@tangle-network/agent-runtime 0.70.0 → 0.70.1

package/dist/chunk-BGOLR66M.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 *\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 DelegateRunCtx,\n  type DetachedWinnerSelection,\n  detachedSessionDelegate,\n} from './mcp/delegates'\nimport type { CoderOutput } from './mcp/detached-coder'\nimport { type CreateKbGateOptions, createKbGate, type FactCandidate } from './mcp/kb-gate'\nimport type { DelegateCodeArgs } from './mcp/types'\nimport {\n  type AuthoredHarness,\n  type Budget,\n  createExecutorRegistry,\n  definePersona,\n  runPersonified,\n  type SandboxClient,\n  type WinnerStrategy,\n  type WorktreeFanoutOptions,\n  type WorktreePatchArtifact,\n  worktreeFanout,\n} from './runtime'\n\n/** @experimental Every delegated-loop mode, for validation + CLI surfaces. */\nexport const DELEGATED_LOOP_MODES = ['code', 'review', 'research', 'audit', 'self-improve'] 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. Pass one to run `review` mode (an approval gate over the candidate). */\n  reviewer?: CoderReviewer\n  /** Winner-selection strategy. Default `highest-score`. */\n  winnerSelection?: DetachedWinnerSelection\n  /** Harnesses for `variants > 1` fanout. */\n  fanoutHarnesses?: string[]\n}\n\n/**\n * @experimental Build a `code`/`review`-mode runner over the sandbox-session coder delegate. Pass a\n * `reviewer` to run `review` mode — an approval gate over the validated candidate.\n */\nexport function coderLoopRunner(options: CoderLoopRunnerOptions): DelegatedLoopRunner<CoderOutput> {\n  const delegate = detachedSessionDelegate({\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/** @experimental Options for the local-repo `code` runner over the GENERIC recursive path. */\nexport interface WorktreeLoopRunnerOptions {\n  /** Absolute path to the local git checkout each worktree is cut from. */\n  repoRoot: string\n  /** The instruction handed to every authored harness (composed under each profile's systemPrompt). */\n  taskPrompt: string\n  /** The supervisor-authored harness profiles — one fanout item (one worktree-CLI leaf) each. */\n  harnesses: ReadonlyArray<AuthoredHarness>\n  /** Conserved budget pool bounding the fanout (equal-k holds by construction). */\n  budget: Budget\n  /** Shell command run in each worktree to derive the tests-PASS signal. */\n  testCmd?: string\n  /** Shell command run in each worktree to derive the typecheck-PASS signal. */\n  typecheckCmd?: string\n  /** Which verification signals the deliverable REQUIRES present-and-passing (default none). */\n  require?: ReadonlyArray<'tests' | 'typecheck'>\n  /** Diff-size cap (lines). */\n  maxDiffLines?: number\n  /** Literal path prefixes the patch must not touch (the secret-floor is always on regardless). */\n  forbiddenPaths?: string[]\n  /** Winner-selection strategy among gated candidates. Default `highest-score`. */\n  winnerStrategy?: WinnerStrategy\n  /** Test seams forwarded to the worktree-CLI leaves so the runner drives offline. */\n  runGit?: WorktreeFanoutOptions['runGit']\n  runHarness?: WorktreeFanoutOptions['runHarness']\n  runCommand?: WorktreeFanoutOptions['runCommand']\n}\n\n/**\n * @experimental\n *\n * `code` mode on the GENERIC recursive path: author one `AgentProfile` per harness, run them as a\n * `worktreeFanout` (N `createWorktreeCliExecutor` leaves, each `gateOnDeliverable`) through\n * `runPersonified` on the keystone Supervisor. This is the local-repo counterpart to\n * {@link coderLoopRunner} (which drives the in-box harness over a `SandboxClient`): no `runLoop`\n * driver, no role-coupled delegate — the harness list is the fanout, the gate is `patchDelivered`,\n * the winner is the shared valid-only selector (NOT `defaultSelectWinner`, whose non-valid fallback\n * would surface an ungated patch). Equal-k holds by the conserved budget pool. Returns the winning\n * patch artifact, or throws when no candidate is delivered (fail loud, never a vacuous done).