@tangle-network/agent-eval 0.216.1 → 0.217.1

package/dist/{canonical-B2RqLofv.js → canonical-Dq2hOQ8L.js}

@@ -1,4 +1,4 @@
-import { s as ValidationError } from "./errors-DVZiqS7K.js";
+import { s as ValidationError } from "./errors-BcOs5MKW.js";
 import { canonicalCandidateBytes, canonicalCandidateJson, sha256Bytes } from "@tangle-network/agent-interface";
 //#region src/ledger-core/canonical.ts
 /**
@@ -125,4 +125,4 @@ function compareCodeUnits(left, right) {
 //#endregion
 export { hashCanonical as a, compareCodeUnits as i, LedgerCanonicalizationError as n, jsonDocument as o, canonicalString as r, LEDGER_HASH_PATTERN as t };
 
-//# sourceMappingURL=canonical-B2RqLofv.js.map
+//# sourceMappingURL=canonical-Dq2hOQ8L.js.map

package/dist/{canonical-B2RqLofv.js.map → canonical-Dq2hOQ8L.js.map}

@@ -1 +1 @@
-{"version":3,"file":"canonical-B2RqLofv.js","names":[],"sources":["../src/ledger-core/canonical.ts"],"sourcesContent":["/**\n * Canonical bytes and digests for durable journal rows.\n *\n * Serialization is RFC 8785 through `@tangle-network/agent-interface` — the\n * stack's single identity scheme — so a value carries the same digest in every\n * package that produces or verifies one.\n *\n * Values canonical JSON cannot represent faithfully are refused, never coerced.\n * `NaN`, `Infinity`, `undefined`, class instances, cycles, and unpaired\n * surrogates each have a lossy encoding that maps distinct records onto one\n * digest, and in a hash chain the digest IS the record's identity: a coercion\n * there is a collision, not a convenience.\n */\n\nimport {\n  canonicalCandidateBytes,\n  canonicalCandidateJson,\n  sha256Bytes,\n} from '@tangle-network/agent-interface'\nimport { ValidationError } from '../errors'\n\nexport type LedgerHash = `sha256:${string}`\n\n/** Shape of every ledger digest: `sha256:` followed by 64 lowercase hex chars. */\nexport const LEDGER_HASH_PATTERN = /^sha256:[a-f0-9]{64}$/\n\n/** A value has no faithful canonical-JSON form, so it cannot be stored or hashed. */\nexport class LedgerCanonicalizationError extends ValidationError {}\n\n/** Canonical JSON encoding (RFC 8785): object keys sorted by UTF-16 code unit,\n * recursively. This is the byte form that is hashed and stored, so equal values\n * always encode identically. */\nexport function canonicalString(value: unknown): string {\n  try {\n    return canonicalCandidateJson(value)\n  } catch (error) {\n    throw canonicalizationError(value, error)\n  }\n}\n\nexport function hashCanonical(value: unknown): LedgerHash {\n  try {\n    return sha256Bytes(canonicalCandidateBytes(value))\n  } catch (error) {\n    throw canonicalizationError(value, error)\n  }\n}\n\n/**\n * The JSON document form of `value` — the value as a JSON file carries it.\n *\n * A JSON document has no `undefined`, so an undefined-valued key is absent\n * from the document and absent from its digest. Use this, and say so, when a\n * record is digested BEFORE being written and re-digested AFTER being read\n * back: without it the two digests disagree for a record carrying an\n * undefined-valued key, and the reader cannot verify the writer.\n *\n * Only that one rule is applied. Every other non-canonical value — `NaN`, a\n * class instance, an `undefined` array item, a cycle — reaches\n * {@link canonicalString} unchanged and is refused there, so this form never\n * turns an ambiguous value into a valid-looking digest the way\n * `JSON.stringify` turns `NaN` into `null`.\n */\nexport function jsonDocument(value: unknown): unknown {\n  if (value === null || typeof value !== 'object') return value\n  if (Array.isArray(value)) return value.map(jsonDocument)\n  const prototype = Object.getPrototypeOf(value)\n  if (prototype !== Object.prototype && prototype !== null) return value\n  // fromEntries creates own properties, including a caller's __proto__ key.\n  return Object.fromEntries(\n    Object.entries(value as Record<string, unknown>)\n      .filter(([, entry]) => entry !== undefined)\n      .map(([key, entry]) => [key, jsonDocument(entry)]),\n  )\n}\n\n/** Name the offending value and path so a refusal is actionable. The RFC 8785\n * encoder above is the sole authority on acceptance; this walk only explains a\n * rejection it already made. */\nfunction canonicalizationError(value: unknown, cause: unknown): LedgerCanonicalizationError {\n  const defect = firstNonCanonicalDefect(value, '$', new Set())\n  const reason = defect !== '' ? defect : cause instanceof Error ? cause.message : String(cause)\n  return new LedgerCanonicalizationError(`value has no canonical JSON form: ${reason}`, {\n    cause,\n  })\n}\n\nfunction firstNonCanonicalDefect(value: unknown, path: string, ancestors: Set<object>): string {\n  if (value === null) return ''\n  if (value === undefined) return `${path} is undefined`\n  const type = typeof value\n  if (type === 'number') {\n    return Number.isFinite(value) ? '' : `${path} is ${String(value)}`\n  }\n  if (type === 'string') {\n    return isWellFormedUtf16(value as string) ? '' : `${path} contains an unpaired surrogate`\n  }\n  if (type === 'boolean') return ''\n  if (type !== 'object') return `${path} is a ${type}`\n  const object = value as object\n  if (ancestors.has(object)) return `${path} closes a reference cycle`\n  const prototype = Object.getPrototypeOf(object)\n  if (!Array.isArray(object) && prototype !== Object.prototype && prototype !== null) {\n    return `${path} is a ${object.constructor?.name ?? 'non-plain'} instance, not plain JSON data`\n  }\n  const nested = new Set(ancestors).add(object)\n  if (Array.isArray(object)) {\n    for (let index = 0; index < object.length; index += 1) {\n      const defect = firstNonCanonicalDefect(object[index], `${path}[${index}]`, nested)\n      if (defect !== '') return defect\n    }\n    return ''\n  }\n  for (const [key, entry] of Object.entries(object)) {\n    if (!isWellFormedUtf16(key)) return `${path}.${key} has a key with an unpaired surrogate`\n    const defect = firstNonCanonicalDefect(entry, `${path}.${key}`, nested)\n    if (defect !== '') return defect\n  }\n  return ''\n}\n\nfunction isWellFormedUtf16(value: string): boolean {\n  for (let index = 0; index < value.length; index += 1) {\n    const code = value.charCodeAt(index)\n    if (code >= 0xd800 && code <= 0xdbff) {\n      const next = value.charCodeAt(index + 1)\n      if (!(next >= 0xdc00 && next <= 0xdfff)) return false\n      index += 1\n    } else if (code >= 0xdc00 && code <= 0xdfff) {\n      return false\n    }\n  }\n  return true\n}\n\n/**\n * Compare two strings by UTF-16 code unit — the order RFC 8785 uses for object\n * keys, and the order any array must be sorted into before it is canonicalized.\n *\n * `String.prototype.localeCompare` reads the host's collation, so it is a\n * property of the machine rather than of the value. RFC 8785 canonicalizes an\n * array BY POSITION, so a `localeCompare` sort in front of `canonicalString` or\n * `hashCanonical` lets the collation decide the digest bytes: the ids\n * `Accuracy, brevity, Clarity` order as `Accuracy,brevity,Clarity` under an\n * en-US collation and as `Accuracy,Clarity,brevity` by code unit, and the two\n * produce different digests for the same data.\n *\n * Use this for any ordering whose result reaches a digest, a stable\n * serialization, or a stored identity. A comparator for a report a human reads\n * is free to keep `localeCompare`.\n */\nexport function compareCodeUnits(left: string, right: string): number {\n  if (left < right) return -1\n  return left > right ? 1 : 0\n}\n"],"mappings":";;;;;;;;;;;;;;;;;AAwBA,MAAa,sBAAsB;;AAGnC,IAAa,8BAAb,cAAiD,gBAAgB,CAAC;;;;AAKlE,SAAgB,gBAAgB,OAAwB;CACtD,IAAI;EACF,OAAO,uBAAuB,KAAK;CACrC,SAAS,OAAO;EACd,MAAM,sBAAsB,OAAO,KAAK;CAC1C;AACF;AAEA,SAAgB,cAAc,OAA4B;CACxD,IAAI;EACF,OAAO,YAAY,wBAAwB,KAAK,CAAC;CACnD,SAAS,OAAO;EACd,MAAM,sBAAsB,OAAO,KAAK;CAC1C;AACF;;;;;;;;;;;;;;;;AAiBA,SAAgB,aAAa,OAAyB;CACpD,IAAI,UAAU,QAAQ,OAAO,UAAU,UAAU,OAAO;CACxD,IAAI,MAAM,QAAQ,KAAK,GAAG,OAAO,MAAM,IAAI,YAAY;CACvD,MAAM,YAAY,OAAO,eAAe,KAAK;CAC7C,IAAI,cAAc,OAAO,aAAa,cAAc,MAAM,OAAO;CAEjE,OAAO,OAAO,YACZ,OAAO,QAAQ,KAAgC,CAAC,CAC7C,QAAQ,GAAG,WAAW,UAAU,KAAA,CAAS,CAAC,CAC1C,KAAK,CAAC,KAAK,WAAW,CAAC,KAAK,aAAa,KAAK,CAAC,CAAC,CACrD;AACF;;;;AAKA,SAAS,sBAAsB,OAAgB,OAA6C;CAC1F,MAAM,SAAS,wBAAwB,OAAO,qBAAK,IAAI,IAAI,CAAC;CAE5D,OAAO,IAAI,4BAA4B,qCADxB,WAAW,KAAK,SAAS,iBAAiB,QAAQ,MAAM,UAAU,OAAO,KAAK,KACP,EACpF,MACF,CAAC;AACH;AAEA,SAAS,wBAAwB,OAAgB,MAAc,WAAgC;CAC7F,IAAI,UAAU,MAAM,OAAO;CAC3B,IAAI,UAAU,KAAA,GAAW,OAAO,GAAG,KAAK;CACxC,MAAM,OAAO,OAAO;CACpB,IAAI,SAAS,UACX,OAAO,OAAO,SAAS,KAAK,IAAI,KAAK,GAAG,KAAK,MAAM,OAAO,KAAK;CAEjE,IAAI,SAAS,UACX,OAAO,kBAAkB,KAAe,IAAI,KAAK,GAAG,KAAK;CAE3D,IAAI,SAAS,WAAW,OAAO;CAC/B,IAAI,SAAS,UAAU,OAAO,GAAG,KAAK,QAAQ;CAC9C,MAAM,SAAS;CACf,IAAI,UAAU,IAAI,MAAM,GAAG,OAAO,GAAG,KAAK;CAC1C,MAAM,YAAY,OAAO,eAAe,MAAM;CAC9C,IAAI,CAAC,MAAM,QAAQ,MAAM,KAAK,cAAc,OAAO,aAAa,cAAc,MAC5E,OAAO,GAAG,KAAK,QAAQ,OAAO,aAAa,QAAQ,YAAY;CAEjE,MAAM,SAAS,IAAI,IAAI,SAAS,CAAC,CAAC,IAAI,MAAM;CAC5C,IAAI,MAAM,QAAQ,MAAM,GAAG;EACzB,KAAK,IAAI,QAAQ,GAAG,QAAQ,OAAO,QAAQ,SAAS,GAAG;GACrD,MAAM,SAAS,wBAAwB,OAAO,QAAQ,GAAG,KAAK,GAAG,MAAM,IAAI,MAAM;GACjF,IAAI,WAAW,IAAI,OAAO;EAC5B;EACA,OAAO;CACT;CACA,KAAK,MAAM,CAAC,KAAK,UAAU,OAAO,QAAQ,MAAM,GAAG;EACjD,IAAI,CAAC,kBAAkB,GAAG,GAAG,OAAO,GAAG,KAAK,GAAG,IAAI;EACnD,MAAM,SAAS,wBAAwB,OAAO,GAAG,KAAK,GAAG,OAAO,MAAM;EACtE,IAAI,WAAW,IAAI,OAAO;CAC5B;CACA,OAAO;AACT;AAEA,SAAS,kBAAkB,OAAwB;CACjD,KAAK,IAAI,QAAQ,GAAG,QAAQ,MAAM,QAAQ,SAAS,GAAG;EACpD,MAAM,OAAO,MAAM,WAAW,KAAK;EACnC,IAAI,QAAQ,SAAU,QAAQ,OAAQ;GACpC,MAAM,OAAO,MAAM,WAAW,QAAQ,CAAC;GACvC,IAAI,EAAE,QAAQ,SAAU,QAAQ,QAAS,OAAO;GAChD,SAAS;EACX,OAAO,IAAI,QAAQ,SAAU,QAAQ,OACnC,OAAO;CAEX;CACA,OAAO;AACT;;;;;;;;;;;;;;;;;AAkBA,SAAgB,iBAAiB,MAAc,OAAuB;CACpE,IAAI,OAAO,OAAO,OAAO;CACzB,OAAO,OAAO,QAAQ,IAAI;AAC5B"}
