@tangle-network/agent-runtime 0.44.0 → 0.46.0

package/dist/{chunk-GFKVVRQ7.js → chunk-GN75RGM6.js}

@@ -3,22 +3,22 @@ import {
 } from "./chunk-FNMGYYSS.js";
 import {
   createDefaultCoderDelegate
-} from "./chunk-SKUZZCHE.js";
+} from "./chunk-I42NHLKX.js";
 import {
   runAnalystLoop
-} from "./chunk-XBUG326M.js";
+} from "./chunk-HNUXAZIJ.js";
 import {
-  optimizePrompt
-} from "./chunk-VOX6Z3II.js";
-import {
-  createDynamicDriver,
+  createDriver,
   runLoop
-} from "./chunk-S7JXV32P.js";
+} from "./chunk-65FQLI4V.js";
 import {
   ConfigError
-} from "./chunk-SQSCRJ7U.js";
+} from "./chunk-GSUO5QS6.js";
 
 // src/loop-runner.ts
+import {
+  selfImprove
+} from "@tangle-network/agent-eval/contract";
 var DELEGATED_LOOP_MODES = [
   "code",
   "review",
@@ -70,10 +70,11 @@ function reviewLoopRunner(options) {
 }
 function dynamicLoopRunner(o) {
   return async (signal) => runLoop({
-    driver: createDynamicDriver({
+    driver: createDriver({
       planner: o.planner,
       ...o.maxIterations !== void 0 ? { maxIterations: o.maxIterations } : {},
-      ...o.maxFanout !== void 0 ? { maxFanout: o.maxFanout } : {}
+      ...o.maxFanout !== void 0 ? { maxFanout: o.maxFanout } : {},
+      ...o.analyze ? { analyze: o.analyze } : {}
     }),
     ...o.agentRun ? { agentRun: o.agentRun } : {},
     ...o.agentRuns ? { agentRuns: o.agentRuns } : {},
@@ -108,7 +109,7 @@ function researchLoopRunner(o) {
   };
 }
 function selfImproveLoopRunner(options) {
-  return async () => optimizePrompt(options);
+  return async () => selfImprove(options);
 }
 function auditLoopRunner(options) {
   return async () => runAnalystLoop(options);
@@ -199,4 +200,4 @@ export {
   runLoopRunnerCli,
   parseLoopRunnerArgv
 };
-//# sourceMappingURL=chunk-GFKVVRQ7.js.map
+//# sourceMappingURL=chunk-GN75RGM6.js.map

package/dist/chunk-GN75RGM6.js.map

@@ -0,0 +1 @@
+{"version":3,"sources":["../src/loop-runner.ts","../src/loop-runner-bin.ts"],"sourcesContent":["/**\n * @experimental\n *\n * `runDelegatedLoop` — the configured delegated loop-runner.\n *\n * One typed entrypoint a worker agent (or a scheduled routine) calls to run a\n * disciplined loop in a chosen MODE, over agent-runtime's hardened engines:\n *\n *   code         → build-in-a-loop via the coder delegate (no-op + secret floor,\n *                  optional reviewer gate, winner-selection)\n *   review       → code mode with a REQUIRED reviewer (the gate is the point)\n *   research     → research-in-a-loop with valid-only KB growth (createKbGate)\n *   audit        → analyze trace/run data → findings (runAnalystLoop, caller-wired)\n *   self-improve → closed-loop text/config optimization (selfImprove, held-out gated)\n *   dynamic      → agent-authored topology (runLoop + createDriver)\n *\n * It is intentionally a thin façade: the value is that EVERY product reuses the\n * one hardened engine instead of forking delegation logic. The dispatcher owns\n * mode routing, timing, fail-loud on an unregistered mode, and a uniform result\n * shape; each mode's engine is a pre-configured runner in the registry (build it\n * with the factories below, or inject your own / a stub).\n */\n\nimport type { Scenario } from '@tangle-network/agent-eval/campaign'\nimport {\n  type SelfImproveOptions,\n  type SelfImproveResult,\n  selfImprove,\n} from '@tangle-network/agent-eval/contract'\nimport { runAnalystLoop } from './analyst-loop'\nimport type { RunAnalystLoopOpts, RunAnalystLoopResult } from './analyst-loop/types'\nimport { ConfigError } from './errors'\nimport {\n  type CoderReviewer,\n  type CoderWinnerSelection,\n  createDefaultCoderDelegate,\n  type DelegateRunCtx,\n} from './mcp/delegates'\nimport { type CreateKbGateOptions, createKbGate, type FactCandidate } from './mcp/kb-gate'\nimport type { DelegateCodeArgs } from './mcp/types'\nimport type { CoderOutput } from './profiles/coder'\nimport {\n  type AgentRunSpec,\n  type CreateDriverOptions,\n  createDriver,\n  type DriverDecision,\n  type LoopResult,\n  type OutputAdapter,\n  runLoop,\n  type SandboxClient,\n  type TopologyPlanner,\n  type Validator,\n} from './runtime'\n\n/** @experimental Every delegated-loop mode, for validation + CLI surfaces. */\nexport const DELEGATED_LOOP_MODES = [\n  'code',\n  'review',\n  'research',\n  'audit',\n  'self-improve',\n  'dynamic',\n] as const\n\n/** @experimental */\nexport type DelegatedLoopMode = (typeof DELEGATED_LOOP_MODES)[number]\n\n/** @experimental Type guard for an untrusted mode string (CLI / config input). */\nexport function isDelegatedLoopMode(value: unknown): value is DelegatedLoopMode {\n  return typeof value === 'string' && (DELEGATED_LOOP_MODES as readonly string[]).includes(value)\n}\n\n/** @experimental A pre-configured loop for one mode. Returns the mode's raw\n *  output; the dispatcher wraps