@percepta/kaizen 0.8.0 → 0.9.0

package/dist/commands/studio.js

@@ -1,43 +0,0 @@
-import { packageRoot, resolveStateDir, workspaceRoot } from "../lib/paths.js";
-import { parseFlags, strFlag } from "../lib/parse-args.js";
-import { existsSync } from "node:fs";
-import { join } from "node:path";
-import { spawn } from "node:child_process";
-//#region src/commands/studio.ts
-async function runStudio(argv) {
-	const { flags } = parseFlags(argv);
-	const port = strFlag(flags, "port") ?? "6789";
-	const root = workspaceRoot();
-	const stateDir = resolveStateDir(root);
-	const dashboardRoot = join(packageRoot(), "dashboard", ".next", "standalone");
-	const serverPath = join(dashboardRoot, "packages", "kaizen", "dashboard", "server.js");
-	if (!existsSync(serverPath)) {
-		process.stderr.write(`kaizen studio: bundled dashboard not found at ${serverPath}.\nBuild the package first with \`pnpm --filter @percepta/kaizen build\`.
-`);
-		return 1;
-	}
-	const env = {
-		...process.env,
-		PORT: port,
-		KAIZEN_WORKSPACE: root,
-		KAIZEN_STATE_DIR: stateDir
-	};
-	process.stdout.write(`kaizen studio: ${root}\n`);
-	process.stdout.write(`state: ${stateDir}\n`);
-	process.stdout.write(`url: http://localhost:${port}\n`);
-	const next = spawn(process.execPath, [serverPath], {
-		cwd: dashboardRoot,
-		env,
-		stdio: "inherit"
-	});
-	const shutdown = () => next.kill("SIGTERM");
-	process.on("SIGINT", shutdown);
-	process.on("SIGTERM", shutdown);
-	return await new Promise((resolve) => {
-		next.on("exit", (code) => resolve(code ?? 0));
-	});
-}
-//#endregion
-export { runStudio };
-
-//# sourceMappingURL=studio.js.map

package/dist/commands/studio.js.map

@@ -1 +0,0 @@
-{"version":3,"file":"studio.js","names":[],"sources":["../../src/commands/studio.ts"],"sourcesContent":["import { spawn } from \"node:child_process\";\nimport { existsSync } from \"node:fs\";\nimport { join } from \"node:path\";\nimport { parseFlags, strFlag } from \"../lib/parse-args.js\";\nimport { packageRoot, resolveStateDir, workspaceRoot } from \"../lib/paths.js\";\n\nexport async function runStudio(argv: string[]): Promise<number> {\n  const { flags } = parseFlags(argv);\n  const port = strFlag(flags, \"port\") ?? \"6789\";\n  const root = workspaceRoot();\n  const stateDir = resolveStateDir(root);\n  const dashboardRoot = join(packageRoot(), \"dashboard\", \".next\", \"standalone\");\n  const serverPath = join(\n    dashboardRoot,\n    \"packages\",\n    \"kaizen\",\n    \"dashboard\",\n    \"server.js\",\n  );\n\n  if (!existsSync(serverPath)) {\n    process.stderr.write(\n      `kaizen studio: bundled dashboard not found at ${serverPath}.\\n` +\n        \"Build the package first with `pnpm --filter @percepta/kaizen build`.\\n\",\n    );\n    return 1;\n  }\n\n  const env = {\n    ...process.env,\n    PORT: port,\n    KAIZEN_WORKSPACE: root,\n    KAIZEN_STATE_DIR: stateDir,\n  };\n\n  process.stdout.write(`kaizen studio: ${root}\\n`);\n  process.stdout.write(`state: ${stateDir}\\n`);\n  process.stdout.write(`url: http://localhost:${port}\\n`);\n\n  const next = spawn(process.execPath, [serverPath], {\n    cwd: dashboardRoot,\n    env,\n    stdio: \"inherit\",\n  });\n\n  const shutdown = () => next.kill(\"SIGTERM\");\n  process.on(\"SIGINT\", shutdown);\n  process.on(\"SIGTERM\", shutdown);\n\n  return await new Promise<number>((resolve) => {\n    next.on(\"exit\", (code) => resolve(code ?? 0));\n  });\n}\n"],"mappings":";;;;;;AAMA,eAAsB,UAAU,MAAiC;CAC/D,MAAM,EAAE,UAAU,WAAW,KAAK;CAClC,MAAM,OAAO,QAAQ,OAAO,OAAO,IAAI;CACvC,MAAM,OAAO,eAAe;CAC5B,MAAM,WAAW,gBAAgB,KAAK;CACtC,MAAM,gBAAgB,KAAK,aAAa,EAAE,aAAa,SAAS,aAAa;CAC7E,MAAM,aAAa,KACjB,eACA,YACA,UACA,aACA,YACD;AAED,KAAI,CAAC,WAAW,WAAW,EAAE;AAC3B,UAAQ,OAAO,MACb,iDAAiD,WAAW;EAE7D;AACD,SAAO;;CAGT,MAAM,MAAM;EACV,GAAG,QAAQ;EACX,MAAM;EACN,kBAAkB;EAClB,kBAAkB;EACnB;AAED,SAAQ,OAAO,MAAM,kBAAkB,KAAK,IAAI;AAChD,SAAQ,OAAO,MAAM,UAAU,SAAS,IAAI;AAC5C,SAAQ,OAAO,MAAM,yBAAyB,KAAK,IAAI;CAEvD,MAAM,OAAO,MAAM,QAAQ,UAAU,CAAC,WAAW,EAAE;EACjD,KAAK;EACL;EACA,OAAO;EACR,CAAC;CAEF,MAAM,iBAAiB,KAAK,KAAK,UAAU;AAC3C,SAAQ,GAAG,UAAU,SAAS;AAC9B,SAAQ,GAAG,WAAW,SAAS;AAE/B,QAAO,MAAM,IAAI,SAAiB,YAAY;AAC5C,OAAK,GAAG,SAAS,SAAS,QAAQ,QAAQ,EAAE,CAAC;GAC7C"}

