@theokit/sdk 5.11.0 → 5.12.0

package/dist/{chunk-R5YPDAKD.cjs.map → chunk-POBZXULR.cjs.map}

@@ -1 +1 @@
-{"version":3,"sources":["../src/internal/persistence/jsonl.ts"],"names":["TheokitAgentError","readFileSync","path","mkdirSync","dirname","openSync","writeSync","closeSync","fstatSync","readSync"],"mappings":";;;;;;AA6BO,IAAM,eAAA,GAAN,cAA8BA,mCAAA,CAAkB;AAAA,EAErD,WAAA,CACE,SACS,IAAA,EACT;AAEA,IAAA,KAAA,CAAM,SAAS,EAAE,IAAA,EAAM,oBAAA,EAAsB,WAAA,EAAa,OAAO,CAAA;AAHxD,IAAA,IAAA,CAAA,IAAA,GAAA,IAAA;AAAA,EAIX;AAAA,EAJW,IAAA;AAAA,EAHO,IAAA,GAAO,iBAAA;AAQ3B;AAEA,SAAS,cAAc,KAAA,EAAkD;AACvE,EAAA,OAAO,OAAO,UAAU,QAAA,IAAY,KAAA,KAAU,QAAQ,CAAC,KAAA,CAAM,QAAQ,KAAK,CAAA;AAC5E;AAGA,SAAS,mBAAmB,IAAA,EAAmD;AAC7E,EAAA,IAAI,IAAA,CAAK,MAAA,KAAW,CAAA,EAAG,OAAO,MAAA;AAC9B,EAAA,IAAI,MAAA;AACJ,EAAA,IAAI;AACF,IAAA,MAAA,GAAS,IAAA,CAAK,MAAM,IAAI,CAAA;AAAA,EAC1B,CAAA,CAAA,MAAQ;AACN,IAAA,OAAO,MAAA;AAAA,EACT;AACA,EAAA,OAAO,aAAA,CAAc,MAAM,CAAA,GAAI,MAAA,GAAS,MAAA;AAC1C;AAQO,SAAS,SAAA,CACd,IAAA,EACA,IAAA,GASI,EAAC,EACA;AACL,EAAA,MAAM,QAAQC,eAAA,CAAa,IAAA,EAAM,MAAM,CAAA,CAAE,MAAM,IAAI,CAAA;AACnD,EAAA,MAAM,MAAW,EAAC;AAClB,EAAA,KAAA,IAAS,IAAI,CAAA,EAAG,CAAA,GAAI,KAAA,CAAM,MAAA,EAAQ,KAAK,CAAA,EAAG;AACxC,IAAA,MAAM,aAAa,CAAA,GAAI,CAAA;AACvB,IAAA,MAAM,IAAA,GAAA,CAAQ,KAAA,CAAM,CAAC,CAAA,IAAK,IAAI,IAAA,EAAK;AACnC,IAAA,IAAI,IAAA,CAAK,WAAW,CAAA,EAAG;AAGvB,IAAA,MAAM,MAAA,GAAS,SAAA;AAAA,MACb,IAAA;AAAA,MACA,UAAA;AAAA,MACA,IAAA,CAAK,2BAAA,KAAgC,IAAA,IAAQ,UAAA,KAAe,KAAA,CAAM;AAAA,KACpE;AACA,IAAA,IAAI,WAAW,MAAA,EAAW;AAC1B,IAAA,GAAA,CAAI,IAAA,CAAK,KAAK,GAAA,GAAM,IAAA,CAAK,IAAI,MAAA,EAAQ,UAAU,IAAK,MAAuB,CAAA;AAAA,EAC7E;AACA,EAAA,OAAO,GAAA;AACT;AASA,SAAS,SAAA,CACP,IAAA,EACA,UAAA,EACA,QAAA,EACqC;AACrC,EAAA,IAAI,MAAA;AACJ,EAAA,IAAI;AACF,IAAA,MAAA,GAAS,IAAA,CAAK,MAAM,IAAI,CAAA;AAAA,EAC1B,CAAA,CAAA,MAAQ;AACN,IAAA,IAAI,UAAU,OAAO,MAAA;AACrB,IAAA,MAAM,IAAI,eAAA,CAAgB,CAAA,KAAA,EAAQ,UAAU,kBAAkB,UAAU,CAAA;AAAA,EAC1E;AACA,EAAA,IAAI,CAAC,aAAA,CAAc,MAAM,CAAA,EAAG;AAC1B,IAAA,MAAM,IAAI,eAAA,CAAgB,CAAA,KAAA,EAAQ,UAAU,uBAAuB,UAAU,CAAA;AAAA,EAC/E;AACA,EAAA,OAAO,MAAA;AACT;AAUO,SAAS,WAAA,CAAYC,QAAc,MAAA,EAAuB;AAgB/D,EAAAC,YAAA,CAAUC,YAAA,CAAQF,MAAI,CAAA,EAAG,EAAE,WAAW,IAAA,EAAM,IAAA,EAAM,KAAO,CAAA;AACzD,EAAA,MAAM,MAAA,GAAS,oBAAA,CAAqBA,MAAI,CAAA,GAAI,IAAA,GAAO,EAAA;AASnD,EAAA,MAAM,EAAA,GAAKG,WAAA,CAASH,MAAA,EAAM,GAAA,EAAK,GAAK,CAAA;AACpC,EAAA,IAAI;AACF,IAAAI,YAAA,CAAU,IAAI,CAAA,EAAG,MAAM,GAAG,IAAA,CAAK,SAAA,CAAU,MAAM,CAAC;AAAA,CAAI,CAAA;AAAA,EACtD,CAAA,SAAE;AACA,IAAAC,YAAA,CAAU,EAAE,CAAA;AAAA,EACd;AACF;AAaA,SAAS,qBAAqB,IAAA,EAAuB;AACnD,EAAA,IAAI,EAAA;AACJ,EAAA,IAAI;AACF,IAAA,EAAA,GAAKF,WAAA,CAAS,MAAM,GAAG,CAAA;AAAA,EACzB,CAAA,CAAA,MAAQ;AACN,IAAA,OAAO,KAAA;AAAA,EACT;AACA,EAAA,IAAI;AACF,IAAA,MAAM,IAAA,GAAOG,YAAA,CAAU,EAAE,CAAA,CAAE,IAAA;AAC3B,IAAA,IAAI,IAAA,KAAS,GAAG,OAAO,KAAA;AACvB,IAAA,MAAM,GAAA,GAAM,MAAA,CAAO,KAAA,CAAM,CAAC,CAAA;AAC1B,IAAAC,WAAA,CAAS,EAAA,EAAI,GAAA,EAAK,CAAA,EAAG,CAAA,EAAG,OAAO,CAAC,CAAA;AAChC,IAAA,OAAO,GAAA,CAAI,CAAC,CAAA,KAAM,EAAA;AAAA,EACpB,CAAA,SAAE;AACA,IAAAF,YAAA,CAAU,EAAE,CAAA;AAAA,EACd;AACF;AAYO,SAAS,YAAA,CACd,MACA,KAAA,EACa;AACb,EAAA,MAAM,IAAA,uBAAW,GAAA,EAAY;AAC7B,EAAA,IAAI,IAAA;AACJ,EAAA,IAAI;AACF,IAAA,IAAA,GAAON,eAAA,CAAa,MAAM,MAAM,CAAA;AAAA,EAClC,CAAA,CAAA,MAAQ;AACN,IAAA,OAAO,IAAA;AAAA,EACT;AACA,EAAA,KAAA,MAAW,OAAA,IAAW,IAAA,CAAK,KAAA,CAAM,IAAI,CAAA,EAAG;AAEtC,IAAA,MAAM,MAAA,GAAS,kBAAA,CAAmB,OAAA,CAAQ,IAAA,EAAM,CAAA;AAChD,IAAA,IAAI,WAAW,MAAA,EAAW;AAC1B,IAAA,MAAM,GAAA,GAAM,MAAM,MAAM,CAAA;AACxB,IAAA,IAAI,OAAO,QAAQ,QAAA,IAAY,GAAA,CAAI,SAAS,CAAA,EAAG,IAAA,CAAK,IAAI,GAAG,CAAA;AAAA,EAC7D;AACA,EAAA,OAAO,IAAA;AACT","file":"chunk-R5YPDAKD.cjs","sourcesContent":["/**\n * Durable JSONL primitives shared by the eval harness (M6).\n *\n * - `loadJsonl` — generic dataset reader (split/trim/skip-blank/parse) with a\n *   line-numbered {@link JsonlParseError}. The dataset SCHEMA is the caller's\n *   concern via `map` (M6 ADR D3) — this module owns only the parse.\n * - `appendJsonl` / `readJsonlIds` — crash-durable, resumable batch persistence\n *   (M6 ADR D1): each record is appended as one whole `\\n`-terminated line the\n *   instant it is produced, and a re-run resumes by skipping already-keyed rows.\n *\n * reference: knowledge-base/references/theocode-eval/lib/swebench-dataset.ts:82\n * (parseJsonl + line-N error) and swebench-batch.ts:113,205 (resume + per-line\n * flush).\n *\n * @internal\n */\nimport {\n  closeSync,\n  fstatSync,\n  mkdirSync,\n  openSync,\n  readFileSync,\n  readSync,\n  writeSync,\n} from \"node:fs\";\nimport { dirname } from \"node:path\";\nimport { TheokitAgentError } from \"../../errors.js\";\n\n/** Raised when a JSONL line is not valid JSON or is not a JSON object. Carries the 1-based line number. */\nexport class JsonlParseError extends TheokitAgentError {\n  override readonly name = \"JsonlParseError\";\n  constructor(\n    message: string,\n    readonly line: number,\n  ) {\n    // Not retryable: the bytes on the line do not change between reads.\n    super(message, { code: \"jsonl_parse_failed\", isRetryable: false });\n  }\n}\n\nfunction isPlainObject(value: unknown): value is Record<string, unknown> {\n  return typeof value === \"object\" && value !== null && !Array.isArray(value);\n}\n\n/** Parse a single trimmed line to a plain object, or `undefined` if blank / invalid / non-object. */\nfunction tryParseObjectLine(line: string): Record<string, unknown> | undefined {\n  if (line.length === 0) return undefined;\n  let parsed: unknown;\n  try {\n    parsed = JSON.parse(line);\n  } catch {\n    return undefined;\n  }\n  return isPlainObject(parsed) ? parsed : undefined;\n}\n\n/**\n * Parse a JSONL file into rows. Blank lines are skipped. A malformed or\n * non-object line throws {@link JsonlParseError} naming the 1-based line. When\n * `map` is provided, each raw object is mapped to the typed row (the SWE-bench\n * schema lives in the caller's `map`, per M6 ADR D3).\n */\nexport function loadJsonl<T = Record<string, unknown>>(\n  path: string,\n  opts: {\n    map?: (raw: Record<string, unknown>, lineNumber: number) => T;\n    /**\n     * M81 — tolerate a truncated LAST line (a crash artifact: the process died mid-write).\n     *\n     * Opt-in on purpose. As a default it would also swallow corruption in the MIDDLE of the file,\n     * turning loud data loss into silent data loss — the wrong trade for a session store.