it in a {@link DelegatedLoopResult}. */\nexport type DelegatedLoopRunner<T = unknown> = (signal: AbortSignal) => Promise<T>\n\n/** @experimental Mode → configured runner. Partial: only register the modes a\n *  given product/routine actually uses. */\nexport type DelegatedLoopRegistry = Partial<Record<DelegatedLoopMode, DelegatedLoopRunner>>\n\n/** @experimental Uniform result — never throws from a registered runner; a\n *  thrown engine becomes `{ ok: false, error }` so a routine can record + move on. */\nexport interface DelegatedLoopResult<T = unknown> {\n  mode: DelegatedLoopMode\n  ok: boolean\n  output?: T\n  error?: string\n  durationMs: number\n}\n\n/** @experimental */\nexport interface RunDelegatedLoopOptions {\n  signal?: AbortSignal\n  /** Clock override for deterministic tests. */\n  now?: () => number\n}\n\n/**\n * @experimental\n *\n * Dispatch a configured loop by mode. Fails loud (throws `ConfigError`) when no\n * runner is registered for the mode — a routine pointed at an unwired mode is a\n * config bug, not a silent no-op. A runner that throws is captured as\n * `{ ok: false }` so unattended runs record the failure rather than crash.\n */\nexport async function runDelegatedLoop<T = unknown>(\n  mode: DelegatedLoopMode,\n  registry: DelegatedLoopRegistry,\n  options: RunDelegatedLoopOptions = {},\n): Promise<DelegatedLoopResult<T>> {\n  const runner = registry[mode] as DelegatedLoopRunner<T> | undefined\n  if (!runner) {\n    throw new ConfigError(\n      `runDelegatedLoop: no runner registered for mode '${mode}' (registered: ${\n        Object.keys(registry).join(', ') || 'none'\n      })`,\n    )\n  }\n  const now = options.now ?? Date.now\n  const signal = options.signal ?? new AbortController().signal\n  const start = now()\n  try {\n    const output = await runner(signal)\n    return { mode, ok: true, output, durationMs: now() - start }\n  } catch (err) {\n    return {\n      mode,\n      ok: false,\n      error: err instanceof Error ? err.message : String(err),\n      durationMs: now() - start,\n    }\n  }\n}\n\n/** @experimental Options for the default `code`/`review` runner. */\nexport interface CoderLoopRunnerOptions {\n  sandboxClient: SandboxClient\n  /** What to build — the delegate args (goal, repoRoot, variants, config, …). */\n  args: DelegateCodeArgs\n  /** Adversarial reviewer. REQUIRED for `review` mode (see `reviewLoopRunner`). */\n  reviewer?: CoderReviewer\n  /** Winner-selection strategy. Default `highest-score`. */\n  winnerSelection?: CoderWinnerSelection\n  /** Harnesses for `variants > 1` fanout. */\n  fanoutHarnesses?: string[]\n}\n\n/** @experimental Build a `code`-mode runner over the hardened coder delegate. */\nexport function coderLoopRunner(options: CoderLoopRunnerOptions): DelegatedLoopRunner<CoderOutput> {\n  const delegate = createDefaultCoderDelegate({\n    sandboxClient: options.sandboxClient,\n    ...(options.reviewer ? { reviewer: options.reviewer } : {}),\n    ...(options.winnerSelection ? { winnerSelection: options.winnerSelection } : {}),\n    ...(options.fanoutHarnesses ? { fanoutHarnesses: options.fanoutHarnesses } : {}),\n  })\n  return async (signal) => {\n    const ctx: DelegateRunCtx = { signal, report: () => {} }\n    return delegate(options.args, ctx)\n  }\n}\n\n/**\n * @experimental\n *\n * `review` mode = `code` with a REQUIRED reviewer. The gate is the whole point,\n * so the type forces a reviewer (a \"review loop\" with no reviewer is a code loop).\n */\nexport function reviewLoopRunner(\n  options: CoderLoopRunnerOptions & { reviewer: CoderReviewer },\n): DelegatedLoopRunner<CoderOutput> {\n  return coderLoopRunner(options)\n}\n\n/** @experimental Options for the default `dynamic` runner. */\nexport interface DynamicLoopRunnerOptions<Task, Output> {\n  sandboxClient: SandboxClient\n  /** The agent-authored topology planner (sync or async; an async planner is where an LLM call goes). */\n  planner: TopologyPlanner<Task, Output>\n  task: Task\n  output: OutputAdapter<Output>\n  validator?: Validator<Output>\n  /** Exactly one of `agentRun` / `agentRuns` (runLoop validates). */\n  agentRun?: AgentRunSpec<Task>\n  agentRuns?: AgentRunSpec<Task>[]\n  maxIterations?: number\n  maxFanout?: number\n  /** Optional trace-analyst hook forwarded to the dynamic driver so the loop runs\n   *  `f(trace, findings)` — see `CreateDriverOptions.analyze`. Caller-side\n   *  seam to `runAnalystLoop`; keeps this runner analyst-free. */\n  analyze?: CreateDriverOptions<Task, Output>['analyze']\n}\n\n/** @experimental `dynamic` mode — agent-authored topology over `runLoop`. */\nexport function dynamicLoopRunner<Task, Output>(\n  o: DynamicLoopRunnerOptions<Task, Output>,\n): DelegatedLoopRunner<LoopResult<Task, Output, DriverDecision>> {\n  return async (signal) =>\n    runLoop<Task, Output, DriverDecision>({\n      driver: createDriver<Task, Output>({\n        planner: o.planner,\n        ...