@tangle-network/agent-eval 0.173.1 → 0.173.3

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 (44) hide show
  1. package/CHANGELOG.md +18 -0
  2. package/dist/analyst/index.d.ts +2 -2
  3. package/dist/analyst/index.js +4 -4
  4. package/dist/{benchmark-command-k-fOGzt8.js → benchmark-command-BY9oscke.js} +7 -7
  5. package/dist/{benchmark-command-k-fOGzt8.js.map → benchmark-command-BY9oscke.js.map} +1 -1
  6. package/dist/benchmarks/index.js +3 -3
  7. package/dist/campaign/index.js +5 -5
  8. package/dist/{campaign-pxS0wmo4.js → campaign-B3kPMU8S.js} +6 -6
  9. package/dist/{campaign-pxS0wmo4.js.map → campaign-B3kPMU8S.js.map} +1 -1
  10. package/dist/{chat-client-Db4bqYfA.js → chat-client-CkmjYlfB.js} +2 -2
  11. package/dist/{chat-client-Db4bqYfA.js.map → chat-client-CkmjYlfB.js.map} +1 -1
  12. package/dist/{chat-json-call-C26igCih.js → chat-json-call-B_Xv2oJK.js} +2 -2
  13. package/dist/{chat-json-call-C26igCih.js.map → chat-json-call-B_Xv2oJK.js.map} +1 -1
  14. package/dist/cli.js +3 -3
  15. package/dist/contract/index.js +5 -5
  16. package/dist/{define-agent-eval-D_i_s69h.js → define-agent-eval-8h3lXXee.js} +3 -3
  17. package/dist/{define-agent-eval-D_i_s69h.js.map → define-agent-eval-8h3lXXee.js.map} +1 -1
  18. package/dist/{dspy-rlm-engine-D5byiHn9.js → dspy-rlm-engine-CF0t2ITD.js} +2 -2
  19. package/dist/{dspy-rlm-engine-D5byiHn9.js.map → dspy-rlm-engine-CF0t2ITD.js.map} +1 -1
  20. package/dist/{external-optimizer-process-Cq_Pg15r.js → external-optimizer-process-BwITA9Jp.js} +2 -2
  21. package/dist/{external-optimizer-process-Cq_Pg15r.js.map → external-optimizer-process-BwITA9Jp.js.map} +1 -1
  22. package/dist/{external-optimizer-subprocess-DgNebftP.js → external-optimizer-subprocess-wBWeoG6A.js} +2 -2
  23. package/dist/{external-optimizer-subprocess-DgNebftP.js.map → external-optimizer-subprocess-wBWeoG6A.js.map} +1 -1
  24. package/dist/index.js +8 -8
  25. package/dist/{llm-client-CxQtdtd6.js → llm-client-CGlSi8sb.js} +2 -1
  26. package/dist/llm-client-CGlSi8sb.js.map +1 -0
  27. package/dist/{llm-judge-B2YxbAJb.js → llm-judge-BfqMFo4h.js} +3 -3
  28. package/dist/{llm-judge-B2YxbAJb.js.map → llm-judge-BfqMFo4h.js.map} +1 -1
  29. package/dist/openapi.json +1 -1
  30. package/dist/{produced-state-7VYDwtkk.js → produced-state-D91uDvQw.js} +3 -3
  31. package/dist/{produced-state-7VYDwtkk.js.map → produced-state-D91uDvQw.js.map} +1 -1
  32. package/dist/{semantic-concept-judge-Ct3QU7t5.js → semantic-concept-judge-Dok7_35a.js} +4 -4
  33. package/dist/{semantic-concept-judge-Ct3QU7t5.js.map → semantic-concept-judge-Dok7_35a.js.map} +1 -1
  34. package/dist/{server-BR6onwZB.js → server-D_cjseFN.js} +2 -2
  35. package/dist/{server-BR6onwZB.js.map → server-D_cjseFN.js.map} +1 -1
  36. package/dist/{skillopt-optimization-method-BzdphODy.js → skillopt-optimization-method-DDw3v3gA.js} +5 -5
  37. package/dist/{skillopt-optimization-method-BzdphODy.js.map → skillopt-optimization-method-DDw3v3gA.js.map} +1 -1
  38. package/dist/supervisor-run/index.d.ts.map +1 -1
  39. package/dist/supervisor-run/index.js +43 -18
  40. package/dist/supervisor-run/index.js.map +1 -1
  41. package/dist/types-gvRsyJLh.d.ts.map +1 -1
  42. package/dist/wire/index.js +1 -1
  43. package/package.json +2 -2
  44. package/dist/llm-client-CxQtdtd6.js.map +0 -1
@@ -1,5 +1,5 @@
1
1
  import { a as resolveModelPricing } from "./metrics-Qv-cpptD.js";
2
- import { _ as removeCredentialEnvironment, a as closeServer, c as waitForActiveHandlers, d as assertExternalOptimizerModelBudget, i as startExternalOptimizerModelProxy, o as listenLocal, r as runWithCleanup, s as sendJsonIfOpen, t as runExternalOptimizerProcess, v as resolveExternalOptimizerCallbackLimits, y as resolveExternalOptimizerProcessLimits } from "./external-optimizer-subprocess-DgNebftP.js";
2
+ import { _ as removeCredentialEnvironment, a as closeServer, c as waitForActiveHandlers, d as assertExternalOptimizerModelBudget, i as startExternalOptimizerModelProxy, o as listenLocal, r as runWithCleanup, s as sendJsonIfOpen, t as runExternalOptimizerProcess, v as resolveExternalOptimizerCallbackLimits, y as resolveExternalOptimizerProcessLimits } from "./external-optimizer-subprocess-wBWeoG6A.js";
3
3
  import { B as parseFindingSubject, P as coerceJson, j as RawAnalystFindingSchema } from "./kind-factory-gP6lDySe.js";
4
4
  import { randomBytes } from "node:crypto";
5
5
  import { createServer } from "node:http";
@@ -462,4 +462,4 @@ function isRecord(value) {
462
462
  //#endregion
463
463
  export { decodeRawFindingArray as n, createDspyRlmTraceEngine as t };
464
464
 
465
- //# sourceMappingURL=dspy-rlm-engine-D5byiHn9.js.map
465
+ //# sourceMappingURL=dspy-rlm-engine-CF0t2ITD.js.map
@@ -1 +1 @@
1
- {"version":3,"file":"dspy-rlm-engine-D5byiHn9.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-CF0t2ITD.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"}
@@ -1,6 +1,6 @@
1
1
  import { n as contentHash } from "./verdict-cache-B3eCVQtY.js";
2
2
  import { m as tryAcquireAtomicFileLock, p as probeAtomicFileLock } from "./ledger-core-PIfjCbKn.js";
3
- import { a as closeServer, c as waitForActiveHandlers, g as isRecord, h as isExternalTextCandidate, o as listenLocal, s as sendJsonIfOpen, v as resolveExternalOptimizerCallbackLimits } from "./external-optimizer-subprocess-DgNebftP.js";