\n     */\n    tolerateTrailingPartialLine?: boolean;\n  } = {},\n): T[] {\n  const lines = readFileSync(path, \"utf8\").split(\"\\n\");\n  const out: T[] = [];\n  for (let i = 0; i < lines.length; i += 1) {\n    const lineNumber = i + 1;\n    const line = (lines[i] ?? \"\").trim();\n    if (line.length === 0) continue;\n    // M81 — `undefined` = a tolerated invalid line (the last one, truncated by a crash). Any other\n    // invalid line has already thrown inside `parseLine`.\n    const parsed = parseLine(\n      line,\n      lineNumber,\n      opts.tolerateTrailingPartialLine === true && lineNumber === lines.length,\n    );\n    if (parsed === undefined) break;\n    out.push(opts.map ? opts.map(parsed, lineNumber) : (parsed as unknown as T));\n  }\n  return out;\n}\n\n/**\n * Parses ONE line, returning `undefined` when it is a tolerated truncated last line.\n *\n * Extracted because `loadJsonl` went over the cognitive-complexity ceiling once it gained the\n * tolerance — and because parsing a line and iterating a file are two responsibilities that were\n * only together out of habit.\n */\nfunction parseLine(\n  line: string,\n  lineNumber: number,\n  tolerate: boolean,\n): Record<string, unknown> | undefined {\n  let parsed: unknown;\n  try {\n    parsed = JSON.parse(line);\n  } catch {\n    if (tolerate) return undefined;\n    throw new JsonlParseError(`line ${lineNumber}: invalid JSON`, lineNumber);\n  }\n  if (!isPlainObject(parsed)) {\n    throw new JsonlParseError(`line ${lineNumber}: not a JSON object`, lineNumber);\n  }\n  return parsed;\n}\n\n/**\n * Append one record as a whole `\\n`-terminated JSON line. Creates the parent dir\n * if missing. `appendFileSync` is synchronous, so within a single Node process\n * the event loop serializes writes and each call writes its line atomically —\n * interleave-safe for the bounded-concurrency batch runner.\n *\n * reference: swebench-batch.ts:192 (mkdir-before-append), :205 (per-line flush).\n */\nexport function appendJsonl(path: string, record: unknown): void {\n  // T2.4 — `0o700` for the same reason the file below is `0o600`, and it is the half M93 missed. A\n  // private file inside a group-writable directory can be replaced wholesale, and the replacement's\n  // mode is whatever the writer chose. Under `umask 002` this directory was born `0775`.\n  //\n  // The directory sits under `~/.theokit`, shared with the credential and trust stores, and the\n  // framework already wrote the diagnosis while fixing a sibling (`@theokit/agents\n  // config/trust-store.ts:157-161`): the mode argument is a no-op on a directory that already\n  // exists, \"and this one is shared with the SDK's transcript root — whoever creates it first sets\n  // the permissions\". Whoever creates it first is usually THIS function, because a session writes a\n  // transcript before it ever touches a credential. So `assertSecureModes` was not wrong to demand a\n  // private directory; this path was wrong to produce one that fails it, depending on run order.