(o.maxIterations !== undefined ? { maxIterations: o.maxIterations } : {}),\n        ...(o.maxFanout !== undefined ? { maxFanout: o.maxFanout } : {}),\n        ...(o.analyze ? { analyze: o.analyze } : {}),\n      }),\n      ...(o.agentRun ? { agentRun: o.agentRun } : {}),\n      ...(o.agentRuns ? { agentRuns: o.agentRuns } : {}),\n      output: o.output,\n      ...(o.validator ? { validator: o.validator } : {}),\n      task: o.task,\n      ctx: { sandboxClient: o.sandboxClient, signal },\n      ...(o.maxIterations !== undefined ? { maxIterations: o.maxIterations } : {}),\n    })\n}\n\n/** @experimental A fact rejected at the KB gate — surfaced, never dropped. */\nexport interface VetoedFact {\n  candidate: FactCandidate\n  vetoedBy?: string\n  reason?: string\n}\n\n/** @experimental */\nexport interface ResearchLoopResult {\n  /** Facts that passed the fail-closed gate — safe to write to the KB. */\n  accepted: FactCandidate[]\n  /** Facts the gate vetoed in the final round — escalate, do not silently drop. */\n  vetoed: VetoedFact[]\n  /** Research rounds actually run. */\n  rounds: number\n}\n\n/** @experimental Options for the default `research` runner. */\nexport interface ResearchLoopRunnerOptions {\n  /**\n   * The research engine (the consumer's web/doc searcher + extractor). Called\n   * each round with the prior round's vetoes so it can re-research the gaps.\n   * Returns fact candidates carrying their grounding (`verbatimPassage` +\n   * `sourceText`).\n   */\n  research: (round: number, vetoed: VetoedFact[]) => Promise<FactCandidate[]>\n  /** Gate config (extra judges, self-artifact kinds, …). The floor is always on. */\n  gate?: CreateKbGateOptions\n  /** Max research rounds (correct-on-veto remediation). Default 1. */\n  maxRounds?: number\n}\n\n/**\n * @experimental `research` mode — research-in-a-loop with valid-only KB growth.\n *\n * Each round: research → gate every candidate (fail-closed; passage MUST be in\n * the source) → accept the clean ones → re-research the vetoed ones next round,\n * up to `maxRounds`. Vetoed facts in the final round are RETURNED (escalate,\n * never silently dropped) so the caller audits vs retries.\n */\nexport function researchLoopRunner(\n  o: ResearchLoopRunnerOptions,\n): DelegatedLoopRunner<ResearchLoopResult> {\n  const gate = createKbGate(o.gate)\n  const maxRounds = Math.max(1, Math.trunc(o.maxRounds ?? 1))\n  return async (signal) => {\n    const accepted: FactCandidate[] = []\n    let vetoed: VetoedFact[] = []\n    let rounds = 0\n    for (let round = 0; round < maxRounds; round += 1) {\n      if (signal.aborted) break\n      rounds += 1\n      const candidates = await o.research(round, vetoed)\n      if (candidates.length === 0) break\n      vetoed = []\n      for (const c of candidates) {\n        const v = await gate(c)\n        if (v.accepted) accepted.push(c)\n        else vetoed.push({ candidate: c, vetoedBy: v.vetoedBy, reason: v.reason })\n      }\n      if (vetoed.length === 0) break\n    }\n    return { accepted, vetoed, rounds }\n  }\n}\n\n/** @experimental `self-improve` mode — agent-eval's one-call closed loop (held-out gated). */\nexport function selfImproveLoopRunner<TScenario extends Scenario, TArtifact>(\n  options: SelfImproveOptions<TScenario, TArtifact>,\n): DelegatedLoopRunner<SelfImproveResult<TScenario, TArtifact>> {\n  return async () => selfImprove<TScenario, TArtifact>(options)\n}\n\n/** @experimental `audit` mode — analyst loop over captured trace/run data. */\nexport function auditLoopRunner<TProposal = unknown, TEdit = unknown>(\n  options: RunAnalystLoopOpts,\n): DelegatedLoopRunner<RunAnalystLoopResult<TProposal, TEdit>> {\n  return async () => runAnalystLoop<TProposal, TEdit>(options)\n}\n","#!/usr/bin/env node\n/**\n * @experimental\n *\n * `agent-runtime-loop` — the schedulable entrypoint for the configured\n * delegated loop-runner. A cron job / routine / Makefile target invokes:\n *\n *   agent-runtime-loop --mode research --config ./loops.config.js\n *\n * The config module wires the registry (with full access to env / creds —\n * which is why the deps live there, not in this generic bin). It must default-\n * export a `DelegatedLoopRegistry`, or a `() => DelegatedLoopRegistry | Promise<…>`.\n * The bin runs the selected mode, prints the `DelegatedLoopResult` as JSON, and\n * exits 0 on `ok`, 1 on a recorded failure, 2 on a usage/config error.