@attest-ai/core 0.5.0 → 0.9.0

package/dist/cjs/index.d.cts

@@ -1,12 +1,12 @@
-import { InitializeResult, Trace, Assertion, EvaluateBatchResult, Step, AssertionResult } from './proto/index.cjs';
-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, InitializeParams, LlmJudgeSpec, PluginSpec, ProtocolError, 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';
+import { InitializeResult, ProtocolLogger, ProtocolDiagnostic, Trace, Assertion, EvaluateBatchResult, DriftReport, SimulatePersona, ConversationMessage, SimulateFaultConfig, Step, AssertionResult } from './proto/index.cjs';
+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, GenerateUserMessageParams, GenerateUserMessageResult, InitializeParams, JudgeAgreement, JudgeMetadata, LlmJudgeSpec, PluginSpec, ProtocolDesyncError, ProtocolDiagnosticBuffer, ProtocolDiagnosticKind, 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, layerForType, previewLine, traceId } from './proto/index.cjs';
 import { ChildProcess } from 'node:child_process';
 import { Interface } from 'node:readline';
 import { AsyncLocalStorage } from 'node:async_hooks';
 import { EventEmitter } from 'node:events';
 
-declare const VERSION = "0.5.0";
-declare const ENGINE_VERSION = "0.5.0";
+declare const VERSION = "0.9.0";
+declare const ENGINE_VERSION = "0.9.0";
 
 declare class EngineManager {
     private enginePath;
@@ -31,6 +31,28 @@ declare class EngineManager {
     get readlineInterface(): Interface | null;
 }
 
+interface AttestClientOptions {
+    /** Logger override. When omitted, warnings only emit if ATTEST_DEBUG_PROTOCOL=1. */
+    logger?: ProtocolLogger;
+    /** Listener invoked for every protocol diagnostic. */
+    onDiagnostic?: (diagnostic: ProtocolDiagnostic) => void;
+    /** Diagnostic ring-buffer capacity. Default 32. */
+    diagnosticBufferSize?: number;
+    /** Number of diagnostics in `desyncWindowMs` that triggers desync rejection. Default 3. */
+    desyncThreshold?: number;
+    /** Window length for desync detection in milliseconds. Default 1000. */
+    desyncWindowMs?: number;
+}
+/**
+ * AttestClient owns request/response correlation over the engine's
+ * NDJSON protocol.
+ *
+ * Desync detection is a one-way latch: once the diagnostic rate breaches
+ * `desyncThreshold` within `desyncWindowMs` the client refuses further
+ * sends for the lifetime of the instance. There is no reset hook by
+ * design — recovery requires constructing a new client over a fresh
+ * engine subprocess.
+ */
 declare class AttestClient {
     private readonly engine;
     private requestId;
@@ -38,10 +60,23 @@ declare class AttestClient {
     private writeChain;
     private readerActive;
     private lineHandler;
-    constructor(engine: EngineManager);
+    private readonly logger;
+    private readonly diagnosticBuffer;
+    private readonly diagnosticListeners;
+    private readonly desyncThreshold;
+    private readonly desyncWindowMs;
+    private desynced;
+    constructor(engine: EngineManager, options?: AttestClientOptions);
+    /** Subscribe to protocol diagnostics. Returns an unsubscribe function. */
+    onProtocolDiagnostic(listener: (diagnostic: ProtocolDiagnostic) => void): () => void;
+    /** Snapshot of the diagnostic ring buffer (most recent last). */
+    protocolDiagnostics(): readonly ProtocolDiagnostic[];
     startReader(): void;
     stopReader(): void;
+    private recordDiagnostic;
+    private maybeTriggerDesync;
     private handleLine;
+    private handleDecodeFailure;
     private failAll;
     sendRequest(method: string, params: Record<string, unknown>, timeoutMs?: number): Promise<unknown>;
     evaluateBatch(trace: Trace, assertions: readonly Assertion[], options?: {
@@ -50,6 +85,12 @@ declare class AttestClient {
     submitPluginResult(traceId: string, pluginName: string, assertionId: string, status: string, score: number, explanation: string, options?: {
         timeout?: number;
     }): Promise<boolean>;
+    queryDrift(assertionId: string, windowSize?: number, options?: {
+        timeout?: number;
+    }): Promise<DriftReport>;
+    generateUserMessage(persona: SimulatePersona, conversationHistory: readonly ConversationMessage[], faultConfig?: SimulateFaultConfig, options?: {
+        timeout?: number;
+    }): Promise<string>;
 }
 
