@tangle-network/agent-eval 0.221.8 → 0.222.0

package/dist/{bounded-process-Cj9YjXbE.js.map → bounded-process-BZrWXIKW.js.map}

@@ -1 +1 @@
-{"version":3,"file":"bounded-process-Cj9YjXbE.js","names":[],"sources":["../src/bounded-process.ts"],"sourcesContent":["/**\n * `runBoundedProcess` — run one command under a wall-clock bound and return\n * what it produced, leaving no descendant behind.\n *\n * The child leads its own process group (`detached` on every non-win32\n * platform), so a deadline or an abort kills the group with\n * `process.kill(-pid, 'SIGKILL')` rather than only the shell that was\n * spawned. That distinction is the whole point of this module: `sh -c\n * 'solver & wait'` puts the real work in a grandchild, and a kill aimed at\n * the shell alone leaves that grandchild running, holding the stdout and\n * stderr pipes open. `close` then never fires, so the deadline produces no\n * result the caller can observe and the descendant outlives the run.\n *\n * Two further guarantees the caller depends on:\n *\n *   - A killed run never reports success. A SIGKILLed child can still close\n *     with exit code 0, so a run killed by the deadline or by an abort has\n *     its exit code forced non-zero.\n *   - A run always resolves. If a grandchild escaped the group and still\n *     holds the pipes, the result is delivered {@link KILL_DRAIN_MS} after\n *     the kill instead of waiting on a `close` that will not come.\n *\n * The captured output is bounded. Once the cap is reached the streams are\n * still drained — the child needs its pipes read to reach `close` — but the\n * bytes are discarded and `outputTruncated` is set.\n */\n\nimport { spawn } from 'node:child_process'\nimport { StringDecoder } from 'node:string_decoder'\n\n/** Wall-clock bound applied when the caller names none. */\nexport const DEFAULT_BOUNDED_PROCESS_TIMEOUT_MS = 10 * 60_000\n\n/** Cap on captured stdout+stderr applied when the caller names none. */\nconst DEFAULT_MAX_OUTPUT_BYTES = 16 * 1024 * 1024\n\n/**\n * Window `close` is given after the process group is killed. A grandchild\n * that escaped the group can hold the pipes open indefinitely; the caller\n * receives its result at the end of this window rather than hanging.\n */\nconst KILL_DRAIN_MS = 2000\n\n/** Exit code reported for a run the wall-clock bound killed (GNU `timeout`). */\nconst TIMEOUT_EXIT_CODE = 124\n\n/** Exit code reported for a run an `AbortSignal` killed (128 + SIGKILL). */\nconst ABORT_EXIT_CODE = 137\n\n/** Exit code reported when the process could not be spawned at all. */\nconst SPAWN_FAILURE_EXIT_CODE = 127\n\n/** Exit code reported when a process closed with no code and was not killed. */\nconst UNKNOWN_EXIT_CODE = 1\n\nexport interface BoundedProcessInput {\n  /**\n   * The command line, interpreted by a shell unless `shell` is `false`. When\n   * {@link BoundedProcessInput.args} is present this is a program path\n   * instead, and no shell reads it.\n   */\n  command: string\n  /**\n   * Argument vector for `command`, delivered to the program as separate\n   * argv entries with no shell between them.\n   *\n   * This is the form to use whenever an argument carries text the caller did\n   * not author — a script body, a path, a pattern. Without it, `interpreter\n   * -flag <script>` can only be expressed by quoting the script into a\n   * shell string, and a quoting bug there is a command injection rather than\n   * a wrong answer. Each entry arrives verbatim, so there is nothing to\n   * quote: `{ command: 'bash', args: ['-n', '-c', body] }` parses `body`\n   * whatever shell metacharacters it holds.\n   *\n   * Two limits are the platform's, not this module's. A NUL byte in an entry\n   * cannot be passed to a process at all, and is reported as a `runnerError`\n   * rather than thrown. A lone surrogate is encoded to UTF-8 as U+FFFD, and a\n   * non-string entry from an untypechecked caller is coerced with `String()`,\n   * so an entry is verbatim exactly when it is a string a UTF-8 argv can hold.\n   *\n   * A shell cannot interpret an argument vector, so `args` and a truthy\n   * `shell` are contradictory. Passing both spawns nothing and reports\n   * `runnerError`.\n   */\n  args?: string[]\n  /**\n   * Text written to the child's standard input. The stream is closed after\n   * the write, so a command that reads until end-of-file sees exactly this\n   * payload and then EOF.\n   *\n   * The child's stdin is closed whether or not this field is present. A pipe\n   * nobody closes is a hang: a command that reads stdin waits for bytes that\n   * never arrive until the deadline kills it, which reads downstream as a\n   * slow command rather than as a caller that forgot to send input.\n   *\n   * The payload is written as UTF-8 and is otherwise verbatim: a NUL byte,\n   * a newline, a quote and a backslash all arrive as themselves, because\n   * nothing parses this text. The one platform limit is the one `args`\n   * already states — a lone surrogate is encoded as U+FFFD, since no UTF-8\n   * stream can carry it.\n   *\n   * A child that exits before reading breaks the pipe. That is the child's\n   * behaviour, not a runner failure, so the run still settles on the child's\n   * own exit code, and the undelivered payload is reported in `runnerError`\n   * with the `stdin not delivered:` prefix. Whether a SMALL payload to a\n   * command that ignores stdin reports this is a race — the write may land in\n   * the pipe buffer before the child exits — so the flag is a statement about\n   * delivery, never about the command. It is not a race for the case that\n   * matters: a payload larger than the pipe buffer that the child never\n   * drains cannot be delivered, and reports every time.\n   */\n  stdin?: string\n  /**\n   * Shell that interprets `command`. `true` (the default) uses the platform\n   * default — `/bin/sh` on POSIX, `cmd.exe` on Windows. A string names the\n   * shell binary, e.g. `'bash'`, which is required for a bash-only construct\n   * such as `[[ ]]` or `$BASH_VERSION`. `false` executes `command` directly\n   * as a program path with no arguments. Ignored — and refused when truthy —\n   * once {@link BoundedProcessInput.args} is present.\n   */\n  shell?: boolean | string\n  /** Working directory. The parent's cwd is inherited when omitted. */\n  cwd?: string\n  /**\n   * Environment for the child. How it combines with the parent environment\n   * is decided by {@link BoundedProcessInput.envMode}.\n   */\n  env?: NodeJS.ProcessEnv\n  /**\n   * `'merge'` (the default) layers `env` on top of `process.env`, so the\n   * child sees every variable this process has plus the named ones.\n   * `'replace'` passes exactly `env` and nothing else, which is what a\n   * grader needs when the claim under test is about the variables the\n   * command could read. `PATH` is not re-added under `'replace'`, and a\n   * command that must find a binary on `PATH` names `PATH` itself.\n   *\n   * With no `PATH` in the child environment the two forms do NOT search the\n   * same directories, so a caller under `'replace'` that names no `PATH` gets\n   * a resolution that depends on which form it picked. The shell form execs\n   * the shell, which applies its own compiled-in default; the argv form\n   * reaches `execvp`, which falls back to `confstr(_CS_PATH)`. Measured on\n   * Linux/glibc: `/usr/local/sbin:/usr/local/bin:/usr/sbin:/usr/bin:/sbin:/bin`\n   * against `/bin:/usr/bin`, so a binary in `/usr/local/bin` resolves on the\n   * shell form and reports ENOENT on the argv form. Name `PATH` and both\n   * forms search exactly what was named.\n   */\n  envMode?: 'merge' | 'replace'\n  /** Wall-clock bound in ms. Defaults to 10 minutes. */\n  timeoutMs?: number\n  /**\n   * Cap on captured stdout+stderr. Reaching it truncates the captured text\n   * and sets `outputTruncated`; it never kills the process. Defaults to\n   * 16 MiB.\n   */\n  maxOutputBytes?: number\n  /**\n   * Kills the process group when it aborts, the same way the deadline does.\n   * A signal that is already aborted when `runBoundedProcess` is called\n   * prevents the spawn entirely — nothing runs.\n   */\n  signal?: AbortSignal\n}\n\nexport interface BoundedProcessResult {\n  /**\n   * The process actually started. False when nothing ran at all: an\n   * already-aborted signal, `args` with a truthy `shell`, an argument the\n   * platform refuses, or a program that does not exist.\n   *\n   * This is the field to branch on, never the text of `runnerError`. A caller\n   * deciding whether a run produced an outcome at all is asking exactly this,\n   * and asking it here is type-checked rather than held by a wording\n   * convention.\n   */\n  spawned: boolean\n  /**\n   * The `stdin` payload reached the child in full. True whenever the caller\n   * supplied none, because nothing could fail to arrive.\n   *\n   * False means the command ran on partial input, which is a fact about the\n   * answer it produced: a checksum over a truncated payload is a wrong answer,\n   * not a failed run. `spawned` stays true and `exitCode` stays the child's.\n   */\n  stdinDelivered: boolean\n  /**\n   * Exit code. Forced non-zero whenever `killedByTimeout` or\n   * `killedBySignal` is set, because a SIGKILLed child can close with 0 and\n   * a killed run must never read as a pass.\n   */\n  exitCode: number\n  /** Captured stdout, truncated at `maxOutputBytes`. */\n  stdout: string\n  /**\n   * Captured stderr, truncated at `maxOutputBytes`. A runner-side failure is\n   * reported in `runnerError`, not appended here; the caller decides whether\n   * the command's own stderr and the runner's diagnosis belong together.\n   */\n  stderr: string\n  /** Wall-clock duration in ms, measured across the whole call. */\n  wallMs: number\n  /** The wall-clock bound was reached and the process group was killed. */\n  killedByTimeout: boolean\n  /** `signal` aborted and the process group was killed. */\n  killedBySignal: boolean\n  /** `maxOutputBytes` was reached and later output was discarded. */\n  outputTruncated: boolean\n  /**\n   * Human-readable diagnosis of a runner-side failure, in one of three shapes:\n   *\n   *   - `not spawned:` — the caller's own bug, such as `args` with a truthy\n   *     `shell`, or an already-aborted signal. No retry helps.\n   *   - `stdin not delivered:` — the command ran and exited on its own, and\n   *     the stdin payload did not reach it in full because the child closed\n   *     the pipe first.\n   *   - anything else — the process could not be spawned (a missing binary, a\n   *     NUL byte in an argument).\n   *\n   * Absent when the command ran and read everything it was sent, however it\n   * exited. The prefixes are for a person reading a log; code reads `spawned`\n   * and `stdinDelivered`, which say the same things and are type-checked.\n   */\n  runnerError?: string\n}\n\nexport async function runBoundedProcess(input: BoundedProcessInput): Promise<BoundedProcessResult> {\n  const start = Date.now()\n  const maxOutputBytes = input.maxOutputBytes ?? DEFAULT_MAX_OUTPUT_BYTES\n  const timeoutMs = input.timeoutMs ?? DEFAULT_BOUNDED_PROCESS_TIMEOUT_MS\n  const envMode = input.envMode ?? 'merge'\n  const env =\n    input.env === undefined && envMode === 'merge'\n      ? undefined\n      : envMode === 'replace'\n        ? { ...input.env }\n        : { ...process.env, ...input.env }\n\n  if (input.signal?.aborted) {\n    return {\n      spawned: false,\n      stdinDelivered: true,\n      exitCode: ABORT_EXIT_CODE,\n      stdout: '',\n      stderr: '',\n      wallMs: Date.now() - start,\n      killedByTimeout: false,\n      killedBySignal: true,\n      outputTruncated: false,\n      runnerError: 'aborted before spawn: the AbortSignal was already aborted',\n    }\n  }\n\n  // Reported rather than thrown, because this module's contract is that a\n  // call always resolves. A caller bug still reads as a failure and never as\n  // a pass: the exit code is non-zero and `runnerError` names the cause.\n  if (input.args !== undefined && input.shell) {\n    return {\n      spawned: false,\n      stdinDelivered: true,\n      exitCode: SPAWN_FAILURE_EXIT_CODE,\n      stdout: '',\n      stderr: '',\n      wallMs: Date.now() - start,\n      killedByTimeout: false,\n      killedBySignal: false,\n      outputTruncated: false,\n      // Shares `SPAWN_FAILURE_EXIT_CODE` with a program that does not exist,\n      // because both mean nothing ran. The `not spawned:` prefix is the stable\n      // discriminator: it marks the caller's own bug, which no retry fixes.\n      runnerError:\n        'not spawned: `args` passes an argument vector directly to the program, so a shell ' +\n        'cannot interpret it — pass `args` or a truthy `shell`, never both',\n    }\n  }\n\n  return await new Promise<BoundedProcessResult>((resolve) => {\n    const spawnOptions = {\n      cwd: input.cwd,\n      env,\n      // Own process group so a kill reaches the whole tree, not just the\n      // shell — otherwise a runaway grandchild keeps the pipes open and\n      // `close` never fires.\n      detached: process.platform !== 'win32',\n    }\n    // The argv form takes the same bounds, the same group and the same kill\n    // path. Two things differ, and both are stated where they are decided:\n    // no shell reads the command, and executable lookup falls back to a\n    // different default `PATH` when the caller names none (see `envMode`).\n    //\n    // `spawn` validates its arguments and THROWS synchronously — a NUL byte in\n    // `command` or in an `args` entry, or an `args` that is not an array from\n    // an untypechecked caller. A throw inside this executor would reject, and\n    // this module's contract is that a call always resolves. A caller that\n    // grades untrusted text must be able to score that text as failed rather\n    // than die on it.\n    let child: ReturnType<typeof spawn>\n    try {\n      child =\n        input.args === undefined\n          ? spawn(input.command, { ...spawnOptions, shell: input.shell ?? true })\n          : spawn(input.command, input.args, { ...spawnOptions, shell: false })\n    } catch (err) {\n      resolve({\n        spawned: false,\n        stdinDelivered: true,\n        exitCode: SPAWN_FAILURE_EXIT_CODE,\n        stdout: '',\n        stderr: '',\n        wallMs: Date.now() - start,\n        killedByTimeout: false,\n        killedBySignal: false,\n        outputTruncated: false,\n        runnerError: String(err),\n      })\n      return\n    }\n    // A pipe nobody closes is a hang, so stdin is always ended — after the\n    // payload when there is one, immediately when there is not. A write that\n    // fails is recorded and never thrown: the child already ran, and this\n    // module's contract is that a call always resolves.\n    let stdinError: string | undefined\n    const childStdin = child.stdin\n    if (childStdin) {\n      childStdin.on('error', (err) => {\n        // Closing an unused pipe races a command that exits at once, and that\n        // broken pipe says nothing about a caller who sent nothing. Only a\n        // payload the caller supplied can fail to be delivered.\n        if (input.stdin !== undefined) stdinError ??= `stdin not delivered: ${String(err)}`\n      })\n      if (input.stdin === undefined) childStdin.end()\n      else childStdin.end(input.stdin, 'utf8')\n    } else if (input.stdin !== undefined) {\n      stdinError = 'stdin not delivered: the child was spawned with no stdin pipe'\n    }\n\n    let stdout = ''\n    let stderr = ''\n    let outputBytes = 0\n    let outputTruncated = false\n    let killedByTimeout = false\n    let killedBySignal = false\n    let settled = false\n\n    // Bound the in-memory buffer so a runaway process can't OOM the caller.\n    // Once the cap is hit the streams are still drained (the child needs them\n    // read to reach `close`) but the bytes are discarded.\n    function capture(s: string): string {\n      if (outputBytes >= maxOutputBytes) {\n        outputTruncated = true\n        return ''\n      }\n      const room = maxOutputBytes - outputBytes\n      if (s.length > room) {\n        outputBytes = maxOutputBytes\n        outputTruncated = true\n        return s.slice(0, room)\n      }\n      outputBytes += s.length\n      return s\n    }\n\n    // Decoded across chunk boundaries. A stream arrives as Buffers split at\n    // arbitrary byte offsets, so decoding each chunk on its own replaces any\n    // multi-byte character that straddles a boundary with U+FFFD — the output\n    // is then corrupted at every 64 KiB of non-ASCII text. The decoder holds\n    // the partial sequence until its remaining bytes arrive.\n    const stdoutDecoder = new StringDecoder('utf8')\n    const stderrDecoder = new StringDecoder('utf8')\n    const decode = (decoder: StringDecoder, d: unknown): string =>\n      Buffer.isBuffer(d) ? decoder.write(d) : String(d)\n\n    const onStdout = (d: unknown) => {\n      const chunk = capture(decode(stdoutDecoder, d))\n      if (chunk) stdout += chunk\n    }\n    const onStderr = (d: unknown) => {\n      const chunk = capture(decode(stderrDecoder, d))\n      if (chunk) stderr += chunk\n    }\n\n    child.stdout?.on('data', onStdout)\n    child.stderr?.on('data', onStderr)\n\n    let forceResolve: ReturnType<typeof setTimeout> | undefined\n\n    const cleanup = () => {\n      clearTimeout(timeout)\n      if (forceResolve) clearTimeout(forceResolve)\n      input.signal?.removeEventListener('abort', onAbort)\n      child.stdout?.off('data', onStdout)\n      child.stderr?.off('data', onStderr)\n      // The stdin `error` listener stays attached on purpose: a stream that\n      // emits `error` with no listener throws, and a late EPIPE after the run\n      // has settled must not take the process down.\n      child.removeAllListeners('close')\n      child.removeAllListeners('error')\n    }\n\n    // Resolve exactly once. `code` comes from `close`; on a kill path the\n    // exit code is forced non-zero so a SIGKILLed child that reports 0 can\n    // never read as a clean pass downstream.\n    const finish = (outcome: { code: number | null; spawned?: boolean; runnerError?: string }) => {\n      if (settled) return\n      settled = true\n      cleanup()\n      const spawned = outcome.spawned ?? true\n      resolve({\n        spawned,\n        // A program that never started was sent nothing, whatever the write to\n        // its doomed pipe reported.\n        stdinDelivered: spawned ? stdinError === undefined : true,\n        exitCode: exitCodeFor(outcome.code, killedByTimeout, killedBySignal),\n        stdout,\n        stderr,\n        wallMs: Date.now() - start,\n        killedByTimeout,\n        killedBySignal,\n        outputTruncated,\n        runnerError: outcome.runnerError ?? stdinError,\n      })\n    }\n\n    // SIGKILL the whole process group (we spawned detached). On a clean close\n    // this never runs.\n    const killTree = () => {\n      try {\n        if (process.platform !== 'win32' && typeof child.pid === 'number') {\n          process.kill(-child.pid, 'SIGKILL')\n        } else {\n          child.kill('SIGKILL')\n        }\n      } catch (err) {\n        // ESRCH means the group exited between the decision to kill and the\n        // signal, which `close` reports on its own. Anything else means the\n        // group may still be running, which the caller must hear about.\n        if ((err as NodeJS.ErrnoException)?.code !== 'ESRCH') {\n          console.warn('[bounded-process] SIGKILL of the process group failed:', err)\n        }\n      }\n    }\n\n    // Give `close` a brief window to fire (flushing any final output) after\n    // the group dies; if an orphaned grandchild keeps the pipes open, resolve\n    // anyway so the caller can't hang on a runaway. The deadline and an abort\n    // can both fire, so the kill happens at most once and the drain window is\n    // not restarted by the second one.\n    let killed = false\n    const killAndDrain = () => {\n      if (killed) return\n      killed = true\n      killTree()\n      forceResolve = setTimeout(() => finish({ code: null }), KILL_DRAIN_MS)\n    }\n\n    function onAbort() {\n      if (settled) return\n      killedBySignal = true\n      killAndDrain()\n    }\n\n    const timeout = setTimeout(() => {\n      killedByTimeout = true\n      killAndDrain()\n    }, timeoutMs)\n\n    input.signal?.addEventListener('abort', onAbort, { once: true })\n\n    child.on('close', (code) => {\n      if (forceResolve) clearTimeout(forceResolve)\n      finish({ code })\n    })\n    child.on('error', (err) => {\n      if (forceResolve) clearTimeout(forceResolve)\n      // `error` on a child means the program never started — a missing binary,\n      // a refused argument. Nothing ran, whatever the stdin write did.\n      finish({ code: SPAWN_FAILURE_EXIT_CODE, spawned: false, runnerError: String(err) })\n    })\n  })\n}\n\n/**\n * A killed run reports the code the child closed with only when that code is\n * already non-zero; otherwise it reports the conventional code for the kind\n * of kill. A run that closed on its own with no code at all reports 1.\n */\nfunction exitCodeFor(\n  code: number | null,\n  killedByTimeout: boolean,\n  killedBySignal: boolean,\n): number {\n  if (code !== null && code !== 0) return code\n  if (killedByTimeout) return TIMEOUT_EXIT_CODE\n  if (killedBySignal) return ABORT_EXIT_CODE\n  return code ?? UNKNOWN_EXIT_CODE\n}\n"],"mappings":";;;;;;;;;;;;;;;;;;;;;;;;;;;;;;AA+BA,MAAa,qCAAqC;;AAGlD,MAAM,2BAA2B;;;;;;AAOjC,MAAM,gBAAgB;;AAGtB,MAAM,oBAAoB;;AAG1B,MAAM,kBAAkB;;AAGxB,MAAM,0BAA0B;;AAGhC,MAAM,oBAAoB;AA2K1B,eAAsB,kBAAkB,OAA2D;CACjG,MAAM,QAAQ,KAAK,IAAI;CACvB,MAAM,iBAAiB,MAAM,kBAAkB;CAC/C,MAAM,YAAY,MAAM,aAAA;CACxB,MAAM,UAAU,MAAM,WAAW;CACjC,MAAM,MACJ,MAAM,QAAQ,KAAA,KAAa,YAAY,UACnC,KAAA,IACA,YAAY,YACV,EAAE,GAAG,MAAM,IAAI,IACf;EAAE,GAAG,QAAQ;EAAK,GAAG,MAAM;CAAI;CAEvC,IAAI,MAAM,QAAQ,SAChB,OAAO;EACL,SAAS;EACT,gBAAgB;EAChB,UAAU;EACV,QAAQ;EACR,QAAQ;EACR,QAAQ,KAAK,IAAI,IAAI;EACrB,iBAAiB;EACjB,gBAAgB;EAChB,iBAAiB;EACjB,aAAa;CACf;CAMF,IAAI,MAAM,SAAS,KAAA,KAAa,MAAM,OACpC,OAAO;EACL,SAAS;EACT,gBAAgB;EAChB,UAAU;EACV,QAAQ;EACR,QAAQ;EACR,QAAQ,KAAK,IAAI,IAAI;EACrB,iBAAiB;EACjB,gBAAgB;EAChB,iBAAiB;EAIjB,aACE;CAEJ;CAGF,OAAO,MAAM,IAAI,SAA+B,YAAY;EAC1D,MAAM,eAAe;GACnB,KAAK,MAAM;GACX;GAIA,UAAU,QAAQ,aAAa;EACjC;EAYA,IAAI;EACJ,IAAI;GACF,QACE,MAAM,SAAS,KAAA,IACX,MAAM,MAAM,SAAS;IAAE,GAAG;IAAc,OAAO,MAAM,SAAS;GAAK,CAAC,IACpE,MAAM,MAAM,SAAS,MAAM,MAAM;IAAE,GAAG;IAAc,OAAO;GAAM,CAAC;EAC1E,SAAS,KAAK;GACZ,QAAQ;IACN,SAAS;IACT,gBAAgB;IAChB,UAAU;IACV,QAAQ;IACR,QAAQ;IACR,QAAQ,KAAK,IAAI,IAAI;IACrB,iBAAiB;IACjB,gBAAgB;IAChB,iBAAiB;IACjB,aAAa,OAAO,GAAG;GACzB,CAAC;GACD;EACF;EAKA,IAAI;EACJ,MAAM,aAAa,MAAM;EACzB,IAAI,YAAY;GACd,WAAW,GAAG,UAAU,QAAQ;IAI9B,IAAI,MAAM,UAAU,KAAA,GAAW,eAAe,wBAAwB,OAAO,GAAG;GAClF,CAAC;GACD,IAAI,MAAM,UAAU,KAAA,GAAW,WAAW,IAAI;QACzC,WAAW,IAAI,MAAM,OAAO,MAAM;EACzC,OAAO,IAAI,MAAM,UAAU,KAAA,GACzB,aAAa;EAGf,IAAI,SAAS;EACb,IAAI,SAAS;EACb,IAAI,cAAc;EAClB,IAAI,kBAAkB;EACtB,IAAI,kBAAkB;EACtB,IAAI,iBAAiB;EACrB,IAAI,UAAU;EAKd,SAAS,QAAQ,GAAmB;GAClC,IAAI,eAAe,gBAAgB;IACjC,kBAAkB;IAClB,OAAO;GACT;GACA,MAAM,OAAO,iBAAiB;GAC9B,IAAI,EAAE,SAAS,MAAM;IACnB,cAAc;IACd,kBAAkB;IAClB,OAAO,EAAE,MAAM,GAAG,IAAI;GACxB;GACA,eAAe,EAAE;GACjB,OAAO;EACT;EAOA,MAAM,gBAAgB,IAAI,cAAc,MAAM;EAC9C,MAAM,gBAAgB,IAAI,cAAc,MAAM;EAC9C,MAAM,UAAU,SAAwB,MACtC,OAAO,SAAS,CAAC,IAAI,QAAQ,MAAM,CAAC,IAAI,OAAO,CAAC;EAElD,MAAM,YAAY,MAAe;GAC/B,MAAM,QAAQ,QAAQ,OAAO,eAAe,CAAC,CAAC;GAC9C,IAAI,OAAO,UAAU;EACvB;EACA,MAAM,YAAY,MAAe;GAC/B,MAAM,QAAQ,QAAQ,OAAO,eAAe,CAAC,CAAC;GAC9C,IAAI,OAAO,UAAU;EACvB;EAEA,MAAM,QAAQ,GAAG,QAAQ,QAAQ;EACjC,MAAM,QAAQ,GAAG,QAAQ,QAAQ;EAEjC,IAAI;EAEJ,MAAM,gBAAgB;GACpB,aAAa,OAAO;GACpB,IAAI,cAAc,aAAa,YAAY;GAC3C,MAAM,QAAQ,oBAAoB,SAAS,OAAO;GAClD,MAAM,QAAQ,IAAI,QAAQ,QAAQ;GAClC,MAAM,QAAQ,IAAI,QAAQ,QAAQ;GAIlC,MAAM,mBAAmB,OAAO;GAChC,MAAM,mBAAmB,OAAO;EAClC;EAKA,MAAM,UAAU,YAA8E;GAC5F,IAAI,SAAS;GACb,UAAU;GACV,QAAQ;GACR,MAAM,UAAU,QAAQ,WAAW;GACnC,QAAQ;IACN;IAGA,gBAAgB,UAAU,eAAe,KAAA,IAAY;IACrD,UAAU,YAAY,QAAQ,MAAM,iBAAiB,cAAc;IACnE;IACA;IACA,QAAQ,KAAK,IAAI,IAAI;IACrB;IACA;IACA;IACA,aAAa,QAAQ,eAAe;GACtC,CAAC;EACH;EAIA,MAAM,iBAAiB;GACrB,IAAI;IACF,IAAI,QAAQ,aAAa,WAAW,OAAO,MAAM,QAAQ,UACvD,QAAQ,KAAK,CAAC,MAAM,KAAK,SAAS;SAElC,MAAM,KAAK,SAAS;GAExB,SAAS,KAAK;IAIZ,IAAK,KAA+B,SAAS,SAC3C,QAAQ,KAAK,0DAA0D,GAAG;GAE9E;EACF;EAOA,IAAI,SAAS;EACb,MAAM,qBAAqB;GACzB,IAAI,QAAQ;GACZ,SAAS;GACT,SAAS;GACT,eAAe,iBAAiB,OAAO,EAAE,MAAM,KAAK,CAAC,GAAG,aAAa;EACvE;EAEA,SAAS,UAAU;GACjB,IAAI,SAAS;GACb,iBAAiB;GACjB,aAAa;EACf;EAEA,MAAM,UAAU,iBAAiB;GAC/B,kBAAkB;GAClB,aAAa;EACf,GAAG,SAAS;EAEZ,MAAM,QAAQ,iBAAiB,SAAS,SAAS,EAAE,MAAM,KAAK,CAAC;EAE/D,MAAM,GAAG,UAAU,SAAS;GAC1B,IAAI,cAAc,aAAa,YAAY;GAC3C,OAAO,EAAE,KAAK,CAAC;EACjB,CAAC;EACD,MAAM,GAAG,UAAU,QAAQ;GACzB,IAAI,cAAc,aAAa,YAAY;GAG3C,OAAO;IAAE,MAAM;IAAyB,SAAS;IAAO,aAAa,OAAO,GAAG;GAAE,CAAC;EACpF,CAAC;CACH,CAAC;AACH;;;;;;AAOA,SAAS,YACP,MACA,iBACA,gBACQ;CACR,IAAI,SAAS,QAAQ,SAAS,GAAG,OAAO;CACxC,IAAI,iBAAiB,OAAO;CAC5B,IAAI,gBAAgB,OAAO;CAC3B,OAAO,QAAQ;AACjB"}
+{"version":3,"file":"bounded-process-BZrWXIKW.js","names":[],"sources":["../src/bounded-process.ts"],"sourcesContent":["/**\n * `runBoundedProcess` — run one command under a wall-clock bound and return\n * what it produced, leaving no descendant behind.\n *\n * The child leads its own process group (`detached` on every non-win32\n * platform), so a deadline or an abort kills the group with\n * `process.kill(-pid, 'SIGKILL')` rather than only the shell that was\n * spawned. That distinction is the whole point of this module: `sh -c\n * 'solver & wait'` puts the real work in a grandchild, and a kill aimed at\n * the shell alone leaves that grandchild running, holding the stdout and\n * stderr pipes open. `close` then never fires, so the deadline produces no\n * result the caller can observe and the descendant outlives the run.\n *\n * Two further guarantees the caller depends on:\n *\n *   - A killed run never reports success. A SIGKILLed child can still close\n *     with exit code 0, so a run killed by the deadline or by an abort has\n *     its exit code forced non-zero.\n *   - A run always resolves. If a grandchild escaped the group and still\n *     holds the pipes, the result is delivered {@link KILL_DRAIN_MS} after\n *     the kill instead of waiting on a `close` that will not come.\n *\n * The captured output is bounded. Once the cap is reached the streams are\n * still drained — the child needs its pipes read to reach `close` — but the\n * bytes are discarded and `outputTruncated` is set.\n */\n\nimport { spawn } from 'node:child_process'\nimport { StringDecoder } from 'node:string_decoder'\n\n/** Wall-clock bound applied when the caller names none. */\nexport const DEFAULT_BOUNDED_PROCESS_TIMEOUT_MS = 10 * 60_000\n\n/** Cap on captured stdout+stderr applied when the caller names none. */\nconst DEFAULT_MAX_OUTPUT_BYTES = 16 * 1024 * 1024\n\n/**\n * Window `close` is given after the process group is killed. A grandchild\n * that escaped the group can hold the pipes open indefinitely; the caller\n * receives its result at the end of this window rather than hanging.\n */\nconst KILL_DRAIN_MS = 2000\n\n/** Exit code reported for a run the wall-clock bound killed (GNU `timeout`). */\nconst TIMEOUT_EXIT_CODE = 124\n\n/** Exit code reported for a run an `AbortSignal` killed (128 + SIGKILL). */\nconst ABORT_EXIT_CODE = 137\n\n/** Exit code reported when the process could not be spawned at all. */\nconst SPAWN_FAILURE_EXIT_CODE = 127\n\n/** Exit code reported when a process closed with no code and was not killed. */\nconst UNKNOWN_EXIT_CODE = 1\n\nexport interface BoundedProcessInput {\n  /**\n   * The command line, interpreted by a shell unless `shell` is `false`. When\n   * {@link BoundedProcessInput.args} is present this is a program path\n   * instead, and no shell reads it.\n   */\n  command: string\n  /**\n   * Argument vector for `command`, delivered to the program as separate\n   * argv entries with no shell between them.\n   *\n   * This is the form to use whenever an argument carries text the caller did\n   * not author — a script body, a path, a pattern. Without it, `interpreter\n   * -flag <script>` can only be expressed by quoting the script into a\n   * shell string, and a quoting bug there is a command injection rather than\n   * a wrong answer. Each entry arrives verbatim, so there is nothing to\n   * quote: `{ command: 'bash', args: ['-n', '-c', body] }` parses `body`\n   * whatever shell metacharacters it holds.\n   *\n   * Two limits are the platform's, not this module's. A NUL byte in an entry\n   * cannot be passed to a process at all, and is reported as a `runnerError`\n   * rather than thrown. A lone surrogate is encoded to UTF-8 as U+FFFD, and a\n   * non-string entry from an untypechecked caller is coerced with `String()`,\n   * so an entry is verbatim exactly when it is a string a UTF-8 argv can hold.\n   *\n   * A shell cannot interpret an argument vector, so `args` and a truthy\n   * `shell` are contradictory. Passing both spawns nothing and reports\n   * `runnerError`.\n   */\n  args?: string[]\n  /**\n   * Text written to the child's standard input. The stream is closed after\n   * the write, so a command that reads until end-of-file sees exactly this\n   * payload and then EOF.\n   *\n   * The child's stdin is closed whether or not this field is present. A pipe\n   * nobody closes is a hang: a command that reads stdin waits for bytes that\n   * never arrive until the deadline kills it, which reads downstream as a\n   * slow command rather than as a caller that forgot to send input.\n   *\n   * The payload is written as UTF-8 and is otherwise verbatim: a NUL byte,\n   * a newline, a quote and a backslash all arrive as themselves, because\n   * nothing parses this text. The one platform limit is the one `args`\n   * already states — a lone surrogate is encoded as U+FFFD, since no UTF-8\n   * stream can carry it.\n   *\n   * A child that exits before reading breaks the pipe. That is the child's\n   * behaviour, not a runner failure, so the run still settles on the child's\n   * own exit code, and the undelivered payload is reported in `runnerError`\n   * with the `stdin not delivered:` prefix. Whether a SMALL payload to a\n   * command that ignores stdin reports this is a race — the write may land in\n   * the pipe buffer before the child exits — so the flag is a statement about\n   * delivery, never about the command. It is not a race for the case that\n   * matters: a payload larger than the pipe buffer that the child never\n   * drains cannot be delivered, and reports every time.\n   */\n  stdin?: string\n  /**\n   * Shell that interprets `command`. `true` (the default) uses the platform\n   * default — `/bin/sh` on POSIX, `cmd.exe` on Windows. A string names the\n   * shell binary, e.g. `'bash'`, which is required for a bash-only construct\n   * such as `[[ ]]` or `$BASH_VERSION`. `false` executes `command` directly\n   * as a program path with no arguments. Ignored — and refused when truthy —\n   * once {@link BoundedProcessInput.args} is present.\n   */\n  shell?: boolean | string\n  /** Working directory. The parent's cwd is inherited when omitted. */\n  cwd?: string\n  /**\n   * Environment for the child. How it combines with the parent environment\n   * is decided by {@link BoundedProcessInput.envMode}.\n   */\n  env?: NodeJS.ProcessEnv\n  /**\n   * `'merge'` (the default) layers `env` on top of `process.env`, so the\n   * child sees every variable this process has plus the named ones.\n   * `'replace'` passes exactly `env` and nothing else, which is what a\n   * grader needs when the claim under test is about the variables the\n   * command could read. `PATH` is not re-added under `'replace'`, and a\n   * command that must find a binary on `PATH` names `PATH` itself.\n   *\n   * With no `PATH` in the child environment the two forms do NOT search the\n   * same directories, so a caller under `'replace'` that names no `PATH` gets\n   * a resolution that depends on which form it picked. The shell form execs\n   * the shell, which applies its own compiled-in default; the argv form\n   * reaches `execvp`, which falls back to `confstr(_CS_PATH)`. Measured on\n   * Linux/glibc: `/usr/local/sbin:/usr/local/bin:/usr/sbin:/usr/bin:/sbin:/bin`\n   * against `/bin:/usr/bin`, so a binary in `/usr/local/bin` resolves on the\n   * shell form and reports ENOENT on the argv form. Name `PATH` and both\n   * forms search exactly what was named.\n   */\n  envMode?: 'merge' | 'replace'\n  /** Wall-clock bound in ms. Defaults to 10 minutes. */\n  timeoutMs?: number\n  /**\n   * Cap on captured stdout+stderr. Reaching it truncates the captured text\n   * and sets `outputTruncated`; it never kills the process. Defaults to\n   * 16 MiB.\n   */\n  maxOutputBytes?: number\n  /**\n   * Kills the process group when it aborts, the same way the deadline does.\n   * A signal that is already aborted when `runBoundedProcess` is called\n   * prevents the spawn entirely — nothing runs.\n   */\n  signal?: AbortSignal\n}\n\nexport interface BoundedProcessResult {\n  /**\n   * The process actually started. False when nothing ran at all: an\n   * already-aborted signal, `args` with a truthy `shell`, an argument the\n   * platform refuses, or a program that does not exist.\n   *\n   * This is the field to branch on, never the text of `runnerError`. A caller\n   * deciding whether a run produced an outcome at all is asking exactly this,\n   * and asking it here is type-checked rather than held by a wording\n   * convention.\n   */\n  spawned: boolean\n  /**\n   * The `stdin` payload reached the child in full. True whenever the caller\n   * supplied none, because nothing could fail to arrive.\n   *\n   * False means the command ran on partial input, which is a fact about the\n   * answer it produced: a checksum over a truncated payload is a wrong answer,\n   * not a failed run. `spawned` stays true and `exitCode` stays the child's.\n   */\n  stdinDelivered: boolean\n  /**\n   * Exit code. Forced non-zero whenever `killedByTimeout` or\n   * `killedBySignal` is set, because a SIGKILLed child can close with 0 and\n   * a killed run must never read as a pass.\n   */\n  exitCode: number\n  /** Captured stdout, truncated at `maxOutputBytes`. */\n  stdout: string\n  /**\n   * Captured stderr, truncated at `maxOutputBytes`. A runner-side failure is\n   * reported in `runnerError`, not appended here; the caller decides whether\n   * the command's own stderr and the runner's diagnosis belong together.\n   */\n  stderr: string\n  /** Wall-clock duration in ms, measured across the whole call. */\n  wallMs: number\n  /** The wall-clock bound was reached and the process group was killed. */\n  killedByTimeout: boolean\n  /** `signal` aborted and the process group was killed. */\n  killedBySignal: boolean\n  /** `maxOutputBytes` was reached and later output was discarded. */\n  outputTruncated: boolean\n  /**\n   * Human-readable diagnosis of a runner-side failure, in one of three shapes:\n   *\n   *   - `not spawned:` — the caller's own bug, such as `args` with a truthy\n   *     `shell`, or an already-aborted signal. No retry helps.\n   *   - `stdin not delivered:` — the command ran and exited on its own, and\n   *     the stdin payload did not reach it in full because the child closed\n   *     the pipe first.\n   *   - anything else — the process could not be spawned (a missing binary, a\n   *     NUL byte in an argument).\n   *\n   * Absent when the command ran and read everything it was sent, however it\n   * exited. The prefixes are for a person reading a log; code reads `spawned`\n   * and `stdinDelivered`, which say the same things and are type-checked.\n   */\n  runnerError?: string\n}\n\nexport async function runBoundedProcess(input: BoundedProcessInput): Promise<BoundedProcessResult> {\n  const start = Date.now()\n  const maxOutputBytes = input.maxOutputBytes ?? DEFAULT_MAX_OUTPUT_BYTES\n  const timeoutMs = input.timeoutMs ?? DEFAULT_BOUNDED_PROCESS_TIMEOUT_MS\n  const envMode = input.envMode ?? 'merge'\n  const env =\n    input.env === undefined && envMode === 'merge'\n      ? undefined\n      : envMode === 'replace'\n        ? { ...input.env }\n        : { ...process.env, ...input.env }\n\n  if (input.signal?.aborted) {\n    return {\n      spawned: false,\n      stdinDelivered: true,\n      exitCode: ABORT_EXIT_CODE,\n      stdout: '',\n      stderr: '',\n      wallMs: Date.now() - start,\n      killedByTimeout: false,\n      killedBySignal: true,\n      outputTruncated: false,\n      runnerError: 'aborted before spawn: the AbortSignal was already aborted',\n    }\n  }\n\n  // Reported rather than thrown, because this module's contract is that a\n  // call always resolves. A caller bug still reads as a failure and never as\n  // a pass: the exit code is non-zero and `runnerError` names the cause.\n  if (input.args !== undefined && input.shell) {\n    return {\n      spawned: false,\n      stdinDelivered: true,\n      exitCode: SPAWN_FAILURE_EXIT_CODE,\n      stdout: '',\n      stderr: '',\n      wallMs: Date.now() - start,\n      killedByTimeout: false,\n      killedBySignal: false,\n      outputTruncated: false,\n      // Shares `SPAWN_FAILURE_EXIT_CODE` with a program that does not exist,\n      // because both mean nothing ran. The `not spawned:` prefix is the stable\n      // discriminator: it marks the caller's own bug, which no retry fixes.\n      runnerError:\n        'not spawned: `args` passes an argument vector directly to the program, so a shell ' +\n        'cannot interpret it — pass `args` or a truthy `shell`, never both',\n    }\n  }\n\n  return await new Promise<BoundedProcessResult>((resolve) => {\n    const spawnOptions = {\n      cwd: input.cwd,\n      env,\n      // Own process group so a kill reaches the whole tree, not just the\n      // shell — otherwise a runaway grandchild keeps the pipes open and\n      // `close` never fires.\n      detached: process.platform !== 'win32',\n    }\n    // The argv form takes the same bounds, the same group and the same kill\n    // path. Two things differ, and both are stated where they are decided:\n    // no shell reads the command, and executable lookup falls back to a\n    // different default `PATH` when the caller names none (see `envMode`).\n    //\n    // `spawn` validates its arguments and THROWS synchronously — a NUL byte in\n    // `command` or in an `args` entry, or an `args` that is not an array from\n    // an untypechecked caller. A throw inside this executor would reject, and\n    // this module's contract is that a call always resolves. A caller that\n    // grades untrusted text must be able to score that text as failed rather\n    // than die on it.\n    let child: ReturnType<typeof spawn>\n    try {\n      child =\n        input.args === undefined\n          ? spawn(input.command, { ...spawnOptions, shell: input.shell ?? true })\n          : spawn(input.command, input.args, { ...spawnOptions, shell: false })\n    } catch (err) {\n      resolve({\n        spawned: false,\n        stdinDelivered: true,\n        exitCode: SPAWN_FAILURE_EXIT_CODE,\n        stdout: '',\n        stderr: '',\n        wallMs: Date.now() - start,\n        killedByTimeout: false,\n        killedBySignal: false,\n        outputTruncated: false,\n        runnerError: String(err),\n      })\n      return\n    }\n    // A pipe nobody closes is a hang, so stdin is always ended — after the\n    // payload when there is one, immediately when there is not. A write that\n    // fails is recorded and never thrown: the child already ran, and this\n    // module's contract is that a call always resolves.\n    let stdinError: string | undefined\n    const childStdin = child.stdin\n    if (childStdin) {\n      childStdin.on('error', (err) => {\n        // Closing an unused pipe races a command that exits at once, and that\n        // broken pipe says nothing about a caller who sent nothing. Only a\n        // payload the caller supplied can fail to be delivered.\n        if (input.stdin !== undefined) stdinError ??= `stdin not delivered: ${String(err)}`\n      })\n      if (input.stdin === undefined) childStdin.end()\n      else childStdin.end(input.stdin, 'utf8')\n    } else if (input.stdin !== undefined) {\n      stdinError = 'stdin not delivered: the child was spawned with no stdin pipe'\n    }\n\n    let stdout = ''\n    let stderr = ''\n    let outputBytes = 0\n    let outputTruncated = false\n    let killedByTimeout = false\n    let killedBySignal = false\n    let settled = false\n\n    // Bound the in-memory buffer so a runaway process can't OOM the caller.\n    // Once the cap is hit the streams are still drained (the child needs them\n    // read to reach `close`) but the bytes are discarded.\n    function capture(s: string): string {\n      if (outputBytes >= maxOutputBytes) {\n        outputTruncated = true\n        return ''\n      }\n      const room = maxOutputBytes - outputBytes\n      if (s.length > room) {\n        outputBytes = maxOutputBytes\n        outputTruncated = true\n        return s.slice(0, room)\n      }\n      outputBytes += s.length\n      return s\n    }\n\n    // Decoded across chunk boundaries. A stream arrives as Buffers split at\n    // arbitrary byte offsets, so decoding each chunk on its own replaces any\n    // multi-byte character that straddles a boundary with U+FFFD — the output\n    // is then corrupted at every 64 KiB of non-ASCII text. The decoder holds\n    // the partial sequence until its remaining bytes arrive.\n    const stdoutDecoder = new StringDecoder('utf8')\n    const stderrDecoder = new StringDecoder('utf8')\n    const decode = (decoder: StringDecoder, d: unknown): string =>\n      Buffer.isBuffer(d) ? decoder.write(d) : String(d)\n\n    const onStdout = (d: unknown) => {\n      const chunk = capture(decode(stdoutDecoder, d))\n      if (chunk) stdout += chunk\n    }\n    const onStderr = (d: unknown) => {\n      const chunk = capture(decode(stderrDecoder, d))\n      if (chunk) stderr += chunk\n    }\n\n    child.stdout?.on('data', onStdout)\n    child.stderr?.on('data', onStderr)\n\n    let forceResolve: ReturnType<typeof setTimeout> | undefined\n\n    const cleanup = () => {\n      clearTimeout(timeout)\n      if (forceResolve) clearTimeout(forceResolve)\n      input.signal?.removeEventListener('abort', onAbort)\n      child.stdout?.off('data', onStdout)\n      child.stderr?.off('data', onStderr)\n      // The stdin `error` listener stays attached on purpose: a stream that\n      // emits `error` with no listener throws, and a late EPIPE after the run\n      // has settled must not take the process down.\n      child.removeAllListeners('close')\n      child.removeAllListeners('error')\n    }\n\n    // Resolve exactly once. `code` comes from `close`; on a kill path the\n    // exit code is forced non-zero so a SIGKILLed child that reports 0 can\n    // never read as a clean pass downstream.\n    const finish = (outcome: { code: number | null; spawned?: boolean; runnerError?: string }) => {\n      if (settled) return\n      settled = true\n      cleanup()\n      const spawned = outcome.spawned ?? true\n      resolve({\n        spawned,\n        // A program that never started was sent nothing, whatever the write to\n        // its doomed pipe reported.\n        stdinDelivered: spawned ? stdinError === undefined : true,\n        exitCode: exitCodeFor(outcome.code, killedByTimeout, killedBySignal),\n        stdout,\n        stderr,\n        wallMs: Date.now() - start,\n        killedByTimeout,\n        killedBySignal,\n        outputTruncated,\n        runnerError: outcome.runnerError ?? stdinError,\n      })\n    }\n\n    // SIGKILL the whole process group (we spawned detached). On a clean close\n    // this never runs.\n    const killTree = () => {\n      try {\n        if (process.platform !== 'win32' && typeof child.pid === 'number') {\n          process.kill(-child.pid, 'SIGKILL')\n        } else {\n          child.kill('SIGKILL')\n        }\n      } catch (err) {\n        // ESRCH means the group exited between the decision to kill and the\n        // signal, which `close` reports on its own. Anything else means the\n        // group may still be running, which the caller must hear about.\n        if ((err as NodeJS.ErrnoException)?.code !== 'ESRCH') {\n          console.warn('[bounded-process] SIGKILL of the process group failed:', err)\n        }\n      }\n    }\n\n    // Give `close` a brief window to fire (flushing any final output) after\n    // the group dies; if an orphaned grandchild keeps the pipes open, resolve\n    // anyway so the caller can't hang on a runaway. The deadline and an abort\n    // can both fire, so the kill happens at most once and the drain window is\n    // not restarted by the second one.\n    let killed = false\n    const killAndDrain = () => {\n      if (killed) return\n      killed = true\n      killTree()\n      forceResolve = setTimeout(() => finish({ code: null }), KILL_DRAIN_MS)\n    }\n\n    function onAbort() {\n      if (settled) return\n      killedBySignal = true\n      killAndDrain()\n    }\n\n    const timeout = setTimeout(() => {\n      killedByTimeout = true\n      killAndDrain()\n    }, timeoutMs)\n\n    input.signal?.addEventListener('abort', onAbort, { once: true })\n\n    child.on('close', (code) => {\n      if (forceResolve) clearTimeout(forceResolve)\n      finish({ code })\n    })\n    child.on('error', (err) => {\n      if (forceResolve) clearTimeout(forceResolve)\n      // `error` on a child means the program never started — a missing binary,\n      // a refused argument. Nothing ran, whatever the stdin write did.\n      finish({ code: SPAWN_FAILURE_EXIT_CODE, spawned: false, runnerError: String(err) })\n    })\n  })\n}\n\n/**\n * A killed run reports the code the child closed with only when that code is\n * already non-zero; otherwise it reports the conventional code for the kind\n * of kill. A run that closed on its own with no code at all reports 1.\n */\nfunction exitCodeFor(\n  code: number | null,\n  killedByTimeout: boolean,\n  killedBySignal: boolean,\n): number {\n  if (code !== null && code !== 0) return code\n  if (killedByTimeout) return TIMEOUT_EXIT_CODE\n  if (killedBySignal) return ABORT_EXIT_CODE\n  return code ?? UNKNOWN_EXIT_CODE\n}\n"],"mappings":";;;;;;;;;;;;;;;;;;;;;;;;;;;;;;AA+BA,MAAa,qCAAqC;;AAGlD,MAAM,2BAA2B;;;;;;AAOjC,MAAM,gBAAgB;;AAGtB,MAAM,oBAAoB;;AAG1B,MAAM,kBAAkB;;AAGxB,MAAM,0BAA0B;;AAGhC,MAAM,oBAAoB;AA2K1B,eAAsB,kBAAkB,OAA2D;CACjG,MAAM,QAAQ,KAAK,IAAI;CACvB,MAAM,iBAAiB,MAAM,kBAAkB;CAC/C,MAAM,YAAY,MAAM,aAAA;CACxB,MAAM,UAAU,MAAM,WAAW;CACjC,MAAM,MACJ,MAAM,QAAQ,KAAA,KAAa,YAAY,UACnC,KAAA,IACA,YAAY,YACV,EAAE,GAAG,MAAM,IAAI,IACf;EAAE,GAAG,QAAQ;EAAK,GAAG,MAAM;CAAI;CAEvC,IAAI,MAAM,QAAQ,SAChB,OAAO;EACL,SAAS;EACT,gBAAgB;EAChB,UAAU;EACV,QAAQ;EACR,QAAQ;EACR,QAAQ,KAAK,IAAI,IAAI;EACrB,iBAAiB;EACjB,gBAAgB;EAChB,iBAAiB;EACjB,aAAa;CACf;CAMF,IAAI,MAAM,SAAS,KAAA,KAAa,MAAM,OACpC,OAAO;EACL,SAAS;EACT,gBAAgB;EAChB,UAAU;EACV,QAAQ;EACR,QAAQ;EACR,QAAQ,KAAK,IAAI,IAAI;EACrB,iBAAiB;EACjB,gBAAgB;EAChB,iBAAiB;EAIjB,aACE;CAEJ;CAGF,OAAO,MAAM,IAAI,SAA+B,YAAY;EAC1D,MAAM,eAAe;GACnB,KAAK,MAAM;GACX;GAIA,UAAU,QAAQ,aAAa;EACjC;EAYA,IAAI;EACJ,IAAI;GACF,QACE,MAAM,SAAS,KAAA,IACX,MAAM,MAAM,SAAS;IAAE,GAAG;IAAc,OAAO,MAAM,SAAS;GAAK,CAAC,IACpE,MAAM,MAAM,SAAS,MAAM,MAAM;IAAE,GAAG;IAAc,OAAO;GAAM,CAAC;EAC1E,SAAS,KAAK;GACZ,QAAQ;IACN,SAAS;IACT,gBAAgB;IAChB,UAAU;IACV,QAAQ;IACR,QAAQ;IACR,QAAQ,KAAK,IAAI,IAAI;IACrB,iBAAiB;IACjB,gBAAgB;IAChB,iBAAiB;IACjB,aAAa,OAAO,GAAG;GACzB,CAAC;GACD;EACF;EAKA,IAAI;EACJ,MAAM,aAAa,MAAM;EACzB,IAAI,YAAY;GACd,WAAW,GAAG,UAAU,QAAQ;IAI9B,IAAI,MAAM,UAAU,KAAA,GAAW,eAAe,wBAAwB,OAAO,GAAG;GAClF,CAAC;GACD,IAAI,MAAM,UAAU,KAAA,GAAW,WAAW,IAAI;QACzC,WAAW,IAAI,MAAM,OAAO,MAAM;EACzC,OAAO,IAAI,MAAM,UAAU,KAAA,GACzB,aAAa;EAGf,IAAI,SAAS;EACb,IAAI,SAAS;EACb,IAAI,cAAc;EAClB,IAAI,kBAAkB;EACtB,IAAI,kBAAkB;EACtB,IAAI,iBAAiB;EACrB,IAAI,UAAU;EAKd,SAAS,QAAQ,GAAmB;GAClC,IAAI,eAAe,gBAAgB;IACjC,kBAAkB;IAClB,OAAO;GACT;GACA,MAAM,OAAO,iBAAiB;GAC9B,IAAI,EAAE,SAAS,MAAM;IACnB,cAAc;IACd,kBAAkB;IAClB,OAAO,EAAE,MAAM,GAAG,IAAI;GACxB;GACA,eAAe,EAAE;GACjB,OAAO;EACT;EAOA,MAAM,gBAAgB,IAAI,cAAc,MAAM;EAC9C,MAAM,gBAAgB,IAAI,cAAc,MAAM;EAC9C,MAAM,UAAU,SAAwB,MACtC,OAAO,SAAS,CAAC,IAAI,QAAQ,MAAM,CAAC,IAAI,OAAO,CAAC;EAElD,MAAM,YAAY,MAAe;GAC/B,MAAM,QAAQ,QAAQ,OAAO,eAAe,CAAC,CAAC;GAC9C,IAAI,OAAO,UAAU;EACvB;EACA,MAAM,YAAY,MAAe;GAC/B,MAAM,QAAQ,QAAQ,OAAO,eAAe,CAAC,CAAC;GAC9C,IAAI,OAAO,UAAU;EACvB;EAEA,MAAM,QAAQ,GAAG,QAAQ,QAAQ;EACjC,MAAM,QAAQ,GAAG,QAAQ,QAAQ;EAEjC,IAAI;EAEJ,MAAM,gBAAgB;GACpB,aAAa,OAAO;GACpB,IAAI,cAAc,aAAa,YAAY;GAC3C,MAAM,QAAQ,oBAAoB,SAAS,OAAO;GAClD,MAAM,QAAQ,IAAI,QAAQ,QAAQ;GAClC,MAAM,QAAQ,IAAI,QAAQ,QAAQ;GAIlC,MAAM,mBAAmB,OAAO;GAChC,MAAM,mBAAmB,OAAO;EAClC;EAKA,MAAM,UAAU,YAA8E;GAC5F,IAAI,SAAS;GACb,UAAU;GACV,QAAQ;GACR,MAAM,UAAU,QAAQ,WAAW;GACnC,QAAQ;IACN;IAGA,gBAAgB,UAAU,eAAe,KAAA,IAAY;IACrD,UAAU,YAAY,QAAQ,MAAM,iBAAiB,cAAc;IACnE;IACA;IACA,QAAQ,KAAK,IAAI,IAAI;IACrB;IACA;IACA;IACA,aAAa,QAAQ,eAAe;GACtC,CAAC;EACH;EAIA,MAAM,iBAAiB;GACrB,IAAI;IACF,IAAI,QAAQ,aAAa,WAAW,OAAO,MAAM,QAAQ,UACvD,QAAQ,KAAK,CAAC,MAAM,KAAK,SAAS;SAElC,MAAM,KAAK,SAAS;GAExB,SAAS,KAAK;IAIZ,IAAK,KAA+B,SAAS,SAC3C,QAAQ,KAAK,0DAA0D,GAAG;GAE9E;EACF;EAOA,IAAI,SAAS;EACb,MAAM,qBAAqB;GACzB,IAAI,QAAQ;GACZ,SAAS;GACT,SAAS;GACT,eAAe,iBAAiB,OAAO,EAAE,MAAM,KAAK,CAAC,GAAG,aAAa;EACvE;EAEA,SAAS,UAAU;GACjB,IAAI,SAAS;GACb,iBAAiB;GACjB,aAAa;EACf;EAEA,MAAM,UAAU,iBAAiB;GAC/B,kBAAkB;GAClB,aAAa;EACf,GAAG,SAAS;EAEZ,MAAM,QAAQ,iBAAiB,SAAS,SAAS,EAAE,MAAM,KAAK,CAAC;EAE/D,MAAM,GAAG,UAAU,SAAS;GAC1B,IAAI,cAAc,aAAa,YAAY;GAC3C,OAAO,EAAE,KAAK,CAAC;EACjB,CAAC;EACD,MAAM,GAAG,UAAU,QAAQ;GACzB,IAAI,cAAc,aAAa,YAAY;GAG3C,OAAO;IAAE,MAAM;IAAyB,SAAS;IAAO,aAAa,OAAO,GAAG;GAAE,CAAC;EACpF,CAAC;CACH,CAAC;AACH;;;;;;AAOA,SAAS,YACP,MACA,iBACA,gBACQ;CACR,IAAI,SAAS,QAAQ,SAAS,GAAG,OAAO;CACxC,IAAI,iBAAiB,OAAO;CAC5B,IAAI,gBAAgB,OAAO;CAC3B,OAAO,QAAQ;AACjB"}