\n */\n\nimport {\n  DELEGATED_LOOP_MODES,\n  type DelegatedLoopMode,\n  type DelegatedLoopRegistry,\n  type DelegatedLoopResult,\n  isDelegatedLoopMode,\n  runDelegatedLoop,\n} from './loop-runner'\n\n/** @experimental Parsed CLI invocation. */\nexport interface LoopRunnerCliArgs {\n  mode: string\n  /** Loads the registry — the bin wires this from `--config`; tests inject a stub. */\n  loadRegistry: () => Promise<DelegatedLoopRegistry> | DelegatedLoopRegistry\n  now?: () => number\n}\n\n/** @experimental */\nexport interface LoopRunnerCliResult {\n  exitCode: number\n  result?: DelegatedLoopResult\n  error?: string\n}\n\n/**\n * @experimental\n *\n * Pure CLI core (no process / argv / IO) so it's unit-testable: validate the\n * mode, load the registry, dispatch, map to an exit code (0 ok / 1 failed /\n * 2 usage). Exported for embedding in custom runners + tests.\n */\nexport async function runLoopRunnerCli(args: LoopRunnerCliArgs): Promise<LoopRunnerCliResult> {\n  if (!isDelegatedLoopMode(args.mode)) {\n    return {\n      exitCode: 2,\n      error: `unknown mode '${args.mode}' (expected one of: ${DELEGATED_LOOP_MODES.join(', ')})`,\n    }\n  }\n  let registry: DelegatedLoopRegistry\n  try {\n    registry = await args.loadRegistry()\n  } catch (err) {\n    return { exitCode: 2, error: `failed to load registry: ${errMsg(err)}` }\n  }\n  if (!registry[args.mode]) {\n    return {\n      exitCode: 2,\n      error: `config registers no runner for mode '${args.mode}' (registered: ${\n        Object.keys(registry).join(', ') || 'none'\n      })`,\n    }\n  }\n  // runDelegatedLoop throws only on a missing runner (guarded above); a failing\n  // engine is captured as { ok: false } → exit 1, not a crash.\n  const result = await runDelegatedLoop(args.mode as DelegatedLoopMode, registry, {\n    ...(args.now ? { now: args.now } : {}),\n  })\n  return { exitCode: result.ok ? 0 : 1, result }\n}\n\n/** Parse `--mode X --config Y` from an argv tail (`process.argv.slice(2)`). */\nexport function parseLoopRunnerArgv(argv: string[]): { mode?: string; config?: string } {\n  const out: { mode?: string; config?: string } = {}\n  for (let i = 0; i < argv.length; i += 1) {\n    const a = argv[i]\n    if (a === '--mode') out.mode = argv[++i]\n    else if (a === '--config') out.config = argv[++i]\n    else if (a?.startsWith('--mode=')) out.mode = a.slice('--mode='.length)\n    else if (a?.startsWith('--config=')) out.config = a.slice('--config='.length)\n  }\n  return out\n}\n\n/** Normalize a config module's default export → a registry. */\nfunction resolveRegistry(mod: unknown): DelegatedLoopRegistry {\n  const def = (mod as { default?: unknown })?.default ?? mod\n  const value = typeof def === 'function' ? (def as () => unknown)() : def\n  return value as DelegatedLoopRegistry\n}\n\nfunction errMsg(err: unknown): string {\n  return err instanceof Error ? err.message : String(err)\n}\n\n/** The argv → IO → exit shell. Kept thin; logic lives in `runLoopRunnerCli`. */\nasync function main(): Promise<void> {\n  const { mode, config } = parseLoopRunnerArgv(process.argv.slice(2))\n  if (!mode || !config) {\n    process.stderr.write(\n      'usage: agent-runtime-loop --mode <mode> --config <module>\\n' +\n        `  modes: ${DELEGATED_LOOP_MODES.join(' | ')}\\n` +\n        '  config: a JS/TS module default-exporting a DelegatedLoopRegistry (or a factory)\\n',\n    )\n    process.exit(2)\n  }\n  const { pathToFileURL } = await import('node:url')\n  const { resolve } = await import('node:path')\n  const cli = await runLoopRunnerCli({\n    mode,\n    loadRegistry: async () => resolveRegistry(await import(pathToFileURL(resolve(config)).href)),\n  })\n  process.stdout.write(`${JSON.stringify(cli.result ?? { error: cli.error }, null, 2)}\\n`)\n  if (cli.error) process.stderr.write(`${cli.error}\\n`)\n  process.exit(cli.exitCode)\n}\n\n// Run only when executed as the bin — never when imported for the testable\n// core, and never when bundled into a runtime that has no `process.argv`\n// (e.g. Cloudflare Workers, where `process` is a shim without `argv`). Reading\n// `process.argv[1]` directly would throw at module load there; `process.argv?.`\n// keeps the guard a no-op instead of crashing the Worker on startup.\nconst invokedScript = typeof process !== 'undefined' ? process.argv?.[1] : undefined\nif (invokedScript && /loop-runner-bin\\.(js|ts|mjs)$/.test(invokedScript)) {\n  void