 declare class TraceBuilder {
@@ -106,6 +147,29 @@ declare class AgentResult {
     get failCount(): number;
 }
 
+/**
+ * Terminal-grade rendering of Attest assertion failures.
+ *
+ * Mirrors the Python attest.cli.diagnostics module so PR comments,
+ * pytest output, and vitest output show the same fields. Use as the
+ * message argument to a vitest `expect(result.passed, ...)` so
+ * reviewers see trace path, expected vs actual, and judge metadata.
+ */
+
+interface RenderOptions {
+    /** Emit ANSI colour escapes. Default false (plain text). */
+    color?: boolean;
+    /** Source location annotation appended after each block. */
+    testFile?: string;
+    /** Render passing assertions in addition to failures. Default false. */
+    includePassing?: boolean;
+}
+declare function renderAssertionFailure(result: AssertionResult, options?: RenderOptions): string;
+declare function renderSummary(result: AgentResult, options?: {
+    color?: boolean;
+}): string;
+declare function renderDiagnostics(result: AgentResult, options?: RenderOptions): string;
+
 declare class TraceTree {
     readonly root: Trace;
     constructor(root: Trace);
@@ -517,7 +581,7 @@ declare class ExpectChain {
      * ```
      */
     outputSimilarTo(reference: string, opts?: {
-        threshold?: number;
+        threshold?: number | "dynamic";
         model?: string;
         soft?: boolean;
     }): this;
@@ -538,7 +602,7 @@ declare class ExpectChain {
      */
     passesJudge(criteria: string, opts?: {
         rubric?: string;
-        threshold?: number;
+        threshold?: number | "dynamic";
         model?: string;
         soft?: boolean;
     }): this;
@@ -805,21 +869,36 @@ declare function tier(level: number): <T extends (...args: never[]) => unknown>(
  * @param options.simulation - Enable simulation mode. All `evaluateBatch` calls
  *   return deterministic pass results without spawning the engine or calling any
  *   LLM. Equivalent to setting `ATTEST_SIMULATION=1` in the environment.
+ * @param options.sampleRate - Sampling rate for continuous evaluation (0.0–1.0).
+ *   Equivalent to setting `ATTEST_SAMPLE_RATE` in the environment.
  */
 declare function config(options: {
     simulation?: boolean;
+    sampleRate?: number;
 }): void;
 /**
  * Returns true when simulation mode is active via either the programmatic API
  * or the `ATTEST_SIMULATION` environment variable.
  */
 declare function isSimulationMode(): boolean;
+/**
+ * Returns the sampling rate for continuous evaluation. Priority:
+ * programmatic `config({ sampleRate })` > `ATTEST_SAMPLE_RATE` env var > 0.
+ */
+declare function getSampleRate(): number;
 /**
  * Reset all programmatic config to defaults. Primarily for use in tests.
  * Environment variables are not affected.
  */
 declare function resetConfig(): void;
 
+/**
+ * Return a deterministic all-pass result for every assertion without
+ * contacting the engine or any LLM provider. Used when simulation mode
+ * is active (`ATTEST_SIMULATION=1` or `config({ simulation: true })`).
+ */
+declare function simulationEvaluateBatch(assertions: readonly Assertion[]): EvaluateBatchResult;
+
 interface PluginResult {
     readonly status: string;
     readonly score: number;
@@ -882,6 +961,64 @@ declare class ContinuousEvalRunner {
     private scheduleNext;
 }
 