package/dist/builder-eval/index.d.ts

@@ -1,8 +1,7 @@
-import { f as Run, o as FailureClass } from "../schema-D6EVKzTB.js";
-import { s as TraceStore } from "../store-B3nWRy3B.js";
-import { o as TraceEmitter } from "../emitter-DABYYTcH.js";
-import { n as SandboxDriver, r as SandboxHarnessResult, t as HarnessConfig } from "../sandbox-harness-BlSOu4LX.js";
-//#region src/test-graded-scenario.d.ts
+import { f as Run, o as FailureClass } from "../schema-CkIvnmLU.js";
+import { s as TraceStore } from "../store-CMs5wGXP.js";
+import { o as TraceEmitter } from "../emitter-mAXC5PSN.js";
+import { n as SandboxDriver, r as SandboxHarnessResult, t as HarnessConfig } from "../sandbox-harness-B3Kx1fOA.js";
 interface TestGradedScenario {
   id: string;
   description?: string;
@@ -22,8 +21,6 @@ interface TestGradedRunResult {
   score: number;
   failureClass?: FailureClass;
 }
-//#endregion
-//#region src/builder-eval/builder-session.d.ts
 export interface BuilderSessionInit {
   projectId: string;
   chatId?: string;
@@ -112,8 +109,6 @@ export declare function resumeBuilderSession(store: TraceStore, projectId: strin
   lastBuildRun?: Run;
   lastAppRuntimeRuns: Run[];
 }>;
-//#endregion
-//#region src/builder-eval/three-layer-eval.d.ts
 export type ProjectKind = 'full' | 'scaffold-only';
 export interface ThreeLayerProjectReport {
   projectId: string;
@@ -144,8 +139,6 @@ export interface ThreeLayerProjectReport {
 export declare function scoreProject(store: TraceStore, projectId: string): Promise<ThreeLayerProjectReport>;
 /** Aggregate scoring across every project in a corpus. */
 export declare function scoreAllProjects(store: TraceStore): Promise<ThreeLayerProjectReport[]>;
-//#endregion
-//#region src/builder-eval/correlation.d.ts
 export interface LayerCorrelation {
   n: number;
   pearson: number;
@@ -160,8 +153,6 @@ export interface CorrelationReport {
   completeProjects: number;
 }
 export declare function correlateLayers(reports: ThreeLayerProjectReport[]): CorrelationReport;
-//#endregion
-//#region src/builder-eval/project-registry.d.ts
 export interface ProjectSummary {
   projectId: string;
   chatCount: number;
@@ -199,5 +190,5 @@ export declare class ProjectRegistry {
   projectTimeline(projectId: string): Promise<ProjectTimelineEntry[]>;
   projectChats(projectId: string): Promise<ChatSummary[]>;
 }
-//#endregion
+export {};
 //# sourceMappingURL=index.d.ts.map