+{"version":3,"file":"canonical-Dq2hOQ8L.js","names":[],"sources":["../src/ledger-core/canonical.ts"],"sourcesContent":["/**\n * Canonical bytes and digests for durable journal rows.\n *\n * Serialization is RFC 8785 through `@tangle-network/agent-interface` — the\n * stack's single identity scheme — so a value carries the same digest in every\n * package that produces or verifies one.\n *\n * Values canonical JSON cannot represent faithfully are refused, never coerced.\n * `NaN`, `Infinity`, `undefined`, class instances, cycles, and unpaired\n * surrogates each have a lossy encoding that maps distinct records onto one\n * digest, and in a hash chain the digest IS the record's identity: a coercion\n * there is a collision, not a convenience.\n */\n\nimport {\n  canonicalCandidateBytes,\n  canonicalCandidateJson,\n  sha256Bytes,\n} from '@tangle-network/agent-interface'\nimport { ValidationError } from '../errors'\n\nexport type LedgerHash = `sha256:${string}`\n\n/** Shape of every ledger digest: `sha256:` followed by 64 lowercase hex chars. */\nexport const LEDGER_HASH_PATTERN = /^sha256:[a-f0-9]{64}$/\n\n/** A value has no faithful canonical-JSON form, so it cannot be stored or hashed. */\nexport class LedgerCanonicalizationError extends ValidationError {}\n\n/** Canonical JSON encoding (RFC 8785): object keys sorted by UTF-16 code unit,\n * recursively. This is the byte form that is hashed and stored, so equal values\n * always encode identically. */\nexport function canonicalString(value: unknown): string {\n  try {\n    return canonicalCandidateJson(value)\n  } catch (error) {\n    throw canonicalizationError(value, error)\n  }\n}\n\nexport function hashCanonical(value: unknown): LedgerHash {\n  try {\n    return sha256Bytes(canonicalCandidateBytes(value))\n  } catch (error) {\n    throw canonicalizationError(value, error)\n  }\n}\n\n/**\n * The JSON document form of `value` — the value as a JSON file carries it.\n *\n * A JSON document has no `undefined`, so an undefined-valued key is absent\n * from the document and absent from its digest. Use this, and say so, when a\n * record is digested BEFORE being written and re-digested AFTER being read\n * back: without it the two digests disagree for a record carrying an\n * undefined-valued key, and the reader cannot verify the writer.\n *\n * Only that one rule is applied. Every other non-canonical value — `NaN`, a\n * class instance, an `undefined` array item, a cycle — reaches\n * {@link canonicalString} unchanged and is refused there, so this form never\n * turns an ambiguous value into a valid-looking digest the way\n * `JSON.stringify` turns `NaN` into `null`.\n */\nexport function jsonDocument(value: unknown): unknown {\n  if (value === null || typeof value !== 'object') return value\n  if (Array.isArray(value)) return value.map(jsonDocument)\n  const prototype = Object.getPrototypeOf(value)\n  if (prototype !== Object.prototype && prototype !== null) return value\n  // fromEntries creates own properties, including a caller's __proto__ key.\n  return Object.fromEntries(\n    Object.entries(value as Record<string, unknown>)\n      .filter(([, entry]) => entry !== undefined)\n      .map(([key, entry]) => [key, jsonDocument(entry)]),\n  )\n}\n\n/** Name the offending value and path so a refusal is actionable. The RFC 8785\n * encoder above is the sole authority on acceptance; this walk only explains a\n * rejection it already made. */\nfunction canonicalizationError(value: unknown, cause: unknown): LedgerCanonicalizationError {\n  const defect = firstNonCanonicalDefect(value, '$', new Set())\n  const reason = defect !== '' ? defect : cause instanceof Error ? cause.message : String(cause)\n  return new LedgerCanonicalizationError(`value has no canonical JSON form: ${reason}`, {\n    cause,\n  })\n}\n\nfunction firstNonCanonicalDefect(value: unknown, path: string, ancestors: Set<object>): string {\n  if (value === null) return ''\n  if (value === undefined) return `${path} is undefined`\n  const type = typeof value\n  if (type === 'number') {\n    return Number.isFinite(value) ? '' : `${path} is ${String(value)}`\n  }\n  if (type === 'string') {\n    return isWellFormedUtf16(value as string) ? '' : `${path} contains an unpaired surrogate`\n  }\n  if (type === 'boolean') return ''\n  if (type !== 'object') return `${path} is a ${type}`\n  const object = value as object\n  if (ancestors.has(object)) return `${path} closes a reference cycle`\n  const prototype = Object.getPrototypeOf(object)\n  if (!Array.isArray(object) && prototype !== Object.prototype && prototype !== null) {\n    return `${path} is a ${object.constructor?.name ?? 'non-plain'} instance, not plain JSON data`\n  }\n  const nested = new Set(ancestors).add(object)\n  if (Array.isArray(object)) {\n    for (let index = 0; index < object.length; index += 1) {\n      const defect = firstNonCanonicalDefect(object[index], `${path}[${index}]`, nested)\n      if (defect !== '') return defect\n    }\n    return ''\n  }\n  for (const [key, entry] of Object.entries(object)) {\n    if (!isWellFormedUtf16(key)) return `${path}.${key} has a key with an unpaired surrogate`\n    const defect = firstNonCanonicalDefect(entry, `${path}.${key}`, nested)\n    if (defect !== '') return defect\n  }\n  return ''\n}\n\nfunction isWellFormedUtf16(value: string): boolean {\n  for (let index = 0; index < value.length; index += 1) {\n    const code = value.charCodeAt(index)\n    if (code >= 0xd800 && code <= 0xdbff) {\n      const next = value.charCodeAt(index + 1)\n      if (!(next >= 0xdc00 && next <= 0xdfff)) return false\n      index += 1\n    } else if (code >= 0xdc00 && code <= 0xdfff) {\n      return false\n    }\n  }\n  return true\n}\n\n/**\n * Compare two strings by UTF-16 code unit — the order RFC 8785 uses for object\n * keys, and the order any array must be sorted into before it is canonicalized.\n *\n * `String.prototype.localeCompare` reads the host's collation, so it is a\n * property of the machine rather than of the value. RFC 8785 canonicalizes an\n * array BY POSITION, so a `localeCompare` sort in front of `canonicalString` or\n * `hashCanonical` lets the collation decide the digest bytes: the ids\n * `Accuracy, brevity, Clarity` order as `Accuracy,brevity,Clarity` under an\n * en-US collation and as `Accuracy,Clarity,brevity` by code unit, and the two\n * produce different digests for the same data.\n *\n * Use this for any ordering whose result reaches a digest, a stable\n * serialization, or a stored identity. A comparator for a report a human reads\n * is free to keep `localeCompare`.\n */\nexport function compareCodeUnits(left: string, right: string): number {\n  if (left < right) return -1\n  return left > right ? 1 : 0\n}\n"],"mappings":";;;;;;;;;;;;;;;;;AAwBA,MAAa,sBAAsB;;AAGnC,IAAa,8BAAb,cAAiD,gBAAgB,CAAC;;;;AAKlE,SAAgB,gBAAgB,OAAwB;CACtD,IAAI;EACF,OAAO,uBAAuB,KAAK;CACrC,SAAS,OAAO;EACd,MAAM,sBAAsB,OAAO,KAAK;CAC1C;AACF;AAEA,SAAgB,cAAc,OAA4B;CACxD,IAAI;EACF,OAAO,YAAY,wBAAwB,KAAK,CAAC;CACnD,SAAS,OAAO;EACd,MAAM,sBAAsB,OAAO,KAAK;CAC1C;AACF;;;;;;;;;;;;;;;;AAiBA,SAAgB,aAAa,OAAyB;CACpD,IAAI,UAAU,QAAQ,OAAO,UAAU,UAAU,OAAO;CACxD,IAAI,MAAM,QAAQ,KAAK,GAAG,OAAO,MAAM,IAAI,YAAY;CACvD,MAAM,YAAY,OAAO,eAAe,KAAK;CAC7C,IAAI,cAAc,OAAO,aAAa,cAAc,MAAM,OAAO;CAEjE,OAAO,OAAO,YACZ,OAAO,QAAQ,KAAgC,CAAC,CAC7C,QAAQ,GAAG,WAAW,UAAU,KAAA,CAAS,CAAC,CAC1C,KAAK,CAAC,KAAK,WAAW,CAAC,KAAK,aAAa,KAAK,CAAC,CAAC,CACrD;AACF;;;;AAKA,SAAS,sBAAsB,OAAgB,OAA6C;CAC1F,MAAM,SAAS,wBAAwB,OAAO,qBAAK,IAAI,IAAI,CAAC;CAE5D,OAAO,IAAI,4BAA4B,qCADxB,WAAW,KAAK,SAAS,iBAAiB,QAAQ,MAAM,UAAU,OAAO,KAAK,KACP,EACpF,MACF,CAAC;AACH;AAEA,SAAS,wBAAwB,OAAgB,MAAc,WAAgC;CAC7F,IAAI,UAAU,MAAM,OAAO;CAC3B,IAAI,UAAU,KAAA,GAAW,OAAO,GAAG,KAAK;CACxC,MAAM,OAAO,OAAO;CACpB,IAAI,SAAS,UACX,OAAO,OAAO,SAAS,KAAK,IAAI,KAAK,GAAG,KAAK,MAAM,OAAO,KAAK;CAEjE,IAAI,SAAS,UACX,OAAO,kBAAkB,KAAe,IAAI,KAAK,GAAG,KAAK;CAE3D,IAAI,SAAS,WAAW,OAAO;CAC/B,IAAI,SAAS,UAAU,OAAO,GAAG,KAAK,QAAQ;CAC9C,MAAM,SAAS;CACf,IAAI,UAAU,IAAI,MAAM,GAAG,OAAO,GAAG,KAAK;CAC1C,MAAM,YAAY,OAAO,eAAe,MAAM;CAC9C,IAAI,CAAC,MAAM,QAAQ,MAAM,KAAK,cAAc,OAAO,aAAa,cAAc,MAC5E,OAAO,GAAG,KAAK,QAAQ,OAAO,aAAa,QAAQ,YAAY;CAEjE,MAAM,SAAS,IAAI,IAAI,SAAS,CAAC,CAAC,IAAI,MAAM;CAC5C,IAAI,MAAM,QAAQ,MAAM,GAAG;EACzB,KAAK,IAAI,QAAQ,GAAG,QAAQ,OAAO,QAAQ,SAAS,GAAG;GACrD,MAAM,SAAS,wBAAwB,OAAO,QAAQ,GAAG,KAAK,GAAG,MAAM,IAAI,MAAM;GACjF,IAAI,WAAW,IAAI,OAAO;EAC5B;EACA,OAAO;CACT;CACA,KAAK,MAAM,CAAC,KAAK,UAAU,OAAO,QAAQ,MAAM,GAAG;EACjD,IAAI,CAAC,kBAAkB,GAAG,GAAG,OAAO,GAAG,KAAK,GAAG,IAAI;EACnD,MAAM,SAAS,wBAAwB,OAAO,GAAG,KAAK,GAAG,OAAO,MAAM;EACtE,IAAI,WAAW,IAAI,OAAO;CAC5B;CACA,OAAO;AACT;AAEA,SAAS,kBAAkB,OAAwB;CACjD,KAAK,IAAI,QAAQ,GAAG,QAAQ,MAAM,QAAQ,SAAS,GAAG;EACpD,MAAM,OAAO,MAAM,WAAW,KAAK;EACnC,IAAI,QAAQ,SAAU,QAAQ,OAAQ;GACpC,MAAM,OAAO,MAAM,WAAW,QAAQ,CAAC;GACvC,IAAI,EAAE,QAAQ,SAAU,QAAQ,QAAS,OAAO;GAChD,SAAS;EACX,OAAO,IAAI,QAAQ,SAAU,QAAQ,OACnC,OAAO;CAEX;CACA,OAAO;AACT;;;;;;;;;;;;;;;;;AAkBA,SAAgB,iBAAiB,MAAc,OAAuB;CACpE,IAAI,OAAO,OAAO,OAAO;CACzB,OAAO,OAAO,QAAQ,IAAI;AAC5B"}

