@tangle-network/agent-eval 0.150.1 → 0.161.0

package/dist/dspy-rlm-engine-DptEII26.js.map

@@ -0,0 +1 @@
+{"version":3,"file":"dspy-rlm-engine-DptEII26.js","names":["isRecord"],"sources":["../src/analyst/finding-codec.ts","../src/analyst/trace-tool-callback.ts","../src/analyst/dspy-rlm-engine.ts"],"sourcesContent":["/**\n * The raw-finding wire contract — one decoder for both sides of the process\n * boundary.\n *\n * A finding array crosses a process boundary twice: an optimizer bridge writes\n * it in Python, TypeScript reads it back. Two decoders written independently\n * accept different shapes, so Python could report a completed investigation\n * whose rows TypeScript then dropped — the paid work disappearing between two\n * \"successes\". This module is the TypeScript half; `finding_contract.json`\n * carries the same rules to the Python half, generated from the schema here.\n *\n * Decoding never invents an empty result. A value that is not an array of rows\n * fails with the exact type it was, and a malformed row is reported with its\n * index and path while its valid siblings survive.\n */\n\nimport { type RawAnalystFinding, RawAnalystFindingSchema } from './finding-signature'\nimport { parseFindingSubject } from './finding-subject'\nimport { coerceJson } from './parse-tolerant'\n\n/**\n * Version of the wire contract both languages implement. Bump when the\n * accepted row shape changes; the Python package pins the same value, and the\n * emitted contract carries it so a mismatched pair is visible.\n */\nexport const FINDING_WIRE_CONTRACT_VERSION = 1\n\n/** Why one row was refused. Stable across languages: Python reports the same codes. */\nexport type FindingRejectionCode = 'not-an-object' | 'schema' | 'invalid-subject' | 'row-limit'\n\n/** A refused row, named precisely enough to repair or report. */\nexport interface RejectedFindingRow {\n  /** Position in the submitted array. */\n  readonly index: number\n  /** Dotted path to the offending field, `''` for the row itself. */\n  readonly path: string\n  readonly code: FindingRejectionCode\n  readonly message: string\n}\n\nexport interface DecodedFindingArray {\n  readonly accepted: RawAnalystFinding[]\n  readonly rejected: RejectedFindingRow[]\n  /**\n   * Set when the value was not a finding array at all — the array itself is\n   * the failure, so `accepted` and `rejected` are both empty and a caller must\n   * not read the result as \"no findings\".\n   */\n  readonly topLevelError?: string\n}\n\n/**\n * Rows past this count are refused rather than validated. A model that emits\n * thousands of rows has lost the plot, and the diagnostics for them would\n * dwarf the answer they came with.\n */\nexport const MAX_FINDING_ROWS = 500\n\n/**\n * Decode a submitted findings value into accepted rows plus per-row\n * diagnostics.\n *\n * Accepts an array of rows, or a string carrying one (a JSON array, possibly\n * fenced). Anything else — an object, a number, null, undefined — sets\n * `topLevelError` naming the type received: a caller that asked for findings\n * and got a number has a defect to report, not an empty result to record.\n */\nexport function decodeRawFindingArray(value: unknown): DecodedFindingArray {\n  const rows = findingRows(value)\n  if (typeof rows === 'string') return { accepted: [], rejected: [], topLevelError: rows }\n\n  const accepted: RawAnalystFinding[] = []\n  const rejected: RejectedFindingRow[] = []\n  for (const [index, row] of rows.entries()) {\n    if (index >= MAX_FINDING_ROWS) {\n      rejected.push({\n        index,\n        path: '',\n        code: 'row-limit',\n        message: `findings array exceeds ${MAX_FINDING_ROWS} rows`,\n      })\n      continue\n    }\n    const decoded = decodeRow(row, index)\n    if ('finding' in decoded) accepted.push(decoded.finding)\n    else rejected.push(decoded.rejection)\n  }\n  return { accepted, rejected }\n}\n\nfunction decodeRow(\n  row: unknown,\n  index: number,\n): { finding: RawAnalystFinding } | { rejection: RejectedFindingRow } {\n  if (row === null || typeof row !== 'object' || Array.isArray(row)) {\n    return {\n      rejection: {\n        index,\n        path: '',\n        code: 'not-an-object',\n        message: `finding row must be an object, received ${describe(row)}`,\n      },\n    }\n  }\n  const parsed = RawAnalystFindingSchema.safeParse(row)\n  if (!parsed.success) {\n    const issue = parsed.error.issues[0]\n    const path = (issue?.path ?? []).join('.')\n    // The schema refines `subject` against the grammar; name that refusal\n    // separately so a repair turn can be told which rule the row broke.\n    const code: FindingRejectionCode =\n      path === 'subject' && parseFindingSubject((row as { subject?: string }).subject) === null\n        ? 'invalid-subject'\n        : 'schema'\n    return {\n      rejection: { index, path, code, message: issue?.message ?? 'row does not match the schema' },\n    }\n  }\n  return { finding: parsed.data }\n}\n\n/**\n * Narrow a submitted value to candidate rows, or return the message naming why\n * it is not a findings array at all.\n */\nfunction findingRows(value: unknown): unknown[] | string {\n  if (Array.isArray(value)) return value\n  if (typeof value === 'string') {\n    const parsed = coerceJson(value)\n    if (parsed === undefined) {\n      return 'findings must be a JSON array; the string received is not JSON'\n    }\n    if (Array.isArray(parsed)) return parsed\n    const unwrapped = findingsProperty(parsed)\n    if (unwrapped !== undefined) return unwrapped\n    return `findings must be a JSON array, received ${describe(parsed)}`\n  }\n  const unwrapped = findingsProperty(value)\n  if (unwrapped !== undefined) return unwrapped\n  return `findings must be an array, received ${describe(value)}`\n}\n\n/** `{ findings: [...] }` is the one wrapper models emit often enough to unwrap. */\nfunction findingsProperty(value: unknown): unknown[] | undefined {\n  if (value === null || typeof value !== 'object' || Array.isArray(value)) return undefined\n  const inner = (value as Record<string, unknown>).findings\n  return Array.isArray(inner) ? inner : undefined\n}\n\nfunction describe(value: unknown): string {\n  if (value === null) return 'null'\n  if (Array.isArray(value)) return 'an array'\n  return `a ${typeof value}`\n}\n\n/** One line per refused row, for a repair prompt or an error message. */\nexport function describeRejectedRows(rejected: readonly RejectedFindingRow[]): string {\n  return rejected\n    .map((row) => `row ${row.index}${row.path ? ` field '${row.path}'` : ''}: ${row.message}`)\n    .join('\\n')\n}\n","import { randomBytes } from 'node:crypto'\nimport { createServer, type IncomingMessage, type ServerResponse } from 'node:http'\nimport {\n  type ExternalOptimizerCallbackLimits,\n  resolveExternalOptimizerCallbackLimits,\n} from '../campaign/external-optimizer-contracts'\nimport { closeServer, listenLocal, sendJson } from '../campaign/external-optimizer-http'\nimport type { TraceAnalysisToolDescriptor } from '../trace-analyst/tools'\n\nexport interface TraceToolCallback {\n  url: string\n  token: string\n  calls: () => number\n  close: () => Promise<void>\n}\n\nexport type TraceToolCallbackLimits = ExternalOptimizerCallbackLimits\n\n/** Expose one bounded trace-tool set only on an authenticated loopback socket. */\nexport async function startTraceToolCallback(args: {\n  tools: readonly TraceAnalysisToolDescriptor[]\n  maxCalls: number\n  /** Trace-tool request/response byte limits. Omitted fields use finite defaults. */\n  limits?: Partial<TraceToolCallbackLimits>\n  signal?: AbortSignal\n}): Promise<TraceToolCallback> {\n  if (!Number.isSafeInteger(args.maxCalls) || args.maxCalls <= 0) {\n    throw new TypeError('trace tool callback maxCalls must be a positive safe integer')\n  }\n  const limits = resolveExternalOptimizerCallbackLimits(args.limits, 'trace tool callback limits')\n  args.signal?.throwIfAborted()\n  const byName = new Map(args.tools.map((tool) => [tool.name, tool]))\n  if (byName.size !== args.tools.length) {\n    throw new Error('trace tool callback received duplicate tool names')\n  }\n\n  const token = randomBytes(32).toString('hex')\n  let calls = 0\n  let accepting = true\n  let closePromise: Promise<void> | undefined\n  const activeControllers = new Set<AbortController>()\n  const activeHandlers = new Set<Promise<void>>()\n  const server = createServer((request, response) => {\n    if (!accepting) {\n      sendJsonIfOpen(response, 503, { error: 'trace tool callback is closing' })\n      return\n    }\n    const controller = new AbortController()\n    const abortRequest = (): void => {\n      request.destroy()\n      response.destroy()\n    }\n    activeControllers.add(controller)\n    controller.signal.addEventListener('abort', abortRequest, { once: true })\n\n    let handler!: Promise<void>\n    handler = handleRequest(request, response, controller.signal).finally(() => {\n      controller.signal.removeEventListener('abort', abortRequest)\n      activeControllers.delete(controller)\n      activeHandlers.delete(handler)\n    })\n    activeHandlers.add(handler)\n    void handler.catch(() => undefined)\n  })\n  const port = await listenLocal(server)\n  const close = (): Promise<void> => {\n    closePromise ??= closeCallback()\n    return closePromise\n  }\n  const onAbort = (): void => {\n    void close().catch(() => undefined)\n  }\n  args.signal?.addEventListener('abort', onAbort, { once: true })\n  if (args.signal?.aborted) onAbort()\n\n  return {\n    url: `http://127.0.0.1:${port}/call`,\n    token,\n    calls: () => calls,\n    close,\n  }\n\n  async function handleRequest(\n    request: IncomingMessage,\n    response: ServerResponse,\n    signal: AbortSignal,\n  ): Promise<void> {\n    try {\n      if (request.method !== 'POST' || request.url !== '/call') {\n        sendJsonIfOpen(response, 404, { error: 'not found' })\n        return\n      }\n      if (request.headers.authorization !== `Bearer ${token}`) {\n        sendJsonIfOpen(response, 401, { error: 'unauthorized' })\n        return\n      }\n      if (calls >= args.maxCalls) {\n        sendJsonIfOpen(response, 429, { error: 'trace tool call limit reached' })\n        return\n      }\n      const body = await readJson(request, limits.maxRequestBytes)\n      if (!isRecord(body) || typeof body.name !== 'string' || !('args' in body)) {\n        sendJsonIfOpen(response, 400, { error: 'name and args are required' })\n        return\n      }\n      const tool = byName.get(body.name)\n      if (!tool) {\n        sendJsonIfOpen(response, 404, { error: `unknown trace tool '${body.name}'` })\n        return\n      }\n      calls += 1\n      const result = await tool.handler(body.args, { signal })\n      const encoded = JSON.stringify({ result })\n      if (Buffer.byteLength(encoded) > limits.maxResponseBytes) {\n        sendJsonIfOpen(response, 413, { error: 'trace tool response too large' })\n        return\n      }\n      response.writeHead(200, {\n        'content-type': 'application/json; charset=utf-8',\n        'content-length': String(Buffer.byteLength(encoded)),\n      })\n      response.end(encoded)\n    } catch (error) {\n      sendJsonIfOpen(response, signal.aborted ? 499 : 400, {\n        error: error instanceof Error ? error.message : String(error),\n      })\n    }\n  }\n\n  async function closeCallback(): Promise<void> {\n    args.signal?.removeEventListener('abort', onAbort)\n    accepting = false\n    const closingServer = closeServer(server)\n    server.closeIdleConnections?.()\n    for (const controller of activeControllers) controller.abort()\n    const [serverResult] = await Promise.allSettled([\n      closingServer,\n      waitForActiveHandlers(activeHandlers),\n    ])\n    if (activeControllers.size !== 0 || activeHandlers.size !== 0) {\n      throw new Error('trace tool callback closed with active requests')\n    }\n    if (serverResult?.status === 'rejected') throw serverResult.reason\n  }\n}\n\nasync function waitForActiveHandlers(activeHandlers: Set<Promise<void>>): Promise<void> {\n  while (activeHandlers.size > 0) {\n    await Promise.allSettled([...activeHandlers])\n  }\n}\n\nfunction readJson(request: IncomingMessage, maxRequestBytes: number): Promise<unknown> {\n  return new Promise((resolve, reject) => {\n    let size = 0\n    const chunks: Buffer[] = []\n    request.on('data', (chunk: Buffer) => {\n      size += chunk.length\n      if (size > maxRequestBytes) {\n        reject(new Error('trace tool request too large'))\n        request.destroy()\n        return\n      }\n      chunks.push(chunk)\n    })\n    request.on('error', reject)\n    request.on('end', () => {\n      try {\n        resolve(JSON.parse(Buffer.concat(chunks).toString('utf8')))\n      } catch (error) {\n        reject(error)\n      }\n    })\n  })\n}\n\nfunction sendJsonIfOpen(response: ServerResponse, status: number, body: unknown): void {\n  if (response.destroyed || response.writableEnded) return\n  sendJson(response, status, body)\n}\n\nfunction isRecord(value: unknown): value is Record<string, unknown> {\n  return typeof value === 'object' && value !== null && !Array.isArray(value)\n}\n","import {\n  assertExternalOptimizerModelBudget,\n  type ExternalOptimizerModelCall,\n  type ExternalOptimizerModelExecutionObservation,\n  type ExternalOptimizerModelProxy,\n  type ExternalOptimizerRunnerCommand,\n  removeCredentialEnvironment,\n  resolveExternalOptimizerCallbackLimits,\n  resolveExternalOptimizerProcessLimits,\n} from '../campaign/external-optimizer-contracts'\nimport { startExternalOptimizerModelProxy } from '../campaign/external-optimizer-model-proxy'\nimport { runWithCleanup } from '../campaign/external-optimizer-resources'\nimport { runExternalOptimizerProcess } from '../campaign/external-optimizer-subprocess'\nimport type { CustomTokenPricing } from '../cost-ledger'\nimport { resolveModelPricing } from '../metrics'\nimport type { TraceAnalysisEngine, TraceAnalysisEngineResult } from './engine'\nimport { decodeRawFindingArray } from './finding-codec'\nimport { startTraceToolCallback, type TraceToolCallbackLimits } from './trace-tool-callback'\n\nconst DEFAULT_TIMEOUT_MS = 10 * 60_000\n// 4096 is below what current coding models emit for a full findings array:\n// glm-5.2 through an OpenAI-compatible gateway returns 8192 and the request is\n// rejected outright (`502 — provider reported 8192 completion tokens,\n// exceeding requested limit 4096`), so the default failed 2/2 smoke cases\n// before any analysis ran. The cap exists to bound spend, and `maxCostUsd`\n// already does that directly, so it starts above what a real completion needs.\nconst DEFAULT_MODEL_OUTPUT_TOKENS = 16_384\nconst DEFAULT_MAX_COST_USD = 1\nconst DEFAULT_MAX_MODEL_REQUEST_BYTES = 16 * 1024 * 1024\nconst DEFAULT_MAX_MODEL_RESPONSE_BYTES = 4 * 1024 * 1024\nconst DEFAULT_TRACE_TOOL_TIMEOUT_MS = 60_000\nconst MAX_TIMER_DELAY_MS = 2_147_483_647\nconst BRIDGE_MODULE = 'agent_eval_rpc.dspy_rlm_bridge'\n/** Bumped whenever this engine's execution behavior changes. */\nconst DSPY_RLM_ENGINE_VERSION = '1.0.0'\n\nexport interface DspyRlmTraceEngineOptions {\n  /** Caller-owned execution path. Agent Eval never receives provider credentials. */\n  call: ExternalOptimizerModelCall\n  /** Stable public identity for the caller-owned path, such as an AgentProfile digest. */\n  callRef: string\n  /** Persist the finite execution record returned for every admitted call. */\n  recordExecution: (observation: ExternalOptimizerModelExecutionObservation) => void\n  model: string\n  /** Exact provider rates. Required when the model is absent from the pricing table. */\n  pricing?: CustomTokenPricing\n  /** Maximum provider spend for one investigation. Default: 1 USD. */\n  maxCostUsd?: number\n  /** Controller response cap. Default: 16384. */\n  maxOutputTokens?: number\n  /**\n   * Thinking tokens one controller turn may bill on top of its completion.\n   * A reasoning model bills these beyond `maxOutputTokens`, so the cost\n   * reservation must cover them. Default: four times the completion cap.\n   */\n  maxReasoningTokens?: number\n  /** Maximum caller-owned model invocations. Default derives from the analysis limits. */\n  maxModelRequests?: number\n  /** Maximum model request bytes. Default: 16 MiB. */\n  maxModelRequestBytes?: number\n  /** Maximum model response bytes. Default: 4 MiB. */\n  maxModelResponseBytes?: number\n  /** Deadline for one caller-owned model invocation. Default: the whole analysis deadline. */\n  modelRequestTimeoutMs?: number\n  /** Trace-tool loopback request/response byte limits. */\n  traceToolLimits?: Partial<TraceToolCallbackLimits>\n  /** Deadline for one Python-to-Node trace-tool call. Default: 60 seconds. */\n  traceToolTimeoutMs?: number\n  /**\n   * How the controller's reasoning and code fields are obtained.\n   *\n   * `tolerant` parses marker output strictly first, then recovers the fields\n   * deterministically from prose plus a fenced code block — the shape coding\n   * models naturally emit — at no extra model cost. `two-step` extracts with a\n   * second call per turn. `chat` accepts marker output only. Default:\n   * `tolerant`.\n   */\n  controlAdapter?: 'chat' | 'two-step' | 'tolerant'\n  /** Python command used to load agent-eval-rpc[dspy]. Default: python. */\n  runner?: ExternalOptimizerRunnerCommand\n  /** Whole investigation deadline. Default: 10 minutes. */\n  timeoutMs?: number\n}\n\n/** Use the official DSPy RLM as a bounded recursive trace-analysis engine. */\nexport function createDspyRlmTraceEngine(options: DspyRlmTraceEngineOptions): TraceAnalysisEngine {\n  assertOptions(options)\n  const maxOutputTokens = options.maxOutputTokens ?? DEFAULT_MODEL_OUTPUT_TOKENS\n  if (!Number.isSafeInteger(maxOutputTokens) || maxOutputTokens <= 0) {\n    throw new TypeError('DSPy RLM maxOutputTokens must be a positive safe integer')\n  }\n  const controlAdapter = options.controlAdapter ?? 'tolerant'\n  const maxReasoningTokens = options.maxReasoningTokens ?? maxOutputTokens * 4\n  if (!Number.isSafeInteger(maxReasoningTokens) || maxReasoningTokens < 0) {\n    throw new TypeError('DSPy RLM maxReasoningTokens must be a non-negative safe integer')\n  }\n  const timeoutMs = options.timeoutMs ?? DEFAULT_TIMEOUT_MS\n  assertTimerDelay(timeoutMs, 'timeoutMs')\n  const maxCostUsd = options.maxCostUsd ?? DEFAULT_MAX_COST_USD\n  if (!Number.isFinite(maxCostUsd) || maxCostUsd <= 0) {\n    throw new TypeError('DSPy RLM maxCostUsd must be positive and finite')\n  }\n  const pricing = options.pricing ?? pricingForModel(options.model)\n  const maxModelRequestBytes = options.maxModelRequestBytes ?? DEFAULT_MAX_MODEL_REQUEST_BYTES\n  const maxModelResponseBytes = options.maxModelResponseBytes ?? DEFAULT_MAX_MODEL_RESPONSE_BYTES\n  const modelRequestTimeoutMs = options.modelRequestTimeoutMs ?? timeoutMs\n  const traceToolTimeoutMs = options.traceToolTimeoutMs ?? DEFAULT_TRACE_TOOL_TIMEOUT_MS\n  assertExternalOptimizerModelBudget(\n    {\n      maxCostUsd,\n      maxRequests: options.maxModelRequests ?? 1,\n      maxRequestBytes: maxModelRequestBytes,\n      maxResponseBytes: maxModelResponseBytes,\n      maxOutputTokensPerRequest: maxOutputTokens,\n      maxReasoningTokensPerRequest: maxReasoningTokens,\n      pricing,\n      requestTimeoutMs: modelRequestTimeoutMs,\n    },\n    'DSPy RLM model limits',\n  )\n  resolveExternalOptimizerCallbackLimits(options.traceToolLimits, 'DSPy RLM trace tool limits')\n  assertTimerDelay(traceToolTimeoutMs, 'traceToolTimeoutMs')\n\n  const runner = sanitizedRunner(options.runner)\n  const processLimits = resolveExternalOptimizerProcessLimits(runner?.limits)\n  return {\n    id: 'dspy-rlm',\n    description: 'Official DSPy RLM with bounded trace tools and metered model calls.',\n    model: options.model,\n    version: DSPY_RLM_ENGINE_VERSION,\n    executionConfig: {\n      bridge_module: BRIDGE_MODULE,\n      call_ref: options.callRef,\n      model: options.model,\n      pricing: { ...pricing },\n      max_cost_usd: maxCostUsd,\n      max_output_tokens: maxOutputTokens,\n      max_reasoning_tokens: maxReasoningTokens,\n      control_adapter: controlAdapter,\n      timeout_ms: timeoutMs,\n      max_model_requests: options.maxModelRequests ?? null,\n      max_request_bytes: maxModelRequestBytes,\n      max_response_bytes: maxModelResponseBytes,\n      model_request_timeout_ms: modelRequestTimeoutMs,\n      trace_tool_limits: options.traceToolLimits ?? null,\n      trace_tool_timeout_ms: traceToolTimeoutMs,\n      process_limits: processLimits,\n      runner: runner ? 