main()\n}\n"],"mappings":";;;;;;;;;;;;;;;;;;AAwBA;AAAA,EAGE;AAAA,OACK;AA2BA,IAAM,uBAAuB;AAAA,EAClC;AAAA,EACA;AAAA,EACA;AAAA,EACA;AAAA,EACA;AAAA,EACA;AACF;AAMO,SAAS,oBAAoB,OAA4C;AAC9E,SAAO,OAAO,UAAU,YAAa,qBAA2C,SAAS,KAAK;AAChG;AAmCA,eAAsB,iBACpB,MACA,UACA,UAAmC,CAAC,GACH;AACjC,QAAM,SAAS,SAAS,IAAI;AAC5B,MAAI,CAAC,QAAQ;AACX,UAAM,IAAI;AAAA,MACR,oDAAoD,IAAI,kBACtD,OAAO,KAAK,QAAQ,EAAE,KAAK,IAAI,KAAK,MACtC;AAAA,IACF;AAAA,EACF;AACA,QAAM,MAAM,QAAQ,OAAO,KAAK;AAChC,QAAM,SAAS,QAAQ,UAAU,IAAI,gBAAgB,EAAE;AACvD,QAAM,QAAQ,IAAI;AAClB,MAAI;AACF,UAAM,SAAS,MAAM,OAAO,MAAM;AAClC,WAAO,EAAE,MAAM,IAAI,MAAM,QAAQ,YAAY,IAAI,IAAI,MAAM;AAAA,EAC7D,SAAS,KAAK;AACZ,WAAO;AAAA,MACL;AAAA,MACA,IAAI;AAAA,MACJ,OAAO,eAAe,QAAQ,IAAI,UAAU,OAAO,GAAG;AAAA,MACtD,YAAY,IAAI,IAAI;AAAA,IACtB;AAAA,EACF;AACF;AAgBO,SAAS,gBAAgB,SAAmE;AACjG,QAAM,WAAW,2BAA2B;AAAA,IAC1C,eAAe,QAAQ;AAAA,IACvB,GAAI,QAAQ,WAAW,EAAE,UAAU,QAAQ,SAAS,IAAI,CAAC;AAAA,IACzD,GAAI,QAAQ,kBAAkB,EAAE,iBAAiB,QAAQ,gBAAgB,IAAI,CAAC;AAAA,IAC9E,GAAI,QAAQ,kBAAkB,EAAE,iBAAiB,QAAQ,gBAAgB,IAAI,CAAC;AAAA,EAChF,CAAC;AACD,SAAO,OAAO,WAAW;AACvB,UAAM,MAAsB,EAAE,QAAQ,QAAQ,MAAM;AAAA,IAAC,EAAE;AACvD,WAAO,SAAS,QAAQ,MAAM,GAAG;AAAA,EACnC;AACF;AAQO,SAAS,iBACd,SACkC;AAClC,SAAO,gBAAgB,OAAO;AAChC;AAsBO,SAAS,kBACd,GAC+D;AAC/D,SAAO,OAAO,WACZ,QAAsC;AAAA,IACpC,QAAQ,aAA2B;AAAA,MACjC,SAAS,EAAE;AAAA,MACX,GAAI,EAAE,kBAAkB,SAAY,EAAE,eAAe,EAAE,cAAc,IAAI,CAAC;AAAA,MAC1E,GAAI,EAAE,cAAc,SAAY,EAAE,WAAW,EAAE,UAAU,IAAI,CAAC;AAAA,MAC9D,GAAI,EAAE,UAAU,EAAE,SAAS,EAAE,QAAQ,IAAI,CAAC;AAAA,IAC5C,CAAC;AAAA,IACD,GAAI,EAAE,WAAW,EAAE,UAAU,EAAE,SAAS,IAAI,CAAC;AAAA,IAC7C,GAAI,EAAE,YAAY,EAAE,WAAW,EAAE,UAAU,IAAI,CAAC;AAAA,IAChD,QAAQ,EAAE;AAAA,IACV,GAAI,EAAE,YAAY,EAAE,WAAW,EAAE,UAAU,IAAI,CAAC;AAAA,IAChD,MAAM,EAAE;AAAA,IACR,KAAK,EAAE,eAAe,EAAE,eAAe,OAAO;AAAA,IAC9C,GAAI,EAAE,kBAAkB,SAAY,EAAE,eAAe,EAAE,cAAc,IAAI,CAAC;AAAA,EAC5E,CAAC;AACL;AA0CO,SAAS,mBACd,GACyC;AACzC,QAAM,OAAO,aAAa,EAAE,IAAI;AAChC,QAAM,YAAY,KAAK,IAAI,GAAG,KAAK,MAAM,EAAE,aAAa,CAAC,CAAC;AAC1D,SAAO,OAAO,WAAW;AACvB,UAAM,WAA4B,CAAC;AACnC,QAAI,SAAuB,CAAC;AAC5B,QAAI,SAAS;AACb,aAAS,QAAQ,GAAG,QAAQ,WAAW,SAAS,GAAG;AACjD,UAAI,OAAO,QAAS;AACpB,gBAAU;AACV,YAAM,aAAa,MAAM,EAAE,SAAS,OAAO,MAAM;AACjD,UAAI,WAAW,WAAW,EAAG;AAC7B,eAAS,CAAC;AACV,iBAAW,KAAK,YAAY;AAC1B,cAAM,IAAI,MAAM,KAAK,CAAC;AACtB,YAAI,EAAE,SAAU,UAAS,KAAK,CAAC;AAAA,YAC1B,QAAO,KAAK,EAAE,WAAW,GAAG,UAAU,EAAE,UAAU,QAAQ,EAAE,OAAO,CAAC;AAAA,MAC3E;AACA,UAAI,OAAO,WAAW,EAAG;AAAA,IAC3B;AACA,WAAO,EAAE,UAAU,QAAQ,OAAO;AAAA,EACpC;AACF;AAGO,SAAS,sBACd,SAC8D;AAC9D,SAAO,YAAY,YAAkC,OAAO;AAC9D;AAGO,SAAS,gBACd,SAC6D;AAC7D,SAAO,YAAY,eAAiC,OAAO;AAC7D;;;ACrPA,eAAsB,iBAAiB,MAAuD;AAC5F,MAAI,CAAC,oBAAoB,KAAK,IAAI,GAAG;AACnC,WAAO;AAAA,MACL,UAAU;AAAA,MACV,OAAO,iBAAiB,KAAK,IAAI,uBAAuB,qBAAqB,KAAK,IAAI,CAAC;AAAA,IACzF;AAAA,EACF;AACA,MAAI;AACJ,MAAI;AACF,eAAW,MAAM,KAAK,aAAa;AAAA,EACrC,SAAS,KAAK;AACZ,WAAO,EAAE,UAAU,GAAG,OAAO,4BAA4B,OAAO,GAAG,CAAC,GAAG;AAAA,EACzE;AACA,MAAI,CAAC,SAAS,KAAK,IAAI,GAAG;AACxB,WAAO;AAAA,MACL,UAAU;AAAA,MACV,OAAO,wCAAwC,KAAK,IAAI,kBACtD,OAAO,KAAK,QAAQ,EAAE,KAAK,IAAI,KAAK,MACtC;AAAA,IACF;AAAA,EACF;AAGA,QAAM,SAAS,MAAM,iBAAiB,KAAK,MAA2B,UAAU;AAAA,IAC9E,GAAI,KAAK,MAAM,EAAE,KAAK,KAAK,IAAI,IAAI,CAAC;AAAA,EACtC,CAAC;AACD,SAAO,EAAE,UAAU,OAAO,KAAK,IAAI,GAAG,OAAO;AAC/C;AAGO,SAAS,oBAAoB,MAAoD;AACtF,QAAM,MAA0C,CAAC;AACjD,WAAS,IAAI,GAAG,IAAI,KAAK,QAAQ,KAAK,GAAG;AACvC,UAAM,IAAI,KAAK,CAAC;AAChB,QAAI,MAAM,SAAU,KAAI,OAAO,KAAK,EAAE,CAAC;AAAA,aAC9B,MAAM,WAAY,KAAI,SAAS,KAAK,EAAE,CAAC;AAAA,aACvC,GAAG,WAAW,SAAS,EAAG,KAAI,OAAO,EAAE,MAAM,UAAU,MAAM;AAAA,aAC7D,GAAG,WAAW,WAAW,EAAG,KAAI,SAAS,EAAE,MAAM,YAAY,MAAM;AAAA,EAC9E;AACA,SAAO;AACT;AAGA,SAAS,gBAAgB,KAAqC;AAC5D,QAAM,MAAO,KAA+B,WAAW;AACvD,QAAM,QAAQ,OAAO,QAAQ,aAAc,IAAsB,IAAI;AACrE,SAAO;AACT;AAEA,SAAS,OAAO,KAAsB;AACpC,SAAO,eAAe,QAAQ,IAAI,UAAU,OAAO,GAAG;AACxD;AAGA,eAAe,OAAsB;AACnC,QAAM,EAAE,MAAM,OAAO,IAAI,oBAAoB,QAAQ,KAAK,MAAM,CAAC,CAAC;AAClE,MAAI,CAAC,QAAQ,CAAC,QAAQ;AACpB,YAAQ,OAAO;AAAA,MACb;AAAA,WACc,qBAAqB,KAAK,KAAK,CAAC;AAAA;AAAA;AAAA,IAEhD;AACA,YAAQ,KAAK,CAAC;AAAA,EAChB;AACA,QAAM,EAAE,cAAc,IAAI,MAAM,OAAO,KAAU;AACjD,QAAM,EAAE,QAAQ,IAAI,MAAM,OAAO,MAAW;AAC5C,QAAM,MAAM,MAAM,iBAAiB;AAAA,IACjC;AAAA,IACA,cAAc,YAAY,gBAAgB,MAAM,OAAO,cAAc,QAAQ,MAAM,CAAC,EAAE,KAAK;AAAA,EAC7F,CAAC;AACD,UAAQ,OAAO,MAAM,GAAG,KAAK,UAAU,IAAI,UAAU,EAAE,OAAO,IAAI,MAAM,GAAG,MAAM,CAAC,CAAC;AAAA,CAAI;AACvF,MAAI,IAAI,MAAO,SAAQ,OAAO,MAAM,GAAG,IAAI,KAAK;AAAA,CAAI;AACpD,UAAQ,KAAK,IAAI,QAAQ;AAC3B;AAOA,IAAM,gBAAgB,OAAO,YAAY,cAAc,QAAQ,OAAO,CAAC,IAAI;AAC3E,IAAI,iBAAiB,gCAAgC,KAAK,aAAa,GAAG;AACxE,OAAK,KAAK;AACZ;","names":[]}

package/dist/chunk-GSUO5QS6.js

@@ -0,0 +1,146 @@
+// src/errors.ts
+import { AgentEvalError } from "@tangle-network/agent-eval";
+import {
+  AgentEvalError as AgentEvalError2,
+  CaptureIntegrityError,
+  ConfigError,
+  JudgeError,
+  NotFoundError,
+  ReplayError,
+  ValidationError,
+  VerificationError
+} from "@tangle-network/agent-eval";
+var SessionMismatchError = class extends AgentEvalError {
+  sessionBackend;
+  requestedBackend;
+  constructor(sessionBackend, requestedBackend, options) {
+    super(
+      "validation",
+      `Cannot resume ${sessionBackend} session with ${requestedBackend} backend`,
+      options
+    );
+    this.sessionBackend = sessionBackend;
+    this.requestedBackend = requestedBackend;
+  }
+};
+var BackendTransportError = class extends AgentEvalError {
+  backend;
+  status;
+  /**
+   * Truncated upstream response body (≤2 KiB) when available. Diagnostic
+   * only — surfaces in `backend_error.error.body` and `final.error.body`
+   * so operators can see "free_tier_limit", "invalid_api_key", etc. without
+   * cracking the log line open.
