when the rate exceeds the configured threshold.\n */\n\nexport type ProtocolDiagnosticKind =\n | \"malformed_json\"\n | \"non_object_response\"\n | \"invalid_jsonrpc_version\"\n | \"missing_id\"\n | \"unknown_id\"\n | \"non_routable_error\";\n\nexport interface ProtocolDiagnostic {\n readonly kind: ProtocolDiagnosticKind;\n readonly message: string;\n readonly rawLine: string;\n readonly timestampMs: number;\n}\n\nexport interface ProtocolLogger {\n warn(message: string, ...args: unknown[]): void;\n error(message: string, ...args: unknown[]): void;\n}\n\nconst DEFAULT_BUFFER_SIZE = 32;\nconst MAX_LINE_PREVIEW = 512;\n\nexport class ProtocolDiagnosticBuffer {\n private readonly capacity: number;\n private readonly entries: ProtocolDiagnostic[] = [];\n\n constructor(capacity: number = DEFAULT_BUFFER_SIZE) {\n if (capacity <= 0) {\n throw new Error(\n `ProtocolDiagnosticBuffer capacity must be positive, got ${capacity}`,\n );\n }\n this.capacity = capacity;\n }\n\n push(diagnostic: ProtocolDiagnostic): void {\n if (this.entries.length >= this.capacity) {\n this.entries.shift();\n }\n this.entries.push(diagnostic);\n }\n\n snapshot(): readonly ProtocolDiagnostic[] {\n return this.entries.slice();\n }\n\n countWithin(windowMs: number, nowMs: number): number {\n const cutoff = nowMs - windowMs;\n let count = 0;\n for (let i = this.entries.length - 1; i >= 0; i -= 1) {\n if (this.entries[i]!.timestampMs >= cutoff) {\n count += 1;\n } else {\n break;\n }\n }\n return count;\n }\n\n clear(): void {\n this.entries.length = 0;\n }\n}\n\nexport function previewLine(line: string): string {\n if (line.length <= MAX_LINE_PREVIEW) return line;\n return line.slice(0, MAX_LINE_PREVIEW) + \"…\";\n}\n","import type { ErrorData, RPCResponse } from \"./types.js\";\nimport { ProtocolError } from \"./errors.js\";\n\n/**\n * Encode a JSON-RPC 2.0 request as an NDJSON line.\n */\nexport function encodeRequest(\n id: number,\n method: string,\n params: Record<string, unknown>,\n): string {\n const msg = {\n jsonrpc: \"2.0\" as const,\n id,\n method,\n params,\n };\n return JSON.stringify(msg) + \"\\n\";\n}\n\n/**\n * Decode a JSON-RPC 2.0 response from an NDJSON line.\n * Throws ProtocolError if the response contains an error.\n * Throws Error for malformed JSON or invalid responses.\n */\nexport function decodeResponse(line: string): RPCResponse {\n const text = line.trim();\n if (!text) {\n throw new Error(\"empty response line\");\n }\n\n let data: unknown;\n try {\n data = JSON.parse(text);\n } catch (e) {\n throw new Error(`malformed JSON response: ${e instanceof Error ? e.message : String(e)}`);\n }\n\n if (typeof data !== \"object\" || data === null || Array.isArray(data)) {\n throw new Error(`expected JSON object, got ${Array.isArray(data) ? \"array\" : typeof data}`);\n }\n\n const obj = data as Record<string, unknown>;\n\n if (obj.jsonrpc !