3
+ import { a as closeServer, c as waitForActiveHandlers, g as isRecord, h as isExternalTextCandidate, o as listenLocal, s as sendJsonIfOpen, v as resolveExternalOptimizerCallbackLimits } from "./external-optimizer-subprocess-wBWeoG6A.js";
4
4
  import { createServer } from "node:http";
5
5
  //#region src/campaign/single-run-lock.ts
6
6
  /**
@@ -291,4 +291,4 @@ function cloneCandidate(candidate) {
291
291
  //#endregion
292
292
  export { acquireSingleRunLock as n, startExternalOptimizerCallback as t };
293
293
 
294
- //# sourceMappingURL=external-optimizer-process-Cq_Pg15r.js.map
294
+ //# sourceMappingURL=external-optimizer-process-BwITA9Jp.js.map
@@ -1 +1 @@
1
- {"version":3,"file":"external-optimizer-process-Cq_Pg15r.js","names":[],"sources":["../src/campaign/single-run-lock.ts","../src/campaign/external-optimizer-callback.ts"],"sourcesContent":["/**\n * Single-run lock for evaluations that share one mutable environment.\n *\n * Two concurrent runs against a shared stateful gym silently corrupt each\n * other: each resets/mutates environment state mid-cell of the other, and\n * every score from both becomes garbage that LOOKS like worker variance\n * (agent-lab R357 burned hours on flip-flopping scores before tracing them\n * to exactly this). The fix is a pid lockfile: refuse to start while a live\n * holder exists, reclaim stale locks whose pid is gone, release only if the\n * lock is still ours.\n *\n * `alsoCheck` exists because independent runners can guard the same shared\n * resource with differently named lockfiles; a runner must respect all of\n * them even though it writes only its own.\n */\n\nimport {\n type AtomicFileLockUnavailable,\n probeAtomicFileLock,\n tryAcquireAtomicFileLock,\n} from '../ledger-core/atomic-file-lock'\n\nexport interface SingleRunLockOptions {\n /** Lockfile this runner writes (and checks). */\n readonly lockPath: string\n /** Other runners' lockfiles guarding the same resource; checked, never written. */\n readonly alsoCheck?: readonly string[]\n /** Install a process 'exit' hook that releases the lock. Default true. */\n readonly releaseOnExit?: boolean\n /** Owner pid recorded in the lockfile metadata. Default process.pid. */\n readonly pid?: number\n}\n\nexport interface SingleRunLock {\n /** Remove the lockfile if this process still owns it. Idempotent. */\n release(): void\n}\n\nfunction assertAvailable(path: string): void {\n const unavailable = probeAtomicFileLock({ lockPath: path })\n if (unavailable) throw unavailableError(path, unavailable)\n}\n\nfunction unavailableError(path: string, unavailable: AtomicFileLockUnavailable): Error {\n if (unavailable.reason === 'recovery') {\n return new Error(\n `single-run lock recovery is already in progress (${path}); refusing a concurrent run`,\n )\n }\n return new Error(\n `single-run lock held by live pid ${unavailable.holder.pid} (${path}); refusing a concurrent run on the shared resource`,\n )\n}\n\n/**\n * Acquire the lock or throw naming the live holder. A stale lock (holder pid\n * no longer running) is reclaimed by one contender. An interrupted reclaim\n * leaves a marker that fails closed instead of admitting overlapping runs.\n */\nexport function acquireSingleRunLock(opts: SingleRunLockOptions): SingleRunLock {\n const pid = opts.pid ?? process.pid\n if (!Number.isSafeInteger(pid) || pid <= 0)\n throw new Error('single-run lock pid must be a positive integer')\n for (const path of opts.alsoCheck ?? []) assertAvailable(path)\n const acquisition = tryAcquireAtomicFileLock({\n lockPath: opts.lockPath,\n pid,\n })\n if (!acquisition.acquired) throw unavailableError(opts.lockPath, acquisition)\n const releaseOnExit = opts.releaseOnExit ?? true\n let released = false\n const release = (): void => {\n if (released) return\n released = true\n if (releaseOnExit) process.off('exit', release)\n try {\n acquisition.lock.release()\n } catch {\n // release is best-effort; a leftover stale lock is reclaimed on next acquire\n }\n }\n try {\n for (const path of opts.alsoCheck ?? []) assertAvailable(path)\n } catch (error) {\n release()\n throw error\n }\n if (releaseOnExit) process.on('exit', release)\n return { release }\n}\n","import { createServer, type IncomingMessage, type ServerResponse } from 'node:http'\nimport { contentHash } from '../verdict-cache'\nimport {\n type ExternalOptimizerCallback,\n type ExternalOptimizerCallbackLimits,\n type ExternalOptimizerEvaluationObservation,\n type ExternalTextCandidate,\n type ExternalTextEvaluationRequest,\n isExternalTextCandidate,\n isRecord,\n resolveExternalOptimizerCallbackLimits,\n} from './external-optimizer-contracts'\nimport {\n closeServer,\n listenLocal,\n sendJsonIfOpen,\n waitForActiveHandlers,\n} from './external-optimizer-http'\n\ntype UnsequencedObservation = ExternalOptimizerEvaluationObservation extends infer T\n ? T extends ExternalOptimizerEvaluationObservation\n ? Omit<T, 'sequence'>\n : never\n : never\n\nexport async function startExternalOptimizerCallback<TResponse>(args: {\n token: string\n maxEvaluations: number\n acceptEvaluation?: () => number | undefined\n evaluate: (request: ExternalTextEvaluationRequest, signal: AbortSignal) => Promise<TResponse>\n observe?: (observation: ExternalOptimizerEvaluationObservation) => void\n /** Loopback JSON byte limits. Omitted fields use finite defaults. */\n limits?: Partial<ExternalOptimizerCallbackLimits>\n signal?: AbortSignal\n}): Promise<ExternalOptimizerCallback> {\n assertCallbackConfig(args)\n const limits = resolveExternalOptimizerCallbackLimits(args.limits)\n args.signal?.throwIfAborted()\n let evaluations = 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 proposedCandidateHashes = new Set<string>()\n let observationSequence = 0\n const server = createServer((request, response) => {\n if (!accepting) {\n sendJsonIfOpen(response, 503, { error: 'external optimizer callback is closing' })\n return\n }\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 = handleCallback(\n request,\n response,\n controller.signal,\n args,\n limits,\n () => {\n const accepted = args.acceptEvaluation ? args.acceptEvaluation() : evaluations + 1\n if (accepted === undefined) return undefined\n if (!Number.isSafeInteger(accepted) || accepted <= 0) {\n throw new Error('external optimizer callback: invalid accepted evaluation count')\n }\n if (accepted > args.maxEvaluations) return undefined\n evaluations += 1\n return accepted\n },\n (observation) => {\n if (!args.observe) return\n args.observe({ ...observation, sequence: ++observationSequence })\n },\n proposedCandidateHashes,\n ).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 ??= closeCallbackServer()\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 return {\n url: `http://127.0.0.1:${port}/evaluate`,\n token: args.token,\n evaluations: () => evaluations,\n close,\n }\n\n async function closeCallbackServer(): 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('external optimizer callback closed with active request work')\n }\n if (serverResult?.status === 'rejected') throw serverResult.reason\n }\n}\n\nasync function handleCallback<TResponse>(\n request: IncomingMessage,\n response: ServerResponse,\n signal: AbortSignal,\n args: {\n token: string\n maxEvaluations: number\n acceptEvaluation?: () => number | undefined\n evaluate: (request: ExternalTextEvaluationRequest, signal: AbortSignal) => Promise<TResponse>\n observe?: (observation: ExternalOptimizerEvaluationObservation) => void\n },\n limits: ExternalOptimizerCallbackLimits,\n nextEvaluation: () => number | undefined,\n observe: (observation: UnsequencedObservation) => void,\n proposedCandidateHashes: Set<string>,\n): Promise<void> {\n try {\n if (request.method !== 'POST' || request.url !== '/evaluate') {\n sendJsonIfOpen(response, 404, { error: 'not found' })\n return\n }\n if (request.headers.authorization !== `Bearer ${args.token}`) {\n sendJsonIfOpen(response, 401, { error: 'unauthorized' })\n return\n }\n let body: unknown\n try {\n body = await readJson(request, limits.maxRequestBytes)\n } catch {\n observe({ kind: 'refusal', reason: 'invalid-request' })\n sendJsonIfOpen(response, 400, { error: 'request body must be bounded valid JSON' })\n return\n }\n if (\n !isRecord(body) ||\n !isExternalTextCandidate(body.candidate) ||\n typeof body.exampleId !== 'string'\n ) {\n observe({ kind: 'refusal', reason: 'invalid-request' })\n sendJsonIfOpen(response, 400, { error: 'candidate and exampleId are required strings' })\n return\n }\n const candidate = cloneCandidate(body.candidate)\n const candidateHash = contentHash({ kind: 'external-text-candidate', candidate })\n if (!proposedCandidateHashes.has(candidateHash)) {\n observe({ kind: 'proposal', candidate, candidateHash })\n proposedCandidateHashes.add(candidateHash)\n }\n const count = nextEvaluation()\n if (count === undefined) {\n observe({\n kind: 'refusal',\n reason: 'evaluation-limit',\n candidate,\n candidateHash,\n exampleId: body.exampleId,\n })\n sendJsonIfOpen(response, 429, { error: 'evaluation limit reached' })\n return\n }\n let result: TResponse\n try {\n result = await args.evaluate({ candidate, exampleId: body.exampleId }, signal)\n } catch (error) {\n observe({\n kind: 'refusal',\n reason: 'evaluation-failed',\n candidate,\n candidateHash,\n exampleId: body.exampleId,\n })\n // The thrown detail is the only diagnostic for a failed evaluation —\n // an opaque 500 costs the caller a blind debugging round.\n sendJsonIfOpen(response, 500, {\n error: `evaluation failed: ${String(error).slice(0, 400)}`,\n })\n return\n }\n let encoded: string\n try {\n encoded = encodeJsonResponse(result)\n } catch {\n observe({\n kind: 'refusal',\n reason: 'evaluation-failed',\n candidate,\n candidateHash,\n exampleId: body.exampleId,\n })\n sendJsonIfOpen(response, 500, { error: 'evaluation response must be valid JSON' })\n return\n }\n if (Buffer.byteLength(encoded) > limits.maxResponseBytes) {\n observe({\n kind: 'refusal',\n reason: 'evaluation-failed',\n candidate,\n candidateHash,\n exampleId: body.exampleId,\n })\n sendJsonIfOpen(response, 413, { error: 'evaluation response too large' })\n return\n }\n observe({\n kind: 'evaluation',\n candidate,\n candidateHash,\n exampleId: body.exampleId,\n evaluationNumber: count,\n response: structuredClone(result),\n })\n sendEncodedJsonIfOpen(response, 