@tangle-network/agent-runtime 0.103.1 → 0.104.0

package/dist/{chunk-AAN2MB2X.js → chunk-3PPVCS75.js}

@@ -5,10 +5,10 @@ import {
   definePersona,
   runPersonified,
   worktreeFanout
-} from "./chunk-3LJF5XSE.js";
+} from "./chunk-ATDFUJYI.js";
 import {
   createExecutorRegistry
-} from "./chunk-EAQ5YRRY.js";
+} from "./chunk-5HNDYWUT.js";
 import {
   runAnalystLoop
 } from "./chunk-FD2MBMOH.js";
@@ -196,4 +196,4 @@ export {
   runLoopRunnerCli,
   parseLoopRunnerArgv
 };
-//# sourceMappingURL=chunk-AAN2MB2X.js.map
+//# sourceMappingURL=chunk-3PPVCS75.js.map

package/dist/chunk-3PPVCS75.js.map

@@ -0,0 +1 @@
+{"version":3,"sources":["../src/loop-runner.ts","../src/loop-runner-bin.ts"],"sourcesContent":["/**\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 on the GENERIC recursive path (worktreeLoopRunner: author one\n *                  `AgentProfile` per harness → worktree-CLI leaves → `patchDelivered` gate)\n *   review       → caller-registered runner — a `code` runner with an approval gate over candidates\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 * @experimental\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 { type CreateKbGateOptions, createKbGate, type FactCandidate } from './mcp/kb-gate'\nimport {\n  type AuthoredHarness,\n  type Budget,\n  createExecutorRegistry,\n  definePersona,\n  runPersonified,\n  type WinnerStrategy,\n  type WorktreeFanoutOptions,\n  type WorktreePatchArtifact,\n  worktreeFanout,\n} from './runtime'\n\n/** All valid delegated-loop mode names — used for validation and CLI surfaces. @experimental */\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/** Type guard — returns true when `value` is a valid `DelegatedLoopMode` string. @experimental */\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 *\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 *\n * @experimental\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 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 *\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. The sandbox-session counterpart that drives the in-box\n * harness over a `SandboxClient` is `detachedSessionDelegate` (`./mcp/delegates`); here there is no\n * `runAgentRounds` driver, no role-coupled delegate — the harness list is the fanout, the gate is\n * `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 *\n * @experimental\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 * `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 *\n * @experimental\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/**\n * `self-improve` mode — agent-eval's one-call closed improvement loop (held-out gated).\n *\n * @experimental\n */\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/**\n * `audit` mode — analyst loop over captured trace/run data.\n *\n * @experimental\n */\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 *\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 * @experimental\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 *\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 *\n * @experimental\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;AAkBA,IAAM,uBAAuB,CAAC,QAAQ,UAAU,YAAY,SAAS,cAAc;AAMnF,SAAS,oBAAoB,OAA4C;AAC9E,SAAO,OAAO,UAAU,YAAa,qBAA2C,SAAS,KAAK;AAChG;AAoCA,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;AA4CO,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;AA4CO,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;AAOO,SAAS,sBACd,SAC8D;AAC9D,SAAO,YAAY,YAAkC,OAAO;AAC9D;AAOO,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-EAQ5YRRY.js → chunk-5HNDYWUT.js}