package/dist/lib/bootstrap.js.map

@@ -1 +0,0 @@
-{"version":3,"file":"bootstrap.js","names":[],"sources":["../../src/lib/bootstrap.ts"],"sourcesContent":["// Paired bootstrap on per-item score deltas.\n// Returns a 95% CI on the mean delta. Auto-promotion rule: lower bound > 0.\n//\n// Why bootstrap and not a t-test:\n//   - No normality assumption (judge scores are bounded, often skewed)\n//   - Works on tiny n (we typically have 30 items)\n//   - Two-sided 95% CI is the right shape for the auto-promotion gate\n//   - Trivially implementable in vanilla JS — no scipy / numpy / jStat dep\n\nexport interface BootstrapResult {\n  mean_delta: number;\n  ci_low: number;\n  ci_high: number;\n  n: number;\n}\n\n/**\n * Bootstrap a 95% CI for the mean of `deltas`.\n * Uses 10000 resamples by default; deterministic for a given `seed`.\n */\nexport function bootstrapMeanCI(\n  deltas: number[],\n  opts: { resamples?: number; seed?: number; alpha?: number } = {},\n): BootstrapResult {\n  const resamples = opts.resamples ?? 10000;\n  const alpha = opts.alpha ?? 0.05;\n  const n = deltas.length;\n  if (n === 0) {\n    return { mean_delta: 0, ci_low: 0, ci_high: 0, n: 0 };\n  }\n  const observed = mean(deltas);\n  const rng = mulberry32(opts.seed ?? 0xc0ffee);\n  const means = new Float64Array(resamples);\n  for (let r = 0; r < resamples; r++) {\n    let sum = 0;\n    for (let i = 0; i < n; i++) {\n      sum += deltas[(rng() * n) | 0];\n    }\n    means[r] = sum / n;\n  }\n  const sorted = Array.from(means).sort((a, b) => a - b);\n  const loIdx = Math.floor((alpha / 2) * resamples);\n  const hiIdx = Math.min(\n    resamples - 1,\n    Math.ceil((1 - alpha / 2) * resamples) - 1,\n  );\n  return {\n    mean_delta: observed,\n    ci_low: sorted[loIdx],\n    ci_high: sorted[hiIdx],\n    n,\n  };\n}\n\n/** Pair item scores from new run with baseline scores by id. Returns deltas only for matched ids. */\nexport function pairDeltas(\n  newScores: Map<string, number>,\n  baselineScores: Map<string, number>,\n): { deltas: number[]; matchedIds: string[] } {\n  const deltas: number[] = [];\n  const matchedIds: string[] = [];\n  for (const [id, s] of newScores) {\n    const b = baselineScores.get(id);\n    if (b === undefined) continue;\n    deltas.push(s - b);\n    matchedIds.push(id);\n  }\n  return { deltas, matchedIds };\n}\n\nfunction mean(xs: number[]): number {\n  if (xs.length === 0) return 0;\n  let s = 0;\n  for (const x of xs) s += x;\n  return s / xs.length;\n}\n\n// Small deterministic PRNG (mulberry32). Public domain.\nfunction mulberry32(seed: number): () => number {\n  let a = seed >>> 0;\n  return function () {\n    a = (a + 0x6d2b79f5) >>> 0;\n    let t = a;\n    t = Math.imul(t ^ (t >>> 15), t | 1);\n    t ^= t + Math.imul(t ^ (t >>> 7), t | 61);\n    return ((t ^ (t >>> 14)) >>> 0) / 4294967296;\n  };\n}\n"],"mappings":";;;;;AAoBA,SAAgB,gBACd,QACA,OAA8D,EAAE,EAC/C;CACjB,MAAM,YAAY,KAAK,aAAa;CACpC,MAAM,QAAQ,KAAK,SAAS;CAC5B,MAAM,IAAI,OAAO;AACjB,KAAI,MAAM,EACR,QAAO;EAAE,YAAY;EAAG,QAAQ;EAAG,SAAS;EAAG,GAAG;EAAG;CAEvD,MAAM,WAAW,KAAK,OAAO;CAC7B,MAAM,MAAM,WAAW,KAAK,QAAQ,SAAS;CAC7C,MAAM,QAAQ,IAAI,aAAa,UAAU;AACzC,MAAK,IAAI,IAAI,GAAG,IAAI,WAAW,KAAK;EAClC,IAAI,MAAM;AACV,OAAK,IAAI,IAAI,GAAG,IAAI,GAAG,IACrB,QAAO,OAAQ,KAAK,GAAG,IAAK;AAE9B,QAAM,KAAK,MAAM;;CAEnB,MAAM,SAAS,MAAM,KAAK,MAAM,CAAC,MAAM,GAAG,MAAM,IAAI,EAAE;CACtD,MAAM,QAAQ,KAAK,MAAO,QAAQ,IAAK,UAAU;CACjD,MAAM,QAAQ,KAAK,IACjB,YAAY,GACZ,KAAK,MAAM,IAAI,QAAQ,KAAK,UAAU,GAAG,EAC1C;AACD,QAAO;EACL,YAAY;EACZ,QAAQ,OAAO;EACf,SAAS,OAAO;EAChB;EACD;;;AAIH,SAAgB,WACd,WACA,gBAC4C;CAC5C,MAAM,SAAmB,EAAE;CAC3B,MAAM,aAAuB,EAAE;AAC/B,MAAK,MAAM,CAAC,IAAI,MAAM,WAAW;EAC/B,MAAM,IAAI,eAAe,IAAI,GAAG;AAChC,MAAI,MAAM,KAAA,EAAW;AACrB,SAAO,KAAK,IAAI,EAAE;AAClB,aAAW,KAAK,GAAG;;AAErB,QAAO;EAAE;EAAQ;EAAY;;AAG/B,SAAS,KAAK,IAAsB;AAClC,KAAI,GAAG,WAAW,EAAG,QAAO;CAC5B,IAAI,IAAI;AACR,MAAK,MAAM,KAAK,GAAI,MAAK;AACzB,QAAO,IAAI,GAAG;;AAIhB,SAAS,WAAW,MAA4B;CAC9C,IAAI,IAAI,SAAS;AACjB,QAAO,WAAY;AACjB,MAAK,IAAI,eAAgB;EACzB,IAAI,IAAI;AACR,MAAI,KAAK,KAAK,IAAK,MAAM,IAAK,IAAI,EAAE;AACpC,OAAK,IAAI,KAAK,KAAK,IAAK,MAAM,GAAI,IAAI,GAAG;AACzC,WAAS,IAAK,MAAM,QAAS,KAAK"}

package/dist/lib/cli.js

@@ -1,20 +0,0 @@
-import { kaizenConfigPath } from "../shared/workspace-paths.js";
-import "./paths.js";
-import { existsSync } from "node:fs";
-//#region src/lib/cli.ts
-function requireKaizenWorkspace(root) {
-	if (existsSync(kaizenConfigPath(root))) return true;
-	process.stderr.write(`no kaizen/config.ts in ${root}. run \`kaizen init\` first.\n`);
-	return false;
-}
-function isSystemId(value) {
-	return /^[a-z][a-z0-9-]*$/.test(value);
-}
-function pad(s, w) {
-	if (s.length >= w) return s.slice(0, w - 1) + " ";
-	return s + " ".repeat(w - s.length);
-}
-//#endregion
-export { isSystemId, pad, requireKaizenWorkspace };
-
-//# sourceMappingURL=cli.js.map

package/dist/lib/cli.js.map

@@ -1 +0,0 @@
-{"version":3,"file":"cli.js","names":[],"sources":["../../src/lib/cli.ts"],"sourcesContent":["import { existsSync } from \"node:fs\";\nimport { kaizenConfigPath } from \"./paths.js\";\n\nexport function requireKaizenWorkspace(root: string): boolean {\n  if (existsSync(kaizenConfigPath(root))) return true;\n  process.stderr.write(\n    `no kaizen/config.ts in ${root}. run \\`kaizen init\\` first.\\n`,\n  );\n  return false;\n}\n\nexport function isSystemId(value: string): boolean {\n  return /^[a-z][a-z0-9-]*$/.test(value);\n}\n\nexport function pad(s: string, w: number): string {\n  if (s.length >= w) return s.slice(0, w - 1) + \" \";\n  return s + \" \".repeat(w - s.length);\n}\n"],"mappings":";;;;AAGA,SAAgB,uBAAuB,MAAuB;AAC5D,KAAI,WAAW,iBAAiB,KAAK,CAAC,CAAE,QAAO;AAC/C,SAAQ,OAAO,MACb,0BAA0B,KAAK,gCAChC;AACD,QAAO;;AAGT,SAAgB,WAAW,OAAwB;AACjD,QAAO,oBAAoB,KAAK,MAAM;;AAGxC,SAAgB,IAAI,GAAW,GAAmB;AAChD,KAAI,EAAE,UAAU,EAAG,QAAO,EAAE,MAAM,GAAG,IAAI,EAAE,GAAG;AAC9C,QAAO,IAAI,IAAI,OAAO,IAAI,EAAE,OAAO"}