\n */\nexport function worktreeLoopRunner(\n  options: WorktreeLoopRunnerOptions,\n): DelegatedLoopRunner<WorktreePatchArtifact> {\n  const shape = worktreeFanout<string>({\n    repoRoot: options.repoRoot,\n    taskPrompt: options.taskPrompt,\n    harnesses: options.harnesses,\n    ...(options.testCmd !== undefined ? { testCmd: options.testCmd } : {}),\n    ...(options.typecheckCmd !== undefined ? { typecheckCmd: options.typecheckCmd } : {}),\n    ...(options.require !== undefined ? { require: options.require } : {}),\n    ...(options.maxDiffLines !== undefined ? { maxDiffLines: options.maxDiffLines } : {}),\n    ...(options.forbiddenPaths !== undefined ? { forbiddenPaths: options.forbiddenPaths } : {}),\n    ...(options.winnerStrategy !== undefined ? { winnerStrategy: options.winnerStrategy } : {}),\n    ...(options.runGit ? { runGit: options.runGit } : {}),\n    ...(options.runHarness ? { runHarness: options.runHarness } : {}),\n    ...(options.runCommand ? { runCommand: options.runCommand } : {}),\n  })\n  // The persona's only role here is to carry the fanout shape onto the Supervisor; each item's\n  // executor is BYO (the gated worktree-CLI leaf), so the registry only needs to pass BYO through.\n  const persona = definePersona<WorktreePatchArtifact>({\n    name: 'worktree-coder',\n    root: { profile: { name: 'worktree-coder' }, harness: null },\n    directive: 'deliver a minimal validated patch on a fresh worktree',\n    context: { role: 'coder' },\n    executors: { registry: createExecutorRegistry() },\n  })\n  return async (signal) => {\n    const result = await runPersonified<string, WorktreePatchArtifact>({\n      persona,\n      shape,\n      task: options.taskPrompt,\n      budget: options.budget,\n      signal,\n    })\n    if (result.kind !== 'winner' || result.out.kind !== 'done') {\n      const blockers =\n        result.kind === 'winner' && result.out.kind === 'blocked'\n          ? result.out.blockers.join('; ')\n          : `supervisor settled ${result.kind}`\n      throw new Error(`worktreeLoopRunner: no delivered patch (${blockers})`)\n    }\n    return result.out.deliverable\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":";;;;;;;;;;;;;;;;;;;;AAuBA;AAAA,EAGE;AAAA,OACK;AA2BA,IAAM,uBAAuB,CAAC,QAAQ,UAAU,YAAY,SAAS,cAAc;AAMnF,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;AAmBO,SAAS,gBAAgB,SAAmE;AACjG,QAAM,WAAW,wBAAwB;AAAA,IACvC,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;AA0CO,SAAS,mBACd,SAC4C;AAC5C,QAAM,QAAQ,eAAuB;AAAA,IACnC,UAAU,QAAQ;AAAA,IAClB,YAAY,QAAQ;AAAA,IACpB,WAAW,QAAQ;AAAA,IACnB,GAAI,QAAQ,YAAY,SAAY,EAAE,SAAS,QAAQ,QAAQ,IAAI,CAAC;AAAA,IACpE,GAAI,QAAQ,iBAAiB,SAAY,EAAE,cAAc,QAAQ,aAAa,IAAI,CAAC;AAAA,IACnF,GAAI,QAAQ,YAAY,SAAY,EAAE,SAAS,QAAQ,QAAQ,IAAI,CAAC;AAAA,IACpE,GAAI,QAAQ,iBAAiB,SAAY,EAAE,cAAc,QAAQ,aAAa,IAAI,CAAC;AAAA,IACnF,GAAI,QAAQ,mBAAmB,SAAY,EAAE,gBAAgB,QAAQ,eAAe,IAAI,CAAC;AAAA,IACzF,GAAI,QAAQ,mBAAmB,SAAY,EAAE,gBAAgB,QAAQ,eAAe,IAAI,CAAC;AAAA,IACzF,GAAI,QAAQ,SAAS,EAAE,QAAQ,QAAQ,OAAO,IAAI,CAAC;AAAA,IACnD,GAAI,QAAQ,aAAa,EAAE,YAAY,QAAQ,WAAW,IAAI,CAAC;AAAA,IAC/D,GAAI,QAAQ,aAAa,EAAE,YAAY,QAAQ,WAAW,IAAI,CAAC;AAAA,EACjE,CAAC;AAGD,QAAM,UAAU,cAAqC;AAAA,IACnD,MAAM;AAAA,IACN,MAAM,EAAE,SAAS,EAAE,MAAM,iBAAiB,GAAG,SAAS,KAAK;AAAA,IAC3D,WAAW;AAAA,IACX,SAAS,EAAE,MAAM,QAAQ;AAAA,IACzB,WAAW,EAAE,UAAU,uBAAuB,EAAE;AAAA,EAClD,CAAC;AACD,SAAO,OAAO,WAAW;AACvB,UAAM,SAAS,MAAM,eAA8C;AAAA,MACjE;AAAA,MACA;AAAA,MACA,MAAM,QAAQ;AAAA,MACd,QAAQ,QAAQ;AAAA,MAChB;AAAA,IACF,CAAC;AACD,QAAI,OAAO,SAAS,YAAY,OAAO,IAAI,SAAS,QAAQ;AAC1D,YAAM,WACJ,OAAO,SAAS,YAAY,OAAO,IAAI,SAAS,YAC5C,OAAO,IAAI,SAAS,KAAK,IAAI,IAC7B,sBAAsB,OAAO,IAAI;AACvC,YAAM,IAAI,MAAM,2CAA2C,QAAQ,GAAG;AAAA,IACxE;AACA,WAAO,OAAO,IAAI;AAAA,EACpB;AACF;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;;;AChRA,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-EDCVUZZC.js → chunk-CRAH5EY2.js}

