npm - @tangle-network/agent-eval - Versions diffs - 0.187.2 → 0.189.0 - Mend

@tangle-network/agent-eval 0.187.2 → 0.189.0

This diff represents the content of publicly available package versions that have been released to one of the supported registries. The information contained in this diff is provided for informational purposes only and reflects changes between package versions as they appear in their respective public registries.

Files changed (279) hide show

package/dist/{dspy-rlm-engine-BE9lCzzL.js.map → dspy-rlm-engine-DLYttKBm.js.map} RENAMED Viewed

	@@ -1 +1 @@
1	- {"version":3,"file":"dspy-rlm-engine-BE9lCzzL.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 {\n closeServer,\n listenLocal,\n sendJsonIfOpen,\n waitForActiveHandlers,\n} 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\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 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 {\n DEFAULT_TRACE_ANALYST_OUTPUT_TOKENS,\n type TraceAnalysisEngine,\n type TraceAnalysisEngineResult,\n} from './engine'\nimport { decodeRawFindingArray } from './finding-codec'\nimport { startTraceToolCallback, type TraceToolCallbackLimits } from './trace-tool-callback'\n\nconst DEFAULT_TIMEOUT_MS = 10 60_000\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: {@link DEFAULT_TRACE_ANALYST_OUTPUT_TOKENS}. /\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_TRACE_ANALYST_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;;;;ACjIA,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,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,SAASA,WAAS,OAAkD;CAClE,OAAO,OAAO,UAAU,YAAY,UAAU,QAAQ,CAAC,MAAM,QAAQ,KAAK;AAC5E;;;AC1JA,MAAM,qBAAqB,KAAK;AAChC,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,mBAAA;CAChC,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"}
1	+ {"version":3,"file":"dspy-rlm-engine-DLYttKBm.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 {\n closeServer,\n listenLocal,\n sendJsonIfOpen,\n waitForActiveHandlers,\n} 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\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 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 {\n DEFAULT_TRACE_ANALYST_OUTPUT_TOKENS,\n type TraceAnalysisEngine,\n type TraceAnalysisEngineResult,\n} from './engine'\nimport { decodeRawFindingArray } from './finding-codec'\nimport { startTraceToolCallback, type TraceToolCallbackLimits } from './trace-tool-callback'\n\nconst DEFAULT_TIMEOUT_MS = 10 60_000\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: {@link DEFAULT_TRACE_ANALYST_OUTPUT_TOKENS}. /\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_TRACE_ANALYST_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;;;;ACjIA,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,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,SAASA,WAAS,OAAkD;CAClE,OAAO,OAAO,UAAU,YAAY,UAAU,QAAQ,CAAC,MAAM,QAAQ,KAAK;AAC5E;;;AC1JA,MAAM,qBAAqB,KAAK;AAChC,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,mBAAA;CAChC,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-DbMXUIVP.js ADDED Viewed

@@ -0,0 +1,334 @@
+import { t as newRecordId } from "./record-id-DUgsK5qp.js";
+import { AsyncLocalStorage } from "node:async_hooks";
+//#region src/trace/emitter.ts
+/**
+* TraceEmitter — hierarchical span builder for one Run. Emitters do NOT
+* share state.
+*
+* Parenting. A new span's parent is, in order: the explicit `parentSpanId`,
+* the innermost handle still open in the current async context, then the span
+* of the enclosing `within`. `within` runs its callback in an async context of
+* its own, so parallel `within` calls each parent their own children with no
+* ids passed by hand. Handles from `span`/`llm`/`tool`/`retrieval`/`sandbox`
+* nest by call order inside one async context; open parallel work with
+* `within`, or pass `parentSpanId`.
+*
+* Capture never throws into the traced run. A store write that fails is
+* counted as dropped, handed to `onCaptureError`, and the counts are written
+* onto the Run as `capture` when it ends, where `assertRunCaptured` reads them.
+* A Run whose records are incomplete therefore says so instead of reading as a
+* shorter run.
