@attest-ai/core 0.6.0 → 1.0.0-rc.1

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.
@@ -1,12 +1,12 @@
1
- import { InitializeResult, Trace, Assertion, EvaluateBatchResult, DriftReport, SimulatePersona, ConversationMessage, SimulateFaultConfig, Step, AssertionResult } from './proto/index.cjs';
2
- export { AgentId, AssertionId, AssertionSpec, ConstraintSpec, ContentSpec, ERR_ASSERTION_ERROR, ERR_ENGINE_ERROR, ERR_INVALID_TRACE, ERR_PROVIDER_ERROR, ERR_SESSION_ERROR, ERR_TIMEOUT, EmbeddingSpec, EngineTimeoutError, ErrorData, EvaluateBatchParams, GenerateUserMessageParams, GenerateUserMessageResult, InitializeParams, LlmJudgeSpec, PluginSpec, ProtocolError, QueryDriftParams, QueryDriftResult, RPCError, RPCRequest, RPCResponse, STATUS_HARD_FAIL, 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, traceId } from './proto/index.cjs';
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';
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 = "0.6.0";
9
- declare const ENGINE_VERSION = "0.6.0";
8
+ declare const VERSION = "1.0.0-rc.1";
9
+ declare const ENGINE_VERSION = "1.0.0-rc.1";
10
10
 
11
11
  declare class EngineManager {
12
12
  private enginePath;
@@ -21,6 +21,8 @@ declare class EngineManager {
21
21
  logLevel?: string;
22
22
  });
23
23
  start(): Promise<InitializeResult>;
24
+ /** Send initialize and check the engine's version and capabilities. */
25
+ private initialize;
24
26
  stop(): Promise<void>;
25
27
  sendRequest(method: string, params: Record<string, unknown>): Promise<unknown>;
26
28
  private sendRequestInternal;
@@ -31,6 +33,28 @@ declare class EngineManager {
31
33
  get readlineInterface(): Interface | null;
32
34
  }
33
35
 
36
+ interface AttestClientOptions {
37
+ /** Logger override. When omitted, warnings only emit if ATTEST_DEBUG_PROTOCOL=1. */
38
+ logger?: ProtocolLogger;
39
+ /** Listener invoked for every protocol diagnostic. */
40
+ onDiagnostic?: (diagnostic: ProtocolDiagnostic) => void;
41
+ /** Diagnostic ring-buffer capacity. Default 32. */
42
+ diagnosticBufferSize?: number;
43
+ /** Number of diagnostics in `desyncWindowMs` that triggers desync rejection. Default 3. */
44
+ desyncThreshold?: number;
45
+ /** Window length for desync detection in milliseconds. Default 1000. */
46
+ desyncWindowMs?: number;
47
+ }
48
+ /**
49
+ * AttestClient owns request/response correlation over the engine's
50
+ * NDJSON protocol.
51
+ *
52
+ * Desync detection is a one-way latch: once the diagnostic rate breaches
53
+ * `desyncThreshold` within `desyncWindowMs` the client refuses further
54
+ * sends for the lifetime of the instance. There is no reset hook by
55
+ * design — recovery requires constructing a new client over a fresh
56
+ * engine subprocess.