'caller-supplied' : 'default',\n      runner_command: runner?.command ?? null,\n    },\n    async analyze(request) {\n      const callback = await startTraceToolCallback({\n        tools: request.tools,\n        maxCalls: request.limits.maxToolCalls,\n        ...(options.traceToolLimits ? { limits: options.traceToolLimits } : {}),\n        ...(request.signal ? { signal: request.signal } : {}),\n      })\n      let modelProxy: ExternalOptimizerModelProxy | undefined\n      const modelExecutions: ExternalOptimizerModelExecutionObservation[] = []\n      const result = await runWithCleanup({\n        label: 'DSPy RLM trace-analysis resources',\n        run: async () => {\n          modelProxy = await startExternalOptimizerModelProxy({\n            call: options.call,\n            callRef: options.callRef,\n            recordExecution: (observation) => {\n              modelExecutions.push(structuredClone(observation))\n              options.recordExecution(observation)\n            },\n            model: options.model,\n            budget: {\n              maxCostUsd,\n              maxRequests: resolveMaxModelRequests(options.maxModelRequests, request.limits),\n              maxRequestBytes: maxModelRequestBytes,\n              maxResponseBytes: maxModelResponseBytes,\n              maxOutputTokensPerRequest: maxOutputTokens,\n              maxReasoningTokensPerRequest: maxReasoningTokens,\n              requestTimeoutMs: modelRequestTimeoutMs,\n              pricing,\n            },\n            costLedger: request.costLedger,\n            channel: 'analyst',\n            phase: request.costPhase,\n            actor: request.analystId,\n            ...(request.costTags ? { tags: request.costTags } : {}),\n            ...(request.signal ? { signal: request.signal } : {}),\n          })\n          request.log?.('trace analyst engine started', {\n            engine: 'dspy-rlm',\n            model: options.model,\n            tools: request.tools.map((tool) => tool.name),\n            limits: request.limits,\n          })\n          const raw = await runExternalOptimizerProcess<unknown>({\n            label: 'DSPy RLM trace analysis',\n            tempPrefix: 'agent-eval-dspy-rlm-',\n            module: BRIDGE_MODULE,\n            input: {\n              operation: 'analyze',\n              question: request.question,\n              instructions: request.instructions,\n              modelProxy: {\n                baseUrl: modelProxy.baseUrl,\n                apiKey: modelProxy.apiKey,\n                model: options.model,\n                maxOutputTokens,\n              },\n              toolCallback: {\n                url: callback.url,\n                token: callback.token,\n                timeoutMs: traceToolTimeoutMs,\n              },\n              toolSpecs: request.tools.map(({ name, description, parameters }) => ({\n                name,\n                description,\n                parameters,\n              })),\n              controlAdapter,\n              limits: {\n                maxIterations: request.limits.maxIterations,\n                maxLlmCalls: request.limits.maxLlmCalls,\n                maxOutputChars: request.limits.maxOutputChars,\n              },\n              // Omitted entirely when the caller supplied none, so a request\n              // without structured inputs sends the payload it always sent.\n              ...(request.taskInputs ? { taskInputs: request.taskInputs } : {}),\n            },\n            ...(runner ? { runner } : {}),\n            additionalArgs: ['--max-input-bytes', String(processLimits.maxInputBytes)],\n            timeoutMs,\n            ...(request.signal ? { signal: request.signal } : {}),\n          })\n          const parsed = parseBridgeOutput(raw, (index, reason) => {\n            request.log?.('finding rejected: bridge row failed schema validation', {\n              engine: 'dspy-rlm',\n              index,\n              reason,\n            })\n          })\n          const successfulCompletions = modelProxy.successfulCompletions()\n          const requestAttempts = modelProxy.requestAttempts()\n          if (parsed.modelCalls !== successfulCompletions) {\n            throw new Error(\n              `DSPy RLM reported ${parsed.modelCalls} model calls, but the provider proxy recorded ${successfulCompletions}`,\n            )\n          }\n          modelProxy.assertExecutionComplete()\n          return {\n            ...parsed,\n            toolCalls: callback.calls(),\n            runtime: {\n              ...parsed.runtime,\n              modelRequestAttempts: requestAttempts,\n              modelSuccessfulCompletions: successfulCompletions,\n              modelExecutions,\n            },\n          } satisfies TraceAnalysisEngineResult\n        },\n        cleanup: async () => {\n          const results = await Promise.allSettled([\n            ...(modelProxy ? [modelProxy.close()] : []),\n            callback.close(),\n          ])\n          const errors = results.flatMap((entry) =>\n            entry.status === 'rejected' ? [entry.reason] : [],\n          )\n          if (errors.length > 0) {\n            throw new AggregateError(errors, 'DSPy RLM resource cleanup failed')\n          }\n        },\n      })\n      request.log?.('trace analyst engine completed', {\n        engine: 'dspy-rlm',\n        model_calls: result.modelCalls,\n        model_request_attempts: result.runtime.modelRequestAttempts,\n        tool_calls: result.toolCalls,\n        findings: result.findings.length,\n      })\n      return result\n    },\n  }\n}\n\nfunction parseBridgeOutput(\n  value: unknown,\n  onRejectedFinding: (index: number, reason: string) => void,\n): Omit<TraceAnalysisEngineResult, 'toolCalls'> {\n  if (!isRecord(value)) throw new Error('DSPy RLM bridge output must be an object')\n  if (typeof value.answer !== 'string' || !value.answer.trim()) {\n    throw new Error('DSPy RLM bridge returned no answer')\n  }\n  // Findings are model output: one malformed row is model noise, not a bridge\n  // fault, and the rest of the paid investigation must survive it. The codec\n  // is the same decoder the Python side runs, so a row this bridge accepts is\n  // a row the optimizer could report, and vice versa.\n  const decoded = decodeRawFindingArray(value.findings)\n  if (decoded.topLevelError !== undefined) {\n    throw new Error(`DSPy RLM bridge findings must be an array: ${decoded.topLevelError}`)\n  }\n  const findings = decoded.accepted\n  const rejectedFindings = decoded.rejected.length\n  for (const rejection of decoded.rejected) {\n    onRejectedFinding(\n      rejection.index,\n      `${rejection.code}${rejection.path ? ` at ${rejection.path}` : ''}: ${rejection.message}`,\n    )\n  }\n  if (!Array.isArray(value.trajectory)) {\n    throw new Error('DSPy RLM bridge trajectory must be an array')\n  }\n  if (!Number.isSafeInteger(value.modelCalls) || (value.modelCalls as number) <= 0) {\n    throw new Error('DSPy RLM bridge modelCalls must be a positive safe integer')\n  }\n  if (!isRecord(value.runtime)) {\n    throw new Error('DSPy RLM bridge runtime must be an object')\n  }\n  return {\n    answer: value.answer,\n    findings,\n    trajectory: value.trajectory,\n    modelCalls: value.modelCalls as number,\n    runtime: { ...value.runtime, rejectedFindings },\n  }\n}\n\nfunction assertOptions(options: DspyRlmTraceEngineOptions): void {\n  for (const [name, value] of [\n    ['callRef', options.callRef],\n    ['model', options.model],\n  ] as const) {\n    if (typeof value !== 'string' || !value.trim() || value !== value.trim()) {\n      throw new TypeError(`DSPy RLM ${name} must be a trimmed non-empty string`)\n    }\n  }\n  if (typeof options.call !== 'function') {\n    throw new TypeError('DSPy RLM call must be a function')\n  }\n  if (typeof options.recordExecution !== 'function') {\n    throw new TypeError('DSPy RLM recordExecution must be a function')\n  }\n  if (\n    options.maxModelRequests !== undefined &&\n    (!Number.isSafeInteger(options.maxModelRequests) || options.maxModelRequests <= 0)\n  ) {\n    throw new TypeError('DSPy RLM maxModelRequests must be a positive safe integer')\n  }\n}\n\nfunction resolveMaxModelRequests(\n  configured: number | undefined,\n  limits: { maxIterations: number; maxLlmCalls: number },\n): number {\n  if (configured !== undefined) return configured\n  const derived = limits.maxIterations + limits.maxLlmCalls + 1\n  if (!Number.isSafeInteger(derived) || derived <= 0) {\n    throw new Error('DSPy RLM analysis limits produce an invalid model request limit')\n  }\n  return derived\n}\n\nfunction assertTimerDelay(value: number, field: string): void {\n  if (!Number.isSafeInteger(value) || value <= 0 || value > MAX_TIMER_DELAY_MS) {\n    throw new TypeError(`DSPy RLM ${field} must be between 1 and ${MAX_TIMER_DELAY_MS}`)\n  }\n}\n\nfunction pricingForModel(model: string): CustomTokenPricing {\n  const pricing = resolveModelPricing(model)\n  if (!pricing) {\n    throw new Error(\n      `no pricing is configured for '${model}'; provide DspyRlmTraceEngineOptions.pricing`,\n    )\n  }\n  return {\n    inputUsdPerMillion: pricing.input * 1_000,\n    outputUsdPerMillion: pricing.output * 1_000,\n  }\n}\n\nfunction sanitizedRunner(\n  runner: ExternalOptimizerRunnerCommand | undefined,\n): ExternalOptimizerRunnerCommand | undefined {\n  if (!runner) return undefined\n  return {\n    ...(runner.command ? { command: runner.command } : {}),\n    ...(runner.args ? { args: runner.args } : {}),\n    ...(runner.env ? { env: removeCredentialEnvironment(runner.env) } : {}),\n    ...(runner.limits ? { limits: { ...runner.limits } } : {}),\n  }\n}\n\nfunction isRecord(value: unknown): value is Record<string, unknown> {\n  return typeof value === 'object' && value !== null && !Array.isArray(value)\n}\n"],"mappings":";;;;;;;;;;;;;;AAmEA,SAAgB,sBAAsB,OAAqC;CACzE,MAAM,OAAO,YAAY,KAAK;CAC9B,IAAI,OAAO,SAAS,UAAU,OAAO;EAAE,UAAU,CAAC;EAAG,UAAU,CAAC;EAAG,eAAe;CAAK;CAEvF,MAAM,WAAgC,CAAC;CACvC,MAAM,WAAiC,CAAC;CACxC,KAAK,MAAM,CAAC,OAAO,QAAQ,KAAK,QAAQ,GAAG;EACzC,IAAI,SAAA,KAA2B;GAC7B,SAAS,KAAK;IACZ;IACA,MAAM;IACN,MAAM;IACN,SAAS;GACX,CAAC;GACD;EACF;EACA,MAAM,UAAU,UAAU,KAAK,KAAK;EACpC,IAAI,aAAa,SAAS,SAAS,KAAK,QAAQ,OAAO;OAClD,SAAS,KAAK,QAAQ,SAAS;CACtC;CACA,OAAO;EAAE;EAAU;CAAS;AAC9B;AAEA,SAAS,UACP,KACA,OACoE;CACpE,IAAI,QAAQ,QAAQ,OAAO,QAAQ,YAAY,MAAM,QAAQ,GAAG,GAC9D,OAAO,EACL,WAAW;EACT;EACA,MAAM;EACN,MAAM;EACN,SAAS,2CAA2C,SAAS,GAAG;CAClE,EACF;CAEF,MAAM,SAAS,wBAAwB,UAAU,GAAG;CACpD,IAAI,CAAC,OAAO,SAAS;EACnB,MAAM,QAAQ,OAAO,MAAM,OAAO;EAClC,MAAM,QAAQ,OAAO,QAAQ,CAAC,EAAA,CAAG,KAAK,GAAG;EAOzC,OAAO,EACL,WAAW;GAAE;GAAO;GAAM,MAJ1B,SAAS,aAAa,oBAAqB,IAA6B,OAAO,MAAM,OACjF,oBACA;GAE4B,SAAS,OAAO,WAAW;EAAgC,EAC7F;CACF;CACA,OAAO,EAAE,SAAS,OAAO,KAAK;AAChC;;;;;AAMA,SAAS,YAAY,OAAoC;CACvD,IAAI,MAAM,QAAQ,KAAK,GAAG,OAAO;CACjC,IAAI,OAAO,UAAU,UAAU;EAC7B,MAAM,SAAS,WAAW,KAAK;EAC/B,IAAI,WAAW,KAAA,GACb,OAAO;EAET,IAAI,MAAM,QAAQ,MAAM,GAAG,OAAO;EAClC,MAAM,YAAY,iBAAiB,MAAM;EACzC,IAAI,cAAc,KAAA,GAAW,OAAO;EACpC,OAAO,2CAA2C,SAAS,MAAM;CACnE;CACA,MAAM,YAAY,iBAAiB,KAAK;CACxC,IAAI,cAAc,KAAA,GAAW,OAAO;CACpC,OAAO,uCAAuC,SAAS,KAAK;AAC9D;;AAGA,SAAS,iBAAiB,OAAuC;CAC/D,IAAI,UAAU,QAAQ,OAAO,UAAU,YAAY,MAAM,QAAQ,KAAK,GAAG,OAAO,KAAA;CAChF,MAAM,QAAS,MAAkC;CACjD,OAAO,MAAM,QAAQ,KAAK,IAAI,QAAQ,KAAA;AACxC;AAEA,SAAS,SAAS,OAAwB;CACxC,IAAI,UAAU,MAAM,OAAO;CAC3B,IAAI,MAAM,QAAQ,KAAK,GAAG,OAAO;CACjC,OAAO,KAAK,OAAO;AACrB;;;;ACtIA,eAAsB,uBAAuB,MAMd;CAC7B,IAAI,CAAC,OAAO,cAAc,KAAK,QAAQ,KAAK,KAAK,YAAY,GAC3D,MAAM,IAAI,UAAU,8DAA8D;CAEpF,MAAM,SAAS,uCAAuC,KAAK,QAAQ,4BAA4B;CAC/F,KAAK,QAAQ,eAAe;CAC5B,MAAM,SAAS,IAAI,IAAI,KAAK,MAAM,KAAK,SAAS,CAAC,KAAK,MAAM,IAAI,CAAC,CAAC;CAClE,IAAI,OAAO,SAAS,KAAK,MAAM,QAC7B,MAAM,IAAI,MAAM,mDAAmD;CAGrE,MAAM,QAAQ,YAAY,EAAE,CAAC,CAAC,SAAS,KAAK;CAC5C,IAAI,QAAQ;CACZ,IAAI,YAAY;CAChB,IAAI;CACJ,MAAM,oCAAoB,IAAI,IAAqB;CACnD,MAAM,iCAAiB,IAAI,IAAmB;CAC9C,MAAM,SAAS,cAAc,SAAS,aAAa;EACjD,IAAI,CAAC,WAAW;GACd,eAAe,UAAU,KAAK,EAAE,OAAO,iCAAiC,CAAC;GACzE;EACF;EACA,MAAM,aAAa,IAAI,gBAAgB;EACvC,MAAM,qBAA2B;GAC/B,QAAQ,QAAQ;GAChB,SAAS,QAAQ;EACnB;EACA,kBAAkB,IAAI,UAAU;EAChC,WAAW,OAAO,iBAAiB,SAAS,cAAc,EAAE,MAAM,KAAK,CAAC;EAExE,IAAI;EACJ,UAAU,cAAc,SAAS,UAAU,WAAW,MAAM,CAAC,CAAC,cAAc;GAC1E,WAAW,OAAO,oBAAoB,SAAS,YAAY;GAC3D,kBAAkB,OAAO,UAAU;GACnC,eAAe,OAAO,OAAO;EAC/B,CAAC;EACD,eAAe,IAAI,OAAO;EAC1B,QAAa,YAAY,KAAA,CAAS;CACpC,CAAC;CACD,MAAM,OAAO,MAAM,YAAY,MAAM;CACrC,MAAM,cAA6B;EACjC,iBAAiB,cAAc;EAC/B,OAAO;CACT;CACA,MAAM,gBAAsB;EAC1B,MAAW,CAAC,CAAC,YAAY,KAAA,CAAS;CACpC;CACA,KAAK,QAAQ,iBAAiB,SAAS,SAAS,EAAE,MAAM,KAAK,CAAC;CAC9D,IAAI,KAAK,QAAQ,SAAS,QAAQ;CAElC,OAAO;EACL,KAAK,oBAAoB,KAAK;EAC9B;EACA,aAAa;EACb;CACF;CAEA,eAAe,cACb,SACA,UACA,QACe;EACf,IAAI;GACF,IAAI,QAAQ,WAAW,UAAU,QAAQ,QAAQ,SAAS;IACxD,eAAe,UAAU,KAAK,EAAE,OAAO,YAAY,CAAC;IACpD;GACF;GACA,IAAI,QAAQ,QAAQ,kBAAkB,UAAU,SAAS;IACvD,eAAe,UAAU,KAAK,EAAE,OAAO,eAAe,CAAC;IACvD;GACF;GACA,IAAI,SAAS,KAAK,UAAU;IAC1B,eAAe,UAAU,KAAK,EAAE,OAAO,gCAAgC,CAAC;IACxE;GACF;GACA,MAAM,OAAO,MAAM,SAAS,SAAS,OAAO,eAAe;GAC3D,IAAI,CAACA,WAAS,IAAI,KAAK,OAAO,KAAK,SAAS,YAAY,EAAE,UAAU,OAAO;IACzE,eAAe,UAAU,KAAK,EAAE,OAAO,6BAA6B,CAAC;IACrE;GACF;GACA,MAAM,OAAO,OAAO,IAAI,KAAK,IAAI;GACjC,IAAI,CAAC,MAAM;IACT,eAAe,UAAU,KAAK,EAAE,OAAO,uBAAuB,KAAK,KAAK,GAAG,CAAC;IAC5E;GACF;GACA,SAAS;GACT,MAAM,SAAS,MAAM,KAAK,QAAQ,KAAK,MAAM,EAAE,OAAO,CAAC;GACvD,MAAM,UAAU,KAAK,UAAU,EAAE,OAAO,CAAC;GACzC,IAAI,OAAO,WAAW,OAAO,IAAI,OAAO,kBAAkB;IACxD,eAAe,UAAU,KAAK,EAAE,OAAO,gCAAgC,CAAC;IACxE;GACF;GACA,SAAS,UAAU,KAAK;IACtB,gBAAgB;IAChB,kBAAkB,OAAO,OAAO,WAAW,OAAO,CAAC;GACrD,CAAC;GACD,SAAS,IAAI,OAAO;EACtB,SAAS,OAAO;GACd,eAAe,UAAU,OAAO,UAAU,MAAM,KAAK,EACnD,OAAO,iBAAiB,QAAQ,MAAM,UAAU,OAAO,KAAK,EAC9D,CAAC;EACH;CACF;CAEA,eAAe,gBAA+B;EAC5C,KAAK,QAAQ,oBAAoB,SAAS,OAAO;EACjD,YAAY;EACZ,MAAM,gBAAgB,YAAY,MAAM;EACxC,OAAO,uBAAuB;EAC9B,KAAK,MAAM,cAAc,mBAAmB,WAAW,MAAM;EAC7D,MAAM,CAAC,gBAAgB,MAAM,QAAQ,WAAW,CAC9C,eACA,sBAAsB,cAAc,CACtC,CAAC;EACD,IAAI,kBAAkB,SAAS,KAAK,eAAe,SAAS,GAC1D,MAAM,IAAI,MAAM,iDAAiD;EAEnE,IAAI,cAAc,WAAW,YAAY,MAAM,aAAa;CAC9D;AACF;AAEA,eAAe,sBAAsB,gBAAmD;CACtF,OAAO,eAAe,OAAO,GAC3B,MAAM,QAAQ,WAAW,CAAC,GAAG,cAAc,CAAC;AAEhD;AAEA,SAAS,SAAS,SAA0B,iBAA2C;CACrF,OAAO,IAAI,SAAS,SAAS,WAAW;EACtC,IAAI,OAAO;EACX,MAAM,SAAmB,CAAC;EAC1B,QAAQ,GAAG,SAAS,UAAkB;GACpC,QAAQ,MAAM;GACd,IAAI,OAAO,iBAAiB;IAC1B,uBAAO,IAAI,MAAM,8BAA8B,CAAC;IAChD,QAAQ,QAAQ;IAChB;GACF;GACA,OAAO,KAAK,KAAK;EACnB,CAAC;EACD,QAAQ,GAAG,SAAS,MAAM;EAC1B,QAAQ,GAAG,aAAa;GACtB,IAAI;IACF,QAAQ,KAAK,MAAM,OAAO,OAAO,MAAM,CAAC,CAAC,SAAS,MAAM,CAAC,CAAC;GAC5D,SAAS,OAAO;IACd,OAAO,KAAK;GACd;EACF,CAAC;CACH,CAAC;AACH;AAEA,SAAS,eAAe,UAA0B,QAAgB,MAAqB;CACrF,IAAI,SAAS,aAAa,SAAS,eAAe;CAClD,SAAS,UAAU,QAAQ,IAAI;AACjC;AAEA,SAASA,WAAS,OAAkD;CAClE,OAAO,OAAO,UAAU,YAAY,UAAU,QAAQ,CAAC,MAAM,QAAQ,KAAK;AAC5E;;;ACpKA,MAAM,qBAAqB,KAAK;AAOhC,MAAM,8BAA8B;AACpC,MAAM,uBAAuB;AAC7B,MAAM,kCAAkC,KAAK,OAAO;AACpD,MAAM,mCAAmC,IAAI,OAAO;AACpD,MAAM,gCAAgC;AACtC,MAAM,qBAAqB;AAC3B,MAAM,gBAAgB;;AAEtB,MAAM,0BAA0B;;AAmDhC,SAAgB,yBAAyB,SAAyD;CAChG,cAAc,OAAO;CACrB,MAAM,kBAAkB,QAAQ,mBAAmB;CACnD,IAAI,CAAC,OAAO,cAAc,eAAe,KAAK,mBAAmB,GAC/D,MAAM,IAAI,UAAU,0DAA0D;CAEhF,MAAM,iBAAiB,QAAQ,kBAAkB;CACjD,MAAM,qBAAqB,QAAQ,sBAAsB,kBAAkB;CAC3E,IAAI,CAAC,OAAO,cAAc,kBAAkB,KAAK,qBAAqB,GACpE,MAAM,IAAI,UAAU,iEAAiE;CAEvF,MAAM,YAAY,QAAQ,aAAa;CACvC,iBAAiB,WAAW,WAAW;CACvC,MAAM,aAAa,QAAQ,cAAc;CACzC,IAAI,CAAC,OAAO,SAAS,UAAU,KAAK,cAAc,GAChD,MAAM,IAAI,UAAU,iDAAiD;CAEvE,MAAM,UAAU,QAAQ,WAAW,gBAAgB,QAAQ,KAAK;CAChE,MAAM,uBAAuB,QAAQ,wBAAwB;CAC7D,MAAM,wBAAwB,QAAQ,yBAAyB;CAC/D,MAAM,wBAAwB,QAAQ,yBAAyB;CAC/D,MAAM,qBAAqB,QAAQ,sBAAsB;CACzD,mCACE;EACE;EACA,aAAa,QAAQ,oBAAoB;EACzC,iBAAiB;EACjB,kBAAkB;EAClB,2BAA2B;EAC3B,8BAA8B;EAC9B;EACA,kBAAkB;CACpB,GACA,uBACF;CACA,uCAAuC,QAAQ,iBAAiB,4BAA4B;CAC5F,iBAAiB,oBAAoB,oBAAoB;CAEzD,MAAM,SAAS,gBAAgB,QAAQ,MAAM;CAC7C,MAAM,gBAAgB,sCAAsC,QAAQ,MAAM;CAC1E,OAAO;EACL,IAAI;EACJ,aAAa;EACb,OAAO,QAAQ;EACf,SAAS;EACT,iBAAiB;GACf,eAAe;GACf,UAAU,QAAQ;GAClB,OAAO,QAAQ;GACf,SAAS,EAAE,GAAG,QAAQ;GACtB,cAAc;GACd,mBAAmB;GACnB,sBAAsB;GACtB,iBAAiB;GACjB,YAAY;GACZ,oBAAoB,QAAQ,oBAAoB;GAChD,mBAAmB;GACnB,oBAAoB;GACpB,0BAA0B;GAC1B,mBAAmB,QAAQ,mBAAmB;GAC9C,uBAAuB;GACvB,gBAAgB;GAChB,QAAQ,SAAS,oBAAoB;GACrC,gBAAgB,QAAQ,WAAW;EACrC;EACA,MAAM,QAAQ,SAAS;GACrB,MAAM,WAAW,MAAM,uBAAuB;IAC5C,OAAO,QAAQ;IACf,UAAU,QAAQ,OAAO;IACzB,GAAI,QAAQ,kBAAkB,EAAE,QAAQ,QAAQ,gBAAgB,IAAI,CAAC;IACrE,GAAI,QAAQ,SAAS,EAAE,QAAQ,QAAQ,OAAO,IAAI,CAAC;GACrD,CAAC;GACD,IAAI;GACJ,MAAM,kBAAgE,CAAC;GACvE,MAAM,SAAS,MAAM,eAAe;IAClC,OAAO;IACP,KAAK,YAAY;KACf,aAAa,MAAM,iCAAiC;MAClD,MAAM,QAAQ;MACd,SAAS,QAAQ;MACjB,kBAAkB,gBAAgB;OAChC,gBAAgB,KAAK,gBAAgB,WAAW,CAAC;OACjD,QAAQ,gBAAgB,WAAW;MACrC;MACA,OAAO,QAAQ;MACf,QAAQ;OACN;OACA,aAAa,wBAAwB,QAAQ,kBAAkB,QAAQ,MAAM;OAC7E,iBAAiB;OACjB,kBAAkB;OAClB,2BAA2B;OAC3B,8BAA8B;OAC9B,kBAAkB;OAClB;MACF;MACA,YAAY,QAAQ;MACpB,SAAS;MACT,OAAO,QAAQ;MACf,OAAO,QAAQ;MACf,GAAI,QAAQ,WAAW,EAAE,MAAM,QAAQ,SAAS,IAAI,CAAC;MACrD,GAAI,QAAQ,SAAS,EAAE,QAAQ,QAAQ,OAAO,IAAI,CAAC;KACrD,CAAC;KACD,QAAQ,MAAM,gCAAgC;MAC5C,QAAQ;MACR,OAAO,QAAQ;MACf,OAAO,QAAQ,MAAM,KAAK,SAAS,KAAK,IAAI;MAC5C,QAAQ,QAAQ;KAClB,CAAC;KAwCD,MAAM,SAAS,kBAAkB,MAvCf,4BAAqC;MACrD,OAAO;MACP,YAAY;MACZ,QAAQ;MACR,OAAO;OACL,WAAW;OACX,UAAU,QAAQ;OAClB,cAAc,QAAQ;OACtB,YAAY;QACV,SAAS,WAAW;QACpB,QAAQ,WAAW;QACnB,OAAO,QAAQ;QACf;OACF;OACA,cAAc;QACZ,KAAK,SAAS;QACd,OAAO,SAAS;QAChB,WAAW;OACb;OACA,WAAW,QAAQ,MAAM,KAAK,EAAE,MAAM,aAAa,kBAAkB;QACnE;QACA;QACA;OACF,EAAE;OACF;OACA,QAAQ;QACN,eAAe,QAAQ,OAAO;QAC9B,aAAa,QAAQ,OAAO;QAC5B,gBAAgB,QAAQ,OAAO;OACjC;OAGA,GAAI,QAAQ,aAAa,EAAE,YAAY,QAAQ,WAAW,IAAI,CAAC;MACjE;MACA,GAAI,SAAS,EAAE,OAAO,IAAI,CAAC;MAC3B,gBAAgB,CAAC,qBAAqB,OAAO,cAAc,aAAa,CAAC;MACzE;MACA,GAAI,QAAQ,SAAS,EAAE,QAAQ,QAAQ,OAAO,IAAI,CAAC;KACrD,CAAC,IACsC,OAAO,WAAW;MACvD,QAAQ,MAAM,yDAAyD;OACrE,QAAQ;OACR;OACA;MACF,CAAC;KACH,CAAC;KACD,MAAM,wBAAwB,WAAW,sBAAsB;KAC/D,MAAM,kBAAkB,WAAW,gBAAgB;KACnD,IAAI,OAAO,eAAe,uBACxB,MAAM,IAAI,MACR,qBAAqB,OAAO,WAAW,gDAAgD,uBACzF;KAEF,WAAW,wBAAwB;KACnC,OAAO;MACL,GAAG;MACH,WAAW,SAAS,MAAM;MAC1B,SAAS;OACP,GAAG,OAAO;OACV,sBAAsB;OACtB,4BAA4B;OAC5B;MACF;KACF;IACF;IACA,SAAS,YAAY;KAKnB,MAAM,UAAS,MAJO,QAAQ,WAAW,CACvC,GAAI,aAAa,CAAC,WAAW,MAAM,CAAC,IAAI,CAAC,GACzC,SAAS,MAAM,CACjB,CAAC,EAAA,CACsB,SAAS,UAC9B,MAAM,WAAW,aAAa,CAAC,MAAM,MAAM,IAAI,CAAC,CAClD;KACA,IAAI,OAAO,SAAS,GAClB,MAAM,IAAI,eAAe,QAAQ,kCAAkC;IAEvE;GACF,CAAC;GACD,QAAQ,MAAM,kCAAkC;IAC9C,QAAQ;IACR,aAAa,OAAO;IACpB,wBAAwB,OAAO,QAAQ;IACvC,YAAY,OAAO;IACnB,UAAU,OAAO,SAAS;GAC5B,CAAC;GACD,OAAO;EACT;CACF;AACF;AAEA,SAAS,kBACP,OACA,mBAC8C;CAC9C,IAAI,CAAC,SAAS,KAAK,GAAG,MAAM,IAAI,MAAM,0CAA0C;CAChF,IAAI,OAAO,MAAM,WAAW,YAAY,CAAC,MAAM,OAAO,KAAK,GACzD,MAAM,IAAI,MAAM,oCAAoC;CAMtD,MAAM,UAAU,sBAAsB,MAAM,QAAQ;CACpD,IAAI,QAAQ,kBAAkB,KAAA,GAC5B,MAAM,IAAI,MAAM,8CAA8C,QAAQ,eAAe;CAEvF,MAAM,WAAW,QAAQ;CACzB,MAAM,mBAAmB,QAAQ,SAAS;CAC1C,KAAK,MAAM,aAAa,QAAQ,UAC9B,kBACE,UAAU,OACV,GAAG,UAAU,OAAO,UAAU,OAAO,OAAO,UAAU,SAAS,GAAG,IAAI,UAAU,SAClF;CAEF,IAAI,CAAC,MAAM,QAAQ,MAAM,UAAU,GACjC,MAAM,IAAI,MAAM,6CAA6C;CAE/D,IAAI,CAAC,OAAO,cAAc,MAAM,UAAU,KAAM,MAAM,cAAyB,GAC7E,MAAM,IAAI,MAAM,4DAA4D;CAE9E,IAAI,CAAC,SAAS,MAAM,OAAO,GACzB,MAAM,IAAI,MAAM,2CAA2C;CAE7D,OAAO;EACL,QAAQ,MAAM;EACd;EACA,YAAY,MAAM;EAClB,YAAY,MAAM;EAClB,SAAS;GAAE,GAAG,MAAM;GAAS;EAAiB;CAChD;AACF;AAEA,SAAS,cAAc,SAA0C;CAC/D,KAAK,MAAM,CAAC,MAAM,UAAU,CAC1B,CAAC,WAAW,QAAQ,OAAO,GAC3B,CAAC,SAAS,QAAQ,KAAK,CACzB,GACE,IAAI,OAAO,UAAU,YAAY,CAAC,MAAM,KAAK,KAAK,UAAU,MAAM,KAAK,GACrE,MAAM,IAAI,UAAU,YAAY,KAAK,oCAAoC;CAG7E,IAAI,OAAO,QAAQ,SAAS,YAC1B,MAAM,IAAI,UAAU,kCAAkC;CAExD,IAAI,OAAO,QAAQ,oBAAoB,YACrC,MAAM,IAAI,UAAU,6CAA6C;CAEnE,IACE,QAAQ,qBAAqB,KAAA,MAC5B,CAAC,OAAO,cAAc,QAAQ,gBAAgB,KAAK,QAAQ,oBAAoB,IAEhF,MAAM,IAAI,UAAU,2DAA2D;AAEnF;AAEA,SAAS,wBACP,YACA,QACQ;CACR,IAAI,eAAe,KAAA,GAAW,OAAO;CACrC,MAAM,UAAU,OAAO,gBAAgB,OAAO,cAAc;CAC5D,IAAI,CAAC,OAAO,cAAc,OAAO,KAAK,WAAW,GAC/C,MAAM,IAAI,MAAM,iEAAiE;CAEnF,OAAO;AACT;AAEA,SAAS,iBAAiB,OAAe,OAAqB;CAC5D,IAAI,CAAC,OAAO,cAAc,KAAK,KAAK,SAAS,KAAK,QAAQ,oBACxD,MAAM,IAAI,UAAU,YAAY,MAAM,yBAAyB,oBAAoB;AAEvF;AAEA,SAAS,gBAAgB,OAAmC;CAC1D,MAAM,UAAU,oBAAoB,KAAK;CACzC,IAAI,CAAC,SACH,MAAM,IAAI,MACR,iCAAiC,MAAM,6CACzC;CAEF,OAAO;EACL,oBAAoB,QAAQ,QAAQ;EACpC,qBAAqB,QAAQ,SAAS;CACxC;AACF;AAEA,SAAS,gBACP,QAC4C;CAC5C,IAAI,CAAC,QAAQ,OAAO,KAAA;CACpB,OAAO;EACL,GAAI,OAAO,UAAU,EAAE,SAAS,OAAO,QAAQ,IAAI,CAAC;EACpD,GAAI,OAAO,OAAO,EAAE,MAAM,OAAO,KAAK,IAAI,CAAC;EAC3C,GAAI,OAAO,MAAM,EAAE,KAAK,4BAA4B,OAAO,GAAG,EAAE,IAAI,CAAC;EACrE,GAAI,OAAO,SAAS,EAAE,QAAQ,EAAE,GAAG,OAAO,OAAO,EAAE,IAAI,CAAC;CAC1D;AACF;AAEA,SAAS,SAAS,OAAkD;CAClE,OAAO,OAAO,UAAU,YAAY,UAAU,QAAQ,CAAC,MAAM,QAAQ,KAAK;AAC5E"}