package/dist/builder-eval/index.d.ts.map

@@ -1 +1 @@
-{"version":3,"file":"index.d.ts","names":[],"sources":["../../src/test-graded-scenario.ts","../../src/builder-eval/builder-session.ts","../../src/builder-eval/three-layer-eval.ts","../../src/builder-eval/correlation.ts","../../src/builder-eval/project-registry.ts"],"mappings":";;;;;UAmBiB;EACf;EACA;EACA,SAAS;;EAET;;EAEA;;EAEA,OAAO;;UAUQ;EACf;EACA,UAAU;EACV,SAAS;EACT;EACA;EACA,eAAe;;;;iBCjBA;EACf;EACA;;EAEA,OAAO;;iBAGQ;EACf,SAAS;EACT,SAAS;;EAET;;iBAGe;EACf,UAAU;;EAEV,SAAS;;qBAGE;UACH;UACA;WACC;WACA;UACD;UACA;UACA;EAER,YAAY,OAAO,YAAY,MAAM,oBAAoB,SAAS;;EAS5D,UAAU,sBAAwC;;MAYpD,WAAW;;;;;EAST,KAAK,SAAS,cAAc;IAAU;IAAe,QAAQ;;;;;;;;EA2B7D,eAAe,SAAS,wBAAwB,QAAQ;;;EAqBxD,gBAAgB,eAAe,qBAAqB;;EAcpD,QAAQ;IAAW;IAAe;IAAgB;MAAmB;;;;;;;;EAerE,gBAAgB,qBAAqB,QAAQ;;;;;;EAsB7C,iBAAiB;IACrB;IACA;IACA;IACA;MACE;MAqBA;MAGA;;;;;;;wBAUgB,qBACpB,OAAO,YACP,oBACC;EACD;EACA,UAAU;EACV,iBAAiB;EACjB,eAAe;EACf,oBAAoB;;;;YC5MV;iBAEK;EACf;;;;;;;EAOA,MAAM;EACN;;EAEA;EACA;;EAEA;EACA;;EAEA;EACA;;;;;;EAMA;;wBAGoB,aACpB,OAAO,YACP,oBACC,QAAQ;;wBAuCW,iBAAiB,OAAO,aAAa,QAAQ;;;iBC/ElD;EACf;EACA;EACA;;iBAGe;;EAEf,cAAc;EACd,gBAAgB;EAChB,iBAAiB;;EAEjB;;wBAGc,gBAAgB,SAAS,4BAA4B;;;iBCpBpD;EACf;EACA;EACA;EACA;EACA;EACA;EACA;IAAkB;IAAe;;;iBAGlB;EACf;EACA;EACA;EACA;EACA;EACA,QAAQ;EACR,UAAU;;EAEV;EACA;EACA;EACA;;iBAGe;EACf,KAAK;EACL;;qBAGW;UACS;EAApB,YAAoB,OAAO;EAErB,gBAAgB,QAAQ;EAgCxB,gBAAgB,oBAAoB,QAAQ;EAgB5C,aAAa,oBAAoB,QAAQ"}
+{"version":3,"file":"index.d.ts","names":[],"sources":["../../src/test-graded-scenario.ts","../../src/builder-eval/builder-session.ts","../../src/builder-eval/three-layer-eval.ts","../../src/builder-eval/correlation.ts","../../src/builder-eval/project-registry.ts"],"mappings":";;;;UAmBiB;EACf;EACA;EACA,SAAS;;EAET;;EAEA;;EAEA,OAAO;;UAUQ;EACf;EACA,UAAU;EACV,SAAS;EACT;EACA;EACA,eAAe;;iBCjBA;EACf;EACA;;EAEA,OAAO;;iBAGQ;EACf,SAAS;EACT,SAAS;;EAET;;iBAGe;EACf,UAAU;;EAEV,SAAS;;qBAGE;UACH;UACA;WACC;WACA;UACD;UACA;UACA;EAEI,YAAA,OAAO,YAAY,MAAM,oBAAoB,SAAS;;EAS5D,UAAU,sBAAwC;;MAYpD,WAAW;;;;;EAST,KAAK,SAAS,cAAc;IAAU;IAAe,QAAQ;;;;;;;;EA2B7D,eAAe,SAAS,wBAAwB,QAAQ;;;EAqBxD,gBAAgB,eAAe,qBAAqB;;EAcpD,QAAQ;IAAW;IAAe;IAAgB;MAAmB;;;;;;;;EAerE,gBAAgB,qBAAqB,QAAQ;;;;;;EAsB7C,iBAAiB;IACrB;IACA;IACA;IACA;MACE;MAqBA;MAGA;;;;;;;wBAUgB,qBACpB,OAAO,YACP,oBACC;EACD;EACA,UAAU;EACV,iBAAiB;EACjB,eAAe;EACf,oBAAoB;;YC5MV;iBAEK;EACf;;;;;;;EAOA,MAAM;EACN;;EAEA;EACA;;EAEA;EACA;;EAEA;EACA;;;;;;EAMA;;wBAGoB,aACpB,OAAO,YACP,oBACC,QAAQ;;wBAuCW,iBAAiB,OAAO,aAAa,QAAQ;iBC/ElD;EACf;EACA;EACA;;iBAGe;;EAEf,cAAc;EACd,gBAAgB;EAChB,iBAAiB;;EAEjB;;wBAGc,gBAAgB,SAAS,4BAA4B;iBCpBpD;EACf;EACA;EACA;EACA;EACA;EACA;EACA;IAAkB;IAAe;;;iBAGlB;EACf;EACA;EACA;EACA;EACA;EACA,QAAQ;EACR,UAAU;;EAEV;EACA;EACA;EACA;;iBAGe;EACf,KAAK;EACL;;qBAGW;UACS;EAAA,YAAA,OAAO;EAErB,gBAAgB,QAAQ;EAgCxB,gBAAgB,oBAAoB,QAAQ;EAgB5C,aAAa,oBAAoB,QAAQ"}

package/dist/builder-eval/index.js

@@ -1,9 +1,9 @@
-import { a as spearmanR, r as pearsonR } from "../descriptive-CVNy48rR.js";
-import { i as judgeSpans } from "../query-B_RZ36an.js";
-import { t as newRecordId } from "../record-id-DUgsK5qp.js";
-import { t as TraceEmitter } from "../emitter-DbMXUIVP.js";
-import { n as runBoundedProcess } from "../bounded-process-Cj9YjXbE.js";
-//#region src/sandbox-harness.ts
+import { a as spearmanR, r as pearsonR } from "../descriptive-cPwITVfy.js";
+import { i as judgeSpans } from "../query-CGXzO9d4.js";
+import { t as newRecordId } from "../record-id-CWZXFh58.js";
+import { t as TraceEmitter } from "../emitter-DhQ65mza.js";
+import { n as runBoundedProcess } from "../bounded-process-BZrWXIKW.js";
+// src/sandbox-harness.ts
 /**
 * SandboxHarness — executes a scenario in an isolated environment and
 * emits a rich SandboxSpan into the trace.
@@ -127,8 +127,7 @@ var SandboxHarness = class {
 		return result;
 	}
 };
-//#endregion
-//#region src/test-graded-scenario.ts
+// src/test-graded-scenario.ts
 async function runTestGradedScenario(scenario, store, options = {}) {
 	const emitter = new TraceEmitter(store);
 	await emitter.startRun({
@@ -169,8 +168,7 @@ function reasonForFailure(result) {
 	}
 	return "no test command";
 }
-//#endregion
-//#region src/builder-eval/builder-session.ts
+// src/builder-eval/builder-session.ts
 /**
 * BuilderSession — ties a builder-of-builders workflow together.
 *
@@ -361,8 +359,7 @@ async function resumeBuilderSession(store, projectId) {
 		lastAppRuntimeRuns: appRuntimeRuns
 	};
 }
-//#endregion
-//#region src/builder-eval/correlation.ts
+// src/builder-eval/correlation.ts
 /**
 * Meta-eval correlation — the highest-leverage signal in the framework.
 *
@@ -406,8 +403,7 @@ function pairwise(reports, a, b) {
 		spearman: spearmanR(xs, ys)
 	};
 }
-//#endregion
-//#region src/builder-eval/project-registry.ts
+// src/builder-eval/project-registry.ts
 var ProjectRegistry = class {
 	store;
 	constructor(store) {
@@ -485,8 +481,7 @@ var ProjectRegistry = class {
 		return out;
 	}
 };
-//#endregion
-//#region src/builder-eval/three-layer-eval.ts
+// src/builder-eval/three-layer-eval.ts
 /**
 * Three-layer evaluation — the canonical scoring breakdown for
 * builder-of-builders workflows.
@@ -552,7 +547,6 @@ async function extractMetaScore(store, builderRunId) {
 	if (meta.score >= 0 && meta.score <= 10) return meta.score / 10;
 	return null;
 }
-//#endregion
 export { BuilderSession, ProjectRegistry, correlateLayers, resumeBuilderSession, scoreAllProjects, scoreProject };
 