200, encoded)\n } catch {\n sendJsonIfOpen(response, 500, { error: 'evaluation failed' })\n }\n}\n\nfunction readJson(request: IncomingMessage, maxRequestBytes: number): Promise<unknown> {\n return new Promise((resolvePromise, 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('callback body 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 resolvePromise(JSON.parse(Buffer.concat(chunks).toString('utf8')))\n } catch (error) {\n reject(error)\n }\n })\n })\n}\n\nfunction encodeJsonResponse(value: unknown): string {\n const encoded = JSON.stringify(value)\n if (encoded === undefined) throw new Error('evaluation response must be JSON-serializable')\n return encoded\n}\n\nfunction sendEncodedJsonIfOpen(response: ServerResponse, status: number, encoded: string): void {\n if (response.destroyed || response.writableEnded) return\n response.writeHead(status, {\n 'content-type': 'application/json; charset=utf-8',\n 'content-length': String(Buffer.byteLength(encoded)),\n })\n response.end(encoded)\n}\n\nfunction assertCallbackConfig(args: {\n token: string\n maxEvaluations: number\n acceptEvaluation?: () => number | undefined\n evaluate: (request: ExternalTextEvaluationRequest, signal: AbortSignal) => Promise<unknown>\n observe?: (observation: ExternalOptimizerEvaluationObservation) => void\n limits?: Partial<ExternalOptimizerCallbackLimits>\n signal?: AbortSignal\n}): void {\n if (typeof args.token !== 'string' || !args.token.trim()) {\n throw new Error('external optimizer callback: token must be non-empty')\n }\n if (!Number.isSafeInteger(args.maxEvaluations) || args.maxEvaluations <= 0) {\n throw new Error('external optimizer callback: maxEvaluations must be a positive safe integer')\n }\n if (args.acceptEvaluation !== undefined && typeof args.acceptEvaluation !== 'function') {\n throw new Error('external optimizer callback: acceptEvaluation must be a function')\n }\n if (typeof args.evaluate !== 'function') {\n throw new Error('external optimizer callback: evaluate must be a function')\n }\n if (args.observe !== undefined && typeof args.observe !== 'function') {\n throw new Error('external optimizer callback: observe must be a function')\n }\n}\n\nfunction cloneCandidate(candidate: ExternalTextCandidate): ExternalTextCandidate {\n return typeof candidate === 'string' ? candidate : { ...candidate }\n}\n"],"mappings":";;;;;;;;;;;;;;;;;;;;AAsCA,SAAS,gBAAgB,MAAoB;CAC3C,MAAM,cAAc,oBAAoB,EAAE,UAAU,KAAK,CAAC;CAC1D,IAAI,aAAa,MAAM,iBAAiB,MAAM,WAAW;AAC3D;AAEA,SAAS,iBAAiB,MAAc,aAA+C;CACrF,IAAI,YAAY,WAAW,YACzB,uBAAO,IAAI,MACT,oDAAoD,KAAK,6BAC3D;CAEF,uBAAO,IAAI,MACT,oCAAoC,YAAY,OAAO,IAAI,IAAI,KAAK,oDACtE;AACF;;;;;;AAOA,SAAgB,qBAAqB,MAA2C;CAC9E,MAAM,MAAM,KAAK,OAAO,QAAQ;CAChC,IAAI,CAAC,OAAO,cAAc,GAAG,KAAK,OAAO,GACvC,MAAM,IAAI,MAAM,gDAAgD;CAClE,KAAK,MAAM,QAAQ,KAAK,aAAa,CAAC,GAAG,gBAAgB,IAAI;CAC7D,MAAM,cAAc,yBAAyB;EAC3C,UAAU,KAAK;EACf;CACF,CAAC;CACD,IAAI,CAAC,YAAY,UAAU,MAAM,iBAAiB,KAAK,UAAU,WAAW;CAC5E,MAAM,gBAAgB,KAAK,iBAAiB;CAC5C,IAAI,WAAW;CACf,MAAM,gBAAsB;EAC1B,IAAI,UAAU;EACd,WAAW;EACX,IAAI,eAAe,QAAQ,IAAI,QAAQ,OAAO;EAC9C,IAAI;GACF,YAAY,KAAK,QAAQ;EAC3B,QAAQ,CAER;CACF;CACA,IAAI;EACF,KAAK,MAAM,QAAQ,KAAK,aAAa,CAAC,GAAG,gBAAgB,IAAI;CAC/D,SAAS,OAAO;EACd,QAAQ;EACR,MAAM;CACR;CACA,IAAI,eAAe,QAAQ,GAAG,QAAQ,OAAO;CAC7C,OAAO,EAAE,QAAQ;AACnB;;;AChEA,eAAsB,+BAA0C,MASzB;CACrC,qBAAqB,IAAI;CACzB,MAAM,SAAS,uCAAuC,KAAK,MAAM;CACjE,KAAK,QAAQ,eAAe;CAC5B,IAAI,cAAc;CAClB,IAAI,YAAY;CAChB,IAAI;CACJ,MAAM,oCAAoB,IAAI,IAAqB;CACnD,MAAM,iCAAiB,IAAI,IAAmB;CAC9C,MAAM,0CAA0B,IAAI,IAAY;CAChD,IAAI,sBAAsB;CAC1B,MAAM,SAAS,cAAc,SAAS,aAAa;EACjD,IAAI,CAAC,WAAW;GACd,eAAe,UAAU,KAAK,EAAE,OAAO,yCAAyC,CAAC;GACjF;EACF;EAEA,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,eACR,SACA,UACA,WAAW,QACX,MACA,cACM;GACJ,MAAM,WAAW,KAAK,mBAAmB,KAAK,iBAAiB,IAAI,cAAc;GACjF,IAAI,aAAa,KAAA,GAAW,OAAO,KAAA;GACnC,IAAI,CAAC,OAAO,cAAc,QAAQ,KAAK,YAAY,GACjD,MAAM,IAAI,MAAM,gEAAgE;GAElF,IAAI,WAAW,KAAK,gBAAgB,OAAO,KAAA;GAC3C,eAAe;GACf,OAAO;EACT,IACC,gBAAgB;GACf,IAAI,CAAC,KAAK,SAAS;GACnB,KAAK,QAAQ;IAAE,GAAG;IAAa,UAAU,EAAE;GAAoB,CAAC;EAClE,GACA,uBACF,CAAC,CAAC,cAAc;GACd,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,oBAAoB;EACrC,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;CAClC,OAAO;EACL,KAAK,oBAAoB,KAAK;EAC9B,OAAO,KAAK;EACZ,mBAAmB;EACnB;CACF;CAEA,eAAe,sBAAqC;EAClD,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,6DAA6D;EAE/E,IAAI,cAAc,WAAW,YAAY,MAAM,aAAa;CAC9D;AACF;AAEA,eAAe,eACb,SACA,UACA,QACA,MAOA,QACA,gBACA,SACA,yBACe;CACf,IAAI;EACF,IAAI,QAAQ,WAAW,UAAU,QAAQ,QAAQ,aAAa;GAC5D,eAAe,UAAU,KAAK,EAAE,OAAO,YAAY,CAAC;GACpD;EACF;EACA,IAAI,QAAQ,QAAQ,kBAAkB,UAAU,KAAK,SAAS;GAC5D,eAAe,UAAU,KAAK,EAAE,OAAO,eAAe,CAAC;GACvD;EACF;EACA,IAAI;EACJ,IAAI;GACF,OAAO,MAAM,SAAS,SAAS,OAAO,eAAe;EACvD,QAAQ;GACN,QAAQ;IAAE,MAAM;IAAW,QAAQ;GAAkB,CAAC;GACtD,eAAe,UAAU,KAAK,EAAE,OAAO,0CAA0C,CAAC;GAClF;EACF;EACA,IACE,CAAC,SAAS,IAAI,KACd,CAAC,wBAAwB,KAAK,SAAS,KACvC,OAAO,KAAK,cAAc,UAC1B;GACA,QAAQ;IAAE,MAAM;IAAW,QAAQ;GAAkB,CAAC;GACtD,eAAe,UAAU,KAAK,EAAE,OAAO,+CAA+C,CAAC;GACvF;EACF;EACA,MAAM,YAAY,eAAe,KAAK,SAAS;EAC/C,MAAM,gBAAgB,YAAY;GAAE,MAAM;GAA2B;EAAU,CAAC;EAChF,IAAI,CAAC,wBAAwB,IAAI,aAAa,GAAG;GAC/C,QAAQ;IAAE,MAAM;IAAY;IAAW;GAAc,CAAC;GACtD,wBAAwB,IAAI,aAAa;EAC3C;EACA,MAAM,QAAQ,eAAe;EAC7B,IAAI,UAAU,KAAA,GAAW;GACvB,QAAQ;IACN,MAAM;IACN,QAAQ;IACR;IACA;IACA,WAAW,KAAK;GAClB,CAAC;GACD,eAAe,UAAU,KAAK,EAAE,OAAO,2BAA2B,CAAC;GACnE;EACF;EACA,IAAI;EACJ,IAAI;GACF,SAAS,MAAM,KAAK,SAAS;IAAE;IAAW,WAAW,KAAK;GAAU,GAAG,MAAM;EAC/E,SAAS,OAAO;GACd,QAAQ;IACN,MAAM;IACN,QAAQ;IACR;IACA;IACA,WAAW,KAAK;GAClB,CAAC;GAGD,eAAe,UAAU,KAAK,EAC5B,OAAO,sBAAsB,OAAO,KAAK,CAAC,CAAC,MAAM,GAAG,GAAG,IACzD,CAAC;GACD;EACF;EACA,IAAI;EACJ,IAAI;GACF,UAAU,mBAAmB,MAAM;EACrC,QAAQ;GACN,QAAQ;IACN,MAAM;IACN,QAAQ;IACR;IACA;IACA,WAAW,KAAK;GAClB,CAAC;GACD,eAAe,UAAU,KAAK,EAAE,OAAO,yCAAyC,CAAC;GACjF;EACF;EACA,IAAI,OAAO,WAAW,OAAO,IAAI,OAAO,kBAAkB;GACxD,QAAQ;IACN,MAAM;IACN,QAAQ;IACR;IACA;IACA,WAAW,KAAK;GAClB,CAAC;GACD,eAAe,UAAU,KAAK,EAAE,OAAO,gCAAgC,CAAC;GACxE;EACF;EACA,QAAQ;GACN,MAAM;GACN;GACA;GACA,WAAW,KAAK;GAChB,kBAAkB;GAClB,UAAU,gBAAgB,MAAM;EAClC,CAAC;EACD,sBAAsB,UAAU,KAAK,OAAO;CAC9C,QAAQ;EACN,eAAe,UAAU,KAAK,EAAE,OAAO,oBAAoB,CAAC;CAC9D;AACF;AAEA,SAAS,SAAS,SAA0B,iBAA2C;CACrF,OAAO,IAAI,SAAS,gBAAgB,WAAW;EAC7C,IAAI,OAAO;EACX,MAAM,SAAmB,CAAC;EAC1B,QAAQ,GAAG,SAAS,UAAkB;GACpC,QAAQ,MAAM;GACd,IAAI,OAAO,iBAAiB;IAC1B,uBAAO,IAAI,MAAM,yBAAyB,CAAC;IAC3C,QAAQ,QAAQ;IAChB;GACF;GACA,OAAO,KAAK,KAAK;EACnB,CAAC;EACD,QAAQ,GAAG,SAAS,MAAM;EAC1B,QAAQ,GAAG,aAAa;GACtB,IAAI;IACF,eAAe,KAAK,MAAM,OAAO,OAAO,MAAM,CAAC,CAAC,SAAS,MAAM,CAAC,CAAC;GACnE,SAAS,OAAO;IACd,OAAO,KAAK;GACd;EACF,CAAC;CACH,CAAC;AACH;AAEA,SAAS,mBAAmB,OAAwB;CAClD,MAAM,UAAU,KAAK,UAAU,KAAK;CACpC,IAAI,YAAY,KAAA,GAAW,MAAM,IAAI,MAAM,+CAA+C;CAC1F,OAAO;AACT;AAEA,SAAS,sBAAsB,UAA0B,QAAgB,SAAuB;CAC9F,IAAI,SAAS,aAAa,SAAS,eAAe;CAClD,SAAS,UAAU,QAAQ;EACzB,gBAAgB;EAChB,kBAAkB,OAAO,OAAO,WAAW,OAAO,CAAC;CACrD,CAAC;CACD,SAAS,IAAI,OAAO;AACtB;AAEA,SAAS,qBAAqB,MAQrB;CACP,IAAI,OAAO,KAAK,UAAU,YAAY,CAAC,KAAK,MAAM,KAAK,GACrD,MAAM,IAAI,MAAM,sDAAsD;CAExE,IAAI,CAAC,OAAO,cAAc,KAAK,cAAc,KAAK,KAAK,kBAAkB,GACvE,MAAM,IAAI,MAAM,6EAA6E;CAE/F,IAAI,KAAK,qBAAqB,KAAA,KAAa,OAAO,KAAK,qBAAqB,YAC1E,MAAM,IAAI,MAAM,kEAAkE;CAEpF,IAAI,OAAO,KAAK,aAAa,YAC3B,MAAM,IAAI,MAAM,0DAA0D;CAE5E,IAAI,KAAK,YAAY,KAAA,KAAa,OAAO,KAAK,YAAY,YACxD,MAAM,IAAI,MAAM,yDAAyD;AAE7E;AAEA,SAAS,eAAe,WAAyD;CAC/E,OAAO,OAAO,cAAc,WAAW,YAAY,EAAE,GAAG,UAAU;AACpE"}
1
+ {"version":3,"file":"external-optimizer-process-BwITA9Jp.js","names":[],"sources":["../src/campaign/single-run-lock.ts","../src/campaign/external-optimizer-callback.ts"],"sourcesContent":["/**\n * Single-run lock for evaluations that share one mutable environment.\n *\n * Two concurrent runs against a shared stateful gym silently corrupt each\n * other: each resets/mutates environment state mid-cell of the other, and\n * every score from both becomes garbage that LOOKS like worker variance\n * (agent-lab R357 burned hours on flip-flopping scores before tracing them\n * to exactly this). The fix is a pid lockfile: refuse to start while a live\n * holder exists, reclaim stale locks whose pid is gone, release only if the\n * lock is still ours.\n *\n * `alsoCheck` exists because independent runners can guard the same shared\n * resource with differently named lockfiles; a runner must respect all of\n * them even though it writes only its own.\n */\n\nimport {\n type AtomicFileLockUnavailable,\n probeAtomicFileLock,\n tryAcquireAtomicFileLock,\n} from '../ledger-core/atomic-file-lock'\n\nexport interface SingleRunLockOptions {\n /** Lockfile this runner writes (and checks). */\n readonly lockPath: string\n /** Other runners' lockfiles guarding the same resource; checked, never written. */\n readonly alsoCheck?: readonly string[]\n /** Install a process 'exit' hook that releases the lock. Default true. */\n readonly releaseOnExit?: boolean\n /** Owner pid recorded in the lockfile metadata. Default process.pid. */\n readonly pid?: number\n}\n\nexport interface SingleRunLock {\n /** Remove the lockfile if this process still owns it. Idempotent. */\n release(): void\n}\n\nfunction assertAvailable(path: string): void {\n const unavailable = probeAtomicFileLock({ lockPath: path })\n if (unavailable) throw unavailableError(path, unavailable)\n}\n\nfunction unavailableError(path: string, unavailable: AtomicFileLockUnavailable): Error {\n if (unavailable.reason === 'recovery') {\n return new Error(\n `single-run lock recovery is already in progress (${path}); refusing a concurrent run`,\n )\n }\n return new Error(\n `single-run lock held by live pid ${unavailable.holder.pid} (${path}); refusing a concurrent run on the shared resource`,\n )\n}\n\n/**\n * Acquire the lock or throw naming the live holder. A stale lock (holder pid\n * no longer running) is reclaimed by one contender. An interrupted reclaim\n * leaves a marker that fails closed instead of admitting overlapping runs.