package/dist/lib/events.js.map

@@ -1 +0,0 @@
-{"version":3,"file":"events.js","names":[],"sources":["../../src/lib/events.ts"],"sourcesContent":["// Event schema for the NDJSON contract between eval scripts and the runner.\n\ninterface BaseEvent {\n  type: string;\n  ts?: number;\n}\n\ninterface StartEvent extends BaseEvent {\n  type: \"start\";\n  n: number;\n  eval_version: number;\n  dataset_version: string;\n  variant?: string;\n}\n\nexport interface ItemEvent extends BaseEvent {\n  type: \"item\";\n  id: string;\n  score: number;\n  breakdown?: Record<string, number>;\n  trace_id?: string | null;\n  why?: string;\n  subgroup?: Record<string, string>;\n}\n\ninterface ProgressEvent extends BaseEvent {\n  type: \"progress\";\n  done: number;\n  total: number;\n}\n\nexport interface CompleteEvent extends BaseEvent {\n  type: \"complete\";\n  score: number;\n  breakdown?: Record<string, number>;\n  n: number;\n  worst_traces?: Array<{ id: string; score: number; trace_id?: string | null }>;\n}\n\ninterface ErrorEvent extends BaseEvent {\n  type: \"error\";\n  message: string;\n}\n\n// Synthesized by the runner, not emitted by the eval.\nexport interface CrashedEvent extends BaseEvent {\n  type: \"crashed\";\n  exit_code: number | null;\n  signal?: string | null;\n  stderr_tail?: string;\n  reason: string;\n}\n\ninterface AbortedEvent extends BaseEvent {\n  type: \"aborted\";\n  reason: string;\n}\n\nexport interface PromotionEvent extends BaseEvent {\n  type: \"promotion\";\n  promoted: boolean;\n  rule:\n    | \"auto\"\n    | \"force\"\n    | \"no_baseline\"\n    | \"version_mismatch\"\n    | \"stat_insufficient\"\n    | \"subgroup_regression\";\n  ci_low?: number;\n  ci_high?: number;\n  mean_delta?: number;\n  n_compared?: number;\n  parent_run_id?: string;\n  details?: string;\n}\n\nexport type Event =\n  | StartEvent\n  | ItemEvent\n  | ProgressEvent\n  | CompleteEvent\n  | ErrorEvent\n  | CrashedEvent\n  | AbortedEvent\n  | PromotionEvent;\n\nexport function validateEvent(value: unknown): Event {\n  if (!value || typeof value !== \"object\") {\n    throw new Error(\"event must be an object\");\n  }\n  const e = value as Record<string, unknown>;\n  if (typeof e.type !== \"string\") {\n    throw new Error(\"event.type must be a string\");\n  }\n\n  switch (e.type) {\n    case \"start\":\n      requireNumber(e, \"n\");\n      requireNumber(e, \"eval_version\");\n      requireString(e, \"dataset_version\");\n      break;\n    case \"item\":\n      requireString(e, \"id\");\n      requireNumber(e, \"score\");\n      break;\n    case \"progress\":\n      requireNumber(e, \"done\");\n      requireNumber(e, \"total\");\n      break;\n    case \"complete\":\n      requireNumber(e, \"score\");\n      requireNumber(e, \"n\");\n      break;\n    case \"error\":\n      requireString(e, \"message\");\n      break;\n    case \"crashed\":\n      requireString(e, \"reason\");\n      break;\n    case \"aborted\":\n      requireString(e, \"reason\");\n      break;\n    case \"promotion\":\n      if (typeof e.promoted !== \"boolean\")\n        throw new Error(\"promotion.promoted must be boolean\");\n      requireString(e, \"rule\");\n      break;\n    default:\n      throw new Error(`unknown event type: ${e.type}`);\n  }\n\n  if (e.ts !== undefined && typeof e.ts !== \"number\") {\n    throw new Error(`${e.type}.ts must be a number when present`);\n  }\n  return e as unknown as Event;\n}\n\nfunction requireString(e: Record<string, unknown>, key: string): void {\n  if (typeof e[key] !== \"string\" || e[key] === \"\") {\n    throw new Error(`${String(e.type)}.${key} must be a non-empty string`);\n  }\n}\n\nfunction requireNumber(e: Record<string, unknown>, key: string): void {\n  if (typeof e[key] !== \"number\" || !Number.isFinite(e[key])) {\n    throw new Error(`${String(e.type)}.${key} must be a finite number`);\n  }\n}\n\nfunction isTerminal(\n  e: Event,\n): e is CompleteEvent | CrashedEvent | AbortedEvent {\n  return e.type === \"complete\" || e.type === \"crashed\" || e.type === \"aborted\";\n}\n\nexport function findTerminal(\n  events: Event[],\n): CompleteEvent | CrashedEvent | AbortedEvent | null {\n  for (let i = events.length - 1; i >= 0; i--) {\n    if (isTerminal(events[i]))\n      return events[i] as CompleteEvent | CrashedEvent | AbortedEvent;\n  }\n  return null;\n}\n\n/** Stream-mode NDJSON parser. Feed `push(chunk)` from a Readable; receive parsed events via `onEvent`. */\nexport class NdjsonReader {\n  private buffer = \"\";\n  public constructor(\n    private onEvent: (e: Event) => void,\n    private onParseError?: (line: string, err: unknown) => void,\n  ) {}\n\n  public push(chunk: string | Buffer): void {\n    this.buffer += typeof chunk === \"string\" ? chunk : chunk.toString(\"utf-8\");\n    let nl: number;\n    while ((nl = this.buffer.indexOf(\"\\n\")) !== -1) {\n      const line = this.buffer.slice(0, nl);\n      this.buffer = this.buffer.slice(nl + 1);\n      if (!line.trim()) continue;\n      try {\n        this.onEvent(validateEvent(JSON.parse(line)));\n      } catch (err) {\n        this.onParseError?.(line, err);\n      }\n    }\n  }\n\n  /** Flush any remaining unterminated content as a best-effort parse. */\n  public end(): void {\n    const tail = this.buffer.trim();\n    this.buffer = \"\";\n    if (!tail) return;\n    try {\n      this.onEvent(validateEvent(JSON.parse(tail)));\n    } catch (err) {\n      this.onParseError?.(tail, err);\n    }\n  }\n}\n"],"mappings":";AAsFA,SAAgB,cAAc,OAAuB;AACnD,KAAI,CAAC,SAAS,OAAO,UAAU,SAC7B,OAAM,IAAI,MAAM,0BAA0B;CAE5C,MAAM,IAAI;AACV,KAAI,OAAO,EAAE,SAAS,SACpB,OAAM,IAAI,MAAM,8BAA8B;AAGhD,SAAQ,EAAE,MAAV;EACE,KAAK;AACH,iBAAc,GAAG,IAAI;AACrB,iBAAc,GAAG,eAAe;AAChC,iBAAc,GAAG,kBAAkB;AACnC;EACF,KAAK;AACH,iBAAc,GAAG,KAAK;AACtB,iBAAc,GAAG,QAAQ;AACzB;EACF,KAAK;AACH,iBAAc,GAAG,OAAO;AACxB,iBAAc,GAAG,QAAQ;AACzB;EACF,KAAK;AACH,iBAAc,GAAG,QAAQ;AACzB,iBAAc,GAAG,IAAI;AACrB;EACF,KAAK;AACH,iBAAc,GAAG,UAAU;AAC3B;EACF,KAAK;AACH,iBAAc,GAAG,SAAS;AAC1B;EACF,KAAK;AACH,iBAAc,GAAG,SAAS;AAC1B;EACF,KAAK;AACH,OAAI,OAAO,EAAE,aAAa,UACxB,OAAM,IAAI,MAAM,qCAAqC;AACvD,iBAAc,GAAG,OAAO;AACxB;EACF,QACE,OAAM,IAAI,MAAM,uBAAuB,EAAE,OAAO;;AAGpD,KAAI,EAAE,OAAO,KAAA,KAAa,OAAO,EAAE,OAAO,SACxC,OAAM,IAAI,MAAM,GAAG,EAAE,KAAK,mCAAmC;AAE/D,QAAO;;AAGT,SAAS,cAAc,GAA4B,KAAmB;AACpE,KAAI,OAAO,EAAE,SAAS,YAAY,EAAE,SAAS,GAC3C,OAAM,IAAI,MAAM,GAAG,OAAO,EAAE,KAAK,CAAC,GAAG,IAAI,6BAA6B;;AAI1E,SAAS,cAAc,GAA4B,KAAmB;AACpE,KAAI,OAAO,EAAE,SAAS,YAAY,CAAC,OAAO,SAAS,EAAE,KAAK,CACxD,OAAM,IAAI,MAAM,GAAG,OAAO,EAAE,KAAK,CAAC,GAAG,IAAI,0BAA0B;;AAIvE,SAAS,WACP,GACkD;AAClD,QAAO,EAAE,SAAS,cAAc,EAAE,SAAS,aAAa,EAAE,SAAS;;AAGrE,SAAgB,aACd,QACoD;AACpD,MAAK,IAAI,IAAI,OAAO,SAAS,GAAG,KAAK,GAAG,IACtC,KAAI,WAAW,OAAO,GAAG,CACvB,QAAO,OAAO;AAElB,QAAO;;;AAIT,IAAa,eAAb,MAA0B;CACxB,SAAiB;CACjB,YACE,SACA,cACA;AAFQ,OAAA,UAAA;AACA,OAAA,eAAA;;CAGV,KAAY,OAA8B;AACxC,OAAK,UAAU,OAAO,UAAU,WAAW,QAAQ,MAAM,SAAS,QAAQ;EAC1E,IAAI;AACJ,UAAQ,KAAK,KAAK,OAAO,QAAQ,KAAK,MAAM,IAAI;GAC9C,MAAM,OAAO,KAAK,OAAO,MAAM,GAAG,GAAG;AACrC,QAAK,SAAS,KAAK,OAAO,MAAM,KAAK,EAAE;AACvC,OAAI,CAAC,KAAK,MAAM,CAAE;AAClB,OAAI;AACF,SAAK,QAAQ,cAAc,KAAK,MAAM,KAAK,CAAC,CAAC;YACtC,KAAK;AACZ,SAAK,eAAe,MAAM,IAAI;;;;;CAMpC,MAAmB;EACjB,MAAM,OAAO,KAAK,OAAO,MAAM;AAC/B,OAAK,SAAS;AACd,MAAI,CAAC,KAAM;AACX,MAAI;AACF,QAAK,QAAQ,cAAc,KAAK,MAAM,KAAK,CAAC,CAAC;WACtC,KAAK;AACZ,QAAK,eAAe,MAAM,IAAI"}