@@ -14,7 +14,7 @@ import {
   DELEGATION_STATUS_DESCRIPTION,
   DELEGATION_STATUS_INPUT_SCHEMA,
   DELEGATION_STATUS_TOOL_NAME
-} from "./chunk-QXWGSDAQ.js";
+} from "./chunk-YFOPWG74.js";
 
 // src/mcp/openai-tools.ts
 function buildTool(name, description, parameters) {
@@ -61,4 +61,4 @@ export {
   mcpToolsForRuntimeMcp,
   mcpToolsForRuntimeMcpSubset
 };
-//# sourceMappingURL=chunk-EDCVUZZC.js.map
+//# sourceMappingURL=chunk-CRAH5EY2.js.map

package/dist/{chunk-L5ZFBVT6.js → chunk-HY4YOQM2.js}

@@ -7,7 +7,7 @@ import {
   runDetachedTurn,
   runLoop,
   selectValidWinner
-} from "./chunk-QXWGSDAQ.js";
+} from "./chunk-YFOPWG74.js";
 import {
   coderProfile,
   coderTaskToPrompt
@@ -472,4 +472,4 @@ export {
   coderTaskFromArgs,
   settleDetachedCoderTurn
 };
-//# sourceMappingURL=chunk-L5ZFBVT6.js.map
+//# sourceMappingURL=chunk-HY4YOQM2.js.map

package/dist/{chunk-ZNQVMMR5.js → chunk-LJZ5GC6C.js}

@@ -1,10 +1,10 @@
 import {
   createFleetWorkspaceExecutor,
   createSiblingSandboxExecutor
-} from "./chunk-L5ZFBVT6.js";
+} from "./chunk-HY4YOQM2.js";
 import {
   runWorktreeHarness
-} from "./chunk-QXWGSDAQ.js";
+} from "./chunk-YFOPWG74.js";
 import {
   buildLoopOtelSpans,
   createOtelExporter
@@ -275,4 +275,4 @@ export {
   createPropagatingTraceEmitter,
   traceContextToEnv
 };
-//# sourceMappingURL=chunk-ZNQVMMR5.js.map
+//# sourceMappingURL=chunk-LJZ5GC6C.js.map

package/dist/chunk-P5OKDSLB.js

@@ -0,0 +1,580 @@
+import {
+  AnalystError,
+  extractLlmCallEvent
+} from "./chunk-VLF5RHEQ.js";
+
+// src/analyst-loop/iterations-to-trace-store.ts
+import {
+  DEFAULT_TRACE_ANALYST_BUDGETS,
+  TRACE_ANALYST_TRUNCATION_MARKER_PREFIX
+} from "@tangle-network/agent-eval";
+var bytesOf = (v) => Buffer.byteLength(JSON.stringify(v) ?? "", "utf8");
+var iso = (ms) => new Date(ms).toISOString();
+function normalizeSignature(message) {
+  return message.replace(/0x[0-9a-fA-F]+/g, "HEX").replace(/[0-9a-fA-F]{8}-[0-9a-fA-F]{4}-[0-9a-fA-F]{4}-[0-9a-fA-F]{4}-[0-9a-fA-F]{12}/g, "UUID").replace(/(\/[\w.-]+){2,}/g, "PATH").replace(/\b\d+(\.\d+)?(ms|s|m|h)\b/g, "DUR").replace(/\b\d+\b/g, "#").replace(/\s+/g, " ").trim().slice(0, 200);
+}
+function spanKindFor(event, agentRunName) {
+  const llm = extractLlmCallEvent(event, agentRunName);
+  if (llm) return { kind: "LLM", model: llm.model ?? null, tool: null };
+  const type = String(event?.type ?? "");
+  if (/tool/i.test(type)) {
+    const d = event?.data;
+    const tool = typeof d?.name === "string" ? d.name : typeof d?.tool === "string" ? d.tool : type;
+    return { kind: "TOOL", model: null, tool };
+  }
+  return { kind: "SPAN", model: null, tool: null };
+}
+function errorMessageOf(event) {
+  const type = String(event?.type ?? "");
+  const d = event?.data;
+  if (/error|fail/i.test(type) || d?.error) {
+    const m = d?.error ?? d?.message;
+    return typeof m === "string" ? m : `${type} error`;
+  }
+  return void 0;
+}
+function projectIteration(iter) {
+  const traceId = iter.events.find((e) => e?.data?.sandboxId)?.data?.sandboxId ?? `iter-${iter.index}`;
+  const start = iso(iter.startedAt);
+  const end = iso(iter.endedAt || iter.startedAt);
+  const durationMs = Math.max(0, (iter.endedAt || iter.startedAt) - iter.startedAt);
+  const rootId = `${traceId}:root`;
+  const iterErrored = Boolean(iter.error) || iter.verdict?.valid === false;