57
+ */
34
58
  declare class AttestClient {
35
59
  private readonly engine;
36
60
  private requestId;
@@ -38,10 +62,23 @@ declare class AttestClient {
38
62
  private writeChain;
39
63
  private readerActive;
40
64
  private lineHandler;
41
- constructor(engine: EngineManager);
65
+ private readonly logger;
66
+ private readonly diagnosticBuffer;
67
+ private readonly diagnosticListeners;
68
+ private readonly desyncThreshold;
69
+ private readonly desyncWindowMs;
70
+ private desynced;
71
+ constructor(engine: EngineManager, options?: AttestClientOptions);
72
+ /** Subscribe to protocol diagnostics. Returns an unsubscribe function. */
73
+ onProtocolDiagnostic(listener: (diagnostic: ProtocolDiagnostic) => void): () => void;
74
+ /** Snapshot of the diagnostic ring buffer (most recent last). */
75
+ protocolDiagnostics(): readonly ProtocolDiagnostic[];
42
76
  startReader(): void;
43
77
  stopReader(): void;
78
+ private recordDiagnostic;
79
+ private maybeTriggerDesync;
44
80
  private handleLine;
81
+ private handleDecodeFailure;
45
82
  private failAll;
46
83
  sendRequest(method: string, params: Record<string, unknown>, timeoutMs?: number): Promise<unknown>;
47
84
  evaluateBatch(trace: Trace, assertions: readonly Assertion[], options?: {
@@ -57,6 +94,12 @@ declare class AttestClient {
57
94
  timeout?: number;
58
95
  }): Promise<string>;
59
96
  }
97
+ /**
98
+ * Throw unless the engine advertised `simulation_mode` at initialize. An
99
+ * engine started without ATTEST_SIMULATION (or older than 1.0) would call
100
+ * real embedding and judge providers in simulation mode.
101
+ */
102
+ declare function requireSimulationCapability(init: InitializeResult | null | undefined): void;
60
103
 
61
104
  declare class TraceBuilder {
62
105
  private traceId;
@@ -104,14 +147,60 @@ declare class AgentResult {
104
147
  readonly totalCost: number;
105
148
  readonly totalDurationMs: number;
106
149
  constructor(trace: Trace, assertionResults?: readonly AssertionResult[], totalCost?: number, totalDurationMs?: number);
150
+ /** True if every assertion has status "pass". */
107
151
  get passed(): boolean;
152
+ /**
153
+ * Test outcome after the whole batch. Any hard_fail fails; otherwise any
154
+ * not_evaluated skips (fails when `strict`); otherwise pass, including
155
+ * when only soft_fail rows are present (see `softFailures`).
156
+ */
157
+ outcome(options?: {
158
+ strict?: boolean;
159
+ }): Outcome;
160
+ /** Failed assertions: hard_fail and soft_fail only. */
108
161
  get failedAssertions(): readonly AssertionResult[];
109
162
  get hardFailures(): readonly AssertionResult[];
110
163
  get softFailures(): readonly AssertionResult[];
164
+ /** Assertions the engine did not evaluate; each carries a `reason`. */
165
+ get notEvaluated(): readonly AssertionResult[];
166
+ /** Count of not_evaluated assertions per reason, in first-seen order. */
167
+ get notEvaluatedReasons(): Readonly<Record<string, number>>;
168
+ /**
169
+ * Passed / evaluated assertions; null when nothing was evaluated.
170
+ * not_evaluated rows are excluded from the denominator.
171
+ */
172
+ get passRate(): number | null;
111
173
  get passCount(): number;
174
+ /** Number of failed assertions (hard_fail + soft_fail). */
112
175
  get failCount(): number;
176
+ /** "not evaluated: <reason>[, <reason> ×n ...]" for skipped tests. */
177
+ skipReason(): string;
113
178
  }
114
179
 
180
+ /**
181
+ * Terminal-grade rendering of Attest assertion failures.
182
+ *
183
+ * Mirrors the Python attest.cli.diagnostics module so PR comments,
184
+ * pytest output, and vitest output show the same fields. Use as the
185
+ * message argument when asserting the verdict, e.g.
186
+ * `expect(result.outcome(), renderDiagnostics(result)).not.toBe("fail")`, so
187
+ * reviewers see trace path, expected vs actual, and judge metadata.