\n  //\n  // Creation-time only, deliberately: repairing a pre-existing directory means `stat` + `chmod` on\n  // every append, and this is the hot path of every session. The pre-existing case is what\n  // `assertSecureModes` is for.\n  mkdirSync(dirname(path), { recursive: true, mode: 0o700 });\n  const prefix = needsLineBreakBefore(path) ? \"\\n\" : \"\";\n  // M93 (adversarial review, H1) — `0o600`. `appendFileSync` takes no `mode`, so the permission came\n  // from the umask: under `umask 022` the transcript was born `0664`, world-readable. The previous\n  // path (`replaceFileAtomic`) pinned `0o600` on purpose — \"holds the FULL in-flight content\n  // (credential snapshots, OAuth tokens)\" (`atomic-write.ts:107`) — and switching to append lost\n  // that silently. The same class had already been caught in the consumer (`atomic-sync.ts`, M88\n  // HIGH-1).\n  //\n  // `mode` only applies on CREATION; a pre-existing file keeps whatever permission it already has.\n  const fd = openSync(path, \"a\", 0o600);\n  try {\n    writeSync(fd, `${prefix}${JSON.stringify(record)}\\n`);\n  } finally {\n    closeSync(fd);\n  }\n}\n\n/**\n * Does the file end without a `\\n`? Then the last line is truncated — a crash mid-append.\n *\n * Without this check the next append **glues** itself onto the middle of the broken line, producing\n * an invalid line the reader discards: the NEW record disappears along with the partial one. The\n * previous path (read-modify-write) self-healed from this because it rewrote the whole file.\n * Measured in M93's adversarial review (H2): after an append over a truncated file, the\n * just-written record was no longer readable.\n *\n * Reads **one byte**, not the file.\n */\nfunction needsLineBreakBefore(path: string): boolean {\n  let fd: number;\n  try {\n    fd = openSync(path, \"r\");\n  } catch {\n    return false; // the file does not exist yet: nothing to splice onto\n  }\n  try {\n    const size = fstatSync(fd).size;\n    if (size === 0) return false;\n    const buf = Buffer.alloc(1);\n    readSync(fd, buf, 0, 1, size - 1);\n    return buf[0] !== 0x0a;\n  } finally {\n    closeSync(fd);\n  }\n}\n\n/**\n * Read the set of keys from an existing JSONL file for which `keyFn(parsed)`\n * returns a non-empty string. Used to resume a crashed batch by skipping rows\n * already persisted with a successful result. A trailing partial line from an\n * interrupted append is tolerated (skipped, not thrown), and a missing file\n * yields an empty set.\n *\n * reference: swebench-batch.ts:113 (readDoneIds), :129 (success-only),\n * :131 (tolerate partial line).\n */\nexport function readJsonlIds(\n  path: string,\n  keyFn: (parsed: Record<string, unknown>) => string | undefined,\n): Set<string> {\n  const done = new Set<string>();\n  let text: string;\n  try {\n    text = readFileSync(path, \"utf8\");\n  } catch {\n    return done; // no file yet → nothing done\n  }\n  for (const rawLine of text.split(\"\\n\")) {\n    // A trailing partial line from an interrupted run parses to undefined → skipped.\n    const parsed = tryParseObjectLine(rawLine.trim());\n    if (parsed === undefined) continue;\n    const key = keyFn(parsed);\n    if (typeof key === \"string\" && key.length > 0) done.add(key);\n  }\n  return done;\n}\n"]}