+  const spans = [
+    {
+      trace_id: traceId,
+      span_id: rootId,
+      parent_span_id: null,
+      name: iter.agentRunName,
+      kind: "AGENT",
+      start_time: start,
+      end_time: end,
+      duration_ms: durationMs,
+      status: iter.error ? "ERROR" : "OK",
+      status_message: iter.error?.message,
+      service_name: "agent-runtime",
+      agent_name: iter.agentRunName,
+      model_name: null,
+      tool_name: null,
+      attributes: {
+        "iteration.index": iter.index,
+        "verdict.valid": iter.verdict?.valid,
+        "verdict.score": iter.verdict?.score,
+        "output.preview": iter.output === void 0 ? void 0 : String(iter.output).slice(0, 2e3)
+      }
+    }
+  ];
+  const models = /* @__PURE__ */ new Set();
+  const tools = /* @__PURE__ */ new Set();
+  iter.events.forEach((event, i) => {
+    const { kind, model, tool } = spanKindFor(event, iter.agentRunName);
+    if (model) models.add(model);
+    if (tool) tools.add(tool);
+    const errMsg = errorMessageOf(event);
+    spans.push({
+      trace_id: traceId,
+      span_id: `${traceId}:e${i}`,
+      parent_span_id: rootId,
+      name: String(event?.type ?? "event"),
+      kind,
+      start_time: start,
+      end_time: end,
+      duration_ms: 0,
+      status: errMsg ? "ERROR" : "OK",
+      status_message: errMsg,
+      service_name: "agent-runtime",
+      agent_name: iter.agentRunName,
+      model_name: model,
+      tool_name: tool,
+      attributes: event?.data ?? {}
+    });
+  });
+  const hasErrors = spans.some((s) => s.status === "ERROR") || iterErrored;
+  const summary = {
+    trace_id: traceId,
+    service_name: "agent-runtime",
+    agent_name: iter.agentRunName,
+    span_count: spans.length,
+    has_errors: hasErrors,
+    start_time: start,
+    end_time: end,
+    duration_ms: durationMs,
+    raw_jsonl_bytes: bytesOf(spans),
+    models: [...models],
+    tools: [...tools]
+  };
+  return { summary, spans, rawBytes: summary.raw_jsonl_bytes };
+}
+function matchesFilters(t, f) {
+  if (!f) return true;
+  if (f.has_errors !== void 0 && t.summary.has_errors !== f.has_errors) return false;
+  if (f.service_names?.length && !f.service_names.includes(t.summary.service_name ?? ""))
+    return false;
+  if (f.agent_names?.length && !f.agent_names.includes(t.summary.agent_name ?? "")) return false;
+  if (f.model_names?.length && !f.model_names.some((m) => t.summary.models.includes(m)))
+    return false;
+  if (f.tool_names?.length && !f.tool_names.some((tn) => t.summary.tools.includes(tn))) return false;
+  if (f.start_time_after && t.summary.start_time < f.start_time_after) return false;
+  if (f.start_time_before && t.summary.start_time > f.start_time_before) return false;
+  if (f.regex_pattern && !new RegExp(f.regex_pattern).test(JSON.stringify(t.spans))) return false;
+  return true;
+}
+function capAttributes(attributes, perAttrCap) {
+  let truncated = 0;
+  const capped = {};
+  for (const [k, v] of Object.entries(attributes)) {
+    const s = typeof v === "string" ? v : JSON.stringify(v);
+    if (typeof s === "string" && s.length > perAttrCap) {
+      truncated += 1;
+      capped[k] = `${TRACE_ANALYST_TRUNCATION_MARKER_PREFIX} ${s.length}b]${s.slice(0, perAttrCap)}`;
+    } else {
+      capped[k] = v;
+    }
+  }
+  return { capped, truncated };
+}
+function iterationsToTraceStore(iterations, budgets = DEFAULT_TRACE_ANALYST_BUDGETS) {
+  if (iterations.length === 0) {
+    throw new AnalystError("iterationsToTraceStore: no iterations to analyze (empty round)");
+  }
+  const traces = iterations.map((it) => projectIteration(it));
+  const byId = new Map(traces.map((t) => [t.summary.trace_id, t]));