+*/
+var TraceEmitter = class {
+	store;
+	context = new AsyncLocalStorage();
+	rootFrame = {
+		parentSpanId: void 0,
+		open: []
+	};
+	_runId;
+	now;
+	id;
+	hooks;
+	hookErrors;
+	onCaptureError;
+	capture = {
+		written: 0,
+		dropped: 0
+	};
+	constructor(store, options = {}) {
+		this.store = store;
+		this.now = options.now ?? (() => Date.now());
+		this.id = options.id ?? (() => newRecordId());
+		this._runId = options.runId ?? this.id();
+		this.hooks = options.onRunComplete ?? [];
+		this.hookErrors = options.hookErrors ?? "swallow";
+		this.onCaptureError = options.onCaptureError ?? warnOnFirstCaptureError(this._runId);
+	}
+	get runId() {
+		return this._runId;
+	}
+	get traceStore() {
+		return this.store;
+	}
+	/** Store writes so far: how many landed, how many were dropped, and the last failure. */
+	captureStats() {
+		return { ...this.capture };
+	}
+	/** The span a new span would be parented to in the current async context. */
+	currentSpanId() {
+		const frame = this.frame();
+		return frame.open[frame.open.length - 1] ?? frame.parentSpanId;
+	}
+	frame() {
+		return this.context.getStore() ?? this.rootFrame;
+	}
+	async write(write, fn) {
+		try {
+			await fn();
+			this.capture.written += 1;
+			return true;
+		} catch (error) {
+			this.capture.dropped += 1;
+			this.capture.lastError = `${write}: ${error instanceof Error ? error.message : String(error)}`;
+			try {
+				this.onCaptureError(error, write);
+			} catch {}
+			return false;
+		}
+	}
+	/** Append a hook after construction (e.g. attach the trace analyst). */
+	addRunCompleteHook(hook) {
+		this.hooks.push(hook);
+	}
+	/**
+	* Begin a Run.
+	*
+	* `scenarioId` is required on the persisted Run shape — every Run downstream
+	* gets a non-empty scenarioId so filters and aggregations stay simple — but
+	* the INPUT here accepts it as optional. When omitted, startRun substitutes
+	* a sensible default (`run.layer ?? run.tags?.['kind'] ?? 'runtime'`) so
+	* runtime / operator / meta-eval runs that have no curated-scenario corpus
+	* to anchor to don't have to invent placeholder strings at the call site.
+	*/
+	async startRun(run) {
+		const scenarioId = run.scenarioId ?? run.layer ?? run.tags?.kind ?? "runtime";
+		const full = {
+			...run,
+			scenarioId,
+			runId: this._runId,
+			startedAt: this.now(),
+			status: "running"
+		};
+		await this.write("appendRun", () => this.store.appendRun(full));
+		return full;
+	}
+	/**
+	* End the Run. The final record carries `capture`: the store writes made
+	* before it and how many were dropped.
+	*/
+	async endRun(outcome) {
+		const status = outcome?.pass === false ? "failed" : "completed";
+		const capture = this.captureStats();
+		await this.write("updateRun", () => this.store.updateRun(this._runId, {
+			endedAt: this.now(),
+			status,
+			outcome,
+			capture
+		}));
+		await this.runHooks({
+			runId: this._runId,
+			emitter: this,
+			store: this.store,
+			outcome,
+			status
+		});
+	}
+	async abortRun(reason) {
+		const outcome = {
+			pass: false,
+			notes: reason
+		};
+		const capture = this.captureStats();
+		await this.write("updateRun", () => this.store.updateRun(this._runId, {
+			endedAt: this.now(),
+			status: "aborted",
+			outcome,
+			capture
+		}));
+		await this.runHooks({
+			runId: this._runId,
+			emitter: this,
+			store: this.store,
+			outcome,
+			status: "aborted"
+		});
+	}
+	async runHooks(ctx) {
+		for (const hook of this.hooks) try {
+			await hook(ctx);
+		} catch (err) {
+			if (this.hookErrors === "throw") throw err;
+			await this.write("appendEvent", () => this.store.appendEvent({
+				eventId: this.id(),
+				runId: this._runId,
+				kind: "log",
+				timestamp: this.now(),
+				payload: {
+					source: "run_complete_hook",
+					error: err instanceof Error ? err.message : String(err)
+				}
+			}));
+		}
+	}
+	async span(init) {
+		return this.open(init, this.frame());
+	}
+	/**
+	* Create a span parented in the current async context. `frame` is the
+	* context the handle stays open in, or null for a `within` span, which
+	* parents its children through its own context instead.