package/dist/lib/fs-utils.js.map

@@ -1 +0,0 @@
-{"version":3,"file":"fs-utils.js","names":[],"sources":["../../src/lib/fs-utils.ts"],"sourcesContent":["import { existsSync, mkdirSync, readFileSync, writeFileSync } from \"node:fs\";\nimport { dirname } from \"node:path\";\n\nexport function ensureDir(p: string): void {\n  mkdirSync(p, { recursive: true });\n}\n\nexport function writeFileSafely(\n  p: string,\n  contents: string,\n  opts: { overwrite?: boolean } = {},\n): boolean {\n  if (existsSync(p) && !opts.overwrite) return false;\n  ensureDir(dirname(p));\n  writeFileSync(p, contents);\n  return true;\n}\n\nexport function applyVars(text: string, vars: Record<string, string>): string {\n  return text.replace(/\\{\\{\\s*(\\w+)\\s*\\}\\}/g, (m, key) =>\n    key in vars ? vars[key] : m,\n  );\n}\n\nexport function appendIfMissing(path: string, line: string): boolean {\n  const existing = existsSync(path) ? readFileSync(path, \"utf-8\") : \"\";\n  if (existing.split(\"\\n\").some((l) => l.trim() === line.trim())) return false;\n  ensureDir(dirname(path));\n  const sep = existing.length === 0 || existing.endsWith(\"\\n\") ? \"\" : \"\\n\";\n  writeFileSync(path, existing + sep + line + \"\\n\");\n  return true;\n}\n"],"mappings":";;;AAGA,SAAgB,UAAU,GAAiB;AACzC,WAAU,GAAG,EAAE,WAAW,MAAM,CAAC;;AAGnC,SAAgB,gBACd,GACA,UACA,OAAgC,EAAE,EACzB;AACT,KAAI,WAAW,EAAE,IAAI,CAAC,KAAK,UAAW,QAAO;AAC7C,WAAU,QAAQ,EAAE,CAAC;AACrB,eAAc,GAAG,SAAS;AAC1B,QAAO;;AAGT,SAAgB,UAAU,MAAc,MAAsC;AAC5E,QAAO,KAAK,QAAQ,yBAAyB,GAAG,QAC9C,OAAO,OAAO,KAAK,OAAO,EAC3B;;AAGH,SAAgB,gBAAgB,MAAc,MAAuB;CACnE,MAAM,WAAW,WAAW,KAAK,GAAG,aAAa,MAAM,QAAQ,GAAG;AAClE,KAAI,SAAS,MAAM,KAAK,CAAC,MAAM,MAAM,EAAE,MAAM,KAAK,KAAK,MAAM,CAAC,CAAE,QAAO;AACvE,WAAU,QAAQ,KAAK,CAAC;AAExB,eAAc,MAAM,YADR,SAAS,WAAW,KAAK,SAAS,SAAS,KAAK,GAAG,KAAK,QAC/B,OAAO,KAAK;AACjD,QAAO"}

package/dist/lib/leaderboard.js.map

@@ -1 +0,0 @@
-{"version":3,"file":"leaderboard.js","names":[],"sources":["../../src/lib/leaderboard.ts"],"sourcesContent":["import { existsSync, readdirSync, statSync } from \"node:fs\";\nimport { join } from \"node:path\";\nimport { systemRunsDir, readJsonIfExists } from \"./run-dir.js\";\n\nexport interface RunSummary {\n  run_id: string;\n  system: string;\n  variant: string;\n  parent_id: string | null;\n  hypothesis: string;\n  status: \"complete\" | \"crashed\" | \"aborted\" | \"running\";\n  score: number | null;\n  n: number | null;\n  promoted: boolean | null;\n  started_at: string;\n  ended_at: string | null;\n  eval_version: number;\n  dataset_version: string;\n  events_path: string;\n  state_path: string;\n  manifest_path: string;\n  linear_issue_id?: string | null;\n  linear_issue_url?: string | null;\n}\n\nexport interface ManifestFile {\n  run_id: string;\n  system: string;\n  variant: string;\n  parent_id: string | null;\n  hypothesis: string;\n  git_sha?: string | null;\n  git_branch?: string | null;\n  worktree_root?: string | null;\n  git_common_dir?: string | null;\n  eval_version: number;\n  dataset_version: string;\n  started_at: string;\n  host?: string;\n  kaizen_version?: string;\n  state_dir?: string;\n  diagnostic?: boolean;\n  linear_issue_id?: string | null;\n  linear_issue_url?: string | null;\n}\n\nexport interface StateFile {\n  run_id: string;\n  status: RunSummary[\"status\"];\n  score: number | null;\n  n_total?: number | null;\n  n_done?: number;\n  promoted?: boolean | null;\n  ended_at?: string | null;\n}\n\nexport function listRuns(stateDir: string, systemId: string): RunSummary[] {\n  const dir = systemRunsDir(stateDir, systemId);\n  if (!existsSync(dir)) return [];\n  const out: RunSummary[] = [];\n  for (const entry of readdirSync(dir)) {\n    const runDir = join(dir, entry);\n    if (!statSync(runDir).isDirectory()) continue;\n    const manifest_path = join(runDir, \"manifest.json\");\n    const state_path = join(runDir, \"state.json\");\n    const events_path = join(runDir, \"events.jsonl\");\n    const manifest = readJsonIfExists<ManifestFile>(manifest_path);\n    if (!manifest) continue;\n    const state = readJsonIfExists<StateFile>(state_path);\n    out.push({\n      run_id: manifest.run_id,\n      system: manifest.system,\n      variant: manifest.variant,\n      parent_id: manifest.parent_id,\n      hypothesis: manifest.hypothesis,\n      status: state?.status ?? \"running\",\n      score: state?.score ?? null,\n      n: state?.n_total ?? null,\n      promoted: state?.promoted ?? null,\n      started_at: manifest.started_at,\n      ended_at: state?.ended_at ?? null,\n      eval_version: manifest.eval_version,\n      dataset_version: manifest.dataset_version,\n      events_path,\n      state_path,\n      manifest_path,\n      linear_issue_id: manifest.linear_issue_id ?? null,\n      linear_issue_url: manifest.linear_issue_url ?? null,\n    });\n  }\n  // Newest first\n  out.sort((a, b) => (b.started_at < a.started_at ? -1 : 1));\n  return out;\n}\n\n/**\n * Promoted baseline = the most recent run that was promoted, under matching versions.\n * Falls back to the highest-scored complete run if nothing has been promoted yet\n * (e.g. the very first run before promotion semantics kick in).\n */\nexport function currentBaseline(\n  runs: RunSummary[],\n  evalVersion: number,\n  datasetVersion: string,\n): RunSummary | null {\n  const eligible = runs.filter(\n    (r) =>\n      r.status === \"complete\" &&\n      r.score !== null &&\n      r.eval_version === evalVersion &&\n      r.dataset_version === datasetVersion,\n  );\n  if (eligible.length === 0) return null;\n  // listRuns sorts newest-first; find the newest with promoted: true.\n  const promoted = eligible.find((r) => r.promoted === true);\n  if (promoted) return promoted;\n  // Fallback: highest-scored eligible run when promotion data is unavailable.\n  let best = eligible[0];\n  for (const r of eligible) {\n    if (r.score! > best.score!) best = r;\n  }\n  return best;\n}\n"],"mappings":";;;;AAwDA,SAAgB,SAAS,UAAkB,UAAgC;CACzE,MAAM,MAAM,cAAc,UAAU,SAAS;AAC7C,KAAI,CAAC,WAAW,IAAI,CAAE,QAAO,EAAE;CAC/B,MAAM,MAAoB,EAAE;AAC5B,MAAK,MAAM,SAAS,YAAY,IAAI,EAAE;EACpC,MAAM,SAAS,KAAK,KAAK,MAAM;AAC/B,MAAI,CAAC,SAAS,OAAO,CAAC,aAAa,CAAE;EACrC,MAAM,gBAAgB,KAAK,QAAQ,gBAAgB;EACnD,MAAM,aAAa,KAAK,QAAQ,aAAa;EAC7C,MAAM,cAAc,KAAK,QAAQ,eAAe;EAChD,MAAM,WAAW,iBAA+B,cAAc;AAC9D,MAAI,CAAC,SAAU;EACf,MAAM,QAAQ,iBAA4B,WAAW;AACrD,MAAI,KAAK;GACP,QAAQ,SAAS;GACjB,QAAQ,SAAS;GACjB,SAAS,SAAS;GAClB,WAAW,SAAS;GACpB,YAAY,SAAS;GACrB,QAAQ,OAAO,UAAU;GACzB,OAAO,OAAO,SAAS;GACvB,GAAG,OAAO,WAAW;GACrB,UAAU,OAAO,YAAY;GAC7B,YAAY,SAAS;GACrB,UAAU,OAAO,YAAY;GAC7B,cAAc,SAAS;GACvB,iBAAiB,SAAS;GAC1B;GACA;GACA;GACA,iBAAiB,SAAS,mBAAmB;GAC7C,kBAAkB,SAAS,oBAAoB;GAChD,CAAC;;AAGJ,KAAI,MAAM,GAAG,MAAO,EAAE,aAAa,EAAE,aAAa,KAAK,EAAG;AAC1D,QAAO;;;;;;;AAQT,SAAgB,gBACd,MACA,aACA,gBACmB;CACnB,MAAM,WAAW,KAAK,QACnB,MACC,EAAE,WAAW,cACb,EAAE,UAAU,QACZ,EAAE,iBAAiB,eACnB,EAAE,oBAAoB,eACzB;AACD,KAAI,SAAS,WAAW,EAAG,QAAO;CAElC,MAAM,WAAW,SAAS,MAAM,MAAM,EAAE,aAAa,KAAK;AAC1D,KAAI,SAAU,QAAO;CAErB,IAAI,OAAO,SAAS;AACpB,MAAK,MAAM,KAAK,SACd,KAAI,EAAE,QAAS,KAAK,MAAQ,QAAO;AAErC,QAAO"}