+  const buildClusters = (set) => {
+    const map = /* @__PURE__ */ new Map();
+    for (const t of set) {
+      for (const s of t.spans) {
+        if (s.status !== "ERROR" || !s.status_message) continue;
+        const sig = normalizeSignature(s.status_message);
+        const c = map.get(sig) ?? {
+          signature: sig,
+          status_message_sample: s.status_message,
+          span_name: s.name,
+          tool_name: s.tool_name,
+          trace_count: 0,
+          span_count: 0,
+          prevalence: 0,
+          exemplar_trace_ids: [],
+          exemplar_span_ids: []
+        };
+        c.span_count += 1;
+        if (!c.exemplar_trace_ids.includes(t.summary.trace_id) && c.exemplar_trace_ids.length < 10) {
+          c.exemplar_trace_ids.push(t.summary.trace_id);
+          c.trace_count += 1;
+        }
+        if (c.exemplar_span_ids.length < 10) c.exemplar_span_ids.push(s.span_id);
+        map.set(sig, c);
+      }
+    }
+    const errorTraces = set.filter((t) => t.summary.has_errors).length || 1;
+    const clusters = [...map.values()].map((c) => ({
+      ...c,
+      prevalence: c.trace_count / errorTraces
+    }));
+    return clusters.sort((a, b) => b.trace_count - a.trace_count);
+  };
+  return {
+    async getOverview(filters) {
+      const set = traces.filter((t) => matchesFilters(t, filters));
+      const services = /* @__PURE__ */ new Set();
+      const agents = /* @__PURE__ */ new Set();
+      const models = /* @__PURE__ */ new Set();
+      const tools = /* @__PURE__ */ new Set();
+      let errorSpans = 0;
+      for (const t of set) {
+        if (t.summary.service_name) services.add(t.summary.service_name);
+        if (t.summary.agent_name) agents.add(t.summary.agent_name);
+        for (const m of t.summary.models) models.add(m);
+        for (const tn of t.summary.tools) tools.add(tn);
+        errorSpans += t.spans.filter((s) => s.status === "ERROR").length;
+      }
+      const times = set.map((t) => t.summary.start_time).sort();
+      return {
+        total_traces: set.length,
+        raw_jsonl_bytes: set.reduce((n, t) => n + t.rawBytes, 0),
+        services: [...services],
+        agents: [...agents],
+        models: [...models],
+        tool_names: [...tools],
+        sample_trace_ids: set.slice(0, 20).map((t) => t.summary.trace_id),
+        errors: {
+          trace_count: set.filter((t) => t.summary.has_errors).length,
+          span_count: errorSpans
+        },
+        error_clusters: buildClusters(set),
+        time_range: times.length ? { earliest: times[0], latest: times[times.length - 1] } : null
+      };
+    },
+    async queryTraces(opts) {
+      const set = traces.filter((t) => matchesFilters(t, opts.filters));
+      const offset = opts.offset ?? 0;
+      const page = set.slice(offset, offset + opts.limit);
+      return {
+        traces: page.map((t) => t.summary),
+        total: set.length,
+        has_more: offset + opts.limit < set.length
+      };
+    },
+    async countTraces(filters) {
+      return traces.filter((t) => matchesFilters(t, filters)).length;
+    },
+    async viewTrace(opts) {
+      const t = byId.get(opts.trace_id);
+      if (!t) return { trace_id: opts.trace_id, spans: [] };
+      const cap = opts.per_attribute_byte_cap ?? budgets.perAttributeViewBudget;
+      const projected = t.spans.map((s) => ({
+        ...s,
+        attributes: capAttributes(s.attributes, cap).capped
+      }));
+      if (bytesOf(projected) > budgets.perCallByteCeiling) {
+        const names = /* @__PURE__ */ new Map();
+        for (const s of t.spans) names.set(s.name, (names.get(s.name) ?? 0) + 1);
+        return {
+          trace_id: opts.trace_id,
+          oversized: {
