@tangle-network/agent-runtime 0.44.0 → 0.45.0

package/dist/{chunk-NRZOXCJK.js → chunk-KSMX62JF.js}

@@ -14,7 +14,7 @@ import {
   DELEGATION_STATUS_DESCRIPTION,
   DELEGATION_STATUS_INPUT_SCHEMA,
   DELEGATION_STATUS_TOOL_NAME
-} from "./chunk-HVYOHJHK.js";
+} from "./chunk-IJGS6J7X.js";
 
 // src/mcp/openai-tools.ts
 function buildTool(name, description, parameters) {
@@ -61,4 +61,4 @@ export {
   mcpToolsForRuntimeMcp,
   mcpToolsForRuntimeMcpSubset
 };
-//# sourceMappingURL=chunk-NRZOXCJK.js.map
+//# sourceMappingURL=chunk-KSMX62JF.js.map

package/dist/{chunk-GFKVVRQ7.js → chunk-NYN5RTLP.js}

@@ -3,22 +3,22 @@ import {
 } from "./chunk-FNMGYYSS.js";
 import {
   createDefaultCoderDelegate
-} from "./chunk-SKUZZCHE.js";
+} from "./chunk-IJ6FGOPO.js";
 import {
   runAnalystLoop
-} from "./chunk-XBUG326M.js";
-import {
-  optimizePrompt
-} from "./chunk-VOX6Z3II.js";
+} from "./chunk-FK53TXOP.js";
 import {
   createDynamicDriver,
   runLoop
-} from "./chunk-S7JXV32P.js";
+} from "./chunk-KEWO4KI6.js";
 import {
   ConfigError
-} from "./chunk-SQSCRJ7U.js";
+} from "./chunk-PRX45WE2.js";
 
 // src/loop-runner.ts
+import {
+  selfImprove
+} from "@tangle-network/agent-eval/contract";
 var DELEGATED_LOOP_MODES = [
   "code",
   "review",
@@ -73,7 +73,8 @@ function dynamicLoopRunner(o) {
     driver: createDynamicDriver({
       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-NYN5RTLP.js.map

package/dist/chunk-NYN5RTLP.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 + createDynamicDriver)\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 CreateDynamicDriverOptions,\n  createDynamicDriver,\n  type DynamicDecision,\n  type LoopResult,\n  type LoopSandboxClient,\n  type OutputAdapter,\n  runLoop,\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: LoopSandboxClient\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: LoopSandboxClient\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 `CreateDynamicDriverOptions.analyze`. Caller-side\n   *  seam to `runAnalystLoop`; keeps this runner analyst-free. */\n  analyze?: CreateDynamicDriverOptions<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, DynamicDecision>> {\n  return async (signal) =>\n    runLoop<Task, Output, DynamicDecision>({\n      driver: createDynamicDriver<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,GACgE;AAChE,SAAO,OAAO,WACZ,QAAuC;AAAA,IACrC,QAAQ,oBAAkC;AAAA,MACxC,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-PRX45WE2.js

@@ -0,0 +1,264 @@
+// 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");
+}
+
+// src/runtime/util.ts
+async function deleteBoxSafe(box) {
+  if (!box || typeof box.delete !== "function") return true;
+  try {
+    await box.delete();
+    return true;
+  } catch {
+    return false;
+  }
+}
+function randomSuffix(len = 8) {
+  return Math.random().toString(36).slice(2, 2 + len);
+}
+function randomUuid() {
+  return crypto.randomUUID();
+}
+function abortError() {
+  const err = new Error("aborted");
+  err.name = "AbortError";
+  return err;
+}
+function throwAbort() {
+  throw abortError();
+}
+function throwIfAborted(signal) {
+  if (signal?.aborted) throw abortError();
+}
+function sleep(ms, signal) {
+  return new Promise((resolve) => {
+    if (signal?.aborted) {
+      resolve();
+      return;
+    }
+    let onAbort;
+    const timer = setTimeout(() => {
+      if (onAbort && signal) signal.removeEventListener("abort", onAbort);
+      resolve();
+    }, ms);
+    if (signal) {
+      onAbort = () => {
+        clearTimeout(timer);
+        resolve();
+      };
+      signal.addEventListener("abort", onAbort, { once: true });
+    }
+  });
+}
+function withTimeout(promise, ms) {
+  return new Promise((resolve) => {
+    const timer = setTimeout(() => resolve(void 0), ms);
+    promise.then(
+      (value) => {
+        clearTimeout(timer);
+        resolve(value);
+      },
+      () => {
+        clearTimeout(timer);
+        resolve(void 0);
+      }
+    );
+  });
+}
+function stringifySafe(value, opts = {}) {
+  let s;
+  try {
+    if (typeof value === "string") {
+      s = value;
+    } else {
+      const json = opts.pretty ? JSON.stringify(value, null, 2) : JSON.stringify(value);
+      s = json ?? String(value);
+    }
+  } catch {
+    s = String(value);
+  }
+  if (opts.max !== void 0 && s.length > opts.max) return `${s.slice(0, opts.max)}\u2026`;
+  return s;
+}
+function zeroTokenUsage() {
+  return { input: 0, output: 0 };
+}
+function addTokenUsage(acc, delta) {
+  acc.input += delta.input ?? 0;
+  acc.output += delta.output ?? 0;
+}
+async function mapWithConcurrency(items, limit, fn) {
+  const bound = Math.max(1, Math.floor(limit));
+  const results = new Array(items.length);
+  let next = 0;
+  let failed = false;
+  const worker = async () => {
+    while (!failed) {
+      const i = next;
+      next += 1;
+      if (i >= items.length) return;
+      try {
+        results[i] = await fn(items[i], i);
+      } catch (err) {
+        failed = true;
+        throw err;
+      }
+    }
+  };
+  const workerCount = Math.min(bound, items.length);
+  await Promise.all(Array.from({ length: workerCount }, () => worker()));
+  return results;
+}
+
+export {
+  SessionMismatchError,
+  BackendTransportError,
+  RuntimeRunStateError,
+  PlannerError,
+  AnalystError,
+  AgentEvalError2 as AgentEvalError,
+  ConfigError,
+  JudgeError,
+  NotFoundError,
+  ValidationError,
+  deleteBoxSafe,
+  randomSuffix,
+  randomUuid,
+  throwAbort,
+  throwIfAborted,
+  sleep,
+  withTimeout,
+  stringifySafe,
+  zeroTokenUsage,
+  addTokenUsage,
+  mapWithConcurrency,
+  extractLlmCallEvent,
+  mapSandboxEvent
+};
+//# sourceMappingURL=chunk-PRX45WE2.js.map