+	*/
+	async open(init, frame) {
+		const spanId = this.id();
+		const parent = init.parentSpanId ?? this.currentSpanId();
+		const span = {
+			spanId,
+			runId: this._runId,
+			startedAt: this.now(),
+			...init,
+			parentSpanId: parent
+		};
+		frame?.open.push(spanId);
+		await this.write("appendSpan", () => this.store.appendSpan(span));
+		return this.handle(span, frame);
+	}
+	handle(span, frame) {
+		const close = () => {
+			if (!frame) return;
+			const idx = frame.open.lastIndexOf(span.spanId);
+			if (idx >= 0) frame.open.splice(idx, 1);
+		};
+		return {
+			span,
+			end: async (patch) => {
+				const endedAt = this.now();
+				close();
+				await this.write("updateSpan", () => this.store.updateSpan(span.spanId, {
+					endedAt,
+					status: "ok",
+					...patch
+				}));
+			},
+			fail: async (error, patch) => {
+				const endedAt = this.now();
+				const errStr = error instanceof Error ? error.message : error;
+				close();
+				await this.write("updateSpan", () => this.store.updateSpan(span.spanId, {
+					endedAt,
+					status: "error",
+					error: errStr,
+					...patch
+				}));
+			}
+		};
+	}
+	llm(init) {
+		return this.span({
+			kind: "llm",
+			...init
+		});
+	}
+	tool(init) {
+		return this.span({
+			kind: "tool",
+			...init
+		});
+	}
+	retrieval(init) {
+		return this.span({
+			kind: "retrieval",
+			...init
+		});
+	}
+	async recordJudge(verdict) {
+		const spanId = this.id();
+		const now = this.now();
+		const full = {
+			spanId,
+			runId: this._runId,
+			kind: "judge",
+			startedAt: now,
+			endedAt: now,
+			status: "ok",
+			...verdict
+		};
+		await this.write("appendSpan", () => this.store.appendSpan(full));
+		return full;
+	}
+	sandbox(init) {
+		return this.span({
+			kind: "sandbox",
+			...init
+		});
+	}
+	async emit(event) {
+		const full = {
+			eventId: this.id(),
+			runId: this._runId,
+			spanId: event.spanId ?? this.currentSpanId(),
+			kind: event.kind,
+			timestamp: this.now(),
+			payload: event.payload ?? {}
+		};
+		await this.write("appendEvent", () => this.store.appendEvent(full));
+		return full;
+	}
+	async recordBudget(entry) {
+		const full = {
+			runId: this._runId,
+			timestamp: entry.timestamp ?? this.now(),
+			dimension: entry.dimension,
+			limit: entry.limit,
+			consumed: entry.consumed,
+			remaining: entry.remaining,
+			breached: entry.breached,
+			spanId: entry.spanId ?? this.currentSpanId()
+		};
+		await this.write("appendBudgetEntry", () => this.store.appendBudgetEntry(full));
+		if (full.breached) await this.emit({
+			kind: "budget_breach",
+			spanId: full.spanId,
+			payload: {
+				dimension: full.dimension,
+				limit: full.limit,
+				consumed: full.consumed
+			}
+		});
+		return full;
+	}
+	async recordArtifact(artifact) {
+		const full = {
+			artifactId: this.id(),
+			runId: this._runId,
+			...artifact
+		};
+		await this.write("appendArtifact", () => this.store.appendArtifact(full));
+		return full;
+	}
+	/**
+	* Runs `fn` inside a span; auto-ends on success, auto-fails on throw.
+	* Returns the fn's return value. Use this for the 95% case.
+	*
+	* `fn` runs in an async context of its own, so spans it opens are parented
+	* to this span even while sibling `within` calls run in parallel.