package/dist/cell-spend-DEWR3nj-.js

@@ -0,0 +1,90 @@
+import { t as isNonnegativeFiniteNumber } from "./nonnegative-number-DvdMxg_f.js";
+//#region src/matrix/cell-spend.ts
+/**
+* Spend carrier for a cell that fails after it has already spent money.
+*
+* `runCell` reports what a cell cost by returning a `CellResult`. A throw
+* carries no result, so the runner cannot know what the cell spent before it
+* failed: the cumulative sum the cost ceiling reads, and `totalCostUsd`, both
+* miss that money. A throw that carries a `CellSpend` closes the gap — the
+* runner bills the failed cell for the spend the carrier declares, and records
+* a cell that declared nothing as `uncaptured` instead of a zero that reads
+* like "this cell spent nothing".
+*
+* The carrier is read structurally, never by `instanceof`, so a throw that
+* crosses a package or realm boundary still bills.
+*/
+/** Property key the carrier occupies on a thrown object. Stable across
+*  versions — two copies of this package must agree on it. */
+const CELL_SPEND_KEY = "__agentEvalCellSpend";
+/**
+* Return `error` carrying `spend`, for a throw site inside `runCell`:
+* `throw withCellSpend(err, { costUsd, durationMs, kind })`.
+*
+* A value that cannot hold the carrier — a thrown string or number, or a
+* frozen, sealed, or otherwise non-extensible object — is wrapped in an
+* `Error` that keeps the original as `cause`, so the spend is never dropped.
+* A later call overwrites an earlier carrier — the outer frame knows more
+* than the inner one.
+*
+* Throws `TypeError` on a non-finite or negative amount. A poisoned amount
+* would reach `cumulativeCost` and disable the cost ceiling for the whole run,
+* because `NaN >= ceiling` is false.
+*/
+function withCellSpend(error, spend) {
+	assertAmount(spend.costUsd, "costUsd");
+	assertAmount(spend.durationMs, "durationMs");
+	if (spend.kind !== "observed" && spend.kind !== "estimated" && spend.kind !== "uncaptured") throw new TypeError(`withCellSpend: kind must be observed, estimated or uncaptured, received ${String(spend.kind)}`);
+	const carrier = {
+		costUsd: spend.costUsd,
+		durationMs: spend.durationMs,
+		kind: spend.kind
+	};
+	if (typeof error === "object" && error !== null && attach(error, carrier)) return error;
+	const wrapper = new Error(`cell failed: ${messageOf(error)}`, { cause: error });
+	attach(wrapper, carrier);
+	return wrapper;
+}
+/** True when the carrier landed. `defineProperty` throws on a non-extensible
+*  target, so the caller must be able to fall back. */
+function attach(target, carrier) {
+	try {
+		Object.defineProperty(target, CELL_SPEND_KEY, {
+			value: carrier,
+			enumerable: false,
+			configurable: true,
+			writable: true
+		});
+		return true;
+	} catch {
+		return false;
+	}
+}
+function messageOf(error) {
+	if (error instanceof Error) return error.message;
+	return String(error);
+}
+/** Read the spend a thrown value carries. `undefined` when the value carries
+*  none, or carries one that does not satisfy `CellSpend` — the runner records
+*  both as `uncaptured`. */
+function readCellSpend(error) {
+	if (typeof error !== "object" || error === null) return void 0;
+	const raw = error[CELL_SPEND_KEY];
+	if (typeof raw !== "object" || raw === null) return void 0;
+	const candidate = raw;
+	if (!isNonnegativeFiniteNumber(candidate.costUsd) || !isNonnegativeFiniteNumber(candidate.durationMs)) return void 0;
+	const kind = candidate.kind;
+	if (kind !== "observed" && kind !== "estimated" && kind !== "uncaptured") return void 0;
+	return {
+		costUsd: candidate.costUsd,
+		durationMs: candidate.durationMs,
+		kind
+	};
+}
+function assertAmount(value, field) {
+	if (!isNonnegativeFiniteNumber(value)) throw new TypeError(`withCellSpend: ${field} must be a finite number >= 0, received ${String(value)}`);
+}
+//#endregion
+export { withCellSpend as n, readCellSpend as t };
+
+//# sourceMappingURL=cell-spend-DEWR3nj-.js.map

package/dist/cell-spend-DEWR3nj-.js.map

@@ -0,0 +1 @@
+{"version":3,"file":"cell-spend-DEWR3nj-.js","names":["isAmount"],"sources":["../src/matrix/cell-spend.ts"],"sourcesContent":["/**\n * Spend carrier for a cell that fails after it has already spent money.\n *\n * `runCell` reports what a cell cost by returning a `CellResult`. A throw\n * carries no result, so the runner cannot know what the cell spent before it\n * failed: the cumulative sum the cost ceiling reads, and `totalCostUsd`, both\n * miss that money. A throw that carries a `CellSpend` closes the gap — the\n * runner bills the failed cell for the spend the carrier declares, and records\n * a cell that declared nothing as `uncaptured` instead of a zero that reads\n * like \"this cell spent nothing\".\n *\n * The carrier is read structurally, never by `instanceof`, so a throw that\n * crosses a package or realm boundary still bills.\n */\n\nimport { isNonnegativeFiniteNumber as isAmount } from '../math/nonnegative-number'\n\n/** Property key the carrier occupies on a thrown object. Stable across\n *  versions — two copies of this package must agree on it. */\nconst CELL_SPEND_KEY = '__agentEvalCellSpend'\n\n/** Money and wall time a cell consumed before it threw. */\nexport interface CellSpend {\n  /** KNOWN subtotal in USD. When `kind` is `uncaptured` the cell spent this\n   *  much AND an unknown amount more, so the value must not be read as the\n   *  cell's total. */\n  costUsd: number\n  /** Wall time the cell consumed before the throw, in milliseconds. */\n  durationMs: number\n  /** `observed` = provider-reported amounts, `estimated` = computed from token\n   *  prices, `uncaptured` = `costUsd` is a subtotal and part of the spend\n   *  could not be measured. */\n  kind: 'observed' | 'estimated' | 'uncaptured'\n}\n\n/**\n * Return `error` carrying `spend`, for a throw site inside `runCell`:\n * `throw withCellSpend(err, { costUsd, durationMs, kind })`.\n *\n * A value that cannot hold the carrier — a thrown string or number, or a\n * frozen, sealed, or otherwise non-extensible object — is wrapped in an\n * `Error` that keeps the original as `cause`, so the spend is never dropped.\n * A later call overwrites an earlier carrier — the outer frame knows more\n * than the inner one.\n *\n * Throws `TypeError` on a non-finite or negative amount. A poisoned amount\n * would reach `cumulativeCost` and disable the cost ceiling for the whole run,\n * because `NaN >= ceiling` is false.\n */\nexport function withCellSpend(error: unknown, spend: CellSpend): unknown {\n  assertAmount(spend.costUsd, 'costUsd')\n  assertAmount(spend.durationMs, 'durationMs')\n  if (spend.kind !== 'observed' && spend.kind !== 'estimated' && spend.kind !== 'uncaptured') {\n    throw new TypeError(\n      `withCellSpend: kind must be observed, estimated or uncaptured, received ${String(spend.kind)}`,\n    )\n  }\n  const carrier: CellSpend = {\n    costUsd: spend.costUsd,\n    durationMs: spend.durationMs,\n    kind: spend.kind,\n  }\n  if (typeof error === 'object' && error !== null && attach(error, carrier)) return error\n  // A primitive cannot hold a property, and a frozen or sealed object refuses\n  // one. Carry the spend on a fresh Error that keeps the original as `cause`\n  // rather than lose it — a dropped carrier bills the cell as uncaptured,\n  // which is the under-billing this module exists to prevent.\n  const wrapper = new Error(`cell failed: ${messageOf(error)}`, { cause: error })\n  attach(wrapper, carrier)\n  return wrapper\n}\n\n/** True when the carrier landed. `defineProperty` throws on a non-extensible\n *  target, so the caller must be able to fall back. */\nfunction attach(target: object, carrier: CellSpend): boolean {\n  try {\n    Object.defineProperty(target, CELL_SPEND_KEY, {\n      value: carrier,\n      enumerable: false,\n      configurable: true,\n      writable: true,\n    })\n    return true\n  } catch {\n    return false\n  }\n}\n\nfunction messageOf(error: unknown): string {\n  if (error instanceof Error) return error.message\n  return String(error)\n}\n\n/** Read the spend a thrown value carries. `undefined` when the value carries\n *  none, or carries one that does not satisfy `CellSpend` — the runner records\n *  both as `uncaptured`. */\nexport function readCellSpend(error: unknown): CellSpend | undefined {\n  if (typeof error !== 'object' || error === null) return undefined\n  const raw = (error as Record<string, unknown>)[CELL_SPEND_KEY]\n  if (typeof raw !== 'object' || raw === null) return undefined\n  const candidate = raw as Record<string, unknown>\n  if (!isAmount(candidate.costUsd) || !isAmount(candidate.durationMs)) return undefined\n  const kind = candidate.kind\n  if (kind !== 'observed' && kind !== 'estimated' && kind !== 'uncaptured') return undefined\n  return { costUsd: candidate.costUsd, durationMs: candidate.durationMs, kind }\n}\n\n/** A number the matrix can add to a running cost or duration: finite and not\n *  negative. Anything else would poison a sum the cost ceiling reads. */\nexport { isNonnegativeFiniteNumber as isAmount } from '../math/nonnegative-number'\n\nfunction assertAmount(value: number, field: string): void {\n  if (!isAmount(value)) {\n    throw new TypeError(\n      `withCellSpend: ${field} must be a finite number >= 0, received ${String(value)}`,\n    )\n  }\n}\n"],"mappings":";;;;;;;;;;;;;;;;;;AAmBA,MAAM,iBAAiB;;;;;;;;;;;;;;;AA8BvB,SAAgB,cAAc,OAAgB,OAA2B;CACvE,aAAa,MAAM,SAAS,SAAS;CACrC,aAAa,MAAM,YAAY,YAAY;CAC3C,IAAI,MAAM,SAAS,cAAc,MAAM,SAAS,eAAe,MAAM,SAAS,cAC5E,MAAM,IAAI,UACR,2EAA2E,OAAO,MAAM,IAAI,GAC9F;CAEF,MAAM,UAAqB;EACzB,SAAS,MAAM;EACf,YAAY,MAAM;EAClB,MAAM,MAAM;CACd;CACA,IAAI,OAAO,UAAU,YAAY,UAAU,QAAQ,OAAO,OAAO,OAAO,GAAG,OAAO;CAKlF,MAAM,UAAU,IAAI,MAAM,gBAAgB,UAAU,KAAK,KAAK,EAAE,OAAO,MAAM,CAAC;CAC9E,OAAO,SAAS,OAAO;CACvB,OAAO;AACT;;;AAIA,SAAS,OAAO,QAAgB,SAA6B;CAC3D,IAAI;EACF,OAAO,eAAe,QAAQ,gBAAgB;GAC5C,OAAO;GACP,YAAY;GACZ,cAAc;GACd,UAAU;EACZ,CAAC;EACD,OAAO;CACT,QAAQ;EACN,OAAO;CACT;AACF;AAEA,SAAS,UAAU,OAAwB;CACzC,IAAI,iBAAiB,OAAO,OAAO,MAAM;CACzC,OAAO,OAAO,KAAK;AACrB;;;;AAKA,SAAgB,cAAc,OAAuC;CACnE,IAAI,OAAO,UAAU,YAAY,UAAU,MAAM,OAAO,KAAA;CACxD,MAAM,MAAO,MAAkC;CAC/C,IAAI,OAAO,QAAQ,YAAY,QAAQ,MAAM,OAAO,KAAA;CACpD,MAAM,YAAY;CAClB,IAAI,CAACA,0BAAS,UAAU,OAAO,KAAK,CAACA,0BAAS,UAAU,UAAU,GAAG,OAAO,KAAA;CAC5E,MAAM,OAAO,UAAU;CACvB,IAAI,SAAS,cAAc,SAAS,eAAe,SAAS,cAAc,OAAO,KAAA;CACjF,OAAO;EAAE,SAAS,UAAU;EAAS,YAAY,UAAU;EAAY;CAAK;AAC9E;AAMA,SAAS,aAAa,OAAe,OAAqB;CACxD,IAAI,CAACA,0BAAS,KAAK,GACjB,MAAM,IAAI,UACR,kBAAkB,MAAM,0CAA0C,OAAO,KAAK,GAChF;AAEJ"}

package/dist/{chat-client-l-MVDlDk.js → chat-client-ldqxYYTN.js}

@@ -1,4 +1,4 @@
-import { t as LlmClient } from "./llm-client-CyBAoYiX.js";
+import { t as LlmClient } from "./llm-client-BshCREn6.js";
 //#region src/analyst/chat-client.ts
 /**
 * Provider-neutral chat contract for every model call made by agent-eval.
@@ -114,4 +114,4 @@ function openAiCompatibleClient(opts) {
 //#endregion
 export { createChatClient as t };
 