 //# sourceMappingURL=index.js.map

package/dist/builder-eval/index.js.map

@@ -1 +1 @@
-{"version":3,"file":"index.js","names":[],"sources":["../../src/sandbox-harness.ts","../../src/test-graded-scenario.ts","../../src/builder-eval/builder-session.ts","../../src/builder-eval/correlation.ts","../../src/builder-eval/project-registry.ts","../../src/builder-eval/three-layer-eval.ts"],"sourcesContent":["/**\n * SandboxHarness — executes a scenario in an isolated environment and\n * emits a rich SandboxSpan into the trace.\n *\n * Two built-in drivers:\n *   - `SubprocessSandboxDriver` — runs the phase in a local cwd with env\n *     vars, through `runBoundedProcess` (deadline, process-group kill,\n *     output cap). Fast, no dependencies, fine for unit tests and CI gates.\n *   - `DockerSandboxDriver` — lifted from tangle-router's sandbox path;\n *     shells out to `docker run`. Stronger isolation, slower startup.\n *\n * Consumers implement `SandboxDriver` for custom backends (Firecracker,\n * Cloudflare sandbox product, etc.). The harness doesn't care which.\n */\n\nimport { runBoundedProcess } from './bounded-process'\n\nimport type { TraceEmitter } from './trace/emitter'\nimport type { SandboxSpan } from './trace/schema'\n\nexport interface HarnessConfig {\n  /** Setup command (e.g. \"pnpm install\"). Non-zero exit fails the run. */\n  setupCommand?: string\n  /** Run command (e.g. \"pnpm build\"). */\n  runCommand?: string\n  /** Test command (e.g. \"pnpm test --run\"). Drives the test count + pass count. */\n  testCommand?: string\n  /** Absolute cwd for the subprocess driver. Ignored by docker driver. */\n  cwd?: string\n  /** Max wall-clock per phase in ms. Default 10 minutes. */\n  timeoutMs?: number\n  /**\n   * Cap on captured stdout+stderr bytes per phase. A runaway process can\n   * otherwise grow the in-memory buffer without bound. Once hit, further\n   * output is dropped and `outputTruncated` is set. Default 16 MiB.\n   */\n  maxOutputBytes?: number\n  /** Image for the docker driver. */\n  image?: string\n  /** Extra env vars (validated; shell-escaped). */\n  env?: Record<string, string>\n  /** Parser for the test output — maps stdout/stderr/exit code → pass count. */\n  testParser?: TestOutputParser\n}\n\nexport interface TestOutputParser {\n  id: string\n  parse(\n    stdout: string,\n    stderr: string,\n    exitCode: number,\n  ): { testsTotal: number; testsPassed: number } | undefined\n}\n\nexport interface SandboxResult {\n  phase: 'setup' | 'run' | 'test'\n  exitCode: number\n  stdout: string\n  stderr: string\n  wallMs: number\n  testsTotal?: number\n  testsPassed?: number\n  /**\n   * True when the process was killed because it exceeded `timeoutMs`. A\n   * SIGKILLed child can still close with exit code 0; callers MUST treat\n   * a timed-out phase as a hard failure regardless of `exitCode`, never\n   * as a pass. `undefined`/`false` means the process completed on its own.\n   */\n  killedByTimeout?: boolean\n  /**\n   * True when captured stdout/stderr hit `maxOutputBytes` and further\n   * output was dropped. The result is still returned (the process was\n   * not killed for this), but downstream parsers see truncated text.\n   */\n  outputTruncated?: boolean\n}\n\nexport interface SandboxDriver {\n  id: string\n  exec(\n    phase: SandboxResult['phase'],\n    command: string,\n    config: HarnessConfig,\n  ): Promise<SandboxResult>\n}\n\n// ── Parsers ──────────────────────────────────────────────────────────\n\n/** Vitest default summary line: \"Tests  X passed | Y failed\". */\nexport const vitestTestParser: TestOutputParser = {\n  id: 'vitest',\n  parse(stdout) {\n    const m = stdout.match(/Tests\\s+(\\d+)\\s+(passed|failed)(?:\\s*\\|\\s*(\\d+)\\s+(passed|failed))?/i)\n    if (!m) return undefined\n    let passed = 0\n    let failed = 0\n    const a = parseInt(m[1]!, 10)\n    const aLabel = m[2]!.toLowerCase()\n    if (aLabel === 'passed') passed += a\n    else failed += a\n    if (m[3] && m[4]) {\n      const b = parseInt(m[3], 10)\n      if (m[4].toLowerCase() === 'passed') passed += b\n      else failed += b\n    }\n    return { testsTotal: passed + failed, testsPassed: passed }\n  },\n}\n\n/** Pytest default: \"collected N items\" + \" X passed, Y failed\". */\nexport const pytestTestParser: TestOutputParser = {\n  id: 'pytest',\n  parse(stdout) {\n    const total = stdout.match(/collected\\s+(\\d+)\\s+items?/i)\n    const passed = stdout.match(/(\\d+)\\s+passed/)\n    if (!total || !passed) return undefined\n    return { testsTotal: parseInt(total[1]!, 10), testsPassed: parseInt(passed[1]!, 10) }\n  },\n}\n\n/** Jest: \"Tests: X passed, Y total\" (and optional failed). */\nexport const jestTestParser: TestOutputParser = {\n  id: 'jest',\n  parse(stdout) {\n    const m = stdout.match(/Tests:\\s+(?:(\\d+)\\s+failed[^,]*,\\s*)?(\\d+)\\s+passed,\\s+(\\d+)\\s+total/i)\n    if (!m) return undefined\n    return { testsTotal: parseInt(m[3]!, 10), testsPassed: parseInt(m[2]!, 10) }\n  },\n}\n\n/** Composite parser — tries a list of parsers in order. */\nexport function composeParsers(...parsers: TestOutputParser[]): TestOutputParser {\n  return {\n    id: parsers.map((p) => p.id).join('|'),\n    parse(stdout, stderr, exitCode) {\n      for (const p of parsers) {\n        const res = p.parse(stdout, stderr, exitCode)\n        if (res) return res\n      }\n      return undefined\n    },\n  }\n}\n\n// ── Drivers ──────────────────────────────────────────────────────────\n\ninterface SubprocessSandboxDriverOptions {\n  /**\n   * Default cwd for all `exec` calls. Used when the per-call `HarnessConfig`\n   * does not set its own `cwd`. Lets callers bind the driver to a working\n   * directory once instead of spreading cwd into every harness config —\n   * useful when the harness config is constructed far from the call site\n   * (e.g. starter-foundry's promoter passes a static HarnessConfig per\n   * family taxonomy but needs a per-run composed-scaffold cwd).\n   */\n  cwd?: string\n  /**\n   * Default env merged into every `exec` call's env (per-call `HarnessConfig.env`\n   * still wins on key collision). Same ergonomic rationale as `cwd` above.\n   */\n  env?: Record<string, string>\n}\n\nexport class SubprocessSandboxDriver implements SandboxDriver {\n  id = 'subprocess'\n  private defaultCwd?: string\n  private defaultEnv?: Record<string, string>\n\n  constructor(options: SubprocessSandboxDriverOptions = {}) {\n    this.defaultCwd = options.cwd\n    this.defaultEnv = options.env\n  }\n\n  async exec(\n    phase: SandboxResult['phase'],\n    command: string,\n    config: HarnessConfig,\n  ): Promise<SandboxResult> {\n    // Per-call config wins; fall back to constructor defaults. Honoring\n    // the constructor `cwd` keeps the subprocess from inheriting Node's\n    // cwd when only the constructor arg is supplied.\n    const effectiveCwd = config.cwd ?? this.defaultCwd\n    // The parent environment is merged here, so the runner is handed the\n    // complete environment and passes it through unchanged.\n    const effectiveEnv = { ...process.env, ...(this.defaultEnv ?? {}), ...(config.env ?? {}) }\n    const result = await runBoundedProcess({\n      command,\n      shell: true,\n      cwd: effectiveCwd,\n      env: effectiveEnv,\n      envMode: 'replace',\n      timeoutMs: config.timeoutMs ?? 10 * 60_000,\n      maxOutputBytes: config.maxOutputBytes ?? 16 * 1024 * 1024,\n    })\n    // A timed-out phase never reaches the parser: the output of a killed\n    // process is a fragment, and a fragment that happens to parse would\n    // report a pass for a run that never finished.\n    const parsed =\n      !result.killedByTimeout && phase === 'test' && config.testParser\n        ? config.testParser.parse(result.stdout, result.stderr, result.exitCode)\n        : undefined\n    return {\n      phase,\n      exitCode: result.exitCode,\n      stdout: result.stdout,\n      stderr: result.runnerError ? result.stderr + result.runnerError : result.stderr,\n      wallMs: result.wallMs,\n      testsTotal: parsed?.testsTotal,\n      testsPassed: parsed?.testsPassed,\n      killedByTimeout: result.killedByTimeout || undefined,\n      outputTruncated: result.outputTruncated || undefined,\n    }\n  }\n}\n\n\n\n\n\n// ── Harness orchestration ────────────────────────────────────────────\n\nexport interface SandboxHarnessResult {\n  passed: boolean\n  setup?: SandboxResult\n  run?: SandboxResult\n  test?: SandboxResult\n  totalWallMs: number\n  /** Final score — 0 when no tests; otherwise testsPassed/testsTotal. */\n  score: number\n}\n\nexport class SandboxHarness {\n  private driver: SandboxDriver\n  constructor(driver: SandboxDriver = new SubprocessSandboxDriver()) {\n    this.driver = driver\n  }\n\n  async run(config: HarnessConfig, emitter: TraceEmitter): Promise<SandboxHarnessResult> {\n    const handle = await emitter.sandbox({\n      name: `sandbox(${this.driver.id})`,\n      image: config.image,\n      command: [config.setupCommand, config.runCommand, config.testCommand]\n        .filter(Boolean)\n        .join(' && '),\n    })\n    const result: SandboxHarnessResult = { passed: false, totalWallMs: 0, score: 0 }\n    try {\n      if (config.setupCommand) {\n        result.setup = await this.driver.exec('setup', config.setupCommand, config)\n        result.totalWallMs += result.setup.wallMs\n        if (result.setup.killedByTimeout || result.setup.exitCode !== 0) {\n          const reason = result.setup.killedByTimeout\n            ? `setup timed out after ${result.setup.wallMs}ms`\n            : `setup failed (exit ${result.setup.exitCode})`\n          await handle.fail(reason, {\n            exitCode: result.setup.exitCode,\n            wallMs: result.totalWallMs,\n          } as Partial<SandboxSpan>)\n          return result\n        }\n      }\n      if (config.runCommand) {\n        result.run = await this.driver.exec('run', config.runCommand, config)\n        result.totalWallMs += result.run.wallMs\n        if (result.run.killedByTimeout || result.run.exitCode !== 0) {\n          const reason = result.run.killedByTimeout\n            ? `run timed out after ${result.run.wallMs}ms`\n            : `run failed (exit ${result.run.exitCode})`\n          await handle.fail(reason, {\n            exitCode: result.run.exitCode,\n            wallMs: result.totalWallMs,\n          } as Partial<SandboxSpan>)\n          return result\n        }\n      }\n      if (config.testCommand) {\n        result.test = await this.driver.exec('test', config.testCommand, config)\n        result.totalWallMs += result.test.wallMs\n        // A timed-out test phase is a hard fail regardless of the exit code\n        // a SIGKILLed child happens to report — never let it score as a pass.\n        const passed = !result.test.killedByTimeout && result.test.exitCode === 0\n        result.passed = passed\n        if (result.test.testsTotal !== undefined && result.test.testsTotal > 0) {\n          result.score = (result.test.testsPassed ?? 0) / result.test.testsTotal\n        } else {\n          result.score = passed ? 1 : 0\n        }\n        await handle.end({\n          exitCode: result.test.exitCode,\n          testsTotal: result.test.testsTotal,\n          testsPassed: result.test.testsPassed,\n          wallMs: result.totalWallMs,\n          status: passed ? 'ok' : 'error',\n        } as Partial<SandboxSpan>)\n      } else {\n        result.passed = true\n        result.score = 1\n        await handle.end({ wallMs: result.totalWallMs } as Partial<SandboxSpan>)\n      }\n    } catch (err) {\n      await handle.fail(err instanceof Error ? err : String(err))\n      throw err\n    }\n    return result\n  }\n}\n","/**\n * TestGradedScenario — a scenario whose score comes from a test suite.\n *\n * This is the SWE-bench pattern generalized. The scenario ships:\n *   - fixture data (setup instructions)\n *   - a test command the harness runs\n *   - optional assertion overrides\n *\n * The runner emits a run, delegates to SandboxHarness, records the\n * outcome, and returns a structured verdict. Consumers bind their own\n * agent execution to this contract.\n */\n\nimport type { HarnessConfig, SandboxDriver, SandboxHarnessResult } from './sandbox-harness'\nimport { SandboxHarness } from './sandbox-harness'\nimport { TraceEmitter } from './trace/emitter'\nimport type { FailureClass, Run } from './trace/schema'\nimport type { TraceStore } from './trace/store'\n\nexport interface TestGradedScenario {\n  id: string\n  description?: string\n  harness: HarnessConfig\n  /** Optional pass threshold in 0..1 (default 1.0 = all tests must pass). */\n  passThreshold?: number\n  /** Provenance for dataset tracking. */\n  datasetVersion?: string\n  /** Free-form tags (difficulty, category, etc.). */\n  tags?: Record<string, string>\n}\n\ninterface TestGradedRunOptions {\n  variantId?: string\n  driver?: SandboxDriver\n  /** Metadata recorded on the Run (codeSha, promptSha, modelFingerprint, seed). */\n  provenance?: Pick<Run, 'codeSha' | 'promptSha' | 'modelFingerprint' | 'seed' | 'envFingerprint'>\n}\n\nexport interface TestGradedRunResult {\n  runId: string\n  scenario: TestGradedScenario\n  harness: SandboxHarnessResult\n  pass: boolean\n  score: number\n  failureClass?: FailureClass\n}\n\nexport async function runTestGradedScenario(\n  scenario: TestGradedScenario,\n  store: TraceStore,\n  options: TestGradedRunOptions = {},\n): Promise<TestGradedRunResult> {\n  const emitter = new TraceEmitter(store)\n  await emitter.startRun({\n    scenarioId: scenario.id,\n    variantId: options.variantId,\n    datasetVersion: scenario.datasetVersion,\n    tags: scenario.tags,\n    ...options.provenance,\n  })\n  const harness = new SandboxHarness(options.driver)\n  const result = await harness.run(scenario.harness, emitter)\n  const threshold = scenario.passThreshold ?? 1.0\n  const pass = result.passed && result.score >= threshold\n  const setupFailed = result.setup !== undefined && result.setup.exitCode !== 0\n  const runFailed = result.run !== undefined && result.run.exitCode !== 0\n  const testFailed = result.test !== undefined && result.test.exitCode !== 0\n  const failureClass: FailureClass | undefined = pass\n    ? 'success'\n    : setupFailed || runFailed\n      ? 'sandbox_failure'\n      : testFailed\n        ? 'format_drift'\n        : 'unknown'\n  await emitter.endRun({\n    pass,\n    score: result.score,\n    failureClass,\n    notes: pass ? undefined : reasonForFailure(result),\n  })\n  return {\n    runId: emitter.runId,\n    scenario,\n    harness: result,\n    pass,\n    score: result.score,\n    failureClass,\n  }\n}\n\nfunction reasonForFailure(result: SandboxHarnessResult): string {\n  if (result.setup && result.setup.exitCode !== 0)\n    return `setup failed: exit ${result.setup.exitCode}`\n  if (result.run && result.run.exitCode !== 0) return `run failed: exit ${result.run.exitCode}`\n  if (result.test) {\n    if (result.test.testsTotal !== undefined) {\n      return `tests: ${result.test.testsPassed ?? 0}/${result.test.testsTotal}`\n    }\n    return `test exit ${result.test.exitCode}`\n  }\n  return 'no test command'\n}\n","/**\n * BuilderSession — ties a builder-of-builders workflow together.\n *\n * Models agent-builder's shape: Project → Chat → Edit → Ship → App →\n * AppAgent. Each layer is a Run (linked via parentRunId). The\n * framework-enforced invariants:\n *\n *   - One Project → many Chats; chatId scopes runs within a project.\n *   - One Chat = one builder Run with `layer='builder'`.\n *   - One Ship = one child Run with `layer='app-build'` + SandboxHarness.\n *   - One AppScenario = one grandchild Run with `layer='app-runtime'`.\n *\n * Consumers obtain a BuilderSession, call `startChat`, drive the\n * builder agent (emitting spans), and call `ship` / `runAppScenario`\n * as the workflow progresses. The session reconstructs itself from\n * trace data via `resume(store, projectId)`.\n */\n\nimport { newRecordId } from '../record-id'\nimport type { HarnessConfig, SandboxDriver, SandboxHarnessResult } from '../sandbox-harness'\nimport { SandboxHarness } from '../sandbox-harness'\nimport type { TestGradedRunResult, TestGradedScenario } from '../test-graded-scenario'\nimport { runTestGradedScenario } from '../test-graded-scenario'\nimport { TraceEmitter } from '../trace/emitter'\nimport type { Run } from '../trace/schema'\nimport type { TraceStore } from '../trace/store'\n\nexport interface BuilderSessionInit {\n  projectId: string\n  chatId?: string\n  /** Free-form: user's task description, project name, etc. Stored on the builder Run. */\n  tags?: Record<string, string>\n}\n\nexport interface ShipOptions {\n  harness: HarnessConfig\n  driver?: SandboxDriver\n  /** scenarioId of this app-build run. Defaults to `${projectId}/build`. */\n  scenarioId?: string\n}\n\nexport interface RunAppScenarioOptions {\n  scenario: TestGradedScenario\n  /** Harness driver override; defaults to the one the session was created with. */\n  driver?: SandboxDriver\n}\n\nexport class BuilderSession {\n  private store: TraceStore\n  private builderEmitter: TraceEmitter\n  readonly projectId: string\n  readonly chatId: string\n  private builderRunId?: string\n  private lastBuildRunId?: string\n  private defaultDriver?: SandboxDriver\n\n  constructor(store: TraceStore, init: BuilderSessionInit, driver?: SandboxDriver) {\n    this.store = store\n    this.projectId = init.projectId\n    this.chatId = init.chatId ?? newRecordId()\n    this.defaultDriver = driver\n    this.builderEmitter = new TraceEmitter(store)\n  }\n\n  /** Start the builder (L0) run for this chat. Returns the runId. */\n  async startChat(scenarioId = `${this.projectId}/chat`): Promise<string> {\n    await this.builderEmitter.startRun({\n      scenarioId,\n      projectId: this.projectId,\n      chatId: this.chatId,\n      layer: 'builder',\n    })\n    this.builderRunId = this.builderEmitter.runId\n    return this.builderRunId\n  }\n\n  /** The emitter for builder-level spans (edits, LLM calls, tool invocations). */\n  get emitter(): TraceEmitter {\n    if (!this.builderRunId) throw new Error('BuilderSession.emitter: call startChat() first')\n    return this.builderEmitter\n  }\n\n  /**\n   * Ship the project's generated app: run the sandbox harness as a child\n   * Run (`layer='app-build'`). Returns the build result + runId.\n   */\n  async ship(options: ShipOptions): Promise<{ runId: string; result: SandboxHarnessResult }> {\n    if (!this.builderRunId) throw new Error('BuilderSession.ship: call startChat() first')\n    const buildEmitter = new TraceEmitter(this.store)\n    await buildEmitter.startRun({\n      scenarioId: options.scenarioId ?? `${this.projectId}/build`,\n      projectId: this.projectId,\n      chatId: this.chatId,\n      parentRunId: this.builderRunId,\n      layer: 'app-build',\n    })\n    const harness = new SandboxHarness(options.driver ?? this.defaultDriver)\n    const result = await harness.run(options.harness, buildEmitter)\n    await buildEmitter.endRun({\n      pass: result.passed,\n      score: result.score,\n      failureClass: result.passed ? 'success' : 'sandbox_failure',\n    })\n    this.lastBuildRunId = buildEmitter.runId\n    return { runId: buildEmitter.runId, result }\n  }\n\n  /**\n   * Run a domain scenario against the just-built app as a grandchild Run\n   * (`layer='app-runtime'`). The `ship` call must precede this so the\n   * parent is set correctly; if no build exists yet the session attaches\n   * directly to the builder run (useful for prototypes).\n   */\n  async runAppScenario(options: RunAppScenarioOptions): Promise<TestGradedRunResult> {\n    const parentRunId = this.lastBuildRunId ?? this.builderRunId\n    if (!parentRunId)\n      throw new Error('BuilderSession.runAppScenario: call startChat() + ship() first')\n    const { scenario, driver } = options\n    const result = await runTestGradedScenario(scenario, this.store, {\n      driver: driver ?? this.defaultDriver,\n      provenance: { codeSha: undefined, promptSha: undefined, modelFingerprint: undefined },\n    })\n    // Attach to the parent chain by updating the stored Run in place.\n    await this.store.updateRun(result.runId, {\n      parentRunId,\n      projectId: this.projectId,\n      chatId: this.chatId,\n      layer: 'app-runtime',\n    })\n    return result\n  }\n\n  /** Record an end-of-chat meta score (judge verdict on whether the builder\n   *  served the user's intent). Accepts a numeric score + optional rationale. */\n  async recordMetaScore(score: number, rationale?: string): Promise<void> {\n    if (!this.builderRunId)\n      throw new Error('BuilderSession.recordMetaScore: call startChat() first')\n    await this.builderEmitter.recordJudge({\n      judgeId: 'builder-meta',\n      targetSpanId: this.builderRunId, // attach to the builder run itself\n      dimension: 'user_intent_satisfaction',\n      score,\n      rationale,\n      name: 'builder-meta',\n    })\n  }\n\n  /** Close the builder Run with a final outcome. */\n  async endChat(outcome: { pass: boolean; score?: number; notes?: string }): Promise<void> {\n    await this.builderEmitter.endRun({\n      pass: outcome.pass,\n      score: outcome.score,\n      notes: outcome.notes,\n    })\n  }\n\n  /**\n   * Inline app-runtime run — for cases where the \"scenario\" isn't a\n   * SWE-bench-style test suite but a live agent interaction (LLM chat,\n   * domain flow). Returns an emitter bound to a fresh Run in the\n   * `app-runtime` layer; caller emits spans inside and calls\n   * `.endRun()` with the final verdict.\n   */\n  async startAppRuntime(scenarioId: string): Promise<TraceEmitter> {\n    const parentRunId = this.lastBuildRunId ?? this.builderRunId\n    if (!parentRunId)\n      throw new Error(\n        'BuilderSession.startAppRuntime: call startChat() + (optionally) ship() first',\n      )\n    const emitter = new TraceEmitter(this.store)\n    await emitter.startRun({\n      scenarioId,\n      projectId: this.projectId,\n      chatId: this.chatId,\n      parentRunId,\n      layer: 'app-runtime',\n    })\n    return emitter\n  }\n\n  /**\n   * Lightweight \"ship marker\" — record an app-build Run with a caller-\n   * provided verdict. Use when there isn't a sandbox harness to run but\n   * you still want to mark the build state at publish time.\n   */\n  async recordShipMarker(args: {\n    pass: boolean\n    score: number\n    scenarioId?: string\n    notes?: string\n  }): Promise<string> {\n    if (!this.builderRunId)\n      throw new Error('BuilderSession.recordShipMarker: call startChat() first')\n    const emitter = new TraceEmitter(this.store)\n    await emitter.startRun({\n      scenarioId: args.scenarioId ?? `${this.projectId}/ship`,\n      projectId: this.projectId,\n      chatId: this.chatId,\n      parentRunId: this.builderRunId,\n      layer: 'app-build',\n    })\n    await emitter.endRun({\n      pass: args.pass,\n      score: args.score,\n      failureClass: args.pass ? 'success' : 'sandbox_failure',\n      notes: args.notes,\n    })\n    this.lastBuildRunId = emitter.runId\n    return emitter.runId\n  }\n\n  get lastBuildRunIdValue(): string | undefined {\n    return this.lastBuildRunId\n  }\n  get builderRunIdValue(): string | undefined {\n    return this.builderRunId\n  }\n}\n\n/**\n * Reconstruct the most recent BuilderSession state for a given project —\n * returns { builderRunId, lastBuildRunId, chatRuns }. For chat-first UIs\n * this is how a resumed session finds its place in the edit history.\n */\nexport async function resumeBuilderSession(\n  store: TraceStore,\n  projectId: string,\n): Promise<{\n  projectId: string\n  chatRuns: Run[]\n  lastBuilderRun?: Run\n  lastBuildRun?: Run\n  lastAppRuntimeRuns: Run[]\n}> {\n  const runs = await store.listRuns({ projectId })\n  const chatRuns = runs\n    .filter((r) => r.layer === 'builder')\n    .sort((a, b) => b.startedAt - a.startedAt)\n  const buildRuns = runs\n    .filter((r) => r.layer === 'app-build')\n    .sort((a, b) => b.startedAt - a.startedAt)\n  const appRuntimeRuns = runs\n    .filter((r) => r.layer === 'app-runtime')\n    .sort((a, b) => b.startedAt - a.startedAt)\n  return {\n    projectId,\n    chatRuns,\n    lastBuilderRun: chatRuns[0],\n    lastBuildRun: buildRuns[0],\n    lastAppRuntimeRuns: appRuntimeRuns,\n  }\n}\n","/**\n * Meta-eval correlation — the highest-leverage signal in the framework.\n *\n * Given a corpus of three-layer project reports, compute how well each\n * pair of layers correlates. The question we care about most:\n *\n *   Does `metaScore` (what the builder thinks it did) predict\n *   `runtimeScore` (what the user actually gets)?\n *\n * If r < ~0.4, the builder's self-scoring is broken — it's optimizing\n * for something other than real-world success. If r > 0.7, meta_score\n * is a usable proxy and can drive CI gates cheaply.\n *\n * Non-parametric rank correlation (Spearman) is also reported because\n * meta scores are often ordinal-ish.\n */\n\nimport { pearsonR, spearmanR } from '../statistics'\nimport type { ThreeLayerProjectReport } from './three-layer-eval'\n\nexport interface LayerCorrelation {\n  n: number\n  pearson: number\n  spearman: number\n}\n\nexport interface CorrelationReport {\n  /** Pairs present in the corpus (layers with ≥ 2 matched data points). */\n  metaVsBuild?: LayerCorrelation\n  metaVsRuntime?: LayerCorrelation\n  buildVsRuntime?: LayerCorrelation\n  /** Number of complete projects (all 3 scores present). */\n  completeProjects: number\n}\n\nexport function correlateLayers(reports: ThreeLayerProjectReport[]): CorrelationReport {\n  const completeProjects = reports.filter((r) => r.complete).length\n  return {\n    metaVsBuild: pairwise(\n      reports,\n      (r) => r.metaScore,\n      (r) => r.buildScore,\n    ),\n    metaVsRuntime: pairwise(\n      reports,\n      (r) => r.metaScore,\n      (r) => r.runtimeScore,\n    ),\n    buildVsRuntime: pairwise(\n      reports,\n      (r) => r.buildScore,\n      (r) => r.runtimeScore,\n    ),\n    completeProjects,\n  }\n}\n\nfunction pairwise(\n  reports: ThreeLayerProjectReport[],\n  a: (r: ThreeLayerProjectReport) => number | null,\n  b: (r: ThreeLayerProjectReport) => number | null,\n): LayerCorrelation | undefined {\n  const xs: number[] = []\n  const ys: number[] = []\n  for (const r of reports) {\n    const x = a(r)\n    const y = b(r)\n    if (x !== null && y !== null && Number.isFinite(x) && Number.isFinite(y)) {\n      xs.push(x)\n      ys.push(y)\n    }\n  }\n  if (xs.length < 2) return undefined\n  return {\n    n: xs.length,\n    pearson: pearsonR(xs, ys),\n    spearman: spearmanR(xs, ys),\n  }\n}\n","/**\n * ProjectRegistry — project-level aggregation over the trace corpus.\n *\n * Thin reader over TraceStore that answers the questions a chat-first,\n * resumable UI needs:\n *   - listProjects() → project IDs with latest activity\n *   - projectTimeline(id) → chats + builds + runtime runs, chronological\n *   - projectChats(id) → chat-level summaries (turn count, outcome)\n *\n * All queries are pure reads; no state duplication.\n */\n\nimport type { Run } from '../trace/schema'\nimport type { TraceStore } from '../trace/store'\n\nexport interface ProjectSummary {\n  projectId: string\n  chatCount: number\n  buildCount: number\n  appRuntimeCount: number\n  lastActivityAt: number\n  latestChatId?: string\n  latestOutcome?: { pass: boolean; score?: number }\n}\n\nexport interface ChatSummary {\n  chatId: string\n  projectId: string\n  builderRunId: string\n  startedAt: number\n  endedAt?: number\n  status: Run['status']\n  outcome?: Run['outcome']\n  /** Counts of spans emitted during the chat. */\n  llmTurns?: number\n  toolCalls?: number\n  buildRunId?: string\n  appRuntimeRunIds: string[]\n}\n\nexport interface ProjectTimelineEntry {\n  run: Run\n  layerBucket: 'chat' | 'build' | 'runtime' | 'other'\n}\n\nexport class ProjectRegistry {\n  constructor(private store: TraceStore) {}\n\n  async listProjects(): Promise<ProjectSummary[]> {\n    const runs = await this.store.listRuns()\n    const byProject = new Map<string, Run[]>()\n    for (const r of runs) {\n      if (!r.projectId) continue\n      const arr = byProject.get(r.projectId) ?? []\n      arr.push(r)\n      byProject.set(r.projectId, arr)\n    }\n    const summaries: ProjectSummary[] = []\n    for (const [projectId, projectRuns] of byProject) {\n      const sorted = projectRuns.slice().sort((a, b) => b.startedAt - a.startedAt)\n      const chats = projectRuns.filter((r) => r.layer === 'builder')\n      const builds = projectRuns.filter((r) => r.layer === 'app-build')\n      const runtimes = projectRuns.filter((r) => r.layer === 'app-runtime')\n      const latest = sorted[0]\n      if (!latest) continue\n      summaries.push({\n        projectId,\n        chatCount: chats.length,\n        buildCount: builds.length,\n        appRuntimeCount: runtimes.length,\n        lastActivityAt: latest.startedAt,\n        latestChatId: chats[0]?.chatId,\n        latestOutcome: latest.outcome\n          ? { pass: latest.outcome.pass ?? false, score: latest.outcome.score }\n          : undefined,\n      })\n    }\n    return summaries.sort((a, b) => b.lastActivityAt - a.lastActivityAt)\n  }\n\n  async projectTimeline(projectId: string): Promise<ProjectTimelineEntry[]> {\n    const runs = await this.store.listRuns({ projectId })\n    const ordered = runs.slice().sort((a, b) => a.startedAt - b.startedAt)\n    return ordered.map((run) => ({\n      run,\n      layerBucket:\n        run.layer === 'builder'\n          ? 'chat'\n          : run.layer === 'app-build'\n            ? 'build'\n            : run.layer === 'app-runtime'\n              ? 'runtime'\n              : 'other',\n    }))\n  }\n\n  async projectChats(projectId: string): Promise<ChatSummary[]> {\n    const builderRuns = (await this.store.listRuns({ projectId, layer: 'builder' })).sort(\n      (a, b) => b.startedAt - a.startedAt,\n    )\n    const childrenFor = async (runId: string) => this.store.listRuns({ parentRunId: runId })\n    const out: ChatSummary[] = []\n    for (const run of builderRuns) {\n      const spans = await this.store.spans({ runId: run.runId })\n      const children = await childrenFor(run.runId)\n      const build = children.find((c) => c.layer === 'app-build')\n      const runtime: string[] = []\n      // Runtime runs can be grandchildren (attached to the build) or siblings\n      // when shipped skipped.