+/**
+ * Calibration helpers — TS parity with engine/internal/assertion/judge/calibration.go
+ * and sdks/python/src/attest/calibration.py.
+ *
+ * Computes Cohen's κ, agreement %, and ROC-AUC over (human, judge) score
+ * pairs and parses CSV/JSONL labels files. Metrics match the engine
+ * byte-for-byte on the same input.
+ */
+interface LabelPair {
+    readonly human: number;
+    readonly judge: number;
+}
+interface AgreementResult {
+    readonly threshold: number;
+    readonly n: number;
+    readonly agreement: number;
+    readonly cohenKappa: number;
+    readonly rocAuc: number;
+}
+interface LabeledRecord {
+    input: string;
+    humanLabel: number;
+    judgeScore: number;
+    judgeKnown: boolean;
+}
+/**
+ * Compute Cohen's κ, agreement %, and ROC-AUC over labeled pairs.
+ *
+ * Throws when pairs is empty or threshold is outside (0, 1). Cohen's κ
+ * degenerates to 0 when all labels land in a single class for both
+ * raters; ROC-AUC is reported as 0 when only one human-class is
+ * present.
+ */
+declare function computeAgreement(pairs: readonly LabelPair[], threshold: number): AgreementResult;
+/**
+ * Parse a 2- or 3-column CSV (input, human_label[, judge_score]).
+ * Header detected when row[0][1] does not parse as a float. Lines
+ * beginning with `#` are skipped.
+ */
+declare function loadLabelsCSV(text: string): LabeledRecord[];
+/**
+ * Parse a newline-delimited JSON file. Each line is an object with the
+ * shape {"input": "...", "human_label": 0.9, "judge_score": 0.8}.
+ * `judge_score` is optional.
+ */
+declare function loadLabelsJSONL(text: string): LabeledRecord[];
+/**
+ * Return the 16-character SHA-256 prefix used by the engine for prompt
+ * hashes. Mirrors promptHash in engine/internal/assertion/judge_eval.go
+ * so SDK-recorded calibration rows align with JudgeMetadata.prompt_hash.
+ */
+declare function promptHash(text: string): string;
+
+declare const FRIENDLY_USER: SimulatePersona;
+declare const ADVERSARIAL_USER: SimulatePersona;
+declare const CONFUSED_USER: SimulatePersona;
+declare const COOPERATIVE_USER: SimulatePersona;
+
 declare class ManualAdapter {
     private readonly agentId;
     constructor(agentId?: string);
@@ -1082,4 +1219,4 @@ interface TraceAdapter {
     traceFromResponse(response: unknown, options?: Record<string, unknown>): Trace;
 }
 
-export { Agent, AgentResult, AlertDispatcher, type AlertPayload, AnthropicAdapter, Assertion, AssertionResult, AttestClient, type AttestPlugin, ContinuousEvalRunner, ENGINE_VERSION, EngineManager, EvaluateBatchResult, ExpectChain, GeminiAdapter, InitializeResult, LangChainAdapter, ManualAdapter, OTelAdapter, OllamaAdapter, OpenAIAdapter, PluginRegistry, type PluginResult, Sampler, Step, TIER_1, TIER_2, TIER_3, Trace, type TraceAdapter, TraceBuilder, TraceTree, VERSION, activeBuilder, agent, attestExpect, config, delegate, executePluginAssertion, isSimulationMode, resetConfig, tier };
+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, 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, resetConfig, simulationEvaluateBatch, tier };