\n */\nexport function acquireSingleRunLock(opts: SingleRunLockOptions): SingleRunLock {\n const pid = opts.pid ?? process.pid\n if (!Number.isSafeInteger(pid) || pid <= 0)\n throw new Error('single-run lock pid must be a positive integer')\n for (const path of opts.alsoCheck ?? []) assertAvailable(path)\n const acquisition = tryAcquireAtomicFileLock({\n lockPath: opts.lockPath,\n pid,\n })\n if (!acquisition.acquired) throw unavailableError(opts.lockPath, acquisition)\n const releaseOnExit = opts.releaseOnExit ?? true\n let released = false\n const release = (): void => {\n if (released) return\n released = true\n if (releaseOnExit) process.off('exit', release)\n try {\n acquisition.lock.release()\n } catch {\n // release is best-effort; a leftover stale lock is reclaimed on next acquire\n }\n }\n try {\n for (const path of opts.alsoCheck ?? []) assertAvailable(path)\n } catch (error) {\n release()\n throw error\n }\n if (releaseOnExit) process.on('exit', release)\n return { release }\n}\n","import { createServer, type IncomingMessage, type ServerResponse } from 'node:http'\nimport { contentHash } from '../verdict-cache'\nimport {\n type ExternalOptimizerCallback,\n type ExternalOptimizerCallbackLimits,\n type ExternalOptimizerEvaluationObservation,\n type ExternalTextCandidate,\n type ExternalTextEvaluationRequest,\n isExternalTextCandidate,\n isRecord,\n resolveExternalOptimizerCallbackLimits,\n} from './external-optimizer-contracts'\nimport {\n closeServer,\n listenLocal,\n sendJsonIfOpen,\n waitForActiveHandlers,\n} from './external-optimizer-http'\n\ntype UnsequencedObservation = ExternalOptimizerEvaluationObservation extends infer T\n ? T extends ExternalOptimizerEvaluationObservation\n ? Omit<T, 'sequence'>\n : never\n : never\n\nexport async function startExternalOptimizerCallback<TResponse>(args: {\n token: string\n maxEvaluations: number\n acceptEvaluation?: () => number | undefined\n evaluate: (request: ExternalTextEvaluationRequest, signal: AbortSignal) => Promise<TResponse>\n observe?: (observation: ExternalOptimizerEvaluationObservation) => void\n /** Loopback JSON byte limits. Omitted fields use finite defaults. */\n limits?: Partial<ExternalOptimizerCallbackLimits>\n signal?: AbortSignal\n}): Promise<ExternalOptimizerCallback> {\n assertCallbackConfig(args)\n const limits = resolveExternalOptimizerCallbackLimits(args.limits)\n args.signal?.throwIfAborted()\n let evaluations = 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 proposedCandidateHashes = new Set<string>()\n let observationSequence = 0\n const server = createServer((request, response) => {\n if (!accepting) {\n sendJsonIfOpen(response, 503, { error: 'external optimizer callback is closing' })\n return\n }\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 = handleCallback(\n request,\n response,\n controller.signal,\n args,\n limits,\n () => {\n const accepted = args.acceptEvaluation ? args.acceptEvaluation() : evaluations + 1\n if (accepted === undefined) return undefined\n if (!Number.isSafeInteger(accepted) || accepted <= 0) {\n throw new Error('external optimizer callback: invalid accepted evaluation count')\n }\n if (accepted > args.maxEvaluations) return undefined\n evaluations += 1\n return accepted\n },\n (observation) => {\n if (!args.observe) return\n args.observe({ ...observation, sequence: ++observationSequence })\n },\n proposedCandidateHashes,\n ).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 ??= closeCallbackServer()\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 return {\n url: `http://127.0.0.1:${port}/evaluate`,\n token: args.token,\n evaluations: () => evaluations,\n close,\n }\n\n async function closeCallbackServer(): 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('external optimizer callback closed with active request work')\n }\n if (serverResult?.status === 'rejected') throw serverResult.reason\n }\n}\n\nasync function handleCallback<TResponse>(\n request: IncomingMessage,\n response: ServerResponse,\n signal: AbortSignal,\n args: {\n token: string\n maxEvaluations: number\n acceptEvaluation?: () => number | undefined\n evaluate: (request: ExternalTextEvaluationRequest, signal: AbortSignal) => Promise<TResponse>\n observe?: (observation: ExternalOptimizerEvaluationObservation) => void\n },\n limits: ExternalOptimizerCallbackLimits,\n nextEvaluation: () => number | undefined,\n observe: (observation: UnsequencedObservation) => void,\n proposedCandidateHashes: Set<string>,\n): Promise<void> {\n try {\n if (request.method !== 'POST' || request.url !== '/evaluate') {\n sendJsonIfOpen(response, 404, { error: 'not found' })\n return\n }\n if (request.headers.authorization !== `Bearer ${args.token}`) {\n sendJsonIfOpen(response, 401, { error: 'unauthorized' })\n return\n }\n let body: unknown\n try {\n body = await readJson(request, limits.maxRequestBytes)\n } catch {\n observe({ kind: 'refusal', reason: 'invalid-request' })\n sendJsonIfOpen(response, 400, { error: 'request body must be bounded valid JSON' })\n return\n }\n if (\n !isRecord(body) ||\n !isExternalTextCandidate(body.candidate) ||\n typeof body.exampleId !== 'string'\n ) {\n observe({ kind: 'refusal', reason: 'invalid-request' })\n sendJsonIfOpen(response, 400, { error: 'candidate and exampleId are required strings' })\n return\n }\n const candidate = cloneCandidate(body.candidate)\n const candidateHash = contentHash({ kind: 'external-text-candidate', candidate })\n if (!proposedCandidateHashes.has(candidateHash)) {\n observe({ kind: 'proposal', candidate, candidateHash })\n proposedCandidateHashes.add(candidateHash)\n }\n const count = nextEvaluation()\n if (count === undefined) {\n observe({\n kind: 'refusal',\n reason: 'evaluation-limit',\n candidate,\n candidateHash,\n exampleId: body.exampleId,\n })\n sendJsonIfOpen(response, 429, { error: 'evaluation limit reached' })\n return\n }\n let result: TResponse\n try {\n result = await args.evaluate({ candidate, exampleId: body.exampleId }, signal)\n } catch (error) {\n observe({\n kind: 'refusal',\n reason: 'evaluation-failed',\n candidate,\n candidateHash,\n exampleId: body.exampleId,\n })\n // The thrown detail is the only diagnostic for a failed evaluation —\n // an opaque 500 costs the caller a blind debugging round.