+            span_count: t.spans.length,
+            top_span_names: [...names.entries()].sort((a, b) => b[1] - a[1]).slice(0, 20),
+            span_response_bytes_max: Math.max(...t.spans.map((s) => bytesOf(s))),
+            error_span_count: t.spans.filter((s) => s.status === "ERROR").length
+          }
+        };
+      }
+      return { trace_id: opts.trace_id, spans: projected };
+    },
+    async viewSpans(opts) {
+      const t = byId.get(opts.trace_id);
+      const cap = opts.per_attribute_byte_cap ?? budgets.perAttributeSpanBudget;
+      const want = new Set(opts.span_ids);
+      const found = (t?.spans ?? []).filter((s) => want.has(s.span_id));
+      let truncated = 0;
+      const spans = found.map((s) => {
+        const { capped, truncated: n } = capAttributes(s.attributes, cap);
+        truncated += n;
+        return { ...s, attributes: capped };
+      });
+      const foundIds = new Set(found.map((s) => s.span_id));
+      return {
+        trace_id: opts.trace_id,
+        spans,
+        missing_span_ids: opts.span_ids.filter((id) => !foundIds.has(id)),
+        truncated_attribute_count: truncated
+      };
+    },
+    async searchTrace(opts) {
+      const t = byId.get(opts.trace_id);
+      const max = opts.max_matches ?? 50;
+      const hits = [];
+      for (const s of t?.spans ?? []) {
+        for (const hit of searchSpanAttrs(s, opts.regex_pattern, budgets.perMatchTextBudget)) {
+          if (hits.length >= max) break;
+          hits.push(hit);
+        }
+      }
+      return {
+        trace_id: opts.trace_id,
+        hits,
+        total_matches: hits.length,
+        has_more: hits.length >= max
+      };
+    },
+    async searchSpan(opts) {
+      const t = byId.get(opts.trace_id);
+      const max = opts.max_matches ?? 50;
+      const span = (t?.spans ?? []).find((s) => s.span_id === opts.span_id);
+      const hits = span ? searchSpanAttrs(span, opts.regex_pattern, budgets.perMatchTextBudget).slice(0, max) : [];
+      return {
+        trace_id: opts.trace_id,
+        span_id: opts.span_id,
+        hits,
+        total_matches: hits.length,
+        has_more: false
+      };
+    }
+  };
+}
+function searchSpanAttrs(span, pattern, textCap) {
+  const re = new RegExp(pattern, "g");
+  const hits = [];
+  for (const [k, v] of Object.entries(span.attributes)) {
+    const text = typeof v === "string" ? v : JSON.stringify(v);
+    if (typeof text !== "string") continue;
+    re.lastIndex = 0;
+    const m = re.exec(text);
+    if (!m) continue;
+    const at = m.index;
+    hits.push({
+      trace_id: span.trace_id,
+      span_id: span.span_id,
+      span_name: span.name,
+      span_kind: span.kind,
+      attribute_path: `attributes.${k}`,
+      matched_text: m[0].slice(0, textCap),
+      context_before: text.slice(Math.max(0, at - textCap / 2), at),
+      context_after: text.slice(at + m[0].length, at + m[0].length + textCap / 2),
+      match_offset: at
+    });
+  }
+  return hits;
+}
+
+// src/analyst-loop/run-analyst-loop.ts
+import { diffFindings } from "@tangle-network/agent-eval";
+async function runAnalystLoop(opts) {
+  const log = opts.log ?? defaultLog;
+  const strategy = opts.priorFindingsStrategy ?? "per-kind";
+  const emit = makeEmitter(opts.onEvent);
+  const startedAt = Date.now();
+  const baselineRunId = resolveBaselineRunId(opts);
+  const priorAll = baselineRunId ? opts.findingsStore?.loadRun(baselineRunId) ?? [] : [];
+  log("baseline resolved", { baselineRunId, prior_findings: priorAll.length });
+  await emit({
+    type: "baseline-resolved",
+    runId: opts.runId,
+    baselineRunId,
+    priorFindingCount: priorAll.length
+  });