== \"2.0\") {\n throw new Error(`invalid jsonrpc version: ${String(obj.jsonrpc)}`);\n }\n\n if (obj.error != null) {\n const err = obj.error as Record<string, unknown>;\n let errorData: ErrorData | undefined;\n\n if (err.data != null) {\n const raw = err.data as Record<string, unknown>;\n errorData = {\n error_type: (raw.error_type as string) ?? \"\",\n retryable: (raw.retryable as boolean) ?? false,\n detail: (raw.detail as string) ?? \"\",\n };\n }\n\n throw new ProtocolError(\n (err.code as number) ?? -1,\n (err.message as string) ?? \"unknown error\",\n errorData,\n );\n }\n\n return obj as unknown as RPCResponse;\n}\n\n/**\n * Extract the result field from a decoded response.\n */\nexport function extractResult(response: RPCResponse): unknown {\n if (response.result === undefined) {\n throw new Error(\"response missing 'result' field\");\n }\n return response.result;\n}\n\n/**\n * Extract the request ID from a decoded response.\n */\nexport function extractId(response: RPCResponse): number {\n if (response.id === undefined) {\n throw new Error(\"response missing 'id' field\");\n }\n return Number(response.id);\n}\n"],"mappings":";;;;;;;;;;;;;;;;;;;;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;;;ACCO,IAAM,gBAAgB;AACtB,IAAM,iBAAiB;AACvB,IAAM,iBAAiB;AACvB,IAAM,kBAAkB;AAGxB,IAAM,cAAc;AACpB,IAAM,mBAAmB;AACzB,IAAM,mBAAmB;AACzB,IAAM,uBAAuB;AAG7B,IAAM,8BAA8B;AACpC,IAAM,iCAAiC;AACvC,IAAM,oBAAoB;AAC1B,IAAM,0BAA0B;AAChC,IAAM,yBAAyB;AAC/B,IAAM,0BAA0B;AAChC,IAAM,yBAAyB;AAG/B,IAAM,eAAe;AACrB,IAAM,eAAe;AACrB,IAAM,eAAe;AAIrB,IAAM,cAAc;AACpB,IAAM,kBAAkB;AACxB,IAAM,aAAa;AACnB,IAAM,eAAe;AACrB,IAAM,iBAAiB;AACvB,IAAM,iBAAiB;AACvB,IAAM,kBAAkB;AACxB,IAAM,cAAc;AAGpB,IAAM,oBAAoB;AAC1B,IAAM,sBAAsB;AAC5B,IAAM,qBAAqB;AAC3B,IAAM,mBAAmB;AACzB,IAAM,cAAc;AACpB,IAAM,oBAAoB;;;AChC1B,SAAS,QAAQ,IAAqB;AAC3C,SAAO;AACT;AAEO,SAAS,YAAY,IAAyB;AACnD,SAAO;AACT;AAEO,SAAS,QAAQ,IAAqB;AAC3C,SAAO;AACT;AAkPO,IAAM,eAAe;AAAA,EAC1B,YAAY;AAAA,EACZ,YAAY;AAAA,EACZ,WAAW;AAAA,EACX,kBAAkB;AAAA,EAClB,oBAAoB;AACtB;AAQO,SAAS,aAAa,eAA2C;AACtE,UAAQ,eAAe;AAAA,IACrB,KAAK;AACH,aAAO;AAAA,IACT,KAAK;AACH,aAAO;AAAA,IACT,KAAK;AAAA,IACL,KAAK;AACH,aAAO;AAAA,IACT,KAAK;AACH,aAAO;AAAA,IACT,KAAK;AACH,aAAO;AAAA,IACT,KAAK;AACH,aAAO;AAAA,IACT,KAAK;AACH,aAAO;AAAA,IACT;AACE,aAAO;AAAA,EACX;AACF;;;ACtSO,IAAM,gBAAN,cAA4B,MAAM;AAAA,EAC9B;AAAA,EACA;AAAA,EACA;AAAA,EAET,YAAY,MAAc,SAAiB,MAAkB;AAC3D,UAAM,OAAO;AACb,SAAK,OAAO;AACZ,SAAK,OAAO;AACZ,SAAK,eAAe;AACpB,SAAK,OAAO;AAAA,EACd;AACF;AAEO,IAAM,qBAAN,cAAiC,MAAM;AAAA,EACnC;AAAA,EACA;AAAA,EAET,YAAY,QAAgB,WAAmB;AAC7C;AAAA,MACE,mBAAmB,MAAM,qBAAqB,SAAS;AAAA,IAEzD;AACA,SAAK,OAAO;AACZ,SAAK,SAAS;AACd,SAAK,YAAY;AAAA,EACnB;AACF;AAQO,IAAM,sBAAN,cAAkC,MAAM;AAAA,EACpC;AAAA,EAET,YAAY,SAAiB,aAA4C;AACvE,UAAM,OAAO;AACb,SAAK,OAAO;AACZ,SAAK,cAAc;AAAA,EACrB;AACF;AAMO,IAAM,sBAAN,cAAkC,MAAM;AAAA,EAC7C,YAAY,SAAiB;AAC3B,UAAM,OAAO;AACb,SAAK,OAAO;AAAA,EACd;AACF;;;AC5BA,IAAM,sBAAsB;AAC5B,IAAM,mBAAmB;AAElB,IAAM,2BAAN,MAA+B;AAAA,EACnB;AAAA,EACA,UAAgC,CAAC;AAAA,EAElD,YAAY,WAAmB,qBAAqB;AAClD,QAAI,YAAY,GAAG;AACjB,YAAM,IAAI;AAAA,QACR,2DAA2D,QAAQ;AAAA,MACrE;AAAA,IACF;AACA,SAAK,WAAW;AAAA,EAClB;AAAA,EAEA,KAAK,YAAsC;AACzC,QAAI,KAAK,QAAQ,UAAU,KAAK,UAAU;AACxC,WAAK,QAAQ,MAAM;AAAA,IACrB;AACA,SAAK,QAAQ,KAAK,UAAU;AAAA,EAC9B;AAAA,EAEA,WAA0C;AACxC,WAAO,KAAK,QAAQ,MAAM;AAAA,EAC5B;AAAA,EAEA,YAAY,UAAkB,OAAuB;AACnD,UAAM,SAAS,QAAQ;AACvB,QAAI,QAAQ;AACZ,aAAS,IAAI,KAAK,QAAQ,SAAS,GAAG,KAAK,GAAG,KAAK,GAAG;AACpD,UAAI,KAAK,QAAQ,CAAC,EAAG,eAAe,QAAQ;AAC1C,iBAAS;AAAA,MACX,OAAO;AACL;AAAA,MACF;AAAA,IACF;AACA,WAAO;AAAA,EACT;AAAA,EAEA,QAAc;AACZ,SAAK,QAAQ,SAAS;AAAA,EACxB;AACF;AAEO,SAAS,YAAY,MAAsB;AAChD,MAAI,KAAK,UAAU,iBAAkB,QAAO;AAC5C,SAAO,KAAK,MAAM,GAAG,gBAAgB,IAAI;AAC3C;;;ACvEO,SAAS,cACd,IACA,QACA,QACQ;AACR,QAAM,MAAM;AAAA,IACV,SAAS;AAAA,IACT;AAAA,IACA;AAAA,IACA;AAAA,EACF;AACA,SAAO,KAAK,UAAU,GAAG,IAAI;AAC/B;AAOO,SAAS,eAAe,MAA2B;AACxD,QAAM,OAAO,KAAK,KAAK;AACvB,MAAI,CAAC,MAAM;AACT,UAAM,IAAI,MAAM,qBAAqB;AAAA,EACvC;AAEA,MAAI;AACJ,MAAI;AACF,WAAO,KAAK,MAAM,IAAI;AAAA,EACxB,SAAS,GAAG;AACV,UAAM,IAAI,MAAM,4BAA4B,aAAa,QAAQ,EAAE,UAAU,OAAO,CAAC,CAAC,EAAE;AAAA,EAC1F;AAEA,MAAI,OAAO,SAAS,YAAY,SAAS,QAAQ,MAAM,QAAQ,IAAI,GAAG;AACpE,UAAM,IAAI,MAAM,6BAA6B,MAAM,QAAQ,IAAI,IAAI,UAAU,OAAO,IAAI,EAAE;AAAA,EAC5F;AAEA,QAAM,MAAM;AAEZ,MAAI,IAAI,YAAY,OAAO;AACzB,UAAM,IAAI,MAAM,4BAA4B,OAAO,IAAI,OAAO,CAAC,EAAE;AAAA,EACnE;AAEA,MAAI,IAAI,SAAS,MAAM;AACrB,UAAM,MAAM,IAAI;AAChB,QAAI;AAEJ,QAAI,IAAI,QAAQ,MAAM;AACpB,YAAM,MAAM,IAAI;AAChB,kBAAY;AAAA,QACV,YAAa,IAAI,cAAyB;AAAA,QAC1C,WAAY,IAAI,aAAyB;AAAA,QACzC,QAAS,IAAI,UAAqB;AAAA,MACpC;AAAA,IACF;AAEA,UAAM,IAAI;AAAA,MACP,IAAI,QAAmB;AAAA,MACvB,IAAI,WAAsB;AAAA,MAC3B;AAAA,IACF;AAAA,EACF;AAEA,SAAO;AACT;AAKO,SAAS,cAAc,UAAgC;AAC5D,MAAI,SAAS,WAAW,QAAW;AACjC,UAAM,IAAI,MAAM,iCAAiC;AAAA,EACnD;AACA,SAAO,SAAS;AAClB;AAKO,SAAS,UAAU,UAA+B;AACvD,MAAI,SAAS,OAAO,QAAW;AAC7B,UAAM,IAAI,MAAM,6BAA6B;AAAA,EAC/C;AACA,SAAO,OAAO,SAAS,EAAE;AAC3B;","names":[]}
|
|
@@ -11,6 +11,8 @@ declare const REASON_PROVIDER_NOT_CONFIGURED: "provider_not_configured";
|
|
|
11
11
|
declare const REASON_SIMULATION: "simulation";
|
|
12
12
|
declare const REASON_MISSING_EVIDENCE: "missing_evidence";
|
|
13
13
|
declare const REASON_SHORT_CIRCUITED: "short_circuited";
|
|
14
|
+
declare const REASON_BUDGET_EXHAUSTED: "budget_exhausted";
|
|
15
|
+
declare const REASON_JUDGE_ABSTAINED: "judge_abstained";
|
|
14
16
|
declare const OUTCOME_PASS: "pass";
|
|
15
17
|
declare const OUTCOME_FAIL: "fail";
|
|
16
18
|
declare const OUTCOME_SKIP: "skip";
|
|
@@ -107,6 +109,30 @@ interface TraceMetadata {
|
|
|
107
109
|
readonly model?: string;
|
|
108
110
|
readonly timestamp?: string;
|
|
109
111
|
}
|
|
112
|
+