\n sendJsonIfOpen(response, 500, {\n error: `evaluation failed: ${String(error).slice(0, 400)}`,\n })\n return\n }\n let encoded: string\n try {\n encoded = encodeJsonResponse(result)\n } catch {\n observe({\n kind: 'refusal',\n reason: 'evaluation-failed',\n candidate,\n candidateHash,\n exampleId: body.exampleId,\n })\n sendJsonIfOpen(response, 500, { error: 'evaluation response must be valid JSON' })\n return\n }\n if (Buffer.byteLength(encoded) > limits.maxResponseBytes) {\n observe({\n kind: 'refusal',\n reason: 'evaluation-failed',\n candidate,\n candidateHash,\n exampleId: body.exampleId,\n })\n sendJsonIfOpen(response, 413, { error: 'evaluation response too large' })\n return\n }\n observe({\n kind: 'evaluation',\n candidate,\n candidateHash,\n exampleId: body.exampleId,\n evaluationNumber: count,\n response: structuredClone(result),\n })\n sendEncodedJsonIfOpen(response, 200, encoded)\n } catch {\n sendJsonIfOpen(response, 500, { error: 'evaluation failed' })\n }\n}\n\nfunction readJson(request: IncomingMessage, maxRequestBytes: number): Promise<unknown> {\n return new Promise((resolvePromise, 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('callback body 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 resolvePromise(JSON.parse(Buffer.concat(chunks).toString('utf8')))\n } catch (error) {\n reject(error)\n }\n })\n })\n}\n\nfunction encodeJsonResponse(value: unknown): string {\n const encoded = JSON.stringify(value)\n if (encoded === undefined) throw new Error('evaluation response must be JSON-serializable')\n return encoded\n}\n\nfunction sendEncodedJsonIfOpen(response: ServerResponse, status: number, encoded: string): void {\n if (response.destroyed || response.writableEnded) return\n response.writeHead(status, {\n 'content-type': 'application/json; charset=utf-8',\n 'content-length': String(Buffer.byteLength(encoded)),\n })\n response.end(encoded)\n}\n\nfunction assertCallbackConfig(args: {\n token: string\n maxEvaluations: number\n acceptEvaluation?: () => number | undefined\n evaluate: (request: ExternalTextEvaluationRequest, signal: AbortSignal) => Promise<unknown>\n observe?: (observation: ExternalOptimizerEvaluationObservation) => void\n limits?: Partial<ExternalOptimizerCallbackLimits>\n signal?: AbortSignal\n}): void {\n if (typeof args.token !== 'string' || !args.token.trim()) {\n throw new Error('external optimizer callback: token must be non-empty')\n }\n if (!Number.isSafeInteger(args.maxEvaluations) || args.maxEvaluations <= 0) {\n throw new Error('external optimizer callback: maxEvaluations must be a positive safe integer')\n }\n if (args.acceptEvaluation !== undefined && typeof args.acceptEvaluation !== 'function') {\n throw new Error('external optimizer callback: acceptEvaluation must be a function')\n }\n if (typeof args.evaluate !== 'function') {\n throw new Error('external optimizer callback: evaluate must be a function')\n }\n if (args.observe !== undefined && typeof args.observe !== 'function') {\n throw new Error('external optimizer callback: observe must be a function')\n }\n}\n\nfunction cloneCandidate(candidate: ExternalTextCandidate): ExternalTextCandidate {\n return typeof candidate === 'string' ? candidate : { ...candidate }\n}\n"],"mappings":";;;;;;;;;;;;;;;;;;;;AAsCA,SAAS,gBAAgB,MAAoB;CAC3C,MAAM,cAAc,oBAAoB,EAAE,UAAU,KAAK,CAAC;CAC1D,IAAI,aAAa,MAAM,iBAAiB,MAAM,WAAW;AAC3D;AAEA,SAAS,iBAAiB,MAAc,aAA+C;CACrF,IAAI,YAAY,WAAW,YACzB,uBAAO,IAAI,MACT,oDAAoD,KAAK,6BAC3D;CAEF,uBAAO,IAAI,MACT,oCAAoC,YAAY,OAAO,IAAI,IAAI,KAAK,oDACtE;AACF;;;;;;AAOA,SAAgB,qBAAqB,MAA2C;CAC9E,MAAM,MAAM,KAAK,OAAO,QAAQ;CAChC,IAAI,CAAC,OAAO,cAAc,GAAG,KAAK,OAAO,GACvC,MAAM,IAAI,MAAM,gDAAgD;CAClE,KAAK,MAAM,QAAQ,KAAK,aAAa,CAAC,GAAG,gBAAgB,IAAI;CAC7D,MAAM,cAAc,yBAAyB;EAC3C,UAAU,KAAK;EACf;CACF,CAAC;CACD,IAAI,CAAC,YAAY,UAAU,MAAM,iBAAiB,KAAK,UAAU,WAAW;CAC5E,MAAM,gBAAgB,KAAK,iBAAiB;CAC5C,IAAI,WAAW;CACf,MAAM,gBAAsB;EAC1B,IAAI,UAAU;EACd,WAAW;EACX,IAAI,eAAe,QAAQ,IAAI,QAAQ,OAAO;EAC9C,IAAI;GACF,YAAY,KAAK,QAAQ;EAC3B,QAAQ,CAER;CACF;CACA,IAAI;EACF,KAAK,MAAM,QAAQ,KAAK,aAAa,CAAC,GAAG,gBAAgB,IAAI;CAC/D,SAAS,OAAO;EACd,QAAQ;EACR,MAAM;CACR;CACA,IAAI,eAAe,QAAQ,GAAG,QAAQ,OAAO;CAC7C,OAAO,EAAE,QAAQ;AACnB;;;AChEA,eAAsB,+BAA0C,MASzB;CACrC,qBAAqB,IAAI;CACzB,MAAM,SAAS,uCAAuC,KAAK,MAAM;CACjE,KAAK,QAAQ,eAAe;CAC5B,IAAI,cAAc;CAClB,IAAI,YAAY;CAChB,IAAI;CACJ,MAAM,oCAAoB,IAAI,IAAqB;CACnD,MAAM,iCAAiB,IAAI,IAAmB;CAC9C,MAAM,0CAA0B,IAAI,IAAY;CAChD,IAAI,sBAAsB;CAC1B,MAAM,SAAS,cAAc,SAAS,aAAa;EACjD,IAAI,CAAC,WAAW;GACd,eAAe,UAAU,KAAK,EAAE,OAAO,yCAAyC,CAAC;GACjF;EACF;EAEA,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,eACR,SACA,UACA,WAAW,QACX,MACA,cACM;GACJ,MAAM,WAAW,KAAK,mBAAmB,KAAK,iBAAiB,IAAI,cAAc;GACjF,IAAI,aAAa,KAAA,GAAW,OAAO,KAAA;GACnC,IAAI,CAAC,OAAO,cAAc,QAAQ,KAAK,YAAY,GACjD,MAAM,IAAI,MAAM,gEAAgE;GAElF,IAAI,WAAW,KAAK,gBAAgB,OAAO,KAAA;GAC3C,eAAe;GACf,OAAO;EACT,IACC