+  const priorFindings = buildPriorFindingsInput(priorAll, strategy, opts.registry.list());
+  const analystResult = await runRegistry(opts, priorFindings, emit);
+  log("analyst run complete", {
+    findings: analystResult.findings.length,
+    cost_usd: analystResult.total_cost_usd,
+    per_analyst: analystResult.per_analyst.map((s) => ({
+      id: s.analyst_id,
+      status: s.status,
+      n: s.findings_count
+    }))
+  });
+  if (opts.findingsStore && analystResult.findings.length > 0) {
+    await opts.findingsStore.append(opts.runId, analystResult.findings);
+    await emit({
+      type: "findings-persisted",
+      runId: opts.runId,
+      count: analystResult.findings.length
+    });
+  }
+  let diff = null;
+  if (baselineRunId && analystResult.findings.length > 0) {
+    diff = diffFindings(
+      priorAll.map((f) => ({ ...f })),
+      analystResult.findings.map((f) => ({ ...f, run_id: opts.runId }))
+    );
+    log("diff vs baseline", {
+      appeared: diff.appeared.length,
+      disappeared: diff.disappeared.length,
+      persisted: diff.persisted.length,
+      changed: diff.changed.length
+    });
+    await emit({
+      type: "diff-computed",
+      runId: opts.runId,
+      baselineRunId,
+      appeared: diff.appeared.length,
+      disappeared: diff.disappeared.length,
+      persisted: diff.persisted.length,
+      changed: diff.changed.length
+    });
+  }
+  let knowledge = null;
+  if (opts.knowledgeAdapter) {
+    knowledge = await runKnowledgeAdapter(opts, analystResult.findings, log, emit);
+  }
+  let improvement = null;
+  if (opts.improvementAdapter) {
+    improvement = await runImprovementAdapter(opts, analystResult.findings, log, emit);
+  }
+  await emit({
+    type: "loop-completed",
+    runId: opts.runId,
+    durationMs: Date.now() - startedAt
+  });
+  return {
+    runId: opts.runId,
+    baselineRunId,
+    analystResult,
+    diff,
+    knowledge,
+    improvement
+  };
+}
+function makeEmitter(onEvent) {
+  if (!onEvent) return async () => {
+  };
+  return async (event) => {
+    await onEvent(event);
+  };
+}
+async function runRegistry(opts, priorFindings, emit) {
+  const reg = opts.registry;
+  if (typeof reg.runStream === "function" && opts.onEvent) {
+    let final = null;
+    for await (const ev of reg.runStream(opts.runId, opts.inputs, { priorFindings })) {
+      await emit({ type: "analyst", runId: opts.runId, event: ev });
+      if (ev.type === "run-completed") final = ev.result;
+    }
+    if (!final) {
+      throw new Error("runAnalystLoop: registry.runStream ended without run-completed event");
+    }
+    return final;
+  }
+  return opts.registry.run(opts.runId, opts.inputs, { priorFindings });
+}
+function resolveBaselineRunId(opts) {
+  if (opts.baselineRunId === null) return null;
+  if (typeof opts.baselineRunId === "string") return opts.baselineRunId;
+  if (!opts.findingsStore) return null;
+  const all = opts.findingsStore.loadAll();
+  let last = null;
+  for (const row of all) {
+    if (row.run_id === opts.runId) continue;
+    last = row.run_id;
+  }
+  return last;
+}
+function buildPriorFindingsInput(prior, strategy, registry) {
+  if (strategy === "none" || prior.length === 0) return void 0;
+  const stripped = prior.map(({ run_id: _run_id, ...rest }) => rest);
+  if (strategy === "wildcard") {
+    return { "*": stripped };
+  }
+  void registry;
+  return stripped;
+}
+async function runKnowledgeAdapter(opts, findings, log, emit) {
+  const adapter = opts.knowledgeAdapter;
+  const batch = await adapter.proposeFromFindings(findings);
+  log("knowledge.proposeFromFindings", {
+    proposals: batch.proposals.length,
+    skipped: batch.skipped,
+    errors: batch.errors.length
+  });
+  await emit({