188
+ */
189
+
190
+ interface RenderOptions {
191
+ /** Emit ANSI colour escapes. Default false (plain text). */
192
+ color?: boolean;
193
+ /** Source location annotation appended after each block. */
194
+ testFile?: string;
195
+ /** Render passing assertions in addition to failures. Default false. */
196
+ includePassing?: boolean;
197
+ }
198
+ declare function renderAssertionFailure(result: AssertionResult, options?: RenderOptions): string;
199
+ declare function renderSummary(result: AgentResult, options?: {
200
+ color?: boolean;
201
+ }): string;
202
+ declare function renderDiagnostics(result: AgentResult, options?: RenderOptions): string;
203
+
115
204
  declare class TraceTree {
116
205
  readonly root: Trace;
117
206
  constructor(root: Trace);
@@ -808,9 +897,11 @@ declare function tier(level: number): <T extends (...args: never[]) => unknown>(
808
897
  /**
809
898
  * Configure global SDK options.
810
899
  *
811
- * @param options.simulation - Enable simulation mode. All `evaluateBatch` calls
812
- * return deterministic pass results without spawning the engine or calling any
813
- * LLM. Equivalent to setting `ATTEST_SIMULATION=1` in the environment.
900
+ * @param options.simulation - Enable simulation mode: the engine evaluates
901
+ * layers 1-4 and reports layers 5-6 as not evaluated (reason "simulation")
902
+ * without calling any provider. The engine reads `ATTEST_SIMULATION` when it
903
+ * starts; the Vitest fixtures set it from this flag. Equivalent to setting
904
+ * `ATTEST_SIMULATION=1` in the environment.
814
905
  * @param options.sampleRate - Sampling rate for continuous evaluation (0.0–1.0).
815
906
  * Equivalent to setting `ATTEST_SAMPLE_RATE` in the environment.
816
907
  */
@@ -823,6 +914,11 @@ declare function config(options: {
823
914
  * or the `ATTEST_SIMULATION` environment variable.
824
915
  */
825
916
  declare function isSimulationMode(): boolean;
917
+ /**
918
+ * True when `ATTEST_SIMULATION` is set for this process (and so for an
919
+ * engine it spawns): "1", "true" or "yes", trimmed and case-insensitive.
920
+ */
921
+ declare function simulationEnvEnabled(): boolean;
826
922
  /**
827
923
  * Returns the sampling rate for continuous evaluation. Priority:
828
924
  * programmatic `config({ sampleRate })` > `ATTEST_SAMPLE_RATE` env var > 0.
@@ -834,13 +930,6 @@ declare function getSampleRate(): number;
834
930
  */
835
931
  declare function resetConfig(): void;
836
932
 
837
- /**
838
- * Return a deterministic all-pass result for every assertion without
839
- * contacting the engine or any LLM provider. Used when simulation mode
840
- * is active (`ATTEST_SIMULATION=1` or `config({ simulation: true })`).
841
- */
842
- declare function simulationEvaluateBatch(assertions: readonly Assertion[]): EvaluateBatchResult;
843
-
844
933
  interface PluginResult {
845
934
  readonly status: string;
846
935
  readonly score: number;
@@ -903,6 +992,59 @@ declare class ContinuousEvalRunner {
903
992
  private scheduleNext;
904
993
  }
905
994
 
995
+ /**
996
+ * Calibration helpers — TS parity with engine/internal/assertion/judge/calibration.go
997
+ * and sdks/python/src/attest/calibration.py.
998
+ *
999
+ * Computes Cohen's κ, agreement %, and ROC-AUC over (human, judge) score
1000
+ * pairs and parses CSV/JSONL labels files. Metrics match the engine
1001
+ * byte-for-byte on the same input.
1002
+ */
1003
+ interface LabelPair {
1004
+ readonly human: number;
1005
+ readonly judge: number;
1006
+ }
1007
+ interface AgreementResult {
1008
+ readonly threshold: number;
1009
+ readonly n: number;
1010
+ readonly agreement: number;
1011
+ readonly cohenKappa: number;
1012
+ readonly rocAuc: number;
1013
+ }
1014
+ interface LabeledRecord {
1015
+ input: string;
1016
+ humanLabel: number;
1017
+ judgeScore: number;
1018
+ judgeKnown: boolean;
1019
+ }
1020
+ /**
1021
+ * Compute Cohen's κ, agreement %, and ROC-AUC over labeled pairs.