\n      if (build) {\n        const grands = await childrenFor(build.runId)\n        for (const g of grands) if (g.layer === 'app-runtime') runtime.push(g.runId)\n      }\n      for (const c of children) if (c.layer === 'app-runtime') runtime.push(c.runId)\n      out.push({\n        chatId: run.chatId ?? run.runId,\n        projectId,\n        builderRunId: run.runId,\n        startedAt: run.startedAt,\n        endedAt: run.endedAt,\n        status: run.status,\n        outcome: run.outcome,\n        llmTurns: spans.filter((s) => s.kind === 'llm').length,\n        toolCalls: spans.filter((s) => s.kind === 'tool').length,\n        buildRunId: build?.runId,\n        appRuntimeRunIds: runtime,\n      })\n    }\n    return out\n  }\n}\n","/**\n * Three-layer evaluation — the canonical scoring breakdown for\n * builder-of-builders workflows.\n *\n *   meta_score:    did the builder understand + satisfy user intent?\n *                  (judge verdict attached to the builder run)\n *   build_score:   did the generated scaffold build + pass its own tests?\n *                  (outcome.score on the app-build child run)\n *   runtime_score: did the generated agent pass its domain scenarios?\n *                  (mean outcome.score over app-runtime grandchild runs)\n *\n * Returns a structured report per project. The cross-layer correlation\n * is the highest-leverage signal the framework computes — if\n * meta_score doesn't predict runtime_score, the builder's self-scoring\n * is broken.\n *\n * Scaffold-only mode: when a project has no `app-runtime` runs (e.g. a\n * scaffold-builder eval that grades compose + build without driving a\n * runtime scenario), `kind` is `'scaffold-only'` and `complete` measures\n * meta + build only. Consumers can tell the two apart without having to\n * interpret null-runtime as either \"not yet computed\" or \"N/A for this\n * project shape\".\n */\n\nimport { judgeSpans } from '../trace/query'\nimport type { Run } from '../trace/schema'\nimport type { TraceStore } from '../trace/store'\n\nexport type ProjectKind = 'full' | 'scaffold-only'\n\nexport interface ThreeLayerProjectReport {\n  projectId: string\n  /**\n   * `'full'` when the project has at least one `app-runtime` run;\n   * `'scaffold-only'` when it only has meta + build layers. Lets\n   * downstream consumers treat a null runtime score as expected\n   * (scaffold-only) vs. missing (full, pipeline broke).\n   */\n  kind: ProjectKind\n  builderRunId?: string\n  /** Judge-verdict score on the builder run (0..1 after normalization). */\n  metaScore: number | null\n  buildRunId?: string\n  /** 0..1 from the sandbox harness (testsPassed / testsTotal). */\n  buildScore: number | null\n  appRuntimeRunIds: string[]\n  /** Mean of outcome.score over app-runtime runs, 0..1. Always null in scaffold-only mode. */\n  runtimeScore: number | null\n  runtimePassRate: number | null\n  /**\n   * Layer-aware completeness:\n   *   - `kind='full'`: all three layers scored\n   *   - `kind='scaffold-only'`: meta + build scored (runtime not applicable)\n   */\n  complete: boolean\n}\n\nexport async function scoreProject(\n  store: TraceStore,\n  projectId: string,\n): Promise<ThreeLayerProjectReport> {\n  const allRuns = await store.listRuns({ projectId })\n  const builder = latestByLayer(allRuns, 'builder')\n  const build = latestByLayer(allRuns, 'app-build')\n  const runtime = allRuns.filter((r) => r.layer === 'app-runtime')\n\n  const metaScore = builder ? await extractMetaScore(store, builder.runId) : null\n  const buildScore = build?.outcome?.score ?? null\n  const runtimeScores = runtime\n    .map((r) => r.outcome?.score)\n    .filter((s): s is number => typeof s === 'number')\n  const runtimeScore =\n    runtimeScores.length > 0\n      ? runtimeScores.reduce((a, b) => a + b, 0) / runtimeScores.length\n      : null\n  const runtimePassed = runtime.filter((r) => r.outcome?.pass === true).length\n  const runtimePassRate = runtime.length > 0 ? runtimePassed / runtime.length : null\n\n  const kind: ProjectKind = runtime.length === 0 ? 'scaffold-only' : 'full'\n  const complete =\n    kind === 'scaffold-only'\n      ? metaScore !== null && buildScore !== null\n      : metaScore !== null && buildScore !== null && runtimeScore !== null\n\n  return {\n    projectId,\n    kind,\n    builderRunId: builder?.runId,\n    metaScore,\n    buildRunId: build?.runId,\n    buildScore,\n    appRuntimeRunIds: runtime.map((r) => r.runId),\n    runtimeScore,\n    runtimePassRate,\n    complete,\n  }\n}\n\n/** Aggregate scoring across every project in a corpus. */\nexport async function scoreAllProjects(store: TraceStore): Promise<ThreeLayerProjectReport[]> {\n  const runs = await store.listRuns()\n  const projectIds = [...new Set(runs.map((r) => r.projectId).filter((p): p is string => !!p))]\n  return Promise.all(projectIds.map((p) => scoreProject(store, p)))\n}\n\nfunction latestByLayer(runs: Run[], layer: Run['layer']): Run | undefined {\n  const filtered = runs.filter((r) => r.layer === layer).sort((a, b) => b.startedAt - a.startedAt)\n  return filtered[0]\n}\n\nasync function extractMetaScore(store: TraceStore, builderRunId: string): Promise<number | null> {\n  const js = await judgeSpans(store, builderRunId)\n  const meta = js.find(\n    (s) => s.judgeId === 'builder-meta' && s.dimension === 'user_intent_satisfaction',\n  )\n  if (!meta) return null\n  // Normalize score to 0..1. Accept 0-1 natively; 0-10 scale is also common.\n  if (meta.score >= 0 && meta.score <= 1) return meta.score\n  if (meta.score >= 0 && meta.score <= 10) return meta.score / 10\n  return null\n}\n"],"mappings":";;;;;;;;;;;;;;;;;;;;AAmKA,IAAa,0BAAb,MAA8D;CAC5D,KAAK;CACL;CACA;CAEA,YAAY,UAA0C,CAAC,GAAG;EACxD,KAAK,aAAa,QAAQ;EAC1B,KAAK,aAAa,QAAQ;CAC5B;CAEA,MAAM,KACJ,OACA,SACA,QACwB;EAIxB,MAAM,eAAe,OAAO,OAAO,KAAK;EAGxC,MAAM,eAAe;GAAE,GAAG,QAAQ;GAAK,GAAI,KAAK,cAAc,CAAC;GAAI,GAAI,OAAO,OAAO,CAAC;EAAG;EACzF,MAAM,SAAS,MAAM,kBAAkB;GACrC;GACA,OAAO;GACP,KAAK;GACL,KAAK;GACL,SAAS;GACT,WAAW,OAAO,aAAa;GAC/B,gBAAgB,OAAO,kBAAkB;EAC3C,CAAC;EAID,MAAM,SACJ,CAAC,OAAO,mBAAmB,UAAU,UAAU,OAAO,aAClD,OAAO,WAAW,MAAM,OAAO,QAAQ,OAAO,QAAQ,OAAO,QAAQ,IACrE,KAAA;EACN,OAAO;GACL;GACA,UAAU,OAAO;GACjB,QAAQ,OAAO;GACf,QAAQ,OAAO,cAAc,OAAO,SAAS,OAAO,cAAc,OAAO;GACzE,QAAQ,OAAO;GACf,YAAY,QAAQ;GACpB,aAAa,QAAQ;GACrB,iBAAiB,OAAO,mBAAmB,KAAA;GAC3C,iBAAiB,OAAO,mBAAmB,KAAA;EAC7C;CACF;AACF;AAkBA,IAAa,iBAAb,MAA4B;CAC1B;CACA,YAAY,SAAwB,IAAI,wBAAwB,GAAG;EACjE,KAAK,SAAS;CAChB;CAEA,MAAM,IAAI,QAAuB,SAAsD;EACrF,MAAM,SAAS,MAAM,QAAQ,QAAQ;GACnC,MAAM,WAAW,KAAK,OAAO,GAAG;GAChC,OAAO,OAAO;GACd,SAAS;IAAC,OAAO;IAAc,OAAO;IAAY,OAAO;GAAW,CAAC,CAClE,OAAO,OAAO,CAAC,CACf,KAAK,MAAM;EAChB,CAAC;EACD,MAAM,SAA+B;GAAE,QAAQ;GAAO,aAAa;GAAG,OAAO;EAAE;EAC/E,IAAI;GACF,IAAI,OAAO,cAAc;IACvB,OAAO,QAAQ,MAAM,KAAK,OAAO,KAAK,SAAS,OAAO,cAAc,MAAM;IAC1E,OAAO,eAAe,OAAO,MAAM;IACnC,IAAI,OAAO,MAAM,mBAAmB,OAAO,MAAM,aAAa,GAAG;KAC/D,MAAM,SAAS,OAAO,MAAM,kBACxB,yBAAyB,OAAO,MAAM,OAAO,MAC7C,sBAAsB,OAAO,MAAM,SAAS;KAChD,MAAM,OAAO,KAAK,QAAQ;MACxB,UAAU,OAAO,MAAM;MACvB,QAAQ,OAAO;KACjB,CAAyB;KACzB,OAAO;IACT;GACF;GACA,IAAI,OAAO,YAAY;IACrB,OAAO,MAAM,MAAM,KAAK,OAAO,KAAK,OAAO,OAAO,YAAY,MAAM;IACpE,OAAO,eAAe,OAAO,IAAI;IACjC,IAAI,OAAO,IAAI,mBAAmB,OAAO,IAAI,aAAa,GAAG;KAC3D,MAAM,SAAS,OAAO,IAAI,kBACtB,uBAAuB,OAAO,IAAI,OAAO,MACzC,oBAAoB,OAAO,IAAI,SAAS;KAC5C,MAAM,OAAO,KAAK,QAAQ;MACxB,UAAU,OAAO,IAAI;MACrB,QAAQ,OAAO;KACjB,CAAyB;KACzB,OAAO;IACT;GACF;GACA,IAAI,OAAO,aAAa;IACtB,OAAO,OAAO,MAAM,KAAK,OAAO,KAAK,QAAQ,OAAO,aAAa,MAAM;IACvE,OAAO,eAAe,OAAO,KAAK;IAGlC,MAAM,SAAS,CAAC,OAAO,KAAK,mBAAmB,OAAO,KAAK,aAAa;IACxE,OAAO,SAAS;IAChB,IAAI,OAAO,KAAK,eAAe,KAAA,KAAa,OAAO,KAAK,aAAa,GACnE,OAAO,SAAS,OAAO,KAAK,eAAe,KAAK,OAAO,KAAK;SAE5D,OAAO,QAAQ,SAAS,IAAI;IAE9B,MAAM,OAAO,IAAI;KACf,UAAU,OAAO,KAAK;KACtB,YAAY,OAAO,KAAK;KACxB,aAAa,OAAO,KAAK;KACzB,QAAQ,OAAO;KACf,QAAQ,SAAS,OAAO;IAC1B,CAAyB;GAC3B,OAAO;IACL,OAAO,SAAS;IAChB,OAAO,QAAQ;IACf,MAAM,OAAO,IAAI,EAAE,QAAQ,OAAO,YAAY,CAAyB;GACzE;EACF,SAAS,KAAK;GACZ,MAAM,OAAO,KAAK,eAAe,QAAQ,MAAM,OAAO,GAAG,CAAC;GAC1D,MAAM;EACR;EACA,OAAO;CACT;AACF;;;AClQA,eAAsB,sBACpB,UACA,OACA,UAAgC,CAAC,GACH;CAC9B,MAAM,UAAU,IAAI,aAAa,KAAK;CACtC,MAAM,QAAQ,SAAS;EACrB,YAAY,SAAS;EACrB,WAAW,QAAQ;EACnB,gBAAgB,SAAS;EACzB,MAAM,SAAS;EACf,GAAG,QAAQ;CACb,CAAC;CAED,MAAM,SAAS,MAAM,IADD,eAAe,QAAQ,MAChB,CAAC,CAAC,IAAI,SAAS,SAAS,OAAO;CAC1D,MAAM,YAAY,SAAS,iBAAiB;CAC5C,MAAM,OAAO,OAAO,UAAU,OAAO,SAAS;CAC9C,MAAM,cAAc,OAAO,UAAU,KAAA,KAAa,OAAO,MAAM,aAAa;CAC5E,MAAM,YAAY,OAAO,QAAQ,KAAA,KAAa,OAAO,IAAI,aAAa;CACtE,MAAM,aAAa,OAAO,SAAS,KAAA,KAAa,OAAO,KAAK,aAAa;CACzE,MAAM,eAAyC,OAC3C,YACA,eAAe,YACb,oBACA,aACE,iBACA;CACR,MAAM,QAAQ,OAAO;EACnB;EACA,OAAO,OAAO;EACd;EACA,OAAO,OAAO,KAAA,IAAY,iBAAiB,MAAM;CACnD,CAAC;CACD,OAAO;EACL,OAAO,QAAQ;EACf;EACA,SAAS;EACT;EACA,OAAO,OAAO;EACd;CACF;AACF;AAEA,SAAS,iBAAiB,QAAsC;CAC9D,IAAI,OAAO,SAAS,OAAO,MAAM,aAAa,GAC5C,OAAO,sBAAsB,OAAO,MAAM;CAC5C,IAAI,OAAO,OAAO,OAAO,IAAI,aAAa,GAAG,OAAO,oBAAoB,OAAO,IAAI;CACnF,IAAI,OAAO,MAAM;EACf,IAAI,OAAO,KAAK,eAAe,KAAA,GAC7B,OAAO,UAAU,OAAO,KAAK,eAAe,EAAE,GAAG,OAAO,KAAK;EAE/D,OAAO,aAAa,OAAO,KAAK;CAClC;CACA,OAAO;AACT;;;;;;;;;;;;;;;;;;;;ACtDA,IAAa,iBAAb,MAA4B;CAC1B;CACA;CACA;CACA;CACA;CACA;CACA;CAEA,YAAY,OAAmB,MAA0B,QAAwB;EAC/E,KAAK,QAAQ;EACb,KAAK,YAAY,KAAK;EACtB,KAAK,SAAS,KAAK,UAAU,YAAY;EACzC,KAAK,gBAAgB;EACrB,KAAK,iBAAiB,IAAI,aAAa,KAAK;CAC9C;;CAGA,MAAM,UAAU,aAAa,GAAG,KAAK,UAAU,QAAyB;EACtE,MAAM,KAAK,eAAe,SAAS;GACjC;GACA,WAAW,KAAK;GAChB,QAAQ,KAAK;GACb,OAAO;EACT,CAAC;EACD,KAAK,eAAe,KAAK,eAAe;EACxC,OAAO,KAAK;CACd;;CAGA,IAAI,UAAwB;EAC1B,IAAI,CAAC,KAAK,cAAc,MAAM,IAAI,MAAM,gDAAgD;EACxF,OAAO,KAAK;CACd;;;;;CAMA,MAAM,KAAK,SAAgF;EACzF,IAAI,CAAC,KAAK,cAAc,MAAM,IAAI,MAAM,6CAA6C;EACrF,MAAM,eAAe,IAAI,aAAa,KAAK,KAAK;EAChD,MAAM,aAAa,SAAS;GAC1B,YAAY,QAAQ,cAAc,GAAG,KAAK,UAAU;GACpD,WAAW,KAAK;GAChB,QAAQ,KAAK;GACb,aAAa,KAAK;GAClB,OAAO;EACT,CAAC;EAED,MAAM,SAAS,MAAM,IADD,eAAe,QAAQ,UAAU,KAAK,aAC/B,CAAC,CAAC,IAAI,QAAQ,SAAS,YAAY;EAC9D,MAAM,aAAa,OAAO;GACxB,MAAM,OAAO;GACb,OAAO,OAAO;GACd,cAAc,OAAO,SAAS,YAAY;EAC5C,CAAC;EACD,KAAK,iBAAiB,aAAa;EACnC,OAAO;GAAE,OAAO,aAAa;GAAO;EAAO;CAC7C;;;;;;;CAQA,MAAM,eAAe,SAA8D;EACjF,MAAM,cAAc,KAAK,kBAAkB,KAAK;EAChD,IAAI,CAAC,aACH,MAAM,IAAI,MAAM,gEAAgE;EAClF,MAAM,EAAE,UAAU,WAAW;EAC7B,MAAM,SAAS,MAAM,sBAAsB,UAAU,KAAK,OAAO;GAC/D,QAAQ,UAAU,KAAK;GACvB,YAAY;IAAE,SAAS,KAAA;IAAW,WAAW,KAAA;IAAW,kBAAkB,KAAA;GAAU;EACtF,CAAC;EAED,MAAM,KAAK,MAAM,UAAU,OAAO,OAAO;GACvC;GACA,WAAW,KAAK;GAChB,QAAQ,KAAK;GACb,OAAO;EACT,CAAC;EACD,OAAO;CACT;;;CAIA,MAAM,gBAAgB,OAAe,WAAmC;EACtE,IAAI,CAAC,KAAK,cACR,MAAM,IAAI,MAAM,wDAAwD;EAC1E,MAAM,KAAK,eAAe,YAAY;GACpC,SAAS;GACT,cAAc,KAAK;GACnB,WAAW;GACX;GACA;GACA,MAAM;EACR,CAAC;CACH;;CAGA,MAAM,QAAQ,SAA2E;EACvF,MAAM,KAAK,eAAe,OAAO;GAC/B,MAAM,QAAQ;GACd,OAAO,QAAQ;GACf,OAAO,QAAQ;EACjB,CAAC;CACH;;;;;;;;CASA,MAAM,gBAAgB,YAA2C;EAC/D,MAAM,cAAc,KAAK,kBAAkB,KAAK;EAChD,IAAI,CAAC,aACH,MAAM,IAAI,MACR,8EACF;EACF,MAAM,UAAU,IAAI,aAAa,KAAK,KAAK;EAC3C,MAAM,QAAQ,SAAS;GACrB;GACA,WAAW,KAAK;GAChB,QAAQ,KAAK;GACb;GACA,OAAO;EACT,CAAC;EACD,OAAO;CACT;;;;;;CAOA,MAAM,iBAAiB,MAKH;EAClB,IAAI,CAAC,KAAK,cACR,MAAM,IAAI,MAAM,yDAAyD;EAC3E,MAAM,UAAU,IAAI,aAAa,KAAK,KAAK;EAC3C,MAAM,QAAQ,SAAS;GACrB,YAAY,KAAK,cAAc,GAAG,KAAK,UAAU;GACjD,WAAW,KAAK;GAChB,QAAQ,KAAK;GACb,aAAa,KAAK;GAClB,OAAO;EACT,CAAC;EACD,MAAM,QAAQ,OAAO;GACnB,MAAM,KAAK;GACX,OAAO,KAAK;GACZ,cAAc,KAAK,OAAO,YAAY;GACtC,OAAO,KAAK;EACd,CAAC;EACD,KAAK,iBAAiB,QAAQ;EAC9B,OAAO,QAAQ;CACjB;CAEA,IAAI,sBAA0C;EAC5C,OAAO,KAAK;CACd;CACA,IAAI,oBAAwC;EAC1C,OAAO,KAAK;CACd;AACF;;;;;;AAOA,eAAsB,qBACpB,OACA,WAOC;CACD,MAAM,OAAO,MAAM,MAAM,SAAS,EAAE,UAAU,CAAC;CAC/C,MAAM,WAAW,KACd,QAAQ,MAAM,EAAE,UAAU,SAAS,CAAC,CACpC,MAAM,GAAG,MAAM,EAAE,YAAY,EAAE,SAAS;CAC3C,MAAM,YAAY,KACf,QAAQ,MAAM,EAAE,UAAU,WAAW,CAAC,CACtC,MAAM,GAAG,MAAM,EAAE,YAAY,EAAE,SAAS;CAC3C,MAAM,iBAAiB,KACpB,QAAQ,MAAM,EAAE,UAAU,aAAa,CAAC,CACxC,MAAM,GAAG,MAAM,EAAE,YAAY,EAAE,SAAS;CAC3C,OAAO;EACL;EACA;EACA,gBAAgB,SAAS;EACzB,cAAc,UAAU;EACxB,oBAAoB;CACtB;AACF;;;;;;;;;;;;;;;;;;;ACxNA,SAAgB,gBAAgB,SAAuD;CACrF,MAAM,mBAAmB,QAAQ,QAAQ,MAAM,EAAE,QAAQ,CAAC,CAAC;CAC3D,OAAO;EACL,aAAa,SACX,UACC,MAAM,EAAE,YACR,MAAM,EAAE,UACX;EACA,eAAe,SACb,UACC,MAAM,EAAE,YACR,MAAM,EAAE,YACX;EACA,gBAAgB,SACd,UACC,MAAM,EAAE,aACR,MAAM,EAAE,YACX;EACA;CACF;AACF;AAEA,SAAS,SACP,SACA,GACA,GAC8B;CAC9B,MAAM,KAAe,CAAC;CACtB,MAAM,KAAe,CAAC;CACtB,KAAK,MAAM,KAAK,SAAS;EACvB,MAAM,IAAI,EAAE,CAAC;EACb,MAAM,IAAI,EAAE,CAAC;EACb,IAAI,MAAM,QAAQ,MAAM,QAAQ,OAAO,SAAS,CAAC,KAAK,OAAO,SAAS,CAAC,GAAG;GACxE,GAAG,KAAK,CAAC;GACT,GAAG,KAAK,CAAC;EACX;CACF;CACA,IAAI,GAAG,SAAS,GAAG,OAAO,KAAA;CAC1B,OAAO;EACL,GAAG,GAAG;EACN,SAAS,SAAS,IAAI,EAAE;EACxB,UAAU,UAAU,IAAI,EAAE;CAC5B;AACF;;;ACjCA,IAAa,kBAAb,MAA6B;CACP;CAApB,YAAY,OAA2B;EAAnB,KAAA,QAAA;CAAoB;CAExC,MAAM,eAA0C;EAC9C,MAAM,OAAO,MAAM,KAAK,MAAM,SAAS;EACvC,MAAM,4BAAY,IAAI,IAAmB;EACzC,KAAK,MAAM,KAAK,MAAM;GACpB,IAAI,CAAC,EAAE,WAAW;GAClB,MAAM,MAAM,UAAU,IAAI,EAAE,SAAS,KAAK,CAAC;GAC3C,IAAI,KAAK,CAAC;GACV,UAAU,IAAI,EAAE,WAAW,GAAG;EAChC;EACA,MAAM,YAA8B,CAAC;EACrC,KAAK,MAAM,CAAC,WAAW,gBAAgB,WAAW;GAChD,MAAM,SAAS,YAAY,MAAM,CAAC,CAAC,MAAM,GAAG,MAAM,EAAE,YAAY,EAAE,SAAS;GAC3E,MAAM,QAAQ,YAAY,QAAQ,MAAM,EAAE,UAAU,SAAS;GAC7D,MAAM,SAAS,YAAY,QAAQ,MAAM,EAAE,UAAU,WAAW;GAChE,MAAM,WAAW,YAAY,QAAQ,MAAM,EAAE,UAAU,aAAa;GACpE,MAAM,SAAS,OAAO;GACtB,IAAI,CAAC,QAAQ;GACb,UAAU,KAAK;IACb;IACA,WAAW,MAAM;IACjB,YAAY,OAAO;IACnB,iBAAiB,SAAS;IAC1B,gBAAgB,OAAO;IACvB,cAAc,MAAM,EAAE,EAAE;IACxB,eAAe,OAAO,UAClB;KAAE,MAAM,OAAO,QAAQ,QAAQ;KAAO,OAAO,OAAO,QAAQ;IAAM,IAClE,KAAA;GACN,CAAC;EACH;EACA,OAAO,UAAU,MAAM,GAAG,MAAM,EAAE,iBAAiB,EAAE,cAAc;CACrE;CAEA,MAAM,gBAAgB,WAAoD;EAGxE,QADgB,MADG,KAAK,MAAM,SAAS,EAAE,UAAU,CAAC,EAAA,CAC/B,MAAM,CAAC,CAAC,MAAM,GAAG,MAAM,EAAE,YAAY,EAAE,SAC/C,CAAC,CAAC,KAAK,SAAS;GAC3B;GACA,aACE,IAAI,UAAU,YACV,SACA,IAAI,UAAU,cACZ,UACA,IAAI,UAAU,gBACZ,YACA;EACZ,EAAE;CACJ;CAEA,MAAM,aAAa,WAA2C;EAC5D,MAAM,eAAe,MAAM,KAAK,MAAM,SAAS;GAAE;GAAW,OAAO;EAAU,CAAC,EAAA,CAAG,MAC9E,GAAG,MAAM,EAAE,YAAY,EAAE,SAC5B;EACA,MAAM,cAAc,OAAO,UAAkB,KAAK,MAAM,SAAS,EAAE,aAAa,MAAM,CAAC;EACvF,MAAM,MAAqB,CAAC;EAC5B,KAAK,MAAM,OAAO,aAAa;GAC7B,MAAM,QAAQ,MAAM,KAAK,MAAM,MAAM,EAAE,OAAO,IAAI,MAAM,CAAC;GACzD,MAAM,WAAW,MAAM,YAAY,IAAI,KAAK;GAC5C,MAAM,QAAQ,SAAS,MAAM,MAAM,EAAE,UAAU,WAAW;GAC1D,MAAM,UAAoB,CAAC;GAG3B,IAAI,OAAO;IACT,MAAM,SAAS,MAAM,YAAY,MAAM,KAAK;IAC5C,KAAK,MAAM,KAAK,QAAQ,IAAI,EAAE,UAAU,eAAe,QAAQ,KAAK,EAAE,KAAK;GAC7E;GACA,KAAK,MAAM,KAAK,UAAU,IAAI,EAAE,UAAU,eAAe,QAAQ,KAAK,EAAE,KAAK;GAC7E,IAAI,KAAK;IACP,QAAQ,IAAI,UAAU,IAAI;IAC1B;IACA,cAAc,IAAI;IAClB,WAAW,IAAI;IACf,SAAS,IAAI;IACb,QAAQ,IAAI;IACZ,SAAS,IAAI;IACb,UAAU,MAAM,QAAQ,MAAM,EAAE,SAAS,KAAK,CAAC,CAAC;IAChD,WAAW,MAAM,QAAQ,MAAM,EAAE,SAAS,MAAM,CAAC,CAAC;IAClD,YAAY,OAAO;IACnB,kBAAkB;GACpB,CAAC;EACH;EACA,OAAO;CACT;AACF;;;;;;;;;;;;;;;;;;;;;;;;;;ACzEA,eAAsB,aACpB,OACA,WACkC;CAClC,MAAM,UAAU,MAAM,MAAM,SAAS,EAAE,UAAU,CAAC;CAClD,MAAM,UAAU,cAAc,SAAS,SAAS;CAChD,MAAM,QAAQ,cAAc,SAAS,WAAW;CAChD,MAAM,UAAU,QAAQ,QAAQ,MAAM,EAAE,UAAU,aAAa;CAE/D,MAAM,YAAY,UAAU,MAAM,iBAAiB,OAAO,QAAQ,KAAK,IAAI;CAC3E,MAAM,aAAa,OAAO,SAAS,SAAS;CAC5C,MAAM,gBAAgB,QACnB,KAAK,MAAM,EAAE,SAAS,KAAK,CAAC,CAC5B,QAAQ,MAAmB,OAAO,MAAM,QAAQ;CACnD,MAAM,eACJ,cAAc,SAAS,IACnB,cAAc,QAAQ,GAAG,MAAM,IAAI,GAAG,CAAC,IAAI,cAAc,SACzD;CACN,MAAM,gBAAgB,QAAQ,QAAQ,MAAM,EAAE,SAAS,SAAS,IAAI,CAAC,CAAC;CACtE,MAAM,kBAAkB,QAAQ,SAAS,IAAI,gBAAgB,QAAQ,SAAS;CAE9E,MAAM,OAAoB,QAAQ,WAAW,IAAI,kBAAkB;CACnE,MAAM,WACJ,SAAS,kBACL,cAAc,QAAQ,eAAe,OACrC,cAAc,QAAQ,eAAe,QAAQ,iBAAiB;CAEpE,OAAO;EACL;EACA;EACA,cAAc,SAAS;EACvB;EACA,YAAY,OAAO;EACnB;EACA,kBAAkB,QAAQ,KAAK,MAAM,EAAE,KAAK;EAC5C;EACA;EACA;CACF;AACF;;AAGA,eAAsB,iBAAiB,OAAuD;CAC5F,MAAM,OAAO,MAAM,MAAM,SAAS;CAClC,MAAM,aAAa,CAAC,GAAG,IAAI,IAAI,KAAK,KAAK,MAAM,EAAE,SAAS,CAAC,CAAC,QAAQ,MAAmB,CAAC,CAAC,CAAC,CAAC,CAAC;CAC5F,OAAO,QAAQ,IAAI,WAAW,KAAK,MAAM,aAAa,OAAO,CAAC,CAAC,CAAC;AAClE;AAEA,SAAS,cAAc,MAAa,OAAsC;CAExE,OADiB,KAAK,QAAQ,MAAM,EAAE,UAAU,KAAK,CAAC,CAAC,MAAM,GAAG,MAAM,EAAE,YAAY,EAAE,SACxE,CAAC,CAAC;AAClB;AAEA,eAAe,iBAAiB,OAAmB,cAA8C;CAE/F,MAAM,QAAO,MADI,WAAW,OAAO,YAAY,EAAA,CAC/B,MACb,MAAM,EAAE,YAAY,kBAAkB,EAAE,cAAc,0BACzD;CACA,IAAI,CAAC,MAAM,OAAO;CAElB,IAAI,KAAK,SAAS,KAAK,KAAK,SAAS,GAAG,OAAO,KAAK;CACpD,IAAI,KAAK,SAAS,KAAK,KAAK,SAAS,IAAI,OAAO,KAAK,QAAQ;CAC7D,OAAO;AACT"}
+{"version":3,"file":"index.js","names":[],"sources":["../../src/sandbox-harness.ts","../../src/test-graded-scenario.ts","../../src/builder-eval/builder-session.ts","../../src/builder-eval/correlation.ts","../../src/builder-eval/project-registry.ts","../../src/builder-eval/three-layer-eval.ts"],"sourcesContent":["/**\n * SandboxHarness — executes a scenario in an isolated environment and\n * emits a rich SandboxSpan into the trace.\n *\n * Two built-in drivers:\n *   - `SubprocessSandboxDriver` — runs the phase in a local cwd with env\n *     vars, through `runBoundedProcess` (deadline, process-group kill,\n *     output cap). Fast, no dependencies, fine for unit tests and CI gates.\n *   - `DockerSandboxDriver` — lifted from tangle-router's sandbox path;\n *     shells out to `docker run`. Stronger isolation, slower startup.\n *\n * Consumers implement `SandboxDriver` for custom backends (Firecracker,\n * Cloudflare sandbox product, etc.). The harness doesn't care which.\n */\n\nimport { runBoundedProcess } from './bounded-process'\n\nimport type { TraceEmitter } from './trace/emitter'\nimport type { SandboxSpan } from './trace/schema'\n\nexport interface HarnessConfig {\n  /** Setup command (e.g. \"pnpm install\"). Non-zero exit fails the run. */\n  setupCommand?: string\n  /** Run command (e.g. \"pnpm build\"). */\n  runCommand?: string\n  /** Test command (e.g. \"pnpm test --run\"). Drives the test count + pass count. */\n  testCommand?: string\n  /** Absolute cwd for the subprocess driver. Ignored by docker driver. */\n  cwd?: string\n  /** Max wall-clock per phase in ms. Default 10 minutes. */\n  timeoutMs?: number\n  /**\n   * Cap on captured stdout+stderr bytes per phase. A runaway process can\n   * otherwise grow the in-memory buffer without bound. Once hit, further\n   * output is dropped and `outputTruncated` is set. Default 16 MiB.\n   */\n  maxOutputBytes?: number\n  /** Image for the docker driver. */\n  image?: string\n  /** Extra env vars (validated; shell-escaped). */\n  env?: Record<string, string>\n  /** Parser for the test output — maps stdout/stderr/exit code → pass count. */\n  testParser?: TestOutputParser\n}\n\nexport interface TestOutputParser {\n  id: string\n  parse(\n    stdout: string,\n    stderr: string,\n    exitCode: number,\n  ): { testsTotal: number; testsPassed: number } | undefined\n}\n\nexport interface SandboxResult {\n  phase: 'setup' | 'run' | 'test'\n  exitCode: number\n  stdout: string\n  stderr: string\n  wallMs: number\n  testsTotal?: number\n  testsPassed?: number\n  /**\n   * True when the process was killed because it exceeded `timeoutMs`. A\n   * SIGKILLed child can still close with exit code 0; callers MUST treat\n   * a timed-out phase as a hard failure regardless of `exitCode`, never\n   * as a pass. `undefined`/`false` means the process completed on its own.\n   */\n  killedByTimeout?: boolean\n  /**\n   * True when captured stdout/stderr hit `maxOutputBytes` and further\n   * output was dropped. The result is still returned (the process was\n   * not killed for this), but downstream parsers see truncated text.\n   */\n  outputTruncated?: boolean\n}\n\nexport interface SandboxDriver {\n  id: string\n  exec(\n    phase: SandboxResult['phase'],\n    command: string,\n    config: HarnessConfig,\n  ): Promise<SandboxResult>\n}\n\n// ── Parsers ──────────────────────────────────────────────────────────\n\n/** Vitest default summary line: \"Tests  X passed | Y failed\". */\nexport const vitestTestParser: TestOutputParser = {\n  id: 'vitest',\n  parse(stdout) {\n    const m = stdout.match(/Tests\\s+(\\d+)\\s+(passed|failed)(?:\\s*\\|\\s*(\\d+)\\s+(passed|failed))?/i)\n    if (!m) return undefined\n    let passed = 0\n    let failed = 0\n    const a = parseInt(m[1]!, 10)\n    const aLabel = m[2]!.toLowerCase()\n    if (aLabel === 'passed') passed += a\n    else failed += a\n    if (m[3] && m[4]) {\n      const b = parseInt(m[3], 10)\n      if (m[4].toLowerCase() === 'passed') passed += b\n      else failed += b\n    }\n    return { testsTotal: passed + failed, testsPassed: passed }\n  },\n}\n\n/** Pytest default: \"collected N items\" + \" X passed, Y failed\". */\nexport const pytestTestParser: TestOutputParser = {\n  id: 'pytest',\n  parse(stdout) {\n    const total = stdout.match(/collected\\s+(\\d+)\\s+items?/i)\n    const passed = stdout.match(/(\\d+)\\s+passed/)\n    if (!total || !passed) return undefined\n    return { testsTotal: parseInt(total[1]!, 10), testsPassed: parseInt(passed[1]!, 10) }\n  },\n}\n\n/** Jest: \"Tests: X passed, Y total\" (and optional failed). */\nexport const jestTestParser: TestOutputParser = {\n  id: 'jest',\n  parse(stdout) {\n    const m = stdout.match(/Tests:\\s+(?:(\\d+)\\s+failed[^,]*,\\s*)?(\\d+)\\s+passed,\\s+(\\d+)\\s+total/i)\n    if (!m) return undefined\n    return { testsTotal: parseInt(m[3]!, 10), testsPassed: parseInt(m[2]!, 10) }\n  },\n}\n\n/** Composite parser — tries a list of parsers in order. */\nexport function composeParsers(...parsers: TestOutputParser[]): TestOutputParser {\n  return {\n    id: parsers.map((p) => p.id).join('|'),\n    parse(stdout, stderr, exitCode) {\n      for (const p of parsers) {\n        const res = p.parse(stdout, stderr, exitCode)\n        if (res) return res\n      }\n      return undefined\n    },\n  }\n}\n\n// ── Drivers ──────────────────────────────────────────────────────────\n\ninterface SubprocessSandboxDriverOptions {\n  /**\n   * Default cwd for all `exec` calls. Used when the per-call `HarnessConfig`\n   * does not set its own `cwd`. Lets callers bind the driver to a working\n   * directory once instead of spreading cwd into every harness config —\n   * useful when the harness config is constructed far from the call site\n   * (e.g. starter-foundry's promoter passes a static HarnessConfig per\n   * family taxonomy but needs a per-run composed-scaffold cwd).\n   */\n  cwd?: string\n  /**\n   * Default env merged into every `exec` call's env (per-call `HarnessConfig.env`\n   * still wins on key collision). Same ergonomic rationale as `cwd` above.\n   */\n  env?: Record<string, string>\n}\n\nexport class SubprocessSandboxDriver implements SandboxDriver {\n  id = 'subprocess'\n  private defaultCwd?: string\n  private defaultEnv?: Record<string, string>\n\n  constructor(options: SubprocessSandboxDriverOptions = {}) {\n    this.defaultCwd = options.cwd\n    this.defaultEnv = options.env\n  }\n\n  async exec(\n    phase: SandboxResult['phase'],\n    command: string,\n    config: HarnessConfig,\n  ): Promise<SandboxResult> {\n    // Per-call config wins; fall back to constructor defaults. Honoring\n    // the constructor `cwd` keeps the subprocess from inheriting Node's\n    // cwd when only the constructor arg is supplied.\n    const effectiveCwd = config.cwd ?? this.defaultCwd\n    // The parent environment is merged here, so the runner is handed the\n    // complete environment and passes it through unchanged.\n    const effectiveEnv = { ...process.env, ...(this.defaultEnv ?? {}), ...(config.env ?? {}) }\n    const result = await runBoundedProcess({\n      command,\n      shell: true,\n      cwd: effectiveCwd,\n      env: effectiveEnv,\n      envMode: 'replace',\n      timeoutMs: config.timeoutMs ?? 10 * 60_000,\n      maxOutputBytes: config.maxOutputBytes ?? 16 * 1024 * 1024,\n    })\n    // A timed-out phase never reaches the parser: the output of a killed\n    // process is a fragment, and a fragment that happens to parse would\n    // report a pass for a run that never finished.\n    const parsed =\n      !result.killedByTimeout && phase === 'test' && config.testParser\n        ? config.testParser.parse(result.stdout, result.stderr, result.exitCode)\n        : undefined\n    return {\n      phase,\n      exitCode: result.exitCode,\n      stdout: result.stdout,\n      stderr: result.runnerError ? result.stderr + result.runnerError : result.stderr,\n      wallMs: result.wallMs,\n      testsTotal: parsed?.testsTotal,\n      testsPassed: parsed?.testsPassed,\n      killedByTimeout: result.killedByTimeout || undefined,\n      outputTruncated: result.outputTruncated || undefined,\n    }\n  }\n}\n\n\n\n\n\n// ── Harness orchestration ────────────────────────────────────────────\n\nexport interface SandboxHarnessResult {\n  passed: boolean\n  setup?: SandboxResult\n  run?: SandboxResult\n  test?: SandboxResult\n  totalWallMs: number\n  /** Final score — 0 when no tests; otherwise testsPassed/testsTotal. */\n  score: number\n}\n\nexport class SandboxHarness {\n  private driver: SandboxDriver\n  constructor(driver: SandboxDriver = new SubprocessSandboxDriver()) {\n    this.driver = driver\n  }\n\n  async run(config: HarnessConfig, emitter: TraceEmitter): Promise<SandboxHarnessResult> {\n    const handle = await emitter.sandbox({\n      name: `sandbox(${this.driver.id})`,\n      image: config.image,\n      command: [config.setupCommand, config.runCommand, config.testCommand]\n        .filter(Boolean)\n        .join(' && '),\n    })\n    const result: SandboxHarnessResult = { passed: false, totalWallMs: 0, score: 0 }\n    try {\n      if (config.setupCommand) {\n        result.setup = await this.driver.exec('setup', config.setupCommand, config)\n        result.totalWallMs += result.setup.wallMs\n        if (result.setup.killedByTimeout || result.setup.exitCode !== 0) {\n          const reason = result.setup.killedByTimeout\n            ? `setup timed out after ${result.setup.wallMs}ms`\n            : `setup failed (exit ${result.setup.exitCode})`\n          await handle.fail(reason, {\n            exitCode: result.setup.exitCode,\n            wallMs: result.totalWallMs,\n          } as Partial<SandboxSpan>)\n          return result\n        }\n      }\n      if (config.runCommand) {\n        result.run = await this.driver.exec('run', config.runCommand, config)\n        result.totalWallMs += result.run.wallMs\n        if (result.run.killedByTimeout || result.run.exitCode !== 0) {\n          const reason = result.run.killedByTimeout\n            ? `run timed out after ${result.run.wallMs}ms`\n            : `run failed (exit ${result.run.exitCode})`\n          await handle.fail(reason, {\n            exitCode: result.run.exitCode,\n            wallMs: result.totalWallMs,\n          } as Partial<SandboxSpan>)\n          return result\n        }\n      }\n      if (config.testCommand) {\n        result.test = await this.driver.exec('test', config.testCommand, config)\n        result.totalWallMs += result.test.wallMs\n        // A timed-out test phase is a hard fail regardless of the exit code\n        // a SIGKILLed child happens to report — never let it score as a pass.\n        const passed = !result.test.killedByTimeout && result.test.exitCode === 0\n        result.passed = passed\n        if (result.test.testsTotal !== undefined && result.test.testsTotal > 0) {\n          result.score = (result.test.testsPassed ?? 0) / result.test.testsTotal\n        } else {\n          result.score = passed ? 1 : 0\n        }\n        await handle.end({\n          exitCode: result.test.exitCode,\n          testsTotal: result.test.testsTotal,\n          testsPassed: result.test.testsPassed,\n          wallMs: result.totalWallMs,\n          status: passed ? 'ok' : 'error',\n        } as Partial<SandboxSpan>)\n      } else {\n        result.passed = true\n        result.score = 1\n        await handle.end({ wallMs: result.totalWallMs } as Partial<SandboxSpan>)\n      }\n    } catch (err) {\n      await handle.fail(err instanceof Error ? err : String(err))\n      throw err\n    }\n    return result\n  }\n}\n","/**\n * TestGradedScenario — a scenario whose score comes from a test suite.\n *\n * This is the SWE-bench pattern generalized. The scenario ships:\n *   - fixture data (setup instructions)\n *   - a test command the harness runs\n *   - optional assertion overrides\n *\n * The runner emits a run, delegates to SandboxHarness, records the\n * outcome, and returns a structured verdict. Consumers bind their own\n * agent execution to this contract.\n */\n\nimport type { HarnessConfig, SandboxDriver, SandboxHarnessResult } from './sandbox-harness'\nimport { SandboxHarness } from './sandbox-harness'\nimport { TraceEmitter } from './trace/emitter'\nimport type { FailureClass, Run } from './trace/schema'\nimport type { TraceStore } from './trace/store'\n\nexport interface TestGradedScenario {\n  id: string\n  description?: string\n  harness: HarnessConfig\n  /** Optional pass threshold in 0..1 (default 1.0 = all tests must pass). */\n  passThreshold?: number\n  /** Provenance for dataset tracking. */\n  datasetVersion?: string\n  /** Free-form tags (difficulty, category, etc.). */\n  tags?: Record<string, string>\n}\n\ninterface TestGradedRunOptions {\n  variantId?: string\n  driver?: SandboxDriver\n  /** Metadata recorded on the Run (codeSha, promptSha, modelFingerprint, seed). */\n  provenance?: Pick<Run, 'codeSha' | 'promptSha' | 'modelFingerprint' | 'seed' | 'envFingerprint'>\n}\n\nexport interface TestGradedRunResult {\n  runId: string\n  scenario: TestGradedScenario\n  harness: SandboxHarnessResult\n  pass: boolean\n  score: number\n  failureClass?: FailureClass\n}\n\nexport async function runTestGradedScenario(\n  scenario: TestGradedScenario,\n  store: TraceStore,\n  options: TestGradedRunOptions = {},\n): Promise<TestGradedRunResult> {\n  const emitter = new TraceEmitter(store)\n  await emitter.startRun({\n    scenarioId: scenario.id,\n    variantId: options.variantId,\n    datasetVersion: scenario.datasetVersion,\n    tags: scenario.tags,\n    ...options.provenance,\n  })\n  const harness = new SandboxHarness(options.driver)\n  const result = await harness.run(scenario.harness, emitter)\n  const threshold = scenario.passThreshold ?? 1.0\n  const pass = result.passed && result.score >= threshold\n  const setupFailed = result.setup !== undefined && result.setup.exitCode !== 0\n  const runFailed = result.run !== undefined && result.run.exitCode !== 0\n  const testFailed = result.test !== undefined && result.test.exitCode !== 0\n  const failureClass: FailureClass | undefined = pass\n    ? 'success'\n    : setupFailed || runFailed\n      ? 'sandbox_failure'\n      : testFailed\n        ? 'format_drift'\n        : 'unknown'\n  await emitter.endRun({\n    pass,\n    score: result.score,\n    failureClass,\n    notes: pass ? undefined : reasonForFailure(result),\n  })\n  return {\n    runId: emitter.runId,\n    scenario,\n    harness: result,\n    pass,\n    score: result.score,\n    failureClass,\n  }\n}\n\nfunction reasonForFailure(result: SandboxHarnessResult): string {\n  if (result.setup && result.setup.exitCode !== 0)\n    return `setup failed: exit ${result.setup.exitCode}`\n  if (result.run && result.run.exitCode !== 0) return `run failed: exit ${result.run.exitCode}`\n  if (result.test) {\n    if (result.test.testsTotal !== undefined) {\n      return `tests: ${result.test.testsPassed ?? 0}/${result.test.testsTotal}`\n    }\n    return `test exit ${result.test.exitCode}`\n  }\n  return 'no test command'\n}\n","/**\n * BuilderSession — ties a builder-of-builders workflow together.\n *\n * Models agent-builder's shape: Project → Chat → Edit → Ship → App →\n * AppAgent. Each layer is a Run (linked via parentRunId). The\n * framework-enforced invariants:\n *\n *   - One Project → many Chats; chatId scopes runs within a project.