package/dist/lib/parse-args.js.map

@@ -1 +0,0 @@
-{"version":3,"file":"parse-args.js","names":[],"sources":["../../src/lib/parse-args.ts"],"sourcesContent":["export type Flags = Record<string, string | boolean>;\n\nexport function parseFlags(argv: string[]): {\n  positional: string[];\n  flags: Flags;\n} {\n  const positional: string[] = [];\n  const flags: Flags = {};\n  for (let i = 0; i < argv.length; i++) {\n    const a = argv[i];\n    if (a.startsWith(\"--\")) {\n      const eq = a.indexOf(\"=\");\n      let key: string, valFromEq: string | undefined;\n      if (eq >= 0) {\n        key = a.slice(2, eq);\n        valFromEq = a.slice(eq + 1);\n      } else {\n        key = a.slice(2);\n      }\n      if (valFromEq !== undefined) {\n        flags[key] = valFromEq;\n        continue;\n      }\n      const next = argv[i + 1];\n      // A flag's value is the next token unless it looks like another flag (starts with `-`).\n      if (next === undefined || next.startsWith(\"-\")) {\n        flags[key] = true;\n      } else {\n        flags[key] = next;\n        i++;\n      }\n    } else if (a.startsWith(\"-\") && a.length > 1) {\n      // Short flag like `-n 3`. Treat similarly to long flags.\n      const key = a.slice(1);\n      const next = argv[i + 1];\n      if (next === undefined || next.startsWith(\"-\")) {\n        flags[key] = true;\n      } else {\n        flags[key] = next;\n        i++;\n      }\n    } else {\n      positional.push(a);\n    }\n  }\n  return { positional, flags };\n}\n\nexport function strFlag(flags: Flags, key: string): string | undefined {\n  const v = flags[key];\n  return typeof v === \"string\" ? v : undefined;\n}\n\nexport function boolFlag(flags: Flags, key: string): boolean {\n  return flags[key] === true || flags[key] === \"true\";\n}\n"],"mappings":";AAEA,SAAgB,WAAW,MAGzB;CACA,MAAM,aAAuB,EAAE;CAC/B,MAAM,QAAe,EAAE;AACvB,MAAK,IAAI,IAAI,GAAG,IAAI,KAAK,QAAQ,KAAK;EACpC,MAAM,IAAI,KAAK;AACf,MAAI,EAAE,WAAW,KAAK,EAAE;GACtB,MAAM,KAAK,EAAE,QAAQ,IAAI;GACzB,IAAI,KAAa;AACjB,OAAI,MAAM,GAAG;AACX,UAAM,EAAE,MAAM,GAAG,GAAG;AACpB,gBAAY,EAAE,MAAM,KAAK,EAAE;SAE3B,OAAM,EAAE,MAAM,EAAE;AAElB,OAAI,cAAc,KAAA,GAAW;AAC3B,UAAM,OAAO;AACb;;GAEF,MAAM,OAAO,KAAK,IAAI;AAEtB,OAAI,SAAS,KAAA,KAAa,KAAK,WAAW,IAAI,CAC5C,OAAM,OAAO;QACR;AACL,UAAM,OAAO;AACb;;aAEO,EAAE,WAAW,IAAI,IAAI,EAAE,SAAS,GAAG;GAE5C,MAAM,MAAM,EAAE,MAAM,EAAE;GACtB,MAAM,OAAO,KAAK,IAAI;AACtB,OAAI,SAAS,KAAA,KAAa,KAAK,WAAW,IAAI,CAC5C,OAAM,OAAO;QACR;AACL,UAAM,OAAO;AACb;;QAGF,YAAW,KAAK,EAAE;;AAGtB,QAAO;EAAE;EAAY;EAAO;;AAG9B,SAAgB,QAAQ,OAAc,KAAiC;CACrE,MAAM,IAAI,MAAM;AAChB,QAAO,OAAO,MAAM,WAAW,IAAI,KAAA;;AAGrC,SAAgB,SAAS,OAAc,KAAsB;AAC3D,QAAO,MAAM,SAAS,QAAQ,MAAM,SAAS"}

package/dist/lib/paths.js

@@ -1,22 +0,0 @@
-import { resolveKaizenStateDir } from "../shared/workspace-paths.js";
-import { dirname, resolve } from "node:path";
-import { fileURLToPath } from "node:url";
-//#region src/lib/paths.ts
-const HERE = dirname(fileURLToPath(import.meta.url));
-function packageRoot() {
-	return resolve(HERE, "..", "..");
-}
-function templatesDir() {
-	return resolve(packageRoot(), "templates");
-}
-function workspaceRoot() {
-	const raw = process.env.KAIZEN_WORKSPACE;
-	return raw ? resolve(raw) : process.cwd();
-}
-function resolveStateDir(workspace) {
-	return resolveKaizenStateDir(workspace);
-}
-//#endregion
-export { packageRoot, resolveStateDir, templatesDir, workspaceRoot };
-
-//# sourceMappingURL=paths.js.map

package/dist/lib/paths.js.map

@@ -1 +0,0 @@
-{"version":3,"file":"paths.js","names":[],"sources":["../../src/lib/paths.ts"],"sourcesContent":["import { dirname, resolve } from \"node:path\";\nimport { fileURLToPath } from \"node:url\";\nimport {\n  defaultKaizenStateDir,\n  kaizenConfigPath,\n  kaizenDir,\n  kaizenSystemDir,\n  kaizenSystemPath,\n  kaizenSystemsDir,\n  primaryWorktreeRoot,\n  resolveKaizenStateDir,\n} from \"../../shared/workspace-paths.js\";\n\nconst HERE = dirname(fileURLToPath(import.meta.url));\n\nexport function packageRoot(): string {\n  // src/lib/paths.ts -> package root\n  return resolve(HERE, \"..\", \"..\");\n}\n\nexport function templatesDir(): string {\n  return resolve(packageRoot(), \"templates\");\n}\n\nexport function workspaceRoot(): string {\n  const raw = process.env.KAIZEN_WORKSPACE;\n  return raw ? resolve(raw) : process.cwd();\n}\n\nexport {\n  defaultKaizenStateDir,\n  kaizenConfigPath,\n  kaizenDir,\n  kaizenSystemDir,\n  kaizenSystemPath,\n  kaizenSystemsDir,\n  primaryWorktreeRoot,\n};\n\nexport function resolveStateDir(workspace: string): string {\n  return resolveKaizenStateDir(workspace);\n}\n"],"mappings":";;;;AAaA,MAAM,OAAO,QAAQ,cAAc,OAAO,KAAK,IAAI,CAAC;AAEpD,SAAgB,cAAsB;AAEpC,QAAO,QAAQ,MAAM,MAAM,KAAK;;AAGlC,SAAgB,eAAuB;AACrC,QAAO,QAAQ,aAAa,EAAE,YAAY;;AAG5C,SAAgB,gBAAwB;CACtC,MAAM,MAAM,QAAQ,IAAI;AACxB,QAAO,MAAM,QAAQ,IAAI,GAAG,QAAQ,KAAK;;AAa3C,SAAgB,gBAAgB,WAA2B;AACzD,QAAO,sBAAsB,UAAU"}

package/dist/lib/promotion.js.map

@@ -1 +0,0 @@
-{"version":3,"file":"promotion.js","names":[],"sources":["../../src/lib/promotion.ts"],"sourcesContent":["import { bootstrapMeanCI, pairDeltas } from \"./bootstrap.js\";\nimport type { Event, ItemEvent, PromotionEvent } from \"./events.js\";\nimport { readNdjson } from \"./run-dir.js\";\n\nexport interface BaselineSummary {\n  run_id: string;\n  variant: string;\n  score: number;\n  events_path: string; // path to events.jsonl, used to lazily read item scores\n}\n\nexport type PromotionDecision = Omit<PromotionEvent, \"type\" | \"ts\">;\n\n/** Read item scores from a run's events.jsonl. Returns Map<item_id, score>. */\nfunction readItemScores(eventsPath: string): Map<string, number> {\n  const events = readNdjson<Event>(eventsPath);\n  const scores = new Map<string, number>();\n  for (const e of events) {\n    if (e.type === \"item\") {\n      const it = e;\n      scores.set(it.id, it.score);\n    }\n  }\n  return scores;\n}\n\nexport interface DecideOpts {\n  /** Set of subgroup keys to guard against regressions. */\n  subgroupKeys?: string[];\n  /** Bootstrap resamples; default 10k. */\n  resamples?: number;\n  /** Seed for determinism. */\n  seed?: number;\n}\n\n/**\n * Decide auto-promotion given the current run's item scores and an optional baseline.\n * Returns a structured decision suitable for emitting as a `promotion` event.\n */\nexport function decidePromotion(\n  currentScores: Map<string, number>,\n  currentItemEvents: ItemEvent[],\n  baseline: BaselineSummary | null,\n  opts: DecideOpts = {},\n): PromotionDecision {\n  if (!baseline) {\n    return {\n      promoted: true,\n      rule: \"no_baseline\",\n      n_compared: 0,\n      details: \"no prior baseline; this run becomes the baseline by default\",\n    };\n  }\n\n  const baselineScores = readItemScores(baseline.events_path);\n  const { deltas, matchedIds } = pairDeltas(currentScores, baselineScores);\n  if (deltas.length < 5) {\n    return {\n      promoted: false,\n      rule: \"stat_insufficient\",\n      n_compared: deltas.length,\n      parent_run_id: baseline.run_id,\n      details: `only ${deltas.length} items paired with baseline ${baseline.run_id}; need ≥5 to bootstrap`,\n    };\n  }\n\n  const ci = bootstrapMeanCI(deltas, {\n    resamples: opts.resamples,\n    seed: opts.seed,\n  });\n  if (!(ci.ci_low > 0)) {\n    return {\n      promoted: false,\n      rule: \"stat_insufficient\",\n      ci_low: ci.ci_low,\n      ci_high: ci.ci_high,\n      mean_delta: ci.mean_delta,\n      n_compared: ci.n,\n      parent_run_id: baseline.run_id,\n      details: `95% CI on mean delta is [${ci.ci_low.toFixed(4)}, ${ci.ci_high.toFixed(4)}]; lower bound is not > 0`,\n    };\n  }\n\n  // Headline gate passed. Now check subgroup guards.\n  if (opts.subgroupKeys && opts.subgroupKeys.length > 0) {\n    const subgroupOf = (id: string, key: string): string | undefined => {\n      for (const ev of currentItemEvents) {\n        if (ev.id === id && ev.subgroup) return ev.subgroup[key];\n      }\n      return undefined;\n    };\n    for (const key of opts.subgroupKeys) {\n      const groups = new Map<string, number[]>();\n      for (let i = 0; i < matchedIds.length; i++) {\n        const sg = subgroupOf(matchedIds[i], key);\n        if (sg === undefined) continue;\n        if (!groups.has(sg)) groups.set(sg, []);\n        groups.get(sg)!.push(deltas[i]);\n      }\n      for (const [sg, sgDeltas] of groups) {\n        if (sgDeltas.length < 5) continue; // not enough to test; ignore\n        const sgCi = bootstrapMeanCI(sgDeltas, {\n          resamples: opts.resamples,\n          seed: opts.seed,\n        });\n        if (sgCi.ci_high < 0) {\n          return {\n            promoted: false,\n            rule: \"subgroup_regression\",\n            ci_low: ci.ci_low,\n            ci_high: ci.ci_high,\n            mean_delta: ci.mean_delta,\n            n_compared: ci.n,\n            parent_run_id: baseline.run_id,\n            details: `subgroup ${key}=${sg} regressed: 95% CI [${sgCi.ci_low.toFixed(4)}, ${sgCi.ci_high.toFixed(4)}]`,\n          };\n        }\n      }\n    }\n  }\n\n  return {\n    promoted: true,\n    rule: \"auto\",\n    ci_low: ci.ci_low,\n    ci_high: ci.ci_high,\n    mean_delta: ci.mean_delta,\n    n_compared: ci.n,\n    parent_run_id: baseline.run_id,\n  };\n}\n"],"mappings":";;;;AAcA,SAAS,eAAe,YAAyC;CAC/D,MAAM,SAAS,WAAkB,WAAW;CAC5C,MAAM,yBAAS,IAAI,KAAqB;AACxC,MAAK,MAAM,KAAK,OACd,KAAI,EAAE,SAAS,QAAQ;EACrB,MAAM,KAAK;AACX,SAAO,IAAI,GAAG,IAAI,GAAG,MAAM;;AAG/B,QAAO;;;;;;AAgBT,SAAgB,gBACd,eACA,mBACA,UACA,OAAmB,EAAE,EACF;AACnB,KAAI,CAAC,SACH,QAAO;EACL,UAAU;EACV,MAAM;EACN,YAAY;EACZ,SAAS;EACV;CAIH,MAAM,EAAE,QAAQ,eAAe,WAAW,eADnB,eAAe,SAAS,YACwB,CAAC;AACxE,KAAI,OAAO,SAAS,EAClB,QAAO;EACL,UAAU;EACV,MAAM;EACN,YAAY,OAAO;EACnB,eAAe,SAAS;EACxB,SAAS,QAAQ,OAAO,OAAO,8BAA8B,SAAS,OAAO;EAC9E;CAGH,MAAM,KAAK,gBAAgB,QAAQ;EACjC,WAAW,KAAK;EAChB,MAAM,KAAK;EACZ,CAAC;AACF,KAAI,EAAE,GAAG,SAAS,GAChB,QAAO;EACL,UAAU;EACV,MAAM;EACN,QAAQ,GAAG;EACX,SAAS,GAAG;EACZ,YAAY,GAAG;EACf,YAAY,GAAG;EACf,eAAe,SAAS;EACxB,SAAS,4BAA4B,GAAG,OAAO,QAAQ,EAAE,CAAC,IAAI,GAAG,QAAQ,QAAQ,EAAE,CAAC;EACrF;AAIH,KAAI,KAAK,gBAAgB,KAAK,aAAa,SAAS,GAAG;EACrD,MAAM,cAAc,IAAY,QAAoC;AAClE,QAAK,MAAM,MAAM,kBACf,KAAI,GAAG,OAAO,MAAM,GAAG,SAAU,QAAO,GAAG,SAAS;;AAIxD,OAAK,MAAM,OAAO,KAAK,cAAc;GACnC,MAAM,yBAAS,IAAI,KAAuB;AAC1C,QAAK,IAAI,IAAI,GAAG,IAAI,WAAW,QAAQ,KAAK;IAC1C,MAAM,KAAK,WAAW,WAAW,IAAI,IAAI;AACzC,QAAI,OAAO,KAAA,EAAW;AACtB,QAAI,CAAC,OAAO,IAAI,GAAG,CAAE,QAAO,IAAI,IAAI,EAAE,CAAC;AACvC,WAAO,IAAI,GAAG,CAAE,KAAK,OAAO,GAAG;;AAEjC,QAAK,MAAM,CAAC,IAAI,aAAa,QAAQ;AACnC,QAAI,SAAS,SAAS,EAAG;IACzB,MAAM,OAAO,gBAAgB,UAAU;KACrC,WAAW,KAAK;KAChB,MAAM,KAAK;KACZ,CAAC;AACF,QAAI,KAAK,UAAU,EACjB,QAAO;KACL,UAAU;KACV,MAAM;KACN,QAAQ,GAAG;KACX,SAAS,GAAG;KACZ,YAAY,GAAG;KACf,YAAY,GAAG;KACf,eAAe,SAAS;KACxB,SAAS,YAAY,IAAI,GAAG,GAAG,sBAAsB,KAAK,OAAO,QAAQ,EAAE,CAAC,IAAI,KAAK,QAAQ,QAAQ,EAAE,CAAC;KACzG;;;;AAMT,QAAO;EACL,UAAU;EACV,MAAM;EACN,QAAQ,GAAG;EACX,SAAS,GAAG;EACZ,YAAY,GAAG;EACf,YAAY,GAAG;EACf,eAAe,SAAS;EACzB"}