/** Trace v2: a step whose span ended in error. */
|
|
113
|
+
interface StepError {
|
|
114
|
+
readonly type: string;
|
|
115
|
+
readonly message?: string;
|
|
116
|
+
}
|
|
117
|
+
/** Trace v2: a recorded model request/response pair. */
|
|
118
|
+
interface Exchange {
|
|
119
|
+
readonly request: Record<string, unknown>;
|
|
120
|
+
readonly response: Record<string, unknown>;
|
|
121
|
+
}
|
|
122
|
+
/** Trace v2: one conversation turn and the indices of its steps. */
|
|
123
|
+
interface Turn {
|
|
124
|
+
readonly index: number;
|
|
125
|
+
readonly span_id?: string;
|
|
126
|
+
readonly input?: unknown;
|
|
127
|
+
readonly output?: unknown;
|
|
128
|
+
readonly step_indices: readonly number[];
|
|
129
|
+
}
|
|
130
|
+
/** Trace v2: control passing from one agent to another at an agent_call step. */
|
|
131
|
+
interface Handoff {
|
|
132
|
+
readonly from_agent: string;
|
|
133
|
+
readonly to_agent: string;
|
|
134
|
+
readonly step_index: number;
|
|
135
|
+
}
|
|
110
136
|
interface Step {
|
|
111
137
|
readonly type: string;
|
|
112
138
|
readonly name: string;
|
|
@@ -118,6 +144,12 @@ interface Step {
|
|
|
118
144
|
readonly ended_at_ms?: number;
|
|
119
145
|
readonly agent_id?: string;
|
|
120
146
|
readonly agent_role?: string;
|
|
147
|
+
readonly span_id?: string;
|
|
148
|
+
readonly parent_span_id?: string;
|
|
149
|
+
readonly tool_call_id?: string;
|
|
150
|
+
readonly retry_of?: number;
|
|
151
|
+
readonly error?: StepError;
|
|
152
|
+
readonly exchange?: Exchange;
|
|
121
153
|
}
|
|
122
154
|
interface Trace {
|
|
123
155
|
readonly trace_id: string;
|
|
@@ -128,6 +160,9 @@ interface Trace {
|
|
|
128
160
|
readonly steps: readonly Step[];
|
|
129
161
|
readonly metadata?: TraceMetadata;
|
|
130
162
|
readonly parent_trace_id?: string;
|
|
163
|
+
readonly conversation_id?: string;
|
|
164
|
+
readonly turns?: readonly Turn[];
|
|
165
|
+
readonly handoffs?: readonly Handoff[];
|
|
131
166
|
}
|
|
132
167
|
interface Assertion {
|
|
133
168
|
readonly assertion_id: string;
|
|
@@ -142,6 +177,7 @@ interface Assertion {
|
|
|
142
177
|
* reviewers need to triage a judge failure.
|
|
143
178
|
*/
|
|
144
179
|
interface JudgeMetadata {
|
|
180
|
+
readonly provider?: string;
|
|
145
181
|
readonly model?: string;
|
|
146
182
|
readonly rubric_name?: string;
|
|
147
183
|
readonly rubric_version?: string;
|
|
@@ -261,14 +297,44 @@ interface InitializeResult {
|
|
|
261
297
|
readonly max_trace_size_bytes?: number;
|
|
262
298
|
readonly max_steps_per_trace?: number;
|
|
263
299
|
}
|
|
300
|
+
/**
|
|
301
|
+
* Limits for one evaluate_batch call on top of the session budget. A judge or
|
|
302
|
+
* embedding call that cannot reserve against them is not sent and its
|
|
303
|
+
* assertion is not_evaluated with reason "budget_exhausted".