package/dist/{emitter-CPBAhxum.js → emitter-BpYFQPj4.js}

@@ -244,23 +244,7 @@ function cryptoRandomId() {
 	if (typeof globalThis.crypto?.randomUUID === "function") return globalThis.crypto.randomUUID();
 	return `${Date.now().toString(36)}-${Math.random().toString(36).slice(2, 10)}`;
 }
-/** Helper to build an LLM span handle args object from a provider-shaped response. */
-function llmSpanFromProvider(args) {
-	return {
-		name: args.name ?? args.model,
-		model: args.model,
-		messages: args.messages,
-		output: args.output,
-		inputTokens: args.usage?.inputTokens,
-		outputTokens: args.usage?.outputTokens,
-		cachedTokens: args.usage?.cachedTokens,
-		cacheWriteTokens: args.usage?.cacheWriteTokens,
-		reasoningTokens: args.usage?.reasoningTokens,
-		costUsd: args.costUsd,
-		finishReason: args.finishReason
-	};
-}
 //#endregion
-export { llmSpanFromProvider as n, TraceEmitter as t };
+export { TraceEmitter as t };
 
-//# sourceMappingURL=emitter-CPBAhxum.js.map
+//# sourceMappingURL=emitter-BpYFQPj4.js.map

package/dist/emitter-BpYFQPj4.js.map

@@ -0,0 +1 @@
+{"version":3,"file":"emitter-BpYFQPj4.js","names":[],"sources":["../src/trace/emitter.ts"],"sourcesContent":["/**\n * TraceEmitter — hierarchical span builder that auto-parents using an\n * internal stack. One emitter per Run; emitters do NOT share state.\n *\n * Convenience methods (`llm`, `tool`, `retrieval`, `judge`, `sandbox`)\n * return a `SpanHandle` with `.end()` / `.fail()` so callers don't\n * have to thread spanIds manually. For async workflows that can't use\n * the stack (e.g. fan-out parallel calls), pass `parentSpanId`\n * explicitly.\n */\n\nimport type {\n  Artifact,\n  BudgetLedgerEntry,\n  EventKind,\n  JudgeSpan,\n  LlmSpan,\n  RetrievalSpan,\n  Run,\n  RunOutcome,\n  SandboxSpan,\n  Span,\n  SpanKind,\n  ToolSpan,\n  TraceEvent,\n} from './schema'\nimport type { TraceStore } from './store'\n\nexport interface SpanHandle<S extends Span = Span> {\n  span: S\n  end(patch?: Partial<S>): Promise<void>\n  fail(error: string | Error, patch?: Partial<S>): Promise<void>\n}\n\nexport interface RunCompleteHookContext {\n  runId: string\n  emitter: TraceEmitter\n  store: TraceStore\n  /** Outcome the caller passed to `endRun` (undefined for `abortRun`). */\n  outcome?: RunOutcome\n  /** Final run status. */\n  status: 'completed' | 'failed' | 'aborted'\n}\n\nexport type RunCompleteHook = (ctx: RunCompleteHookContext) => Promise<void> | void\n\nexport interface TraceEmitterOptions {\n  runId?: string\n  /** Inject a clock for deterministic tests. */\n  now?: () => number\n  /** Inject an id generator for deterministic tests. */\n  id?: () => string\n  /**\n   * Hooks fired after `endRun` / `abortRun` writes the final run state.\n   * Designed for trace-analyst auto-execution, integrity assertions, and\n   * outbound notifications. Hooks run sequentially in the order supplied.\n   *\n   * By default a hook that throws is swallowed and logged as a `note` event\n   * on the run — auto-orchestration must not crash the underlying flow.\n   * Set `hookErrors: 'throw'` to propagate.\n   */\n  onRunComplete?: RunCompleteHook[]\n  /** `'swallow'` (default) | `'throw'`. */\n  hookErrors?: 'swallow' | 'throw'\n}\n\nexport class TraceEmitter {\n  private store: TraceStore\n  private stack: string[] = []\n  private _runId: string\n  private now: () => number\n  private id: () => string\n  private hooks: RunCompleteHook[]\n  private hookErrors: 'swallow' | 'throw'\n\n  constructor(store: TraceStore, options: TraceEmitterOptions = {}) {\n    this.store = store\n    this.now = options.now ?? (() => Date.now())\n    this.id = options.id ?? (() => cryptoRandomId())\n    this._runId = options.runId ?? this.id()\n    this.hooks = options.onRunComplete ?? []\n    this.hookErrors = options.hookErrors ?? 'swallow'\n  }\n\n  get runId(): string {\n    return this._runId\n  }\n\n  get traceStore(): TraceStore {\n    return this.store\n  }\n\n  /** Append a hook after construction (e.g. attach the trace analyst). */\n  addRunCompleteHook(hook: RunCompleteHook): void {\n    this.hooks.push(hook)\n  }\n\n  // ── Run lifecycle ──────────────────────────────────────────────────\n\n  /**\n   * Begin a Run.\n   *\n   * `scenarioId` is required on the persisted Run shape — every Run downstream\n   * gets a non-empty scenarioId so filters and aggregations stay simple — but\n   * the INPUT here accepts it as optional. When omitted, startRun substitutes\n   * a sensible default (`run.layer ?? run.tags?.['kind'] ?? 'runtime'`) so\n   * runtime / operator / meta-eval runs that have no curated-scenario corpus\n   * to anchor to don't have to invent placeholder strings at the call site.\n   */\n  async startRun(\n    run: Omit<Run, 'runId' | 'scenarioId' | 'startedAt' | 'status'> & { scenarioId?: string },\n  ): Promise<Run> {\n    const scenarioId = run.scenarioId ?? run.layer ?? run.tags?.kind ?? 'runtime'\n    const full: Run = {\n      ...run,\n      scenarioId,\n      runId: this._runId,\n      startedAt: this.now(),\n      status: 'running',\n    }\n    await this.store.appendRun(full)\n    return full\n  }\n\n  async endRun(outcome?: RunOutcome): Promise<void> {\n    const status: 'completed' | 'failed' = outcome?.pass === false ? 'failed' : 'completed'\n    await this.store.updateRun(this._runId, { endedAt: this.now(), status, outcome })\n    await this.runHooks({ runId: this._runId, emitter: this, store: this.store, outcome, status })\n  }\n\n  async abortRun(reason: string): Promise<void> {\n    const outcome = { pass: false, notes: reason }\n    await this.store.updateRun(this._runId, {\n      endedAt: this.now(),\n      status: 'aborted',\n      outcome,\n    })\n    await this.runHooks({\n      runId: this._runId,\n      emitter: this,\n      store: this.store,\n      outcome,\n      status: 'aborted',\n    })\n  }\n\n  private async runHooks(ctx: RunCompleteHookContext): Promise<void> {\n    for (const hook of this.hooks) {\n      try {\n        await hook(ctx)\n      } catch (err) {\n        if (this.hookErrors === 'throw') throw err\n        try {\n          await this.store.appendEvent({\n            eventId: this.id(),\n            runId: this._runId,\n            kind: 'log',\n            timestamp: this.now(),\n            payload: {\n              source: 'run_complete_hook',\n              error: err instanceof Error ? err.message : String(err),\n            },\n          })\n        } catch {\n          // best-effort\n        }\n      }\n    }\n  }\n\n  // ── Generic span ───────────────────────────────────────────────────\n\n  async span<S extends Span = Span>(\n    init: {\n      kind: SpanKind\n      name: string\n      parentSpanId?: string\n      attributes?: Record<string, unknown>\n    } & Partial<Omit<S, 'spanId' | 'runId' | 'startedAt' | 'kind' | 'name'>>,\n  ): Promise<SpanHandle<S>> {\n    const spanId = this.id()\n    const parent = init.parentSpanId ?? this.stack[this.stack.length - 1]\n    const span = {\n      spanId,\n      parentSpanId: parent,\n      runId: this._runId,\n      startedAt: this.now(),\n      ...init,\n    } as unknown as S\n    await this.store.appendSpan(span)\n    this.stack.push(spanId)\n    return this.handle<S>(span)\n  }\n\n  private handle<S extends Span>(span: S): SpanHandle<S> {\n    return {\n      span,\n      end: async (patch?: Partial<S>) => {\n        const endedAt = this.now()\n        await this.store.updateSpan(span.spanId, {\n          endedAt,\n          status: 'ok',\n          ...patch,\n        } as Partial<Span>)\n        this.pop(span.spanId)\n      },\n      fail: async (error: string | Error, patch?: Partial<S>) => {\n        const endedAt = this.now()\n        const errStr = error instanceof Error ? error.message : error\n        await this.store.updateSpan(span.spanId, {\n          endedAt,\n          status: 'error',\n          error: errStr,\n          ...patch,\n        } as Partial<Span>)\n        this.pop(span.spanId)\n      },\n    }\n  }\n\n  private pop(spanId: string): void {\n    const idx = this.stack.lastIndexOf(spanId)\n    if (idx >= 0) this.stack.splice(idx, 1)\n  }\n\n  // ── Typed span conveniences ────────────────────────────────────────\n\n  llm(\n    init: Omit<LlmSpan, 'spanId' | 'runId' | 'kind' | 'startedAt'>,\n  ): Promise<SpanHandle<LlmSpan>> {\n    return this.span<LlmSpan>({ kind: 'llm', ...init })\n  }\n\n  tool(\n    init: Omit<ToolSpan, 'spanId' | 'runId' | 'kind' | 'startedAt'>,\n  ): Promise<SpanHandle<ToolSpan>> {\n    return this.span<ToolSpan>({ kind: 'tool', ...init })\n  }\n\n  retrieval(\n    init: Omit<RetrievalSpan, 'spanId' | 'runId' | 'kind' | 'startedAt'>,\n  ): Promise<SpanHandle<RetrievalSpan>> {\n    return this.span<RetrievalSpan>({ kind: 'retrieval', ...init })\n  }\n\n  async recordJudge(\n    verdict: Omit<JudgeSpan, 'spanId' | 'runId' | 'kind' | 'startedAt' | 'endedAt'>,\n  ): Promise<JudgeSpan> {\n    const spanId = this.id()\n    const now = this.now()\n    const full: JudgeSpan = {\n      spanId,\n      runId: this._runId,\n      kind: 'judge',\n      startedAt: now,\n      endedAt: now,\n      status: 'ok',\n      ...verdict,\n    }\n    await this.store.appendSpan(full)\n    return full\n  }\n\n  sandbox(\n    init: Omit<SandboxSpan, 'spanId' | 'runId' | 'kind' | 'startedAt'>,\n  ): Promise<SpanHandle<SandboxSpan>> {\n    return this.span<SandboxSpan>({ kind: 'sandbox', ...init })\n  }\n\n  // ── Events ─────────────────────────────────────────────────────────\n\n  async emit(event: {\n    kind: EventKind\n    spanId?: string\n    payload?: Record<string, unknown>\n  }): Promise<TraceEvent> {\n    const full: TraceEvent = {\n      eventId: this.id(),\n      runId: this._runId,\n      spanId: event.spanId ?? this.stack[this.stack.length - 1],\n      kind: event.kind,\n      timestamp: this.now(),\n      payload: event.payload ?? {},\n    }\n    await this.store.appendEvent(full)\n    return full\n  }\n\n  // ── Budget ledger ──────────────────────────────────────────────────\n\n  async recordBudget(\n    entry: Omit<BudgetLedgerEntry, 'runId' | 'timestamp'> & { timestamp?: number },\n  ): Promise<BudgetLedgerEntry> {\n    const full: BudgetLedgerEntry = {\n      runId: this._runId,\n      timestamp: entry.timestamp ?? this.now(),\n      dimension: entry.dimension,\n      limit: entry.limit,\n      consumed: entry.consumed,\n      remaining: entry.remaining,\n      breached: entry.breached,\n      spanId: entry.spanId ?? this.stack[this.stack.length - 1],\n    }\n    await this.store.appendBudgetEntry(full)\n    if (full.breached) {\n      await this.emit({\n        kind: 'budget_breach',\n        spanId: full.spanId,\n        payload: { dimension: full.dimension, limit: full.limit, consumed: full.consumed },\n      })\n    }\n    return full\n  }\n\n  // ── Artifacts ──────────────────────────────────────────────────────\n\n  async recordArtifact(artifact: Omit<Artifact, 'artifactId' | 'runId'>): Promise<Artifact> {\n    const full: Artifact = { artifactId: this.id(), runId: this._runId, ...artifact }\n    await this.store.appendArtifact(full)\n    return full\n  }\n\n  // ── Nested composition ─────────────────────────────────────────────\n\n  /**\n   * Runs `fn` inside a span; auto-ends on success, auto-fails on throw.\n   * Returns the fn's return value. Use this for the 95% case.\n   */\n  async within<T>(\n    init: Parameters<TraceEmitter['span']>[0],\n    fn: (handle: SpanHandle) => Promise<T>,\n  ): Promise<T> {\n    const handle = await this.span(init)\n    try {\n      const result = await fn(handle)\n      await handle.end()\n      return result\n    } catch (err) {\n      await handle.fail(err instanceof Error ? err : String(err))\n      throw err\n    }\n  }\n}\n\n// Helpers -------------------------------------------------------------\n\nfunction cryptoRandomId(): string {\n  if (typeof globalThis.crypto?.randomUUID === 'function') return globalThis.crypto.randomUUID()\n  return `${Date.now().toString(36)}-${Math.random().toString(36).slice(2, 10)}`\n}\n"],"mappings":";AAkEA,IAAa,eAAb,MAA0B;CACxB;CACA,QAA0B,CAAC;CAC3B;CACA;CACA;CACA;CACA;CAEA,YAAY,OAAmB,UAA+B,CAAC,GAAG;EAChE,KAAK,QAAQ;EACb,KAAK,MAAM,QAAQ,cAAc,KAAK,IAAI;EAC1C,KAAK,KAAK,QAAQ,aAAa,eAAe;EAC9C,KAAK,SAAS,QAAQ,SAAS,KAAK,GAAG;EACvC,KAAK,QAAQ,QAAQ,iBAAiB,CAAC;EACvC,KAAK,aAAa,QAAQ,cAAc;CAC1C;CAEA,IAAI,QAAgB;EAClB,OAAO,KAAK;CACd;CAEA,IAAI,aAAyB;EAC3B,OAAO,KAAK;CACd;;CAGA,mBAAmB,MAA6B;EAC9C,KAAK,MAAM,KAAK,IAAI;CACtB;;;;;;;;;;;CAcA,MAAM,SACJ,KACc;EACd,MAAM,aAAa,IAAI,cAAc,IAAI,SAAS,IAAI,MAAM,QAAQ;EACpE,MAAM,OAAY;GAChB,GAAG;GACH;GACA,OAAO,KAAK;GACZ,WAAW,KAAK,IAAI;GACpB,QAAQ;EACV;EACA,MAAM,KAAK,MAAM,UAAU,IAAI;EAC/B,OAAO;CACT;CAEA,MAAM,OAAO,SAAqC;EAChD,MAAM,SAAiC,SAAS,SAAS,QAAQ,WAAW;EAC5E,MAAM,KAAK,MAAM,UAAU,KAAK,QAAQ;GAAE,SAAS,KAAK,IAAI;GAAG;GAAQ;EAAQ,CAAC;EAChF,MAAM,KAAK,SAAS;GAAE,OAAO,KAAK;GAAQ,SAAS;GAAM,OAAO,KAAK;GAAO;GAAS;EAAO,CAAC;CAC/F;CAEA,MAAM,SAAS,QAA+B;EAC5C,MAAM,UAAU;GAAE,MAAM;GAAO,OAAO;EAAO;EAC7C,MAAM,KAAK,MAAM,UAAU,KAAK,QAAQ;GACtC,SAAS,KAAK,IAAI;GAClB,QAAQ;GACR;EACF,CAAC;EACD,MAAM,KAAK,SAAS;GAClB,OAAO,KAAK;GACZ,SAAS;GACT,OAAO,KAAK;GACZ;GACA,QAAQ;EACV,CAAC;CACH;CAEA,MAAc,SAAS,KAA4C;EACjE,KAAK,MAAM,QAAQ,KAAK,OACtB,IAAI;GACF,MAAM,KAAK,GAAG;EAChB,SAAS,KAAK;GACZ,IAAI,KAAK,eAAe,SAAS,MAAM;GACvC,IAAI;IACF,MAAM,KAAK,MAAM,YAAY;KAC3B,SAAS,KAAK,GAAG;KACjB,OAAO,KAAK;KACZ,MAAM;KACN,WAAW,KAAK,IAAI;KACpB,SAAS;MACP,QAAQ;MACR,OAAO,eAAe,QAAQ,IAAI,UAAU,OAAO,GAAG;KACxD;IACF,CAAC;GACH,QAAQ,CAER;EACF;CAEJ;CAIA,MAAM,KACJ,MAMwB;EACxB,MAAM,SAAS,KAAK,GAAG;EAEvB,MAAM,OAAO;GACX;GACA,cAHa,KAAK,gBAAgB,KAAK,MAAM,KAAK,MAAM,SAAS;GAIjE,OAAO,KAAK;GACZ,WAAW,KAAK,IAAI;GACpB,GAAG;EACL;EACA,MAAM,KAAK,MAAM,WAAW,IAAI;EAChC,KAAK,MAAM,KAAK,MAAM;EACtB,OAAO,KAAK,OAAU,IAAI;CAC5B;CAEA,OAA+B,MAAwB;EACrD,OAAO;GACL;GACA,KAAK,OAAO,UAAuB;IACjC,MAAM,UAAU,KAAK,IAAI;IACzB,MAAM,KAAK,MAAM,WAAW,KAAK,QAAQ;KACvC;KACA,QAAQ;KACR,GAAG;IACL,CAAkB;IAClB,KAAK,IAAI,KAAK,MAAM;GACtB;GACA,MAAM,OAAO,OAAuB,UAAuB;IACzD,MAAM,UAAU,KAAK,IAAI;IACzB,MAAM,SAAS,iBAAiB,QAAQ,MAAM,UAAU;IACxD,MAAM,KAAK,MAAM,WAAW,KAAK,QAAQ;KACvC;KACA,QAAQ;KACR,OAAO;KACP,GAAG;IACL,CAAkB;IAClB,KAAK,IAAI,KAAK,MAAM;GACtB;EACF;CACF;CAEA,IAAY,QAAsB;EAChC,MAAM,MAAM,KAAK,MAAM,YAAY,MAAM;EACzC,IAAI,OAAO,GAAG,KAAK,MAAM,OAAO,KAAK,CAAC;CACxC;CAIA,IACE,MAC8B;EAC9B,OAAO,KAAK,KAAc;GAAE,MAAM;GAAO,GAAG;EAAK,CAAC;CACpD;CAEA,KACE,MAC+B;EAC/B,OAAO,KAAK,KAAe;GAAE,MAAM;GAAQ,GAAG;EAAK,CAAC;CACtD;CAEA,UACE,MACoC;EACpC,OAAO,KAAK,KAAoB;GAAE,MAAM;GAAa,GAAG;EAAK,CAAC;CAChE;CAEA,MAAM,YACJ,SACoB;EACpB,MAAM,SAAS,KAAK,GAAG;EACvB,MAAM,MAAM,KAAK,IAAI;EACrB,MAAM,OAAkB;GACtB;GACA,OAAO,KAAK;GACZ,MAAM;GACN,WAAW;GACX,SAAS;GACT,QAAQ;GACR,GAAG;EACL;EACA,MAAM,KAAK,MAAM,WAAW,IAAI;EAChC,OAAO;CACT;CAEA,QACE,MACkC;EAClC,OAAO,KAAK,KAAkB;GAAE,MAAM;GAAW,GAAG;EAAK,CAAC;CAC5D;CAIA,MAAM,KAAK,OAIa;EACtB,MAAM,OAAmB;GACvB,SAAS,KAAK,GAAG;GACjB,OAAO,KAAK;GACZ,QAAQ,MAAM,UAAU,KAAK,MAAM,KAAK,MAAM,SAAS;GACvD,MAAM,MAAM;GACZ,WAAW,KAAK,IAAI;GACpB,SAAS,MAAM,WAAW,CAAC;EAC7B;EACA,MAAM,KAAK,MAAM,YAAY,IAAI;EACjC,OAAO;CACT;CAIA,MAAM,aACJ,OAC4B;EAC5B,MAAM,OAA0B;GAC9B,OAAO,KAAK;GACZ,WAAW,MAAM,aAAa,KAAK,IAAI;GACvC,WAAW,MAAM;GACjB,OAAO,MAAM;GACb,UAAU,MAAM;GAChB,WAAW,MAAM;GACjB,UAAU,MAAM;GAChB,QAAQ,MAAM,UAAU,KAAK,MAAM,KAAK,MAAM,SAAS;EACzD;EACA,MAAM,KAAK,MAAM,kBAAkB,IAAI;EACvC,IAAI,KAAK,UACP,MAAM,KAAK,KAAK;GACd,MAAM;GACN,QAAQ,KAAK;GACb,SAAS;IAAE,WAAW,KAAK;IAAW,OAAO,KAAK;IAAO,UAAU,KAAK;GAAS;EACnF,CAAC;EAEH,OAAO;CACT;CAIA,MAAM,eAAe,UAAqE;EACxF,MAAM,OAAiB;GAAE,YAAY,KAAK,GAAG;GAAG,OAAO,KAAK;GAAQ,GAAG;EAAS;EAChF,MAAM,KAAK,MAAM,eAAe,IAAI;EACpC,OAAO;CACT;;;;;CAQA,MAAM,OACJ,MACA,IACY;EACZ,MAAM,SAAS,MAAM,KAAK,KAAK,IAAI;EACnC,IAAI;GACF,MAAM,SAAS,MAAM,GAAG,MAAM;GAC9B,MAAM,OAAO,IAAI;GACjB,OAAO;EACT,SAAS,KAAK;GACZ,MAAM,OAAO,KAAK,eAAe,QAAQ,MAAM,OAAO,GAAG,CAAC;GAC1D,MAAM;EACR;CACF;AACF;AAIA,SAAS,iBAAyB;CAChC,IAAI,OAAO,WAAW,QAAQ,eAAe,YAAY,OAAO,WAAW,OAAO,WAAW;CAC7F,OAAO,GAAG,KAAK,IAAI,CAAC,CAAC,SAAS,EAAE,EAAE,GAAG,KAAK,OAAO,CAAC,CAAC,SAAS,EAAE,CAAC,CAAC,MAAM,GAAG,EAAE;AAC7E"}

package/dist/{emitter-DGQGoLyj.d.ts → emitter-D_jYSGRd.d.ts}

@@ -1,5 +1,5 @@
-import { C as TraceEvent, S as ToolSpan, _ as Span, c as JudgeSpan, d as RetrievalSpan, f as Run, g as SandboxSpan, i as EventKind, l as LlmSpan, m as RunOutcome, n as BudgetLedgerEntry, t as Artifact, u as Message, y as SpanKind } from "./schema-BtVldJ3T.js";
-import { s as TraceStore } from "./store-CT9YIIve.js";
+import { C as TraceEvent, S as ToolSpan, _ as Span, c as JudgeSpan, d as RetrievalSpan, f as Run, g as SandboxSpan, i as EventKind, l as LlmSpan, m as RunOutcome, n as BudgetLedgerEntry, t as Artifact, y as SpanKind } from "./schema-Bjgdsn73.js";
+import { s as TraceStore } from "./store-B06JdC56.js";
 //#region src/trace/emitter.d.ts
 interface SpanHandle<S extends Span = Span> {
   span: S;
@@ -92,22 +92,6 @@ declare class TraceEmitter {
    */
   within<T>(init: Parameters<TraceEmitter['span']>[0], fn: (handle: SpanHandle) => Promise<T>): Promise<T>;
 }
-/** Helper to build an LLM span handle args object from a provider-shaped response. */
-declare function llmSpanFromProvider(args: {
-  name?: string;
-  model: string;
-  messages: Message[];
-  output: string;
-  usage?: {
-    inputTokens?: number;
-    outputTokens?: number;
-    cachedTokens?: number;
-    cacheWriteTokens?: number;
-    reasoningTokens?: number;
-  };
-  costUsd?: number;
-  finishReason?: string;
-}): Omit<LlmSpan, 'spanId' | 'runId' | 'kind' | 'startedAt'>;
 //#endregion
-export { TraceEmitterOptions as a, TraceEmitter as i, RunCompleteHookContext as n, llmSpanFromProvider as o, SpanHandle as r, RunCompleteHook as t };
-//# sourceMappingURL=emitter-DGQGoLyj.d.ts.map
+export { TraceEmitterOptions as a, TraceEmitter as i, RunCompleteHookContext as n, SpanHandle as r, RunCompleteHook as t };
+//# sourceMappingURL=emitter-D_jYSGRd.d.ts.map

package/dist/{emitter-DGQGoLyj.d.ts.map → emitter-D_jYSGRd.d.ts.map}

@@ -1 +1 @@
-{"version":3,"file":"emitter-DGQGoLyj.d.ts","names":[],"sources":["../src/trace/emitter.ts"],"mappings":";;;UA6BiB,WAAW,UAAU,OAAO;EAC3C,MAAM;EACN,IAAI,QAAQ,QAAQ,KAAK;EACzB,KAAK,gBAAgB,OAAO,QAAQ,QAAQ,KAAK;;UAGlC;EACf;EACA,SAAS;EACT,OAAO;;EAEP,UAAU;;EAEV;;KAGU,mBAAmB,KAAK,2BAA2B;UAE9C;EACf;;EAEA;;EAEA;;;;;;;;;;EAUA,gBAAgB;;EAEhB;;cAGW;UACH;UACA;UACA;UACA;UACA;UACA;UACA;EAER,YAAY,OAAO,YAAY,UAAS;MASpC;MAIA,cAAc;;EAKlB,mBAAmB,MAAM;;;;;;;;;;;EAgBnB,SACJ,KAAK,KAAK;IAA0D;MACnE,QAAQ;EAaL,OAAO,UAAU,aAAa;EAM9B,SAAS,iBAAiB;UAgBlB;EA0BR,KAAK,UAAU,OAAO,MAC1B;IACE,MAAM;IACN;IACA;IACA,aAAa;MACX,QAAQ,KAAK,0DAChB,QAAQ,WAAW;UAed;UA0BA;EAOR,IACE,MAAM,KAAK,sDACV,QAAQ,WAAW;EAItB,KACE,MAAM,KAAK,uDACV,QAAQ,WAAW;EAItB,UACE,MAAM,KAAK,4DACV,QAAQ,WAAW;EAIhB,YACJ,SAAS,KAAK,oEACb,QAAQ;EAgBX,QACE,MAAM,KAAK,0DACV,QAAQ,WAAW;EAMhB,KAAK;IACT,MAAM;IACN;IACA,UAAU;MACR,QAAQ;EAeN,aACJ,OAAO,KAAK;IAA8C;MACzD,QAAQ;EAwBL,eAAe,UAAU,KAAK,oCAAoC,QAAQ;;;;;EAY1E,OAAO,GACX,MAAM,WAAW,0BACjB,KAAK,QAAQ,eAAe,QAAQ,KACnC,QAAQ;;;iBAqBG,oBAAoB;EAClC;EACA;EACA,UAAU;EACV;EACA;IACE;IACA;IACA;IACA;IACA;;EAEF;EACA;IACE,KAAK"}
+{"version":3,"file":"emitter-D_jYSGRd.d.ts","names":[],"sources":["../src/trace/emitter.ts"],"mappings":";;;UA4BiB,WAAW,UAAU,OAAO;EAC3C,MAAM;EACN,IAAI,QAAQ,QAAQ,KAAK;EACzB,KAAK,gBAAgB,OAAO,QAAQ,QAAQ,KAAK;;UAGlC;EACf;EACA,SAAS;EACT,OAAO;;EAEP,UAAU;;EAEV;;KAGU,mBAAmB,KAAK,2BAA2B;UAE9C;EACf;;EAEA;;EAEA;;;;;;;;;;EAUA,gBAAgB;;EAEhB;;cAGW;UACH;UACA;UACA;UACA;UACA;UACA;UACA;EAER,YAAY,OAAO,YAAY,UAAS;MASpC;MAIA,cAAc;;EAKlB,mBAAmB,MAAM;;;;;;;;;;;EAgBnB,SACJ,KAAK,KAAK;IAA0D;MACnE,QAAQ;EAaL,OAAO,UAAU,aAAa;EAM9B,SAAS,iBAAiB;UAgBlB;EA0BR,KAAK,UAAU,OAAO,MAC1B;IACE,MAAM;IACN;IACA;IACA,aAAa;MACX,QAAQ,KAAK,0DAChB,QAAQ,WAAW;UAed;UA0BA;EAOR,IACE,MAAM,KAAK,sDACV,QAAQ,WAAW;EAItB,KACE,MAAM,KAAK,uDACV,QAAQ,WAAW;EAItB,UACE,MAAM,KAAK,4DACV,QAAQ,WAAW;EAIhB,YACJ,SAAS,KAAK,oEACb,QAAQ;EAgBX,QACE,MAAM,KAAK,0DACV,QAAQ,WAAW;EAMhB,KAAK;IACT,MAAM;IACN;IACA,UAAU;MACR,QAAQ;EAeN,aACJ,OAAO,KAAK;IAA8C;MACzD,QAAQ;EAwBL,eAAe,UAAU,KAAK,oCAAoC,QAAQ;;;;;EAY1E,OAAO,GACX,MAAM,WAAW,0BACjB,KAAK,QAAQ,eAAe,QAAQ,KACnC,QAAQ"}