@@ -18,7 +18,7 @@ import {
   runWorktreeChecks,
   runWorktreeHarness,
   withDriverExecutor
-} from "./chunk-6WZZXQV5.js";
+} from "./chunk-2X65HYMI.js";
 import {
   FileResultBlobStore,
   FileSpawnJournal,
@@ -29,7 +29,7 @@ import {
 import {
   providerAsExecutor,
   resolveAgentEnvironmentProvider
-} from "./chunk-M6MD6JBS.js";
+} from "./chunk-KRBFHMV6.js";
 import {
   addTokenUsage,
   deleteBoxSafe,
@@ -69,6 +69,17 @@ function gateOnDeliverable(inner, deliverable) {
     runtime: inner.runtime,
     ...inner.budgetExempt !== void 0 ? { budgetExempt: inner.budgetExempt } : {},
     ...inner.deliver ? { deliver: (m) => inner.deliver?.(m) } : {},
+    // Forward the OPTIONAL live-observation surfaces so a gated worker stays supervisable
+    // mid-flight. The scope captures `progress`/`traceSource` at spawn and only if the executor
+    // exposes them; a gate that drops them makes `observe_agent` read `recentActivity:[]` and
+    // loses online detection for a running worker (it saw the tokens/turns the fold derives, but
+    // none of the harness's own turn count or tool activity). `metered` is a driver-executor's
+    // OWN-inference subtree total, re-homed by the parent on settle — dropping it would leak a
+    // gated sub-driver's inference out of the journal. Preserve the "undefined ⟺ not implemented"
+    // contract: only expose a method the inner executor actually implements.
+    ...inner.progress ? { progress: () => inner.progress?.() } : {},
+    ...inner.traceSource ? { traceSource: () => inner.traceSource?.() } : {},
+    ...inner.metered ? { metered: () => inner.metered?.() } : {},
     execute(task, signal) {
       const r = inner.execute(task, signal);
       if (isAsyncIterable(r)) {
@@ -369,19 +380,19 @@ async function destroyBounded(box) {
 // src/runtime/run-loop.ts
 var DEFAULT_MAX_ITERATIONS = 10;
 var DEFAULT_MAX_CONCURRENCY = 4;
-async function runLoop(options) {
+async function runAgentRounds(options) {
   const specs = resolveAgentRuns(options);
   const maxIterations = options.maxIterations ?? DEFAULT_MAX_ITERATIONS;
   if (!Number.isFinite(maxIterations) || maxIterations <= 0) {
-    throw new ValidationError("runLoop: maxIterations must be > 0");
+    throw new ValidationError("runAgentRounds: maxIterations must be > 0");
   }
   const maxConcurrency = options.maxConcurrency ?? DEFAULT_MAX_CONCURRENCY;
   if (!Number.isFinite(maxConcurrency) || maxConcurrency <= 0) {
-    throw new ValidationError("runLoop: maxConcurrency must be > 0");
+    throw new ValidationError("runAgentRounds: maxConcurrency must be > 0");
   }
   const sandboxStreaming = options.lineage?.streaming ?? "sse";
   if (!options.ctx?.sandboxClient || typeof options.ctx.sandboxClient.create !== "function") {
-    throw new ValidationError("runLoop: ctx.sandboxClient.create is required");
+    throw new ValidationError("runAgentRounds: ctx.sandboxClient.create is required");
   }
   const now = options.now ?? Date.now;
   const runId = options.runId ?? `loop-${randomSuffix()}`;
@@ -559,12 +570,13 @@ async function runLoop(options) {
     if (lineageState) await lineageState.lineage.teardown();
   }
 }
+var runLoop = runAgentRounds;
 async function setUpLineage(options, maxConcurrency, recordMount) {
   const lineageOpts = options.lineage;
   if (!lineageOpts || !lineageOpts.sessionContinuity && !lineageOpts.forkFanout) return void 0;
   if (options.onWorkerBox) {
     throw new ValidationError(
-      "runLoop: `lineage` and `onWorkerBox` both own worker boxes \u2014 pass only one"
+      "runAgentRounds: `lineage` and `onWorkerBox` both own worker boxes \u2014 pass only one"
     );
   }
   const capabilities = await probeSandboxCapabilities(options.ctx.sandboxClient);
@@ -584,12 +596,14 @@ function planLineageRound(state, specs, slice, parentIndex, signal) {
   const parent = parentIndex !== void 0 ? state.handles.get(parentIndex) : void 0;
   const promptFor = (offset) => {
     const spec = specs[offset % specs.length];
-    if (!spec) throw new ValidationError("runLoop: no AgentRunSpec available for lineage iteration");
+    if (!spec)
+      throw new ValidationError("runAgentRounds: no AgentRunSpec available for lineage iteration");
     return spec.taskToPrompt(slice[offset]);
   };
   const specAt = (offset) => {
     const spec = specs[offset % specs.length];
-    if (!spec) throw new ValidationError("runLoop: no AgentRunSpec available for lineage iteration");
+    if (!spec)
+      throw new ValidationError("runAgentRounds: no AgentRunSpec available for lineage iteration");
     return spec;
   };