-//# sourceMappingURL=chat-client-l-MVDlDk.js.map
+//# sourceMappingURL=chat-client-ldqxYYTN.js.map

package/dist/{chat-client-l-MVDlDk.js.map → chat-client-ldqxYYTN.js.map}

@@ -1 +1 @@
-{"version":3,"file":"chat-client-l-MVDlDk.js","names":[],"sources":["../src/analyst/chat-client.ts"],"sourcesContent":["/**\n * Provider-neutral chat contract for every model call made by agent-eval.\n *\n * The caller owns model execution. Evaluation code receives canonical requests\n * and results without importing a provider SDK, and agent-eval reads no\n * environment variable to find a provider credential: every transport is bound\n * at the call site.\n *\n * Two ways to bind one. `custom` and `sandbox-sdk` take a `chat` function the\n * caller wrote. `openai-compatible` takes an endpoint and a credential the\n * caller passes as values, and agent-eval issues the HTTP request itself.\n *\n * WHY THE THIRD ONE EXISTS. Every consumer that wanted a plain OpenAI-style\n * endpoint had to hand-write the same fetch loop — retry on 429 and 5xx,\n * `json_schema` degrade, fenced-JSON stripping, and, load-bearing here, the\n * `servedModel` echo that `assertServedModel` and `assertCrossFamilyServed`\n * read. A hand-written transport that omits `servedModel` makes both identity\n * checks read `unreported`, so the two-vendor judge rule they enforce silently\n * measures nothing. Measured motive: agent-eval 0.172.1 exported no such\n * transport, so a two-model judge panel in discovery-lab reached\n * `createChatClient({ transport: 'custom', baseUrl, apiKey })` — fields that\n * exist on no transport option — and every call threw\n * `TypeError: opts.chat is not a function`, which surfaced as\n * `aggregateJudgeVerdicts: all 2 judges failed`.\n *\n * The credential still never comes from agent-eval's environment. `baseUrl`\n * and the bearer are required arguments with no default and no fallback, so a\n * misconfigured caller gets a refusal at construction, never a call to an\n * endpoint it did not name.\n */\n\nimport {\n  type LlmCallRequest,\n  type LlmCallResult,\n  LlmClient,\n  type LlmClientOptions,\n} from '../llm-client'\n\n/**\n * Unified chat interface using the package's canonical LLM request and result.\n */\nexport interface ChatClient {\n  /** Display name of the bound transport, included in telemetry. */\n  readonly transport: ChatTransport\n  /** Default model when the caller omits one. */\n  readonly defaultModel?: string\n  /** Total provider attempts this transport can make for one chat call. */\n  readonly maximumAttempts?: number\n\n  /** Implementations must enforce `req.maxTokens` when it is present. */\n  chat(req: ChatRequest, opts?: ChatCallOpts): Promise<ChatResponse>\n}\n\nexport type ChatTransport =\n  | 'sandbox-sdk' // box.streamPrompt() — chat completion via sandbox SDK\n  | 'custom' // caller-adapted SDK or transport\n  | 'openai-compatible' // caller-named /v1 endpoint, driven by this package\n  | 'mock' // test-time injection\n\nexport interface ChatRequest extends Omit<LlmCallRequest, 'model'> {\n  /** Optional — falls back to ChatClient.defaultModel. */\n  model?: string\n}\n\nexport type ChatResponse = LlmCallResult\n\nexport interface ChatCallOpts {\n  /** Cancel the in-flight request. */\n  signal?: AbortSignal\n  /** Hard USD ceiling for this single call (informational; the underlying transport may not enforce). */\n  maxCostUsd?: number\n  /** Correlation tag carried into request headers when the transport allows. */\n  correlationId?: string\n  /** Stable provider idempotency key for retries/redrives of one paid call. */\n  idempotencyKey?: string\n}\n\n// ── Factory ─────────────────────────────────────────────────────────\n\nexport type CreateChatClientOpts =\n  | SandboxSdkTransportOpts\n  | CustomTransportOpts\n  | OpenAiCompatibleTransportOpts\n  | MockTransportOpts\n\ninterface BaseTransportOpts {\n  defaultModel?: string\n  /** Total provider attempts. Required for opaque transports used in capped runs. */\n  maximumAttempts?: number\n}\n\n/**\n * Sandbox-SDK transport. The caller supplies a canonical chat function for an\n * already-configured Sandbox handle, so agent-eval does not import the SDK.\n */\nexport interface SandboxSdkTransportOpts extends BaseTransportOpts {\n  transport: 'sandbox-sdk'\n  chat: (req: ChatRequest, opts?: ChatCallOpts) => Promise<ChatResponse>\n}\n\n/** Caller-adapted SDK or transport returning the canonical ChatResponse shape. */\nexport interface CustomTransportOpts extends BaseTransportOpts {\n  transport: 'custom'\n  chat: (req: ChatRequest, opts?: ChatCallOpts) => Promise<ChatResponse>\n}\n\n/**\n * OpenAI-compatible HTTP transport: the caller names a `/v1` endpoint and\n * hands over a bearer, and this package drives `POST {baseUrl}/chat/completions`.\n *\n * Use this instead of hand-rolling a `custom` transport around `fetch`. The\n * result carries `servedModel` — the model id the provider echoed, verbatim —\n * which is the only field `assertServedModel` and `assertCrossFamilyServed`\n * can read to witness a gateway answering from a different model than the one\n * requested.\n *\n * `baseUrl` ends at the `/v1` prefix; the `/chat/completions` path is this\n * package's to append. A `baseUrl` that already carries the path is a\n * construction error, not a request to a doubled URL.\n *\n * Exactly one credential form is required — `apiKey`, `bearer`, or\n * `authHeader`. There is no environment fallback: agent-eval never goes\n * looking for a key.\n */\nexport interface OpenAiCompatibleTransportOpts\n  extends BaseTransportOpts,\n    Pick<\n      LlmClientOptions,\n      | 'assertServedModel'\n      | 'customTokenPricing'\n      | 'deadlineMs'\n      | 'defaultTimeoutMs'\n      | 'fetch'\n      | 'jsonPayloadMode'\n      | 'jsonSchemaTransport'\n      | 'minIntervalMs'\n      | 'provider'\n      | 'rawSink'\n      | 'signal'\n      | 'thinking'\n    > {\n  transport: 'openai-compatible'\n  /** Endpoint ending at the `/v1` prefix. Required — there is no default endpoint. */\n  baseUrl: string\n  /** Bearer credential. One of `apiKey`, `bearer`, or `authHeader` is required. */\n  apiKey?: string\n  /** Bearer credential, alternate spelling. */\n  bearer?: string\n  /** Non-bearer authorization header, for endpoints that want their own scheme. */\n  authHeader?: { name: string; value: string }\n}\n\n/**\n * Mock transport for tests. The handler receives the request and returns\n * whatever the test wants. No retries, no JSON-schema degrade.\n */\nexport interface MockTransportOpts extends BaseTransportOpts {\n  transport: 'mock'\n  handler: (req: ChatRequest, opts?: ChatCallOpts) => Promise<ChatResponse>\n}\n\n/**\n * Build a ChatClient bound to a specific transport. The returned client\n * is safe to share across analysts in a single registry run.\n */\nexport function createChatClient(opts: CreateChatClientOpts): ChatClient {\n  switch (opts.transport) {\n    case 'sandbox-sdk':\n      return {\n        transport: 'sandbox-sdk',\n        defaultModel: opts.defaultModel,\n        maximumAttempts: opts.maximumAttempts,\n        chat: async (req, callOpts) => opts.chat(resolveModel(req, opts.defaultModel), callOpts),\n      }\n    case 'custom':\n      return {\n        transport: 'custom',\n        defaultModel: opts.defaultModel,\n        maximumAttempts: opts.maximumAttempts,\n        chat: async (req, callOpts) => opts.chat(resolveModel(req, opts.defaultModel), callOpts),\n      }\n    case 'openai-compatible':\n      return openAiCompatibleClient(opts)\n    case 'mock':\n      return {\n        transport: 'mock',\n        defaultModel: opts.defaultModel,\n        maximumAttempts: 1,\n        chat: async (req, callOpts) => opts.handler(resolveModel(req, opts.defaultModel), callOpts),\n      }\n  }\n}\n\nfunction resolveModel(req: ChatRequest, defaultModel: string | undefined): ChatRequest {\n  if (req.model) return req\n  if (!defaultModel) {\n    throw new Error(\n      'ChatClient.chat: no model on request and no defaultModel on the client. ' +\n        'Either pass req.model or bind defaultModel at createChatClient().',\n    )\n  }\n  return { ...req, model: defaultModel }\n}\n\n/**\n * Build the OpenAI-compatible client, refusing a half-configured one first.\n *\n * Both refusals are construction-time on purpose. A missing credential or a\n * doubled path is a caller mistake that a request would turn into a 401 or a\n * 404 from an endpoint the caller did not intend to reach — the failure then\n * reads as a provider problem instead of a configuration one.\n */\nfunction openAiCompatibleClient(opts: OpenAiCompatibleTransportOpts): ChatClient {\n  const baseUrl = opts.baseUrl?.trim().replace(/\\/+$/, '') ?? ''\n  if (!baseUrl) {\n    throw new Error(\n      \"createChatClient({ transport: 'openai-compatible' }): baseUrl is required and has no default.