package/dist/{engine-BLzhNzoY.d.ts → engine-Cu5qD5Fc.d.ts}

@@ -1,11 +1,11 @@
 import { c as CostLedgerHandle } from "./cost-ledger-DbQdN3nO.js";
-import { a as RunRecord } from "./run-record-DVV82Gwh.js";
-import { S as BoundedTraceAnalysisStoreOptions, p as EvidenceRef, w as TraceAnalysisStore } from "./types-B2NsbrNy.js";
-import { n as AnalystRegistry } from "./registry-BQwrSYpC.js";
-import "./types-DdFNuyxQ.js";
-import { r as DatasetScenario } from "./dataset-CJjKqQfA.js";
-import "./summary-report-B__Y5ub3.js";
-import { o as InsightReport, r as ExecutionInsight, t as CostProvenanceSummary } from "./insight-report-BeT8KCgI.js";
+import { a as RunRecord } from "./run-record-VVy4T9OW.js";
+import { S as BoundedTraceAnalysisStoreOptions, p as EvidenceRef, w as TraceAnalysisStore } from "./types-D9ssmxKL.js";
+import { n as AnalystRegistry } from "./registry-8You7OK1.js";
+import "./types-D4s7Z6nq.js";
+import { r as DatasetScenario } from "./dataset-DQqhOCPt.js";
+import "./summary-report-CC07PhEL.js";
+import { o as InsightReport, r as ExecutionInsight, t as CostProvenanceSummary } from "./insight-report-DRe8LB6d.js";
 import { z } from "zod";
 //#region src/contract/analyze-runs.d.ts
 interface AnalyzeRunsOptions {
@@ -102,7 +102,6 @@ declare function traceAnalystFunctionGroup(options: BuildTraceAnalysisToolsOptio
 };
 //#endregion
 //#region src/analyst/finding-signature.d.ts
-declare const ANALYST_SEVERITIES: readonly ['critical', 'high', 'medium', 'low', 'info'];
 declare const RawAnalystEvidenceSchema: z.ZodObject<{
   uri: z.ZodString;
   excerpt: z.ZodOptional<z.ZodString>;
@@ -149,7 +148,6 @@ interface TraceAnalystLimits {
   /** Maximum controller output retained between iterations. */
   maxOutputChars: number;
 }
-declare const DEFAULT_TRACE_ANALYST_LIMITS: Readonly<TraceAnalystLimits>;
 interface TraceAnalysisEngineRequest {
   analystId: string;
   question: string;
@@ -213,5 +211,5 @@ interface TraceAnalysisEngine {
 }
 declare function resolveTraceAnalystLimits(limits: Partial<TraceAnalystLimits> | undefined): TraceAnalystLimits;
 //#endregion
-export { analyzeRuns as C, SummarizeExecutionOptions as S, TraceAnalysisToolDescriptor as _, TraceAnalystLimits as a, AnalyzeRunsOptions as b, RAW_FINDING_SCHEMA_PROMPT as c, RawAnalystFinding as d, RawAnalystFindingSchema as f, TRACE_ANALYST_TOOL_NAMESPACE as g, BuildTraceAnalysisToolsOptions as h, TraceAnalysisEngineResult as i, RawAnalystEvidence as l, parseRawFinding as m, TraceAnalysisEngine as n, resolveTraceAnalystLimits as o, evidenceRefsFromRawFinding as p, TraceAnalysisEngineRequest as r, ANALYST_SEVERITIES as s, DEFAULT_TRACE_ANALYST_LIMITS as t, RawAnalystEvidenceSchema as u, buildTraceAnalysisToolDescriptors as v, summarizeExecution as w, ExecutionReport as x, traceAnalystFunctionGroup as y };
-//# sourceMappingURL=engine-BLzhNzoY.d.ts.map
+export { AnalyzeRunsOptions as _, resolveTraceAnalystLimits as a, analyzeRuns as b, RawAnalystFinding as c, parseRawFinding as d, BuildTraceAnalysisToolsOptions as f, traceAnalystFunctionGroup as g, buildTraceAnalysisToolDescriptors as h, TraceAnalystLimits as i, RawAnalystFindingSchema as l, TraceAnalysisToolDescriptor as m, TraceAnalysisEngineRequest as n, RAW_FINDING_SCHEMA_PROMPT as o, TRACE_ANALYST_TOOL_NAMESPACE as p, TraceAnalysisEngineResult as r, RawAnalystEvidence as s, TraceAnalysisEngine as t, evidenceRefsFromRawFinding as u, ExecutionReport as v, summarizeExecution as x, SummarizeExecutionOptions as y };
+//# sourceMappingURL=engine-Cu5qD5Fc.d.ts.map

package/dist/{engine-BLzhNzoY.d.ts.map → engine-Cu5qD5Fc.d.ts.map}

@@ -1 +1 @@
-{"version":3,"file":"engine-BLzhNzoY.d.ts","names":[],"sources":["../src/contract/analyze-runs.ts","../src/eval-tools.ts","../src/trace-analyst/tools.ts","../src/analyst/finding-signature.ts","../src/analyst/engine.ts"],"mappings":";;;;;;;;;;UAoEiB;;EAEf,MAAM;;;EAGN;;;;;EAKA;EACA;;;EAGA,kBAAkB;;;EAGlB,UAAU;;;;EAIV;IACE;IACA,cAAc;;;;;EAKhB,cAAc;IAAQ;IAAe;IAAe;;;EAEpD;;;;EAIA;;;;;;;;;EASA,eAAe;;;EAGf;;UAGe;EACf,MAAM;EACN;;UAGe;EACf,WAAW;EACX,gBAAgB;;;iBAIF,mBAAmB,MAAM,4BAA4B;iBAgB/C,YAAY,MAAM,qBAAqB,QAAQ;;;;UCpHpD;EACf;EACA;EACA,YAAY;EACZ,UAAU,eAAe;IAAQ,SAAS;QAAkB;;;;cClBjD;UAEI,uCAAuC;EACtD,OAAO;;UAGQ,oCAAoC;EACnD,kBAAkB;;;iBAIJ,kCACd,SAAS,iCACR;iBAkHa,0BAA0B,SAAS;EACjD;EACA;EACA;EACA;EACA,WAAW;;;;cCtIA;cAEA,0BAAwB,EAAA;;;GAK1B,EAAA,KAAA;KAEC,qBAAqB,EAAE,aAAa;cAiBnC,yBAAuB,EAAA;;;;;;;;;;;;;;;;;GAKzB,EAAA,KAAA;KAEC,oBAAoB,EAAE,aAAa;;;;;;cAOlC;;iBAYG,2BAA2B,SAAS,oBAAoB;iBAQxD,gBACd,cACA,OAAO,aAAa,SAAS,mCAC5B;;;;UCtEc;;EAEf;;EAEA;;EAEA;;EAEA;;cAGW,8BAA8B,SAAS;UAOnC;EACf;EACA;EACA;EACA,gBAAgB;;;;;;;;;;;EAWhB,aAAa,SAAS;EACtB,QAAQ;EACR,YAAY;EACZ;EACA,WAAW;EACX,SAAS;EACT,OAAO,iBAAiB,SAAS;;;UAIlB;EACf;EACA,UAAU;;EAEV;;;;;EAKA;;EAEA;EACA,SAAS;;;;;;;;;UAUM;WACN;WACA;WACA;;WAEA;;;;;;;;WAQA,iBAAiB,SAAS;EACnC,QAAQ,SAAS,6BAA6B,QAAQ;;iBAGxC,0BACd,QAAQ,QAAQ,kCACf"}
+{"version":3,"file":"engine-Cu5qD5Fc.d.ts","names":[],"sources":["../src/contract/analyze-runs.ts","../src/eval-tools.ts","../src/trace-analyst/tools.ts","../src/analyst/finding-signature.ts","../src/analyst/engine.ts"],"mappings":";;;;;;;;;;UAoEiB;;EAEf,MAAM;;;EAGN;;;;;EAKA;EACA;;;EAGA,kBAAkB;;;EAGlB,UAAU;;;;EAIV;IACE;IACA,cAAc;;;;;EAKhB,cAAc;IAAQ;IAAe;IAAe;;;EAEpD;;;;EAIA;;;;;;;;;EASA,eAAe;;;EAGf;;UAGe;EACf,MAAM;EACN;;UAGe;EACf,WAAW;EACX,gBAAgB;;;iBAIF,mBAAmB,MAAM,4BAA4B;iBAgB/C,YAAY,MAAM,qBAAqB,QAAQ;;;;UCpHpD;EACf;EACA;EACA,YAAY;EACZ,UAAU,eAAe;IAAQ,SAAS;QAAkB;;;;cClBjD;UAEI,uCAAuC;EACtD,OAAO;;UAGQ,oCAAoC;EACnD,kBAAkB;;;iBAIJ,kCACd,SAAS,iCACR;iBAkHa,0BAA0B,SAAS;EACjD;EACA;EACA;EACA;EACA,WAAW;;;;cCpIP,0BAAwB,EAAA;;;GAKnB,EAAA,KAAA;KAEC,qBAAqB,EAAE,aAAa;cAiBnC,yBAAuB,EAAA;;;;;;;;;;;;;;;;;GAKzB,EAAA,KAAA;KAEC,oBAAoB,EAAE,aAAa;;;;;;cAOlC;;iBAYG,2BAA2B,SAAS,oBAAoB;iBAQxD,gBACd,cACA,OAAO,aAAa,SAAS,mCAC5B;;;;UCtEc;;EAEf;;EAEA;;EAEA;;EAEA;;UAUe;EACf;EACA;EACA;EACA,gBAAgB;;;;;;;;;;;EAWhB,aAAa,SAAS;EACtB,QAAQ;EACR,YAAY;EACZ;EACA,WAAW;EACX,SAAS;EACT,OAAO,iBAAiB,SAAS;;;UAIlB;EACf;EACA,UAAU;;EAEV;;;;;EAKA;;EAEA;EACA,SAAS;;;;;;;;;UAUM;WACN;WACA;WACA;;WAEA;;;;;;;;WAQA,iBAAiB,SAAS;EACnC,QAAQ,SAAS,6BAA6B,QAAQ;;iBAGxC,0BACd,QAAQ,QAAQ,kCACf"}

package/dist/{eval-campaign-C4jmuM-b.js → eval-campaign-BsXWL2-2.js}

@@ -1,11 +1,10 @@
-import { r as hashJson, t as canonicalize } from "./pre-registration-DakwTRXk.js";
+import { n as hashJson } from "./pre-registration-KN9jkh58.js";
 import { buildAgentProfileCell, verifyAgentProfileCell } from "./profile-cell.js";
-import { c as validateRunRecord } from "./run-record-D2lDdSAz.js";
-import { i as researchReport } from "./summary-report-Blysd6Z2.js";
-import { t as TraceEmitter } from "./emitter-CPBAhxum.js";
+import { c as validateRunRecord } from "./run-record-BC0ebuRP.js";
+import { i as researchReport } from "./summary-report-BI5hUtvK.js";
+import { t as TraceEmitter } from "./emitter-BpYFQPj4.js";
 import { t as FileSystemRawProviderSink } from "./raw-provider-sink-BQd7mzyT.js";
 import { n as assertRunCaptured, t as RunIntegrityError } from "./integrity-Cy9WHAtb.js";
-import { a as assertLlmRoute } from "./llm-client-Bg32RW0j.js";
 //#region src/eval-campaign.ts
 /**
 * EvalCampaign — opinionated matrix runner that wires the four
@@ -21,12 +20,10 @@ import { a as assertLlmRoute } from "./llm-client-Bg32RW0j.js";
 * `EvalCampaign` is the structural fix — consumers don't wire the
 * integrity surface themselves; the campaign owns it. Specifically:
 *
-*   - calls `assertLlmRoute` once at preflight before any work runs
 *   - constructs a per-run `TraceStore` and `RawProviderSink` via factories
-*   - constructs the `TraceEmitter` with `onRunComplete: [analyst hook]`
-*   - hands the runner an `LlmClientOptions` pre-wired with the sink and
-*     trace context — the runner can't accidentally call an LLM without
-*     capturing the raw HTTP envelope
+*   - builds the run's `ChatClient` through `chatFactory`, handing it the
+*     run's raw sink and trace context — a transport built any other way has
+*     no raw HTTP envelope, and `assertRunCaptured` says so
 *   - calls `assertRunCaptured` after every `endRun` and routes failures
 *     through a configurable policy (`throw` / `mark_failed` / `log`)
 *   - assembles per-run `RunRecord`s and runs `researchReport` at the end
@@ -44,19 +41,14 @@ import { a as assertLlmRoute } from "./llm-client-Bg32RW0j.js";
 *   - Distributed/cluster execution (concurrency is local async)
 *   - Adaptive sampling / sequential interim looks
 *   - Resume from partial state across crashes
-*   - LLM-call retry beyond what `LlmClient` already does
+*   - LLM-call retry beyond what the caller's transport already does
 */
 const DEFAULT_INTEGRITY = {
 	llmSpansMin: 1,
 	requireRawCoverageOfLlmSpans: true,
 	requireOutcome: true
 };
-const DEFAULT_ROUTE = {
-	requireExplicitBaseUrl: true,
-	requireAuth: true
-};
 async function runEvalCampaign(opts) {
-	assertLlmRoute(opts.llmOpts, opts.routeRequirements ?? DEFAULT_ROUTE);
 	if (opts.variants.length === 0) throw new Error("runEvalCampaign: variants must be non-empty.");
 	if (opts.scenarios.length === 0) throw new Error("runEvalCampaign: scenarios must be non-empty.");
 	const variantIds = /* @__PURE__ */ new Set();
@@ -84,21 +76,19 @@ async function runEvalCampaign(opts) {
 	};
 	const onIntegrityFailure = opts.onIntegrityFailure ?? "mark_failed";
 	const now = opts.now ?? (() => Date.now());
-	const baseUrl = (opts.llmOpts.baseUrl ?? "").replace(/\/+$/, "");
-	const provider = opts.llmOpts.provider ?? null;
+	const executionRef = opts.executionRef ?? null;
 	const preregistrationHash = opts.preregistrationHash ?? null;
 	const rawSinkFactory = opts.rawSinkFactory ?? defaultRawSinkFactory(opts.workDir);
-	const campaignFingerprint = await hashJson(canonicalize({
+	const campaignFingerprint = await hashJson({
 		campaignId: opts.campaignId,
 		variants: opts.variants.map((v) => v.id).sort(),
 		scenarios: opts.scenarios.map((s) => s.scenarioId).sort(),
 		seeds: [...seeds].sort((a, b) => a - b),
 		splitTag,
 		comparator: opts.report?.comparator ?? null,
-		baseUrl,
-		provider,
+		executionRef,
 		preregistrationHash
-	}));
+	});
 	const cells = [];
 	for (const variant of opts.variants) for (const scenario of opts.scenarios) for (const seed of seeds) cells.push({
 		variant,
@@ -157,11 +147,10 @@ async function runEvalCampaign(opts) {
 			onRunComplete: opts.onRunComplete
 		});
 		openRuns.set(runId, emitter);
-		const llmOpts = {
-			...opts.llmOpts,
+		const chat = opts.chatFactory({
 			rawSink,
-			traceContext: { runId }
-		};
+			runId
+		});
 		const ctx = {
 			runId,
 			experimentId: opts.campaignId,
@@ -174,7 +163,7 @@ async function runEvalCampaign(opts) {
 			emitter,
 			store,
 			rawSink,
-			llmOpts
+			chat
 		};
 		try {
 			const wallStart = now();
@@ -347,4 +336,4 @@ function defaultRunId(params) {
 //#endregion
 export { runEvalCampaign as t };
 
-//# sourceMappingURL=eval-campaign-C4jmuM-b.js.map
+//# sourceMappingURL=eval-campaign-BsXWL2-2.js.map