package/dist/cjs/proto/index.cjs

@@ -27,6 +27,8 @@ __export(proto_exports, {
   ERR_SESSION_ERROR: () => ERR_SESSION_ERROR,
   ERR_TIMEOUT: () => ERR_TIMEOUT,
   EngineTimeoutError: () => EngineTimeoutError,
+  ProtocolDesyncError: () => ProtocolDesyncError,
+  ProtocolDiagnosticBuffer: () => ProtocolDiagnosticBuffer,
   ProtocolError: () => ProtocolError,
   STATUS_HARD_FAIL: () => STATUS_HARD_FAIL,
   STATUS_PASS: () => STATUS_PASS,
@@ -49,6 +51,8 @@ __export(proto_exports, {
   encodeRequest: () => encodeRequest,
   extractId: () => extractId,
   extractResult: () => extractResult,
+  layerForType: () => layerForType,
+  previewLine: () => previewLine,
   traceId: () => traceId
 });
 module.exports = __toCommonJS(proto_exports);
@@ -86,6 +90,27 @@ function assertionId(id) {
 function agentId(id) {
   return id;
 }
+function layerForType(assertionType) {
+  switch (assertionType) {
+    case "schema":
+      return 1;
+    case "constraint":
+      return 2;
+    case "trace":
+    case "trace_tree":
+      return 3;
+    case "content":
+      return 4;
+    case "embedding":
+      return 5;
+    case "llm_judge":
+      return 6;
+    case "plugin":
+      return 8;
+    default:
+      return 0;
+  }
+}
 
 // src/proto/errors.ts
 var ProtocolError = class extends Error {
@@ -112,6 +137,58 @@ var EngineTimeoutError = class extends Error {
     this.timeoutMs = timeoutMs;
   }
 };
+var ProtocolDesyncError = class extends Error {
+  diagnostics;
+  constructor(message, diagnostics) {
+    super(message);
+    this.name = "ProtocolDesyncError";
+    this.diagnostics = diagnostics;
+  }
+};
+
+// src/proto/diagnostics.ts
+var DEFAULT_BUFFER_SIZE = 32;
+var MAX_LINE_PREVIEW = 512;
+var ProtocolDiagnosticBuffer = class {
+  capacity;
+  entries = [];
+  constructor(capacity = DEFAULT_BUFFER_SIZE) {
+    if (capacity <= 0) {
+      throw new Error(
+        `ProtocolDiagnosticBuffer capacity must be positive, got ${capacity}`
+      );
+    }
+    this.capacity = capacity;
+  }
+  push(diagnostic) {
+    if (this.entries.length >= this.capacity) {
+      this.entries.shift();
+    }
+    this.entries.push(diagnostic);
+  }
+  snapshot() {
+    return this.entries.slice();
+  }
+  countWithin(windowMs, nowMs) {
+    const cutoff = nowMs - windowMs;
+    let count = 0;
+    for (let i = this.entries.length - 1; i >= 0; i -= 1) {
+      if (this.entries[i].timestampMs >= cutoff) {
+        count += 1;
+      } else {
+        break;
+      }
+    }
+    return count;
+  }
+  clear() {
+    this.entries.length = 0;
+  }
+};
+function previewLine(line) {
+  if (line.length <= MAX_LINE_PREVIEW) return line;
+  return line.slice(0, MAX_LINE_PREVIEW) + "\u2026";
+}
 
 // src/proto/codec.ts
 function encodeRequest(id, method, params) {
@@ -181,6 +258,8 @@ function extractId(response) {
   ERR_SESSION_ERROR,
   ERR_TIMEOUT,
   EngineTimeoutError,
+  ProtocolDesyncError,
+  ProtocolDiagnosticBuffer,
   ProtocolError,
   STATUS_HARD_FAIL,
   STATUS_PASS,
@@ -203,6 +282,8 @@ function extractId(response) {
   encodeRequest,
   extractId,
   extractResult,
+  layerForType,
+  previewLine,
   traceId
 });
 //# sourceMappingURL=index.cjs.map

package/dist/cjs/proto/index.cjs.map

@@ -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;\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;\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","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":[]}
+{"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;\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  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\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;;;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;AAgMO,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;;;ACtOO,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":[]}