+{"version":3,"sources":["../src/internal/persistence/jsonl.ts"],"names":["TheokitAgentError","readFileSync","path","mkdirSync","dirname","openSync","writeSync","closeSync","fstatSync","readSync"],"mappings":";;;;;;AA6BO,IAAM,eAAA,GAAN,cAA8BA,mCAAA,CAAkB;AAAA,EAErD,WAAA,CACE,SACS,IAAA,EACT;AAEA,IAAA,KAAA,CAAM,SAAS,EAAE,IAAA,EAAM,oBAAA,EAAsB,WAAA,EAAa,OAAO,CAAA;AAHxD,IAAA,IAAA,CAAA,IAAA,GAAA,IAAA;AAAA,EAIX;AAAA,EAJW,IAAA;AAAA,EAHO,IAAA,GAAO,iBAAA;AAQ3B;AAEA,SAAS,cAAc,KAAA,EAAkD;AACvE,EAAA,OAAO,OAAO,UAAU,QAAA,IAAY,KAAA,KAAU,QAAQ,CAAC,KAAA,CAAM,QAAQ,KAAK,CAAA;AAC5E;AAGA,SAAS,mBAAmB,IAAA,EAAmD;AAC7E,EAAA,IAAI,IAAA,CAAK,MAAA,KAAW,CAAA,EAAG,OAAO,MAAA;AAC9B,EAAA,IAAI,MAAA;AACJ,EAAA,IAAI;AACF,IAAA,MAAA,GAAS,IAAA,CAAK,MAAM,IAAI,CAAA;AAAA,EAC1B,CAAA,CAAA,MAAQ;AACN,IAAA,OAAO,MAAA;AAAA,EACT;AACA,EAAA,OAAO,aAAA,CAAc,MAAM,CAAA,GAAI,MAAA,GAAS,MAAA;AAC1C;AAQO,SAAS,SAAA,CACd,IAAA,EACA,IAAA,GASI,EAAC,EACA;AACL,EAAA,MAAM,QAAQC,eAAA,CAAa,IAAA,EAAM,MAAM,CAAA,CAAE,MAAM,IAAI,CAAA;AACnD,EAAA,MAAM,MAAW,EAAC;AAClB,EAAA,KAAA,IAAS,IAAI,CAAA,EAAG,CAAA,GAAI,KAAA,CAAM,MAAA,EAAQ,KAAK,CAAA,EAAG;AACxC,IAAA,MAAM,aAAa,CAAA,GAAI,CAAA;AACvB,IAAA,MAAM,IAAA,GAAA,CAAQ,KAAA,CAAM,CAAC,CAAA,IAAK,IAAI,IAAA,EAAK;AACnC,IAAA,IAAI,IAAA,CAAK,WAAW,CAAA,EAAG;AAGvB,IAAA,MAAM,MAAA,GAAS,SAAA;AAAA,MACb,IAAA;AAAA,MACA,UAAA;AAAA,MACA,IAAA,CAAK,2BAAA,KAAgC,IAAA,IAAQ,UAAA,KAAe,KAAA,CAAM;AAAA,KACpE;AACA,IAAA,IAAI,WAAW,MAAA,EAAW;AAC1B,IAAA,GAAA,CAAI,IAAA,CAAK,KAAK,GAAA,GAAM,IAAA,CAAK,IAAI,MAAA,EAAQ,UAAU,IAAK,MAAuB,CAAA;AAAA,EAC7E;AACA,EAAA,OAAO,GAAA;AACT;AASA,SAAS,SAAA,CACP,IAAA,EACA,UAAA,EACA,QAAA,EACqC;AACrC,EAAA,IAAI,MAAA;AACJ,EAAA,IAAI;AACF,IAAA,MAAA,GAAS,IAAA,CAAK,MAAM,IAAI,CAAA;AAAA,EAC1B,CAAA,CAAA,MAAQ;AACN,IAAA,IAAI,UAAU,OAAO,MAAA;AACrB,IAAA,MAAM,IAAI,eAAA,CAAgB,CAAA,KAAA,EAAQ,UAAU,kBAAkB,UAAU,CAAA;AAAA,EAC1E;AACA,EAAA,IAAI,CAAC,aAAA,CAAc,MAAM,CAAA,EAAG;AAC1B,IAAA,MAAM,IAAI,eAAA,CAAgB,CAAA,KAAA,EAAQ,UAAU,uBAAuB,UAAU,CAAA;AAAA,EAC/E;AACA,EAAA,OAAO,MAAA;AACT;AAUO,SAAS,WAAA,CAAYC,QAAc,MAAA,EAAuB;AAgB/D,EAAAC,YAAA,CAAUC,YAAA,CAAQF,MAAI,CAAA,EAAG,EAAE,WAAW,IAAA,EAAM,IAAA,EAAM,KAAO,CAAA;AACzD,EAAA,MAAM,MAAA,GAAS,oBAAA,CAAqBA,MAAI,CAAA,GAAI,IAAA,GAAO,EAAA;AASnD,EAAA,MAAM,EAAA,GAAKG,WAAA,CAASH,MAAA,EAAM,GAAA,EAAK,GAAK,CAAA;AACpC,EAAA,IAAI;AACF,IAAAI,YAAA,CAAU,IAAI,CAAA,EAAG,MAAM,GAAG,IAAA,CAAK,SAAA,CAAU,MAAM,CAAC;AAAA,CAAI,CAAA;AAAA,EACtD,CAAA,SAAE;AACA,IAAAC,YAAA,CAAU,EAAE,CAAA;AAAA,EACd;AACF;AAaA,SAAS,qBAAqB,IAAA,EAAuB;AACnD,EAAA,IAAI,EAAA;AACJ,EAAA,IAAI;AACF,IAAA,EAAA,GAAKF,WAAA,CAAS,MAAM,GAAG,CAAA;AAAA,EACzB,CAAA,CAAA,MAAQ;AACN,IAAA,OAAO,KAAA;AAAA,EACT;AACA,EAAA,IAAI;AACF,IAAA,MAAM,IAAA,GAAOG,YAAA,CAAU,EAAE,CAAA,CAAE,IAAA;AAC3B,IAAA,IAAI,IAAA,KAAS,GAAG,OAAO,KAAA;AACvB,IAAA,MAAM,GAAA,GAAM,MAAA,CAAO,KAAA,CAAM,CAAC,CAAA;AAC1B,IAAAC,WAAA,CAAS,EAAA,EAAI,GAAA,EAAK,CAAA,EAAG,CAAA,EAAG,OAAO,CAAC,CAAA;AAChC,IAAA,OAAO,GAAA,CAAI,CAAC,CAAA,KAAM,EAAA;AAAA,EACpB,CAAA,SAAE;AACA,IAAAF,YAAA,CAAU,EAAE,CAAA;AAAA,EACd;AACF;AAYO,SAAS,YAAA,CACd,MACA,KAAA,EACa;AACb,EAAA,MAAM,IAAA,uBAAW,GAAA,EAAY;AAC7B,EAAA,IAAI,IAAA;AACJ,EAAA,IAAI;AACF,IAAA,IAAA,GAAON,eAAA,CAAa,MAAM,MAAM,CAAA;AAAA,EAClC,CAAA,CAAA,MAAQ;AACN,IAAA,OAAO,IAAA;AAAA,EACT;AACA,EAAA,KAAA,MAAW,OAAA,IAAW,IAAA,CAAK,KAAA,CAAM,IAAI,CAAA,EAAG;AAEtC,IAAA,MAAM,MAAA,GAAS,kBAAA,CAAmB,OAAA,CAAQ,IAAA,EAAM,CAAA;AAChD,IAAA,IAAI,WAAW,MAAA,EAAW;AAC1B,IAAA,MAAM,GAAA,GAAM,MAAM,MAAM,CAAA;AACxB,IAAA,IAAI,OAAO,QAAQ,QAAA,IAAY,GAAA,CAAI,SAAS,CAAA,EAAG,IAAA,CAAK,IAAI,GAAG,CAAA;AAAA,EAC7D;AACA,EAAA,OAAO,IAAA;AACT","file":"chunk-POBZXULR.cjs","sourcesContent":["/**\n * Durable JSONL primitives shared by the eval harness (M6).