\",\n    )\n  }\n  if (/\\/chat\\/completions$/.test(baseUrl)) {\n    throw new Error(\n      \"createChatClient({ transport: 'openai-compatible' }): baseUrl ends at the /v1 prefix; \" +\n        `this package appends /chat/completions. Got ${baseUrl}`,\n    )\n  }\n  if (!opts.apiKey?.trim() && !opts.bearer?.trim() && !opts.authHeader) {\n    throw new Error(\n      \"createChatClient({ transport: 'openai-compatible' }): one of apiKey, bearer, or authHeader \" +\n        'is required. agent-eval reads no environment variable for a credential.',\n    )\n  }\n\n  const clientOptions: LlmClientOptions = {\n    baseUrl,\n    ...(opts.apiKey ? { apiKey: opts.apiKey } : {}),\n    ...(opts.bearer ? { bearer: opts.bearer } : {}),\n    ...(opts.authHeader ? { authHeader: opts.authHeader } : {}),\n    ...(opts.maximumAttempts === undefined ? {} : { maximumAttempts: opts.maximumAttempts }),\n    ...(opts.assertServedModel === undefined ? {} : { assertServedModel: opts.assertServedModel }),\n    ...(opts.customTokenPricing === undefined\n      ? {}\n      : { customTokenPricing: opts.customTokenPricing }),\n    ...(opts.deadlineMs === undefined ? {} : { deadlineMs: opts.deadlineMs }),\n    ...(opts.defaultTimeoutMs === undefined ? {} : { defaultTimeoutMs: opts.defaultTimeoutMs }),\n    ...(opts.fetch === undefined ? {} : { fetch: opts.fetch }),\n    ...(opts.jsonPayloadMode === undefined ? {} : { jsonPayloadMode: opts.jsonPayloadMode }),\n    ...(opts.jsonSchemaTransport === undefined\n      ? {}\n      : { jsonSchemaTransport: opts.jsonSchemaTransport }),\n    ...(opts.minIntervalMs === undefined ? {} : { minIntervalMs: opts.minIntervalMs }),\n    ...(opts.provider === undefined ? {} : { provider: opts.provider }),\n    ...(opts.rawSink === undefined ? {} : { rawSink: opts.rawSink }),\n    ...(opts.signal === undefined ? {} : { signal: opts.signal }),\n    ...(opts.thinking === undefined ? {} : { thinking: opts.thinking }),\n  }\n  const client = new LlmClient(clientOptions)\n\n  return {\n    transport: 'openai-compatible',\n    ...(opts.defaultModel ? { defaultModel: opts.defaultModel } : {}),\n    maximumAttempts: client.maximumAttempts,\n    chat: async (req, callOpts) => {\n      const resolved = resolveModel(req, opts.defaultModel)\n      return client.call(resolved as LlmCallRequest, {\n        ...(callOpts?.signal ? { signal: callOpts.signal } : {}),\n        ...(callOpts?.idempotencyKey ? { idempotencyKey: callOpts.idempotencyKey } : {}),\n      })\n    },\n  }\n}\n"],"mappings":";;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;AAqKA,SAAgB,iBAAiB,MAAwC;CACvE,QAAQ,KAAK,WAAb;EACE,KAAK,eACH,OAAO;GACL,WAAW;GACX,cAAc,KAAK;GACnB,iBAAiB,KAAK;GACtB,MAAM,OAAO,KAAK,aAAa,KAAK,KAAK,aAAa,KAAK,KAAK,YAAY,GAAG,QAAQ;EACzF;EACF,KAAK,UACH,OAAO;GACL,WAAW;GACX,cAAc,KAAK;GACnB,iBAAiB,KAAK;GACtB,MAAM,OAAO,KAAK,aAAa,KAAK,KAAK,aAAa,KAAK,KAAK,YAAY,GAAG,QAAQ;EACzF;EACF,KAAK,qBACH,OAAO,uBAAuB,IAAI;EACpC,KAAK,QACH,OAAO;GACL,WAAW;GACX,cAAc,KAAK;GACnB,iBAAiB;GACjB,MAAM,OAAO,KAAK,aAAa,KAAK,QAAQ,aAAa,KAAK,KAAK,YAAY,GAAG,QAAQ;EAC5F;CACJ;AACF;AAEA,SAAS,aAAa,KAAkB,cAA+C;CACrF,IAAI,IAAI,OAAO,OAAO;CACtB,IAAI,CAAC,cACH,MAAM,IAAI,MACR,2IAEF;CAEF,OAAO;EAAE,GAAG;EAAK,OAAO;CAAa;AACvC;;;;;;;;;AAUA,SAAS,uBAAuB,MAAiD;CAC/E,MAAM,UAAU,KAAK,SAAS,KAAK,CAAC,CAAC,QAAQ,QAAQ,EAAE,KAAK;CAC5D,IAAI,CAAC,SACH,MAAM,IAAI,MACR,+FACF;CAEF,IAAI,uBAAuB,KAAK,OAAO,GACrC,MAAM,IAAI,MACR,qIACiD,SACnD;CAEF,IAAI,CAAC,KAAK,QAAQ,KAAK,KAAK,CAAC,KAAK,QAAQ,KAAK,KAAK,CAAC,KAAK,YACxD,MAAM,IAAI,MACR,oKAEF;CAGF,MAAM,gBAAkC;EACtC;EACA,GAAI,KAAK,SAAS,EAAE,QAAQ,KAAK,OAAO,IAAI,CAAC;EAC7C,GAAI,KAAK,SAAS,EAAE,QAAQ,KAAK,OAAO,IAAI,CAAC;EAC7C,GAAI,KAAK,aAAa,EAAE,YAAY,KAAK,WAAW,IAAI,CAAC;EACzD,GAAI,KAAK,oBAAoB,KAAA,IAAY,CAAC,IAAI,EAAE,iBAAiB,KAAK,gBAAgB;EACtF,GAAI,KAAK,sBAAsB,KAAA,IAAY,CAAC,IAAI,EAAE,mBAAmB,KAAK,kBAAkB;EAC5F,GAAI,KAAK,uBAAuB,KAAA,IAC5B,CAAC,IACD,EAAE,oBAAoB,KAAK,mBAAmB;EAClD,GAAI,KAAK,eAAe,KAAA,IAAY,CAAC,IAAI,EAAE,YAAY,KAAK,WAAW;EACvE,GAAI,KAAK,qBAAqB,KAAA,IAAY,CAAC,IAAI,EAAE,kBAAkB,KAAK,iBAAiB;EACzF,GAAI,KAAK,UAAU,KAAA,IAAY,CAAC,IAAI,EAAE,OAAO,KAAK,MAAM;EACxD,GAAI,KAAK,oBAAoB,KAAA,IAAY,CAAC,IAAI,EAAE,iBAAiB,KAAK,gBAAgB;EACtF,GAAI,KAAK,wBAAwB,KAAA,IAC7B,CAAC,IACD,EAAE,qBAAqB,KAAK,oBAAoB;EACpD,GAAI,KAAK,kBAAkB,KAAA,IAAY,CAAC,IAAI,EAAE,eAAe,KAAK,cAAc;EAChF,GAAI,KAAK,aAAa,KAAA,IAAY,CAAC,IAAI,EAAE,UAAU,KAAK,SAAS;EACjE,GAAI,KAAK,YAAY,KAAA,IAAY,CAAC,IAAI,EAAE,SAAS,KAAK,QAAQ;EAC9D,GAAI,KAAK,WAAW,KAAA,IAAY,CAAC,IAAI,EAAE,QAAQ,KAAK,OAAO;EAC3D,GAAI,KAAK,aAAa,KAAA,IAAY,CAAC,IAAI,EAAE,UAAU,KAAK,SAAS;CACnE;CACA,MAAM,SAAS,IAAI,UAAU,aAAa;CAE1C,OAAO;EACL,WAAW;EACX,GAAI,KAAK,eAAe,EAAE,cAAc,KAAK,aAAa,IAAI,CAAC;EAC/D,iBAAiB,OAAO;EACxB,MAAM,OAAO,KAAK,aAAa;GAC7B,MAAM,WAAW,aAAa,KAAK,KAAK,YAAY;GACpD,OAAO,OAAO,KAAK,UAA4B;IAC7C,GAAI,UAAU,SAAS,EAAE,QAAQ,SAAS,OAAO,IAAI,CAAC;IACtD,GAAI,UAAU,iBAAiB,EAAE,gBAAgB,SAAS,eAAe,IAAI,CAAC;GAChF,CAAC;EACH;CACF;AACF"}
+{"version":3,"file":"chat-client-ldqxYYTN.js","names":[],"sources":["../src/analyst/chat-client.ts"],"sourcesContent":["/**\n * Provider-neutral chat contract for every model call made by agent-eval.\n *\n * The caller owns model execution. Evaluation code receives canonical requests\n * and results without importing a provider SDK, and agent-eval reads no\n * environment variable to find a provider credential: every transport is bound\n * at the call site.\n *\n * Two ways to bind one. `custom` and `sandbox-sdk` take a `chat` function the\n * caller wrote. `openai-compatible` takes an endpoint and a credential the\n * caller passes as values, and agent-eval issues the HTTP request itself.\n *\n * WHY THE THIRD ONE EXISTS. Every consumer that wanted a plain OpenAI-style\n * endpoint had to hand-write the same fetch loop — retry on 429 and 5xx,\n * `json_schema` degrade, fenced-JSON stripping, and, load-bearing here, the\n * `servedModel` echo that `assertServedModel` and `assertCrossFamilyServed`\n * read. A hand-written transport that omits `servedModel` makes both identity\n * checks read `unreported`, so the two-vendor judge rule they enforce silently\n * measures nothing. Measured motive: agent-eval 0.172.1 exported no such\n * transport, so a two-model judge panel in discovery-lab reached\n * `createChatClient({ transport: 'custom', baseUrl, apiKey })` — fields that\n * exist on no transport option — and every call threw\n * `TypeError: opts.chat is not a function`, which surfaced as\n * `aggregateJudgeVerdicts: all 2 judges failed`.\n *\n * The credential still never comes from agent-eval's environment. `baseUrl`\n * and the bearer are required arguments with no default and no fallback, so a\n * misconfigured caller gets a refusal at construction, never a call to an\n * endpoint it did not name.\n */\n\nimport {\n  type LlmCallRequest,\n  type LlmCallResult,\n  LlmClient,\n  type LlmClientOptions,\n} from '../llm-client'\n\n/**\n * Unified chat interface using the package's canonical LLM request and result.\n */\nexport interface ChatClient {\n  /** Display name of the bound transport, included in telemetry. */\n  readonly transport: ChatTransport\n  /** Default model when the caller omits one. */\n  readonly defaultModel?: string\n  /** Total provider attempts this transport can make for one chat call. */\n  readonly maximumAttempts?: number\n\n  /** Implementations must enforce `req.maxTokens` when it is present. */\n  chat(req: ChatRequest, opts?: ChatCallOpts): Promise<ChatResponse>\n}\n\nexport type ChatTransport =\n  | 'sandbox-sdk' // box.streamPrompt() — chat completion via sandbox SDK\n  | 'custom' // caller-adapted SDK or transport\n  | 'openai-compatible' // caller-named /v1 endpoint, driven by this package\n  | 'mock' // test-time injection\n\nexport interface ChatRequest extends Omit<LlmCallRequest, 'model'> {\n  /** Optional — falls back to ChatClient.defaultModel. */\n  model?: string\n}\n\nexport type ChatResponse = LlmCallResult\n\nexport interface ChatCallOpts {\n  /** Cancel the in-flight request. */\n  signal?: AbortSignal\n  /** Hard USD ceiling for this single call (informational; the underlying transport may not enforce). */\n  maxCostUsd?: number\n  /** Correlation tag carried into request headers when the transport allows. */\n  correlationId?: string\n  /** Stable provider idempotency key for retries/redrives of one paid call. */\n  idempotencyKey?: string\n}\n\n// ── Factory ─────────────────────────────────────────────────────────\n\nexport type CreateChatClientOpts =\n  | SandboxSdkTransportOpts\n  | CustomTransportOpts\n  | OpenAiCompatibleTransportOpts\n  | MockTransportOpts\n\ninterface BaseTransportOpts {\n  defaultModel?: string\n  /** Total provider attempts. Required for opaque transports used in capped runs. */\n  maximumAttempts?: number\n}\n\n/**\n * Sandbox-SDK transport. The caller supplies a canonical chat function for an\n * already-configured Sandbox handle, so agent-eval does not import the SDK.\n */\nexport interface SandboxSdkTransportOpts extends BaseTransportOpts {\n  transport: 'sandbox-sdk'\n  chat: (req: ChatRequest, opts?: ChatCallOpts) => Promise<ChatResponse>\n}\n\n/** Caller-adapted SDK or transport returning the canonical ChatResponse shape. */\nexport interface CustomTransportOpts extends BaseTransportOpts {\n  transport: 'custom'\n  chat: (req: ChatRequest, opts?: ChatCallOpts) => Promise<ChatResponse>\n}\n\n/**\n * OpenAI-compatible HTTP transport: the caller names a `/v1` endpoint and\n * hands over a bearer, and this package drives `POST {baseUrl}/chat/completions`.