\n *   - One Chat = one builder Run with `layer='builder'`.\n *   - One Ship = one child Run with `layer='app-build'` + SandboxHarness.\n *   - One AppScenario = one grandchild Run with `layer='app-runtime'`.\n *\n * Consumers obtain a BuilderSession, call `startChat`, drive the\n * builder agent (emitting spans), and call `ship` / `runAppScenario`\n * as the workflow progresses. The session reconstructs itself from\n * trace data via `resume(store, projectId)`.\n */\n\nimport { newRecordId } from '../record-id'\nimport type { HarnessConfig, SandboxDriver, SandboxHarnessResult } from '../sandbox-harness'\nimport { SandboxHarness } from '../sandbox-harness'\nimport type { TestGradedRunResult, TestGradedScenario } from '../test-graded-scenario'\nimport { runTestGradedScenario } from '../test-graded-scenario'\nimport { TraceEmitter } from '../trace/emitter'\nimport type { Run } from '../trace/schema'\nimport type { TraceStore } from '../trace/store'\n\nexport interface BuilderSessionInit {\n  projectId: string\n  chatId?: string\n  /** Free-form: user's task description, project name, etc. Stored on the builder Run. */\n  tags?: Record<string, string>\n}\n\nexport interface ShipOptions {\n  harness: HarnessConfig\n  driver?: SandboxDriver\n  /** scenarioId of this app-build run. Defaults to `${projectId}/build`. */\n  scenarioId?: string\n}\n\nexport interface RunAppScenarioOptions {\n  scenario: TestGradedScenario\n  /** Harness driver override; defaults to the one the session was created with. */\n  driver?: SandboxDriver\n}\n\nexport class BuilderSession {\n  private store: TraceStore\n  private builderEmitter: TraceEmitter\n  readonly projectId: string\n  readonly chatId: string\n  private builderRunId?: string\n  private lastBuildRunId?: string\n  private defaultDriver?: SandboxDriver\n\n  constructor(store: TraceStore, init: BuilderSessionInit, driver?: SandboxDriver) {\n    this.store = store\n    this.projectId = init.projectId\n    this.chatId = init.chatId ?? newRecordId()\n    this.defaultDriver = driver\n    this.builderEmitter = new TraceEmitter(store)\n  }\n\n  /** Start the builder (L0) run for this chat. Returns the runId. */\n  async startChat(scenarioId = `${this.projectId}/chat`): Promise<string> {\n    await this.builderEmitter.startRun({\n      scenarioId,\n      projectId: this.projectId,\n      chatId: this.chatId,\n      layer: 'builder',\n    })\n    this.builderRunId = this.builderEmitter.runId\n    return this.builderRunId\n  }\n\n  /** The emitter for builder-level spans (edits, LLM calls, tool invocations). */\n  get emitter(): TraceEmitter {\n    if (!this.builderRunId) throw new Error('BuilderSession.emitter: call startChat() first')\n    return this.builderEmitter\n  }\n\n  /**\n   * Ship the project's generated app: run the sandbox harness as a child\n   * Run (`layer='app-build'`). Returns the build result + runId.\n   */\n  async ship(options: ShipOptions): Promise<{ runId: string; result: SandboxHarnessResult }> {\n    if (!this.builderRunId) throw new Error('BuilderSession.ship: call startChat() first')\n    const buildEmitter = new TraceEmitter(this.store)\n    await buildEmitter.startRun({\n      scenarioId: options.scenarioId ?? `${this.projectId}/build`,\n      projectId: this.projectId,\n      chatId: this.chatId,\n      parentRunId: this.builderRunId,\n      layer: 'app-build',\n    })\n    const harness = new SandboxHarness(options.driver ?? this.defaultDriver)\n    const result = await harness.run(options.harness, buildEmitter)\n    await buildEmitter.endRun({\n      pass: result.passed,\n      score: result.score,\n      failureClass: result.passed ? 'success' : 'sandbox_failure',\n    })\n    this.lastBuildRunId = buildEmitter.runId\n    return { runId: buildEmitter.runId, result }\n  }\n\n  /**\n   * Run a domain scenario against the just-built app as a grandchild Run\n   * (`layer='app-runtime'`). The `ship` call must precede this so the\n   * parent is set correctly; if no build exists yet the session attaches\n   * directly to the builder run (useful for prototypes).\n   */\n  async runAppScenario(options: RunAppScenarioOptions): Promise<TestGradedRunResult> {\n    const parentRunId = this.lastBuildRunId ?? this.builderRunId\n    if (!parentRunId)\n      throw new Error('BuilderSession.runAppScenario: call startChat() + ship() first')\n    const { scenario, driver } = options\n    const result = await runTestGradedScenario(scenario, this.store, {\n      driver: driver ?? this.defaultDriver,\n      provenance: { codeSha: undefined, promptSha: undefined, modelFingerprint: undefined },\n    })\n    // Attach to the parent chain by updating the stored Run in place.\n    await this.store.updateRun(result.runId, {\n      parentRunId,\n      projectId: this.projectId,\n      chatId: this.chatId,\n      layer: 'app-runtime',\n    })\n    return result\n  }\n\n  /** Record an end-of-chat meta score (judge verdict on whether the builder\n   *  served the user's intent). Accepts a numeric score + optional rationale. */\n  async recordMetaScore(score: number, rationale?: string): Promise<void> {\n    if (!this.builderRunId)\n      throw new Error('BuilderSession.recordMetaScore: call startChat() first')\n    await this.builderEmitter.recordJudge({\n      judgeId: 'builder-meta',\n      targetSpanId: this.builderRunId, // attach to the builder run itself\n      dimension: 'user_intent_satisfaction',\n      score,\n      rationale,\n      name: 'builder-meta',\n    })\n  }\n\n  /** Close the builder Run with a final outcome. */\n  async endChat(outcome: { pass: boolean; score?: number; notes?: string }): Promise<void> {\n    await this.builderEmitter.endRun({\n      pass: outcome.pass,\n      score: outcome.score,\n      notes: outcome.notes,\n    })\n  }\n\n  /**\n   * Inline app-runtime run — for cases where the \"scenario\" isn't a\n   * SWE-bench-style test suite but a live agent interaction (LLM chat,\n   * domain flow). Returns an emitter bound to a fresh Run in the\n   * `app-runtime` layer; caller emits spans inside and calls\n   * `.endRun()` with the final verdict.\n   */\n  async startAppRuntime(scenarioId: string): Promise<TraceEmitter> {\n    const parentRunId = this.lastBuildRunId ?? this.builderRunId\n    if (!parentRunId)\n      throw new Error(\n        'BuilderSession.startAppRuntime: call startChat() + (optionally) ship() first',\n      )\n    const emitter = new TraceEmitter(this.store)\n    await emitter.startRun({\n      scenarioId,\n      projectId: this.projectId,\n      chatId: this.chatId,\n      parentRunId,\n      layer: 'app-runtime',\n    })\n    return emitter\n  }\n\n  /**\n   * Lightweight \"ship marker\" — record an app-build Run with a caller-\n   * provided verdict. Use when there isn't a sandbox harness to run but\n   * you still want to mark the build state at publish time.\n   */\n  async recordShipMarker(args: {\n    pass: boolean\n    score: number\n    scenarioId?: string\n    notes?: string\n  }): Promise<string> {\n    if (!this.builderRunId)\n      throw new Error('BuilderSession.recordShipMarker: call startChat() first')\n    const emitter = new TraceEmitter(this.store)\n    await emitter.startRun({\n      scenarioId: args.scenarioId ?? `${this.projectId}/ship`,\n      projectId: this.projectId,\n      chatId: this.chatId,\n      parentRunId: this.builderRunId,\n      layer: 'app-build',\n    })\n    await emitter.endRun({\n      pass: args.pass,\n      score: args.score,\n      failureClass: args.pass ? 'success' : 'sandbox_failure',\n      notes: args.notes,\n    })\n    this.lastBuildRunId = emitter.runId\n    return emitter.runId\n  }\n\n  get lastBuildRunIdValue(): string | undefined {\n    return this.lastBuildRunId\n  }\n  get builderRunIdValue(): string | undefined {\n    return this.builderRunId\n  }\n}\n\n/**\n * Reconstruct the most recent BuilderSession state for a given project —\n * returns { builderRunId, lastBuildRunId, chatRuns }. For chat-first UIs\n * this is how a resumed session finds its place in the edit history.\n */\nexport async function resumeBuilderSession(\n  store: TraceStore,\n  projectId: string,\n): Promise<{\n  projectId: string\n  chatRuns: Run[]\n  lastBuilderRun?: Run\n  lastBuildRun?: Run\n  lastAppRuntimeRuns: Run[]\n}> {\n  const runs = await store.listRuns({ projectId })\n  const chatRuns = runs\n    .filter((r) => r.layer === 'builder')\n    .sort((a, b) => b.startedAt - a.startedAt)\n  const buildRuns = runs\n    .filter((r) => r.layer === 'app-build')\n    .sort((a, b) => b.startedAt - a.startedAt)\n  const appRuntimeRuns = runs\n    .filter((r) => r.layer === 'app-runtime')\n    .sort((a, b) => b.startedAt - a.startedAt)\n  return {\n    projectId,\n    chatRuns,\n    lastBuilderRun: chatRuns[0],\n    lastBuildRun: buildRuns[0],\n    lastAppRuntimeRuns: appRuntimeRuns,\n  }\n}\n","/**\n * Meta-eval correlation — the highest-leverage signal in the framework.\n *\n * Given a corpus of three-layer project reports, compute how well each\n * pair of layers correlates. The question we care about most:\n *\n *   Does `metaScore` (what the builder thinks it did) predict\n *   `runtimeScore` (what the user actually gets)?\n *\n * If r < ~0.4, the builder's self-scoring is broken — it's optimizing\n * for something other than real-world success. If r > 0.7, meta_score\n * is a usable proxy and can drive CI gates cheaply.\n *\n * Non-parametric rank correlation (Spearman) is also reported because\n * meta scores are often ordinal-ish.\n */\n\nimport { pearsonR, spearmanR } from '../statistics'\nimport type { ThreeLayerProjectReport } from './three-layer-eval'\n\nexport interface LayerCorrelation {\n  n: number\n  pearson: number\n  spearman: number\n}\n\nexport interface CorrelationReport {\n  /** Pairs present in the corpus (layers with ≥ 2 matched data points). */\n  metaVsBuild?: LayerCorrelation\n  metaVsRuntime?: LayerCorrelation\n  buildVsRuntime?: LayerCorrelation\n  /** Number of complete projects (all 3 scores present). */\n  completeProjects: number\n}\n\nexport function correlateLayers(reports: ThreeLayerProjectReport[]): CorrelationReport {\n  const completeProjects = reports.filter((r) => r.complete).length\n  return {\n    metaVsBuild: pairwise(\n      reports,\n      (r) => r.metaScore,\n      (r) => r.buildScore,\n    ),\n    metaVsRuntime: pairwise(\n      reports,\n      (r) => r.metaScore,\n      (r) => r.runtimeScore,\n    ),\n    buildVsRuntime: pairwise(\n      reports,\n      (r) => r.buildScore,\n      (r) => r.runtimeScore,\n    ),\n    completeProjects,\n  }\n}\n\nfunction pairwise(\n  reports: ThreeLayerProjectReport[],\n  a: (r: ThreeLayerProjectReport) => number | null,\n  b: (r: ThreeLayerProjectReport) => number | null,\n): LayerCorrelation | undefined {\n  const xs: number[] = []\n  const ys: number[] = []\n  for (const r of reports) {\n    const x = a(r)\n    const y = b(r)\n    if (x !== null && y !== null && Number.isFinite(x) && Number.isFinite(y)) {\n      xs.push(x)\n      ys.push(y)\n    }\n  }\n  if (xs.length < 2) return undefined\n  return {\n    n: xs.length,\n    pearson: pearsonR(xs, ys),\n    spearman: spearmanR(xs, ys),\n  }\n}\n","/**\n * ProjectRegistry — project-level aggregation over the trace corpus.\n *\n * Thin reader over TraceStore that answers the questions a chat-first,\n * resumable UI needs:\n *   - listProjects() → project IDs with latest activity\n *   - projectTimeline(id) → chats + builds + runtime runs, chronological\n *   - projectChats(id) → chat-level summaries (turn count, outcome)\n *\n * All queries are pure reads; no state duplication.\n */\n\nimport type { Run } from '../trace/schema'\nimport type { TraceStore } from '../trace/store'\n\nexport interface ProjectSummary {\n  projectId: string\n  chatCount: number\n  buildCount: number\n  appRuntimeCount: number\n  lastActivityAt: number\n  latestChatId?: string\n  latestOutcome?: { pass: boolean; score?: number }\n}\n\nexport interface ChatSummary {\n  chatId: string\n  projectId: string\n  builderRunId: string\n  startedAt: number\n  endedAt?: number\n  status: Run['status']\n  outcome?: Run['outcome']\n  /** Counts of spans emitted during the chat. */\n  llmTurns?: number\n  toolCalls?: number\n  buildRunId?: string\n  appRuntimeRunIds: string[]\n}\n\nexport interface ProjectTimelineEntry {\n  run: Run\n  layerBucket: 'chat' | 'build' | 'runtime' | 'other'\n}\n\nexport class ProjectRegistry {\n  constructor(private store: TraceStore) {}\n\n  async listProjects(): Promise<ProjectSummary[]> {\n    const runs = await this.store.listRuns()\n    const byProject = new Map<string, Run[]>()\n    for (const r of runs) {\n      if (!r.projectId) continue\n      const arr = byProject.get(r.projectId) ?? []\n      arr.push(r)\n      byProject.set(r.projectId, arr)\n    }\n    const summaries: ProjectSummary[] = []\n    for (const [projectId, projectRuns] of byProject) {\n      const sorted = projectRuns.slice().sort((a, b) => b.startedAt - a.startedAt)\n      const chats = projectRuns.filter((r) => r.layer === 'builder')\n      const builds = projectRuns.filter((r) => r.layer === 'app-build')\n      const runtimes = projectRuns.filter((r) => r.layer === 'app-runtime')\n      const latest = sorted[0]\n      if (!latest) continue\n      summaries.push({\n        projectId,\n        chatCount: chats.length,\n        buildCount: builds.length,\n        appRuntimeCount: runtimes.length,\n        lastActivityAt: latest.startedAt,\n        latestChatId: chats[0]?.chatId,\n        latestOutcome: latest.outcome\n          ? { pass: latest.outcome.pass ?? false, score: latest.outcome.score }\n          : undefined,\n      })\n    }\n    return summaries.sort((a, b) => b.lastActivityAt - a.lastActivityAt)\n  }\n\n  async projectTimeline(projectId: string): Promise<ProjectTimelineEntry[]> {\n    const runs = await this.store.listRuns({ projectId })\n    const ordered = runs.slice().sort((a, b) => a.startedAt - b.startedAt)\n    return ordered.map((run) => ({\n      run,\n      layerBucket:\n        run.layer === 'builder'\n          ? 'chat'\n          : run.layer === 'app-build'\n            ? 'build'\n            : run.layer === 'app-runtime'\n              ? 'runtime'\n              : 'other',\n    }))\n  }\n\n  async projectChats(projectId: string): Promise<ChatSummary[]> {\n    const builderRuns = (await this.store.listRuns({ projectId, layer: 'builder' })).sort(\n      (a, b) => b.startedAt - a.startedAt,\n    )\n    const childrenFor = async (runId: string) => this.store.listRuns({ parentRunId: runId })\n    const out: ChatSummary[] = []\n    for (const run of builderRuns) {\n      const spans = await this.store.spans({ runId: run.runId })\n      const children = await childrenFor(run.runId)\n      const build = children.find((c) => c.layer === 'app-build')\n      const runtime: string[] = []\n      // Runtime runs can be grandchildren (attached to the build) or siblings\n      // when shipped skipped.\n      if (build) {\n        const grands = await childrenFor(build.runId)\n        for (const g of grands) if (g.layer === 'app-runtime') runtime.push(g.runId)\n      }\n      for (const c of children) if (c.layer === 'app-runtime') runtime.push(c.runId)\n      out.push({\n        chatId: run.chatId ?? run.runId,\n        projectId,\n        builderRunId: run.runId,\n        startedAt: run.startedAt,\n        endedAt: run.endedAt,\n        status: run.status,\n        outcome: run.outcome,\n        llmTurns: spans.filter((s) => s.kind === 'llm').length,\n        toolCalls: spans.filter((s) => s.kind === 'tool').length,\n        buildRunId: build?.runId,\n        appRuntimeRunIds: runtime,\n      })\n    }\n    return out\n  }\n}\n","/**\n * Three-layer evaluation — the canonical scoring breakdown for\n * builder-of-builders workflows.\n *\n *   meta_score:    did the builder understand + satisfy user intent?\n *                  (judge verdict attached to the builder run)\n *   build_score:   did the generated scaffold build + pass its own tests?\n *                  (outcome.score on the app-build child run)\n *   runtime_score: did the generated agent pass its domain scenarios?\n *                  (mean outcome.score over app-runtime grandchild runs)\n *\n * Returns a structured report per project. The cross-layer correlation\n * is the highest-leverage signal the framework computes — if\n * meta_score doesn't predict runtime_score, the builder's self-scoring\n * is broken.\n *\n * Scaffold-only mode: when a project has no `app-runtime` runs (e.g. a\n * scaffold-builder eval that grades compose + build without driving a\n * runtime scenario), `kind` is `'scaffold-only'` and `complete` measures\n * meta + build only. Consumers can tell the two apart without having to\n * interpret null-runtime as either \"not yet computed\" or \"N/A for this\n * project shape\".\n */\n\nimport { judgeSpans } from '../trace/query'\nimport type { Run } from '../trace/schema'\nimport type { TraceStore } from '../trace/store'\n\nexport type ProjectKind = 'full' | 'scaffold-only'\n\nexport interface ThreeLayerProjectReport {\n  projectId: string\n  /**\n   * `'full'` when the project has at least one `app-runtime` run;\n   * `'scaffold-only'` when it only has meta + build layers. Lets\n   * downstream consumers treat a null runtime score as expected\n   * (scaffold-only) vs. missing (full, pipeline broke).\n   */\n  kind: ProjectKind\n  builderRunId?: string\n  /** Judge-verdict score on the builder run (0..1 after normalization). */\n  metaScore: number | null\n  buildRunId?: string\n  /** 0..1 from the sandbox harness (testsPassed / testsTotal). */\n  buildScore: number | null\n  appRuntimeRunIds: string[]\n  /** Mean of outcome.score over app-runtime runs, 0..1. Always null in scaffold-only mode. */\n  runtimeScore: number | null\n  runtimePassRate: number | null\n  /**\n   * Layer-aware completeness:\n   *   - `kind='full'`: all three layers scored\n   *   - `kind='scaffold-only'`: meta + build scored (runtime not applicable)\n   */\n  complete: boolean\n}\n\nexport async function scoreProject(\n  store: TraceStore,\n  projectId: string,\n): Promise<ThreeLayerProjectReport> {\n  const allRuns = await store.listRuns({ projectId })\n  const builder = latestByLayer(allRuns, 'builder')\n  const build = latestByLayer(allRuns, 'app-build')\n  const runtime = allRuns.filter((r) => r.layer === 'app-runtime')\n\n  const metaScore = builder ? await extractMetaScore(store, builder.runId) : null\n  const buildScore = build?.outcome?.score ?? null\n  const runtimeScores = runtime\n    .map((r) => r.outcome?.score)\n    .filter((s): s is number => typeof s === 'number')\n  const runtimeScore =\n    runtimeScores.length > 0\n      ? runtimeScores.reduce((a, b) => a + b, 0) / runtimeScores.length\n      : null\n  const runtimePassed = runtime.filter((r) => r.outcome?.pass === true).length\n  const runtimePassRate = runtime.length > 0 ? runtimePassed / runtime.length : null\n\n  const kind: ProjectKind = runtime.length === 0 ? 'scaffold-only' : 'full'\n  const complete =\n    kind === 'scaffold-only'\n      ? metaScore !== null && buildScore !== null\n      : metaScore !== null && buildScore !== null && runtimeScore !== null\n\n  return {\n    projectId,\n    kind,\n    builderRunId: builder?.runId,\n    metaScore,\n    buildRunId: build?.runId,\n    buildScore,\n    appRuntimeRunIds: runtime.map((r) => r.runId),\n    runtimeScore,\n    runtimePassRate,\n    complete,\n  }\n}\n\n/** Aggregate scoring across every project in a corpus. */\nexport async function scoreAllProjects(store: TraceStore): Promise<ThreeLayerProjectReport[]> {\n  const runs = await store.listRuns()\n  const projectIds = [...new Set(runs.map((r) => r.projectId).filter((p): p is string => !!p))]\n  return Promise.all(projectIds.map((p) => scoreProject(store, p)))\n}\n\nfunction latestByLayer(runs: Run[], layer: Run['layer']): Run | undefined {\n  const filtered = runs.filter((r) => r.layer === layer).sort((a, b) => b.startedAt - a.startedAt)\n  return filtered[0]\n}\n\nasync function extractMetaScore(store: TraceStore, builderRunId: string): Promise<number | null> {\n  const js = await judgeSpans(store, builderRunId)\n  const meta = js.find(\n    (s) => s.judgeId === 'builder-meta' && s.dimension === 'user_intent_satisfaction',\n  )\n  if (!meta) return null\n  // Normalize score to 0..1. Accept 0-1 natively; 0-10 scale is also common.\n  if (meta.score >= 0 && meta.score <= 1) return meta.score\n  if (meta.score >= 0 && meta.score <= 10) return meta.score / 10\n  return null\n}\n"],"mappings":";;;;;;;;;;;;;;;;;;;;AAmKA,IAAa,0BAAb,MAA8D;CAC5D,KAAK;CACL;CACA;CAEA,YAAY,UAA0C,CAAC,GAAG;EACxD,KAAK,aAAa,QAAQ;EAC1B,KAAK,aAAa,QAAQ;CAC5B;CAEA,MAAM,KACJ,OACA,SACA,QACwB;EAIxB,MAAM,eAAe,OAAO,OAAO,KAAK;EAGxC,MAAM,eAAe;GAAE,GAAG,QAAQ;GAAK,GAAI,KAAK,cAAc,CAAC;GAAI,GAAI,OAAO,OAAO,CAAC;EAAG;EACzF,MAAM,SAAS,MAAM,kBAAkB;GACrC;GACA,OAAO;GACP,KAAK;GACL,KAAK;GACL,SAAS;GACT,WAAW,OAAO,aAAa;GAC/B,gBAAgB,OAAO,kBAAkB;EAC3C,CAAC;EAID,MAAM,SACJ,CAAC,OAAO,mBAAmB,UAAU,UAAU,OAAO,aAClD,OAAO,WAAW,MAAM,OAAO,QAAQ,OAAO,QAAQ,OAAO,QAAQ,IACrE,KAAA;EACN,OAAO;GACL;GACA,UAAU,OAAO;GACjB,QAAQ,OAAO;GACf,QAAQ,OAAO,cAAc,OAAO,SAAS,OAAO,cAAc,OAAO;GACzE,QAAQ,OAAO;GACf,YAAY,QAAQ;GACpB,aAAa,QAAQ;GACrB,iBAAiB,OAAO,mBAAmB,KAAA;GAC3C,iBAAiB,OAAO,mBAAmB,KAAA;EAC7C;CACF;AACF;AAkBA,IAAa,iBAAb,MAA4B;CAC1B;CACA,YAAY,SAAwB,IAAI,wBAAwB,GAAG;EACjE,KAAK,SAAS;CAChB;CAEA,MAAM,IAAI,QAAuB,SAAsD;EACrF,MAAM,SAAS,MAAM,QAAQ,QAAQ;GACnC,MAAM,WAAW,KAAK,OAAO,GAAG;GAChC,OAAO,OAAO;GACd,SAAS;IAAC,OAAO;IAAc,OAAO;IAAY,OAAO;GAAW,CAAC,CAClE,OAAO,OAAO,CAAC,CACf,KAAK,MAAM;EAChB,CAAC;EACD,MAAM,SAA+B;GAAE,QAAQ;GAAO,aAAa;GAAG,OAAO;EAAE;EAC/E,IAAI;GACF,IAAI,OAAO,cAAc;IACvB,OAAO,QAAQ,MAAM,KAAK,OAAO,KAAK,SAAS,OAAO,cAAc,MAAM;IAC1E,OAAO,eAAe,OAAO,MAAM;IACnC,IAAI,OAAO,MAAM,mBAAmB,OAAO,MAAM,aAAa,GAAG;KAC/D,MAAM,SAAS,OAAO,MAAM,kBACxB,yBAAyB,OAAO,MAAM,OAAO,MAC7C,sBAAsB,OAAO,MAAM,SAAS;KAChD,MAAM,OAAO,KAAK,QAAQ;MACxB,UAAU,OAAO,MAAM;MACvB,QAAQ,OAAO;KACjB,CAAyB;KACzB,OAAO;IACT;GACF;GACA,IAAI,OAAO,YAAY;IACrB,OAAO,MAAM,MAAM,KAAK,OAAO,KAAK,OAAO,OAAO,YAAY,MAAM;IACpE,OAAO,eAAe,OAAO,IAAI;IACjC,IAAI,OAAO,IAAI,mBAAmB,OAAO,IAAI,aAAa,GAAG;KAC3D,MAAM,SAAS,OAAO,IAAI,kBACtB,uBAAuB,OAAO,IAAI,OAAO,MACzC,oBAAoB,OAAO,IAAI,SAAS;KAC5C,MAAM,OAAO,KAAK,QAAQ;MACxB,UAAU,OAAO,IAAI;MACrB,QAAQ,OAAO;KACjB,CAAyB;KACzB,OAAO;IACT;GACF;GACA,IAAI,OAAO,aAAa;IACtB,OAAO,OAAO,MAAM,KAAK,OAAO,KAAK,QAAQ,OAAO,aAAa,MAAM;IACvE,OAAO,eAAe,OAAO,KAAK;IAGlC,MAAM,SAAS,CAAC,OAAO,KAAK,mBAAmB,OAAO,KAAK,aAAa;IACxE,OAAO,SAAS;IAChB,IAAI,OAAO,KAAK,eAAe,KAAA,KAAa,OAAO,KAAK,aAAa,GACnE,OAAO,SAAS,OAAO,KAAK,eAAe,KAAK,OAAO,KAAK;SAE5D,OAAO,QAAQ,SAAS,IAAI;IAE9B,MAAM,OAAO,IAAI;KACf,UAAU,OAAO,KAAK;KACtB,YAAY,OAAO,KAAK;KACxB,aAAa,OAAO,KAAK;KACzB,QAAQ,OAAO;KACf,QAAQ,SAAS,OAAO;IAC1B,CAAyB;GAC3B,OAAO;IACL,OAAO,SAAS;IAChB,OAAO,QAAQ;IACf,MAAM,OAAO,IAAI,EAAE,QAAQ,OAAO,YAAY,CAAyB;GACzE;EACF,SAAS,KAAK;GACZ,MAAM,OAAO,KAAK,eAAe,QAAQ,MAAM,OAAO,GAAG,CAAC;GAC1D,MAAM;EACR;EACA,OAAO;CACT;AACF;;AClQA,eAAsB,sBACpB,UACA,OACA,UAAgC,CAAC,GACH;CAC9B,MAAM,UAAU,IAAI,aAAa,KAAK;CACtC,MAAM,QAAQ,SAAS;EACrB,YAAY,SAAS;EACrB,WAAW,QAAQ;EACnB,gBAAgB,SAAS;EACzB,MAAM,SAAS;EACf,GAAG,QAAQ;CACb,CAAC;CAED,MAAM,SAAS,MAAM,IADD,eAAe,QAAQ,MAChB,CAAC,CAAC,IAAI,SAAS,SAAS,OAAO;CAC1D,MAAM,YAAY,SAAS,iBAAiB;CAC5C,MAAM,OAAO,OAAO,UAAU,OAAO,SAAS;CAC9C,MAAM,cAAc,OAAO,UAAU,KAAA,KAAa,OAAO,MAAM,aAAa;CAC5E,MAAM,YAAY,OAAO,QAAQ,KAAA,KAAa,OAAO,IAAI,aAAa;CACtE,MAAM,aAAa,OAAO,SAAS,KAAA,KAAa,OAAO,KAAK,aAAa;CACzE,MAAM,eAAyC,OAC3C,YACA,eAAe,YACb,oBACA,aACE,iBACA;CACR,MAAM,QAAQ,OAAO;EACnB;EACA,OAAO,OAAO;EACd;EACA,OAAO,OAAO,KAAA,IAAY,iBAAiB,MAAM;CACnD,CAAC;CACD,OAAO;EACL,OAAO,QAAQ;EACf;EACA,SAAS;EACT;EACA,OAAO,OAAO;EACd;CACF;AACF;AAEA,SAAS,iBAAiB,QAAsC;CAC9D,IAAI,OAAO,SAAS,OAAO,MAAM,aAAa,GAC5C,OAAO,sBAAsB,OAAO,MAAM;CAC5C,IAAI,OAAO,OAAO,OAAO,IAAI,aAAa,GAAG,OAAO,oBAAoB,OAAO,IAAI;CACnF,IAAI,OAAO,MAAM;EACf,IAAI,OAAO,KAAK,eAAe,KAAA,GAC7B,OAAO,UAAU,OAAO,KAAK,eAAe,EAAE,GAAG,OAAO,KAAK;EAE/D,OAAO,aAAa,OAAO,KAAK;CAClC;CACA,OAAO;AACT;;;;;;;;;;;;;;;;;;;ACtDA,IAAa,iBAAb,MAA4B;CAC1B;CACA;CACA;CACA;CACA;CACA;CACA;CAEA,YAAY,OAAmB,MAA0B,QAAwB;EAC/E,KAAK,QAAQ;EACb,KAAK,YAAY,KAAK;EACtB,KAAK,SAAS,KAAK,UAAU,YAAY;EACzC,KAAK,gBAAgB;EACrB,KAAK,iBAAiB,IAAI,aAAa,KAAK;CAC9C;;CAGA,MAAM,UAAU,aAAa,GAAG,KAAK,UAAU,QAAyB;EACtE,MAAM,KAAK,eAAe,SAAS;GACjC;GACA,WAAW,KAAK;GAChB,QAAQ,KAAK;GACb,OAAO;EACT,CAAC;EACD,KAAK,eAAe,KAAK,eAAe;EACxC,OAAO,KAAK;CACd;;CAGA,IAAI,UAAwB;EAC1B,IAAI,CAAC,KAAK,cAAc,MAAM,IAAI,MAAM,gDAAgD;EACxF,OAAO,KAAK;CACd;;;;;CAMA,MAAM,KAAK,SAAgF;EACzF,IAAI,CAAC,KAAK,cAAc,MAAM,IAAI,MAAM,6CAA6C;EACrF,MAAM,eAAe,IAAI,aAAa,KAAK,KAAK;EAChD,MAAM,aAAa,SAAS;GAC1B,YAAY,QAAQ,cAAc,GAAG,KAAK,UAAU;GACpD,WAAW,KAAK;GAChB,QAAQ,KAAK;GACb,aAAa,KAAK;GAClB,OAAO;EACT,CAAC;EAED,MAAM,SAAS,MAAM,IADD,eAAe,QAAQ,UAAU,KAAK,aAC/B,CAAC,CAAC,IAAI,QAAQ,SAAS,YAAY;EAC9D,MAAM,aAAa,OAAO;GACxB,MAAM,OAAO;GACb,OAAO,OAAO;GACd,cAAc,OAAO,SAAS,YAAY;EAC5C,CAAC;EACD,KAAK,iBAAiB,aAAa;EACnC,OAAO;GAAE,OAAO,aAAa;GAAO;EAAO;CAC7C;;;;;;;CAQA,MAAM,eAAe,SAA8D;EACjF,MAAM,cAAc,KAAK,kBAAkB,KAAK;EAChD,IAAI,CAAC,aACH,MAAM,IAAI,MAAM,gEAAgE;EAClF,MAAM,EAAE,UAAU,WAAW;EAC7B,MAAM,SAAS,MAAM,sBAAsB,UAAU,KAAK,OAAO;GAC/D,QAAQ,UAAU,KAAK;GACvB,YAAY;IAAE,SAAS,KAAA;IAAW,WAAW,KAAA;IAAW,kBAAkB,KAAA;GAAU;EACtF,CAAC;EAED,MAAM,KAAK,MAAM,UAAU,OAAO,OAAO;GACvC;GACA,WAAW,KAAK;GAChB,QAAQ,KAAK;GACb,OAAO;EACT,CAAC;EACD,OAAO;CACT;;;CAIA,MAAM,gBAAgB,OAAe,WAAmC;EACtE,IAAI,CAAC,KAAK,cACR,MAAM,IAAI,MAAM,wDAAwD;EAC1E,MAAM,KAAK,eAAe,YAAY;GACpC,SAAS;GACT,cAAc,KAAK;GACnB,WAAW;GACX;GACA;GACA,MAAM;EACR,CAAC;CACH;;CAGA,MAAM,QAAQ,SAA2E;EACvF,MAAM,KAAK,eAAe,OAAO;GAC/B,MAAM,QAAQ;GACd,OAAO,QAAQ;GACf,OAAO,QAAQ;EACjB,CAAC;CACH;;;;;;;;CASA,MAAM,gBAAgB,YAA2C;EAC/D,MAAM,cAAc,KAAK,kBAAkB,KAAK;EAChD,IAAI,CAAC,aACH,MAAM,IAAI,MACR,8EACF;EACF,MAAM,UAAU,IAAI,aAAa,KAAK,KAAK;EAC3C,MAAM,QAAQ,SAAS;GACrB;GACA,WAAW,KAAK;GAChB,QAAQ,KAAK;GACb;GACA,OAAO;EACT,CAAC;EACD,OAAO;CACT;;;;;;CAOA,MAAM,iBAAiB,MAKH;EAClB,IAAI,CAAC,KAAK,cACR,MAAM,IAAI,MAAM,yDAAyD;EAC3E,MAAM,UAAU,IAAI,aAAa,KAAK,KAAK;EAC3C,MAAM,QAAQ,SAAS;GACrB,YAAY,KAAK,cAAc,GAAG,KAAK,UAAU;GACjD,WAAW,KAAK;GAChB,QAAQ,KAAK;GACb,aAAa,KAAK;GAClB,OAAO;EACT,CAAC;EACD,MAAM,QAAQ,OAAO;GACnB,MAAM,KAAK;GACX,OAAO,KAAK;GACZ,cAAc,KAAK,OAAO,YAAY;GACtC,OAAO,KAAK;EACd,CAAC;EACD,KAAK,iBAAiB,QAAQ;EAC9B,OAAO,QAAQ;CACjB;CAEA,IAAI,sBAA0C;EAC5C,OAAO,KAAK;CACd;CACA,IAAI,oBAAwC;EAC1C,OAAO,KAAK;CACd;AACF;;;;;;AAOA,eAAsB,qBACpB,OACA,WAOC;CACD,MAAM,OAAO,MAAM,MAAM,SAAS,EAAE,UAAU,CAAC;CAC/C,MAAM,WAAW,KACd,QAAQ,MAAM,EAAE,UAAU,SAAS,CAAC,CACpC,MAAM,GAAG,MAAM,EAAE,YAAY,EAAE,SAAS;CAC3C,MAAM,YAAY,KACf,QAAQ,MAAM,EAAE,UAAU,WAAW,CAAC,CACtC,MAAM,GAAG,MAAM,EAAE,YAAY,EAAE,SAAS;CAC3C,MAAM,iBAAiB,KACpB,QAAQ,MAAM,EAAE,UAAU,aAAa,CAAC,CACxC,MAAM,GAAG,MAAM,EAAE,YAAY,EAAE,SAAS;CAC3C,OAAO;EACL;EACA;EACA,gBAAgB,SAAS;EACzB,cAAc,UAAU;EACxB,oBAAoB;CACtB;AACF;;;;;;;;;;;;;;;;;;ACxNA,SAAgB,gBAAgB,SAAuD;CACrF,MAAM,mBAAmB,QAAQ,QAAQ,MAAM,EAAE,QAAQ,CAAC,CAAC;CAC3D,OAAO;EACL,aAAa,SACX,UACC,MAAM,EAAE,YACR,MAAM,EAAE,UACX;EACA,eAAe,SACb,UACC,MAAM,EAAE,YACR,MAAM,EAAE,YACX;EACA,gBAAgB,SACd,UACC,MAAM,EAAE,aACR,MAAM,EAAE,YACX;EACA;CACF;AACF;AAEA,SAAS,SACP,SACA,GACA,GAC8B;CAC9B,MAAM,KAAe,CAAC;CACtB,MAAM,KAAe,CAAC;CACtB,KAAK,MAAM,KAAK,SAAS;EACvB,MAAM,IAAI,EAAE,CAAC;EACb,MAAM,IAAI,EAAE,CAAC;EACb,IAAI,MAAM,QAAQ,MAAM,QAAQ,OAAO,SAAS,CAAC,KAAK,OAAO,SAAS,CAAC,GAAG;GACxE,GAAG,KAAK,CAAC;GACT,GAAG,KAAK,CAAC;EACX;CACF;CACA,IAAI,GAAG,SAAS,GAAG,OAAO,KAAA;CAC1B,OAAO;EACL,GAAG,GAAG;EACN,SAAS,SAAS,IAAI,EAAE;EACxB,UAAU,UAAU,IAAI,EAAE;CAC5B;AACF;;ACjCA,IAAa,kBAAb,MAA6B;CACP;CAApB,YAAY,OAA2B;EAAnB,KAAA,QAAA;CAAoB;CAExC,MAAM,eAA0C;EAC9C,MAAM,OAAO,MAAM,KAAK,MAAM,SAAS;EACvC,MAAM,4BAAY,IAAI,IAAmB;EACzC,KAAK,MAAM,KAAK,MAAM;GACpB,IAAI,CAAC,EAAE,WAAW;GAClB,MAAM,MAAM,UAAU,IAAI,EAAE,SAAS,KAAK,CAAC;GAC3C,IAAI,KAAK,CAAC;GACV,UAAU,IAAI,EAAE,WAAW,GAAG;EAChC;EACA,MAAM,YAA8B,CAAC;EACrC,KAAK,MAAM,CAAC,WAAW,gBAAgB,WAAW;GAChD,MAAM,SAAS,YAAY,MAAM,CAAC,CAAC,MAAM,GAAG,MAAM,EAAE,YAAY,EAAE,SAAS;GAC3E,MAAM,QAAQ,YAAY,QAAQ,MAAM,EAAE,UAAU,SAAS;GAC7D,MAAM,SAAS,YAAY,QAAQ,MAAM,EAAE,UAAU,WAAW;GAChE,MAAM,WAAW,YAAY,QAAQ,MAAM,EAAE,UAAU,aAAa;GACpE,MAAM,SAAS,OAAO;GACtB,IAAI,CAAC,QAAQ;GACb,UAAU,KAAK;IACb;IACA,WAAW,MAAM;IACjB,YAAY,OAAO;IACnB,iBAAiB,SAAS;IAC1B,gBAAgB,OAAO;IACvB,cAAc,MAAM,EAAE,EAAE;IACxB,eAAe,OAAO,UAClB;KAAE,MAAM,OAAO,QAAQ,QAAQ;KAAO,OAAO,OAAO,QAAQ;IAAM,IAClE,KAAA;GACN,CAAC;EACH;EACA,OAAO,UAAU,MAAM,GAAG,MAAM,EAAE,iBAAiB,EAAE,cAAc;CACrE;CAEA,MAAM,gBAAgB,WAAoD;EAGxE,QADgB,MADG,KAAK,MAAM,SAAS,EAAE,UAAU,CAAC,EAAA,CAC/B,MAAM,CAAC,CAAC,MAAM,GAAG,MAAM,EAAE,YAAY,EAAE,SAC/C,CAAC,CAAC,KAAK,SAAS;GAC3B;GACA,aACE,IAAI,UAAU,YACV,SACA,IAAI,UAAU,cACZ,UACA,IAAI,UAAU,gBACZ,YACA;EACZ,EAAE;CACJ;CAEA,MAAM,aAAa,WAA2C;EAC5D,MAAM,eAAe,MAAM,KAAK,MAAM,SAAS;GAAE;GAAW,OAAO;EAAU,CAAC,EAAA,CAAG,MAC9E,GAAG,MAAM,EAAE,YAAY,EAAE,SAC5B;EACA,MAAM,cAAc,OAAO,UAAkB,KAAK,MAAM,SAAS,EAAE,aAAa,MAAM,CAAC;EACvF,MAAM,MAAqB,CAAC;EAC5B,KAAK,MAAM,OAAO,aAAa;GAC7B,MAAM,QAAQ,MAAM,KAAK,MAAM,MAAM,EAAE,OAAO,IAAI,MAAM,CAAC;GACzD,MAAM,WAAW,MAAM,YAAY,IAAI,KAAK;GAC5C,MAAM,QAAQ,SAAS,MAAM,MAAM,EAAE,UAAU,WAAW;GAC1D,MAAM,UAAoB,CAAC;GAG3B,IAAI,OAAO;IACT,MAAM,SAAS,MAAM,YAAY,MAAM,KAAK;IAC5C,KAAK,MAAM,KAAK,QAAQ,IAAI,EAAE,UAAU,eAAe,QAAQ,KAAK,EAAE,KAAK;GAC7E;GACA,KAAK,MAAM,KAAK,UAAU,IAAI,EAAE,UAAU,eAAe,QAAQ,KAAK,EAAE,KAAK;GAC7E,IAAI,KAAK;IACP,QAAQ,IAAI,UAAU,IAAI;IAC1B;IACA,cAAc,IAAI;IAClB,WAAW,IAAI;IACf,SAAS,IAAI;IACb,QAAQ,IAAI;IACZ,SAAS,IAAI;IACb,UAAU,MAAM,QAAQ,MAAM,EAAE,SAAS,KAAK,CAAC,CAAC;IAChD,WAAW,MAAM,QAAQ,MAAM,EAAE,SAAS,MAAM,CAAC,CAAC;IAClD,YAAY,OAAO;IACnB,kBAAkB;GACpB,CAAC;EACH;EACA,OAAO;CACT;AACF;;;;;;;;;;;;;;;;;;;;;;;;;ACzEA,eAAsB,aACpB,OACA,WACkC;CAClC,MAAM,UAAU,MAAM,MAAM,SAAS,EAAE,UAAU,CAAC;CAClD,MAAM,UAAU,cAAc,SAAS,SAAS;CAChD,MAAM,QAAQ,cAAc,SAAS,WAAW;CAChD,MAAM,UAAU,QAAQ,QAAQ,MAAM,EAAE,UAAU,aAAa;CAE/D,MAAM,YAAY,UAAU,MAAM,iBAAiB,OAAO,QAAQ,KAAK,IAAI;CAC3E,MAAM,aAAa,OAAO,SAAS,SAAS;CAC5C,MAAM,gBAAgB,QACnB,KAAK,MAAM,EAAE,SAAS,KAAK,CAAC,CAC5B,QAAQ,MAAmB,OAAO,MAAM,QAAQ;CACnD,MAAM,eACJ,cAAc,SAAS,IACnB,cAAc,QAAQ,GAAG,MAAM,IAAI,GAAG,CAAC,IAAI,cAAc,SACzD;CACN,MAAM,gBAAgB,QAAQ,QAAQ,MAAM,EAAE,SAAS,SAAS,IAAI,CAAC,CAAC;CACtE,MAAM,kBAAkB,QAAQ,SAAS,IAAI,gBAAgB,QAAQ,SAAS;CAE9E,MAAM,OAAoB,QAAQ,WAAW,IAAI,kBAAkB;CACnE,MAAM,WACJ,SAAS,kBACL,cAAc,QAAQ,eAAe,OACrC,cAAc,QAAQ,eAAe,QAAQ,iBAAiB;CAEpE,OAAO;EACL;EACA;EACA,cAAc,SAAS;EACvB;EACA,YAAY,OAAO;EACnB;EACA,kBAAkB,QAAQ,KAAK,MAAM,EAAE,KAAK;EAC5C;EACA;EACA;CACF;AACF;;AAGA,eAAsB,iBAAiB,OAAuD;CAC5F,MAAM,OAAO,MAAM,MAAM,SAAS;CAClC,MAAM,aAAa,CAAC,GAAG,IAAI,IAAI,KAAK,KAAK,MAAM,EAAE,SAAS,CAAC,CAAC,QAAQ,MAAmB,CAAC,CAAC,CAAC,CAAC,CAAC;CAC5F,OAAO,QAAQ,IAAI,WAAW,KAAK,MAAM,aAAa,OAAO,CAAC,CAAC,CAAC;AAClE;AAEA,SAAS,cAAc,MAAa,OAAsC;CAExE,OADiB,KAAK,QAAQ,MAAM,EAAE,UAAU,KAAK,CAAC,CAAC,MAAM,GAAG,MAAM,EAAE,YAAY,EAAE,SACxE,CAAC,CAAC;AAClB;AAEA,eAAe,iBAAiB,OAAmB,cAA8C;CAE/F,MAAM,QAAO,MADI,WAAW,OAAO,YAAY,EAAA,CAC/B,MACb,MAAM,EAAE,YAAY,kBAAkB,EAAE,cAAc,0BACzD;CACA,IAAI,CAAC,MAAM,OAAO;CAElB,IAAI,KAAK,SAAS,KAAK,KAAK,SAAS,GAAG,OAAO,KAAK;CACpD,IAAI,KAAK,SAAS,KAAK,KAAK,SAAS,IAAI,OAAO,KAAK,QAAQ;CAC7D,OAAO;AACT"}