,gBAAgB;GACf,IAAI,CAAC,KAAK,SAAS;GACnB,KAAK,QAAQ;IAAE,GAAG;IAAa,UAAU,EAAE;GAAoB,CAAC;EAClE,GACA,uBACF,CAAC,CAAC,cAAc;GACd,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,oBAAoB;EACrC,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;CAClC,OAAO;EACL,KAAK,oBAAoB,KAAK;EAC9B,OAAO,KAAK;EACZ,mBAAmB;EACnB;CACF;CAEA,eAAe,sBAAqC;EAClD,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,6DAA6D;EAE/E,IAAI,cAAc,WAAW,YAAY,MAAM,aAAa;CAC9D;AACF;AAEA,eAAe,eACb,SACA,UACA,QACA,MAOA,QACA,gBACA,SACA,yBACe;CACf,IAAI;EACF,IAAI,QAAQ,WAAW,UAAU,QAAQ,QAAQ,aAAa;GAC5D,eAAe,UAAU,KAAK,EAAE,OAAO,YAAY,CAAC;GACpD;EACF;EACA,IAAI,QAAQ,QAAQ,kBAAkB,UAAU,KAAK,SAAS;GAC5D,eAAe,UAAU,KAAK,EAAE,OAAO,eAAe,CAAC;GACvD;EACF;EACA,IAAI;EACJ,IAAI;GACF,OAAO,MAAM,SAAS,SAAS,OAAO,eAAe;EACvD,QAAQ;GACN,QAAQ;IAAE,MAAM;IAAW,QAAQ;GAAkB,CAAC;GACtD,eAAe,UAAU,KAAK,EAAE,OAAO,0CAA0C,CAAC;GAClF;EACF;EACA,IACE,CAAC,SAAS,IAAI,KACd,CAAC,wBAAwB,KAAK,SAAS,KACvC,OAAO,KAAK,cAAc,UAC1B;GACA,QAAQ;IAAE,MAAM;IAAW,QAAQ;GAAkB,CAAC;GACtD,eAAe,UAAU,KAAK,EAAE,OAAO,+CAA+C,CAAC;GACvF;EACF;EACA,MAAM,YAAY,eAAe,KAAK,SAAS;EAC/C,MAAM,gBAAgB,YAAY;GAAE,MAAM;GAA2B;EAAU,CAAC;EAChF,IAAI,CAAC,wBAAwB,IAAI,aAAa,GAAG;GAC/C,QAAQ;IAAE,MAAM;IAAY;IAAW;GAAc,CAAC;GACtD,wBAAwB,IAAI,aAAa;EAC3C;EACA,MAAM,QAAQ,eAAe;EAC7B,IAAI,UAAU,KAAA,GAAW;GACvB,QAAQ;IACN,MAAM;IACN,QAAQ;IACR;IACA;IACA,WAAW,KAAK;GAClB,CAAC;GACD,eAAe,UAAU,KAAK,EAAE,OAAO,2BAA2B,CAAC;GACnE;EACF;EACA,IAAI;EACJ,IAAI;GACF,SAAS,MAAM,KAAK,SAAS;IAAE;IAAW,WAAW,KAAK;GAAU,GAAG,MAAM;EAC/E,SAAS,OAAO;GACd,QAAQ;IACN,MAAM;IACN,QAAQ;IACR;IACA;IACA,WAAW,KAAK;GAClB,CAAC;GAGD,eAAe,UAAU,KAAK,EAC5B,OAAO,sBAAsB,OAAO,KAAK,CAAC,CAAC,MAAM,GAAG,GAAG,IACzD,CAAC;GACD;EACF;EACA,IAAI;EACJ,IAAI;GACF,UAAU,mBAAmB,MAAM;EACrC,QAAQ;GACN,QAAQ;IACN,MAAM;IACN,QAAQ;IACR;IACA;IACA,WAAW,KAAK;GAClB,CAAC;GACD,eAAe,UAAU,KAAK,EAAE,OAAO,yCAAyC,CAAC;GACjF;EACF;EACA,IAAI,OAAO,WAAW,OAAO,IAAI,OAAO,kBAAkB;GACxD,QAAQ;IACN,MAAM;IACN,QAAQ;IACR;IACA;IACA,WAAW,KAAK;GAClB,CAAC;GACD,eAAe,UAAU,KAAK,EAAE,OAAO,gCAAgC,CAAC;GACxE;EACF;EACA,QAAQ;GACN,MAAM;GACN;GACA;GACA,WAAW,KAAK;GAChB,kBAAkB;GAClB,UAAU,gBAAgB,MAAM;EAClC,CAAC;EACD,sBAAsB,UAAU,KAAK,OAAO;CAC9C,QAAQ;EACN,eAAe,UAAU,KAAK,EAAE,OAAO,oBAAoB,CAAC;CAC9D;AACF;AAEA,SAAS,SAAS,SAA0B,iBAA2C;CACrF,OAAO,IAAI,SAAS,gBAAgB,WAAW;EAC7C,IAAI,OAAO;EACX,MAAM,SAAmB,CAAC;EAC1B,QAAQ,GAAG,SAAS,UAAkB;GACpC,QAAQ,MAAM;GACd,IAAI,OAAO,iBAAiB;IAC1B,uBAAO,IAAI,MAAM,yBAAyB,CAAC;IAC3C,QAAQ,QAAQ;IAChB;GACF;GACA,OAAO,KAAK,KAAK;EACnB,CAAC;EACD,QAAQ,GAAG,SAAS,MAAM;EAC1B,QAAQ,GAAG,aAAa;GACtB,IAAI;IACF,eAAe,KAAK,MAAM,OAAO,OAAO,MAAM,CAAC,CAAC,SAAS,MAAM,CAAC,CAAC;GACnE,SAAS,OAAO;IACd,OAAO,KAAK;GACd;EACF,CAAC;CACH,CAAC;AACH;AAEA,SAAS,mBAAmB,OAAwB;CAClD,MAAM,UAAU,KAAK,UAAU,KAAK;CACpC,IAAI,YAAY,KAAA,GAAW,MAAM,IAAI,MAAM,+CAA+C;CAC1F,OAAO;AACT;AAEA,SAAS,sBAAsB,UAA0B,QAAgB,SAAuB;CAC9F,IAAI,SAAS,aAAa,SAAS,eAAe;CAClD,SAAS,UAAU,QAAQ;EACzB,gBAAgB;EAChB,kBAAkB,OAAO,OAAO,WAAW,OAAO,CAAC;CACrD,CAAC;CACD,SAAS,IAAI,OAAO;AACtB;AAEA,SAAS,qBAAqB,MAQrB;CACP,IAAI,OAAO,KAAK,UAAU,YAAY,CAAC,KAAK,MAAM,KAAK,GACrD,MAAM,IAAI,MAAM,sDAAsD;CAExE,IAAI,CAAC,OAAO,cAAc,KAAK,cAAc,KAAK,KAAK,kBAAkB,GACvE,MAAM,IAAI,MAAM,6EAA6E;CAE/F,IAAI,KAAK,qBAAqB,KAAA,KAAa,OAAO,KAAK,qBAAqB,YAC1E,MAAM,IAAI,MAAM,kEAAkE;CAEpF,IAAI,OAAO,KAAK,aAAa,YAC3B,MAAM,IAAI,MAAM,0DAA0D;CAE5E,IAAI,KAAK,YAAY,KAAA,KAAa,OAAO,KAAK,YAAY,YACxD,MAAM,IAAI,MAAM,yDAAyD;AAE7E;AAEA,SAAS,eAAe,WAAyD;CAC/E,OAAO,OAAO,cAAc,WAAW,YAAY,EAAE,GAAG,UAAU;AACpE"}
@@ -2,7 +2,7 @@ import { s as ValidationError } from "./errors-Dngq5h35.js";
2
2
  import { t as canonicalJson } from "./verdict-cache-B3eCVQtY.js";
3
3
  import { c as costForTokenPricing, i as CostLedger } from "./cost-ledger-B1qx30B4.js";
4
4
  import { a as appendLedgerLine, o as tryWithLedgerFileLock } from "./ledger-core-PIfjCbKn.js";
5
- import { f as ModelSubstitutionError, g as assertServedModelPolicy, h as assertServedModel } from "./llm-client-CxQtdtd6.js";
5
+ import { f as ModelSubstitutionError, g as assertServedModelPolicy, h as assertServedModel } from "./llm-client-CGlSi8sb.js";
6
6
  import { createRequire } from "node:module";
7
7
  import { randomBytes } from "node:crypto";
8
8
  import { isAbsolute, join, resolve } from "node:path";
@@ -1915,4 +1915,4 @@ function summarizeProcessOutput(capture, max = 4e3) {
1915
1915
  //#endregion
1916
1916
  export { inMemoryCampaignStorage as C, SearchLedgerIntegrityError as D, SearchLedgerError as E, fsCampaignStorage as S, SearchLedgerConflictError as T, removeCredentialEnvironment as _, closeServer as a, safePathComponent as b, waitForActiveHandlers as c, assertExternalOptimizerModelBudget as d, assertJsonValue as f, isRecord as g, isExternalTextCandidate as h, startExternalOptimizerModelProxy as i, DEFAULT_EXTERNAL_OPTIMIZER_CALLBACK_LIMITS as l, isCandidateText as m, closeExternalOptimizerResources as n, listenLocal as o, assertNoCredentialValues as p, runWithCleanup as r, sendJsonIfOpen as s, runExternalOptimizerProcess as t, DEFAULT_EXTERNAL_OPTIMIZER_PROCESS_LIMITS as u, resolveExternalOptimizerCallbackLimits as v, SEARCH_LEDGER_FILE_CONTEXT as w, createRunCostLedger as x, resolveExternalOptimizerProcessLimits as y };
1917
1917
 
1918
- //# sourceMappingURL=external-optimizer-subprocess-DgNebftP.js.map
1918
+ //# sourceMappingURL=external-optimizer-subprocess-wBWeoG6A.js.map