package/dist/cjs/proto/index.d.cts

@@ -65,7 +65,7 @@ interface EmbeddingSpec {
     readonly type: "embedding";
     readonly target: string;
     readonly reference: string;
-    readonly threshold: number;
+    readonly threshold: number | "dynamic";
     readonly model?: string;
     readonly soft: boolean;
 }
@@ -74,7 +74,7 @@ interface LlmJudgeSpec {
     readonly target: string;
     readonly criteria: string;
     readonly rubric: string;
-    readonly threshold: number;
+    readonly threshold: number | "dynamic";
     readonly model?: string;
     readonly soft: boolean;
 }
@@ -125,6 +125,46 @@ interface Assertion {
     readonly spec: Record<string, unknown>;
     readonly request_id?: string;
 }
+/**
+ * Audit trail for an LLM-judge evaluation. Mirrors
+ * engine/pkg/types/assertion.go::JudgeMetadata so report v2 payloads
+ * round-trip without losing the model / rubric / variance evidence
+ * reviewers need to triage a judge failure.
+ */
+interface JudgeMetadata {
+    readonly model?: string;
+    readonly rubric_name?: string;
+    readonly rubric_version?: string;
+    readonly prompt_hash?: string;
+    readonly sample_scores?: readonly number[];
+    readonly score_mean?: number;
+    readonly score_stddev?: number;
+    readonly high_variance?: boolean;
+    readonly bias_probes?: readonly BiasProbe[];
+    readonly calibration?: JudgeAgreement;
+}
+/**
+ * One judge-bias probe result. Score is the score the judge produced on
+ * the mutated input; delta is `score - baseline`. A well-calibrated
+ * judge has |delta| close to zero. Mirrors
+ * engine/pkg/types/assertion.go::BiasProbe.
+ */
+interface BiasProbe {
+    readonly name: string;
+    readonly score: number;
+    readonly delta: number;
+}
+/**
+ * Calibration agreement summary keyed by
+ * (rubric_name, rubric_version, prompt_hash). Mirrors
+ * engine/pkg/types/assertion.go::JudgeAgreement.
+ */
+interface JudgeAgreement {
+    readonly label_count: number;
+    readonly agreement: number;
+    readonly cohen_kappa: number;
+    readonly roc_auc?: number;
+}
 interface AssertionResult {
     readonly assertion_id: string;
     readonly status: string;
@@ -133,7 +173,21 @@ interface AssertionResult {
     readonly cost?: number;
     readonly duration_ms?: number;
     readonly request_id?: string;
+    readonly threshold_source?: string;
+    readonly layer?: number;
+    readonly type?: string;
+    readonly trace_node_path?: string;
+    readonly expected?: string;
+    readonly actual?: string;
+    readonly suggested_action?: string;
+    readonly judge_metadata?: JudgeMetadata;
 }
+/**
+ * Map an assertion-type string to its pipeline layer (1–8). Returns 0
+ * when the type is unknown so callers can route uncategorised results
+ * into their own bucket.
+ */
+declare function layerForType(assertionType: string | undefined): number;
 interface ErrorData {
     readonly error_type: string;
     readonly retryable: boolean;
@@ -182,6 +236,13 @@ interface EvaluateBatchResult {
     readonly results: readonly AssertionResult[];
     readonly total_cost?: number;
     readonly total_duration_ms?: number;
+    /**
+     * True when this result was produced by an SDK simulation shim rather
+     * than a real engine evaluation. Engine-produced results always carry
+     * simulated=false. CI gates can fail builds that accidentally evaluate
+     * against the simulator by checking this flag.
+     */
+    readonly simulated?: boolean;
 }
 interface ShutdownResult {
     readonly sessions_completed: number;
@@ -193,6 +254,77 @@ interface SubmitPluginResultParams {
     readonly assertion_id: string;
     readonly result: AssertionResult;
 }
+interface DriftReport {
+    readonly assertion_id: string;
+    readonly mean: number;
+    readonly stddev: number;
+    readonly count: number;
+    readonly latest_score: number;
+    readonly deviation: number;
+    readonly status: string;
+}
+interface QueryDriftParams {
+    readonly assertion_id: string;
+    readonly window_size: number;
+}
+interface QueryDriftResult {
+    readonly report: DriftReport;
+}
+interface SimulatePersona {
+    readonly name: string;
+    readonly system_prompt: string;
+    readonly style: string;
+    readonly temperature: number;
+    readonly max_tokens?: number;
+}
+interface SimulateFaultConfig {
+    readonly error_rate: number;
+    readonly latency_jitter_ms: number;
+    readonly content_corruption: boolean;
+    readonly timeout_after_ms: number;
+}
+interface ConversationMessage {
+    readonly role: string;
+    readonly content: string;
+}
+interface GenerateUserMessageParams {
+    readonly persona: SimulatePersona;
+    readonly conversation_history: readonly ConversationMessage[];
+    readonly fault_config?: SimulateFaultConfig;
+}
+interface GenerateUserMessageResult {
+    readonly message: string;
+}
+
+/**
+ * Protocol diagnostic types for client-side observability.
+ *
+ * The reader loop emits diagnostics whenever it cannot route a response
+ * line: malformed JSON, missing id, unknown id, or non-routable protocol
+ * errors. Diagnostics are buffered in a bounded ring (32 entries) and
+ * trigger a desync error when the rate exceeds the configured threshold.
+ */
+type ProtocolDiagnosticKind = "malformed_json" | "non_object_response" | "invalid_jsonrpc_version" | "missing_id" | "unknown_id" | "non_routable_error";
+interface ProtocolDiagnostic {
+    readonly kind: ProtocolDiagnosticKind;
+    readonly message: string;
+    readonly rawLine: string;
+    readonly timestampMs: number;
+}
+interface ProtocolLogger {
+    warn(message: string, ...args: unknown[]): void;
+    error(message: string, ...args: unknown[]): void;
+}
+declare class ProtocolDiagnosticBuffer {
+    private readonly capacity;
+    private readonly entries;
+    constructor(capacity?: number);
+    push(diagnostic: ProtocolDiagnostic): void;
+    snapshot(): readonly ProtocolDiagnostic[];
+    countWithin(windowMs: number, nowMs: number): number;
+    clear(): void;
+}
+declare function previewLine(line: string): string;
 