package/dist/chunk-PRX45WE2.js.map

@@ -0,0 +1 @@
+{"version":3,"sources":["../src/errors.ts","../src/runtime/sandbox-events.ts","../src/runtime/util.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","/**\n * @experimental\n *\n * Internal loop-kernel utilities shared across the kernel, drivers, and the\n * sandbox-acquire layer. Not part of the public barrel surface.\n */\n\nimport type { SandboxInstance } from '@tangle-network/sandbox'\nimport type { LoopTokenUsage } from './types'\n\n/**\n * Best-effort sandbox delete. Skips instances without a `delete` (test fakes);\n * swallows errors (the platform reaps on expiry). Returns `false` when delete\n * threw, `true` otherwise, so callers can surface a leak if they choose.\n */\nexport async function deleteBoxSafe(box: SandboxInstance | undefined): Promise<boolean> {\n  if (!box || typeof (box as { delete?: unknown }).delete !== 'function') return true\n  try {\n    await box.delete()\n    return true\n  } catch {\n    return false\n  }\n}\n\n/** Short base36 id for trace correlation. Not cryptographic, not collision-free. */\nexport function randomSuffix(len = 8): string {\n  return Math.random()\n    .toString(36)\n    .slice(2, 2 + len)\n}\n\n/** Collision-resistant id for sandbox naming (find-by-name recovery must be unique). */\nexport function randomUuid(): string {\n  return crypto.randomUUID()\n}\n\n/** Construct an AbortError. Downstream code pattern-matches on `err.name`. */\nexport function abortError(): Error {\n  const err = new Error('aborted')\n  err.name = 'AbortError'\n  return err\n}\n\n/** Throw an AbortError. */\nexport function throwAbort(): never {\n  throw abortError()\n}\n\n/** Throw if the signal is already aborted; otherwise no-op. */\nexport function throwIfAborted(signal: AbortSignal | undefined): void {\n  if (signal?.aborted) throw abortError()\n}\n\n/**\n * Sleep that resolves early on abort and always clears its timer so it never\n * keeps the event loop alive. Resolves (does not reject) on abort — callers\n * re-check the signal explicitly after the sleep.\n */\nexport function sleep(ms: number, signal?: AbortSignal): Promise<void> {\n  return new Promise((resolve) => {\n    if (signal?.aborted) {\n      resolve()\n      return\n    }\n    let onAbort: (() => void) | undefined\n    const timer = setTimeout(() => {\n      if (onAbort && signal) signal.removeEventListener('abort', onAbort)\n      resolve()\n    }, ms)\n    if (signal) {\n      onAbort = () => {\n        clearTimeout(timer)\n        resolve()\n      }\n      signal.addEventListener('abort', onAbort, { once: true })\n    }\n  })\n}\n\n/**\n * Race a promise against a timeout. Resolves with the value if it settles in\n * time, otherwise resolves with `undefined`. Always clears the timer.\n */\nexport function withTimeout<T>(promise: Promise<T>, ms: number): Promise<T | undefined> {\n  return new Promise<T | undefined>((resolve) => {\n    const timer = setTimeout(() => resolve(undefined), ms)\n    promise.then(\n      (value) => {\n        clearTimeout(timer)\n        resolve(value)\n      },\n      () => {\n        clearTimeout(timer)\n        resolve(undefined)\n      },\n    )\n  })\n}\n\ninterface StringifyOptions {\n  /** Pretty-print with 2-space indent. Default false (compact). */\n  pretty?: boolean\n  /** Truncate to this many chars, appending `…`. Default unbounded. */\n  max?: number\n}\n\n/**\n * `JSON.stringify` with a `String()` fallback on throw (cyclic / non-JSON).\n * Strings pass through unstringified so a preview of a string output is the\n * string itself, not a quoted re-encoding.\n */\nexport function stringifySafe(value: unknown, opts: StringifyOptions = {}): string {\n  let s: string\n  try {\n    if (typeof value === 'string') {\n      s = value\n    } else {\n      const json = opts.pretty ? JSON.stringify(value, null, 2) : JSON.stringify(value)\n      s = json ?? String(value)\n    }\n  } catch {\n    s = String(value)\n  }\n  if (opts.max !== undefined && s.length > opts.max) return `${s.slice(0, opts.max)}…`\n  return s\n}\n\n/** A fresh zero token-usage accumulator. */\nexport function zeroTokenUsage(): LoopTokenUsage {\n  return { input: 0, output: 0 }\n}\n\n/** Add `delta` into `acc` in place. Missing fields count as zero. */\nexport function addTokenUsage(acc: LoopTokenUsage, delta: Partial<LoopTokenUsage>): void {\n  acc.input += delta.input ?? 0\n  acc.output += delta.output ?? 0\n}\n\n/**\n * Map `items` through `fn` with at most `limit` calls in flight at once,\n * preserving input order in the result. On the first `fn` rejection no NEW\n * items are picked up; already-in-flight calls are awaited, then the first\n * error is rethrown. `limit` is clamped to ≥ 1.\n *\n * Used where a burst of provisioning (e.g. forking N child boxes) must respect\n * the loop's concurrency bound instead of firing all N at once.\n */\nexport async function mapWithConcurrency<T, R>(\n  items: readonly T[],\n  limit: number,\n  fn: (item: T, index: number) => Promise<R>,\n): Promise<R[]> {\n  const bound = Math.max(1, Math.floor(limit))\n  const results = new Array<R>(items.length)\n  let next = 0\n  let failed = false\n  const worker = async (): Promise<void> => {\n    while (!failed) {\n      const i = next\n      next += 1\n      if (i >= items.length) return\n      try {\n        results[i] = await fn(items[i] as T, i)\n      } catch (err) {\n        failed = true\n        throw err\n      }\n    }\n  }\n  const workerCount = Math.min(bound, items.length)\n  await Promise.all(Array.from({ length: workerCount }, () => worker()))\n  return results\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;;;ACpIA,eAAsB,cAAc,KAAoD;AACtF,MAAI,CAAC,OAAO,OAAQ,IAA6B,WAAW,WAAY,QAAO;AAC/E,MAAI;AACF,UAAM,IAAI,OAAO;AACjB,WAAO;AAAA,EACT,QAAQ;AACN,WAAO;AAAA,EACT;AACF;AAGO,SAAS,aAAa,MAAM,GAAW;AAC5C,SAAO,KAAK,OAAO,EAChB,SAAS,EAAE,EACX,MAAM,GAAG,IAAI,GAAG;AACrB;AAGO,SAAS,aAAqB;AACnC,SAAO,OAAO,WAAW;AAC3B;AAGO,SAAS,aAAoB;AAClC,QAAM,MAAM,IAAI,MAAM,SAAS;AAC/B,MAAI,OAAO;AACX,SAAO;AACT;AAGO,SAAS,aAAoB;AAClC,QAAM,WAAW;AACnB;AAGO,SAAS,eAAe,QAAuC;AACpE,MAAI,QAAQ,QAAS,OAAM,WAAW;AACxC;AAOO,SAAS,MAAM,IAAY,QAAqC;AACrE,SAAO,IAAI,QAAQ,CAAC,YAAY;AAC9B,QAAI,QAAQ,SAAS;AACnB,cAAQ;AACR;AAAA,IACF;AACA,QAAI;AACJ,UAAM,QAAQ,WAAW,MAAM;AAC7B,UAAI,WAAW,OAAQ,QAAO,oBAAoB,SAAS,OAAO;AAClE,cAAQ;AAAA,IACV,GAAG,EAAE;AACL,QAAI,QAAQ;AACV,gBAAU,MAAM;AACd,qBAAa,KAAK;AAClB,gBAAQ;AAAA,MACV;AACA,aAAO,iBAAiB,SAAS,SAAS,EAAE,MAAM,KAAK,CAAC;AAAA,IAC1D;AAAA,EACF,CAAC;AACH;AAMO,SAAS,YAAe,SAAqB,IAAoC;AACtF,SAAO,IAAI,QAAuB,CAAC,YAAY;AAC7C,UAAM,QAAQ,WAAW,MAAM,QAAQ,MAAS,GAAG,EAAE;AACrD,YAAQ;AAAA,MACN,CAAC,UAAU;AACT,qBAAa,KAAK;AAClB,gBAAQ,KAAK;AAAA,MACf;AAAA,MACA,MAAM;AACJ,qBAAa,KAAK;AAClB,gBAAQ,MAAS;AAAA,MACnB;AAAA,IACF;AAAA,EACF,CAAC;AACH;AAcO,SAAS,cAAc,OAAgB,OAAyB,CAAC,GAAW;AACjF,MAAI;AACJ,MAAI;AACF,QAAI,OAAO,UAAU,UAAU;AAC7B,UAAI;AAAA,IACN,OAAO;AACL,YAAM,OAAO,KAAK,SAAS,KAAK,UAAU,OAAO,MAAM,CAAC,IAAI,KAAK,UAAU,KAAK;AAChF,UAAI,QAAQ,OAAO,KAAK;AAAA,IAC1B;AAAA,EACF,QAAQ;AACN,QAAI,OAAO,KAAK;AAAA,EAClB;AACA,MAAI,KAAK,QAAQ,UAAa,EAAE,SAAS,KAAK,IAAK,QAAO,GAAG,EAAE,MAAM,GAAG,KAAK,GAAG,CAAC;AACjF,SAAO;AACT;AAGO,SAAS,iBAAiC;AAC/C,SAAO,EAAE,OAAO,GAAG,QAAQ,EAAE;AAC/B;AAGO,SAAS,cAAc,KAAqB,OAAsC;AACvF,MAAI,SAAS,MAAM,SAAS;AAC5B,MAAI,UAAU,MAAM,UAAU;AAChC;AAWA,eAAsB,mBACpB,OACA,OACA,IACc;AACd,QAAM,QAAQ,KAAK,IAAI,GAAG,KAAK,MAAM,KAAK,CAAC;AAC3C,QAAM,UAAU,IAAI,MAAS,MAAM,MAAM;AACzC,MAAI,OAAO;AACX,MAAI,SAAS;AACb,QAAM,SAAS,YAA2B;AACxC,WAAO,CAAC,QAAQ;AACd,YAAM,IAAI;AACV,cAAQ;AACR,UAAI,KAAK,MAAM,OAAQ;AACvB,UAAI;AACF,gBAAQ,CAAC,IAAI,MAAM,GAAG,MAAM,CAAC,GAAQ,CAAC;AAAA,MACxC,SAAS,KAAK;AACZ,iBAAS;AACT,cAAM;AAAA,MACR;AAAA,IACF;AAAA,EACF;AACA,QAAM,cAAc,KAAK,IAAI,OAAO,MAAM,MAAM;AAChD,QAAM,QAAQ,IAAI,MAAM,KAAK,EAAE,QAAQ,YAAY,GAAG,MAAM,OAAO,CAAC,CAAC;AACrE,SAAO;AACT;","names":["AgentEvalError"]}