package/dist/{calibration-CTwxpJ-t.js → calibration-6usfVYYo.js}

@@ -1,6 +1,6 @@
-import { s as runMetricExtractor } from "./query-B_RZ36an.js";
-import { i as reduceOutcomeMetric, t as assertUniqueObservationIds } from "./outcome-observations-FLmRpgYa.js";
-//#region src/meta-eval/calibration.ts
+import { s as runMetricExtractor } from "./query-CGXzO9d4.js";
+import { i as reduceOutcomeMetric, t as assertUniqueObservationIds } from "./outcome-observations-D4snOtUb.js";
+// src/meta-eval/calibration.ts
 /**
 * Calibration curve — binned "if eval says X, what does reality show?"
 *
@@ -191,7 +191,6 @@ function validateCalibrationRequest(evalMetric, outcomeMetric, options) {
 	if (options.binning !== void 0 && !["equal-width", "equal-frequency"].includes(options.binning)) throw new Error("calibration binning must be equal-width or equal-frequency");
 	if (options.range !== void 0 && (!Number.isFinite(options.range.lo) || !Number.isFinite(options.range.hi) || !Number.isFinite(options.range.hi - options.range.lo) || options.range.hi < options.range.lo)) throw new Error("calibration range must have finite ordered bounds");
 }
-//#endregion
 export { rankedProbabilityScore as a, multiclassBrierScore as i, calibrationCurve as n, calibrationFromPairs as r, brierScore as t };
 
-//# sourceMappingURL=calibration-CTwxpJ-t.js.map
+//# sourceMappingURL=calibration-6usfVYYo.js.map