+   */
+  body;
+  constructor(backend, message, options) {
+    super("config", message, options);
+    this.backend = backend;
+    this.status = options?.status;
+    this.body = options?.body;
+  }
+};
+var RuntimeRunStateError = class extends AgentEvalError {
+  constructor(message, options) {
+    super("validation", message, options);
+  }
+};
+var PlannerError = class extends AgentEvalError {
+  constructor(message, options) {
+    super("validation", message, options);
+  }
+};
+var AnalystError = class extends AgentEvalError {
+  constructor(message, options) {
+    super("validation", message, options);
+  }
+};
+
+// src/runtime/sandbox-events.ts
+function extractLlmCallEvent(event, agentRunName) {
+  if (!event || typeof event !== "object") return void 0;
+  const type = String(event.type ?? "");
+  const data = event.data && typeof event.data === "object" ? event.data : {};
+  if (type === "llm_call" || type === "cost.usage" || type === "usage") {
+    return buildLlmCall(data, agentRunName);
+  }
+  if (type === "message.completed" || type === "result" || type === "final") {
+    const usage = data.usage;
+    if (!usage || typeof usage !== "object") return void 0;
+    return buildLlmCall({ ...usage, model: data.model ?? usage.model }, agentRunName);
+  }
+  if (type === "done") {
+    const usage = data.tokenUsage;
+    if (!usage || typeof usage !== "object") return void 0;
+    const out = pickFiniteNumber(usage, ["outputTokens", "completion_tokens", "tokensOut"]);
+    const reasoning = pickFiniteNumber(usage, ["reasoningTokens"]);
+    const mergedOut = out !== void 0 || reasoning !== void 0 ? (out ?? 0) + (reasoning ?? 0) : void 0;
+    return buildLlmCall(
+      {
+        inputTokens: usage.inputTokens,
+        outputTokens: mergedOut,
+        totalCostUsd: data.totalCostUsd,
+        model: data.model ?? usage.model
+      },
+      agentRunName
+    );
+  }
+  return void 0;
+}
+function buildLlmCall(data, agentRunName) {
+  const tokensIn = pickFiniteNumber(data, ["tokensIn", "inputTokens", "prompt_tokens"]);
+  const tokensOut = pickFiniteNumber(data, ["tokensOut", "outputTokens", "completion_tokens"]);
+  const costUsd = pickFiniteNumber(data, ["costUsd", "totalCostUsd", "cost_usd", "cost"]);
+  if (tokensIn === void 0 && tokensOut === void 0 && costUsd === void 0) {
+    return void 0;
+  }
+  const model = typeof data.model === "string" && data.model.length > 0 ? data.model : agentRunName;
+  const event = {
+    type: "llm_call",
+    model
+  };
+  if (tokensIn !== void 0) event.tokensIn = tokensIn;
+  if (tokensOut !== void 0) event.tokensOut = tokensOut;
+  if (costUsd !== void 0) event.costUsd = costUsd;
+  return event;
+}
+function pickFiniteNumber(data, keys) {
+  for (const key of keys) {
+    const value = data[key];
+    if (typeof value === "number" && Number.isFinite(value)) return value;
+  }
+  return void 0;
+}
+function mapSandboxEvent(event, opts = {}) {
+  if (!event || typeof event !== "object") return void 0;
+  const type = String(event.type ?? "");
+  const data = event.data && typeof event.data === "object" ? event.data : {};
+  if (type === "message.part.updated") {
+    const part = data.part && typeof data.part === "object" ? data.part : {};
+    const partType = String(part.type ?? "");
+    const delta = typeof data.delta === "string" ? data.delta : void 0;
+    const text = delta ?? (typeof part.text === "string" ? part.text : void 0);
+    if (text === void 0) return void 0;
+    if (partType === "text") return { type: "text_delta", text };
+    if (partType === "reasoning" || partType === "thinking")
+      return { type: "reasoning_delta", text };
+    return void 0;
+  }
+  return extractLlmCallEvent(event, opts.agentRunName ?? "agent");
+}
+
+export {
+  SessionMismatchError,
+  BackendTransportError,
+  RuntimeRunStateError,
+  PlannerError,
+  AnalystError,
+  AgentEvalError2 as AgentEvalError,
+  ConfigError,
+  JudgeError,
+  NotFoundError,
+  ValidationError,
+  extractLlmCallEvent,
+  mapSandboxEvent
+};
+//# sourceMappingURL=chunk-GSUO5QS6.js.map

package/dist/chunk-GSUO5QS6.js.map

@@ -0,0 +1 @@
+{"version":3,"sources":["../src/errors.ts","../src/runtime/sandbox-events.ts"],"sourcesContent":["/**\n * @stable\n *\n * Error taxonomy for `@tangle-network/agent-runtime`.\n *\n * Public contract: every error this package throws as part of its consumer-\n * facing API either extends `AgentEvalError` (re-exported here for ergonomic\n * `instanceof` checks at the runtime boundary) or extends one of the\n * runtime-specific subclasses below.\n *\n * Internal invariant guards (`throw new Error('this should never happen')`)\n * remain plain `Error` — they are programmer-mistake assertions, not\n * consumer-catchable contract failures.\n *\n * Subclassing strategy: where a runtime-specific failure maps cleanly to an\n * agent-eval code (validation, config, not_found), we re-use the agent-eval\n * subclass. Runtime-only failure modes (session resume against the wrong\n * backend, backend transport errors) get fresh subclasses that still carry an\n * `AgentEvalErrorCode` so cross-package handlers can pattern-match without\n * importing the runtime.\n */\n\nimport { AgentEvalError } from '@tangle-network/agent-eval'\n\nexport {\n  AgentEvalError,\n  type AgentEvalErrorCode,\n  CaptureIntegrityError,\n  ConfigError,\n  JudgeError,\n  NotFoundError,\n  ReplayError,\n  ValidationError,\n  VerificationError,\n} from '@tangle-network/agent-eval'\n\n/**\n * @stable\n *\n * Caller asked to resume a session against a backend whose `kind` does not\n * match the session's recorded backend. This is a routing bug — the same\n * session id was reused across two different backend implementations — and\n * is not retryable without picking the right backend.