package/dist/eval-campaign-BsXWL2-2.js.map

@@ -0,0 +1 @@
+{"version":3,"file":"eval-campaign-BsXWL2-2.js","names":[],"sources":["../src/eval-campaign.ts"],"sourcesContent":["/**\n * EvalCampaign — opinionated matrix runner that wires the four\n * capture-integrity directives by construction.\n *\n * The canonical benchmark shape — matrix runner → for each\n * (variant, scenario, seed) → start a TraceEmitter → call LLMs → end the\n * run → analyze — has a bug class at the integration boundary: raw\n * events not captured, route silently wrong, integrity not asserted,\n * analyst never run. The directives in `SKILL.md § Capture integrity`\n * are the mitigations.\n *\n * `EvalCampaign` is the structural fix — consumers don't wire the\n * integrity surface themselves; the campaign owns it. Specifically:\n *\n *   - constructs a per-run `TraceStore` and `RawProviderSink` via factories\n *   - builds the run's `ChatClient` through `chatFactory`, handing it the\n *     run's raw sink and trace context — a transport built any other way has\n *     no raw HTTP envelope, and `assertRunCaptured` says so\n *   - calls `assertRunCaptured` after every `endRun` and routes failures\n *     through a configurable policy (`throw` / `mark_failed` / `log`)\n *   - assembles per-run `RunRecord`s and runs `researchReport` at the end\n *     so the campaign artifact is launch-decision-grade by default\n *   - embeds the campaign fingerprint (a SHA-256 over the canonicalised\n *     run set) and optional `preregistrationHash` in the report\n *\n * The runner contract is intentionally narrow: produce a `CampaignRunOutcome`\n * given a fully-wired `CampaignRunContext`. Everything orchestration-shaped\n * lives in the campaign. This is the inversion-of-control point — consumers\n * stop writing matrix runners and start writing scenario-runners.\n *\n * Out of scope for v1 (tracked in `docs/research-report-methodology.md`):\n *\n *   - Distributed/cluster execution (concurrency is local async)\n *   - Adaptive sampling / sequential interim looks\n *   - Resume from partial state across crashes\n *   - LLM-call retry beyond what the caller's transport already does\n */\n\nimport {\n  type AgentProfileCell,\n  type AgentProfileCellInput,\n  buildAgentProfileCell,\n  verifyAgentProfileCell,\n} from './agent-profile-cell'\nimport type { ChatClient } from './analyst/chat-client'\nimport { hashJson } from './pre-registration'\nimport type {\n  JudgeScoresRecord,\n  RunCostProvenance,\n  RunJudgeMetadata,\n  RunOutcome,\n  RunRecord,\n  RunSplitTag,\n  RunTaskFailure,\n  RunTokenUsage,\n} from './run-record'\nimport { validateRunRecord } from './run-record'\nimport { type ResearchReport, type ResearchReportOptions, researchReport } from './summary-report'\nimport type { RunCompleteHook } from './trace/emitter'\nimport { TraceEmitter } from './trace/emitter'\nimport {\n  assertRunCaptured,\n  RunIntegrityError,\n  type RunIntegrityExpectations,\n  type RunIntegrityReport,\n} from './trace/integrity'\nimport { FileSystemRawProviderSink, type RawProviderSink } from './trace/raw-provider-sink'\nimport type { TraceStore } from './trace/store'\n\n// ── Public types ─────────────────────────────────────────────────────────\n\nexport interface CampaignVariant<V> {\n  id: string\n  payload: V\n}\n\nexport interface CampaignScenario {\n  scenarioId: string\n  /** Free-form metadata propagated to runs and reports. */\n  tags?: Record<string, string>\n}\n\nexport interface CampaignRunContext<V> {\n  /** Stable run id. The campaign generates this; the runner does not. */\n  runId: string\n  /** Logical experiment id (campaignId by default; overridable per-run via opts). */\n  experimentId: string\n  variant: V\n  variantId: string\n  scenarioId: string\n  scenarioTags: Record<string, string>\n  seed: number\n  splitTag: RunSplitTag\n  /**\n   * The TraceEmitter for this run, with `onRunComplete` hooks pre-wired\n   * (analyst auto-execution if configured, plus integrity check). The\n   * runner MUST call `emitter.startRun` before doing any work and either\n   * `emitter.endRun` or `emitter.abortRun` before returning.\n   */\n  emitter: TraceEmitter\n  store: TraceStore\n  rawSink: RawProviderSink\n  /**\n   * The run's model transport, built by `chatFactory` with this run's\n   * `rawSink` and `runId` already bound.\n   */\n  chat: ChatClient\n}\n\n/** What the campaign binds into the run's transport. */\nexport interface CampaignChatWiring {\n  /**\n   * Raw provider sink for this run. Bind it into the transport: the campaign's\n   * integrity check requires every LLM span to carry a matching raw request\n   * event, so a transport built without it fails `assertRunCaptured`.\n   */\n  rawSink: RawProviderSink\n  runId: string\n}\n\ninterface CampaignRunOutcomeFields {\n  /** Did the run pass? Mirrors `RunOutcome.pass` semantics. */\n  pass: boolean\n  /** Score for the run on its split. Maps to `searchScore` or `holdoutScore`. */\n  score: number\n  /** Cost in USD, or null when the runner could not capture it. */\n  costUsd: number | null\n  /** Source of the cost amount. */\n  costProvenance: RunCostProvenance\n  tokenUsage: RunTokenUsage\n  /** Snapshot model id (e.g. `claude-sonnet-4-6@2025-04-15`). */\n  model: string\n  /** sha256 of the effective prompt sent to the model. */\n  promptHash: string\n  /** sha256 of the effective config (model, temperature, tools, judges, splits). */\n  configHash: string\n  /** Optional extra numeric metrics to land in `outcome.raw`. */\n  raw?: Record<string, number>\n  /** Optional judge metadata when a judge was used. */\n  judgeMetadata?: RunJudgeMetadata\n  /**\n   * Optional per-judge / per-dim breakdown for ensemble-judged runs.\n   * Propagated to `outcome.judgeScores` on the resulting `RunRecord`.\n   * Single-judge or scalar-only runs leave this unset.\n   */\n  judgeScores?: JudgeScoresRecord\n  /**\n   * Agent profile cell observed by the runner. When supplied, it overrides\n   * `EvalCampaignOptions.agentProfile` for this run and must match the\n   * outcome's `model` and `promptHash`.\n   */\n  agentProfile?: AgentProfileCell | AgentProfileCellInput\n}\n\n/** Campaign result with the same task-failure invariant as `RunRecord`. */\nexport type CampaignRunOutcome = CampaignRunOutcomeFields & RunTaskFailure\n\nexport type CampaignRunner<V> = (ctx: CampaignRunContext<V>) => Promise<CampaignRunOutcome>\n\nexport type CampaignIntegrityPolicy = 'throw' | 'mark_failed' | 'log'\n\nexport interface EvalCampaignOptions<V> {\n  /**\n   * Stable id for the campaign. Used as the default `experimentId` on\n   * every run, and folded into the campaign fingerprint.\n   */\n  campaignId: string\n  variants: CampaignVariant<V>[]\n  scenarios: CampaignScenario[]\n  /** Default `[0, 1, 2]`. */\n  seeds?: number[]\n  /** Default `'holdout'` — the split that anchors a launch decision. */\n  splitTag?: RunSplitTag\n  /** Git SHA the campaign is run against. Mandatory; `RunRecord` rejects unset. */\n  commitSha: string\n  /**\n   * Build the model transport for one run. agent-eval executes no paid model:\n   * the caller owns the transport and the credential never enters this\n   * package. The campaign calls this once per run and passes the run's raw\n   * provider sink and `runId`, so a transport that binds them captures the\n   * raw HTTP envelope `assertRunCaptured` checks for.\n   */\n  chatFactory: (wiring: CampaignChatWiring) => ChatClient\n  /**\n   * Caller-declared identity of the execution route, folded into the campaign\n   * fingerprint so two campaigns run against different endpoints do not share\n   * one identity. agent-eval no longer knows the endpoint; the owner of\n   * execution names it.\n   */\n  executionRef?: string\n  /**\n   * Per-run TraceStore factory. Common shape: a fresh store per run keyed\n   * on `runId`. Implementations that share a store across the campaign\n   * are valid — the campaign only writes through `emitter`.\n   */\n  storeFactory: (params: CampaignFactoryParams) => TraceStore\n  /**\n   * Per-run RawProviderSink factory. Defaults to `FileSystemRawProviderSink`\n   * rooted at `${workDir}/raw-events/${runId}` if `workDir` is supplied;\n   * otherwise required. Forensic capture is non-negotiable in a campaign\n   * run — pass `NoopRawProviderSink` explicitly if you want to opt out.\n   */\n  rawSinkFactory?: (params: CampaignFactoryParams) => RawProviderSink\n  /**\n   * Filesystem root for default `rawSinkFactory`. Ignored if\n   * `rawSinkFactory` is supplied.\n   */\n  workDir?: string\n  /**\n   * Extra `onRunComplete` hooks the campaign appends (after its own\n   * integrity-check hook). Pass `traceAnalystOnRunComplete(...)` here.\n   */\n  onRunComplete?: RunCompleteHook[]\n  /**\n   * Per-run integrity expectations. Defaults to:\n   *   `{ llmSpansMin: 1, requireRawCoverageOfLlmSpans: true, requireOutcome: true }`.\n   * Override (e.g. `{ llmSpansMin: 0 }`) for runs that don't call LLMs.\n   */\n  integrity?: RunIntegrityExpectations\n  /** Behaviour when integrity fails. Default `'mark_failed'`. */\n  onIntegrityFailure?: CampaignIntegrityPolicy\n  /**\n   * Per-run runner. Receives a fully-wired context; produces an outcome\n   * the campaign converts into a `RunRecord`.\n   */\n  runner: CampaignRunner<V>\n  /**\n   * If set, the campaign computes `researchReport` at the end. `comparator`\n   * is a `variantId`. Other fields are forwarded verbatim.\n   */\n  report?: { comparator?: string } & Omit<\n    ResearchReportOptions,\n    'comparator' | 'preregistrationHash' | 'generatedAt'\n  >\n  /**\n   * Hash of a signed `HypothesisManifest` (see `pre-registration.ts`).\n   * Embedded in the campaign fingerprint and the research report.\n   */\n  preregistrationHash?: string\n  /** Local concurrency. Default `1` (sequential). */\n  concurrency?: number\n  /**\n   * Override the time source. Tests pass a mock to make wallMs deterministic.\n   */\n  now?: () => number\n  /** Override the runId generator. Tests pin this. */\n  runId?: (params: CampaignFactoryParams) => string\n  /**\n   * Agent profile cell for campaign runs. Static profiles can pass an object;\n   * routers or variant-specific harnesses can pass a factory. The campaign\n   * stamps the built cell onto every `RunRecord` and rejects profile/model or\n   * profile/prompt contradictions.\n   */\n  agentProfile?:\n    | AgentProfileCell\n    | AgentProfileCellInput\n    | ((\n        params: CampaignFactoryParams & {\n          variant: V\n          scenarioTags: Record<string, string>\n        },\n      ) =>\n        | AgentProfileCell\n        | AgentProfileCellInput\n        | Promise<AgentProfileCell | AgentProfileCellInput>)\n}\n\nexport interface CampaignFactoryParams {\n  campaignId: string\n  runId: string\n  variantId: string\n  scenarioId: string\n  seed: number\n}\n\nexport interface FailedRun {\n  runId: string\n  variantId: string\n  scenarioId: string\n  seed: number\n  reason: string\n  error?: string\n}\n\nexport interface EvalCampaignResult {\n  campaignId: string\n  /** SHA-256 over canonicalised `(variantIds, scenarioIds, seeds, comparator, splitTag, executionRef, preregistrationHash)`. */\n  campaignFingerprint: string\n  preregistrationHash: string | null\n  /** Successful runs only. Failed runs land in `failedRuns`. */\n  runs: RunRecord[]\n  /** Integrity reports for every successful run. */\n  integrityReports: RunIntegrityReport[]\n  failedRuns: FailedRun[]\n  /** Computed when `report` is set on options. */\n  report?: ResearchReport\n  startedAt: string\n  endedAt: string\n}\n\n// ── Implementation ───────────────────────────────────────────────────────\n\nconst DEFAULT_INTEGRITY: RunIntegrityExpectations = {\n  llmSpansMin: 1,\n  requireRawCoverageOfLlmSpans: true,\n  requireOutcome: true,\n}\n\nexport async function runEvalCampaign<V>(\n  opts: EvalCampaignOptions<V>,\n): Promise<EvalCampaignResult> {\n  // ── Preflight ──────────────────────────────────────────────────────\n  if (opts.variants.length === 0) {\n    throw new Error('runEvalCampaign: variants must be non-empty.')\n  }\n  if (opts.scenarios.length === 0) {\n    throw new Error('runEvalCampaign: scenarios must be non-empty.')\n  }\n  const variantIds = new Set<string>()\n  for (const v of opts.variants) {\n    if (variantIds.has(v.id)) {\n      throw new Error(`runEvalCampaign: duplicate variant id \"${v.id}\".`)\n    }\n    variantIds.add(v.id)\n  }\n  const scenarioIds = new Set<string>()\n  for (const s of opts.scenarios) {\n    if (scenarioIds.has(s.scenarioId)) {\n      throw new Error(`runEvalCampaign: duplicate scenarioId \"${s.scenarioId}\".`)\n    }\n    scenarioIds.add(s.scenarioId)\n  }\n  if (opts.report?.comparator && !variantIds.has(opts.report.comparator)) {\n    throw new Error(\n      `runEvalCampaign: report.comparator \"${opts.report.comparator}\" is not a configured variantId.`,\n    )\n  }\n  if (!opts.commitSha) {\n    throw new Error('runEvalCampaign: commitSha is required (every RunRecord needs it).')\n  }\n\n  const seeds = opts.seeds ?? [0, 1, 2]\n  const splitTag: RunSplitTag = opts.splitTag ?? 'holdout'\n  const concurrency = Math.max(1, opts.concurrency ?? 1)\n  const integrity = { ...DEFAULT_INTEGRITY, ...(opts.integrity ?? {}) }\n  const onIntegrityFailure: CampaignIntegrityPolicy = opts.onIntegrityFailure ?? 'mark_failed'\n  const now = opts.now ?? (() => Date.now())\n  const executionRef = opts.executionRef ?? null\n  const preregistrationHash = opts.preregistrationHash ?? null\n\n  const rawSinkFactory = opts.rawSinkFactory ?? defaultRawSinkFactory(opts.workDir)\n\n  // ── Fingerprint ────────────────────────────────────────────────────\n  const campaignFingerprint = await hashJson({\n    campaignId: opts.campaignId,\n    variants: opts.variants.map((v) => v.id).sort(),\n    scenarios: opts.scenarios.map((s) => s.scenarioId).sort(),\n    seeds: [...seeds].sort((a, b) => a - b),\n    splitTag,\n    comparator: opts.report?.comparator ?? null,\n    executionRef,\n    preregistrationHash,\n  })\n\n  // ── Plan the matrix ────────────────────────────────────────────────\n  type Cell = { variant: CampaignVariant<V>; scenario: CampaignScenario; seed: number }\n  const cells: Cell[] = []\n  for (const variant of opts.variants) {\n    for (const scenario of opts.scenarios) {\n      for (const seed of seeds) {\n        cells.push({ variant, scenario, seed })\n      }\n    }\n  }\n\n  const startedAt = new Date(now()).toISOString()\n  const runs: RunRecord[] = []\n  const integrityReports: RunIntegrityReport[] = []\n  const failedRuns: FailedRun[] = []\n\n  // ── Execute (bounded-concurrency worker pool) ──────────────────────\n  // A genuine (non-CellExecutionError) error from any worker is a bug, not a\n  // run-level failure. We must NOT let `Promise.all` reject mid-flight and\n  // orphan the other workers' in-progress runs (emitters never finalized,\n  // sinks/handles leak, partial work silently discarded). Instead: capture the\n  // first genuine error, stop dispatching new cells so in-flight workers wind\n  // down, finalize every still-open run, then re-throw the aggregated error.\n  let cursor = 0\n  let aborting = false\n  const genuineErrors: unknown[] = []\n  // Emitters for runs that are currently mid-flight, keyed by runId. A worker\n  // registers its emitter before invoking the runner and removes it once the\n  // run is finalized (via endRun/abortRun, success or failure). Anything left\n  // here after the pool settles is an orphan we must abort.\n  const openRuns = new Map<string, TraceEmitter>()\n\n  async function worker(): Promise<void> {\n    while (!aborting) {\n      const i = cursor++\n      if (i >= cells.length) return\n      const cell = cells[i]!\n      try {\n        const result = await runOneCell(cell)\n        runs.push(result.record)\n        integrityReports.push(result.integrity)\n      } catch (err) {\n        if (err instanceof CellExecutionError) {\n          failedRuns.push(err.failed)\n          if (err.integrity) integrityReports.push(err.integrity)\n        } else {\n          // Genuine bug — not a runner failure, not an integrity failure.\n          // Capture it and stop dispatching so peers wind down gracefully;\n          // re-thrown after the pool settles. Do not surface here — that would\n          // reject Promise.all and orphan the other workers' runs.\n          genuineErrors.push(err)\n          aborting = true\n          return\n        }\n      }\n    }\n  }\n\n  async function runOneCell(\n    cell: Cell,\n  ): Promise<{ record: RunRecord; integrity: RunIntegrityReport }> {\n    const runId = (opts.runId ?? defaultRunId)({\n      campaignId: opts.campaignId,\n      runId: '', // unused by default generator\n      variantId: cell.variant.id,\n      scenarioId: cell.scenario.scenarioId,\n      seed: cell.seed,\n    })\n    const factoryParams: CampaignFactoryParams = {\n      campaignId: opts.campaignId,\n      runId,\n      variantId: cell.variant.id,\n      scenarioId: cell.scenario.scenarioId,\n      seed: cell.seed,\n    }\n    const store = opts.storeFactory(factoryParams)\n    const rawSink = rawSinkFactory(factoryParams)\n\n    const emitter = new TraceEmitter(store, {\n      runId,\n      now: opts.now,\n      onRunComplete: opts.onRunComplete,\n    })\n    // Track this run as open so a genuine error elsewhere in the pool can\n    // finalize it instead of orphaning it. Removed in the finally below.