package/dist/{chunk-3HMHSN22.js → chunk-QR4UUC5P.js}

@@ -1,7 +1,3 @@
-import {
-  createFanoutVoteDriver
-} from "./chunk-PY6NMZYX.js";
-
 // src/profiles/coder.ts
 var DEFAULT_MAX_DIFF_LINES = 400;
 function coderProfile(options = {}) {
@@ -38,7 +34,11 @@ function multiHarnessCoderFanout(options = {}) {
     return agentRunSpec;
   });
   const { output, validator } = coderProfile();
-  const driver = createFanoutVoteDriver({ n: harnesses.length });
+  const driver = {
+    name: "fanout",
+    plan: async (task, history) => history.length === 0 ? agentRuns.map(() => task) : [],
+    decide: (history) => history.some((i) => i.verdict?.valid === true) ? "pick-winner" : "fail"
+  };
   return { agentRuns, output, validator, driver };
 }
 var DEFAULT_CODER_SYSTEM_PROMPT = [
@@ -261,4 +261,4 @@ export {
   multiHarnessCoderFanout,
   createCoderValidator
 };
-//# sourceMappingURL=chunk-3HMHSN22.js.map
+//# sourceMappingURL=chunk-QR4UUC5P.js.map

package/dist/chunk-QR4UUC5P.js.map

@@ -0,0 +1 @@
+{"version":3,"sources":["../src/profiles/coder.ts"],"sourcesContent":["/**\n * @experimental\n *\n * `coderProfile` — opinionated preset for code-modification tasks.\n *\n * The agent is told to:\n *   - work on a fresh branch inside the sandbox workspace\n *   - keep the patch minimal (under `maxDiffLines`)\n *   - avoid `forbiddenPaths`\n *   - run `testCmd` and `typecheckCmd`\n *   - emit a final JSON result the output adapter parses\n *\n * The profile is stateless and agent-agnostic — `harness` selects the\n * sandbox-SDK backend (`claude-code`, `codex`, `opencode/*`). For\n * heterogeneous fanout, use `multiHarnessCoderFanout`.\n */\n\nimport type { AgentProfile, SandboxEvent } from '@tangle-network/sandbox'\nimport type {\n  AgentRunSpec,\n  DefaultVerdict,\n  Driver,\n  OutputAdapter,\n  Validator,\n} from '../runtime/types'\n\nconst DEFAULT_MAX_DIFF_LINES = 400\n\n/** @experimental */\nexport interface CoderTask {\n  /** What the agent must accomplish. Free-form prose. */\n  goal: string\n  /** Absolute path inside the sandbox where the repo lives. */\n  repoRoot: string\n  /** Default `main`. The branch the agent diffs against. */\n  baseBranch?: string\n  /** Default `pnpm test --run`. */\n  testCmd?: string\n  /** Default `pnpm typecheck`. */\n  typecheckCmd?: string\n  /** Files the agent may inspect for context. Surfaced verbatim in the prompt. */\n  contextFiles?: string[]\n  /**\n   * Paths the agent must not touch. Validator hard-fails on any match.\n   * Use glob-free literal path prefixes for unambiguous enforcement.\n   */\n  forbiddenPaths?: string[]\n  /** Default 400. Hard cap; validator hard-fails when exceeded. */\n  maxDiffLines?: number\n}\n\n/** @experimental */\nexport interface CoderOutput {\n  /** Branch the agent wrote the patch on. */\n  branch: string\n  /** Unified diff (`git diff <base>..HEAD`). */\n  patch: string\n  testResult: { passed: boolean; output: string }\n  typecheckResult: { passed: boolean; output: string }\n  diffStats: { filesChanged: number; insertions: number; deletions: number }\n  /** Optional reviewer commentary surfaced by the agent. */\n  reviewerNotes?: string\n}\n\n/** @experimental */\nexport interface CoderProfileOptions {\n  /** Sandbox-SDK backend.type. Default `'claude-code'`. */\n  harness?: string\n  /** Default model id passed in `AgentProfile.model.default`. */\n  model?: string\n  /** Custom system prompt replacement. Default = built-in coder preset. */\n  systemPrompt?: string\n  /** Stable name for `AgentRunSpec.name`. Default = `coder-${harness}`. */\n  name?: string\n}\n\n/**\n * Build a coder preset.\n *\n * `validator` enforces test + typecheck + a 400-line default diff cap. For\n * per-task `forbiddenPaths` / `maxDiffLines` enforcement, pass `task` here\n * — the returned validator closes over its constraints. Without a task\n * the validator falls back to the default cap and skips path enforcement.\n *\n * @experimental\n */\nexport function coderProfile(options: CoderProfileOptions & { task?: CoderTask } = {}): {\n  profile: AgentProfile\n  taskToPrompt: (task: CoderTask) => string\n  output: OutputAdapter<CoderOutput>\n  validator: Validator<CoderOutput>\n  agentRunSpec: AgentRunSpec<CoderTask>\n} {\n  const harness = options.harness ?? 