package/dist/lib/prompt.js.map

@@ -1 +0,0 @@
-{"version":3,"file":"prompt.js","names":[],"sources":["../../src/lib/prompt.ts"],"sourcesContent":["import { stdin, stdout } from \"node:process\";\nimport { createInterface } from \"node:readline/promises\";\n\nexport async function prompt(\n  question: string,\n  defaultValue?: string,\n): Promise<string> {\n  const rl = createInterface({ input: stdin, output: stdout });\n  const suffix = defaultValue ? ` [${defaultValue}]` : \"\";\n  const answer = (await rl.question(`${question}${suffix}: `)).trim();\n  rl.close();\n  return answer || (defaultValue ?? \"\");\n}\n\nexport async function promptChoice(\n  question: string,\n  choices: string[],\n  defaultValue?: string,\n): Promise<string> {\n  const rl = createInterface({ input: stdin, output: stdout });\n  stdout.write(`${question}\\n`);\n  for (const [i, c] of choices.entries()) {\n    const marker = c === defaultValue ? \"*\" : \" \";\n    stdout.write(`  ${marker} ${i + 1}) ${c}\\n`);\n  }\n  const raw = (await rl.question(`choose [${defaultValue ?? \"1\"}]: `)).trim();\n  rl.close();\n  if (!raw) return defaultValue ?? choices[0];\n  const n = Number(raw);\n  if (Number.isInteger(n) && n >= 1 && n <= choices.length)\n    return choices[n - 1];\n  if (choices.includes(raw)) return raw;\n  return defaultValue ?? choices[0];\n}\n"],"mappings":";;;AAGA,eAAsB,OACpB,UACA,cACiB;CACjB,MAAM,KAAK,gBAAgB;EAAE,OAAO;EAAO,QAAQ;EAAQ,CAAC;CAC5D,MAAM,SAAS,eAAe,KAAK,aAAa,KAAK;CACrD,MAAM,UAAU,MAAM,GAAG,SAAS,GAAG,WAAW,OAAO,IAAI,EAAE,MAAM;AACnE,IAAG,OAAO;AACV,QAAO,WAAW,gBAAgB;;AAGpC,eAAsB,aACpB,UACA,SACA,cACiB;CACjB,MAAM,KAAK,gBAAgB;EAAE,OAAO;EAAO,QAAQ;EAAQ,CAAC;AAC5D,QAAO,MAAM,GAAG,SAAS,IAAI;AAC7B,MAAK,MAAM,CAAC,GAAG,MAAM,QAAQ,SAAS,EAAE;EACtC,MAAM,SAAS,MAAM,eAAe,MAAM;AAC1C,SAAO,MAAM,KAAK,OAAO,GAAG,IAAI,EAAE,IAAI,EAAE,IAAI;;CAE9C,MAAM,OAAO,MAAM,GAAG,SAAS,WAAW,gBAAgB,IAAI,KAAK,EAAE,MAAM;AAC3E,IAAG,OAAO;AACV,KAAI,CAAC,IAAK,QAAO,gBAAgB,QAAQ;CACzC,MAAM,IAAI,OAAO,IAAI;AACrB,KAAI,OAAO,UAAU,EAAE,IAAI,KAAK,KAAK,KAAK,QAAQ,OAChD,QAAO,QAAQ,IAAI;AACrB,KAAI,QAAQ,SAAS,IAAI,CAAE,QAAO;AAClC,QAAO,gBAAgB,QAAQ"}

package/dist/lib/run-dir.js.map

@@ -1 +0,0 @@
-{"version":3,"file":"run-dir.js","names":[],"sources":["../../src/lib/run-dir.ts"],"sourcesContent":["import { randomBytes } from \"node:crypto\";\nimport {\n  appendFileSync,\n  closeSync,\n  existsSync,\n  fsyncSync,\n  mkdirSync,\n  openSync,\n  readFileSync,\n  renameSync,\n  unlinkSync,\n  writeFileSync,\n} from \"node:fs\";\nimport { dirname, join } from \"node:path\";\n\nexport function systemRunsDir(stateDir: string, systemId: string): string {\n  return join(stateDir, \"runs\", systemId);\n}\n\nexport function runDir(\n  stateDir: string,\n  systemId: string,\n  runId: string,\n): string {\n  return join(systemRunsDir(stateDir, systemId), runId);\n}\n\nexport function hypothesesPath(stateDir: string, systemId: string): string {\n  return join(stateDir, \"hypotheses\", `${systemId}.jsonl`);\n}\n\nexport function generateRunId(): string {\n  // r-<4 hex chars><timestamp suffix>: short, sortable-ish, collision-resistant enough for our scale.\n  const ts = Date.now().toString(36);\n  const rand = randomBytes(2).toString(\"hex\");\n  return `r-${rand}${ts.slice(-6)}`;\n}\n\nexport function ensureDir(p: string): void {\n  mkdirSync(p, { recursive: true });\n}\n\n/** Atomic write: write to a temp file in the same dir, fsync, rename over the target. */\nfunction writeFileAtomic(path: string, contents: string | Uint8Array): void {\n  ensureDir(dirname(path));\n  const tmp = `${path}.tmp.${process.pid}.${randomBytes(3).toString(\"hex\")}`;\n  const fd = openSync(tmp, \"w\", 0o644);\n  try {\n    writeFileSync(fd, contents);\n    fsyncSync(fd);\n  } finally {\n    closeSync(fd);\n  }\n  renameSync(tmp, path);\n}\n\nexport function writeJsonAtomic(path: string, value: unknown): void {\n  writeFileAtomic(path, JSON.stringify(value, null, 2) + \"\\n\");\n}\n\n/** Append a single NDJSON line. POSIX guarantees writes ≤ PIPE_BUF with O_APPEND are atomic. */\nexport function appendNdjsonLine(path: string, value: unknown): void {\n  ensureDir(dirname(path));\n  appendFileSync(path, JSON.stringify(value) + \"\\n\");\n}\n\nfunction readJson<T>(path: string): T {\n  return JSON.parse(readFileSync(path, \"utf-8\")) as T;\n}\n\nexport function readJsonIfExists<T>(path: string): T | null {\n  if (!existsSync(path)) return null;\n  return readJson<T>(path);\n}\n\n/** Read an NDJSON file as an array of parsed objects. Tolerates a trailing partial line. */\nexport function readNdjson<T>(path: string): T[] {\n  if (!existsSync(path)) return [];\n  const text = readFileSync(path, \"utf-8\");\n  const out: T[] = [];\n  for (const line of text.split(\"\\n\")) {\n    if (!line) continue;\n    try {\n      out.push(JSON.parse(line) as T);\n    } catch {\n      // Last line may be a partial write if a process is still appending. Skip silently.\n    }\n  }\n  return out;\n}\n\n// --- lock file ---\n\nexport interface LockInfo {\n  pid: number;\n  startedAt: string;\n}\n\nfunction lockPath(runDirPath: string): string {\n  return join(runDirPath, \".lock\");\n}\n\nexport function writeLock(runDirPath: string): void {\n  const info: LockInfo = {\n    pid: process.pid,\n    startedAt: new Date().toISOString(),\n  };\n  writeJsonAtomic(lockPath(runDirPath), info);\n}\n\nexport function readLock(runDirPath: string): LockInfo | null {\n  return readJsonIfExists<LockInfo>(lockPath(runDirPath));\n}\n\nexport function clearLock(runDirPath: string): void {\n  const p = lockPath(runDirPath);\n  if (existsSync(p)) unlinkSync(p);\n}\n\n/** True if the PID in the lock file is still alive on this machine. */\nexport function isPidAlive(pid: number): boolean {\n  try {\n    process.kill(pid, 0);\n    return true;\n  } catch (err) {\n    const code = (err as NodeJS.ErrnoException).code;\n    return code === \"EPERM\"; // exists but we can't signal — still alive\n  }\n}\n"],"mappings":";;;;AAeA,SAAgB,cAAc,UAAkB,UAA0B;AACxE,QAAO,KAAK,UAAU,QAAQ,SAAS;;AAGzC,SAAgB,OACd,UACA,UACA,OACQ;AACR,QAAO,KAAK,cAAc,UAAU,SAAS,EAAE,MAAM;;AAGvD,SAAgB,eAAe,UAAkB,UAA0B;AACzE,QAAO,KAAK,UAAU,cAAc,GAAG,SAAS,QAAQ;;AAG1D,SAAgB,gBAAwB;CAEtC,MAAM,KAAK,KAAK,KAAK,CAAC,SAAS,GAAG;AAElC,QAAO,KADM,YAAY,EAAE,CAAC,SAAS,MACrB,GAAG,GAAG,MAAM,GAAG;;AAGjC,SAAgB,UAAU,GAAiB;AACzC,WAAU,GAAG,EAAE,WAAW,MAAM,CAAC;;;AAInC,SAAS,gBAAgB,MAAc,UAAqC;AAC1E,WAAU,QAAQ,KAAK,CAAC;CACxB,MAAM,MAAM,GAAG,KAAK,OAAO,QAAQ,IAAI,GAAG,YAAY,EAAE,CAAC,SAAS,MAAM;CACxE,MAAM,KAAK,SAAS,KAAK,KAAK,IAAM;AACpC,KAAI;AACF,gBAAc,IAAI,SAAS;AAC3B,YAAU,GAAG;WACL;AACR,YAAU,GAAG;;AAEf,YAAW,KAAK,KAAK;;AAGvB,SAAgB,gBAAgB,MAAc,OAAsB;AAClE,iBAAgB,MAAM,KAAK,UAAU,OAAO,MAAM,EAAE,GAAG,KAAK;;;AAI9D,SAAgB,iBAAiB,MAAc,OAAsB;AACnE,WAAU,QAAQ,KAAK,CAAC;AACxB,gBAAe,MAAM,KAAK,UAAU,MAAM,GAAG,KAAK;;AAGpD,SAAS,SAAY,MAAiB;AACpC,QAAO,KAAK,MAAM,aAAa,MAAM,QAAQ,CAAC;;AAGhD,SAAgB,iBAAoB,MAAwB;AAC1D,KAAI,CAAC,WAAW,KAAK,CAAE,QAAO;AAC9B,QAAO,SAAY,KAAK;;;AAI1B,SAAgB,WAAc,MAAmB;AAC/C,KAAI,CAAC,WAAW,KAAK,CAAE,QAAO,EAAE;CAChC,MAAM,OAAO,aAAa,MAAM,QAAQ;CACxC,MAAM,MAAW,EAAE;AACnB,MAAK,MAAM,QAAQ,KAAK,MAAM,KAAK,EAAE;AACnC,MAAI,CAAC,KAAM;AACX,MAAI;AACF,OAAI,KAAK,KAAK,MAAM,KAAK,CAAM;UACzB;;AAIV,QAAO;;AAUT,SAAS,SAAS,YAA4B;AAC5C,QAAO,KAAK,YAAY,QAAQ;;AAGlC,SAAgB,UAAU,YAA0B;CAClD,MAAM,OAAiB;EACrB,KAAK,QAAQ;EACb,4BAAW,IAAI,MAAM,EAAC,aAAa;EACpC;AACD,iBAAgB,SAAS,WAAW,EAAE,KAAK;;AAG7C,SAAgB,SAAS,YAAqC;AAC5D,QAAO,iBAA2B,SAAS,WAAW,CAAC;;AAGzD,SAAgB,UAAU,YAA0B;CAClD,MAAM,IAAI,SAAS,WAAW;AAC9B,KAAI,WAAW,EAAE,CAAE,YAAW,EAAE;;;AAIlC,SAAgB,WAAW,KAAsB;AAC/C,KAAI;AACF,UAAQ,KAAK,KAAK,EAAE;AACpB,SAAO;UACA,KAAK;AAEZ,SADc,IAA8B,SAC5B"}