\n    openRuns.set(runId, emitter)\n\n    const chat = opts.chatFactory({ rawSink, runId })\n\n    const ctx: CampaignRunContext<V> = {\n      runId,\n      experimentId: opts.campaignId,\n      variant: cell.variant.payload,\n      variantId: cell.variant.id,\n      scenarioId: cell.scenario.scenarioId,\n      scenarioTags: cell.scenario.tags ?? {},\n      seed: cell.seed,\n      splitTag,\n      emitter,\n      store,\n      rawSink,\n      chat,\n    }\n\n    try {\n      const wallStart = now()\n      let outcome: CampaignRunOutcome\n      try {\n        outcome = await opts.runner(ctx)\n      } catch (err) {\n        const message = err instanceof Error ? err.message : String(err)\n        // The runner threw mid-execution. Abort the run so the emitter\n        // finalizes. The only benign abortRun failure is \"the runner never\n        // started the run\" (nothing to finalize). A store-write failure (disk\n        // full, FS error) is a genuine diagnostic — surface it rather than\n        // masking it as a plain runner failure.\n        await finalizeAbort(emitter, runId, message)\n        throw new CellExecutionError({\n          runId,\n          variantId: cell.variant.id,\n          scenarioId: cell.scenario.scenarioId,\n          seed: cell.seed,\n          reason: 'runner_threw',\n          error: message,\n        })\n      }\n      const wallMs = now() - wallStart\n\n      const integrityReport = await assertRunCaptured(store, runId, { ...integrity, rawSink })\n      if (!integrityReport.ok) {\n        switch (onIntegrityFailure) {\n          case 'throw':\n            throw new RunIntegrityError(integrityReport)\n          case 'mark_failed':\n            throw new CellExecutionError(\n              {\n                runId,\n                variantId: cell.variant.id,\n                scenarioId: cell.scenario.scenarioId,\n                seed: cell.seed,\n                reason: 'integrity_failed',\n                error: integrityReport.issues.map((i) => i.code).join(', '),\n              },\n              integrityReport,\n            )\n          case 'log':\n            // Caller wants the run admitted with a flagged report; fall through.\n            break\n        }\n      }\n\n      const recordOutcome: RunOutcome = {\n        raw: outcome.raw ?? {},\n      }\n      if (splitTag === 'holdout') recordOutcome.holdoutScore = outcome.score\n      else recordOutcome.searchScore = outcome.score\n      if (outcome.judgeScores !== undefined) recordOutcome.judgeScores = outcome.judgeScores\n\n      const record: RunRecord = {\n        runId,\n        experimentId: opts.campaignId,\n        candidateId: cell.variant.id,\n        seed: cell.seed,\n        model: outcome.model,\n        promptHash: outcome.promptHash,\n        configHash: outcome.configHash,\n        commitSha: opts.commitSha,\n        wallMs,\n        costUsd: outcome.costUsd,\n        costProvenance: outcome.costProvenance,\n        tokenUsage: outcome.tokenUsage,\n        terminalOutcome: 'succeeded',\n        judgeMetadata: outcome.judgeMetadata,\n        outcome: recordOutcome,\n        ...(outcome.failureClass ? { failureClass: outcome.failureClass } : {}),\n        failureMode: outcome.failureMode,\n        splitTag,\n        scenarioId: cell.scenario.scenarioId,\n      }\n      const profileSource =\n        outcome.agentProfile ??\n        (typeof opts.agentProfile === 'function'\n          ? await opts.agentProfile({\n              campaignId: opts.campaignId,\n              runId,\n              variantId: cell.variant.id,\n              scenarioId: cell.scenario.scenarioId,\n              seed: cell.seed,\n              variant: cell.variant.payload,\n              scenarioTags: cell.scenario.tags ?? {},\n            })\n          : opts.agentProfile)\n      if (profileSource !== undefined) {\n        const agentProfile = await resolveAgentProfileCell(profileSource)\n        assertAgentProfileMatchesRun(agentProfile, outcome.model, outcome.promptHash)\n        record.agentProfile = agentProfile\n      }\n      return { record: validateRunRecord(record), integrity: integrityReport }\n    } finally {\n      // This run's worker has finished with it (success, run-level failure, or\n      // genuine error). It is no longer the pool's job to finalize — drop it so\n      // the post-settle sweep doesn't double-abort a finalized run.\n      openRuns.delete(runId)\n    }\n  }\n\n  const workers = Array.from({ length: Math.min(concurrency, cells.length) }, () => worker())\n  // allSettled (not all): a genuine error in one worker must not reject the\n  // pool mid-flight and orphan the others. Each worker captures its own\n  // genuine error into `genuineErrors` and returns; we re-throw below.\n  await Promise.allSettled(workers)\n\n  // Finalize any run still open after the pool wound down. With the\n  // stop-dispatch flag these are the runs that were mid-flight in peer workers\n  // when the first genuine error fired — abort them so their emitters finalize\n  // (hooks fire, store records a terminal status) instead of leaking.\n  for (const [runId, emitter] of openRuns) {\n    await finalizeAbort(emitter, runId, 'campaign aborted: genuine error in a sibling run')\n  }\n  openRuns.clear()\n\n  if (genuineErrors.length > 0) {\n    throw genuineErrors.length === 1\n      ? genuineErrors[0]\n      : new AggregateError(\n          genuineErrors,\n          `runEvalCampaign: ${genuineErrors.length} runs failed with genuine (non-run-level) errors`,\n        )\n  }\n\n  // ── Optional research report ───────────────────────────────────────\n  let report: ResearchReport | undefined\n  if (opts.report) {\n    const reportOpts: ResearchReportOptions = {\n      ...opts.report,\n      comparator: opts.report.comparator,\n      split: splitTag === 'dev' ? 'search' : splitTag,\n      generatedAt: new Date(now()).toISOString(),\n      preregistrationHash: preregistrationHash ?? undefined,\n    }\n    report = await researchReport(runs, reportOpts)\n  }\n\n  const endedAt = new Date(now()).toISOString()\n\n  return {\n    campaignId: opts.campaignId,\n    campaignFingerprint,\n    preregistrationHash,\n    runs,\n    integrityReports,\n    failedRuns,\n    report,\n    startedAt,\n    endedAt,\n  }\n}\n\n// ── Internal ─────────────────────────────────────────────────────────────\n\nclass CellExecutionError extends Error {\n  readonly failed: FailedRun\n  readonly integrity?: RunIntegrityReport\n  constructor(failed: FailedRun, integrity?: RunIntegrityReport) {\n    super(`cell ${failed.variantId}/${failed.scenarioId}@${failed.seed} failed: ${failed.reason}`)\n    this.failed = failed\n    this.integrity = integrity\n  }\n}\n\n/**\n * Abort a run whose owning work threw or was orphaned by a sibling's genuine\n * error, finalizing the emitter so hooks fire and the store records a terminal\n * status. Safe to call unconditionally: it only aborts runs that are still\n * `running`. Two no-op cases are intentional and benign:\n *\n *   - the run was never started (absent from the store) — nothing to finalize\n *   - the run is already terminal (`completed` / `failed` / `aborted`) —\n *     finalizing again would overwrite the real outcome (e.g. flip a passed run\n *     to `aborted` with `{ pass: false, notes: reason }`), so we leave it alone\n *\n * Any OTHER failure of the store read or the abort write (disk full, FS fault,\n * backend down) is a genuine diagnostic and propagates rather than being\n * swallowed.\n */\nexport async function finalizeAbort(\n  emitter: TraceEmitter,\n  runId: string,\n  reason: string,\n): Promise<void> {\n  const existing = await emitter.traceStore.getRun(runId)\n  if (existing === undefined) return // run never started; nothing to abort\n  if (existing.status !== 'running') return // already finalized; never overwrite a real outcome\n  await emitter.abortRun(reason)\n}\n\nfunction defaultRawSinkFactory(workDir: string | undefined) {\n  return (params: CampaignFactoryParams): RawProviderSink => {\n    if (!workDir) {\n      throw new Error(\n        'runEvalCampaign: rawSinkFactory not supplied and workDir not set. Pass either to enable raw provider capture, or pass `new NoopRawProviderSink()` via rawSinkFactory to opt out explicitly.',\n      )\n    }\n    return new FileSystemRawProviderSink({\n      dir: `${workDir}/raw-events/${params.runId}`,\n    })\n  }\n}\n\nasync function resolveAgentProfileCell(\n  input: AgentProfileCell | AgentProfileCellInput,\n): Promise<AgentProfileCell> {\n  if (isAgentProfileCell(input)) {\n    if (!(await verifyAgentProfileCell(input))) {\n      throw new Error(`runEvalCampaign: agentProfile.cellId does not match its content`)\n    }\n    return input\n  }\n  return buildAgentProfileCell(input)\n}\n\nfunction isAgentProfileCell(\n  input: AgentProfileCell | AgentProfileCellInput,\n): input is AgentProfileCell {\n  return 'schemaVersion' in input && 'cellId' in input\n}\n\nfunction assertAgentProfileMatchesRun(\n  profile: AgentProfileCell,\n  model: string,\n  promptHash: string,\n): void {\n  if (profile.model !== undefined && profile.model !== model) {\n    throw new Error(\n      `runEvalCampaign: agentProfile.model \"${profile.model}\" does not match outcome.model \"${model}\"`,\n    )\n  }\n  if (profile.promptHash !== undefined && profile.promptHash !== promptHash) {\n    throw new Error(\n      `runEvalCampaign: agentProfile.promptHash \"${profile.promptHash}\" does not match outcome.promptHash \"${promptHash}\"`,\n    )\n  }\n}\n\nfunction defaultRunId(params: CampaignFactoryParams): string {\n  // Stable across re-runs: fingerprint of (campaignId, variantId, scenarioId, seed).\n  // Caller can override via opts.runId for non-deterministic IDs.\n  const base = `${params.campaignId}::${params.variantId}::${params.scenarioId}::${params.seed}`\n  // Lightweight hex: we don't need crypto-grade here, just stability + uniqueness.\n  let h1 = 0x811c9dc5\n  let h2 = 0x12345678\n  for (let i = 0; i < base.length; i++) {\n    const c = base.charCodeAt(i)\n    h1 = Math.imul(h1 ^ c, 0x01000193) >>> 0\n    h2 = Math.imul(h2 ^ c, 0x9e3779b1) >>> 0\n  }\n  return `run-${h1.toString(16).padStart(8, '0')}${h2.toString(16).padStart(8, '0')}`\n}\n"],"mappings":";;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;AA8SA,MAAM,oBAA8C;CAClD,aAAa;CACb,8BAA8B;CAC9B,gBAAgB;AAClB;AAEA,eAAsB,gBACpB,MAC6B;CAE7B,IAAI,KAAK,SAAS,WAAW,GAC3B,MAAM,IAAI,MAAM,8CAA8C;CAEhE,IAAI,KAAK,UAAU,WAAW,GAC5B,MAAM,IAAI,MAAM,+CAA+C;CAEjE,MAAM,6BAAa,IAAI,IAAY;CACnC,KAAK,MAAM,KAAK,KAAK,UAAU;EAC7B,IAAI,WAAW,IAAI,EAAE,EAAE,GACrB,MAAM,IAAI,MAAM,0CAA0C,EAAE,GAAG,GAAG;EAEpE,WAAW,IAAI,EAAE,EAAE;CACrB;CACA,MAAM,8BAAc,IAAI,IAAY;CACpC,KAAK,MAAM,KAAK,KAAK,WAAW;EAC9B,IAAI,YAAY,IAAI,EAAE,UAAU,GAC9B,MAAM,IAAI,MAAM,0CAA0C,EAAE,WAAW,GAAG;EAE5E,YAAY,IAAI,EAAE,UAAU;CAC9B;CACA,IAAI,KAAK,QAAQ,cAAc,CAAC,WAAW,IAAI,KAAK,OAAO,UAAU,GACnE,MAAM,IAAI,MACR,uCAAuC,KAAK,OAAO,WAAW,iCAChE;CAEF,IAAI,CAAC,KAAK,WACR,MAAM,IAAI,MAAM,oEAAoE;CAGtF,MAAM,QAAQ,KAAK,SAAS;EAAC;EAAG;EAAG;CAAC;CACpC,MAAM,WAAwB,KAAK,YAAY;CAC/C,MAAM,cAAc,KAAK,IAAI,GAAG,KAAK,eAAe,CAAC;CACrD,MAAM,YAAY;EAAE,GAAG;EAAmB,GAAI,KAAK,aAAa,CAAC;CAAG;CACpE,MAAM,qBAA8C,KAAK,sBAAsB;CAC/E,MAAM,MAAM,KAAK,cAAc,KAAK,IAAI;CACxC,MAAM,eAAe,KAAK,gBAAgB;CAC1C,MAAM,sBAAsB,KAAK,uBAAuB;CAExD,MAAM,iBAAiB,KAAK,kBAAkB,sBAAsB,KAAK,OAAO;CAGhF,MAAM,sBAAsB,MAAM,SAAS;EACzC,YAAY,KAAK;EACjB,UAAU,KAAK,SAAS,KAAK,MAAM,EAAE,EAAE,CAAC,CAAC,KAAK;EAC9C,WAAW,KAAK,UAAU,KAAK,MAAM,EAAE,UAAU,CAAC,CAAC,KAAK;EACxD,OAAO,CAAC,GAAG,KAAK,CAAC,CAAC,MAAM,GAAG,MAAM,IAAI,CAAC;EACtC;EACA,YAAY,KAAK,QAAQ,cAAc;EACvC;EACA;CACF,CAAC;CAID,MAAM,QAAgB,CAAC;CACvB,KAAK,MAAM,WAAW,KAAK,UACzB,KAAK,MAAM,YAAY,KAAK,WAC1B,KAAK,MAAM,QAAQ,OACjB,MAAM,KAAK;EAAE;EAAS;EAAU;CAAK,CAAC;CAK5C,MAAM,YAAY,IAAI,KAAK,IAAI,CAAC,CAAC,CAAC,YAAY;CAC9C,MAAM,OAAoB,CAAC;CAC3B,MAAM,mBAAyC,CAAC;CAChD,MAAM,aAA0B,CAAC;CASjC,IAAI,SAAS;CACb,IAAI,WAAW;CACf,MAAM,gBAA2B,CAAC;CAKlC,MAAM,2BAAW,IAAI,IAA0B;CAE/C,eAAe,SAAwB;EACrC,OAAO,CAAC,UAAU;GAChB,MAAM,IAAI;GACV,IAAI,KAAK,MAAM,QAAQ;GACvB,MAAM,OAAO,MAAM;GACnB,IAAI;IACF,MAAM,SAAS,MAAM,WAAW,IAAI;IACpC,KAAK,KAAK,OAAO,MAAM;IACvB,iBAAiB,KAAK,OAAO,SAAS;GACxC,SAAS,KAAK;IACZ,IAAI,eAAe,oBAAoB;KACrC,WAAW,KAAK,IAAI,MAAM;KAC1B,IAAI,IAAI,WAAW,iBAAiB,KAAK,IAAI,SAAS;IACxD,OAAO;KAKL,cAAc,KAAK,GAAG;KACtB,WAAW;KACX;IACF;GACF;EACF;CACF;CAEA,eAAe,WACb,MAC+D;EAC/D,MAAM,SAAS,KAAK,SAAS,aAAA,CAAc;GACzC,YAAY,KAAK;GACjB,OAAO;GACP,WAAW,KAAK,QAAQ;GACxB,YAAY,KAAK,SAAS;GAC1B,MAAM,KAAK;EACb,CAAC;EACD,MAAM,gBAAuC;GAC3C,YAAY,KAAK;GACjB;GACA,WAAW,KAAK,QAAQ;GACxB,YAAY,KAAK,SAAS;GAC1B,MAAM,KAAK;EACb;EACA,MAAM,QAAQ,KAAK,aAAa,aAAa;EAC7C,MAAM,UAAU,eAAe,aAAa;EAE5C,MAAM,UAAU,IAAI,aAAa,OAAO;GACtC;GACA,KAAK,KAAK;GACV,eAAe,KAAK;EACtB,CAAC;EAGD,SAAS,IAAI,OAAO,OAAO;EAE3B,MAAM,OAAO,KAAK,YAAY;GAAE;GAAS;EAAM,CAAC;EAEhD,MAAM,MAA6B;GACjC;GACA,cAAc,KAAK;GACnB,SAAS,KAAK,QAAQ;GACtB,WAAW,KAAK,QAAQ;GACxB,YAAY,KAAK,SAAS;GAC1B,cAAc,KAAK,SAAS,QAAQ,CAAC;GACrC,MAAM,KAAK;GACX;GACA;GACA;GACA;GACA;EACF;EAEA,IAAI;GACF,MAAM,YAAY,IAAI;GACtB,IAAI;GACJ,IAAI;IACF,UAAU,MAAM,KAAK,OAAO,GAAG;GACjC,SAAS,KAAK;IACZ,MAAM,UAAU,eAAe,QAAQ,IAAI,UAAU,OAAO,GAAG;IAM/D,MAAM,cAAc,SAAS,OAAO,OAAO;IAC3C,MAAM,IAAI,mBAAmB;KAC3B;KACA,WAAW,KAAK,QAAQ;KACxB,YAAY,KAAK,SAAS;KAC1B,MAAM,KAAK;KACX,QAAQ;KACR,OAAO;IACT,CAAC;GACH;GACA,MAAM,SAAS,IAAI,IAAI;GAEvB,MAAM,kBAAkB,MAAM,kBAAkB,OAAO,OAAO;IAAE,GAAG;IAAW;GAAQ,CAAC;GACvF,IAAI,CAAC,gBAAgB,IACnB,QAAQ,oBAAR;IACE,KAAK,SACH,MAAM,IAAI,kBAAkB,eAAe;IAC7C,KAAK,eACH,MAAM,IAAI,mBACR;KACE;KACA,WAAW,KAAK,QAAQ;KACxB,YAAY,KAAK,SAAS;KAC1B,MAAM,KAAK;KACX,QAAQ;KACR,OAAO,gBAAgB,OAAO,KAAK,MAAM,EAAE,IAAI,CAAC,CAAC,KAAK,IAAI;IAC5D,GACA,eACF;IACF,KAAK,OAEH;GACJ;GAGF,MAAM,gBAA4B,EAChC,KAAK,QAAQ,OAAO,CAAC,EACvB;GACA,IAAI,aAAa,WAAW,cAAc,eAAe,QAAQ;QAC5D,cAAc,cAAc,QAAQ;GACzC,IAAI,QAAQ,gBAAgB,KAAA,GAAW,cAAc,cAAc,QAAQ;GAE3E,MAAM,SAAoB;IACxB;IACA,cAAc,KAAK;IACnB,aAAa,KAAK,QAAQ;IAC1B,MAAM,KAAK;IACX,OAAO,QAAQ;IACf,YAAY,QAAQ;IACpB,YAAY,QAAQ;IACpB,WAAW,KAAK;IAChB;IACA,SAAS,QAAQ;IACjB,gBAAgB,QAAQ;IACxB,YAAY,QAAQ;IACpB,iBAAiB;IACjB,eAAe,QAAQ;IACvB,SAAS;IACT,GAAI,QAAQ,eAAe,EAAE,cAAc,QAAQ,aAAa,IAAI,CAAC;IACrE,aAAa,QAAQ;IACrB;IACA,YAAY,KAAK,SAAS;GAC5B;GACA,MAAM,gBACJ,QAAQ,iBACP,OAAO,KAAK,iBAAiB,aAC1B,MAAM,KAAK,aAAa;IACtB,YAAY,KAAK;IACjB;IACA,WAAW,KAAK,QAAQ;IACxB,YAAY,KAAK,SAAS;IAC1B,MAAM,KAAK;IACX,SAAS,KAAK,QAAQ;IACtB,cAAc,KAAK,SAAS,QAAQ,CAAC;GACvC,CAAC,IACD,KAAK;GACX,IAAI,kBAAkB,KAAA,GAAW;IAC/B,MAAM,eAAe,MAAM,wBAAwB,aAAa;IAChE,6BAA6B,cAAc,QAAQ,OAAO,QAAQ,UAAU;IAC5E,OAAO,eAAe;GACxB;GACA,OAAO;IAAE,QAAQ,kBAAkB,MAAM;IAAG,WAAW;GAAgB;EACzE,UAAU;GAIR,SAAS,OAAO,KAAK;EACvB;CACF;CAEA,MAAM,UAAU,MAAM,KAAK,EAAE,QAAQ,KAAK,IAAI,aAAa,MAAM,MAAM,EAAE,SAAS,OAAO,CAAC;CAI1F,MAAM,QAAQ,WAAW,OAAO;CAMhC,KAAK,MAAM,CAAC,OAAO,YAAY,UAC7B,MAAM,cAAc,SAAS,OAAO,kDAAkD;CAExF,SAAS,MAAM;CAEf,IAAI,cAAc,SAAS,GACzB,MAAM,cAAc,WAAW,IAC3B,cAAc,KACd,IAAI,eACF,eACA,oBAAoB,cAAc,OAAO,iDAC3C;CAIN,IAAI;CACJ,IAAI,KAAK,QAQP,SAAS,MAAM,eAAe,MAAM;EANlC,GAAG,KAAK;EACR,YAAY,KAAK,OAAO;EACxB,OAAO,aAAa,QAAQ,WAAW;EACvC,aAAa,IAAI,KAAK,IAAI,CAAC,CAAC,CAAC,YAAY;EACzC,qBAAqB,uBAAuB,KAAA;CAED,CAAC;CAGhD,MAAM,UAAU,IAAI,KAAK,IAAI,CAAC,CAAC,CAAC,YAAY;CAE5C,OAAO;EACL,YAAY,KAAK;EACjB;EACA;EACA;EACA;EACA;EACA;EACA;EACA;CACF;AACF;AAIA,IAAM,qBAAN,cAAiC,MAAM;CACrC;CACA;CACA,YAAY,QAAmB,WAAgC;EAC7D,MAAM,QAAQ,OAAO,UAAU,GAAG,OAAO,WAAW,GAAG,OAAO,KAAK,WAAW,OAAO,QAAQ;EAC7F,KAAK,SAAS;EACd,KAAK,YAAY;CACnB;AACF;;;;;;;;;;;;;;;;AAiBA,eAAsB,cACpB,SACA,OACA,QACe;CACf,MAAM,WAAW,MAAM,QAAQ,WAAW,OAAO,KAAK;CACtD,IAAI,aAAa,KAAA,GAAW;CAC5B,IAAI,SAAS,WAAW,WAAW;CACnC,MAAM,QAAQ,SAAS,MAAM;AAC/B;AAEA,SAAS,sBAAsB,SAA6B;CAC1D,QAAQ,WAAmD;EACzD,IAAI,CAAC,SACH,MAAM,IAAI,MACR,6LACF;EAEF,OAAO,IAAI,0BAA0B,EACnC,KAAK,GAAG,QAAQ,cAAc,OAAO,QACvC,CAAC;CACH;AACF;AAEA,eAAe,wBACb,OAC2B;CAC3B,IAAI,mBAAmB,KAAK,GAAG;EAC7B,IAAI,CAAE,MAAM,uBAAuB,KAAK,GACtC,MAAM,IAAI,MAAM,iEAAiE;EAEnF,OAAO;CACT;CACA,OAAO,sBAAsB,KAAK;AACpC;AAEA,SAAS,mBACP,OAC2B;CAC3B,OAAO,mBAAmB,SAAS,YAAY;AACjD;AAEA,SAAS,6BACP,SACA,OACA,YACM;CACN,IAAI,QAAQ,UAAU,KAAA,KAAa,QAAQ,UAAU,OACnD,MAAM,IAAI,MACR,wCAAwC,QAAQ,MAAM,kCAAkC,MAAM,EAChG;CAEF,IAAI,QAAQ,eAAe,KAAA,KAAa,QAAQ,eAAe,YAC7D,MAAM,IAAI,MACR,6CAA6C,QAAQ,WAAW,uCAAuC,WAAW,EACpH;AAEJ;AAEA,SAAS,aAAa,QAAuC;CAG3D,MAAM,OAAO,GAAG,OAAO,WAAW,IAAI,OAAO,UAAU,IAAI,OAAO,WAAW,IAAI,OAAO;CAExF,IAAI,KAAK;CACT,IAAI,KAAK;CACT,KAAK,IAAI,IAAI,GAAG,IAAI,KAAK,QAAQ,KAAK;EACpC,MAAM,IAAI,KAAK,WAAW,CAAC;EAC3B,KAAK,KAAK,KAAK,KAAK,GAAG,QAAU,MAAM;EACvC,KAAK,KAAK,KAAK,KAAK,GAAG,UAAU,MAAM;CACzC;CACA,OAAO,OAAO,GAAG,SAAS,EAAE,CAAC,CAAC,SAAS,GAAG,GAAG,IAAI,GAAG,SAAS,EAAE,CAAC,CAAC,SAAS,GAAG,GAAG;AAClF"}