'claude-code'\n  const name = options.name ?? `coder-${harness}`\n  const systemPrompt = options.systemPrompt ?? DEFAULT_CODER_SYSTEM_PROMPT\n  const profile: AgentProfile = {\n    name,\n    description: 'Code-modification agent. Minimal-diff worktree-based coder.',\n    prompt: { systemPrompt },\n    model: options.model ? { default: options.model } : undefined,\n    tools: { git: true, fs: true, shell: true, test_runner: true },\n    metadata: { backendType: harness, role: 'coder' },\n  }\n  const output: OutputAdapter<CoderOutput> = { parse: parseCoderEvents }\n  const validator: Validator<CoderOutput> = options.task\n    ? createCoderValidator(options.task)\n    : createCoderValidator({\n        goal: '',\n        repoRoot: '',\n        forbiddenPaths: [],\n        maxDiffLines: DEFAULT_MAX_DIFF_LINES,\n      })\n  const agentRunSpec: AgentRunSpec<CoderTask> = {\n    name,\n    profile,\n    taskToPrompt: formatCoderPrompt,\n  }\n  return { profile, taskToPrompt: formatCoderPrompt, output, validator, agentRunSpec }\n}\n\n/** @experimental */\nexport interface MultiHarnessCoderFanoutOptions {\n  /**\n   * Sandbox-SDK backend.type identifiers, one per parallel agent. Default:\n   * `['claude-code', 'codex', 'opencode/zai-coding-plan/glm-5.1']`.\n   */\n  harnesses?: string[]\n  /** Optional per-harness model override. Indexed parallel to `harnesses`. */\n  models?: (string | undefined)[]\n}\n\n/** @experimental */\nexport function multiHarnessCoderFanout(options: MultiHarnessCoderFanoutOptions = {}): {\n  agentRuns: AgentRunSpec<CoderTask>[]\n  output: OutputAdapter<CoderOutput>\n  validator: Validator<CoderOutput>\n  driver: Driver<CoderTask, CoderOutput, 'pick-winner' | 'fail'>\n} {\n  const harnesses =\n    options.harnesses && options.harnesses.length > 0\n      ? options.harnesses\n      : ['claude-code', 'codex', 'opencode/zai-coding-plan/glm-5.1']\n  const models = options.models ?? []\n  const agentRuns = harnesses.map((harness, i) => {\n    const { agentRunSpec } = coderProfile({ harness, model: models[i] })\n    return agentRunSpec\n  })\n  const { output, validator } = coderProfile()\n  const driver: Driver<CoderTask, CoderOutput, 'pick-winner' | 'fail'> = {\n    name: 'fanout',\n    plan: async (task, history) => (history.length === 0 ? agentRuns.map(() => task) : []),\n    decide: (history) => (history.some((i) => i.verdict?.valid === true) ? 'pick-winner' : 'fail'),\n  }\n  return { agentRuns, output, validator, driver }\n}\n\nconst DEFAULT_CODER_SYSTEM_PROMPT = [\n  'You are a coder agent operating inside an isolated sandbox workspace.',\n  'Your job is to deliver a minimal, correct patch for the user-supplied goal.',\n  '',\n  'Hard rules:',\n  '  1. Work on a fresh branch off the supplied base. Do not mutate the base branch.',\n  '  2. Never touch a forbidden path. The user will list them explicitly.',\n  '  3. Keep the diff under the max-diff cap. Prefer the smallest change that ships.',\n  '  4. Run the supplied test and typecheck commands before declaring done.',\n  '  5. If either command fails, fix the cause — do not weaken the test or hide the error.',\n  '',\n  'When you finish, emit a single final structured message of the shape:',\n  '  ```json',\n  '  { \"branch\": \"<branch-name>\",',\n  '    \"patch\": \"<unified-diff>\",',\n  '    \"testResult\": { \"passed\": <bool>, \"output\": \"<stdout/stderr>\" },',\n  '    \"typecheckResult\": { \"passed\": <bool>, \"output\": \"<stdout/stderr>\" },',\n  '    \"diffStats\": { \"filesChanged\": <int>, \"insertions\": <int>, \"deletions\": <int> },',\n  '    \"reviewerNotes\": \"<optional commentary>\" }',\n  '  ```',\n].join('\\n')\n\nfunction formatCoderPrompt(task: CoderTask): string {\n  const base = task.baseBranch ?? 'main'\n  const testCmd = task.testCmd ?? 'pnpm test --run'\n  const typecheckCmd = task.typecheckCmd ?? 'pnpm typecheck'\n  const maxDiff = task.maxDiffLines ?? DEFAULT_MAX_DIFF_LINES\n  const forbidden = task.forbiddenPaths?.length ? task.forbiddenPaths.join(', ') : '(none)'\n  const context = task.contextFiles?.length\n    ? task.contextFiles.map((f) => `  - ${f}`).join('\\n')\n    : '  (none)'\n  return [\n    `Goal: ${task.goal}`,\n    `Repo: ${task.repoRoot}`,\n    `Base branch: ${base}`,\n    `Run tests with: ${testCmd}`,\n    `Run typecheck with: ${typecheckCmd}`,\n    `Forbidden paths: ${forbidden}`,\n    `Max diff lines: ${maxDiff}`,\n    'Context files:',\n    context,\n    '',\n    'Produce a minimal patch on a fresh branch. Run tests and typecheck before',\n    'returning. Emit the final JSON result block exactly as instructed.',\n  ].join('\\n')\n}\n\n/**\n * Walk the event stream and return the last structured `coder.result` payload.\n *\n * The agent is instructed to emit a JSON block; in practice the sandbox SDK\n * lifts the structured payload onto `data.result` of a `result` / `final`\n * event. When the event stream does not contain a structured result, the\n * adapter scans text deltas for a fenced JSON block matching the expected\n * keys. Both shapes converge on `CoderOutput`.