1022
+ *
1023
+ * Throws when pairs is empty or threshold is outside (0, 1). Cohen's κ
1024
+ * degenerates to 0 when all labels land in a single class for both
1025
+ * raters; ROC-AUC is reported as 0 when only one human-class is
1026
+ * present.
1027
+ */
1028
+ declare function computeAgreement(pairs: readonly LabelPair[], threshold: number): AgreementResult;
1029
+ /**
1030
+ * Parse a 2- or 3-column CSV (input, human_label[, judge_score]).
1031
+ * Header detected when row[0][1] does not parse as a float. Lines
1032
+ * beginning with `#` are skipped.
1033
+ */
1034
+ declare function loadLabelsCSV(text: string): LabeledRecord[];
1035
+ /**
1036
+ * Parse a newline-delimited JSON file. Each line is an object with the
1037
+ * shape {"input": "...", "human_label": 0.9, "judge_score": 0.8}.
1038
+ * `judge_score` is optional.
1039
+ */
1040
+ declare function loadLabelsJSONL(text: string): LabeledRecord[];
1041
+ /**
1042
+ * Return the 16-character SHA-256 prefix used by the engine for prompt
1043
+ * hashes. Mirrors promptHash in engine/internal/assertion/judge_eval.go
1044
+ * so SDK-recorded calibration rows align with JudgeMetadata.prompt_hash.
1045
+ */
1046
+ declare function promptHash(text: string): string;
1047
+
906
1048
  declare const FRIENDLY_USER: SimulatePersona;
907
1049
  declare const ADVERSARIAL_USER: SimulatePersona;
908
1050
  declare const CONFUSED_USER: SimulatePersona;
@@ -1108,4 +1250,4 @@ interface TraceAdapter {
1108
1250
  traceFromResponse(response: unknown, options?: Record<string, unknown>): Trace;
1109
1251
  }
1110
1252
 
1111
- export { ADVERSARIAL_USER, Agent, AgentResult, AlertDispatcher, type AlertPayload, AnthropicAdapter, Assertion, AssertionResult, AttestClient, type AttestPlugin, CONFUSED_USER, COOPERATIVE_USER, ContinuousEvalRunner, ConversationMessage, DriftReport, ENGINE_VERSION, EngineManager, EvaluateBatchResult, ExpectChain, FRIENDLY_USER, GeminiAdapter, InitializeResult, LangChainAdapter, ManualAdapter, OTelAdapter, OllamaAdapter, OpenAIAdapter, PluginRegistry, type PluginResult, Sampler, SimulateFaultConfig, SimulatePersona, Step, TIER_1, TIER_2, TIER_3, Trace, type TraceAdapter, TraceBuilder, TraceTree, VERSION, activeBuilder, agent, attestExpect, config, delegate, executePluginAssertion, getSampleRate, isSimulationMode, resetConfig, simulationEvaluateBatch, tier };
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 };
@@ -27,8 +27,20 @@ __export(proto_exports, {
27
27
  ERR_SESSION_ERROR: () => ERR_SESSION_ERROR,
28
28
  ERR_TIMEOUT: () => ERR_TIMEOUT,
29
29
  EngineTimeoutError: () => EngineTimeoutError,
30
+ FailureClass: () => FailureClass,
31
+ OUTCOME_FAIL: () => OUTCOME_FAIL,
32
+ OUTCOME_PASS: () => OUTCOME_PASS,
33
+ OUTCOME_SKIP: () => OUTCOME_SKIP,
34
+ ProtocolDesyncError: () => ProtocolDesyncError,
35
+ ProtocolDiagnosticBuffer: () => ProtocolDiagnosticBuffer,
30
36
  ProtocolError: () => ProtocolError,
37
+ REASON_MISSING_EVIDENCE: () => REASON_MISSING_EVIDENCE,
38