package/dist/{exact-types-ccQAyut1.d.ts → exact-types-qnexxJ1Z.d.ts}

@@ -1,4 +1,4 @@
-import { l as AnalystRunResult, s as AnalystRunEvent, t as Analyst, u as AnalystRunSummary } from "./types-B2NsbrNy.js";
+import { l as AnalystRunResult, s as AnalystRunEvent, t as Analyst, u as AnalystRunSummary } from "./types-D9ssmxKL.js";
 import { z } from "zod";
 //#region src/analyst/exact-types.d.ts
 /** Analyst metadata required before the exact registry path will execute it. */
@@ -231,4 +231,4 @@ declare const exactExecutionPlanSchema: z.ZodObject<{
 }, z.core.$strict>;
 //#endregion
 export { ExactAnalystRunPolicySnapshot as a, ExactAnalystSnapshot as c, ExactExecutionComponentSnapshot as d, ExactAnalystRunEvent as i, ExactCapableAnalyst as l, ExactAnalystExecutionPlanSnapshot as n, ExactAnalystRunResult as o, ExactAnalystRunCompletion as r, ExactAnalystRunSummary as s, ExactAnalystBudgetSnapshot as t, ExactExecutionComponentIdentity as u };
-//# sourceMappingURL=exact-types-ccQAyut1.d.ts.map
+//# sourceMappingURL=exact-types-qnexxJ1Z.d.ts.map

package/dist/{exact-types-ccQAyut1.d.ts.map → exact-types-qnexxJ1Z.d.ts.map}

@@ -1 +1 @@
-{"version":3,"file":"exact-types-ccQAyut1.d.ts","names":[],"sources":["../src/analyst/exact-types.ts"],"mappings":";;;;UAMiB,oBAAoB,0BAA0B,QAAQ;;WAE5D,iBAAiB,SAAS;;KAGhC,aAAa,KAAK,0BAA0B,mBACpC,aAAa,UACtB,+BACc,aAAa,IAAI,aAAa,EAAE,WAC5C;KAEM,uBAAuB,aAAa,EAAE,aAAa;KACnD,6BAA6B,aAAa,EAAE,aAAa;;UAGpD;EACf;EACA;EACA,QAAQ,SAAS;;;UAIF;EACf;EACA;EACA;;KAyBU,gCAAgC,aAAa,EAAE,aAAa;KAC5D,oCAAoC,aAC9C,EAAE,aAAa;KAGL,yBAAyB;;EAEnC;;KAGU;EACN;;EAEA;EACA;IAAS;IAAe;;;UAGb,8BAA8B;EAC7C,aAAa;EACb,gBAAgB;EAChB,YAAY;;;KAIF,wBACP,QAAQ;EAAmB;;EAC1B,gBAAgB;MAEjB,KAAK,QAAQ;EAAmB;;EAC/B,SAAS;KAEX,QAAQ;EAAmB;MAC1B,KAAK,QAAQ;EAAmB;;EAC/B,SAAS;MAEV,KAAK,QAAQ;EAAmB;;EAC/B,QAAQ;;cAwCR,uBAAqB,EAAA;;;;;;;;;;;;;;;;;;;;;;;;;GAOzB,EAAA,KAAA;cAII,sBAAoB,EAAA,uBAAA,EAAA;;;;;;;;;;;GAaxB,EAAA,KAAA;cAiBI,sBAAoB,EAAA;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;GAYxB,EAAA,KAAA;cAEI,0BAAwB,EAAA;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;GA+E1B,EAAA,KAAA"}
+{"version":3,"file":"exact-types-qnexxJ1Z.d.ts","names":[],"sources":["../src/analyst/exact-types.ts"],"mappings":";;;;UAMiB,oBAAoB,0BAA0B,QAAQ;;WAE5D,iBAAiB,SAAS;;KAGhC,aAAa,KAAK,0BAA0B,mBACpC,aAAa,UACtB,+BACc,aAAa,IAAI,aAAa,EAAE,WAC5C;KAEM,uBAAuB,aAAa,EAAE,aAAa;KACnD,6BAA6B,aAAa,EAAE,aAAa;;UAGpD;EACf;EACA;EACA,QAAQ,SAAS;;;UAIF;EACf;EACA;EACA;;KAyBU,gCAAgC,aAAa,EAAE,aAAa;KAC5D,oCAAoC,aAC9C,EAAE,aAAa;KAGL,yBAAyB;;EAEnC;;KAGU;EACN;;EAEA;EACA;IAAS;IAAe;;;UAGb,8BAA8B;EAC7C,aAAa;EACb,gBAAgB;EAChB,YAAY;;;KAIF,wBACP,QAAQ;EAAmB;;EAC1B,gBAAgB;MAEjB,KAAK,QAAQ;EAAmB;;EAC/B,SAAS;KAEX,QAAQ;EAAmB;MAC1B,KAAK,QAAQ;EAAmB;;EAC/B,SAAS;MAEV,KAAK,QAAQ;EAAmB;;EAC/B,QAAQ;;cAwCR,uBAAqB,EAAA;;;;;;;;;;;;;;;;;;;;;;;;;GAOzB,EAAA,KAAA;cAII,sBAAoB,EAAA,uBAAA,EAAA;;;;;;;;;;;GAaxB,EAAA,KAAA;cAiBI,sBAAoB,EAAA;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;GAYxB,EAAA,KAAA;cAEI,0BAAwB,EAAA;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;GA+E1B,EAAA,KAAA"}

package/dist/{exec-y-DCLqK7.js → exec-D9WpA2p-.js}

@@ -199,6 +199,6 @@ function wrapActionForExec(action, cwd) {
 	return `cd ${`'${cwd.replaceAll("'", `'\\''`)}'`} && printf %s ${b64} | base64 -d | sh`;
 }
 //#endregion
-export { unreadableExitCount as _, TIMEOUT_OBSERVATION_MARKER as a, decodeRecordedTurns as c, isElidedField as d, isRecordedTimeout as f, parseRecordedReturncode as g, parseObservationOutput as h, SUBMIT_ACTION_SIGNATURE as i, deriveFailureSignature as l, isSubmitOnlyAction as m, FORMAT_ERROR_OBSERVATION_PREFIX as n, assertReplayableTrajectory as o, isSubmitAction as p, RECORDED_ELISION_PATTERN as r, classifyObservation as s, wrapActionForExec as t, finalRecordedOutcome as u };
+export { assertReplayableTrajectory as a, deriveFailureSignature as c, isRecordedTimeout as d, isSubmitAction as f, unreadableExitCount as g, parseRecordedReturncode as h, TIMEOUT_OBSERVATION_MARKER as i, finalRecordedOutcome as l, parseObservationOutput as m, FORMAT_ERROR_OBSERVATION_PREFIX as n, classifyObservation as o, isSubmitOnlyAction as p, SUBMIT_ACTION_SIGNATURE as r, decodeRecordedTurns as s, wrapActionForExec as t, isElidedField as u };
 