+	*/
+	async within(init, fn) {
+		const handle = await this.open(init, null);
+		const frame = {
+			parentSpanId: handle.span.spanId,
+			open: []
+		};
+		try {
+			const result = await this.context.run(frame, () => fn(handle));
+			await handle.end();
+			return result;
+		} catch (err) {
+			await handle.fail(err instanceof Error ? err : String(err));
+			throw err;
+		}
+	}
+};
+function warnOnFirstCaptureError(runId) {
+	let warned = false;
+	return (error, write) => {
+		if (warned) return;
+		warned = true;
+		const reason = error instanceof Error ? error.message : String(error);
+		globalThis.process?.emitWarning?.(`TraceEmitter run ${runId}: ${write} failed (${reason}); this and later failed writes are counted in the run's capture report.`, { code: "AGENT_EVAL_TRACE_CAPTURE" });
+	};
+}
+//#endregion
+export { TraceEmitter as t };
+//# sourceMappingURL=emitter-DbMXUIVP.js.map

package/dist/emitter-DbMXUIVP.js.map ADDED Viewed

@@ -0,0 +1 @@

+ {"version":3,"file":"emitter-DbMXUIVP.js","names":[],"sources":["../src/trace/emitter.ts"],"sourcesContent":["/**\n * TraceEmitter — hierarchical span builder for one Run. Emitters do NOT\n * share state.\n *\n * Parenting. A new span's parent is, in order: the explicit `parentSpanId`,\n * the innermost handle still open in the current async context, then the span\n * of the enclosing `within`. `within` runs its callback in an async context of\n * its own, so parallel `within` calls each parent their own children with no\n * ids passed by hand. Handles from `span`/`llm`/`tool`/`retrieval`/`sandbox`\n * nest by call order inside one async context; open parallel work with\n * `within`, or pass `parentSpanId`.\n *\n * Capture never throws into the traced run. A store write that fails is\n * counted as dropped, handed to `onCaptureError`, and the counts are written\n * onto the Run as `capture` when it ends, where `assertRunCaptured` reads them.\n * A Run whose records are incomplete therefore says so instead of reading as a\n * shorter run.\n */\n\nimport { AsyncLocalStorage } from 'node:async_hooks'\nimport { newRecordId } from '../record-id'\nimport type {\n Artifact,\n BudgetLedgerEntry,\n EventKind,\n JudgeSpan,\n LlmSpan,\n RetrievalSpan,\n Run,\n RunCaptureReport,\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\n/** The store write that failed, for `onCaptureError`. */\nexport type CaptureWrite =\n | 'appendRun'\n | 'updateRun'\n | 'appendSpan'\n | 'updateSpan'\n | 'appendEvent'\n | 'appendBudgetEntry'\n | 'appendArtifact'\n\nexport type CaptureErrorHandler = (error: unknown, write: CaptureWrite) => void\n\n/** Parent state for one async context: open handles, then the enclosing `within` span. */\ninterface ParentFrame {\n parentSpanId: string | undefined\n open: string[]\n}\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 * Called once per failed store write. The write is already counted as\n * dropped; the handler reports it. Default: one process warning per emitter\n * (code `AGENT_EVAL_TRACE_CAPTURE`) on the first failure. A handler that\n * throws is ignored, because reporting must not break the traced run either.