+ REASON_PROVIDER_NOT_CONFIGURED: () => REASON_PROVIDER_NOT_CONFIGURED,
39
+ REASON_PROVIDER_UNAVAILABLE: () => REASON_PROVIDER_UNAVAILABLE,
40
+ REASON_SHORT_CIRCUITED: () => REASON_SHORT_CIRCUITED,
41
+ REASON_SIMULATION: () => REASON_SIMULATION,
31
42
  STATUS_HARD_FAIL: () => STATUS_HARD_FAIL,
43
+ STATUS_NOT_EVALUATED: () => STATUS_NOT_EVALUATED,
32
44
  STATUS_PASS: () => STATUS_PASS,
33
45
  STATUS_SOFT_FAIL: () => STATUS_SOFT_FAIL,
34
46
  STEP_AGENT_CALL: () => STEP_AGENT_CALL,
@@ -49,6 +61,8 @@ __export(proto_exports, {
49
61
  encodeRequest: () => encodeRequest,
50
62
  extractId: () => extractId,
51
63
  extractResult: () => extractResult,
64
+ layerForType: () => layerForType,
65
+ previewLine: () => previewLine,
52
66
  traceId: () => traceId
53
67
  });
54
68
  module.exports = __toCommonJS(proto_exports);
@@ -61,6 +75,15 @@ var STEP_AGENT_CALL = "agent_call";
61
75
  var STATUS_PASS = "pass";
62
76
  var STATUS_SOFT_FAIL = "soft_fail";
63
77
  var STATUS_HARD_FAIL = "hard_fail";
78
+ var STATUS_NOT_EVALUATED = "not_evaluated";
79
+ var REASON_PROVIDER_UNAVAILABLE = "provider_unavailable";
80
+ var REASON_PROVIDER_NOT_CONFIGURED = "provider_not_configured";
81
+ var REASON_SIMULATION = "simulation";
82
+ var REASON_MISSING_EVIDENCE = "missing_evidence";
83
+ var REASON_SHORT_CIRCUITED = "short_circuited";
84
+ var OUTCOME_PASS = "pass";
85
+ var OUTCOME_FAIL = "fail";
86
+ var OUTCOME_SKIP = "skip";
64
87
  var TYPE_SCHEMA = "schema";
65
88
  var TYPE_CONSTRAINT = "constraint";
66
89
  var TYPE_TRACE = "trace";
@@ -86,6 +109,34 @@ function assertionId(id) {
86
109
  function agentId(id) {
87
110
  return id;
88
111
  }
112
+ var FailureClass = {
113
+ BrokenCode: "broken_code",
114
+ FlakyJudge: "flaky_judge",
115
+ BadRubric: "bad_rubric",
116
+ MissingTraceData: "missing_trace_data",
117
+ StochasticVariance: "stochastic_variance"
118
+ };
119
+ function layerForType(assertionType) {
120
+ switch (assertionType) {
121
+ case "schema":
122
+ return 1;
123
+ case "constraint":
124
+ return 2;
125
+ case "trace":
126
+ case "trace_tree":
127
+ return 3;
128
+ case "content":
129
+ return 4;
130
+ case "embedding":
131
+ return 5;
132
+ case "llm_judge":
133
+ return 6;
134
+ case "plugin":
135
+ return 8;
136
+ default:
137
+ return 0;
138
+ }
139
+ }
89
140
 
90
141
  // src/proto/errors.ts
91
142
  var ProtocolError = class extends Error {
@@ -112,6 +163,58 @@ var EngineTimeoutError = class extends Error {
112
163
  this.timeoutMs = timeoutMs;
113
164
  }
114
165
  };
166
+ var ProtocolDesyncError = class extends Error {
167
+ diagnostics;
168
+ constructor(message, diagnostics) {
169
+ super(message);
170
+ this.name = "ProtocolDesyncError";
171
+ this.diagnostics = diagnostics;
172
+ }
173
+ };
174
+
175
+ // src/proto/diagnostics.ts
176
+ var DEFAULT_BUFFER_SIZE = 32;
177
+ var MAX_LINE_PREVIEW = 512;
178
+ var ProtocolDiagnosticBuffer = class {
179
+ capacity;
180
+ entries = [];
181
+ constructor(capacity = DEFAULT_BUFFER_SIZE) {
182
