@tangle-network/agent-eval 1.0.18 → 1.0.20

package/dist/{single-run-lock-BGlW3OSk.js.map → single-run-lock-CUj2tjxT.js.map}

@@ -1 +1 @@
-{"version":3,"file":"single-run-lock-BGlW3OSk.js","names":[],"sources":["../src/campaign/single-run-lock.ts"],"sourcesContent":["/**\n * Single-run lock for evaluations that share one mutable environment.\n *\n * Two concurrent runs against a shared stateful gym silently corrupt each\n * other: each resets/mutates environment state mid-cell of the other, and\n * every score from both becomes garbage that LOOKS like worker variance\n * (agent-lab R357 burned hours on flip-flopping scores before tracing them\n * to exactly this). The fix is a holder lockfile: refuse to start while the\n * holder may be alive, reclaim a lock whose holder is proved dead\n * (`lockHolderVerdict`), release only if the lock is still ours.\n *\n * `alsoCheck` exists because independent runners can guard the same shared\n * resource with differently named lockfiles; a runner must respect all of\n * them even though it writes only its own.\n */\n\nimport {\n  type AtomicFileLockUnavailable,\n  probeAtomicFileLock,\n  tryAcquireAtomicFileLock,\n} from '../ledger-core/atomic-file-lock'\n\nexport interface SingleRunLockOptions {\n  /** Lockfile this runner writes (and checks). */\n  readonly lockPath: string\n  /** Other runners' lockfiles guarding the same resource; checked, never written. */\n  readonly alsoCheck?: readonly string[]\n  /** Install a process 'exit' hook that releases the lock. Default true. */\n  readonly releaseOnExit?: boolean\n  /** Owner pid recorded in the lockfile metadata. Default process.pid. */\n  readonly pid?: number\n}\n\nexport interface SingleRunLock {\n  /** Remove the lockfile if this process still owns it. Idempotent. */\n  release(): void\n}\n\nfunction assertAvailable(path: string): void {\n  const unavailable = probeAtomicFileLock({ lockPath: path })\n  if (unavailable) throw unavailableError(path, unavailable)\n}\n\nfunction unavailableError(path: string, unavailable: AtomicFileLockUnavailable): Error {\n  if (unavailable.reason === 'recovery') {\n    return new Error(\n      `single-run lock recovery is already in progress (${path}); refusing a concurrent run`,\n    )\n  }\n  return new Error(\n    `single-run lock held by live pid ${unavailable.holder.pid} (${path}); refusing a concurrent run on the shared resource`,\n  )\n}\n\n/**\n * Acquire the lock or throw naming the holder. A lock whose holder is proved\n * dead is reclaimed by one contender. An interrupted reclaim\n * leaves a marker that fails closed instead of admitting overlapping runs.\n */\nexport function acquireSingleRunLock(opts: SingleRunLockOptions): SingleRunLock {\n  const pid = opts.pid ?? process.pid\n  if (!Number.isSafeInteger(pid) || pid <= 0)\n    throw new Error('single-run lock pid must be a positive integer')\n  for (const path of opts.alsoCheck ?? []) assertAvailable(path)\n  const acquisition = tryAcquireAtomicFileLock({\n    lockPath: opts.lockPath,\n    pid,\n  })\n  if (!acquisition.acquired) throw unavailableError(opts.lockPath, acquisition)\n  const releaseOnExit = opts.releaseOnExit ?? true\n  let released = false\n  const release = (): void => {\n    if (released) return\n    released = true\n    if (releaseOnExit) process.off('exit', release)\n    try {\n      acquisition.lock.release()\n    } catch {\n      // release is best-effort; a leftover stale lock is reclaimed on next acquire\n    }\n  }\n  try {\n    for (const path of opts.alsoCheck ?? []) assertAvailable(path)\n  } catch (error) {\n    release()\n    throw error\n  }\n  if (releaseOnExit) process.on('exit', release)\n  return { release }\n}\n"],"mappings":";;;;;;;;;;;;;;;;;AAsCA,SAAS,gBAAgB,MAAoB;CAC3C,MAAM,cAAc,oBAAoB,EAAE,UAAU,KAAK,CAAC;CAC1D,IAAI,aAAa,MAAM,iBAAiB,MAAM,WAAW;AAC3D;AAEA,SAAS,iBAAiB,MAAc,aAA+C;CACrF,IAAI,YAAY,WAAW,YACzB,uBAAO,IAAI,MACT,oDAAoD,KAAK,6BAC3D;CAEF,uBAAO,IAAI,MACT,oCAAoC,YAAY,OAAO,IAAI,IAAI,KAAK,oDACtE;AACF;;;;;;AAOA,SAAgB,qBAAqB,MAA2C;CAC9E,MAAM,MAAM,KAAK,OAAO,QAAQ;CAChC,IAAI,CAAC,OAAO,cAAc,GAAG,KAAK,OAAO,GACvC,MAAM,IAAI,MAAM,gDAAgD;CAClE,KAAK,MAAM,QAAQ,KAAK,aAAa,CAAC,GAAG,gBAAgB,IAAI;CAC7D,MAAM,cAAc,yBAAyB;EAC3C,UAAU,KAAK;EACf;CACF,CAAC;CACD,IAAI,CAAC,YAAY,UAAU,MAAM,iBAAiB,KAAK,UAAU,WAAW;CAC5E,MAAM,gBAAgB,KAAK,iBAAiB;CAC5C,IAAI,WAAW;CACf,MAAM,gBAAsB;EAC1B,IAAI,UAAU;EACd,WAAW;EACX,IAAI,eAAe,QAAQ,IAAI,QAAQ,OAAO;EAC9C,IAAI;GACF,YAAY,KAAK,QAAQ;EAC3B,QAAQ,CAER;CACF;CACA,IAAI;EACF,KAAK,MAAM,QAAQ,KAAK,aAAa,CAAC,GAAG,gBAAgB,IAAI;CAC/D,SAAS,OAAO;EACd,QAAQ;EACR,MAAM;CACR;CACA,IAAI,eAAe,QAAQ,GAAG,QAAQ,OAAO;CAC7C,OAAO,EAAE,QAAQ;AACnB"}
+{"version":3,"file":"single-run-lock-CUj2tjxT.js","names":[],"sources":["../src/campaign/single-run-lock.ts"],"sourcesContent":["/**\n * Single-run lock for evaluations that share one mutable environment.\n *\n * Two concurrent runs against a shared stateful gym silently corrupt each\n * other: each resets/mutates environment state mid-cell of the other, and\n * every score from both becomes garbage that LOOKS like worker variance\n * (agent-lab R357 burned hours on flip-flopping scores before tracing them\n * to exactly this). The fix is a holder lockfile: refuse to start while the\n * holder may be alive, reclaim a lock whose holder is proved dead\n * (`lockHolderVerdict`), release only if the lock is still ours.\n *\n * `alsoCheck` exists because independent runners can guard the same shared\n * resource with differently named lockfiles; a runner must respect all of\n * them even though it writes only its own.\n */\n\nimport {\n  type AtomicFileLockUnavailable,\n  probeAtomicFileLock,\n  tryAcquireAtomicFileLock,\n} from '../ledger-core/atomic-file-lock'\n\nexport interface SingleRunLockOptions {\n  /** Lockfile this runner writes (and checks). */\n  readonly lockPath: string\n  /** Other runners' lockfiles guarding the same resource; checked, never written. */\n  readonly alsoCheck?: readonly string[]\n  /** Install a process 'exit' hook that releases the lock. Default true. */\n  readonly releaseOnExit?: boolean\n  /** Owner pid recorded in the lockfile metadata. Default process.pid. */\n  readonly pid?: number\n}\n\nexport interface SingleRunLock {\n  /** Remove the lockfile if this process still owns it. Idempotent. */\n  release(): void\n}\n\nfunction assertAvailable(path: string): void {\n  const unavailable = probeAtomicFileLock({ lockPath: path })\n  if (unavailable) throw unavailableError(path, unavailable)\n}\n\nfunction unavailableError(path: string, unavailable: AtomicFileLockUnavailable): Error {\n  if (unavailable.reason === 'recovery') {\n    return new Error(\n      `single-run lock recovery is already in progress (${path}); refusing a concurrent run`,\n    )\n  }\n  return new Error(\n    `single-run lock held by live pid ${unavailable.holder.pid} (${path}); refusing a concurrent run on the shared resource`,\n  )\n}\n\n/**\n * Acquire the lock or throw naming the holder. A lock whose holder is proved\n * dead is reclaimed by one contender. An interrupted reclaim\n * leaves a marker that fails closed instead of admitting overlapping runs.