   if (slice.length === 1 && parent && state.options.sessionContinuity) {
@@ -615,7 +629,7 @@ function planLineageRound(state, specs, slice, parentIndex, signal) {
         const branches = await ensureForked();
         const branch = branches[offset];
         if (!branch)
-          throw new ValidationError("runLoop: lineage fork produced no branch for offset");
+          throw new ValidationError("runAgentRounds: lineage fork produced no branch for offset");
         return branch;
       }
     }));
@@ -673,9 +687,9 @@ async function runBatch(args) {
 async function executeIteration(args) {
   const slot = args.iterations[args.item.index];
   if (!slot)
-    throw new ValidationError(`runLoop: missing iteration slot at index ${args.item.index}`);
+    throw new ValidationError(`runAgentRounds: missing iteration slot at index ${args.item.index}`);
   const spec = args.specs[args.item.index % args.specs.length];
-  if (!spec) throw new ValidationError("runLoop: no AgentRunSpec available for iteration");
+  if (!spec) throw new ValidationError("runAgentRounds: no AgentRunSpec available for iteration");
   slot.startedAt = args.now();
   slot.agentRunName = spec.name ?? spec.profile.name ?? "agent";
   await emitTrace(args.ctx.traceEmitter, {
@@ -986,11 +1000,11 @@ function defaultSelectWinner(iterations) {
 }
 function resolveAgentRuns(options) {
   if (options.agentRun && options.agentRuns) {
-    throw new ValidationError("runLoop: pass exactly one of `agentRun` or `agentRuns`");
+    throw new ValidationError("runAgentRounds: pass exactly one of `agentRun` or `agentRuns`");
   }
   if (options.agentRun) return [options.agentRun];
   if (options.agentRuns && options.agentRuns.length > 0) return options.agentRuns;
-  throw new ValidationError("runLoop: `agentRun` or non-empty `agentRuns` is required");
+  throw new ValidationError("runAgentRounds: `agentRun` or non-empty `agentRuns` is required");
 }
 function isTerminalDecision(decision) {
   return decision === "stop" || decision === "pick-winner" || decision === "fail" || decision === "done";
@@ -1301,9 +1315,9 @@ var piExecutor = (spec, ctx) => {
   };
   return {
     runtime: PI_RUNTIME,
-    // pi owns the queue; `deliver` only routes. A forceful message becomes pi's `steer` (which
-    // pi delivers after the current tool batch and before the next model call); a queued one
-    // becomes `follow_up` (after the turn finishes). Never throws — the inbox contract.
+    // pi owns the queue; `deliver` only routes through its state-safe `prompt` command. Its
+    // streaming behavior chooses steer versus follow-up atomically in pi, rather than trusting
+    // this adapter's delayed view of whether the current run has already ended.
     deliver: (m) => inbox.deliver(m),
     progress: () => ({
       turns: state.turns,
@@ -1361,6 +1375,7 @@ async function* streamPiSession(args) {
   };
   const stdoutLines = readJsonLines(proc, (value) => {
     const ev = value;
+    if (ev.type === "agent_start") idle = false;
     if (ev.type === "agent_end") idle = true;
     events.push(ev);
     notify();
@@ -1396,7 +1411,7 @@ ${taskText(args.task)}` : taskText(args.task);
     sendCommand(proc, { type: "prompt", message: opening });
     state.note = "turn 0";
     for (; ; ) {
-      forwardPending(proc, inbox, activity);
+      if (!idle) forwardPending(proc, inbox, activity);
       while (events.length > 0) {
         const ev = events.shift();
         for (const usage of projectPiEvent(ev, args, tokens)) {
@@ -1445,8 +1460,9 @@ function forwardPending(proc, inbox, activity) {
   const pending = inbox.drain();
   for (const m of pending) {
     sendCommand(proc, {
-      type: m.interrupt ? "steer" : "follow_up",
-      message: renderOne(m)
+      type: "prompt",
+      message: renderOne(m),
+      streamingBehavior: m.interrupt ? "steer" : "followUp"
     });
     activity.push({
       at: Date.now(),
@@ -2299,7 +2315,7 @@ async function* streamSandboxLeaf(args) {
     ...args.seam.lineage !== void 0 ? { lineage: args.seam.lineage } : {}
   };
   try {
-    const result = await runLoop(loopOptions);
+    const result = await runAgentRounds(loopOptions);
     const out = result.winner?.output ?? { events: [] };
     const verdict = result.winner?.verdict;
     const spent = {
@@ -6310,6 +6326,7 @@ export {
   acquireSandbox,
   probeSandboxCapabilities,
   createSandboxLineage,
+  runAgentRounds,
   runLoop,
   createSandboxForSpec,
   defaultSelectWinner,
@@ -6404,4 +6421,4 @@ export {
   workerFromBackend,
   supervise
 };
-//# sourceMappingURL=chunk-EAQ5YRRY.js.map
+//# sourceMappingURL=chunk-5HNDYWUT.js.map