+    type: "knowledge-proposed",
+    runId: opts.runId,
+    proposalCount: batch.proposals.length,
+    skipped: batch.skipped,
+    errors: batch.errors.length
+  });
+  const auto = opts.autoApply?.knowledge ?? false;
+  const threshold = opts.autoApply?.knowledgeConfidenceThreshold ?? 0.85;
+  if (!auto || !adapter.apply) {
+    await emit({
+      type: "knowledge-applied",
+      runId: opts.runId,
+      writtenCount: 0,
+      withheldForReview: batch.proposals.length
+    });
+    return {
+      proposals: batch.proposals,
+      applied: [],
+      skipped: batch.skipped,
+      errors: batch.errors,
+      withheld_for_review: batch.proposals.length
+    };
+  }
+  const findingsById = new Map(findings.map((f) => [f.finding_id, f]));
+  const safe = [];
+  let withheld = 0;
+  for (const p of batch.proposals) {
+    const src = p.sourceFindingId ? findingsById.get(p.sourceFindingId) : void 0;
+    if (!src) {
+      withheld += 1;
+      continue;
+    }
+    if (src.confidence < threshold) {
+      withheld += 1;
+      continue;
+    }
+    safe.push(p);
+  }
+  const result = await adapter.apply(safe);
+  log("knowledge.apply", {
+    applied: result.written.length,
+    withheld_for_review: withheld,
+    warnings: result.warnings.length
+  });
+  await emit({
+    type: "knowledge-applied",
+    runId: opts.runId,
+    writtenCount: result.written.length,
+    withheldForReview: withheld
+  });
+  return {
+    proposals: batch.proposals,
+    applied: result.written,
+    skipped: batch.skipped,
+    errors: batch.errors,
+    withheld_for_review: withheld
+  };
+}
+async function runImprovementAdapter(opts, findings, log, emit) {
+  const adapter = opts.improvementAdapter;
+  const batch = await adapter.proposeFromFindings(findings);
+  log("improvement.proposeFromFindings", {
+    edits: batch.edits.length,
+    skipped: batch.skipped,
+    errors: batch.errors.length
+  });
+  await emit({
+    type: "improvement-proposed",
+    runId: opts.runId,
+    editCount: batch.edits.length,
+    skipped: batch.skipped,
+    errors: batch.errors.length
+  });
+  const auto = opts.autoApply?.improvement ?? false;
+  const threshold = opts.autoApply?.improvementConfidenceThreshold ?? 0.9;
+  if (!auto || !adapter.apply) {
+    await emit({
+      type: "improvement-applied",
+      runId: opts.runId,
+      appliedCount: 0,
+      withheldForReview: batch.edits.length
+    });
+    return {
+      edits: batch.edits,
+      applied: [],
+      skipped: batch.skipped,
+      errors: batch.errors,
+      withheld_for_review: batch.edits.length
+    };
+  }
+  const findingsById = new Map(findings.map((f) => [f.finding_id, f]));
+  const safe = [];
+  let withheld = 0;
+  for (const e of batch.edits) {
+    const src = e.sourceFindingId ? findingsById.get(e.sourceFindingId) : void 0;
+    if (!src || src.confidence < threshold) {
+      withheld += 1;
+      continue;
+    }
+    safe.push(e);
+  }
+  const result = await adapter.apply(safe);
+  log("improvement.apply", {
+    applied: result.applied.length,
+    withheld_for_review: withheld,
+    warnings: result.warnings.length
+  });
+  await emit({
+    type: "improvement-applied",
+    runId: opts.runId,
+    appliedCount: result.applied.length,
+    withheldForReview: withheld
+  });
+  return {
+    edits: batch.edits,
+    applied: result.applied,
+    skipped: batch.skipped,
+    errors: batch.errors,
+    withheld_for_review: withheld
+  };
+}
+function defaultLog(msg, fields) {
+  if (fields) console.log(`[analyst-loop] ${msg}`, fields);
+  else console.log(`[analyst-loop] ${msg}`);
+}
+
+export {
+  iterationsToTraceStore,
+  runAnalystLoop
+};
+//# sourceMappingURL=chunk-P5OKDSLB.js.map