\n */\nfunction parseCoderEvents(events: SandboxEvent[]): CoderOutput {\n  for (let i = events.length - 1; i >= 0; i -= 1) {\n    const event = events[i]\n    if (!event) continue\n    const type = String(event.type ?? '')\n    const data = isRecord(event.data) ? event.data : {}\n    if (type === 'result' || type === 'final' || type === 'coder.result') {\n      const direct = coerceCoderOutput(data.result ?? data.output ?? data)\n      if (direct) return direct\n    }\n  }\n  // Fallback: scan text deltas in reverse for a fenced JSON block.\n  for (let i = events.length - 1; i >= 0; i -= 1) {\n    const event = events[i]\n    if (!event) continue\n    const data = isRecord(event.data) ? event.data : {}\n    const text = pickString(data.text) ?? pickString(data.delta)\n    if (!text) continue\n    const fenced = extractFencedJson(text)\n    if (!fenced) continue\n    const coerced = coerceCoderOutput(fenced)\n    if (coerced) return coerced\n  }\n  return {\n    branch: '',\n    patch: '',\n    testResult: { passed: false, output: '' },\n    typecheckResult: { passed: false, output: '' },\n    diffStats: { filesChanged: 0, insertions: 0, deletions: 0 },\n  }\n}\n\n/**\n * Build a validator that closes over a specific `CoderTask`'s constraints.\n *\n * Checks in order:\n *   1. Forbidden-path: any `+++` / `---` header in the patch matching a\n *      path prefix in `task.forbiddenPaths` fails hard.\n *   2. Diff size: line count above `task.maxDiffLines` (default 400) fails\n *      hard; below cap, the score shrinks linearly.\n *   3. Tests: `output.testResult.passed` must be `true`.\n *   4. Typecheck: `output.typecheckResult.passed` must be `true`.\n *\n * Aggregate score: `0.5 * tests + 0.3 * typecheck + 0.2 * (1 - diffLines/maxDiff)`.\n * `valid` is the conjunction of all four.\n *\n * @experimental\n */\n/**\n * Default-on safety floor (folded from the ai-trading-blueprint delegation\n * MCP): a coder patch that touches a credential-shaped path is rejected\n * regardless of `forbiddenPaths` config. Catches `.env`, private keys,\n * keystores, wallets, and the common secret/credential JSON files.\n */\nconst SECRET_PATH_RE =\n  /(^|\\/)(\\.env(\\.|$)|.*\\.(pem|key|p12|pfx|keystore|wallet)|id_rsa|id_ed25519|secrets?\\.json|credentials?\\.json)$/i\n\nexport function createCoderValidator(task: CoderTask): Validator<CoderOutput> {\n  const maxDiff = task.maxDiffLines ?? DEFAULT_MAX_DIFF_LINES\n  const forbidden = task.forbiddenPaths ?? []\n  return {\n    async validate(output) {\n      const scores: Record<string, number> = {}\n      const notes: string[] = []\n      let pass = true\n\n      const touched = touchedPathsFromPatch(output.patch)\n\n      // No-op rejection: an empty patch can trivially \"pass\" tests/typecheck\n      // (nothing changed) yet does no work — never a valid coder result.\n      if (touched.length === 0 || output.patch.trim().length === 0) {\n        pass = false\n        scores.nonEmpty = 0\n        notes.push('empty patch — no files changed')\n      } else {\n        scores.nonEmpty = 1\n      }\n\n      // Secret-path floor: always-on, independent of `forbiddenPaths`.\n      const touchedSecrets = touched.filter((p) => SECRET_PATH_RE.test(p))\n      if (touchedSecrets.length > 0) {\n        pass = false\n        scores.noSecrets = 0\n        notes.push(`touched secret-shaped paths: ${touchedSecrets.join(', ')}`)\n      } else {\n        scores.noSecrets = 1\n      }\n\n      const touchedForbidden = forbidden.filter((path) => {\n        const prefix = path.endsWith('/') ? path : `${path}/`\n        const exact = prefix.slice(0, -1)\n        return touched.some((p) => p === exact || p.startsWith(prefix))\n      })\n      if (touchedForbidden.length > 0) {\n        pass = false\n        scores.forbiddenPath = 0\n        notes.push(`touched forbidden paths: ${touchedForbidden.join(', ')}`)\n      } else {\n        scores.forbiddenPath = 1\n      }\n\n      const diffLines = countDiffLines(output.patch)\n      if (diffLines > maxDiff) {\n        pass = false\n        scores.diffSize = 0\n        notes.push(`diff ${diffLines} lines exceeds cap ${maxDiff}`)\n      } else {\n        scores.diffSize = maxDiff === 0 ? 0 : Math.max(0, 1 - diffLines / maxDiff)\n      }\n\n      scores.tests = output.testResult.passed ? 1 : 0\n      scores.typecheck = output.typecheckResult.passed ? 