\n */\nexport class SessionMismatchError extends AgentEvalError {\n  readonly sessionBackend: string\n  readonly requestedBackend: string\n\n  constructor(sessionBackend: string, requestedBackend: string, options?: { cause?: unknown }) {\n    super(\n      'validation',\n      `Cannot resume ${sessionBackend} session with ${requestedBackend} backend`,\n      options,\n    )\n    this.sessionBackend = sessionBackend\n    this.requestedBackend = requestedBackend\n  }\n}\n\n/**\n * @stable\n *\n * A backend transport call (HTTP, gRPC, sidecar IPC) failed with a non-success\n * status. Distinct from `JudgeError` (which is structural / unrecoverable)\n * because backend failures are sometimes retryable and consumers may want to\n * branch on the upstream status code.\n */\nexport class BackendTransportError extends AgentEvalError {\n  readonly backend: string\n  readonly status?: number\n  /**\n   * Truncated upstream response body (≤2 KiB) when available. Diagnostic\n   * only — surfaces in `backend_error.error.body` and `final.error.body`\n   * so operators can see \"free_tier_limit\", \"invalid_api_key\", etc. without\n   * cracking the log line open.\n   */\n  readonly body?: string\n\n  constructor(\n    backend: string,\n    message: string,\n    options?: { cause?: unknown; status?: number; body?: string },\n  ) {\n    super('config', message, options)\n    this.backend = backend\n    this.status = options?.status\n    this.body = options?.body\n  }\n}\n\n/**\n * @stable\n *\n * A runtime-run lifecycle method was called in an order the state machine does\n * not allow: `persist()` before `complete()`, `complete()` twice, etc.\n */\nexport class RuntimeRunStateError extends AgentEvalError {\n  constructor(message: string, options?: { cause?: unknown }) {\n    super('validation', message, options)\n  }\n}\n\n/**\n * @stable\n *\n * The dynamic-loop planner returned an unusable topology move — the LLM emitted\n * no parseable envelope, an unknown `kind`, or a structurally-invalid move\n * (e.g. a fanout with zero tasks). This is a structural failure of the\n * agent-authored topology, not a config mistake: the planner ran but its output\n * cannot drive the kernel. Carries `validation` so cross-package handlers can\n * pattern-match without importing the runtime. Fail loud — never substitute a\n * default move, or the loop silently runs a topology nobody chose.\n */\nexport class PlannerError extends AgentEvalError {\n  constructor(message: string, options?: { cause?: unknown }) {\n    super('validation', message, options)\n  }\n}\n\n/**\n * The analyst loop could not read or run over a round's trace — e.g. an empty round\n * (no iterations to analyze) or a malformed trace projection. Fail loud: a silent empty\n * store would mask a broken capture path and the driver would steer on nothing.\n */\nexport class AnalystError extends AgentEvalError {\n  constructor(message: string, options?: { cause?: unknown }) {\n    super('validation', message, options)\n  }\n}\n","/**\n * Sandbox-event → runtime-event mapping.\n *\n * The sandbox SDK emits a polymorphic `SandboxEvent = { type, data, id? }`\n * whose `type` vocabulary is backend-determined (opencode, etc.) rather than\n * enumerated by the SDK. Two consumers project it:\n *   - the loop kernel's cost ledger (`extractLlmCallEvent`) — sums usage off\n *     every cost-bearing event, regardless of stream shape;\n *   - the `AgentRuntime.act` streaming contract (`mapSandboxEvent`) — projects\n *     incremental events to the `RuntimeStreamEvent` chat-UX vocabulary.\n *\n * Both live here so the empirically-observed `type` vocabulary has one home.\n */\n\nimport type { SandboxEvent } from '@tangle-network/sandbox'\nimport type { RuntimeStreamEvent } from '../types'\n\n/**\n * Extract a `RuntimeStreamEvent`-shaped `llm_call` from a sandbox event when\n * the event carries usage/cost data. Returns `undefined` for non-cost events\n * so the kernel can iterate the full stream without branching.\n *\n * Canonical cost-carrying types observed in the wild:\n *   - `llm_call` — `data: { model, tokensIn, tokensOut, costUsd, ... }`\n *   - `message.completed` / `result` — `data: { usage: { inputTokens,\n *      outputTokens, totalCostUsd? } }`\n *   - `cost.usage` / `usage` — same shape under a dedicated type\n *\n * Numeric coercion is strict: `Number.isFinite` gates every accumulator write\n * so a sentinel `NaN` from a misbehaving backend cannot poison the ledger.\n */\nexport function extractLlmCallEvent(\n  event: SandboxEvent,\n  agentRunName: string,\n): (RuntimeStreamEvent & { type: 'llm_call' }) | undefined {\n  if (!event || typeof event !== 'object') return undefined\n  const type = String(event.type ?? '')\n  const data =\n    event.data && typeof event.data === 'object'\n      ? (event.data as Record<string, unknown>)\n      : ({} as Record<string, unknown>)\n\n  if (type === 'llm_call' || type === 'cost.usage' || type === 'usage') {\n    return buildLlmCall(data, agentRunName)\n  }\n  if (type === 'message.completed' || type === 'result' || type === 'final') {\n    const usage = data.usage as Record<string, unknown> | undefined\n    if (!usage || typeof usage !== 'object') return undefined\n    return buildLlmCall({ ...usage, model: data.model ?? usage.model }, agentRunName)\n  }\n  // sandbox 0.4.0 terminal event: `data = { tokenUsage: { inputTokens, outputTokens,\n  // reasoningTokens, cacheReadInputTokens }, totalCostUsd }`. Usage lives under\n  // `tokenUsage` (not `usage`) and the cost is top-level — neither matched the\n  // branches above, so an in-process loopDispatch run reported {0,0} and the\n  // backend-integrity guard misread a real run as a stub. Reasoning tokens are\n  // billed output (reasoning models), so they fold into the output count.\n  if (type === 'done') {\n    const usage = data.tokenUsage as Record<string, unknown> | undefined\n    if (!usage || typeof usage !== 'object') return undefined\n    const out = pickFiniteNumber(usage, ['outputTokens', 'completion_tokens', 'tokensOut'])\n    const reasoning = pickFiniteNumber(usage, ['reasoningTokens'])\n    const mergedOut =\n      out !== undefined || reasoning !== undefined ? (out ?? 0) + (reasoning ?? 