\n */\nexport function acquireSingleRunLock(opts: SingleRunLockOptions): SingleRunLock {\n  const pid = opts.pid ?? process.pid\n  if (!Number.isSafeInteger(pid) || pid <= 0)\n    throw new Error('single-run lock pid must be a positive integer')\n  for (const path of opts.alsoCheck ?? []) assertAvailable(path)\n  const acquisition = tryAcquireAtomicFileLock({\n    lockPath: opts.lockPath,\n    pid,\n  })\n  if (!acquisition.acquired) throw unavailableError(opts.lockPath, acquisition)\n  const releaseOnExit = opts.releaseOnExit ?? true\n  let released = false\n  const release = (): void => {\n    if (released) return\n    released = true\n    if (releaseOnExit) process.off('exit', release)\n    try {\n      acquisition.lock.release()\n    } catch {\n      // release is best-effort; a leftover stale lock is reclaimed on next acquire\n    }\n  }\n  try {\n    for (const path of opts.alsoCheck ?? []) assertAvailable(path)\n  } catch (error) {\n    release()\n    throw error\n  }\n  if (releaseOnExit) process.on('exit', release)\n  return { release }\n}\n"],"mappings":";;;;;;;;;;;;;;;;;AAsCA,SAAS,gBAAgB,MAAoB;CAC3C,MAAM,cAAc,oBAAoB,EAAE,UAAU,KAAK,CAAC;CAC1D,IAAI,aAAa,MAAM,iBAAiB,MAAM,WAAW;AAC3D;AAEA,SAAS,iBAAiB,MAAc,aAA+C;CACrF,IAAI,YAAY,WAAW,YACzB,uBAAO,IAAI,MACT,oDAAoD,KAAK,6BAC3D;CAEF,uBAAO,IAAI,MACT,oCAAoC,YAAY,OAAO,IAAI,IAAI,KAAK,oDACtE;AACF;;;;;;AAOA,SAAgB,qBAAqB,MAA2C;CAC9E,MAAM,MAAM,KAAK,OAAO,QAAQ;CAChC,IAAI,CAAC,OAAO,cAAc,GAAG,KAAK,OAAO,GACvC,MAAM,IAAI,MAAM,gDAAgD;CAClE,KAAK,MAAM,QAAQ,KAAK,aAAa,CAAC,GAAG,gBAAgB,IAAI;CAC7D,MAAM,cAAc,yBAAyB;EAC3C,UAAU,KAAK;EACf;CACF,CAAC;CACD,IAAI,CAAC,YAAY,UAAU,MAAM,iBAAiB,KAAK,UAAU,WAAW;CAC5E,MAAM,gBAAgB,KAAK,iBAAiB;CAC5C,IAAI,WAAW;CACf,MAAM,gBAAsB;EAC1B,IAAI,UAAU;EACd,WAAW;EACX,IAAI,eAAe,QAAQ,IAAI,QAAQ,OAAO;EAC9C,IAAI;GACF,YAAY,KAAK,QAAQ;EAC3B,QAAQ,CAER;CACF;CACA,IAAI;EACF,KAAK,MAAM,QAAQ,KAAK,aAAa,CAAC,GAAG,gBAAgB,IAAI;CAC/D,SAAS,OAAO;EACd,QAAQ;EACR,MAAM;CACR;CACA,IAAI,eAAe,QAAQ,GAAG,QAAQ,OAAO;CAC7C,OAAO,EAAE,QAAQ;AACnB"}

package/dist/{skillopt-optimization-method-D72XZZfR.js → skillopt-optimization-method-DYwEndON.js}

@@ -1,20 +1,20 @@
 import { a as hashCanonical } from "./canonical-CicJlINA.js";
-import { r as meanOrZero } from "./mean-D1d6gedE.js";
-import { t as decidePairedPromotion } from "./paired-promotion-decision-CiueaTTj.js";
-import { a as heldoutSignificance, o as pairHoldout } from "./statistical-heldout-NBxC7jGo.js";
-import { d as mapConcurrent } from "./journal-C1uLJi0w.js";
+import { lt as meanOrZero } from "./statistics-DCZIC-qT.js";
+import { t as decidePairedPromotion } from "./paired-promotion-decision-Nt4CkkOT.js";
+import { a as heldoutSignificance, o as pairHoldout } from "./statistical-heldout-D6i58CJ5.js";
+import { d as mapConcurrent } from "./journal-CTX5uK-n.js";
 import { n as contentHash, t as canonicalJson } from "./verdict-cache-BVH-fl18.js";
-import { $ as resolveRunDir, C as surfaceDiff, G as assertIndependentEvaluationSplit, J as exposeFinalEvidence, K as captureFinalEvidencePolicy, R as captureJudge, T as assertSearchHistoryAdmissionOptions, Y as reserveFinalEvidence, Z as runCampaign, b as SearchRecorder, c as readGepaCandidatePopulationArtifact, d as optimizationTokenUsageFromSummary, l as combineComparisonCosts, o as executeOptimizationMethod, q as evaluationUnitMap, s as assertGepaCandidatePopulationSummary, u as costFromLedgerSummary, w as surfaceNode, x as developmentClaim } from "./llm-judge-CHLvjbVR.js";
-import { D as assertCompleteCampaign, T as assertCampaignDesign, f as campaignBreakdown, t as createCampaignEvidenceReceipt } from "./campaign-evidence-CcA7a67m.js";