\n *\n * Use this instead of hand-rolling a `custom` transport around `fetch`. The\n * result carries `servedModel` — the model id the provider echoed, verbatim —\n * which is the only field `assertServedModel` and `assertCrossFamilyServed`\n * can read to witness a gateway answering from a different model than the one\n * requested.\n *\n * `baseUrl` ends at the `/v1` prefix; the `/chat/completions` path is this\n * package's to append. A `baseUrl` that already carries the path is a\n * construction error, not a request to a doubled URL.\n *\n * Exactly one credential form is required — `apiKey`, `bearer`, or\n * `authHeader`. There is no environment fallback: agent-eval never goes\n * looking for a key.\n */\nexport interface OpenAiCompatibleTransportOpts\n  extends BaseTransportOpts,\n    Pick<\n      LlmClientOptions,\n      | 'assertServedModel'\n      | 'customTokenPricing'\n      | 'deadlineMs'\n      | 'defaultTimeoutMs'\n      | 'fetch'\n      | 'jsonPayloadMode'\n      | 'jsonSchemaTransport'\n      | 'minIntervalMs'\n      | 'provider'\n      | 'rawSink'\n      | 'signal'\n      | 'thinking'\n    > {\n  transport: 'openai-compatible'\n  /** Endpoint ending at the `/v1` prefix. Required — there is no default endpoint. */\n  baseUrl: string\n  /** Bearer credential. One of `apiKey`, `bearer`, or `authHeader` is required. */\n  apiKey?: string\n  /** Bearer credential, alternate spelling. */\n  bearer?: string\n  /** Non-bearer authorization header, for endpoints that want their own scheme. */\n  authHeader?: { name: string; value: string }\n}\n\n/**\n * Mock transport for tests. The handler receives the request and returns\n * whatever the test wants. No retries, no JSON-schema degrade.\n */\nexport interface MockTransportOpts extends BaseTransportOpts {\n  transport: 'mock'\n  handler: (req: ChatRequest, opts?: ChatCallOpts) => Promise<ChatResponse>\n}\n\n/**\n * Build a ChatClient bound to a specific transport. The returned client\n * is safe to share across analysts in a single registry run.\n */\nexport function createChatClient(opts: CreateChatClientOpts): ChatClient {\n  switch (opts.transport) {\n    case 'sandbox-sdk':\n      return {\n        transport: 'sandbox-sdk',\n        defaultModel: opts.defaultModel,\n        maximumAttempts: opts.maximumAttempts,\n        chat: async (req, callOpts) => opts.chat(resolveModel(req, opts.defaultModel), callOpts),\n      }\n    case 'custom':\n      return {\n        transport: 'custom',\n        defaultModel: opts.defaultModel,\n        maximumAttempts: opts.maximumAttempts,\n        chat: async (req, callOpts) => opts.chat(resolveModel(req, opts.defaultModel), callOpts),\n      }\n    case 'openai-compatible':\n      return openAiCompatibleClient(opts)\n    case 'mock':\n      return {\n        transport: 'mock',\n        defaultModel: opts.defaultModel,\n        maximumAttempts: 1,\n        chat: async (req, callOpts) => opts.handler(resolveModel(req, opts.defaultModel), callOpts),\n      }\n  }\n}\n\nfunction resolveModel(req: ChatRequest, defaultModel: string | undefined): ChatRequest {\n  if (req.model) return req\n  if (!defaultModel) {\n    throw new Error(\n      'ChatClient.chat: no model on request and no defaultModel on the client. ' +\n        'Either pass req.model or bind defaultModel at createChatClient().',\n    )\n  }\n  return { ...req, model: defaultModel }\n}\n\n/**\n * Build the OpenAI-compatible client, refusing a half-configured one first.\n *\n * Both refusals are construction-time on purpose. A missing credential or a\n * doubled path is a caller mistake that a request would turn into a 401 or a\n * 404 from an endpoint the caller did not intend to reach — the failure then\n * reads as a provider problem instead of a configuration one.\n */\nfunction openAiCompatibleClient(opts: OpenAiCompatibleTransportOpts): ChatClient {\n  const baseUrl = opts.baseUrl?.trim().replace(/\\/+$/, '') ?? ''\n  if (!baseUrl) {\n    throw new Error(\n      \"createChatClient({ transport: 'openai-compatible' }): baseUrl is required and has no default.\",\n    )\n  }\n  if (/\\/chat\\/completions$/.test(baseUrl)) {\n    throw new Error(\n      \"createChatClient({ transport: 'openai-compatible' }): baseUrl ends at the /v1 prefix; \" +\n        `this package appends /chat/completions. Got ${baseUrl}`,\n    )\n  }\n  if (!opts.apiKey?.trim() && !opts.bearer?.trim() && !opts.authHeader) {\n    throw new Error(\n      \"createChatClient({ transport: 'openai-compatible' }): one of apiKey, bearer, or authHeader \" +\n        'is required. agent-eval reads no environment variable for a credential.',\n    )\n  }\n\n  const clientOptions: LlmClientOptions = {\n    baseUrl,\n    ...(opts.apiKey ? { apiKey: opts.apiKey } : {}),\n    ...(opts.bearer ? { bearer: opts.bearer } : {}),\n    ...(opts.authHeader ? { authHeader: opts.authHeader } : {}),\n    ...(opts.maximumAttempts === undefined ? {} : { maximumAttempts: opts.maximumAttempts }),\n    ...(opts.assertServedModel === undefined ? {} : { assertServedModel: opts.assertServedModel }),\n    ...(opts.customTokenPricing === undefined\n      ? {}\n      : { customTokenPricing: opts.customTokenPricing }),\n    ...(opts.deadlineMs === undefined ? {} : { deadlineMs: opts.deadlineMs }),\n    ...(opts.defaultTimeoutMs === undefined ? {} : { defaultTimeoutMs: opts.defaultTimeoutMs }),\n    ...(opts.fetch === undefined ? {} : { fetch: opts.fetch }),\n    ...(opts.jsonPayloadMode === undefined ? {} : { jsonPayloadMode: opts.jsonPayloadMode }),\n    ...(opts.jsonSchemaTransport === undefined\n      ? {}\n      : { jsonSchemaTransport: opts.jsonSchemaTransport }),\n    ...(opts.minIntervalMs === undefined ? {} : { minIntervalMs: opts.minIntervalMs }),\n    ...(opts.provider === undefined ? {} : { provider: opts.provider }),\n    ...(opts.rawSink === undefined ? {} : { rawSink: opts.rawSink }),\n    ...(opts.signal === undefined ? {} : { signal: opts.signal }),\n    ...(opts.thinking === undefined ? {} : { thinking: opts.thinking }),\n  }\n  const client = new LlmClient(clientOptions)\n\n  return {\n    transport: 'openai-compatible',\n    ...(opts.defaultModel ? { defaultModel: opts.defaultModel } : {}),\n    maximumAttempts: client.maximumAttempts,\n    chat: async (req, callOpts) => {\n      const resolved = resolveModel(req, opts.defaultModel)\n      return client.call(resolved as LlmCallRequest, {\n        ...(callOpts?.signal ? { signal: callOpts.signal } : {}),\n        ...(callOpts?.idempotencyKey ? { idempotencyKey: callOpts.idempotencyKey } : {}),\n      })\n    },\n  }\n}\n"],"mappings":";;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;AAqKA,SAAgB,iBAAiB,MAAwC;CACvE,QAAQ,KAAK,WAAb;EACE,KAAK,eACH,OAAO;GACL,WAAW;GACX,cAAc,KAAK;GACnB,iBAAiB,KAAK;GACtB,MAAM,OAAO,KAAK,aAAa,KAAK,KAAK,aAAa,KAAK,KAAK,YAAY,GAAG,QAAQ;EACzF;EACF,KAAK,UACH,OAAO;GACL,WAAW;GACX,cAAc,KAAK;GACnB,iBAAiB,KAAK;GACtB,MAAM,OAAO,KAAK,aAAa,KAAK,KAAK,aAAa,KAAK,KAAK,YAAY,GAAG,QAAQ;EACzF;EACF,KAAK,qBACH,OAAO,uBAAuB,IAAI;EACpC,KAAK,QACH,OAAO;GACL,WAAW;GACX,cAAc,KAAK;GACnB,iBAAiB;GACjB,MAAM,OAAO,KAAK,aAAa,KAAK,QAAQ,aAAa,KAAK,KAAK,YAAY,GAAG,QAAQ;EAC5F;CACJ;AACF;AAEA,SAAS,aAAa,KAAkB,cAA+C;CACrF,IAAI,IAAI,OAAO,OAAO;CACtB,IAAI,CAAC,cACH,MAAM,IAAI,MACR,2IAEF;CAEF,OAAO;EAAE,GAAG;EAAK,OAAO;CAAa;AACvC;;;;;;;;;AAUA,SAAS,uBAAuB,MAAiD;CAC/E,MAAM,UAAU,KAAK,SAAS,KAAK,CAAC,CAAC,QAAQ,QAAQ,EAAE,KAAK;CAC5D,IAAI,CAAC,SACH,MAAM,IAAI,MACR,+FACF;CAEF,IAAI,uBAAuB,KAAK,OAAO,GACrC,MAAM,IAAI,MACR,qIACiD,SACnD;CAEF,IAAI,CAAC,KAAK,QAAQ,KAAK,KAAK,CAAC,KAAK,QAAQ,KAAK,KAAK,CAAC,KAAK,YACxD,MAAM,IAAI,MACR,oKAEF;CAGF,MAAM,gBAAkC;EACtC;EACA,GAAI,KAAK,SAAS,EAAE,QAAQ,KAAK,OAAO,IAAI,CAAC;EAC7C,GAAI,KAAK,SAAS,EAAE,QAAQ,KAAK,OAAO,IAAI,CAAC;EAC7C,GAAI,KAAK,aAAa,EAAE,YAAY,KAAK,WAAW,IAAI,CAAC;EACzD,GAAI,KAAK,oBAAoB,KAAA,IAAY,CAAC,IAAI,EAAE,iBAAiB,KAAK,gBAAgB;EACtF,GAAI,KAAK,sBAAsB,KAAA,IAAY,CAAC,IAAI,EAAE,mBAAmB,KAAK,kBAAkB;EAC5F,GAAI,KAAK,uBAAuB,KAAA,IAC5B,CAAC,IACD,EAAE,oBAAoB,KAAK,mBAAmB;EAClD,GAAI,KAAK,eAAe,KAAA,IAAY,CAAC,IAAI,EAAE,YAAY,KAAK,WAAW;EACvE,GAAI,KAAK,qBAAqB,KAAA,IAAY,CAAC,IAAI,EAAE,kBAAkB,KAAK,iBAAiB;EACzF,GAAI,KAAK,UAAU,KAAA,IAAY,CAAC,IAAI,EAAE,OAAO,KAAK,MAAM;EACxD,GAAI,KAAK,oBAAoB,KAAA,IAAY,CAAC,IAAI,EAAE,iBAAiB,KAAK,gBAAgB;EACtF,GAAI,KAAK,wBAAwB,KAAA,IAC7B,CAAC,IACD,EAAE,qBAAqB,KAAK,oBAAoB;EACpD,GAAI,KAAK,kBAAkB,KAAA,IAAY,CAAC,IAAI,EAAE,eAAe,KAAK,cAAc;EAChF,GAAI,KAAK,aAAa,KAAA,IAAY,CAAC,IAAI,EAAE,UAAU,KAAK,SAAS;EACjE,GAAI,KAAK,YAAY,KAAA,IAAY,CAAC,IAAI,EAAE,SAAS,KAAK,QAAQ;EAC9D,GAAI,KAAK,WAAW,KAAA,IAAY,CAAC,IAAI,EAAE,QAAQ,KAAK,OAAO;EAC3D,GAAI,KAAK,aAAa,KAAA,IAAY,CAAC,IAAI,EAAE,UAAU,KAAK,SAAS;CACnE;CACA,MAAM,SAAS,IAAI,UAAU,aAAa;CAE1C,OAAO;EACL,WAAW;EACX,GAAI,KAAK,eAAe,EAAE,cAAc,KAAK,aAAa,IAAI,CAAC;EAC/D,iBAAiB,OAAO;EACxB,MAAM,OAAO,KAAK,aAAa;GAC7B,MAAM,WAAW,aAAa,KAAK,KAAK,YAAY;GACpD,OAAO,OAAO,KAAK,UAA4B;IAC7C,GAAI,UAAU,SAAS,EAAE,QAAQ,SAAS,OAAO,IAAI,CAAC;IACtD,GAAI,UAAU,iBAAiB,EAAE,gBAAgB,SAAS,eAAe,IAAI,CAAC;GAChF,CAAC;EACH;CACF;AACF"}