\n *\n * - `loadJsonl` — generic dataset reader (split/trim/skip-blank/parse) with a\n *   line-numbered {@link JsonlParseError}. The dataset SCHEMA is the caller's\n *   concern via `map` (M6 ADR D3) — this module owns only the parse.\n * - `appendJsonl` / `readJsonlIds` — crash-durable, resumable batch persistence\n *   (M6 ADR D1): each record is appended as one whole `\\n`-terminated line the\n *   instant it is produced, and a re-run resumes by skipping already-keyed rows.\n *\n * reference: knowledge-base/references/theocode-eval/lib/swebench-dataset.ts:82\n * (parseJsonl + line-N error) and swebench-batch.ts:113,205 (resume + per-line\n * flush).\n *\n * @internal\n */\nimport {\n  closeSync,\n  fstatSync,\n  mkdirSync,\n  openSync,\n  readFileSync,\n  readSync,\n  writeSync,\n} from \"node:fs\";\nimport { dirname } from \"node:path\";\nimport { TheokitAgentError } from \"../../errors.js\";\n\n/** Raised when a JSONL line is not valid JSON or is not a JSON object. Carries the 1-based line number. */\nexport class JsonlParseError extends TheokitAgentError {\n  override readonly name = \"JsonlParseError\";\n  constructor(\n    message: string,\n    readonly line: number,\n  ) {\n    // Not retryable: the bytes on the line do not change between reads.\n    super(message, { code: \"jsonl_parse_failed\", isRetryable: false });\n  }\n}\n\nfunction isPlainObject(value: unknown): value is Record<string, unknown> {\n  return typeof value === \"object\" && value !== null && !Array.isArray(value);\n}\n\n/** Parse a single trimmed line to a plain object, or `undefined` if blank / invalid / non-object. */\nfunction tryParseObjectLine(line: string): Record<string, unknown> | undefined {\n  if (line.length === 0) return undefined;\n  let parsed: unknown;\n  try {\n    parsed = JSON.parse(line);\n  } catch {\n    return undefined;\n  }\n  return isPlainObject(parsed) ? parsed : undefined;\n}\n\n/**\n * Parse a JSONL file into rows. Blank lines are skipped. A malformed or\n * non-object line throws {@link JsonlParseError} naming the 1-based line. When\n * `map` is provided, each raw object is mapped to the typed row (the SWE-bench\n * schema lives in the caller's `map`, per M6 ADR D3).\n */\nexport function loadJsonl<T = Record<string, unknown>>(\n  path: string,\n  opts: {\n    map?: (raw: Record<string, unknown>, lineNumber: number) => T;\n    /**\n     * M81 — tolerate a truncated LAST line (a crash artifact: the process died mid-write).\n     *\n     * Opt-in on purpose. As a default it would also swallow corruption in the MIDDLE of the file,\n     * turning loud data loss into silent data loss — the wrong trade for a session store.