\n */\n onCaptureError?: CaptureErrorHandler\n}\n\nexport class TraceEmitter {\n private store: TraceStore\n private readonly context = new AsyncLocalStorage<ParentFrame>()\n private readonly rootFrame: ParentFrame = { parentSpanId: undefined, open: [] }\n private _runId: string\n private now: () => number\n private id: () => string\n private hooks: RunCompleteHook[]\n private hookErrors: 'swallow' | 'throw'\n private onCaptureError: CaptureErrorHandler\n private capture: RunCaptureReport = { written: 0, dropped: 0 }\n\n constructor(store: TraceStore, options: TraceEmitterOptions = {}) {\n this.store = store\n this.now = options.now ?? (() => Date.now())\n this.id = options.id ?? (() => newRecordId())\n this._runId = options.runId ?? this.id()\n this.hooks = options.onRunComplete ?? []\n this.hookErrors = options.hookErrors ?? 'swallow'\n this.onCaptureError = options.onCaptureError ?? warnOnFirstCaptureError(this._runId)\n }\n\n get runId(): string {\n return this._runId\n }\n\n get traceStore(): TraceStore {\n return this.store\n }\n\n /** Store writes so far: how many landed, how many were dropped, and the last failure. */\n captureStats(): RunCaptureReport {\n return { ...this.capture }\n }\n\n /** The span a new span would be parented to in the current async context. */\n currentSpanId(): string | undefined {\n const frame = this.frame()\n return frame.open[frame.open.length - 1] ?? frame.parentSpanId\n }\n\n private frame(): ParentFrame {\n return this.context.getStore() ?? this.rootFrame\n }\n\n private async write(write: CaptureWrite, fn: () => Promise<unknown>): Promise<boolean> {\n try {\n await fn()\n this.capture.written += 1\n return true\n } catch (error) {\n this.capture.dropped += 1\n this.capture.lastError = `${write}: ${error instanceof Error ? error.message : String(error)}`\n try {\n this.onCaptureError(error, write)\n } catch {\n // A failing reporter must not reach the traced run.\n }\n return false\n }\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.write('appendRun', () => this.store.appendRun(full))\n return full\n }\n\n /**\n * End the Run. The final record carries `capture`: the store writes made\n * before it and how many were dropped.\n */\n async endRun(outcome?: RunOutcome): Promise<void> {\n const status: 'completed' | 'failed' = outcome?.pass === false ? 'failed' : 'completed'\n const capture = this.captureStats()\n await this.write('updateRun', () =>\n this.store.updateRun(this._runId, { endedAt: this.now(), status, outcome, capture }),\n )\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 const capture = this.captureStats()\n await this.write('updateRun', () =>\n this.store.updateRun(this._runId, {\n endedAt: this.now(),\n status: 'aborted',\n outcome,\n capture,\n }),\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 await this.write('appendEvent', () =>\n 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 )\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 return this.open<S>(init, this.frame())\n }\n\n /**\n * Create a span parented in the current async context. `frame` is the\n * context the handle stays open in, or null for a `within` span, which\n * parents its children through its own context instead.