package/dist/{chat-json-call-puMArmtW.js → chat-json-call-DavzexSp.js}

@@ -1,4 +1,5 @@
-import { a as maximumChargeForLlmRequest, b as extractJsonPayload, i as costReceiptFromLlmError, r as costReceiptFromLlm } from "./llm-client-CyBAoYiX.js";
+import { t as extractJsonPayload } from "./json-BhdsJaOQ.js";
+import { a as maximumChargeForLlmRequest, i as costReceiptFromLlmError, r as costReceiptFromLlm } from "./llm-client-BshCREn6.js";
 //#region src/chat-json-call.ts
 /** Parse a JSON answer out of a model response. The transport may fence it. */
 function parseJsonAnswer(response, actor) {
@@ -60,4 +61,4 @@ async function paidJsonChat(input) {
 //#endregion
 export { paidJsonChat as n, paidChat as t };
 
-//# sourceMappingURL=chat-json-call-puMArmtW.js.map
+//# sourceMappingURL=chat-json-call-DavzexSp.js.map

package/dist/chat-json-call-DavzexSp.js.map

@@ -0,0 +1 @@
+{"version":3,"file":"chat-json-call-DavzexSp.js","names":[],"sources":["../src/chat-json-call.ts"],"sourcesContent":["/**\n * One paid model call through a caller-owned `ChatClient`, metered by the\n * cost ledger.\n *\n * agent-eval executes no paid model: the transport is supplied by the caller\n * and the credential never enters this package. What stays here is the\n * accounting around the call — the priced maximum reserved before it runs, the\n * stable call id forwarded as the provider idempotency key, the receipt\n * settled from the response, and an honest unknown-usage receipt when the\n * transport failed.\n *\n * The judges, the wire judge endpoint, and the in-process trace-analysis\n * engine all make the same call in the same order; this is the one copy of\n * that sequence. `paidChat` is the sequence; `paidJsonChat` adds the JSON\n * answer parse that a structured caller needs.\n */\n\nimport type { ChatClient, ChatResponse } from './analyst/chat-client'\nimport type { CostChannel, CostLedgerHandle, CostReceipt, CustomTokenPricing } from './cost-ledger'\nimport {\n  costReceiptFromLlm,\n  costReceiptFromLlmError,\n  extractJsonPayload,\n  type LlmCallRequest,\n  maximumChargeForLlmRequest,\n} from './llm-client'\n\ninterface PaidChatInput {\n  /** Caller-owned transport. One `chat()` call. */\n  chat: ChatClient\n  /** The exact canonical request, including its JSON-mode or schema fields. */\n  request: LlmCallRequest\n  ledger: CostLedgerHandle\n  channel: CostChannel\n  phase: string\n  actor: string\n  tags?: Record<string, string>\n  signal?: AbortSignal\n  /** Endpoint rates used when the transport reports no billed amount. */\n  pricing?: CustomTokenPricing\n}\n\ntype PaidJsonChatInput = PaidChatInput\n\ntype PaidChatResult =\n  | { succeeded: true; response: ChatResponse; receipt: CostReceipt }\n  | { succeeded: false; error: Error; receipt?: CostReceipt }\n\ntype PaidJsonChatResult<T> =\n  | { succeeded: true; value: T; response: ChatResponse; receipt: CostReceipt }\n  | { succeeded: false; error: Error; receipt?: CostReceipt }\n\n/** Parse a JSON answer out of a model response. The transport may fence it. */\nfunction parseJsonAnswer<T>(response: ChatResponse, actor: string): T {\n  try {\n    return JSON.parse(extractJsonPayload(response.content)) as T\n  } catch (error) {\n    throw new Error(\n      `${actor}: model answer was not JSON — ${error instanceof Error ? error.message : String(error)}`,\n    )\n  }\n}\n\n/** Reserve, execute, and settle one caller-owned chat call. */\nexport async function paidChat(input: PaidChatInput): Promise<PaidChatResult> {\n  const paid = await input.ledger.runPaidCall({\n    channel: input.channel,\n    phase: input.phase,\n    actor: input.actor,\n    model: input.request.model,\n    ...(input.tags && Object.keys(input.tags).length > 0 ? { tags: input.tags } : {}),\n    maximumCharge: maximumChargeForLlmRequest(input.request, {\n      ...(input.chat.maximumAttempts === undefined\n        ? {}\n        : { maximumAttempts: input.chat.maximumAttempts }),\n      ...(input.pricing ? { customTokenPricing: input.pricing } : {}),\n    }),\n    ...(input.signal ? { signal: input.signal } : {}),\n    execute: (signal, callId) => input.chat.chat(input.request, { signal, idempotencyKey: callId }),\n    receipt: (response) => costReceiptFromLlm(response, input.pricing),\n    receiptFromError: (error) => costReceiptFromLlmError(error, input.pricing),\n  })\n  if (!paid.succeeded) {\n    return {\n      succeeded: false,\n      error: paid.error,\n      ...(paid.receipt ? { receipt: paid.receipt } : {}),\n    }\n  }\n  return { succeeded: true, response: paid.value, receipt: paid.receipt }\n}\n\nexport async function paidJsonChat<T>(input: PaidJsonChatInput): Promise<PaidJsonChatResult<T>> {\n  const paid = await paidChat(input)\n  if (!paid.succeeded) return paid\n  // The call completed and was billed. A malformed answer is a contract\n  // failure AFTER the money was spent, so it keeps the settled receipt instead\n  // of reporting the spend as unknown.\n  try {\n    return {\n      succeeded: true,\n      value: parseJsonAnswer<T>(paid.response, input.actor),\n      response: paid.response,\n      receipt: paid.receipt,\n    }\n  } catch (error) {\n    return {\n      succeeded: false,\n      error: error instanceof Error ? error : new Error(String(error)),\n      receipt: paid.receipt,\n    }\n  }\n}\n"],"mappings":";;;;AAqDA,SAAS,gBAAmB,UAAwB,OAAkB;CACpE,IAAI;EACF,OAAO,KAAK,MAAM,mBAAmB,SAAS,OAAO,CAAC;CACxD,SAAS,OAAO;EACd,MAAM,IAAI,MACR,GAAG,MAAM,gCAAgC,iBAAiB,QAAQ,MAAM,UAAU,OAAO,KAAK,GAChG;CACF;AACF;;AAGA,eAAsB,SAAS,OAA+C;CAC5E,MAAM,OAAO,MAAM,MAAM,OAAO,YAAY;EAC1C,SAAS,MAAM;EACf,OAAO,MAAM;EACb,OAAO,MAAM;EACb,OAAO,MAAM,QAAQ;EACrB,GAAI,MAAM,QAAQ,OAAO,KAAK,MAAM,IAAI,CAAC,CAAC,SAAS,IAAI,EAAE,MAAM,MAAM,KAAK,IAAI,CAAC;EAC/E,eAAe,2BAA2B,MAAM,SAAS;GACvD,GAAI,MAAM,KAAK,oBAAoB,KAAA,IAC/B,CAAC,IACD,EAAE,iBAAiB,MAAM,KAAK,gBAAgB;GAClD,GAAI,MAAM,UAAU,EAAE,oBAAoB,MAAM,QAAQ,IAAI,CAAC;EAC/D,CAAC;EACD,GAAI,MAAM,SAAS,EAAE,QAAQ,MAAM,OAAO,IAAI,CAAC;EAC/C,UAAU,QAAQ,WAAW,MAAM,KAAK,KAAK,MAAM,SAAS;GAAE;GAAQ,gBAAgB;EAAO,CAAC;EAC9F,UAAU,aAAa,mBAAmB,UAAU,MAAM,OAAO;EACjE,mBAAmB,UAAU,wBAAwB,OAAO,MAAM,OAAO;CAC3E,CAAC;CACD,IAAI,CAAC,KAAK,WACR,OAAO;EACL,WAAW;EACX,OAAO,KAAK;EACZ,GAAI,KAAK,UAAU,EAAE,SAAS,KAAK,QAAQ,IAAI,CAAC;CAClD;CAEF,OAAO;EAAE,WAAW;EAAM,UAAU,KAAK;EAAO,SAAS,KAAK;CAAQ;AACxE;AAEA,eAAsB,aAAgB,OAA0D;CAC9F,MAAM,OAAO,MAAM,SAAS,KAAK;CACjC,IAAI,CAAC,KAAK,WAAW,OAAO;CAI5B,IAAI;EACF,OAAO;GACL,WAAW;GACX,OAAO,gBAAmB,KAAK,UAAU,MAAM,KAAK;GACpD,UAAU,KAAK;GACf,SAAS,KAAK;EAChB;CACF,SAAS,OAAO;EACd,OAAO;GACL,WAAW;GACX,OAAO,iBAAiB,QAAQ,QAAQ,IAAI,MAAM,OAAO,KAAK,CAAC;GAC/D,SAAS,KAAK;EAChB;CACF;AACF"}

package/dist/{claim-q4HQhK1G.js → claim-BPVuyHuy.js}

@@ -1,5 +1,5 @@
-import { s as ValidationError } from "./errors-DVZiqS7K.js";
-import { i as compareCodeUnits } from "./canonical-B2RqLofv.js";
+import { s as ValidationError } from "./errors-BcOs5MKW.js";
+import { i as compareCodeUnits } from "./canonical-Dq2hOQ8L.js";
 import { r as readField } from "./predicate-FdKmElA_.js";
 import { z } from "zod";
 //#region src/experiment/claim.ts
@@ -51,4 +51,4 @@ function summarizeEvaluationUnits(claim, rows) {
 //#endregion
 export { evaluationClaimSchema as n, summarizeEvaluationUnits as r, defineEvaluationClaim as t };
 
-//# sourceMappingURL=claim-q4HQhK1G.js.map
+//# sourceMappingURL=claim-BPVuyHuy.js.map

package/dist/{claim-q4HQhK1G.js.map → claim-BPVuyHuy.js.map}

@@ -1 +1 @@
-{"version":3,"file":"claim-q4HQhK1G.js","names":[],"sources":["../src/experiment/claim.ts"],"sourcesContent":["import { z } from 'zod'\nimport { ValidationError } from '../errors'\nimport { compareCodeUnits } from '../ledger-core/canonical'\nimport type { EvidenceRecord } from './ast/evidence'\nimport { readField } from './ast/predicate'\n\nconst nonEmpty = z\n  .string()\n  .min(1)\n  .refine((value) => value.trim() === value)\n\n/** The claim's schema, shared by every record that binds a claim. */\nexport const evaluationClaimSchema = z\n  .object({\n    use: z.enum(['development', 'comparison', 'certification']),\n    population: z.object({ id: nonEmpty, description: nonEmpty }).strict(),\n    samplingFrame: nonEmpty,\n    independentUnit: nonEmpty,\n    generalization: z.enum(['fixed-roster', 'new-units']),\n    minimumEffect: z.number().finite().positive().optional(),\n  })\n  .strict()\n\n/** The population and independent observations a measured result can describe. */\nexport type EvaluationClaim = z.infer<typeof evaluationClaimSchema>\n\n/** Validate a claim before binding it to a sealed experiment or final evidence. */\nexport function defineEvaluationClaim(input: EvaluationClaim): EvaluationClaim {\n  const parsed = evaluationClaimSchema.safeParse(input)\n  if (!parsed.success)\n    throw new ValidationError(`invalid evaluation claim: ${parsed.error.message}`)\n  return Object.freeze({ ...parsed.data, population: Object.freeze(parsed.data.population) })\n}\n\nexport interface EvaluationUnitSummary {\n  observations: number\n  independentUnits: number\n  units: Array<{ id: string; observations: number }>\n}\n\n/** Repetitions retain their denominator without becoming additional independent units. */\nexport function summarizeEvaluationUnits(\n  claim: EvaluationClaim,\n  rows: readonly object[],\n): EvaluationUnitSummary {\n  const validated = defineEvaluationClaim(claim)\n  const counts = new Map<string, number>()\n  for (const [index, row] of rows.entries()) {\n    if (row === null || typeof row !== 'object' || Array.isArray(row)) {\n      throw new ValidationError(`evaluation claim: row ${index} must be an object`)\n    }\n    const value = readField(row as EvidenceRecord, validated.independentUnit)\n    if (typeof value !== 'string' || !value.trim() || value.trim() !== value) {\n      throw new ValidationError(\n        `evaluation claim: row ${index} needs a nonempty string at '${validated.independentUnit}'`,\n      )\n    }\n    counts.set(value, (counts.get(value) ?? 0) + 1)\n  }\n  return {\n    observations: rows.length,\n    independentUnits: counts.size,\n    units: [...counts]\n      .sort(([left], [right]) => compareCodeUnits(left, right))\n      .map(([id, observations]) => ({ id, observations })),\n  }\n}\n"],"mappings":";;;;;AAMA,MAAM,WAAW,EACd,OAAO,CAAC,CACR,IAAI,CAAC,CAAC,CACN,QAAQ,UAAU,MAAM,KAAK,MAAM,KAAK;;AAG3C,MAAa,wBAAwB,EAClC,OAAO;CACN,KAAK,EAAE,KAAK;EAAC;EAAe;EAAc;CAAe,CAAC;CAC1D,YAAY,EAAE,OAAO;EAAE,IAAI;EAAU,aAAa;CAAS,CAAC,CAAC,CAAC,OAAO;CACrE,eAAe;CACf,iBAAiB;CACjB,gBAAgB,EAAE,KAAK,CAAC,gBAAgB,WAAW,CAAC;CACpD,eAAe,EAAE,OAAO,CAAC,CAAC,OAAO,CAAC,CAAC,SAAS,CAAC,CAAC,SAAS;AACzD,CAAC,CAAC,CACD,OAAO;;AAMV,SAAgB,sBAAsB,OAAyC;CAC7E,MAAM,SAAS,sBAAsB,UAAU,KAAK;CACpD,IAAI,CAAC,OAAO,SACV,MAAM,IAAI,gBAAgB,6BAA6B,OAAO,MAAM,SAAS;CAC/E,OAAO,OAAO,OAAO;EAAE,GAAG,OAAO;EAAM,YAAY,OAAO,OAAO,OAAO,KAAK,UAAU;CAAE,CAAC;AAC5F;;AASA,SAAgB,yBACd,OACA,MACuB;CACvB,MAAM,YAAY,sBAAsB,KAAK;CAC7C,MAAM,yBAAS,IAAI,IAAoB;CACvC,KAAK,MAAM,CAAC,OAAO,QAAQ,KAAK,QAAQ,GAAG;EACzC,IAAI,QAAQ,QAAQ,OAAO,QAAQ,YAAY,MAAM,QAAQ,GAAG,GAC9D,MAAM,IAAI,gBAAgB,yBAAyB,MAAM,mBAAmB;EAE9E,MAAM,QAAQ,UAAU,KAAuB,UAAU,eAAe;EACxE,IAAI,OAAO,UAAU,YAAY,CAAC,MAAM,KAAK,KAAK,MAAM,KAAK,MAAM,OACjE,MAAM,IAAI,gBACR,yBAAyB,MAAM,+BAA+B,UAAU,gBAAgB,EAC1F;EAEF,OAAO,IAAI,QAAQ,OAAO,IAAI,KAAK,KAAK,KAAK,CAAC;CAChD;CACA,OAAO;EACL,cAAc,KAAK;EACnB,kBAAkB,OAAO;EACzB,OAAO,CAAC,GAAG,MAAM,CAAC,CACf,MAAM,CAAC,OAAO,CAAC,WAAW,iBAAiB,MAAM,KAAK,CAAC,CAAC,CACxD,KAAK,CAAC,IAAI,mBAAmB;GAAE;GAAI;EAAa,EAAE;CACvD;AACF"}
+{"version":3,"file":"claim-BPVuyHuy.js","names":[],"sources":["../src/experiment/claim.ts"],"sourcesContent":["import { z } from 'zod'\nimport { ValidationError } from '../errors'\nimport { compareCodeUnits } from '../ledger-core/canonical'\nimport type { EvidenceRecord } from './ast/evidence'\nimport { readField } from './ast/predicate'\n\nconst nonEmpty = z\n  .string()\n  .min(1)\n  .refine((value) => value.trim() === value)\n\n/** The claim's schema, shared by every record that binds a claim. */\nexport const evaluationClaimSchema = z\n  .object({\n    use: z.enum(['development', 'comparison', 'certification']),\n    population: z.object({ id: nonEmpty, description: nonEmpty }).strict(),\n    samplingFrame: nonEmpty,\n    independentUnit: nonEmpty,\n    generalization: z.enum(['fixed-roster', 'new-units']),\n    minimumEffect: z.number().finite().positive().optional(),\n  })\n  .strict()\n\n/** The population and independent observations a measured result can describe. */\nexport type EvaluationClaim = z.infer<typeof evaluationClaimSchema>\n\n/** Validate a claim before binding it to a sealed experiment or final evidence. */\nexport function defineEvaluationClaim(input: EvaluationClaim): EvaluationClaim {\n  const parsed = evaluationClaimSchema.safeParse(input)\n  if (!parsed.success)\n    throw new ValidationError(`invalid evaluation claim: ${parsed.error.message}`)\n  return Object.freeze({ ...parsed.data, population: Object.freeze(parsed.data.population) })\n}\n\nexport interface EvaluationUnitSummary {\n  observations: number\n  independentUnits: number\n  units: Array<{ id: string; observations: number }>\n}\n\n/** Repetitions retain their denominator without becoming additional independent units. */\nexport function summarizeEvaluationUnits(\n  claim: EvaluationClaim,\n  rows: readonly object[],\n): EvaluationUnitSummary {\n  const validated = defineEvaluationClaim(claim)\n  const counts = new Map<string, number>()\n  for (const [index, row] of rows.entries()) {\n    if (row === null || typeof row !== 'object' || Array.isArray(row)) {\n      throw new ValidationError(`evaluation claim: row ${index} must be an object`)\n    }\n    const value = readField(row as EvidenceRecord, validated.independentUnit)\n    if (typeof value !== 'string' || !value.trim() || value.trim() !== value) {\n      throw new ValidationError(\n        `evaluation claim: row ${index} needs a nonempty string at '${validated.independentUnit}'`,\n      )\n    }\n    counts.set(value, (counts.get(value) ?? 0) + 1)\n  }\n  return {\n    observations: rows.length,\n    independentUnits: counts.size,\n    units: [...counts]\n      .sort(([left], [right]) => compareCodeUnits(left, right))\n      .map(([id, observations]) => ({ id, observations })),\n  }\n}\n"],"mappings":";;;;;AAMA,MAAM,WAAW,EACd,OAAO,CAAC,CACR,IAAI,CAAC,CAAC,CACN,QAAQ,UAAU,MAAM,KAAK,MAAM,KAAK;;AAG3C,MAAa,wBAAwB,EAClC,OAAO;CACN,KAAK,EAAE,KAAK;EAAC;EAAe;EAAc;CAAe,CAAC;CAC1D,YAAY,EAAE,OAAO;EAAE,IAAI;EAAU,aAAa;CAAS,CAAC,CAAC,CAAC,OAAO;CACrE,eAAe;CACf,iBAAiB;CACjB,gBAAgB,EAAE,KAAK,CAAC,gBAAgB,WAAW,CAAC;CACpD,eAAe,EAAE,OAAO,CAAC,CAAC,OAAO,CAAC,CAAC,SAAS,CAAC,CAAC,SAAS;AACzD,CAAC,CAAC,CACD,OAAO;;AAMV,SAAgB,sBAAsB,OAAyC;CAC7E,MAAM,SAAS,sBAAsB,UAAU,KAAK;CACpD,IAAI,CAAC,OAAO,SACV,MAAM,IAAI,gBAAgB,6BAA6B,OAAO,MAAM,SAAS;CAC/E,OAAO,OAAO,OAAO;EAAE,GAAG,OAAO;EAAM,YAAY,OAAO,OAAO,OAAO,KAAK,UAAU;CAAE,CAAC;AAC5F;;AASA,SAAgB,yBACd,OACA,MACuB;CACvB,MAAM,YAAY,sBAAsB,KAAK;CAC7C,MAAM,yBAAS,IAAI,IAAoB;CACvC,KAAK,MAAM,CAAC,OAAO,QAAQ,KAAK,QAAQ,GAAG;EACzC,IAAI,QAAQ,QAAQ,OAAO,QAAQ,YAAY,MAAM,QAAQ,GAAG,GAC9D,MAAM,IAAI,gBAAgB,yBAAyB,MAAM,mBAAmB;EAE9E,MAAM,QAAQ,UAAU,KAAuB,UAAU,eAAe;EACxE,IAAI,OAAO,UAAU,YAAY,CAAC,MAAM,KAAK,KAAK,MAAM,KAAK,MAAM,OACjE,MAAM,IAAI,gBACR,yBAAyB,MAAM,+BAA+B,UAAU,gBAAgB,EAC1F;EAEF,OAAO,IAAI,QAAQ,OAAO,IAAI,KAAK,KAAK,KAAK,CAAC;CAChD;CACA,OAAO;EACL,cAAc,KAAK;EACnB,kBAAkB,OAAO;EACzB,OAAO,CAAC,GAAG,MAAM,CAAC,CACf,MAAM,CAAC,OAAO,CAAC,WAAW,iBAAiB,MAAM,KAAK,CAAC,CAAC,CACxD,KAAK,CAAC,IAAI,mBAAmB;GAAE;GAAI;EAAa,EAAE;CACvD;AACF"}