-//# sourceMappingURL=exec-y-DCLqK7.js.map
+//# sourceMappingURL=exec-D9WpA2p-.js.map

package/dist/exec-D9WpA2p-.js.map

@@ -0,0 +1 @@
+{"version":3,"file":"exec-D9WpA2p-.js","names":[],"sources":["../src/trajectory-replay/steps.ts","../src/trajectory-replay/exec.ts"],"sourcesContent":["/**\n * Recorded shell-trajectory steps and the observation grammar they carry.\n *\n * A recorded trajectory is the action/observation sequence an agent actually\n * ran. Scaffolds that execute one shell command per step (mini-SWE and the\n * CodeTracer-normalized corpora built from it) tag each observation with the\n * command's returncode and its combined output:\n *\n *     <returncode>2</returncode>\n *     <output>\n *     …command output…\n *     </output>\n *\n * That is one of four shapes a recorded turn carries, and the other three are\n * not command results at all:\n *\n *   - a timeout notice, when the environment killed the command at its bound;\n *   - a format-error notice, when the scaffold rejected the turn and ran\n *     nothing;\n *   - an elision marker `$<hex>`, when the published dump dropped the string.\n *\n * A turn also carries no observation when it is the run's last turn, because\n * the scaffold records an observation only when it hands one back to the model.\n *\n * The parsers here are the only place that grammar is decoded. Everything\n * downstream — replay verdicts, corpus enumeration, admission funnels, fix\n * prompts — reads the returncode, the output, and the failure signature through\n * these functions. A second decoder elsewhere is how a corpus reads as\n * unreplayable when it is not.\n */\n\n/**\n * One step of a recorded shell trajectory. Structural: any richer step record\n * (file refs, thinking text, tool type) satisfies it.\n */\nexport interface RecordedTrajectoryStep {\n  /** 1-based position in the trajectory. */\n  readonly step_id: number\n  readonly action: string\n  /** Null when the step recorded no observation (terminal submit steps). */\n  readonly observation: string | null\n}\n\n/** Recorded returncode of a step, or null when the observation carries none. */\nexport function parseRecordedReturncode(observation: string | null): number | null {\n  if (!observation) return null\n  const m = /<returncode>(-?\\d+)<\\/returncode>/.exec(observation)\n  return m ? Number(m[1]) : null\n}\n\n/** Text between the observation's <output> tags, or the raw observation when\n *  the tags are absent. */\nexport function parseObservationOutput(observation: string | null): string {\n  if (!observation) return ''\n  const m = /<output>\\n?([\\s\\S]*?)\\n?<\\/output>/.exec(observation)\n  return m ? m[1]! : observation\n}\n\n/**\n * Stable failure-signature candidate: the first line of the recorded output\n * that contains the word \"error\". Null when no such line exists — a verdict\n * then falls back to returncode-only matching and says so.\n * Pass an explicit signature to override (compiler quote glyphs vary with\n * locale, so a hand-picked ASCII substring is often more robust).\n */\nexport function deriveFailureSignature(observation: string | null): string | null {\n  const line = parseObservationOutput(observation)\n    .split('\\n')\n    .find((l) => /\\berror\\b/i.test(l))\n  return line ? line.trim().slice(0, 200) : null\n}\n\n/** mini-SWE's end-of-run submit convention: the agent echoes this sentinel\n *  and dumps the diff. A label on this step marks a bad SUBMIT DECISION, not a\n *  failed command — there is no executable failure to reproduce, so it is\n *  never a counterfactual replay target. */\nexport const SUBMIT_ACTION_SIGNATURE = 'COMPLETE_TASK_AND_SUBMIT_FINAL_OUTPUT'\n\nexport function isSubmitAction(action: string): boolean {\n  return action.includes(SUBMIT_ACTION_SIGNATURE)\n}\n\n/**\n * True when the action is the sentinel echo and nothing else.\n *\n * The distinction decides whether a step may be dropped. An agent is told to\n * issue the sentinel alone, and 5.7% of recorded runs end on a command that\n * writes files or edits them and then echoes it. Dropping such a step because\n * it holds the sentinel would remove the run's last state change from the\n * replay, so the recorded end state and the replayed one would differ.\n */\nexport function isSubmitOnlyAction(action: string): boolean {\n  return /^echo\\s+(?:\"COMPLETE_TASK_AND_SUBMIT_FINAL_OUTPUT\"|'COMPLETE_TASK_AND_SUBMIT_FINAL_OUTPUT'|COMPLETE_TASK_AND_SUBMIT_FINAL_OUTPUT)$/.test(\n    action.trim(),\n  )\n}\n\n// ── Fields the published dump dropped ────────────────────────────────\n\n/**\n * A field the dump replaced with a counter instead of its text.\n *\n * The counter is hexadecimal and rises by one per dropped string in document\n * order, so a decimal-only reading (`$12`) accepts every marker that carries a\n * letter (`$3a`) as if it were real text. A command read that way replays as\n * the literal two-to-four characters `$3a`, which is not the command the run\n * executed.\n *\n * Nothing in the dump maps a marker back to its text: the same marker carries\n * different content in different rows, so there is no dictionary to read. A\n * field that matches is unrecoverable, so a caller rejects the row that holds\n * it rather than replaying the marker.\n */\nconst RECORDED_ELISION_PATTERN = /^\\$[0-9a-f]+$/\n\nexport function isElidedField(value: string | null | undefined): boolean {\n  return typeof value === 'string' && RECORDED_ELISION_PATTERN.test(value)\n}\n\n// ── Observation grammar ──────────────────────────────────────────────\n\n/** Substring the timeout notice always carries, whatever the command was. */\nexport const TIMEOUT_OBSERVATION_MARKER = 'timed out and has been killed'\n\n/** Opening of the notice the scaffold writes when a turn held no single action. */\nexport const FORMAT_ERROR_OBSERVATION_PREFIX =\n  'Please always provide EXACTLY ONE action in triple backticks, found '\n\n/**\n * What a recorded observation is.\n *\n * `command-result` is the only kind that carries an exit status. `timeout` and\n * `format-error` are the scaffold speaking rather than a command: the first\n * says the environment killed the command, the second says no command ran at\n * all. `elided` and `absent` carry no information about the step, and\n * `unreadable` is a shape this grammar does not know — never assumed to be\n * anything.\n */\nexport type RecordedObservationKind =\n  | 'command-result'\n  | 'timeout'\n  | 'format-error'\n  | 'elided'\n  | 'absent'\n  | 'unreadable'\n\nexport function classifyObservation(observation: string | null): RecordedObservationKind {\n  if (observation === null) return 'absent'\n  if (isElidedField(observation)) return 'elided'\n  if (parseRecordedReturncode(observation) !== null) return 'command-result'\n  if (observation.startsWith(FORMAT_ERROR_OBSERVATION_PREFIX)) return 'format-error'\n  if (observation.includes(TIMEOUT_OBSERVATION_MARKER)) return 'timeout'\n  return 'unreadable'\n}\n\n/** True when the recording shows the environment killed this step at its\n *  wall-clock bound. Such a step carries no returncode, so no replay can\n *  confirm or contradict it. */\nexport function isRecordedTimeout(observation: string | null): boolean {\n  return classifyObservation(observation) === 'timeout'\n}\n\n// ── Turns to steps ───────────────────────────────────────────────────\n\n/**\n * One turn as the published trajectory dump holds it.\n *\n * A turn is not a step: the system prompt, the task statement and every turn\n * the scaffold rejected are turns that executed nothing.\n */\nexport interface RecordedTrajectoryTurn {\n  readonly src?: string | null\n  readonly msg?: string | null\n  readonly tools?: readonly { readonly cmd?: string | null }[] | null\n  readonly obs?: string | null\n}\n\nexport interface DecodedTrajectory {\n  /** Executed commands in recorded order, each holding its OWN observation. */\n  readonly steps: readonly RecordedTrajectoryStep[]\n  /** Turns the scaffold rejected before anything ran. */\n  readonly formatErrorTurns: number\n  /**\n   * Executed commands whose text the dump dropped.\n   *\n   * The step stays in `steps` carrying the marker, because the run did execute\n   * a command there and a shorter list would misreport the trajectory. The\n   * marker is not a command, so any count above zero means this trajectory\n   * cannot be replayed: gate on it, or call `assertReplayableTrajectory`.\n   */\n  readonly elidedCommands: number\n  /** True when the run's last turn was the submit sentinel with no observation. */\n  readonly endedOnSubmitSentinel: boolean\n  /** Turns carrying an observation this grammar cannot read. */\n  readonly unreadableTurns: number\n}\n\n/**\n * Pair every recorded command with the observation of its own turn.\n *\n * Collecting commands and observations into two lists and zipping them is the\n * decode that looks right and is not: a turn the scaffold rejected carries an\n * observation of its own, so from the first such turn onward every observation\n * belongs to a different command than the one it is read against.\n *\n * A rejected turn executed nothing, so it is never a step — including when the\n * dump recorded a command for it. The scaffold rejects a turn holding several\n * bash blocks and runs none of them, while the dump keeps one of the blocks in\n * the command field. Replaying that field would execute a command the recorded\n * run did not, which is a worse corpus than a smaller one.\n *\n * A rejected turn is recognised by its observation, so a rejected turn whose\n * observation the dump elided is indistinguishable from an executed command\n * whose observation it elided. Both read as a step. Nothing in the dump\n * separates them, and the row-level defence is the share of unreadable exits a\n * caller admits: a row with no elided observation cannot hold this case at all.\n *\n * A trailing step that echoes the sentinel and nothing else, with no\n * observation, is dropped from `steps` and reported as `endedOnSubmitSentinel`.\n * The scaffold records an observation only when it hands one to the model, and\n * the sentinel ends the run, so the missing observation is the end of the\n * transcript rather than a gap in it. Echoing the sentinel changes no state, so\n * the recorded end state is the state the step before it left.\n *\n * A step that DID get an observation stays a step: the run continued past it.\n * So does a step that echoes the sentinel after doing real work — its state\n * change is part of the recorded end state, and with no observation its exit is\n * unknown, which `finalRecordedOutcome` reports rather than hides.\n */\nexport function decodeRecordedTurns(turns: readonly RecordedTrajectoryTurn[]): DecodedTrajectory {\n  const steps: RecordedTrajectoryStep[] = []\n  let formatErrorTurns = 0\n  let unreadableTurns = 0\n  for (const turn of turns) {\n    const command = turn.tools?.[0]?.cmd ?? null\n    const observation = turn.obs ?? null\n    const kind = classifyObservation(observation)\n    if (kind === 'format-error') {\n      formatErrorTurns += 1\n      continue\n    }\n    if (command === null) {\n      if (kind === 'unreadable' || kind === 'elided') unreadableTurns += 1\n      continue\n    }\n    steps.push({ step_id: steps.length + 1, action: command, observation })\n  }\n  const last = steps[steps.length - 1]\n  const endedOnSubmitSentinel =\n    last !== undefined && last.observation === null && isSubmitOnlyAction(last.action)\n  if (endedOnSubmitSentinel) steps.pop()\n  return {\n    steps,\n    formatErrorTurns,\n    elidedCommands: steps.filter((step) => isElidedField(step.action)).length,\n    endedOnSubmitSentinel,\n    unreadableTurns,\n  }\n}\n\n// ── The state a trajectory ended in ──────────────────────────────────\n\n/**\n * How the recorded run's last executed command ended.\n *\n * `killed` is a measured outcome, not a missing one: the environment stopped\n * the command at its bound and wrote a notice instead of an exit status.\n * `unreadable` names the observation kind that blocked the read, so a funnel\n * can report which shape cost it the row.\n */\nexport type RecordedFinalOutcome =\n  | { readonly kind: 'returncode'; readonly value: number }\n  | { readonly kind: 'killed' }\n  | { readonly kind: 'unreadable'; readonly reason: RecordedObservationKind }\n\n/**\n * The outcome of the last step, or `null` when the trajectory has no steps.\n *\n * Reads only the last step. `decodeRecordedTurns` has already removed the turns\n * that executed nothing, so the last step is the last command the run ran.\n */\nexport function finalRecordedOutcome(\n  steps: readonly RecordedTrajectoryStep[],\n): RecordedFinalOutcome | null {\n  const last = steps[steps.length - 1]\n  if (last === undefined) return null\n  const kind = classifyObservation(last.observation)\n  if (kind === 'command-result') {\n    return { kind: 'returncode', value: parseRecordedReturncode(last.observation)! }\n  }\n  if (kind === 'timeout') return { kind: 'killed' }\n  return { kind: 'unreadable', reason: kind }\n}\n\n/**\n * Steps whose recorded exit a replay cannot check.\n *\n * A killed step counts: the recording holds no exit status to compare a replay\n * against, so agreement on it cannot be measured either way.\n */\nexport function unreadableExitCount(steps: readonly RecordedTrajectoryStep[]): number {\n  return steps.filter((step) => classifyObservation(step.observation) !== 'command-result').length\n}\n\n/**\n * Throw unless every recorded command survived the dump.\n *\n * A replay executes `steps` verbatim, so one elision marker in an action means\n * the replay runs the literal two-to-four characters of the marker instead of\n * the command the run executed. The guard is here so a replayer needs one call\n * rather than a field check it can forget.\n */\nexport function assertReplayableTrajectory(decoded: DecodedTrajectory): void {\n  if (decoded.elidedCommands === 0) return\n  const markers = decoded.steps\n    .filter((step) => isElidedField(step.action))\n    .map((step) => step.action)\n  throw new Error(\n    `trajectory holds ${decoded.elidedCommands} command(s) the dump elided (${markers.slice(0, 5).join(', ')}); ` +\n      'no dictionary maps a marker back to its text, so this trajectory cannot be replayed',\n  )\n}\n","/**\n * The execution boundary replay runs across.\n *\n * A replay needs one thing from its environment: a session that runs a shell\n * command inside the trajectory's own image and reports the exit code and\n * output. That is the whole contract. Concrete backends — a sandbox platform\n * client, a docker exec, an SSH shell — live with the consumer that owns the\n * infrastructure, so this package depends on no sandbox client.\n */\n\nexport interface ReplayExecResult {\n  exitCode: number\n  stdout: string\n  stderr: string\n}\n\nexport interface ReplayExecSession {\n  exec(command: string, timeoutMs: number): Promise<ReplayExecResult>\n  close(): Promise<void>\n}\n\nexport interface ReplayExecBackend {\n  /** One fresh execution environment per call; the caller closes it. */\n  open(): Promise<ReplayExecSession>\n}\n\n/** Builds a backend pinned to one image. Callers that resolve images\n *  internally (batch, corpus wire, finding verification) take this instead of\n *  a backend, so every case runs against its own image. */\nexport type ReplayExecBackendFactory = (image: string) => ReplayExecBackend\n\n/**\n * mini-SWE runs every action as a fresh /bin/sh subshell from a fixed\n * workdir. Reproduce that exactly — and stay quote-proof for arbitrary\n * recorded actions — by piping the base64 of the action into `sh` after\n * cd-ing to the workdir. Exit code is sh's, i.e. the action's.\n */\nexport function wrapActionForExec(action: string, cwd: string): string {\n  const b64 = Buffer.from(action, 'utf8').toString('base64')\n  const quotedCwd = `'${cwd.replaceAll(\"'\", `'\\\\''`)}'`\n  return `cd ${quotedCwd} && printf %s ${b64} | base64 -d | sh`\n}\n"],"mappings":";;AA4CA,SAAgB,wBAAwB,aAA2C;CACjF,IAAI,CAAC,aAAa,OAAO;CACzB,MAAM,IAAI,oCAAoC,KAAK,WAAW;CAC9D,OAAO,IAAI,OAAO,EAAE,EAAE,IAAI;AAC5B;;;AAIA,SAAgB,uBAAuB,aAAoC;CACzE,IAAI,CAAC,aAAa,OAAO;CACzB,MAAM,IAAI,qCAAqC,KAAK,WAAW;CAC/D,OAAO,IAAI,EAAE,KAAM;AACrB;;;;;;;;AASA,SAAgB,uBAAuB,aAA2C;CAChF,MAAM,OAAO,uBAAuB,WAAW,CAAC,CAC7C,MAAM,IAAI,CAAC,CACX,MAAM,MAAM,aAAa,KAAK,CAAC,CAAC;CACnC,OAAO,OAAO,KAAK,KAAK,CAAC,CAAC,MAAM,GAAG,GAAG,IAAI;AAC5C;;;;;AAMA,MAAa,0BAA0B;AAEvC,SAAgB,eAAe,QAAyB;CACtD,OAAO,OAAO,SAAS,uBAAuB;AAChD;;;;;;;;;;AAWA,SAAgB,mBAAmB,QAAyB;CAC1D,OAAO,qIAAqI,KAC1I,OAAO,KAAK,CACd;AACF;;;;;;;;;;;;;;;AAkBA,MAAM,2BAA2B;AAEjC,SAAgB,cAAc,OAA2C;CACvE,OAAO,OAAO,UAAU,YAAY,yBAAyB,KAAK,KAAK;AACzE;;AAKA,MAAa,6BAA6B;;AAG1C,MAAa,kCACX;AAoBF,SAAgB,oBAAoB,aAAqD;CACvF,IAAI,gBAAgB,MAAM,OAAO;CACjC,IAAI,cAAc,WAAW,GAAG,OAAO;CACvC,IAAI,wBAAwB,WAAW,MAAM,MAAM,OAAO;CAC1D,IAAI,YAAY,WAAA,sEAA0C,GAAG,OAAO;CACpE,IAAI,YAAY,SAAA,+BAAmC,GAAG,OAAO;CAC7D,OAAO;AACT;;;;AAKA,SAAgB,kBAAkB,aAAqC;CACrE,OAAO,oBAAoB,WAAW,MAAM;AAC9C;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;AAqEA,SAAgB,oBAAoB,OAA6D;CAC/F,MAAM,QAAkC,CAAC;CACzC,IAAI,mBAAmB;CACvB,IAAI,kBAAkB;CACtB,KAAK,MAAM,QAAQ,OAAO;EACxB,MAAM,UAAU,KAAK,QAAQ,EAAE,EAAE,OAAO;EACxC,MAAM,cAAc,KAAK,OAAO;EAChC,MAAM,OAAO,oBAAoB,WAAW;EAC5C,IAAI,SAAS,gBAAgB;GAC3B,oBAAoB;GACpB;EACF;EACA,IAAI,YAAY,MAAM;GACpB,IAAI,SAAS,gBAAgB,SAAS,UAAU,mBAAmB;GACnE;EACF;EACA,MAAM,KAAK;GAAE,SAAS,MAAM,SAAS;GAAG,QAAQ;GAAS;EAAY,CAAC;CACxE;CACA,MAAM,OAAO,MAAM,MAAM,SAAS;CAClC,MAAM,wBACJ,SAAS,KAAA,KAAa,KAAK,gBAAgB,QAAQ,mBAAmB,KAAK,MAAM;CACnF,IAAI,uBAAuB,MAAM,IAAI;CACrC,OAAO;EACL;EACA;EACA,gBAAgB,MAAM,QAAQ,SAAS,cAAc,KAAK,MAAM,CAAC,CAAC,CAAC;EACnE;EACA;CACF;AACF;;;;;;;AAuBA,SAAgB,qBACd,OAC6B;CAC7B,MAAM,OAAO,MAAM,MAAM,SAAS;CAClC,IAAI,SAAS,KAAA,GAAW,OAAO;CAC/B,MAAM,OAAO,oBAAoB,KAAK,WAAW;CACjD,IAAI,SAAS,kBACX,OAAO;EAAE,MAAM;EAAc,OAAO,wBAAwB,KAAK,WAAW;CAAG;CAEjF,IAAI,SAAS,WAAW,OAAO,EAAE,MAAM,SAAS;CAChD,OAAO;EAAE,MAAM;EAAc,QAAQ;CAAK;AAC5C;;;;;;;AAQA,SAAgB,oBAAoB,OAAkD;CACpF,OAAO,MAAM,QAAQ,SAAS,oBAAoB,KAAK,WAAW,MAAM,gBAAgB,CAAC,CAAC;AAC5F;;;;;;;;;AAUA,SAAgB,2BAA2B,SAAkC;CAC3E,IAAI,QAAQ,mBAAmB,GAAG;CAClC,MAAM,UAAU,QAAQ,MACrB,QAAQ,SAAS,cAAc,KAAK,MAAM,CAAC,CAAC,CAC5C,KAAK,SAAS,KAAK,MAAM;CAC5B,MAAM,IAAI,MACR,oBAAoB,QAAQ,eAAe,+BAA+B,QAAQ,MAAM,GAAG,CAAC,CAAC,CAAC,KAAK,IAAI,EAAE,uFAE3G;AACF;;;;;;;;;AC5RA,SAAgB,kBAAkB,QAAgB,KAAqB;CACrE,MAAM,MAAM,OAAO,KAAK,QAAQ,MAAM,CAAC,CAAC,SAAS,QAAQ;CAEzD,OAAO,MAAM,IADS,IAAI,WAAW,KAAK,OAAO,EAAE,GAC5B,gBAAgB,IAAI;AAC7C"}