\n     */\n    tolerateTrailingPartialLine?: boolean;\n  } = {},\n): T[] {\n  const lines = readFileSync(path, \"utf8\").split(\"\\n\");\n  const out: T[] = [];\n  for (let i = 0; i < lines.length; i += 1) {\n    const lineNumber = i + 1;\n    const line = (lines[i] ?? \"\").trim();\n    if (line.length === 0) continue;\n    // M81 — `undefined` = a tolerated invalid line (the last one, truncated by a crash). Any other\n    // invalid line has already thrown inside `parseLine`.\n    const parsed = parseLine(\n      line,\n      lineNumber,\n      opts.tolerateTrailingPartialLine === true && lineNumber === lines.length,\n    );\n    if (parsed === undefined) break;\n    out.push(opts.map ? opts.map(parsed, lineNumber) : (parsed as unknown as T));\n  }\n  return out;\n}\n\n/**\n * Parses ONE line, returning `undefined` when it is a tolerated truncated last line.\n *\n * Extracted because `loadJsonl` went over the cognitive-complexity ceiling once it gained the\n * tolerance — and because parsing a line and iterating a file are two responsibilities that were\n * only together out of habit.\n */\nfunction parseLine(\n  line: string,\n  lineNumber: number,\n  tolerate: boolean,\n): Record<string, unknown> | undefined {\n  let parsed: unknown;\n  try {\n    parsed = JSON.parse(line);\n  } catch {\n    if (tolerate) return undefined;\n    throw new JsonlParseError(`line ${lineNumber}: invalid JSON`, lineNumber);\n  }\n  if (!isPlainObject(parsed)) {\n    throw new JsonlParseError(`line ${lineNumber}: not a JSON object`, lineNumber);\n  }\n  return parsed;\n}\n\n/**\n * Append one record as a whole `\\n`-terminated JSON line. Creates the parent dir\n * if missing. `appendFileSync` is synchronous, so within a single Node process\n * the event loop serializes writes and each call writes its line atomically —\n * interleave-safe for the bounded-concurrency batch runner.\n *\n * reference: swebench-batch.ts:192 (mkdir-before-append), :205 (per-line flush).\n */\nexport function appendJsonl(path: string, record: unknown): void {\n  // T2.4 — `0o700` for the same reason the file below is `0o600`, and it is the half M93 missed. A\n  // private file inside a group-writable directory can be replaced wholesale, and the replacement's\n  // mode is whatever the writer chose. Under `umask 002` this directory was born `0775`.\n  //\n  // The directory sits under `~/.theokit`, shared with the credential and trust stores, and the\n  // framework already wrote the diagnosis while fixing a sibling (`@theokit/agents\n  // config/trust-store.ts:157-161`): the mode argument is a no-op on a directory that already\n  // exists, \"and this one is shared with the SDK's transcript root — whoever creates it first sets\n  // the permissions\". Whoever creates it first is usually THIS function, because a session writes a\n  // transcript before it ever touches a credential. So `assertSecureModes` was not wrong to demand a\n  // private directory; this path was wrong to produce one that fails it, depending on run order.\n  //\n  // Creation-time only, deliberately: repairing a pre-existing directory means `stat` + `chmod` on\n  // every append, and this is the hot path of every session. The pre-existing case is what\n  // `assertSecureModes` is for.