\n */\n private async open<S extends Span>(\n init: { kind: SpanKind; name: string; parentSpanId?: string },\n frame: ParentFrame | null,\n ): Promise<SpanHandle<S>> {\n const spanId = this.id()\n const parent = init.parentSpanId ?? this.currentSpanId()\n const span = {\n spanId,\n runId: this._runId,\n startedAt: this.now(),\n ...init,\n parentSpanId: parent,\n } as unknown as S\n frame?.open.push(spanId)\n await this.write('appendSpan', () => this.store.appendSpan(span))\n return this.handle<S>(span, frame)\n }\n\n private handle<S extends Span>(span: S, frame: ParentFrame | null): SpanHandle<S> {\n const close = () => {\n if (!frame) return\n const idx = frame.open.lastIndexOf(span.spanId)\n if (idx >= 0) frame.open.splice(idx, 1)\n }\n return {\n span,\n end: async (patch?: Partial<S>) => {\n const endedAt = this.now()\n close()\n await this.write('updateSpan', () =>\n this.store.updateSpan(span.spanId, {\n endedAt,\n status: 'ok',\n ...patch,\n } as Partial<Span>),\n )\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 close()\n await this.write('updateSpan', () =>\n this.store.updateSpan(span.spanId, {\n endedAt,\n status: 'error',\n error: errStr,\n ...patch,\n } as Partial<Span>),\n )\n },\n }\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.write('appendSpan', () => 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.currentSpanId(),\n kind: event.kind,\n timestamp: this.now(),\n payload: event.payload ?? {},\n }\n await this.write('appendEvent', () => 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.currentSpanId(),\n }\n await this.write('appendBudgetEntry', () => 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.write('appendArtifact', () => 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 * `fn` runs in an async context of its own, so spans it opens are parented\n * to this span even while sibling `within` calls run in parallel.\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.open<Span>(init, null)\n const frame: ParentFrame = { parentSpanId: handle.span.spanId, open: [] }\n try {\n const result = await this.context.run(frame, () => 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 warnOnFirstCaptureError(runId: string): CaptureErrorHandler {\n let warned = false\n return (error, write) => {\n if (warned) return\n warned = true\n const reason = error instanceof Error ? error.message : String(error)\n globalThis.process?.emitWarning?.(\n `TraceEmitter run ${runId}: ${write} failed (${reason}); this and later failed writes are counted in the run's capture report.`,\n { code: 'AGENT_EVAL_TRACE_CAPTURE' },\n )\n }\n}\n"],"mappings":";;;;;;;;;;;;;;;;;;;;;AAsGA,IAAa,eAAb,MAA0B;CACxB;CACA,UAA2B,IAAI,kBAA+B;CAC9D,YAA0C;EAAE,cAAc,KAAA;EAAW,MAAM,CAAC;CAAE;CAC9E;CACA;CACA;CACA;CACA;CACA;CACA,UAAoC;EAAE,SAAS;EAAG,SAAS;CAAE;CAE7D,YAAY,OAAmB,UAA+B,CAAC,GAAG;EAChE,KAAK,QAAQ;EACb,KAAK,MAAM,QAAQ,cAAc,KAAK,IAAI;EAC1C,KAAK,KAAK,QAAQ,aAAa,YAAY;EAC3C,KAAK,SAAS,QAAQ,SAAS,KAAK,GAAG;EACvC,KAAK,QAAQ,QAAQ,iBAAiB,CAAC;EACvC,KAAK,aAAa,QAAQ,cAAc;EACxC,KAAK,iBAAiB,QAAQ,kBAAkB,wBAAwB,KAAK,MAAM;CACrF;CAEA,IAAI,QAAgB;EAClB,OAAO,KAAK;CACd;CAEA,IAAI,aAAyB;EAC3B,OAAO,KAAK;CACd;;CAGA,eAAiC;EAC/B,OAAO,EAAE,GAAG,KAAK,QAAQ;CAC3B;;CAGA,gBAAoC;EAClC,MAAM,QAAQ,KAAK,MAAM;EACzB,OAAO,MAAM,KAAK,MAAM,KAAK,SAAS,MAAM,MAAM;CACpD;CAEA,QAA6B;EAC3B,OAAO,KAAK,QAAQ,SAAS,KAAK,KAAK;CACzC;CAEA,MAAc,MAAM,OAAqB,IAA8C;EACrF,IAAI;GACF,MAAM,GAAG;GACT,KAAK,QAAQ,WAAW;GACxB,OAAO;EACT,SAAS,OAAO;GACd,KAAK,QAAQ,WAAW;GACxB,KAAK,QAAQ,YAAY,GAAG,MAAM,IAAI,iBAAiB,QAAQ,MAAM,UAAU,OAAO,KAAK;GAC3F,IAAI;IACF,KAAK,eAAe,OAAO,KAAK;GAClC,QAAQ,CAER;GACA,OAAO;EACT;CACF;;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,mBAAmB,KAAK,MAAM,UAAU,IAAI,CAAC;EAC9D,OAAO;CACT;;;;;CAMA,MAAM,OAAO,SAAqC;EAChD,MAAM,SAAiC,SAAS,SAAS,QAAQ,WAAW;EAC5E,MAAM,UAAU,KAAK,aAAa;EAClC,MAAM,KAAK,MAAM,mBACf,KAAK,MAAM,UAAU,KAAK,QAAQ;GAAE,SAAS,KAAK,IAAI;GAAG;GAAQ;GAAS;EAAQ,CAAC,CACrF;EACA,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,UAAU,KAAK,aAAa;EAClC,MAAM,KAAK,MAAM,mBACf,KAAK,MAAM,UAAU,KAAK,QAAQ;GAChC,SAAS,KAAK,IAAI;GAClB,QAAQ;GACR;GACA;EACF,CAAC,CACH;EACA,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,MAAM,KAAK,MAAM,qBACf,KAAK,MAAM,YAAY;IACrB,SAAS,KAAK,GAAG;IACjB,OAAO,KAAK;IACZ,MAAM;IACN,WAAW,KAAK,IAAI;IACpB,SAAS;KACP,QAAQ;KACR,OAAO,eAAe,QAAQ,IAAI,UAAU,OAAO,GAAG;IACxD;GACF,CAAC,CACH;EACF;CAEJ;CAIA,MAAM,KACJ,MAMwB;EACxB,OAAO,KAAK,KAAQ,MAAM,KAAK,MAAM,CAAC;CACxC;;;;;;CAOA,MAAc,KACZ,MACA,OACwB;EACxB,MAAM,SAAS,KAAK,GAAG;EACvB,MAAM,SAAS,KAAK,gBAAgB,KAAK,cAAc;EACvD,MAAM,OAAO;GACX;GACA,OAAO,KAAK;GACZ,WAAW,KAAK,IAAI;GACpB,GAAG;GACH,cAAc;EAChB;EACA,OAAO,KAAK,KAAK,MAAM;EACvB,MAAM,KAAK,MAAM,oBAAoB,KAAK,MAAM,WAAW,IAAI,CAAC;EAChE,OAAO,KAAK,OAAU,MAAM,KAAK;CACnC;CAEA,OAA+B,MAAS,OAA0C;EAChF,MAAM,cAAc;GAClB,IAAI,CAAC,OAAO;GACZ,MAAM,MAAM,MAAM,KAAK,YAAY,KAAK,MAAM;GAC9C,IAAI,OAAO,GAAG,MAAM,KAAK,OAAO,KAAK,CAAC;EACxC;EACA,OAAO;GACL;GACA,KAAK,OAAO,UAAuB;IACjC,MAAM,UAAU,KAAK,IAAI;IACzB,MAAM;IACN,MAAM,KAAK,MAAM,oBACf,KAAK,MAAM,WAAW,KAAK,QAAQ;KACjC;KACA,QAAQ;KACR,GAAG;IACL,CAAkB,CACpB;GACF;GACA,MAAM,OAAO,OAAuB,UAAuB;IACzD,MAAM,UAAU,KAAK,IAAI;IACzB,MAAM,SAAS,iBAAiB,QAAQ,MAAM,UAAU;IACxD,MAAM;IACN,MAAM,KAAK,MAAM,oBACf,KAAK,MAAM,WAAW,KAAK,QAAQ;KACjC;KACA,QAAQ;KACR,OAAO;KACP,GAAG;IACL,CAAkB,CACpB;GACF;EACF;CACF;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,oBAAoB,KAAK,MAAM,WAAW,IAAI,CAAC;EAChE,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,cAAc;GAC3C,MAAM,MAAM;GACZ,WAAW,KAAK,IAAI;GACpB,SAAS,MAAM,WAAW,CAAC;EAC7B;EACA,MAAM,KAAK,MAAM,qBAAqB,KAAK,MAAM,YAAY,IAAI,CAAC;EAClE,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,cAAc;EAC7C;EACA,MAAM,KAAK,MAAM,2BAA2B,KAAK,MAAM,kBAAkB,IAAI,CAAC;EAC9E,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,wBAAwB,KAAK,MAAM,eAAe,IAAI,CAAC;EACxE,OAAO;CACT;;;;;;;;CAWA,MAAM,OACJ,MACA,IACY;EACZ,MAAM,SAAS,MAAM,KAAK,KAAW,MAAM,IAAI;EAC/C,MAAM,QAAqB;GAAE,cAAc,OAAO,KAAK;GAAQ,MAAM,CAAC;EAAE;EACxE,IAAI;GACF,MAAM,SAAS,MAAM,KAAK,QAAQ,IAAI,aAAa,GAAG,MAAM,CAAC;GAC7D,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,wBAAwB,OAAoC;CACnE,IAAI,SAAS;CACb,QAAQ,OAAO,UAAU;EACvB,IAAI,QAAQ;EACZ,SAAS;EACT,MAAM,SAAS,iBAAiB,QAAQ,MAAM,UAAU,OAAO,KAAK;EACpE,WAAW,SAAS,cAClB,oBAAoB,MAAM,IAAI,MAAM,WAAW,OAAO,2EACtD,EAAE,MAAM,2BAA2B,CACrC;CACF;AACF"}