0) : undefined\n    return buildLlmCall(\n      {\n        inputTokens: usage.inputTokens,\n        outputTokens: mergedOut,\n        totalCostUsd: data.totalCostUsd,\n        model: data.model ?? usage.model,\n      },\n      agentRunName,\n    )\n  }\n  return undefined\n}\n\nfunction buildLlmCall(\n  data: Record<string, unknown>,\n  agentRunName: string,\n): (RuntimeStreamEvent & { type: 'llm_call' }) | undefined {\n  const tokensIn = pickFiniteNumber(data, ['tokensIn', 'inputTokens', 'prompt_tokens'])\n  const tokensOut = pickFiniteNumber(data, ['tokensOut', 'outputTokens', 'completion_tokens'])\n  const costUsd = pickFiniteNumber(data, ['costUsd', 'totalCostUsd', 'cost_usd', 'cost'])\n  if (tokensIn === undefined && tokensOut === undefined && costUsd === undefined) {\n    return undefined\n  }\n  const model = typeof data.model === 'string' && data.model.length > 0 ? data.model : agentRunName\n  const event: RuntimeStreamEvent & { type: 'llm_call' } = {\n    type: 'llm_call',\n    model,\n  }\n  if (tokensIn !== undefined) event.tokensIn = tokensIn\n  if (tokensOut !== undefined) event.tokensOut = tokensOut\n  if (costUsd !== undefined) event.costUsd = costUsd\n  return event\n}\n\nfunction pickFiniteNumber(data: Record<string, unknown>, keys: string[]): number | undefined {\n  for (const key of keys) {\n    const value = data[key]\n    if (typeof value === 'number' && Number.isFinite(value)) return value\n  }\n  return undefined\n}\n\n/**\n * Project one `SandboxEvent` onto the `RuntimeStreamEvent` chat-UX vocabulary,\n * for runtimes that bridge a sandbox `streamPrompt` into the\n * `AgentRuntime.act` streaming contract. Returns `undefined` for events that\n * have no faithful projection — the raw stream is preserved separately for the\n * `OutputAdapter`, so an unmapped event never loses data.\n *\n * Mapped (the task-optional incremental variants — no synthesized task\n * lifecycle, no guessed tool-part shapes):\n *   - `message.part.updated` text part → `text_delta`\n *   - `message.part.updated` reasoning/thinking part → `reasoning_delta`\n *   - cost-bearing events → `llm_call` (shared with the ledger extractor)\n *\n * The opencode backend emits incremental text as\n * `{ type: 'message.part.updated', data: { part: { type, text }, delta } }`;\n * `delta` is the increment, `part.text` the running accumulation.\n */\nexport function mapSandboxEvent(\n  event: SandboxEvent,\n  opts: { agentRunName?: string } = {},\n): RuntimeStreamEvent | undefined {\n  if (!event || typeof event !== 'object') return undefined\n  const type = String(event.type ?? '')\n  const data =\n    event.data && typeof event.data === 'object'\n      ? (event.data as Record<string, unknown>)\n      : ({} as Record<string, unknown>)\n\n  if (type === 'message.part.updated') {\n    const part =\n      data.part && typeof data.part === 'object' ? (data.part as Record<string, unknown>) : {}\n    const partType = String(part.type ?? '')\n    const delta = typeof data.delta === 'string' ? data.delta : undefined\n    const text = delta ?? (typeof part.text === 'string' ? part.text : undefined)\n    if (text === undefined) return undefined\n    if (partType === 'text') return { type: 'text_delta', text }\n    if (partType === 'reasoning' || partType === 'thinking')\n      return { type: 'reasoning_delta', text }\n    return undefined\n  }\n\n  return extractLlmCallEvent(event, opts.agentRunName ?? 'agent')\n}\n"],"mappings":";AAsBA,SAAS,sBAAsB;AAE/B;AAAA,EACE,kBAAAA;AAAA,EAEA;AAAA,EACA;AAAA,EACA;AAAA,EACA;AAAA,EACA;AAAA,EACA;AAAA,EACA;AAAA,OACK;AAUA,IAAM,uBAAN,cAAmC,eAAe;AAAA,EAC9C;AAAA,EACA;AAAA,EAET,YAAY,gBAAwB,kBAA0B,SAA+B;AAC3F;AAAA,MACE;AAAA,MACA,iBAAiB,cAAc,iBAAiB,gBAAgB;AAAA,MAChE;AAAA,IACF;AACA,SAAK,iBAAiB;AACtB,SAAK,mBAAmB;AAAA,EAC1B;AACF;AAUO,IAAM,wBAAN,cAAoC,eAAe;AAAA,EAC/C;AAAA,EACA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA,EAOA;AAAA,EAET,YACE,SACA,SACA,SACA;AACA,UAAM,UAAU,SAAS,OAAO;AAChC,SAAK,UAAU;AACf,SAAK,SAAS,SAAS;AACvB,SAAK,OAAO,SAAS;AAAA,EACvB;AACF;AAQO,IAAM,uBAAN,cAAmC,eAAe;AAAA,EACvD,YAAY,SAAiB,SAA+B;AAC1D,UAAM,cAAc,SAAS,OAAO;AAAA,EACtC;AACF;AAaO,IAAM,eAAN,cAA2B,eAAe;AAAA,EAC/C,YAAY,SAAiB,SAA+B;AAC1D,UAAM,cAAc,SAAS,OAAO;AAAA,EACtC;AACF;AAOO,IAAM,eAAN,cAA2B,eAAe;AAAA,EAC/C,YAAY,SAAiB,SAA+B;AAC1D,UAAM,cAAc,SAAS,OAAO;AAAA,EACtC;AACF;;;ACjGO,SAAS,oBACd,OACA,cACyD;AACzD,MAAI,CAAC,SAAS,OAAO,UAAU,SAAU,QAAO;AAChD,QAAM,OAAO,OAAO,MAAM,QAAQ,EAAE;AACpC,QAAM,OACJ,MAAM,QAAQ,OAAO,MAAM,SAAS,WAC/B,MAAM,OACN,CAAC;AAER,MAAI,SAAS,cAAc,SAAS,gBAAgB,SAAS,SAAS;AACpE,WAAO,aAAa,MAAM,YAAY;AAAA,EACxC;AACA,MAAI,SAAS,uBAAuB,SAAS,YAAY,SAAS,SAAS;AACzE,UAAM,QAAQ,KAAK;AACnB,QAAI,CAAC,SAAS,OAAO,UAAU,SAAU,QAAO;AAChD,WAAO,aAAa,EAAE,GAAG,OAAO,OAAO,KAAK,SAAS,MAAM,MAAM,GAAG,YAAY;AAAA,EAClF;AAOA,MAAI,SAAS,QAAQ;AACnB,UAAM,QAAQ,KAAK;AACnB,QAAI,CAAC,SAAS,OAAO,UAAU,SAAU,QAAO;AAChD,UAAM,MAAM,iBAAiB,OAAO,CAAC,gBAAgB,qBAAqB,WAAW,CAAC;AACtF,UAAM,YAAY,iBAAiB,OAAO,CAAC,iBAAiB,CAAC;AAC7D,UAAM,YACJ,QAAQ,UAAa,cAAc,UAAa,OAAO,MAAM,aAAa,KAAK;AACjF,WAAO;AAAA,MACL;AAAA,QACE,aAAa,MAAM;AAAA,QACnB,cAAc;AAAA,QACd,cAAc,KAAK;AAAA,QACnB,OAAO,KAAK,SAAS,MAAM;AAAA,MAC7B;AAAA,MACA;AAAA,IACF;AAAA,EACF;AACA,SAAO;AACT;AAEA,SAAS,aACP,MACA,cACyD;AACzD,QAAM,WAAW,iBAAiB,MAAM,CAAC,YAAY,eAAe,eAAe,CAAC;AACpF,QAAM,YAAY,iBAAiB,MAAM,CAAC,aAAa,gBAAgB,mBAAmB,CAAC;AAC3F,QAAM,UAAU,iBAAiB,MAAM,CAAC,WAAW,gBAAgB,YAAY,MAAM,CAAC;AACtF,MAAI,aAAa,UAAa,cAAc,UAAa,YAAY,QAAW;AAC9E,WAAO;AAAA,EACT;AACA,QAAM,QAAQ,OAAO,KAAK,UAAU,YAAY,KAAK,MAAM,SAAS,IAAI,KAAK,QAAQ;AACrF,QAAM,QAAmD;AAAA,IACvD,MAAM;AAAA,IACN;AAAA,EACF;AACA,MAAI,aAAa,OAAW,OAAM,WAAW;AAC7C,MAAI,cAAc,OAAW,OAAM,YAAY;AAC/C,MAAI,YAAY,OAAW,OAAM,UAAU;AAC3C,SAAO;AACT;AAEA,SAAS,iBAAiB,MAA+B,MAAoC;AAC3F,aAAW,OAAO,MAAM;AACtB,UAAM,QAAQ,KAAK,GAAG;AACtB,QAAI,OAAO,UAAU,YAAY,OAAO,SAAS,KAAK,EAAG,QAAO;AAAA,EAClE;AACA,SAAO;AACT;AAmBO,SAAS,gBACd,OACA,OAAkC,CAAC,GACH;AAChC,MAAI,CAAC,SAAS,OAAO,UAAU,SAAU,QAAO;AAChD,QAAM,OAAO,OAAO,MAAM,QAAQ,EAAE;AACpC,QAAM,OACJ,MAAM,QAAQ,OAAO,MAAM,SAAS,WAC/B,MAAM,OACN,CAAC;AAER,MAAI,SAAS,wBAAwB;AACnC,UAAM,OACJ,KAAK,QAAQ,OAAO,KAAK,SAAS,WAAY,KAAK,OAAmC,CAAC;AACzF,UAAM,WAAW,OAAO,KAAK,QAAQ,EAAE;AACvC,UAAM,QAAQ,OAAO,KAAK,UAAU,WAAW,KAAK,QAAQ;AAC5D,UAAM,OAAO,UAAU,OAAO,KAAK,SAAS,WAAW,KAAK,OAAO;AACnE,QAAI,SAAS,OAAW,QAAO;AAC/B,QAAI,aAAa,OAAQ,QAAO,EAAE,MAAM,cAAc,KAAK;AAC3D,QAAI,aAAa,eAAe,aAAa;AAC3C,aAAO,EAAE,MAAM,mBAAmB,KAAK;AACzC,WAAO;AAAA,EACT;AAEA,SAAO,oBAAoB,OAAO,KAAK,gBAAgB,OAAO;AAChE;","names":["AgentEvalError"]}