package/dist/cli.js

@@ -1,19 +1,20 @@
 #!/usr/bin/env node
-import { l as operatorYield, u as INSUFFICIENT_FROM } from "./pareto-Cm6ndrWN.js";
-import { i as replaySearchLedgerText } from "./search-ledger-CVUGLDZp.js";
-import { t as renderSearchSummary } from "./search-summary-BQCSWf1g.js";
-import { c as hostedTenantFromEnv, g as SearchRunKindSchema, r as shipSearchLedger, t as SearchShipConflictError, u as SEARCH_LEDGER_BATCH_MAX_BYTES } from "./search-shipper-BOSD46qv.js";
-import { i as runRolloutReleaseCli } from "./hf-dataset-BR2yhuNn.js";
-import { t as createChatClient } from "./chat-client-l-MVDlDk.js";
-import { t as runAnalystBenchmarkCommand } from "./benchmark-command-By0vh8q7.js";
-import { a as runRpcBatch, o as runRpcOnce, p as handleVersion, r as startServerAsync, s as buildOpenApi } from "./server-e9oWO_bN.js";
-import { C as editCredit, S as front, g as formatLandscape, l as metaSearch, n as taskMatrix, o as skillManifold, p as landscape, r as formatSkillManifold, s as renderMetaSearchText, t as tree, w as editCreditText, y as surfaceTextEdits } from "./tree-C1XYeXvy.js";
-import { t as runSupervisorRunCommand } from "./report-command-SC9nqb4M.js";
+import { c as errorMessage } from "./errors-BcOs5MKW.js";
+import { l as operatorYield, u as INSUFFICIENT_FROM } from "./pareto-BzsxidFD.js";
+import { i as replaySearchLedgerText } from "./search-ledger-TUvB5gWH.js";
+import { t as renderSearchSummary } from "./search-summary-DQIvJsSB.js";
+import { l as SEARCH_LEDGER_BATCH_MAX_BYTES, m as SearchRunKindSchema, r as shipSearchLedger, s as hostedTenantFromEnv, t as SearchShipConflictError } from "./search-shipper-Cw2z3-td.js";
+import { n as runRolloutReleaseCli } from "./hf-dataset-D9f7-oX7.js";
+import { t as createChatClient } from "./chat-client-ldqxYYTN.js";
+import { t as runAnalystBenchmarkCommand } from "./benchmark-command-xQK8nOI6.js";
+import { i as buildOpenApi, l as handleVersion, o as runRpcBatch, r as startServerAsync, s as runRpcOnce } from "./server-DoVbqqtJ.js";
+import { b as editCreditText, f as landscape, g as surfaceTextEdits, l as metaSearch, m as formatLandscape, n as taskMatrix, o as skillManifold, r as formatSkillManifold, s as renderMetaSearchText, t as tree, v as front, y as editCredit } from "./tree-BGRglkKd.js";
+import { i as analyzeSupervisorRun, n as renderSupervisorRunHeadline, r as renderSupervisorRunMarkdown } from "./render-C-w2EZ6p.js";
 import { createHash } from "node:crypto";
 import { existsSync, readFileSync, writeFileSync } from "node:fs";
-import { dirname, join } from "node:path";
+import { dirname, join, resolve } from "node:path";
 import { fileURLToPath } from "node:url";
-import { open, readFile } from "node:fs/promises";
+import { open, readFile, stat } from "node:fs/promises";
 //#region src/cli-config.ts
 /**
 * Provider configuration for the `agent-eval` binary.
@@ -435,6 +436,106 @@ async function firstLineSearchId(path) {
 	}
 }
 //#endregion
+//#region src/supervisor-run/report-command.ts
+/**
+* `agent-eval supervisor-run report <runDir>` takes one run directory and
+* prints the report the module already renders. The command reads through
+* `analyzeSupervisorRun`, so a Runtime run directory takes the
+* same path and the status comes from the record `terminal-record.ts` names.
+*
+* Exit codes follow the other self-parsing subcommands: 2 for a usage error,
+* 1 when the directory cannot be read (missing, not a directory, or a corrupt
+* journal or terminal record), 0 after the report was written to stdout.
+*/
+const FORMATS = [
+	"headline",
+	"markdown",
+	"json"
+];
+const SUPERVISOR_RUN_USAGE = "usage: agent-eval supervisor-run report <runDir> [--format headline|markdown|json]";
+const PROCESS_IO = {
+	stdout: (text) => {
+		process.stdout.write(text);
+	},
+	stderr: (text) => {
+		process.stderr.write(text);
+	}
+};
+function isFormat(value) {
+	return FORMATS.includes(value);
+}
+function parseReportArgs(argv) {
+	let runDir = null;
+	let format = "headline";
+	for (let i = 0; i < argv.length; i++) {
+		const arg = argv[i];
+		if (arg === "--format") {
+			const raw = argv[++i];
+			if (raw === void 0 || !isFormat(raw)) throw new Error(`--format expects one of ${FORMATS.join("|")}, got ${JSON.stringify(raw ?? "")}`);
+			format = raw;
+		} else if (arg.startsWith("--format=")) {
+			const raw = arg.slice(9);
+			if (!isFormat(raw)) throw new Error(`--format expects one of ${FORMATS.join("|")}, got ${JSON.stringify(raw)}`);
+			format = raw;
+		} else if (arg.startsWith("--")) throw new Error(`unknown flag "${arg}"`);
+		else if (runDir === null) runDir = arg;
+		else throw new Error(`unexpected argument "${arg}"`);
+	}
+	if (runDir === null) throw new Error(SUPERVISOR_RUN_USAGE);
+	return {
+		runDir,
+		format
+	};
+}
+/**
+* CLI driver for `agent-eval supervisor-run`. `argv` is everything after the
+* command name, so `['report', runDir, ...flags]`. Returns the process exit code.
+*/
+async function runSupervisorRunCommand(argv, io = PROCESS_IO) {
+	const [verb, ...rest] = argv;
+	if (verb === void 0 || verb === "--help" || verb === "-h" || verb === "help") {
+		io.stdout(`${SUPERVISOR_RUN_USAGE}\n`);
+		return verb === void 0 ? 2 : 0;
+	}
+	if (verb !== "report") {
+		io.stderr(`unknown supervisor-run subcommand "${verb}"\n${SUPERVISOR_RUN_USAGE}\n`);
+		return 2;
+	}
+	if (rest.includes("--help") || rest.includes("-h")) {
+		io.stdout(`${SUPERVISOR_RUN_USAGE}\n`);
+		return 0;
+	}
+	let args;
+	try {
+		args = parseReportArgs(rest);
+	} catch (error) {
+		io.stderr(`${errorMessage(error)}\n`);
+		return 2;
+	}
+	const runDir = resolve(args.runDir);
+	const entry = await stat(runDir).catch(() => null);
+	if (entry === null || !entry.isDirectory()) {
+		io.stderr(`[agent-eval] supervisor-run report: ${runDir} is not a directory\n`);
+		return 1;
+	}
+	try {
+		const report = await analyzeSupervisorRun(runDir);
+		switch (args.format) {
+			case "headline":
+				io.stdout(`${renderSupervisorRunHeadline(report)}\n`);
+				break;
+			case "markdown":
+				io.stdout(renderSupervisorRunMarkdown(report));
+				break;
+			case "json": io.stdout(`${JSON.stringify(report, null, 2)}\n`);
+		}
+		return 0;
+	} catch (error) {
+		io.stderr(`[agent-eval] supervisor-run report failed for ${runDir}: ${errorMessage(error)}\n`);
+		return 1;
+	}
+}
+//#endregion
 //#region src/cli.ts
 /**
 * agent-eval CLI.