|
|
304
|
+
*/
|
|
305
|
+
interface BatchBudget {
|
|
306
|
+
readonly max_judge_calls?: number;
|
|
307
|
+
readonly max_cost_usd?: number;
|
|
308
|
+
}
|
|
264
309
|
interface EvaluateBatchParams {
|
|
265
310
|
readonly trace: Trace;
|
|
266
311
|
readonly assertions: readonly Assertion[];
|
|
312
|
+
/** Omitted when no per-batch budget applies. */
|
|
313
|
+
readonly budget?: BatchBudget;
|
|
314
|
+
}
|
|
315
|
+
/**
|
|
316
|
+
* Attest-initiated model usage (embeddings, judges, judge repeats, bias
|
|
317
|
+
* probes, simulated-user messages). Agent calls not routed through attest
|
|
318
|
+
* are not counted.
|
|
319
|
+
*/
|
|
320
|
+
interface Usage {
|
|
321
|
+
readonly model_calls: number;
|
|
322
|
+
/** Judge-provider requests, the part of model_calls that max_judge_calls limits. */
|
|
323
|
+
readonly judge_calls: number;
|
|
324
|
+
readonly input_tokens: number;
|
|
325
|
+
readonly output_tokens: number;
|
|
326
|
+
readonly cached_tokens: number;
|
|
327
|
+
readonly cost_usd: number;
|
|
328
|
+
readonly retry_cost_usd: number;
|
|
329
|
+
readonly unpriced_calls: number;
|
|
330
|
+
readonly budget_refused: number;
|
|
267
331
|
}
|
|
268
332
|
interface EvaluateBatchResult {
|
|
269
333
|
readonly results: readonly AssertionResult[];
|
|
270
334
|
readonly total_cost?: number;
|
|
271
335
|
readonly total_duration_ms?: number;
|
|
336
|
+
/** Always sent by the engine; validated by AttestClient.evaluateBatch. */
|
|
337
|
+
readonly usage: Usage;
|
|
272
338
|
/**
|
|
273
339
|
* True when this result was produced by an SDK simulation shim rather
|
|
274
340
|
* than a real engine evaluation. Engine-produced results always carry
|
|
@@ -380,6 +446,13 @@ declare class ProtocolDesyncError extends Error {
|
|
|
380
446
|
readonly diagnostics: readonly ProtocolDiagnostic[];
|
|
381
447
|
constructor(message: string, diagnostics: readonly ProtocolDiagnostic[]);
|
|
382
448
|
}
|
|
449
|
+
/**
|
|
450
|
+
* The engine rejected an OTLP/JSON export (`attest-engine trace convert`). The message is
|
|
451
|
+
* the engine's error, which names the offending span or field.
|
|
452
|
+
*/
|
|
453
|
+
declare class OTLPConversionError extends Error {
|
|
454
|
+
constructor(message: string);
|
|
455
|
+
}
|
|
383
456
|
|
|
384
457
|
/**
|
|
385
458
|
* Encode a JSON-RPC 2.0 request as an NDJSON line.