package/dist/{calibration-CTwxpJ-t.js.map → calibration-6usfVYYo.js.map}

@@ -1 +1 @@
-{"version":3,"file":"calibration-CTwxpJ-t.js","names":[],"sources":["../src/meta-eval/calibration.ts"],"sourcesContent":["/**\n * Calibration curve — binned \"if eval says X, what does reality show?\"\n *\n * Companion to correlationStudy. Raw correlation is a single number;\n * the calibration curve shows *where* the eval is well-calibrated vs\n * overconfident / underconfident. Buckets the eval metric, computes\n * mean outcome per bucket, reports expected-calibration-error (ECE).\n */\n\nimport { runMetricExtractor } from '../trace/query'\nimport type { TraceStore } from '../trace/store'\nimport type { EvalMetricSpec } from './correlation-study'\nimport { assertUniqueObservationIds, reduceOutcomeMetric } from './outcome-observations'\nimport type { DeploymentOutcome, OutcomeStore } from './outcome-store'\n\nexport interface CalibrationBin {\n  lower: number\n  upper: number\n  n: number\n  evalMean: number\n  outcomeMean: number\n  /** |outcomeMean − evalMean|; contributes to ECE weighted by n/total. */\n  gap: number\n}\n\nexport interface CalibrationReport {\n  evalMetric: string\n  outcomeMetric: string\n  n: number\n  bins: CalibrationBin[]\n  /** Expected Calibration Error — Σ (n_i/N) × |outcomeMean_i − evalMean_i|. */\n  ece: number\n  /** Largest observed difference between a bin's mean score and mean outcome. */\n  maxGap: number\n}\n\nexport interface CalibrationOptions {\n  /** Positive integer; empty bins are omitted. Default 10. */\n  bins?: number\n  /** Equal-width (fixed bin edges) or equal-frequency (quantile bins). */\n  binning?: 'equal-width' | 'equal-frequency'\n  /** Clip eval values to [lo, hi] before binning. */\n  range?: { lo: number; hi: number }\n}\n\nexport interface CalibrationPair {\n  evalScore: number\n  outcome: number\n}\n\nexport async function calibrationCurve(\n  traceStore: TraceStore,\n  outcomeStore: OutcomeStore,\n  evalMetric: EvalMetricSpec,\n  outcomeMetric: string,\n  options: CalibrationOptions = {},\n): Promise<CalibrationReport | null> {\n  const settings = {\n    ...options,\n    range: options.range === undefined ? undefined : { ...options.range },\n  }\n  validateCalibrationRequest(evalMetric.id, outcomeMetric, settings)\n  const extract = evalMetric.extract ?? runMetricExtractor(evalMetric.id)\n  const metricId = evalMetric.id\n  const runs = await traceStore.listRuns()\n  assertUniqueObservationIds(\n    runs.map((run) => run.runId),\n    'runId',\n  )\n  const outcomes = await outcomeStore.list()\n  const byRun = new Map<string, DeploymentOutcome[]>()\n  for (const o of outcomes) {\n    const arr = byRun.get(o.runId) ?? []\n    arr.push(o)\n    byRun.set(o.runId, arr)\n  }\n\n  const pairs: Array<{ x: number; y: number }> = []\n  for (const run of runs) {\n    const os = byRun.get(run.runId)\n    if (!os?.length) continue\n    const x = await extract(run, traceStore)\n    if (x === null || !Number.isFinite(x)) continue\n    const y = reduceOutcomeMetric(os, outcomeMetric, 'latest')\n    if (y === null) continue\n    pairs.push({ x, y })\n  }\n  if (pairs.length < 2) return null\n\n  return calibrationFromPairs(\n    pairs.map((p) => ({ evalScore: p.x, outcome: p.y })),\n    metricId,\n    outcomeMetric,\n    settings,\n  )\n}\n\n/** Measure already joined observations without constructing trace and outcome stores. */\nexport function calibrationFromPairs(\n  inputPairs: readonly CalibrationPair[],\n  evalMetric: string,\n  outcomeMetric: string,\n  options: CalibrationOptions = {},\n): CalibrationReport | null {\n  validateCalibrationRequest(evalMetric, outcomeMetric, options)\n  for (const [index, pair] of inputPairs.entries()) {\n    if (\n      pair === null ||\n      typeof pair !== 'object' ||\n      !Number.isFinite(pair.evalScore) ||\n      !Number.isFinite(pair.outcome)\n    ) {\n      throw new Error(`calibration pair ${index} must contain finite evalScore and outcome values`)\n    }\n  }\n  const pairs = inputPairs\n  if (pairs.length < 2) return null\n\n  const numBins = options.bins ?? 10\n  const binning = options.binning ?? 'equal-width'\n  const xs = pairs.map((p) => p.evalScore)\n  const lo = options.range?.lo ?? Math.min(...xs)\n  const hi = options.range?.hi ?? Math.max(...xs)\n  const span = hi - lo\n  if (!Number.isFinite(span)) throw new Error('calibration range span must be finite')\n  const clipped = pairs.map((pair) => ({\n    ...pair,\n    evalScore: Math.min(hi, Math.max(lo, pair.evalScore)),\n  }))\n\n  const bins: CalibrationBin[] = []\n  if (span === 0 || clipped.every((pair) => pair.evalScore === clipped[0]!.evalScore)) {\n    bins.push(toBin(clipped))\n  } else if (binning === 'equal-frequency') {\n    const sorted = [...clipped].sort((a, b) => a.evalScore - b.evalScore)\n    const count = Math.min(numBins, sorted.length)\n    for (let i = 0; i < count; i++) {\n      const start = Math.floor((i * sorted.length) / count)\n      const end = Math.floor(((i + 1) * sorted.length) / count)\n      bins.push(toBin(sorted.slice(start, end)))\n    }\n  } else {\n    const groups = new Map<number, CalibrationPair[]>()\n    for (const pair of clipped) {\n      const index = Math.min(numBins - 1, Math.floor(((pair.evalScore - lo) / span) * numBins))\n      const group = groups.get(index) ?? []\n      group.push(pair)\n      groups.set(index, group)\n    }\n    for (const [index, chunk] of [...groups].sort(([a], [b]) => a - b)) {\n      bins.push(toBin(chunk, lo + span * (index / numBins), lo + span * ((index + 1) / numBins)))\n    }\n  }\n\n  const total = bins.reduce((a, b) => a + b.n, 0)\n  const ece = bins.reduce((a, b) => a + (b.n / total) * b.gap, 0)\n  const maxGap = bins.reduce((a, b) => Math.max(a, b.gap), 0)\n\n  return { evalMetric, outcomeMetric, n: pairs.length, bins, ece, maxGap }\n}\n\n/**\n * Mean squared error of probability forecasts against binary outcomes, Σ(p − y)²/N; always\n * forecasting 0.5 scores 0.25. For the reliability table and ECE use `calibrationFromPairs`\n * with `range: { lo: 0, hi: 1 }` (`binning: 'equal-frequency'` gives equal-mass bins).\n */\nexport function brierScore(pairs: readonly CalibrationPair[]): number {\n  if (!pairs.length) throw new Error('Brier score requires at least one pair')\n  let total = 0\n  for (const [index, pair] of pairs.entries()) {\n    if (\n      pair === null ||\n      typeof pair !== 'object' ||\n      !(pair.evalScore >= 0 && pair.evalScore <= 1) ||\n      (pair.outcome !== 0 && pair.outcome !== 1)\n    ) {\n      throw new Error(`Brier pair ${index} needs a probability in [0, 1] and an outcome of 0 or 1`)\n    }\n    total += (pair.evalScore - pair.outcome) ** 2\n  }\n  return total / pairs.length\n}\n\n/** One categorical forecast: a distribution over named outcomes and the outcome that occurred. */\nexport interface CategoricalForecast {\n  /** Probability per outcome name. Must sum to 1 within 1e-6 and include `outcome`. */\n  probabilities: Readonly<Record<string, number>>\n  outcome: string\n}\n\n/**\n * Multiclass Brier score in its sum form, mean over forecasts of Σₖ(pₖ − yₖ)², range [0, 2].\n * With two outcomes it is twice the binary `brierScore`. A uniform forecast over K outcomes\n * scores 1 − 1/K. The distribution is not renormalized here: repair provider rounding before\n * scoring (`normalizeSystemOneResult` in `/systemone/protocol`), so a defective answer stays\n * visible instead of being scored as if it were valid.\n */\nexport function multiclassBrierScore(forecasts: readonly CategoricalForecast[]): number {\n  if (!forecasts.length) throw new Error('multiclass Brier score requires at least one forecast')\n  let total = 0\n  for (const [index, forecast] of forecasts.entries()) {\n    const entries = categoricalEntries(forecast, index, 'multiclass Brier')\n    for (const [name, probability] of entries) {\n      total += (probability - (name === forecast.outcome ? 1 : 0)) ** 2\n    }\n  }\n  return total / forecasts.length\n}\n\n/** One ordinal forecast: probabilities in level order, lowest level first. */\nexport interface OrdinalForecast {\n  probabilities: readonly number[]\n  /** Index into `probabilities` of the level that occurred. */\n  outcome: number\n}\n\n/**\n * Ranked probability score, mean over forecasts of Σₖ₌₁ᴷ⁻¹(Fₖ − Oₖ)² / (K − 1), where F and O are\n * the cumulative forecast and observation, range [0, 1]. Unlike multiclass Brier it charges\n * less for mass on a level next to the outcome than for mass far from it, which is the right\n * penalty for score questions. With two levels it equals the binary Brier score.\n */\nexport function rankedProbabilityScore(forecasts: readonly OrdinalForecast[]): number {\n  if (!forecasts.length) throw new Error('ranked probability score requires at least one forecast')\n  let total = 0\n  for (const [index, forecast] of forecasts.entries()) {\n    const probabilities = forecast?.probabilities\n    if (!Array.isArray(probabilities) || probabilities.length < 2) {\n      throw new Error(`RPS forecast ${index} needs at least two level probabilities`)\n    }\n    assertDistribution(probabilities, `RPS forecast ${index}`)\n    if (\n      !Number.isInteger(forecast.outcome) ||\n      forecast.outcome < 0 ||\n      forecast.outcome >= probabilities.length\n    ) {\n      throw new Error(\n        `RPS forecast ${index} outcome must index one of its ${probabilities.length} levels`,\n      )\n    }\n    let cumulative = 0\n    let sum = 0\n    for (let level = 0; level < probabilities.length - 1; level++) {\n      cumulative += probabilities[level]!\n      sum += (cumulative - (level >= forecast.outcome ? 1 : 0)) ** 2\n    }\n    total += sum / (probabilities.length - 1)\n  }\n  return total / forecasts.length\n}\n\nfunction categoricalEntries(\n  forecast: CategoricalForecast,\n  index: number,\n  label: string,\n): [string, number][] {\n  if (forecast === null || typeof forecast !== 'object' || typeof forecast.outcome !== 'string') {\n    throw new Error(`${label} forecast ${index} needs probabilities and a string outcome`)\n  }\n  const probabilities = forecast.probabilities\n  if (probabilities === null || typeof probabilities !== 'object' || Array.isArray(probabilities)) {\n    throw new Error(`${label} forecast ${index} probabilities must be an object keyed by outcome`)\n  }\n  const entries = Object.entries(probabilities)\n  if (entries.length < 2) throw new Error(`${label} forecast ${index} needs at least two outcomes`)\n  if (!entries.some(([name]) => name === forecast.outcome)) {\n    throw new Error(\n      `${label} forecast ${index} has no probability for outcome ${JSON.stringify(forecast.outcome)}`,\n    )\n  }\n  assertDistribution(\n    entries.map(([, probability]) => probability),\n    `${label} forecast ${index}`,\n  )\n  return entries\n}\n\nfunction assertDistribution(probabilities: readonly number[], label: string): void {\n  let sum = 0\n  for (const probability of probabilities) {\n    if (!(probability >= 0 && probability <= 1)) {\n      throw new Error(`${label} has a probability outside [0, 1]: ${probability}`)\n    }\n    sum += probability\n  }\n  if (Math.abs(sum - 1) > 1e-6) throw new Error(`${label} probabilities sum to ${sum}, not 1`)\n}\n\nfunction toBin(chunk: CalibrationPair[], lower?: number, upper?: number): CalibrationBin {\n  const xs = chunk.map((c) => c.evalScore)\n  const ys = chunk.map((c) => c.outcome)\n  const evalMean = mean(xs)\n  const outcomeMean = mean(ys)\n  return {\n    lower: lower ?? Math.min(...xs),\n    upper: upper ?? Math.max(...xs),\n    n: chunk.length,\n    evalMean,\n    outcomeMean,\n    gap: Math.abs(outcomeMean - evalMean),\n  }\n}\n\nfunction mean(xs: number[]): number {\n  return xs.reduce((sum, value) => sum + value / xs.length, 0)\n}\n\nfunction validateCalibrationRequest(\n  evalMetric: string,\n  outcomeMetric: string,\n  options: CalibrationOptions,\n): void {\n  assertUniqueObservationIds([evalMetric], 'eval metric')\n  assertUniqueObservationIds([outcomeMetric], 'outcome metric')\n  if (evalMetric.trim() !== evalMetric || outcomeMetric.trim() !== outcomeMetric) {\n    throw new Error('calibration metric identities must not have surrounding whitespace')\n  }\n  if (options.bins !== undefined && (!Number.isSafeInteger(options.bins) || options.bins < 1)) {\n    throw new Error('calibration bins must be a positive safe integer')\n  }\n  if (\n    options.binning !== undefined &&\n    !['equal-width', 'equal-frequency'].includes(options.binning)\n  ) {\n    throw new Error('calibration binning must be equal-width or equal-frequency')\n  }\n  if (\n    options.range !== undefined &&\n    (!Number.isFinite(options.range.lo) ||\n      !Number.isFinite(options.range.hi) ||\n      !Number.isFinite(options.range.hi - options.range.lo) ||\n      options.range.hi < options.range.lo)\n  ) {\n    throw new Error('calibration range must have finite ordered bounds')\n  }\n}\n"],"mappings":";;;;;;;;;;;AAkDA,eAAsB,iBACpB,YACA,cACA,YACA,eACA,UAA8B,CAAC,GACI;CACnC,MAAM,WAAW;EACf,GAAG;EACH,OAAO,QAAQ,UAAU,KAAA,IAAY,KAAA,IAAY,EAAE,GAAG,QAAQ,MAAM;CACtE;CACA,2BAA2B,WAAW,IAAI,eAAe,QAAQ;CACjE,MAAM,UAAU,WAAW,WAAW,mBAAmB,WAAW,EAAE;CACtE,MAAM,WAAW,WAAW;CAC5B,MAAM,OAAO,MAAM,WAAW,SAAS;CACvC,2BACE,KAAK,KAAK,QAAQ,IAAI,KAAK,GAC3B,OACF;CACA,MAAM,WAAW,MAAM,aAAa,KAAK;CACzC,MAAM,wBAAQ,IAAI,IAAiC;CACnD,KAAK,MAAM,KAAK,UAAU;EACxB,MAAM,MAAM,MAAM,IAAI,EAAE,KAAK,KAAK,CAAC;EACnC,IAAI,KAAK,CAAC;EACV,MAAM,IAAI,EAAE,OAAO,GAAG;CACxB;CAEA,MAAM,QAAyC,CAAC;CAChD,KAAK,MAAM,OAAO,MAAM;EACtB,MAAM,KAAK,MAAM,IAAI,IAAI,KAAK;EAC9B,IAAI,CAAC,IAAI,QAAQ;EACjB,MAAM,IAAI,MAAM,QAAQ,KAAK,UAAU;EACvC,IAAI,MAAM,QAAQ,CAAC,OAAO,SAAS,CAAC,GAAG;EACvC,MAAM,IAAI,oBAAoB,IAAI,eAAe,QAAQ;EACzD,IAAI,MAAM,MAAM;EAChB,MAAM,KAAK;GAAE;GAAG;EAAE,CAAC;CACrB;CACA,IAAI,MAAM,SAAS,GAAG,OAAO;CAE7B,OAAO,qBACL,MAAM,KAAK,OAAO;EAAE,WAAW,EAAE;EAAG,SAAS,EAAE;CAAE,EAAE,GACnD,UACA,eACA,QACF;AACF;;AAGA,SAAgB,qBACd,YACA,YACA,eACA,UAA8B,CAAC,GACL;CAC1B,2BAA2B,YAAY,eAAe,OAAO;CAC7D,KAAK,MAAM,CAAC,OAAO,SAAS,WAAW,QAAQ,GAC7C,IACE,SAAS,QACT,OAAO,SAAS,YAChB,CAAC,OAAO,SAAS,KAAK,SAAS,KAC/B,CAAC,OAAO,SAAS,KAAK,OAAO,GAE7B,MAAM,IAAI,MAAM,oBAAoB,MAAM,kDAAkD;CAGhG,MAAM,QAAQ;CACd,IAAI,MAAM,SAAS,GAAG,OAAO;CAE7B,MAAM,UAAU,QAAQ,QAAQ;CAChC,MAAM,UAAU,QAAQ,WAAW;CACnC,MAAM,KAAK,MAAM,KAAK,MAAM,EAAE,SAAS;CACvC,MAAM,KAAK,QAAQ,OAAO,MAAM,KAAK,IAAI,GAAG,EAAE;CAC9C,MAAM,KAAK,QAAQ,OAAO,MAAM,KAAK,IAAI,GAAG,EAAE;CAC9C,MAAM,OAAO,KAAK;CAClB,IAAI,CAAC,OAAO,SAAS,IAAI,GAAG,MAAM,IAAI,MAAM,uCAAuC;CACnF,MAAM,UAAU,MAAM,KAAK,UAAU;EACnC,GAAG;EACH,WAAW,KAAK,IAAI,IAAI,KAAK,IAAI,IAAI,KAAK,SAAS,CAAC;CACtD,EAAE;CAEF,MAAM,OAAyB,CAAC;CAChC,IAAI,SAAS,KAAK,QAAQ,OAAO,SAAS,KAAK,cAAc,QAAQ,EAAE,CAAE,SAAS,GAChF,KAAK,KAAK,MAAM,OAAO,CAAC;MACnB,IAAI,YAAY,mBAAmB;EACxC,MAAM,SAAS,CAAC,GAAG,OAAO,CAAC,CAAC,MAAM,GAAG,MAAM,EAAE,YAAY,EAAE,SAAS;EACpE,MAAM,QAAQ,KAAK,IAAI,SAAS,OAAO,MAAM;EAC7C,KAAK,IAAI,IAAI,GAAG,IAAI,OAAO,KAAK;GAC9B,MAAM,QAAQ,KAAK,MAAO,IAAI,OAAO,SAAU,KAAK;GACpD,MAAM,MAAM,KAAK,OAAQ,IAAI,KAAK,OAAO,SAAU,KAAK;GACxD,KAAK,KAAK,MAAM,OAAO,MAAM,OAAO,GAAG,CAAC,CAAC;EAC3C;CACF,OAAO;EACL,MAAM,yBAAS,IAAI,IAA+B;EAClD,KAAK,MAAM,QAAQ,SAAS;GAC1B,MAAM,QAAQ,KAAK,IAAI,UAAU,GAAG,KAAK,OAAQ,KAAK,YAAY,MAAM,OAAQ,OAAO,CAAC;GACxF,MAAM,QAAQ,OAAO,IAAI,KAAK,KAAK,CAAC;GACpC,MAAM,KAAK,IAAI;GACf,OAAO,IAAI,OAAO,KAAK;EACzB;EACA,KAAK,MAAM,CAAC,OAAO,UAAU,CAAC,GAAG,MAAM,CAAC,CAAC,MAAM,CAAC,IAAI,CAAC,OAAO,IAAI,CAAC,GAC/D,KAAK,KAAK,MAAM,OAAO,KAAK,QAAQ,QAAQ,UAAU,KAAK,SAAS,QAAQ,KAAK,QAAQ,CAAC;CAE9F;CAEA,MAAM,QAAQ,KAAK,QAAQ,GAAG,MAAM,IAAI,EAAE,GAAG,CAAC;CAC9C,MAAM,MAAM,KAAK,QAAQ,GAAG,MAAM,IAAK,EAAE,IAAI,QAAS,EAAE,KAAK,CAAC;CAC9D,MAAM,SAAS,KAAK,QAAQ,GAAG,MAAM,KAAK,IAAI,GAAG,EAAE,GAAG,GAAG,CAAC;CAE1D,OAAO;EAAE;EAAY;EAAe,GAAG,MAAM;EAAQ;EAAM;EAAK;CAAO;AACzE;;;;;;AAOA,SAAgB,WAAW,OAA2C;CACpE,IAAI,CAAC,MAAM,QAAQ,MAAM,IAAI,MAAM,wCAAwC;CAC3E,IAAI,QAAQ;CACZ,KAAK,MAAM,CAAC,OAAO,SAAS,MAAM,QAAQ,GAAG;EAC3C,IACE,SAAS,QACT,OAAO,SAAS,YAChB,EAAE,KAAK,aAAa,KAAK,KAAK,aAAa,MAC1C,KAAK,YAAY,KAAK,KAAK,YAAY,GAExC,MAAM,IAAI,MAAM,cAAc,MAAM,wDAAwD;EAE9F,UAAU,KAAK,YAAY,KAAK,YAAY;CAC9C;CACA,OAAO,QAAQ,MAAM;AACvB;;;;;;;;AAgBA,SAAgB,qBAAqB,WAAmD;CACtF,IAAI,CAAC,UAAU,QAAQ,MAAM,IAAI,MAAM,uDAAuD;CAC9F,IAAI,QAAQ;CACZ,KAAK,MAAM,CAAC,OAAO,aAAa,UAAU,QAAQ,GAAG;EACnD,MAAM,UAAU,mBAAmB,UAAU,OAAO,kBAAkB;EACtE,KAAK,MAAM,CAAC,MAAM,gBAAgB,SAChC,UAAU,eAAe,SAAS,SAAS,UAAU,IAAI,OAAO;CAEpE;CACA,OAAO,QAAQ,UAAU;AAC3B;;;;;;;AAeA,SAAgB,uBAAuB,WAA+C;CACpF,IAAI,CAAC,UAAU,QAAQ,MAAM,IAAI,MAAM,yDAAyD;CAChG,IAAI,QAAQ;CACZ,KAAK,MAAM,CAAC,OAAO,aAAa,UAAU,QAAQ,GAAG;EACnD,MAAM,gBAAgB,UAAU;EAChC,IAAI,CAAC,MAAM,QAAQ,aAAa,KAAK,cAAc,SAAS,GAC1D,MAAM,IAAI,MAAM,gBAAgB,MAAM,wCAAwC;EAEhF,mBAAmB,eAAe,gBAAgB,OAAO;EACzD,IACE,CAAC,OAAO,UAAU,SAAS,OAAO,KAClC,SAAS,UAAU,KACnB,SAAS,WAAW,cAAc,QAElC,MAAM,IAAI,MACR,gBAAgB,MAAM,iCAAiC,cAAc,OAAO,QAC9E;EAEF,IAAI,aAAa;EACjB,IAAI,MAAM;EACV,KAAK,IAAI,QAAQ,GAAG,QAAQ,cAAc,SAAS,GAAG,SAAS;GAC7D,cAAc,cAAc;GAC5B,QAAQ,cAAc,SAAS,SAAS,UAAU,IAAI,OAAO;EAC/D;EACA,SAAS,OAAO,cAAc,SAAS;CACzC;CACA,OAAO,QAAQ,UAAU;AAC3B;AAEA,SAAS,mBACP,UACA,OACA,OACoB;CACpB,IAAI,aAAa,QAAQ,OAAO,aAAa,YAAY,OAAO,SAAS,YAAY,UACnF,MAAM,IAAI,MAAM,GAAG,MAAM,YAAY,MAAM,0CAA0C;CAEvF,MAAM,gBAAgB,SAAS;CAC/B,IAAI,kBAAkB,QAAQ,OAAO,kBAAkB,YAAY,MAAM,QAAQ,aAAa,GAC5F,MAAM,IAAI,MAAM,GAAG,MAAM,YAAY,MAAM,kDAAkD;CAE/F,MAAM,UAAU,OAAO,QAAQ,aAAa;CAC5C,IAAI,QAAQ,SAAS,GAAG,MAAM,IAAI,MAAM,GAAG,MAAM,YAAY,MAAM,6BAA6B;CAChG,IAAI,CAAC,QAAQ,MAAM,CAAC,UAAU,SAAS,SAAS,OAAO,GACrD,MAAM,IAAI,MACR,GAAG,MAAM,YAAY,MAAM,kCAAkC,KAAK,UAAU,SAAS,OAAO,GAC9F;CAEF,mBACE,QAAQ,KAAK,GAAG,iBAAiB,WAAW,GAC5C,GAAG,MAAM,YAAY,OACvB;CACA,OAAO;AACT;AAEA,SAAS,mBAAmB,eAAkC,OAAqB;CACjF,IAAI,MAAM;CACV,KAAK,MAAM,eAAe,eAAe;EACvC,IAAI,EAAE,eAAe,KAAK,eAAe,IACvC,MAAM,IAAI,MAAM,GAAG,MAAM,qCAAqC,aAAa;EAE7E,OAAO;CACT;CACA,IAAI,KAAK,IAAI,MAAM,CAAC,IAAI,MAAM,MAAM,IAAI,MAAM,GAAG,MAAM,wBAAwB,IAAI,QAAQ;AAC7F;AAEA,SAAS,MAAM,OAA0B,OAAgB,OAAgC;CACvF,MAAM,KAAK,MAAM,KAAK,MAAM,EAAE,SAAS;CACvC,MAAM,KAAK,MAAM,KAAK,MAAM,EAAE,OAAO;CACrC,MAAM,WAAW,KAAK,EAAE;CACxB,MAAM,cAAc,KAAK,EAAE;CAC3B,OAAO;EACL,OAAO,SAAS,KAAK,IAAI,GAAG,EAAE;EAC9B,OAAO,SAAS,KAAK,IAAI,GAAG,EAAE;EAC9B,GAAG,MAAM;EACT;EACA;EACA,KAAK,KAAK,IAAI,cAAc,QAAQ;CACtC;AACF;AAEA,SAAS,KAAK,IAAsB;CAClC,OAAO,GAAG,QAAQ,KAAK,UAAU,MAAM,QAAQ,GAAG,QAAQ,CAAC;AAC7D;AAEA,SAAS,2BACP,YACA,eACA,SACM;CACN,2BAA2B,CAAC,UAAU,GAAG,aAAa;CACtD,2BAA2B,CAAC,aAAa,GAAG,gBAAgB;CAC5D,IAAI,WAAW,KAAK,MAAM,cAAc,cAAc,KAAK,MAAM,eAC/D,MAAM,IAAI,MAAM,oEAAoE;CAEtF,IAAI,QAAQ,SAAS,KAAA,MAAc,CAAC,OAAO,cAAc,QAAQ,IAAI,KAAK,QAAQ,OAAO,IACvF,MAAM,IAAI,MAAM,kDAAkD;CAEpE,IACE,QAAQ,YAAY,KAAA,KACpB,CAAC,CAAC,eAAe,iBAAiB,CAAC,CAAC,SAAS,QAAQ,OAAO,GAE5D,MAAM,IAAI,MAAM,4DAA4D;CAE9E,IACE,QAAQ,UAAU,KAAA,MACjB,CAAC,OAAO,SAAS,QAAQ,MAAM,EAAE,KAChC,CAAC,OAAO,SAAS,QAAQ,MAAM,EAAE,KACjC,CAAC,OAAO,SAAS,QAAQ,MAAM,KAAK,QAAQ,MAAM,EAAE,KACpD,QAAQ,MAAM,KAAK,QAAQ,MAAM,KAEnC,MAAM,IAAI,MAAM,mDAAmD;AAEvE"}
+{"version":3,"file":"calibration-6usfVYYo.js","names":[],"sources":["../src/meta-eval/calibration.ts"],"sourcesContent":["/**\n * Calibration curve — binned \"if eval says X, what does reality show?\"\n *\n * Companion to correlationStudy. Raw correlation is a single number;\n * the calibration curve shows *where* the eval is well-calibrated vs\n * overconfident / underconfident. Buckets the eval metric, computes\n * mean outcome per bucket, reports expected-calibration-error (ECE).\n */\n\nimport { runMetricExtractor } from '../trace/query'\nimport type { TraceStore } from '../trace/store'\nimport type { EvalMetricSpec } from './correlation-study'\nimport { assertUniqueObservationIds, reduceOutcomeMetric } from './outcome-observations'\nimport type { DeploymentOutcome, OutcomeStore } from './outcome-store'\n\nexport interface CalibrationBin {\n  lower: number\n  upper: number\n  n: number\n  evalMean: number\n  outcomeMean: number\n  /** |outcomeMean − evalMean|; contributes to ECE weighted by n/total. */\n  gap: number\n}\n\nexport interface CalibrationReport {\n  evalMetric: string\n  outcomeMetric: string\n  n: number\n  bins: CalibrationBin[]\n  /** Expected Calibration Error — Σ (n_i/N) × |outcomeMean_i − evalMean_i|. */\n  ece: number\n  /** Largest observed difference between a bin's mean score and mean outcome. */\n  maxGap: number\n}\n\nexport interface CalibrationOptions {\n  /** Positive integer; empty bins are omitted. Default 10. */\n  bins?: number\n  /** Equal-width (fixed bin edges) or equal-frequency (quantile bins). */\n  binning?: 'equal-width' | 'equal-frequency'\n  /** Clip eval values to [lo, hi] before binning. */\n  range?: { lo: number; hi: number }\n}\n\nexport interface CalibrationPair {\n  evalScore: number\n  outcome: number\n}\n\nexport async function calibrationCurve(\n  traceStore: TraceStore,\n  outcomeStore: OutcomeStore,\n  evalMetric: EvalMetricSpec,\n  outcomeMetric: string,\n  options: CalibrationOptions = {},\n): Promise<CalibrationReport | null> {\n  const settings = {\n    ...options,\n    range: options.range === undefined ? undefined : { ...options.range },\n  }\n  validateCalibrationRequest(evalMetric.id, outcomeMetric, settings)\n  const extract = evalMetric.extract ?? runMetricExtractor(evalMetric.id)\n  const metricId = evalMetric.id\n  const runs = await traceStore.listRuns()\n  assertUniqueObservationIds(\n    runs.map((run) => run.runId),\n    'runId',\n  )\n  const outcomes = await outcomeStore.list()\n  const byRun = new Map<string, DeploymentOutcome[]>()\n  for (const o of outcomes) {\n    const arr = byRun.get(o.runId) ?? []\n    arr.push(o)\n    byRun.set(o.runId, arr)\n  }\n\n  const pairs: Array<{ x: number; y: number }> = []\n  for (const run of runs) {\n    const os = byRun.get(run.runId)\n    if (!os?.length) continue\n    const x = await extract(run, traceStore)\n    if (x === null || !Number.isFinite(x)) continue\n    const y = reduceOutcomeMetric(os, outcomeMetric, 'latest')\n    if (y === null) continue\n    pairs.push({ x, y })\n  }\n  if (pairs.length < 2) return null\n\n  return calibrationFromPairs(\n    pairs.map((p) => ({ evalScore: p.x, outcome: p.y })),\n    metricId,\n    outcomeMetric,\n    settings,\n  )\n}\n\n/** Measure already joined observations without constructing trace and outcome stores. */\nexport function calibrationFromPairs(\n  inputPairs: readonly CalibrationPair[],\n  evalMetric: string,\n  outcomeMetric: string,\n  options: CalibrationOptions = {},\n): CalibrationReport | null {\n  validateCalibrationRequest(evalMetric, outcomeMetric, options)\n  for (const [index, pair] of inputPairs.entries()) {\n    if (\n      pair === null ||\n      typeof pair !== 'object' ||\n      !Number.isFinite(pair.evalScore) ||\n      !Number.isFinite(pair.outcome)\n    ) {\n      throw new Error(`calibration pair ${index} must contain finite evalScore and outcome values`)\n    }\n  }\n  const pairs = inputPairs\n  if (pairs.length < 2) return null\n\n  const numBins = options.bins ?? 10\n  const binning = options.binning ?? 'equal-width'\n  const xs = pairs.map((p) => p.evalScore)\n  const lo = options.range?.lo ?? Math.min(...xs)\n  const hi = options.range?.hi ?? Math.max(...xs)\n  const span = hi - lo\n  if (!Number.isFinite(span)) throw new Error('calibration range span must be finite')\n  const clipped = pairs.map((pair) => ({\n    ...pair,\n    evalScore: Math.min(hi, Math.max(lo, pair.evalScore)),\n  }))\n\n  const bins: CalibrationBin[] = []\n  if (span === 0 || clipped.every((pair) => pair.evalScore === clipped[0]!.evalScore)) {\n    bins.push(toBin(clipped))\n  } else if (binning === 'equal-frequency') {\n    const sorted = [...clipped].sort((a, b) => a.evalScore - b.evalScore)\n    const count = Math.min(numBins, sorted.length)\n    for (let i = 0; i < count; i++) {\n      const start = Math.floor((i * sorted.length) / count)\n      const end = Math.floor(((i + 1) * sorted.length) / count)\n      bins.push(toBin(sorted.slice(start, end)))\n    }\n  } else {\n    const groups = new Map<number, CalibrationPair[]>()\n    for (const pair of clipped) {\n      const index = Math.min(numBins - 1, Math.floor(((pair.evalScore - lo) / span) * numBins))\n      const group = groups.get(index) ?? []\n      group.push(pair)\n      groups.set(index, group)\n    }\n    for (const [index, chunk] of [...groups].sort(([a], [b]) => a - b)) {\n      bins.push(toBin(chunk, lo + span * (index / numBins), lo + span * ((index + 1) / numBins)))\n    }\n  }\n\n  const total = bins.reduce((a, b) => a + b.n, 0)\n  const ece = bins.reduce((a, b) => a + (b.n / total) * b.gap, 0)\n  const maxGap = bins.reduce((a, b) => Math.max(a, b.gap), 0)\n\n  return { evalMetric, outcomeMetric, n: pairs.length, bins, ece, maxGap }\n}\n\n/**\n * Mean squared error of probability forecasts against binary outcomes, Σ(p − y)²/N; always\n * forecasting 0.5 scores 0.25. For the reliability table and ECE use `calibrationFromPairs`\n * with `range: { lo: 0, hi: 1 }` (`binning: 'equal-frequency'` gives equal-mass bins).\n */\nexport function brierScore(pairs: readonly CalibrationPair[]): number {\n  if (!pairs.length) throw new Error('Brier score requires at least one pair')\n  let total = 0\n  for (const [index, pair] of pairs.entries()) {\n    if (\n      pair === null ||\n      typeof pair !== 'object' ||\n      !(pair.evalScore >= 0 && pair.evalScore <= 1) ||\n      (pair.outcome !== 0 && pair.outcome !== 1)\n    ) {\n      throw new Error(`Brier pair ${index} needs a probability in [0, 1] and an outcome of 0 or 1`)\n    }\n    total += (pair.evalScore - pair.outcome) ** 2\n  }\n  return total / pairs.length\n}\n\n/** One categorical forecast: a distribution over named outcomes and the outcome that occurred. */\nexport interface CategoricalForecast {\n  /** Probability per outcome name. Must sum to 1 within 1e-6 and include `outcome`. */\n  probabilities: Readonly<Record<string, number>>\n  outcome: string\n}\n\n/**\n * Multiclass Brier score in its sum form, mean over forecasts of Σₖ(pₖ − yₖ)², range [0, 2].\n * With two outcomes it is twice the binary `brierScore`. A uniform forecast over K outcomes\n * scores 1 − 1/K. The distribution is not renormalized here: repair provider rounding before\n * scoring (`normalizeSystemOneResult` in `/systemone/protocol`), so a defective answer stays\n * visible instead of being scored as if it were valid.\n */\nexport function multiclassBrierScore(forecasts: readonly CategoricalForecast[]): number {\n  if (!forecasts.length) throw new Error('multiclass Brier score requires at least one forecast')\n  let total = 0\n  for (const [index, forecast] of forecasts.entries()) {\n    const entries = categoricalEntries(forecast, index, 'multiclass Brier')\n    for (const [name, probability] of entries) {\n      total += (probability - (name === forecast.outcome ? 1 : 0)) ** 2\n    }\n  }\n  return total / forecasts.length\n}\n\n/** One ordinal forecast: probabilities in level order, lowest level first. */\nexport interface OrdinalForecast {\n  probabilities: readonly number[]\n  /** Index into `probabilities` of the level that occurred. */\n  outcome: number\n}\n\n/**\n * Ranked probability score, mean over forecasts of Σₖ₌₁ᴷ⁻¹(Fₖ − Oₖ)² / (K − 1), where F and O are\n * the cumulative forecast and observation, range [0, 1]. Unlike multiclass Brier it charges\n * less for mass on a level next to the outcome than for mass far from it, which is the right\n * penalty for score questions. With two levels it equals the binary Brier score.\n */\nexport function rankedProbabilityScore(forecasts: readonly OrdinalForecast[]): number {\n  if (!forecasts.length) throw new Error('ranked probability score requires at least one forecast')\n  let total = 0\n  for (const [index, forecast] of forecasts.entries()) {\n    const probabilities = forecast?.probabilities\n    if (!Array.isArray(probabilities) || probabilities.length < 2) {\n      throw new Error(`RPS forecast ${index} needs at least two level probabilities`)\n    }\n    assertDistribution(probabilities, `RPS forecast ${index}`)\n    if (\n      !Number.isInteger(forecast.outcome) ||\n      forecast.outcome < 0 ||\n      forecast.outcome >= probabilities.length\n    ) {\n      throw new Error(\n        `RPS forecast ${index} outcome must index one of its ${probabilities.length} levels`,\n      )\n    }\n    let cumulative = 0\n    let sum = 0\n    for (let level = 0; level < probabilities.length - 1; level++) {\n      cumulative += probabilities[level]!\n      sum += (cumulative - (level >= forecast.outcome ? 1 : 0)) ** 2\n    }\n    total += sum / (probabilities.length - 1)\n  }\n  return total / forecasts.length\n}\n\nfunction categoricalEntries(\n  forecast: CategoricalForecast,\n  index: number,\n  label: string,\n): [string, number][] {\n  if (forecast === null || typeof forecast !== 'object' || typeof forecast.outcome !== 'string') {\n    throw new Error(`${label} forecast ${index} needs probabilities and a string outcome`)\n  }\n  const probabilities = forecast.probabilities\n  if (probabilities === null || typeof probabilities !== 'object' || Array.isArray(probabilities)) {\n    throw new Error(`${label} forecast ${index} probabilities must be an object keyed by outcome`)\n  }\n  const entries = Object.entries(probabilities)\n  if (entries.length < 2) throw new Error(`${label} forecast ${index} needs at least two outcomes`)\n  if (!entries.some(([name]) => name === forecast.outcome)) {\n    throw new Error(\n      `${label} forecast ${index} has no probability for outcome ${JSON.stringify(forecast.outcome)}`,\n    )\n  }\n  assertDistribution(\n    entries.map(([, probability]) => probability),\n    `${label} forecast ${index}`,\n  )\n  return entries\n}\n\nfunction assertDistribution(probabilities: readonly number[], label: string): void {\n  let sum = 0\n  for (const probability of probabilities) {\n    if (!(probability >= 0 && probability <= 1)) {\n      throw new Error(`${label} has a probability outside [0, 1]: ${probability}`)\n    }\n    sum += probability\n  }\n  if (Math.abs(sum - 1) > 1e-6) throw new Error(`${label} probabilities sum to ${sum}, not 1`)\n}\n\nfunction toBin(chunk: CalibrationPair[], lower?: number, upper?: number): CalibrationBin {\n  const xs = chunk.map((c) => c.evalScore)\n  const ys = chunk.map((c) => c.outcome)\n  const evalMean = mean(xs)\n  const outcomeMean = mean(ys)\n  return {\n    lower: lower ?? Math.min(...xs),\n    upper: upper ?? Math.max(...xs),\n    n: chunk.length,\n    evalMean,\n    outcomeMean,\n    gap: Math.abs(outcomeMean - evalMean),\n  }\n}\n\nfunction mean(xs: number[]): number {\n  return xs.reduce((sum, value) => sum + value / xs.length, 0)\n}\n\nfunction validateCalibrationRequest(\n  evalMetric: string,\n  outcomeMetric: string,\n  options: CalibrationOptions,\n): void {\n  assertUniqueObservationIds([evalMetric], 'eval metric')\n  assertUniqueObservationIds([outcomeMetric], 'outcome metric')\n  if (evalMetric.trim() !== evalMetric || outcomeMetric.trim() !== outcomeMetric) {\n    throw new Error('calibration metric identities must not have surrounding whitespace')\n  }\n  if (options.bins !== undefined && (!Number.isSafeInteger(options.bins) || options.bins < 1)) {\n    throw new Error('calibration bins must be a positive safe integer')\n  }\n  if (\n    options.binning !== undefined &&\n    !['equal-width', 'equal-frequency'].includes(options.binning)\n  ) {\n    throw new Error('calibration binning must be equal-width or equal-frequency')\n  }\n  if (\n    options.range !== undefined &&\n    (!Number.isFinite(options.range.lo) ||\n      !Number.isFinite(options.range.hi) ||\n      !Number.isFinite(options.range.hi - options.range.lo) ||\n      options.range.hi < options.range.lo)\n  ) {\n    throw new Error('calibration range must have finite ordered bounds')\n  }\n}\n"],"mappings":";;;;;;;;;;;AAkDA,eAAsB,iBACpB,YACA,cACA,YACA,eACA,UAA8B,CAAC,GACI;CACnC,MAAM,WAAW;EACf,GAAG;EACH,OAAO,QAAQ,UAAU,KAAA,IAAY,KAAA,IAAY,EAAE,GAAG,QAAQ,MAAM;CACtE;CACA,2BAA2B,WAAW,IAAI,eAAe,QAAQ;CACjE,MAAM,UAAU,WAAW,WAAW,mBAAmB,WAAW,EAAE;CACtE,MAAM,WAAW,WAAW;CAC5B,MAAM,OAAO,MAAM,WAAW,SAAS;CACvC,2BACE,KAAK,KAAK,QAAQ,IAAI,KAAK,GAC3B,OACF;CACA,MAAM,WAAW,MAAM,aAAa,KAAK;CACzC,MAAM,wBAAQ,IAAI,IAAiC;CACnD,KAAK,MAAM,KAAK,UAAU;EACxB,MAAM,MAAM,MAAM,IAAI,EAAE,KAAK,KAAK,CAAC;EACnC,IAAI,KAAK,CAAC;EACV,MAAM,IAAI,EAAE,OAAO,GAAG;CACxB;CAEA,MAAM,QAAyC,CAAC;CAChD,KAAK,MAAM,OAAO,MAAM;EACtB,MAAM,KAAK,MAAM,IAAI,IAAI,KAAK;EAC9B,IAAI,CAAC,IAAI,QAAQ;EACjB,MAAM,IAAI,MAAM,QAAQ,KAAK,UAAU;EACvC,IAAI,MAAM,QAAQ,CAAC,OAAO,SAAS,CAAC,GAAG;EACvC,MAAM,IAAI,oBAAoB,IAAI,eAAe,QAAQ;EACzD,IAAI,MAAM,MAAM;EAChB,MAAM,KAAK;GAAE;GAAG;EAAE,CAAC;CACrB;CACA,IAAI,MAAM,SAAS,GAAG,OAAO;CAE7B,OAAO,qBACL,MAAM,KAAK,OAAO;EAAE,WAAW,EAAE;EAAG,SAAS,EAAE;CAAE,EAAE,GACnD,UACA,eACA,QACF;AACF;;AAGA,SAAgB,qBACd,YACA,YACA,eACA,UAA8B,CAAC,GACL;CAC1B,2BAA2B,YAAY,eAAe,OAAO;CAC7D,KAAK,MAAM,CAAC,OAAO,SAAS,WAAW,QAAQ,GAC7C,IACE,SAAS,QACT,OAAO,SAAS,YAChB,CAAC,OAAO,SAAS,KAAK,SAAS,KAC/B,CAAC,OAAO,SAAS,KAAK,OAAO,GAE7B,MAAM,IAAI,MAAM,oBAAoB,MAAM,kDAAkD;CAGhG,MAAM,QAAQ;CACd,IAAI,MAAM,SAAS,GAAG,OAAO;CAE7B,MAAM,UAAU,QAAQ,QAAQ;CAChC,MAAM,UAAU,QAAQ,WAAW;CACnC,MAAM,KAAK,MAAM,KAAK,MAAM,EAAE,SAAS;CACvC,MAAM,KAAK,QAAQ,OAAO,MAAM,KAAK,IAAI,GAAG,EAAE;CAC9C,MAAM,KAAK,QAAQ,OAAO,MAAM,KAAK,IAAI,GAAG,EAAE;CAC9C,MAAM,OAAO,KAAK;CAClB,IAAI,CAAC,OAAO,SAAS,IAAI,GAAG,MAAM,IAAI,MAAM,uCAAuC;CACnF,MAAM,UAAU,MAAM,KAAK,UAAU;EACnC,GAAG;EACH,WAAW,KAAK,IAAI,IAAI,KAAK,IAAI,IAAI,KAAK,SAAS,CAAC;CACtD,EAAE;CAEF,MAAM,OAAyB,CAAC;CAChC,IAAI,SAAS,KAAK,QAAQ,OAAO,SAAS,KAAK,cAAc,QAAQ,EAAE,CAAE,SAAS,GAChF,KAAK,KAAK,MAAM,OAAO,CAAC;MACnB,IAAI,YAAY,mBAAmB;EACxC,MAAM,SAAS,CAAC,GAAG,OAAO,CAAC,CAAC,MAAM,GAAG,MAAM,EAAE,YAAY,EAAE,SAAS;EACpE,MAAM,QAAQ,KAAK,IAAI,SAAS,OAAO,MAAM;EAC7C,KAAK,IAAI,IAAI,GAAG,IAAI,OAAO,KAAK;GAC9B,MAAM,QAAQ,KAAK,MAAO,IAAI,OAAO,SAAU,KAAK;GACpD,MAAM,MAAM,KAAK,OAAQ,IAAI,KAAK,OAAO,SAAU,KAAK;GACxD,KAAK,KAAK,MAAM,OAAO,MAAM,OAAO,GAAG,CAAC,CAAC;EAC3C;CACF,OAAO;EACL,MAAM,yBAAS,IAAI,IAA+B;EAClD,KAAK,MAAM,QAAQ,SAAS;GAC1B,MAAM,QAAQ,KAAK,IAAI,UAAU,GAAG,KAAK,OAAQ,KAAK,YAAY,MAAM,OAAQ,OAAO,CAAC;GACxF,MAAM,QAAQ,OAAO,IAAI,KAAK,KAAK,CAAC;GACpC,MAAM,KAAK,IAAI;GACf,OAAO,IAAI,OAAO,KAAK;EACzB;EACA,KAAK,MAAM,CAAC,OAAO,UAAU,CAAC,GAAG,MAAM,CAAC,CAAC,MAAM,CAAC,IAAI,CAAC,OAAO,IAAI,CAAC,GAC/D,KAAK,KAAK,MAAM,OAAO,KAAK,QAAQ,QAAQ,UAAU,KAAK,SAAS,QAAQ,KAAK,QAAQ,CAAC;CAE9F;CAEA,MAAM,QAAQ,KAAK,QAAQ,GAAG,MAAM,IAAI,EAAE,GAAG,CAAC;CAC9C,MAAM,MAAM,KAAK,QAAQ,GAAG,MAAM,IAAK,EAAE,IAAI,QAAS,EAAE,KAAK,CAAC;CAC9D,MAAM,SAAS,KAAK,QAAQ,GAAG,MAAM,KAAK,IAAI,GAAG,EAAE,GAAG,GAAG,CAAC;CAE1D,OAAO;EAAE;EAAY;EAAe,GAAG,MAAM;EAAQ;EAAM;EAAK;CAAO;AACzE;;;;;;AAOA,SAAgB,WAAW,OAA2C;CACpE,IAAI,CAAC,MAAM,QAAQ,MAAM,IAAI,MAAM,wCAAwC;CAC3E,IAAI,QAAQ;CACZ,KAAK,MAAM,CAAC,OAAO,SAAS,MAAM,QAAQ,GAAG;EAC3C,IACE,SAAS,QACT,OAAO,SAAS,YAChB,EAAE,KAAK,aAAa,KAAK,KAAK,aAAa,MAC1C,KAAK,YAAY,KAAK,KAAK,YAAY,GAExC,MAAM,IAAI,MAAM,cAAc,MAAM,wDAAwD;EAE9F,UAAU,KAAK,YAAY,KAAK,YAAY;CAC9C;CACA,OAAO,QAAQ,MAAM;AACvB;;;;;;;;AAgBA,SAAgB,qBAAqB,WAAmD;CACtF,IAAI,CAAC,UAAU,QAAQ,MAAM,IAAI,MAAM,uDAAuD;CAC9F,IAAI,QAAQ;CACZ,KAAK,MAAM,CAAC,OAAO,aAAa,UAAU,QAAQ,GAAG;EACnD,MAAM,UAAU,mBAAmB,UAAU,OAAO,kBAAkB;EACtE,KAAK,MAAM,CAAC,MAAM,gBAAgB,SAChC,UAAU,eAAe,SAAS,SAAS,UAAU,IAAI,OAAO;CAEpE;CACA,OAAO,QAAQ,UAAU;AAC3B;;;;;;;AAeA,SAAgB,uBAAuB,WAA+C;CACpF,IAAI,CAAC,UAAU,QAAQ,MAAM,IAAI,MAAM,yDAAyD;CAChG,IAAI,QAAQ;CACZ,KAAK,MAAM,CAAC,OAAO,aAAa,UAAU,QAAQ,GAAG;EACnD,MAAM,gBAAgB,UAAU;EAChC,IAAI,CAAC,MAAM,QAAQ,aAAa,KAAK,cAAc,SAAS,GAC1D,MAAM,IAAI,MAAM,gBAAgB,MAAM,wCAAwC;EAEhF,mBAAmB,eAAe,gBAAgB,OAAO;EACzD,IACE,CAAC,OAAO,UAAU,SAAS,OAAO,KAClC,SAAS,UAAU,KACnB,SAAS,WAAW,cAAc,QAElC,MAAM,IAAI,MACR,gBAAgB,MAAM,iCAAiC,cAAc,OAAO,QAC9E;EAEF,IAAI,aAAa;EACjB,IAAI,MAAM;EACV,KAAK,IAAI,QAAQ,GAAG,QAAQ,cAAc,SAAS,GAAG,SAAS;GAC7D,cAAc,cAAc;GAC5B,QAAQ,cAAc,SAAS,SAAS,UAAU,IAAI,OAAO;EAC/D;EACA,SAAS,OAAO,cAAc,SAAS;CACzC;CACA,OAAO,QAAQ,UAAU;AAC3B;AAEA,SAAS,mBACP,UACA,OACA,OACoB;CACpB,IAAI,aAAa,QAAQ,OAAO,aAAa,YAAY,OAAO,SAAS,YAAY,UACnF,MAAM,IAAI,MAAM,GAAG,MAAM,YAAY,MAAM,0CAA0C;CAEvF,MAAM,gBAAgB,SAAS;CAC/B,IAAI,kBAAkB,QAAQ,OAAO,kBAAkB,YAAY,MAAM,QAAQ,aAAa,GAC5F,MAAM,IAAI,MAAM,GAAG,MAAM,YAAY,MAAM,kDAAkD;CAE/F,MAAM,UAAU,OAAO,QAAQ,aAAa;CAC5C,IAAI,QAAQ,SAAS,GAAG,MAAM,IAAI,MAAM,GAAG,MAAM,YAAY,MAAM,6BAA6B;CAChG,IAAI,CAAC,QAAQ,MAAM,CAAC,UAAU,SAAS,SAAS,OAAO,GACrD,MAAM,IAAI,MACR,GAAG,MAAM,YAAY,MAAM,kCAAkC,KAAK,UAAU,SAAS,OAAO,GAC9F;CAEF,mBACE,QAAQ,KAAK,GAAG,iBAAiB,WAAW,GAC5C,GAAG,MAAM,YAAY,OACvB;CACA,OAAO;AACT;AAEA,SAAS,mBAAmB,eAAkC,OAAqB;CACjF,IAAI,MAAM;CACV,KAAK,MAAM,eAAe,eAAe;EACvC,IAAI,EAAE,eAAe,KAAK,eAAe,IACvC,MAAM,IAAI,MAAM,GAAG,MAAM,qCAAqC,aAAa;EAE7E,OAAO;CACT;CACA,IAAI,KAAK,IAAI,MAAM,CAAC,IAAI,MAAM,MAAM,IAAI,MAAM,GAAG,MAAM,wBAAwB,IAAI,QAAQ;AAC7F;AAEA,SAAS,MAAM,OAA0B,OAAgB,OAAgC;CACvF,MAAM,KAAK,MAAM,KAAK,MAAM,EAAE,SAAS;CACvC,MAAM,KAAK,MAAM,KAAK,MAAM,EAAE,OAAO;CACrC,MAAM,WAAW,KAAK,EAAE;CACxB,MAAM,cAAc,KAAK,EAAE;CAC3B,OAAO;EACL,OAAO,SAAS,KAAK,IAAI,GAAG,EAAE;EAC9B,OAAO,SAAS,KAAK,IAAI,GAAG,EAAE;EAC9B,GAAG,MAAM;EACT;EACA;EACA,KAAK,KAAK,IAAI,cAAc,QAAQ;CACtC;AACF;AAEA,SAAS,KAAK,IAAsB;CAClC,OAAO,GAAG,QAAQ,KAAK,UAAU,MAAM,QAAQ,GAAG,QAAQ,CAAC;AAC7D;AAEA,SAAS,2BACP,YACA,eACA,SACM;CACN,2BAA2B,CAAC,UAAU,GAAG,aAAa;CACtD,2BAA2B,CAAC,aAAa,GAAG,gBAAgB;CAC5D,IAAI,WAAW,KAAK,MAAM,cAAc,cAAc,KAAK,MAAM,eAC/D,MAAM,IAAI,MAAM,oEAAoE;CAEtF,IAAI,QAAQ,SAAS,KAAA,MAAc,CAAC,OAAO,cAAc,QAAQ,IAAI,KAAK,QAAQ,OAAO,IACvF,MAAM,IAAI,MAAM,kDAAkD;CAEpE,IACE,QAAQ,YAAY,KAAA,KACpB,CAAC,CAAC,eAAe,iBAAiB,CAAC,CAAC,SAAS,QAAQ,OAAO,GAE5D,MAAM,IAAI,MAAM,4DAA4D;CAE9E,IACE,QAAQ,UAAU,KAAA,MACjB,CAAC,OAAO,SAAS,QAAQ,MAAM,EAAE,KAChC,CAAC,OAAO,SAAS,QAAQ,MAAM,EAAE,KACjC,CAAC,OAAO,SAAS,QAAQ,MAAM,KAAK,QAAQ,MAAM,EAAE,KACpD,QAAQ,MAAM,KAAK,QAAQ,MAAM,KAEnC,MAAM,IAAI,MAAM,mDAAmD;AAEvE"}