\n  mkdirSync(dirname(path), { recursive: true, mode: 0o700 });\n  const prefix = needsLineBreakBefore(path) ? \"\\n\" : \"\";\n  // M93 (adversarial review, H1) — `0o600`. `appendFileSync` takes no `mode`, so the permission came\n  // from the umask: under `umask 022` the transcript was born `0664`, world-readable. The previous\n  // path (`replaceFileAtomic`) pinned `0o600` on purpose — \"holds the FULL in-flight content\n  // (credential snapshots, OAuth tokens)\" (`atomic-write.ts:107`) — and switching to append lost\n  // that silently. The same class had already been caught in the consumer (`atomic-sync.ts`, M88\n  // HIGH-1).\n  //\n  // `mode` only applies on CREATION; a pre-existing file keeps whatever permission it already has.\n  const fd = openSync(path, \"a\", 0o600);\n  try {\n    writeSync(fd, `${prefix}${JSON.stringify(record)}\\n`);\n  } finally {\n    closeSync(fd);\n  }\n}\n\n/**\n * Does the file end without a `\\n`? Then the last line is truncated — a crash mid-append.\n *\n * Without this check the next append **glues** itself onto the middle of the broken line, producing\n * an invalid line the reader discards: the NEW record disappears along with the partial one. The\n * previous path (read-modify-write) self-healed from this because it rewrote the whole file.\n * Measured in M93's adversarial review (H2): after an append over a truncated file, the\n * just-written record was no longer readable.\n *\n * Reads **one byte**, not the file.\n */\nfunction needsLineBreakBefore(path: string): boolean {\n  let fd: number;\n  try {\n    fd = openSync(path, \"r\");\n  } catch {\n    return false; // the file does not exist yet: nothing to splice onto\n  }\n  try {\n    const size = fstatSync(fd).size;\n    if (size === 0) return false;\n    const buf = Buffer.alloc(1);\n    readSync(fd, buf, 0, 1, size - 1);\n    return buf[0] !== 0x0a;\n  } finally {\n    closeSync(fd);\n  }\n}\n\n/**\n * Read the set of keys from an existing JSONL file for which `keyFn(parsed)`\n * returns a non-empty string. Used to resume a crashed batch by skipping rows\n * already persisted with a successful result. A trailing partial line from an\n * interrupted append is tolerated (skipped, not thrown), and a missing file\n * yields an empty set.\n *\n * reference: swebench-batch.ts:113 (readDoneIds), :129 (success-only),\n * :131 (tolerate partial line).\n */\nexport function readJsonlIds(\n  path: string,\n  keyFn: (parsed: Record<string, unknown>) => string | undefined,\n): Set<string> {\n  const done = new Set<string>();\n  let text: string;\n  try {\n    text = readFileSync(path, \"utf8\");\n  } catch {\n    return done; // no file yet → nothing done\n  }\n  for (const rawLine of text.split(\"\\n\")) {\n    // A trailing partial line from an interrupted run parses to undefined → skipped.\n    const parsed = tryParseObjectLine(rawLine.trim());\n    if (parsed === undefined) continue;\n    const key = keyFn(parsed);\n    if (typeof key === \"string\" && key.length > 0) done.add(key);\n  }\n  return done;\n}\n"]}