1 : 0\n      if (!output.testResult.passed) {\n        pass = false\n        notes.push('tests failed')\n      }\n      if (!output.typecheckResult.passed) {\n        pass = false\n        notes.push('typecheck failed')\n      }\n\n      const score = 0.5 * scores.tests + 0.3 * scores.typecheck + 0.2 * scores.diffSize\n      const verdict: DefaultVerdict = {\n        valid: pass,\n        score: Number.isFinite(score) ? score : 0,\n        scores,\n      }\n      if (notes.length > 0) verdict.notes = notes.join('; ')\n      return verdict\n    },\n  }\n}\n\nfunction touchedPathsFromPatch(patch: string): string[] {\n  const out = new Set<string>()\n  for (const line of patch.split(/\\r?\\n/)) {\n    if (line.startsWith('+++ ') || line.startsWith('--- ')) {\n      const rest = line.slice(4).trim()\n      if (rest === '/dev/null') continue\n      const stripped = rest.startsWith('a/') || rest.startsWith('b/') ? rest.slice(2) : rest\n      out.add(stripped)\n    }\n  }\n  return [...out]\n}\n\nfunction countDiffLines(patch: string): number {\n  let count = 0\n  for (const line of patch.split(/\\r?\\n/)) {\n    if (\n      (line.startsWith('+') || line.startsWith('-')) &&\n      !line.startsWith('+++') &&\n      !line.startsWith('---')\n    ) {\n      count += 1\n    }\n  }\n  return count\n}\n\nfunction isRecord(value: unknown): value is Record<string, unknown> {\n  return value !== null && typeof value === 'object' && !Array.isArray(value)\n}\n\nfunction pickString(value: unknown): string | undefined {\n  return typeof value === 'string' && value.length > 0 ? value : undefined\n}\n\nfunction extractFencedJson(text: string): unknown | undefined {\n  const match = text.match(/```(?:json)?\\s*([\\s\\S]*?)```/i)\n  if (!match) return undefined\n  const body = (match[1] ?? '').trim()\n  if (!body) return undefined\n  try {\n    return JSON.parse(body)\n  } catch {\n    return undefined\n  }\n}\n\nfunction coerceCoderOutput(value: unknown): CoderOutput | undefined {\n  if (!isRecord(value)) return undefined\n  const branch = pickString(value.branch)\n  const patch = pickString(value.patch) ?? ''\n  if (branch === undefined) return undefined\n  const testResult = coerceCmdResult(value.testResult)\n  const typecheckResult = coerceCmdResult(value.typecheckResult)\n  const diffStats = coerceDiffStats(value.diffStats)\n  return {\n    branch,\n    patch,\n    testResult,\n    typecheckResult,\n    diffStats,\n    reviewerNotes: pickString(value.reviewerNotes),\n  }\n}\n\nfunction coerceCmdResult(value: unknown): { passed: boolean; output: string } {\n  if (!isRecord(value)) return { passed: false, output: '' }\n  return {\n    passed: value.passed === true,\n    output: pickString(value.output) ?? '',\n  }\n}\n\nfunction coerceDiffStats(value: unknown): {\n  filesChanged: number\n  insertions: number\n  deletions: number\n} {\n  if (!isRecord(value)) return { filesChanged: 0, insertions: 0, deletions: 0 }\n  return {\n    filesChanged: toFiniteInt(value.filesChanged),\n    insertions: toFiniteInt(value.insertions),\n    deletions: toFiniteInt(value.deletions),\n  }\n}\n\nfunction toFiniteInt(value: unknown): number {\n  if (typeof value !== 'number') return 0\n  if (!Number.isFinite(value)) return 0\n  return Math.max(0, Math.trunc(value))\n}\n"],"mappings":";AA0BA,IAAM,yBAAyB;AA4DxB,SAAS,aAAa,UAAsD,CAAC,GAMlF;AACA,QAAM,UAAU,QAAQ,WAAW;AACnC,QAAM,OAAO,QAAQ,QAAQ,SAAS,OAAO;AAC7C,QAAM,eAAe,QAAQ,gBAAgB;AAC7C,QAAM,UAAwB;AAAA,IAC5B;AAAA,IACA,aAAa;AAAA,IACb,QAAQ,EAAE,aAAa;AAAA,IACvB,OAAO,QAAQ,QAAQ,EAAE,SAAS,QAAQ,MAAM,IAAI;AAAA,IACpD,OAAO,EAAE,KAAK,MAAM,IAAI,MAAM,OAAO,MAAM,aAAa,KAAK;AAAA,IAC7D,UAAU,EAAE,aAAa,SAAS,MAAM,QAAQ;AAAA,EAClD;AACA,QAAM,SAAqC,EAAE,OAAO,iBAAiB;AACrE,QAAM,YAAoC,QAAQ,OAC9C,qBAAqB,QAAQ,IAAI,IACjC,qBAAqB;AAAA,IACnB,MAAM;AAAA,IACN,UAAU;AAAA,IACV,gBAAgB,CAAC;AAAA,IACjB,cAAc;AAAA,EAChB,CAAC;AACL,QAAM,eAAwC;AAAA,IAC5C;AAAA,IACA;AAAA,IACA,cAAc;AAAA,EAChB;AACA,SAAO,EAAE,SAAS,cAAc,mBAAmB,QAAQ,WAAW,aAAa;AACrF;AAcO,SAAS,wBAAwB,UAA0C,CAAC,GAKjF;AACA,QAAM,YACJ,QAAQ,aAAa,QAAQ,UAAU,SAAS,IAC5C,QAAQ,YACR,CAAC,eAAe,SAAS,kCAAkC;AACjE,QAAM,SAAS,QAAQ,UAAU,CAAC;AAClC,QAAM,YAAY,UAAU,IAAI,CAAC,SAAS,MAAM;AAC9C,UAAM,EAAE,aAAa,IAAI,aAAa,EAAE,SAAS,OAAO,OAAO,CAAC,EAAE,CAAC;AACnE,WAAO;AAAA,EACT,CAAC;AACD,QAAM,EAAE,QAAQ,UAAU,IAAI,aAAa;AAC3C,QAAM,SAAiE;AAAA,IACrE,MAAM;AAAA,IACN,MAAM,OAAO,MAAM,YAAa,QAAQ,WAAW,IAAI,UAAU,IAAI,MAAM,IAAI,IAAI,CAAC;AAAA,IACpF,QAAQ,CAAC,YAAa,QAAQ,KAAK,CAAC,MAAM,EAAE,SAAS,UAAU,IAAI,IAAI,gBAAgB;AAAA,EACzF;AACA,SAAO,EAAE,WAAW,QAAQ,WAAW,OAAO;AAChD;AAEA,IAAM,8BAA8B;AAAA,EAClC;AAAA,EACA;AAAA,EACA;AAAA,EACA;AAAA,EACA;AAAA,EACA;AAAA,EACA;AAAA,EACA;AAAA,EACA;AAAA,EACA;AAAA,EACA;AAAA,EACA;AAAA,EACA;AAAA,EACA;AAAA,EACA;AAAA,EACA;AAAA,EACA;AAAA,EACA;AAAA,EACA;AACF,EAAE,KAAK,IAAI;AAEX,SAAS,kBAAkB,MAAyB;AAClD,QAAM,OAAO,KAAK,cAAc;AAChC,QAAM,UAAU,KAAK,WAAW;AAChC,QAAM,eAAe,KAAK,gBAAgB;AAC1C,QAAM,UAAU,KAAK,gBAAgB;AACrC,QAAM,YAAY,KAAK,gBAAgB,SAAS,KAAK,eAAe,KAAK,IAAI,IAAI;AACjF,QAAM,UAAU,KAAK,cAAc,SAC/B,KAAK,aAAa,IAAI,CAAC,MAAM,OAAO,CAAC,EAAE,EAAE,KAAK,IAAI,IAClD;AACJ,SAAO;AAAA,IACL,SAAS,KAAK,IAAI;AAAA,IAClB,SAAS,KAAK,QAAQ;AAAA,IACtB,gBAAgB,IAAI;AAAA,IACpB,mBAAmB,OAAO;AAAA,IAC1B,uBAAuB,YAAY;AAAA,IACnC,oBAAoB,SAAS;AAAA,IAC7B,mBAAmB,OAAO;AAAA,IAC1B;AAAA,IACA;AAAA,IACA;AAAA,IACA;AAAA,IACA;AAAA,EACF,EAAE,KAAK,IAAI;AACb;AAWA,SAAS,iBAAiB,QAAqC;AAC7D,WAAS,IAAI,OAAO,SAAS,GAAG,KAAK,GAAG,KAAK,GAAG;AAC9C,UAAM,QAAQ,OAAO,CAAC;AACtB,QAAI,CAAC,MAAO;AACZ,UAAM,OAAO,OAAO,MAAM,QAAQ,EAAE;AACpC,UAAM,OAAO,SAAS,MAAM,IAAI,IAAI,MAAM,OAAO,CAAC;AAClD,QAAI,SAAS,YAAY,SAAS,WAAW,SAAS,gBAAgB;AACpE,YAAM,SAAS,kBAAkB,KAAK,UAAU,KAAK,UAAU,IAAI;AACnE,UAAI,OAAQ,QAAO;AAAA,IACrB;AAAA,EACF;AAEA,WAAS,IAAI,OAAO,SAAS,GAAG,KAAK,GAAG,KAAK,GAAG;AAC9C,UAAM,QAAQ,OAAO,CAAC;AACtB,QAAI,CAAC,MAAO;AACZ,UAAM,OAAO,SAAS,MAAM,IAAI,IAAI,MAAM,OAAO,CAAC;AAClD,UAAM,OAAO,WAAW,KAAK,IAAI,KAAK,WAAW,KAAK,KAAK;AAC3D,QAAI,CAAC,KAAM;AACX,UAAM,SAAS,kBAAkB,IAAI;AACrC,QAAI,CAAC,OAAQ;AACb,UAAM,UAAU,kBAAkB,MAAM;AACxC,QAAI,QAAS,QAAO;AAAA,EACtB;AACA,SAAO;AAAA,IACL,QAAQ;AAAA,IACR,OAAO;AAAA,IACP,YAAY,EAAE,QAAQ,OAAO,QAAQ,GAAG;AAAA,IACxC,iBAAiB,EAAE,QAAQ,OAAO,QAAQ,GAAG;AAAA,IAC7C,WAAW,EAAE,cAAc,GAAG,YAAY,GAAG,WAAW,EAAE;AAAA,EAC5D;AACF;AAwBA,IAAM,iBACJ;AAEK,SAAS,qBAAqB,MAAyC;AAC5E,QAAM,UAAU,KAAK,gBAAgB;AACrC,QAAM,YAAY,KAAK,kBAAkB,CAAC;AAC1C,SAAO;AAAA,IACL,MAAM,SAAS,QAAQ;AACrB,YAAM,SAAiC,CAAC;AACxC,YAAM,QAAkB,CAAC;AACzB,UAAI,OAAO;AAEX,YAAM,UAAU,sBAAsB,OAAO,KAAK;AAIlD,UAAI,QAAQ,WAAW,KAAK,OAAO,MAAM,KAAK,EAAE,WAAW,GAAG;AAC5D,eAAO;AACP,eAAO,WAAW;AAClB,cAAM,KAAK,qCAAgC;AAAA,MAC7C,OAAO;AACL,eAAO,WAAW;AAAA,MACpB;AAGA,YAAM,iBAAiB,QAAQ,OAAO,CAAC,MAAM,eAAe,KAAK,CAAC,CAAC;AACnE,UAAI,eAAe,SAAS,GAAG;AAC7B,eAAO;AACP,eAAO,YAAY;AACnB,cAAM,KAAK,gCAAgC,eAAe,KAAK,IAAI,CAAC,EAAE;AAAA,MACxE,OAAO;AACL,eAAO,YAAY;AAAA,MACrB;AAEA,YAAM,mBAAmB,UAAU,OAAO,CAAC,SAAS;AAClD,cAAM,SAAS,KAAK,SAAS,GAAG,IAAI,OAAO,GAAG,IAAI;AAClD,cAAM,QAAQ,OAAO,MAAM,GAAG,EAAE;AAChC,eAAO,QAAQ,KAAK,CAAC,MAAM,MAAM,SAAS,EAAE,WAAW,MAAM,CAAC;AAAA,MAChE,CAAC;AACD,UAAI,iBAAiB,SAAS,GAAG;AAC/B,eAAO;AACP,eAAO,gBAAgB;AACvB,cAAM,KAAK,4BAA4B,iBAAiB,KAAK,IAAI,CAAC,EAAE;AAAA,MACtE,OAAO;AACL,eAAO,gBAAgB;AAAA,MACzB;AAEA,YAAM,YAAY,eAAe,OAAO,KAAK;AAC7C,UAAI,YAAY,SAAS;AACvB,eAAO;AACP,eAAO,WAAW;AAClB,cAAM,KAAK,QAAQ,SAAS,sBAAsB,OAAO,EAAE;AAAA,MAC7D,OAAO;AACL,eAAO,WAAW,YAAY,IAAI,IAAI,KAAK,IAAI,GAAG,IAAI,YAAY,OAAO;AAAA,MAC3E;AAEA,aAAO,QAAQ,OAAO,WAAW,SAAS,IAAI;AAC9C,aAAO,YAAY,OAAO,gBAAgB,SAAS,IAAI;AACvD,UAAI,CAAC,OAAO,WAAW,QAAQ;AAC7B,eAAO;AACP,cAAM,KAAK,cAAc;AAAA,MAC3B;AACA,UAAI,CAAC,OAAO,gBAAgB,QAAQ;AAClC,eAAO;AACP,cAAM,KAAK,kBAAkB;AAAA,MAC/B;AAEA,YAAM,QAAQ,MAAM,OAAO,QAAQ,MAAM,OAAO,YAAY,MAAM,OAAO;AACzE,YAAM,UAA0B;AAAA,QAC9B,OAAO;AAAA,QACP,OAAO,OAAO,SAAS,KAAK,IAAI,QAAQ;AAAA,QACxC;AAAA,MACF;AACA,UAAI,MAAM,SAAS,EAAG,SAAQ,QAAQ,MAAM,KAAK,IAAI;AACrD,aAAO;AAAA,IACT;AAAA,EACF;AACF;AAEA,SAAS,sBAAsB,OAAyB;AACtD,QAAM,MAAM,oBAAI,IAAY;AAC5B,aAAW,QAAQ,MAAM,MAAM,OAAO,GAAG;AACvC,QAAI,KAAK,WAAW,MAAM,KAAK,KAAK,WAAW,MAAM,GAAG;AACtD,YAAM,OAAO,KAAK,MAAM,CAAC,EAAE,KAAK;AAChC,UAAI,SAAS,YAAa;AAC1B,YAAM,WAAW,KAAK,WAAW,IAAI,KAAK,KAAK,WAAW,IAAI,IAAI,KAAK,MAAM,CAAC,IAAI;AAClF,UAAI,IAAI,QAAQ;AAAA,IAClB;AAAA,EACF;AACA,SAAO,CAAC,GAAG,GAAG;AAChB;AAEA,SAAS,eAAe,OAAuB;AAC7C,MAAI,QAAQ;AACZ,aAAW,QAAQ,MAAM,MAAM,OAAO,GAAG;AACvC,SACG,KAAK,WAAW,GAAG,KAAK,KAAK,WAAW,GAAG,MAC5C,CAAC,KAAK,WAAW,KAAK,KACtB,CAAC,KAAK,WAAW,KAAK,GACtB;AACA,eAAS;AAAA,IACX;AAAA,EACF;AACA,SAAO;AACT;AAEA,SAAS,SAAS,OAAkD;AAClE,SAAO,UAAU,QAAQ,OAAO,UAAU,YAAY,CAAC,MAAM,QAAQ,KAAK;AAC5E;AAEA,SAAS,WAAW,OAAoC;AACtD,SAAO,OAAO,UAAU,YAAY,MAAM,SAAS,IAAI,QAAQ;AACjE;AAEA,SAAS,kBAAkB,MAAmC;AAC5D,QAAM,QAAQ,KAAK,MAAM,+BAA+B;AACxD,MAAI,CAAC,MAAO,QAAO;AACnB,QAAM,QAAQ,MAAM,CAAC,KAAK,IAAI,KAAK;AACnC,MAAI,CAAC,KAAM,QAAO;AAClB,MAAI;AACF,WAAO,KAAK,MAAM,IAAI;AAAA,EACxB,QAAQ;AACN,WAAO;AAAA,EACT;AACF;AAEA,SAAS,kBAAkB,OAAyC;AAClE,MAAI,CAAC,SAAS,KAAK,EAAG,QAAO;AAC7B,QAAM,SAAS,WAAW,MAAM,MAAM;AACtC,QAAM,QAAQ,WAAW,MAAM,KAAK,KAAK;AACzC,MAAI,WAAW,OAAW,QAAO;AACjC,QAAM,aAAa,gBAAgB,MAAM,UAAU;AACnD,QAAM,kBAAkB,gBAAgB,MAAM,eAAe;AAC7D,QAAM,YAAY,gBAAgB,MAAM,SAAS;AACjD,SAAO;AAAA,IACL;AAAA,IACA;AAAA,IACA;AAAA,IACA;AAAA,IACA;AAAA,IACA,eAAe,WAAW,MAAM,aAAa;AAAA,EAC/C;AACF;AAEA,SAAS,gBAAgB,OAAqD;AAC5E,MAAI,CAAC,SAAS,KAAK,EAAG,QAAO,EAAE,QAAQ,OAAO,QAAQ,GAAG;AACzD,SAAO;AAAA,IACL,QAAQ,MAAM,WAAW;AAAA,IACzB,QAAQ,WAAW,MAAM,MAAM,KAAK;AAAA,EACtC;AACF;AAEA,SAAS,gBAAgB,OAIvB;AACA,MAAI,CAAC,SAAS,KAAK,EAAG,QAAO,EAAE,cAAc,GAAG,YAAY,GAAG,WAAW,EAAE;AAC5E,SAAO;AAAA,IACL,cAAc,YAAY,MAAM,YAAY;AAAA,IAC5C,YAAY,YAAY,MAAM,UAAU;AAAA,IACxC,WAAW,YAAY,MAAM,SAAS;AAAA,EACxC;AACF;AAEA,SAAS,YAAY,OAAwB;AAC3C,MAAI,OAAO,UAAU,SAAU,QAAO;AACtC,MAAI,CAAC,OAAO,SAAS,KAAK,EAAG,QAAO;AACpC,SAAO,KAAK,IAAI,GAAG,KAAK,MAAM,KAAK,CAAC;AACtC;","names":[]}