+import { $ as resolveRunDir, C as surfaceDiff, G as assertIndependentEvaluationSplit, J as exposeFinalEvidence, K as captureFinalEvidencePolicy, R as captureJudge, T as assertSearchHistoryAdmissionOptions, Y as reserveFinalEvidence, Z as runCampaign, b as SearchRecorder, c as readGepaCandidatePopulationArtifact, d as optimizationTokenUsageFromSummary, l as combineComparisonCosts, o as executeOptimizationMethod, q as evaluationUnitMap, s as assertGepaCandidatePopulationSummary, u as costFromLedgerSummary, w as surfaceNode, x as developmentClaim } from "./llm-judge-Bnxuwjxl.js";
+import { D as assertCompleteCampaign, T as assertCampaignDesign, f as campaignBreakdown, t as createCampaignEvidenceReceipt } from "./campaign-evidence-BuKlRzSA.js";
 import { t as deepFreezeCanonicalJson } from "./deep-freeze-CK8-2W7p.js";
-import { M as fsCampaignStorage, _ as resolveExternalOptimizerCallbackLimits, d as assertExternalOptimizerModelBudget, f as assertJsonValue, g as removeCredentialEnvironment, h as isExternalTextCandidate, i as startExternalOptimizerModelProxy, j as createRunCostLedger, m as isCandidateText, n as closeExternalOptimizerResources, p as assertNoCredentialValues, r as runWithCleanup, t as runExternalOptimizerProcess, v as resolveExternalOptimizerProcessLimits, y as safePathComponent } from "./external-optimizer-subprocess-B44kOfZb.js";
+import { M as fsCampaignStorage, _ as resolveExternalOptimizerCallbackLimits, d as assertExternalOptimizerModelBudget, f as assertJsonValue, g as removeCredentialEnvironment, h as isExternalTextCandidate, i as startExternalOptimizerModelProxy, j as createRunCostLedger, m as isCandidateText, n as closeExternalOptimizerResources, p as assertNoCredentialValues, r as runWithCleanup, t as runExternalOptimizerProcess, v as resolveExternalOptimizerProcessLimits, y as safePathComponent } from "./external-optimizer-subprocess-_C_2yriN.js";
 import { s as surfaceContentHash } from "./json-recovery-DK4MDm5r.js";
 import { r as summarizeEvaluationUnits, t as defineEvaluationClaim } from "./claim-Dod_mjk3.js";
-import { n as openSearchLedger } from "./search-ledger-CV_bQQed.js";
-import { t as acquireSingleRunLock } from "./single-run-lock-BGlW3OSk.js";
+import { n as openSearchLedger } from "./search-ledger-Dk-yk5Li.js";
+import { t as acquireSingleRunLock } from "./single-run-lock-CUj2tjxT.js";
 import { n as isRecord } from "./json-DtvZNAU9.js";
-import { c as combineAbortSignals, f as assertServedModelPolicy } from "./llm-client-BdSiZ4qz.js";
-import { t as startExternalOptimizerCallback } from "./external-optimizer-process-za-SsIap.js";
+import { c as combineAbortSignals, f as assertServedModelPolicy } from "./llm-client-DMMSg-Ro.js";
+import { t as startExternalOptimizerCallback } from "./external-optimizer-process-2ImGH_LR.js";
 import { createHash, randomBytes } from "node:crypto";
 import { mkdir } from "node:fs/promises";
 import { isDeepStrictEqual } from "node:util";
@@ -3223,4 +3223,4 @@ async function finishSkillOptRun(context, result, outputDir, modelProxy) {
 }
 export { importExternalEvaluations as a, composeGate as c, readExternalOptimizerObservationArtifact as d, externalSurface as i, externalTextOptimizationMethod as l, compareOptimizationMethods as n, importGepaPopulation as o, gepaOptimizationMethod as r, heldOutGate as s, skillOptOptimizationMethod as t, decodeExternalTextCandidate as u };
 
-//# sourceMappingURL=skillopt-optimization-method-D72XZZfR.js.map
+//# sourceMappingURL=skillopt-optimization-method-DYwEndON.js.map