+ if (capacity <= 0) {
183
+ throw new Error(
184
+ `ProtocolDiagnosticBuffer capacity must be positive, got ${capacity}`
185
+ );
186
+ }
187
+ this.capacity = capacity;
188
+ }
189
+ push(diagnostic) {
190
+ if (this.entries.length >= this.capacity) {
191
+ this.entries.shift();
192
+ }
193
+ this.entries.push(diagnostic);
194
+ }
195
+ snapshot() {
196
+ return this.entries.slice();
197
+ }
198
+ countWithin(windowMs, nowMs) {
199
+ const cutoff = nowMs - windowMs;
200
+ let count = 0;
201
+ for (let i = this.entries.length - 1; i >= 0; i -= 1) {
202
+ if (this.entries[i].timestampMs >= cutoff) {
203
+ count += 1;
204
+ } else {
205
+ break;
206
+ }
207
+ }
208
+ return count;
209
+ }
210
+ clear() {
211
+ this.entries.length = 0;
212
+ }
213
+ };
214
+ function previewLine(line) {
215
+ if (line.length <= MAX_LINE_PREVIEW) return line;
216
+ return line.slice(0, MAX_LINE_PREVIEW) + "\u2026";
217
+ }
115
218
 
116
219
  // src/proto/codec.ts
117
220
  function encodeRequest(id, method, params) {
@@ -181,8 +284,20 @@ function extractId(response) {
181
284
  ERR_SESSION_ERROR,
182
285
  ERR_TIMEOUT,
183
286
  EngineTimeoutError,
287
+ FailureClass,
288
+ OUTCOME_FAIL,
289
+ OUTCOME_PASS,
290
+ OUTCOME_SKIP,
291
+ ProtocolDesyncError,
292
+ ProtocolDiagnosticBuffer,
184
293
  ProtocolError,
294
+ REASON_MISSING_EVIDENCE,
295
+ REASON_PROVIDER_NOT_CONFIGURED,
296
+ REASON_PROVIDER_UNAVAILABLE,
297
+ REASON_SHORT_CIRCUITED,
298
+ REASON_SIMULATION,
185
299
  STATUS_HARD_FAIL,
300
+ STATUS_NOT_EVALUATED,
186
301
  STATUS_PASS,
187
302
  STATUS_SOFT_FAIL,
188
303
  STEP_AGENT_CALL,
@@ -203,6 +318,8 @@ function extractId(response) {
203
318
  encodeRequest,
204
319
  extractId,
205
320
  extractResult,
321
+ layerForType,
322
+ previewLine,
206
323
  traceId
207
324
  });
208
325
  //# sourceMappingURL=index.cjs.map
@@ -1 +1 @@
1
- {"version":3,"sources":["../../../src/proto/index.ts","../../../src/proto/constants.ts","../../../src/proto/types.ts","../../../src/proto/errors.ts","../../../src/proto/codec.ts"],"sourcesContent":["export * from \"./constants.js\";\nexport * from \"./types.js\";\nexport * from \"./errors.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;\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\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}\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\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\";\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","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;;;ACCO,IAAM,gBAAgB;AACtB,IAAM,iBAAiB;AACvB,IAAM,iBAAiB;AACvB,IAAM,kBAAkB;AAGxB,IAAM,cAAc;AACpB,IAAM,mBAAmB;AACzB,IAAM,mBAAmB;AAGzB,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;;;AChB1B,SAAS,QAAQ,IAAqB;AAC3C,SAAO;AACT;AAEO,SAAS,YAAY,IAAyB;AACnD,SAAO;AACT;AAEO,SAAS,QAAQ,IAAqB;AAC3C,SAAO;AACT;;;ACnBO,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;;;ACvBO,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;\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":[]}