package/dist/chunk-WIR4HOOJ.js

@@ -0,0 +1,27 @@
+// src/profiles/ui-auditor/substrate.ts
+var UI_LENSES = [
+  "consistency",
+  "hierarchy",
+  "layout",
+  "ux-flow",
+  "duplication",
+  "accessibility",
+  "responsive",
+  "states",
+  "content",
+  "interaction",
+  "performance-perceived",
+  "other"
+];
+var UI_FINDING_SEVERITIES = [
+  "critical",
+  "high",
+  "med",
+  "low"
+];
+
+export {
+  UI_LENSES,
+  UI_FINDING_SEVERITIES
+};
+//# sourceMappingURL=chunk-WIR4HOOJ.js.map

package/dist/chunk-WIR4HOOJ.js.map

@@ -0,0 +1 @@
+{"version":3,"sources":["../src/profiles/ui-auditor/substrate.ts"],"sourcesContent":["/**\n * @experimental\n *\n * UI audit finding shapes — the unit of evidence a contributor can act on.\n *\n * A finding describes a single, actionable UI problem: lens, severity,\n * route, observation, impact, suggested fix, and screenshot evidence.\n * Findings are produced by the auditor profile, persisted by the issue\n * writer as self-contained GitHub-issue Markdown, and surfaced over MCP.\n *\n * The shapes are deliberately constraining — the validator + writer\n * hard-fail on missing screenshot evidence, missing lens, missing title.\n */\n\n/**\n * Canonical audit lenses. Each lens scopes a finding to a single class of\n * problem so a single audit pass can iterate them without pile-on findings\n * under a generic label.\n */\nexport type UiLens =\n  | 'consistency'\n  | 'hierarchy'\n  | 'layout'\n  | 'ux-flow'\n  | 'duplication'\n  | 'accessibility'\n  | 'responsive'\n  | 'states'\n  | 'content'\n  | 'interaction'\n  | 'performance-perceived'\n  | 'other'\n\n/** Frozen tuple of lenses for validation + iteration. */\nexport const UI_LENSES: readonly UiLens[] = [\n  'consistency',\n  'hierarchy',\n  'layout',\n  'ux-flow',\n  'duplication',\n  'accessibility',\n  'responsive',\n  'states',\n  'content',\n  'interaction',\n  'performance-perceived',\n  'other',\n] as const\n\n/**\n * Severity scale.\n *   - `critical` — blocks a core task or is an accessibility blocker.\n *   - `high`     — confusing, broken-looking, or noticeable friction.\n *   - `med`      — visible polish issue, would be caught in code review.\n *   - `low`      — nitpick worth fixing eventually.\n */\nexport type UiFindingSeverity = 'low' | 'med' | 'high' | 'critical'\n\n/** Frozen severity tuple, ordered worst → least bad for sort/report. */\nexport const UI_FINDING_SEVERITIES: readonly UiFindingSeverity[] = [\n  'critical',\n  'high',\n  'med',\n  'low',\n] as const\n\n/** Pointer to a screenshot referenced by a finding (workspace-relative path). */\nexport interface UiFindingScreenshot {\n  path: string\n  viewport?: string\n  label?: string\n}\n\n/**\n * A single UI audit finding — the unit of work a contributor can act on.\n *\n * Every field except the documented optionals is required. The auditor\n * validator + writer hard-fail on missing screenshot evidence, missing\n * lens, missing title, etc.\n */\nexport interface UiFinding {\n  /** Monotonic id assigned by the writer when persisting. Optional in-transit. */\n  id?: number\n  title: string\n  lens: UiLens\n  severity: UiFindingSeverity\n  /** Logical route the finding was observed on (e.g. `home`, `checkout-step-2`). */\n  route: string\n  /** Fully qualified URL the finding was observed at. */\n  url?: string\n  /** Viewport string the offending capture was taken at (e.g. `1280x800`). */\n  viewport?: string\n  /** CSS selector pinning the offending element, when one can be identified. */\n  selector?: string\n  /** 1–3 sentences describing what the screenshot shows that is wrong. */\n  observation: string\n  /** Who is affected and how. */\n  impact: string\n  /** A specific change a contributor could apply without asking back. */\n  suggestedFix: string\n  /** Optional explicit reproduction steps. Writer synthesizes from route/url/selector when omitted. */\n  reproSteps?: string\n  /** Free-form tags. */\n  tags?: readonly string[]\n  /** Screenshot references — must be non-empty for actionable findings. */\n  screenshots: readonly UiFindingScreenshot[]\n  /** Cross-references to similar findings already on file, by id. */\n  similarTo?: readonly number[]\n  /** ISO-8601 creation timestamp set by the writer when persisted. */\n  createdAt?: string\n}\n"],"mappings":";AAkCO,IAAM,YAA+B;AAAA,EAC1C;AAAA,EACA;AAAA,EACA;AAAA,EACA;AAAA,EACA;AAAA,EACA;AAAA,EACA;AAAA,EACA;AAAA,EACA;AAAA,EACA;AAAA,EACA;AAAA,EACA;AACF;AAYO,IAAM,wBAAsD;AAAA,EACjE;AAAA,EACA;AAAA,EACA;AAAA,EACA;AACF;","names":[]}