 declare class ProtocolError extends Error {
     readonly code: number;
@@ -205,6 +337,16 @@ declare class EngineTimeoutError extends Error {
     readonly timeoutMs: number;
     constructor(method: string, timeoutMs: number);
 }
+/**
+ * Raised when the reader loop detects that engine output has lost
+ * framing or contains repeated unparseable lines. All in-flight requests
+ * are rejected with this error so callers see a visible failure rather
+ * than a hung promise.
+ */
+declare class ProtocolDesyncError extends Error {
+    readonly diagnostics: readonly ProtocolDiagnostic[];
+    constructor(message: string, diagnostics: readonly ProtocolDiagnostic[]);
+}
 
 /**
  * Encode a JSON-RPC 2.0 request as an NDJSON line.
@@ -225,4 +367,4 @@ declare function extractResult(response: RPCResponse): unknown;
  */
 declare function extractId(response: RPCResponse): number;
 
-export { type AgentId, type Assertion, type AssertionId, type AssertionResult, type AssertionSpec, type ConstraintSpec, type ContentSpec, 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 InitializeParams, type InitializeResult, type LlmJudgeSpec, type PluginSpec, ProtocolError, type RPCError, type RPCRequest, type RPCResponse, STATUS_HARD_FAIL, STATUS_PASS, STATUS_SOFT_FAIL, STEP_AGENT_CALL, STEP_LLM_CALL, STEP_RETRIEVAL, STEP_TOOL_CALL, type SchemaSpec, type ShutdownResult, 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, traceId };
+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, type GenerateUserMessageParams, type GenerateUserMessageResult, type InitializeParams, type InitializeResult, type JudgeAgreement, type JudgeMetadata, type LlmJudgeSpec, type PluginSpec, ProtocolDesyncError, type ProtocolDiagnostic, ProtocolDiagnosticBuffer, type ProtocolDiagnosticKind, ProtocolError, type ProtocolLogger, type QueryDriftParams, type QueryDriftResult, type RPCError, type RPCRequest, type RPCResponse, STATUS_HARD_FAIL, 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 };