|
|
@@ -400,4 +473,4 @@ declare function extractResult(response: RPCResponse): unknown;
|
|
|
400
473
|
*/
|
|
401
474
|
declare function extractId(response: RPCResponse): number;
|
|
402
475
|
|
|
403
|
-
export { type AgentId, type Assertion, type AssertionId, type AssertionResult, type AssertionSpec, type BiasProbe, type ConstraintSpec, type ContentSpec, type ConversationMessage, type DriftReport, ERR_ASSERTION_ERROR, ERR_ENGINE_ERROR, ERR_INVALID_TRACE, ERR_PROVIDER_ERROR, ERR_SESSION_ERROR, ERR_TIMEOUT, type EmbeddingSpec, EngineTimeoutError, type ErrorData, type EvaluateBatchParams, type EvaluateBatchResult, FailureClass, type GenerateUserMessageParams, type GenerateUserMessageResult, type InitializeParams, type InitializeResult, type JudgeAgreement, type JudgeMetadata, type LlmJudgeSpec, OUTCOME_FAIL, OUTCOME_PASS, OUTCOME_SKIP, type Outcome, type PluginSpec, ProtocolDesyncError, type ProtocolDiagnostic, ProtocolDiagnosticBuffer, type ProtocolDiagnosticKind, ProtocolError, type ProtocolLogger, type QueryDriftParams, type QueryDriftResult, REASON_MISSING_EVIDENCE, REASON_PROVIDER_NOT_CONFIGURED, REASON_PROVIDER_UNAVAILABLE, REASON_SHORT_CIRCUITED, REASON_SIMULATION, type RPCError, type RPCRequest, type RPCResponse, STATUS_HARD_FAIL, STATUS_NOT_EVALUATED, STATUS_PASS, STATUS_SOFT_FAIL, STEP_AGENT_CALL, STEP_LLM_CALL, STEP_RETRIEVAL, STEP_TOOL_CALL, type SchemaSpec, type ShutdownResult, type SimulateFaultConfig, type SimulatePersona, type Step, type SubmitPluginResultParams, TYPE_CONSTRAINT, TYPE_CONTENT, TYPE_EMBEDDING, TYPE_LLM_JUDGE, TYPE_PLUGIN, TYPE_SCHEMA, TYPE_TRACE, TYPE_TRACE_TREE, type Trace, type TraceId, type TraceMetadata, type TraceSpec, type TraceTreeSpec, agentId, assertionId, decodeResponse, encodeRequest, extractId, extractResult, layerForType, previewLine, traceId };
|
|
476
|
+
export { type AgentId, type Assertion, type AssertionId, type AssertionResult, type AssertionSpec, type BatchBudget, type BiasProbe, type ConstraintSpec, type ContentSpec, type ConversationMessage, type DriftReport, ERR_ASSERTION_ERROR, ERR_ENGINE_ERROR, ERR_INVALID_TRACE, ERR_PROVIDER_ERROR, ERR_SESSION_ERROR, ERR_TIMEOUT, type EmbeddingSpec, EngineTimeoutError, type ErrorData, type EvaluateBatchParams, type EvaluateBatchResult, type Exchange, FailureClass, type GenerateUserMessageParams, type GenerateUserMessageResult, type Handoff, type InitializeParams, type InitializeResult, type JudgeAgreement, type JudgeMetadata, type LlmJudgeSpec, OTLPConversionError, OUTCOME_FAIL, OUTCOME_PASS, OUTCOME_SKIP, type Outcome, type PluginSpec, ProtocolDesyncError, type ProtocolDiagnostic, ProtocolDiagnosticBuffer, type ProtocolDiagnosticKind, ProtocolError, type ProtocolLogger, type QueryDriftParams, type QueryDriftResult, REASON_BUDGET_EXHAUSTED, REASON_JUDGE_ABSTAINED, REASON_MISSING_EVIDENCE, REASON_PROVIDER_NOT_CONFIGURED, REASON_PROVIDER_UNAVAILABLE, REASON_SHORT_CIRCUITED, REASON_SIMULATION, type RPCError, type RPCRequest, type RPCResponse, STATUS_HARD_FAIL, STATUS_NOT_EVALUATED, STATUS_PASS, STATUS_SOFT_FAIL, STEP_AGENT_CALL, STEP_LLM_CALL, STEP_RETRIEVAL, STEP_TOOL_CALL, type SchemaSpec, type ShutdownResult, type SimulateFaultConfig, type SimulatePersona, type Step, type StepError, type SubmitPluginResultParams, TYPE_CONSTRAINT, TYPE_CONTENT, TYPE_EMBEDDING, TYPE_LLM_JUDGE, TYPE_PLUGIN, TYPE_SCHEMA, TYPE_TRACE, TYPE_TRACE_TREE, type Trace, type TraceId, type TraceMetadata, type TraceSpec, type TraceTreeSpec, type Turn, type Usage, agentId, assertionId, decodeResponse, encodeRequest, extractId, extractResult, layerForType, previewLine, traceId };
|
package/dist/esm/cli.js
CHANGED
|
@@ -7,8 +7,8 @@ import * as os2 from "os";
|
|
|
7
7
|
import { fileURLToPath } from "url";
|
|
8
8
|
|
|
9
9
|
// src/version.ts
|
|
10
|
-
var VERSION = "1.0.0-rc.
|
|
11
|
-
var ENGINE_VERSION = "1.0.0-rc.
|
|
10
|
+
var VERSION = "1.0.0-rc.2";
|
|
11
|
+
var ENGINE_VERSION = "1.0.0-rc.2";
|
|
12
12
|
|
|
13
13
|
// src/engine-manager.ts
|
|
14
14
|
import { spawn, spawnSync } from "child_process";
|