package/dist/lib/runner.js.map

@@ -1 +0,0 @@
-{"version":3,"file":"runner.js","names":[],"sources":["../../src/lib/runner.ts"],"sourcesContent":["import { type ChildProcess, execSync, spawn } from \"node:child_process\";\nimport { createWriteStream, existsSync, writeFileSync } from \"node:fs\";\nimport { createRequire } from \"node:module\";\nimport { arch, hostname, platform } from \"node:os\";\nimport { extname, join } from \"node:path\";\nimport type { Readable } from \"node:stream\";\nimport type { LinearIssueLink } from \"../../shared/linear-issue.js\";\nimport {\n  type CompleteEvent,\n  type CrashedEvent,\n  type Event,\n  type ItemEvent,\n  NdjsonReader,\n  type PromotionEvent,\n} from \"./events.js\";\nimport {\n  type ManifestFile,\n  type StateFile,\n  currentBaseline,\n  listRuns,\n} from \"./leaderboard.js\";\nimport { resolveStateDir } from \"./paths.js\";\nimport { decidePromotion } from \"./promotion.js\";\nimport {\n  appendNdjsonLine,\n  clearLock,\n  ensureDir,\n  generateRunId,\n  hypothesesPath,\n  isPidAlive,\n  systemRunsDir,\n  readJsonIfExists,\n  readLock,\n  runDir,\n  writeJsonAtomic,\n  writeLock,\n} from \"./run-dir.js\";\nimport { type SystemDef, loadSystem, resolveEvalPath } from \"./system.js\";\n\nexport interface RunOptions {\n  workspace: string;\n  systemId: string;\n  variant: string;\n  parent: string | null;\n  hypothesis: string;\n  diagnostic: boolean;\n  noAutoPromote: boolean;\n  maxItems: number | null;\n  kaizenVersion: string;\n  linearIssue: LinearIssueLink | null;\n}\n\nexport interface RunResult {\n  runId: string;\n  status: \"complete\" | \"crashed\" | \"aborted\";\n  score: number | null;\n  promoted: boolean | null;\n  exitCode: number;\n}\n\nconst FAILURE_K = 10;\nconst require = createRequire(import.meta.url);\n\nexport async function runExperiment(opts: RunOptions): Promise<RunResult> {\n  const stateDir = resolveStateDir(opts.workspace);\n\n  // 0. Reap stale runs across this system (lock present, PID dead, no terminal event).\n  reapStaleRuns(stateDir, opts.systemId);\n\n  // 1. Load system + resolve eval interpreter.\n  const system = loadSystem(opts.workspace, opts.systemId);\n  const evalAbs = resolveEvalPath(opts.workspace, system);\n  if (!existsSync(evalAbs)) {\n    throw new Error(\n      `eval script not found at ${evalAbs} (declared as run_eval: ${system.frontmatter.run_eval} in kaizen/systems/${opts.systemId}/system.md)`,\n    );\n  }\n  const interp = detectInterpreter(evalAbs);\n\n  // 2. Generate run_id, create dir, write manifest, take lock.\n  const runId = generateRunId();\n  const dir = runDir(stateDir, opts.systemId, runId);\n  ensureDir(dir);\n  const startedAt = new Date().toISOString();\n  const manifest: ManifestFile = {\n    run_id: runId,\n    system: opts.systemId,\n    variant: opts.variant,\n    parent_id: opts.parent,\n    hypothesis: opts.hypothesis,\n    git_sha: gitSha(opts.workspace),\n    git_branch: gitBranch(opts.workspace),\n    worktree_root: gitTopLevel(opts.workspace),\n    git_common_dir: gitCommonDir(opts.workspace),\n    eval_version: system.frontmatter.eval_version,\n    dataset_version: system.frontmatter.dataset_version,\n    started_at: startedAt,\n    host: `${platform()}-${arch()} (${hostname()})`,\n    kaizen_version: opts.kaizenVersion,\n    state_dir: stateDir,\n    diagnostic: opts.diagnostic,\n    linear_issue_id: opts.linearIssue?.id ?? null,\n    linear_issue_url: opts.linearIssue?.url ?? null,\n  };\n  writeJsonAtomic(join(dir, \"manifest.json\"), manifest);\n  writeLock(dir);\n\n  const eventsPath = join(dir, \"events.jsonl\");\n  const stdoutPath = join(dir, \"stdout.log\");\n  const stderrPath = join(dir, \"stderr.log\");\n\n  // 3. Build initial state and write it.\n  const state: StateFile & {\n    system: string;\n    variant: string;\n    eval_version: number;\n    dataset_version: string;\n    started_at: string;\n    updated_at: string;\n    n_done: number;\n  } = {\n    run_id: runId,\n    system: opts.systemId,\n    variant: opts.variant,\n    status: \"running\",\n    score: null,\n    n_total: null,\n    n_done: 0,\n    promoted: null,\n    started_at: startedAt,\n    updated_at: startedAt,\n    ended_at: null,\n    eval_version: system.frontmatter.eval_version,\n    dataset_version: system.frontmatter.dataset_version,\n  };\n  writeJsonAtomic(join(dir, \"state.json\"), state);\n\n  // 4. Spawn the eval with fd 3 piped for events.\n  // For shebang scripts (cmd === evalAbs) we omit evalAbs from args, since\n  // spawn already passes cmd as argv[0]. Including it would duplicate the\n  // path as a positional and break argparse-based eval scripts.\n  const evalArgs = [\n    ...interp.preArgs,\n    ...(interp.cmd === evalAbs ? [] : [evalAbs]),\n    \"--variant\",\n    opts.variant,\n    \"--dataset\",\n    system.frontmatter.dataset_version,\n    \"--out-fd\",\n    \"3\",\n  ];\n  const effectiveMax = opts.diagnostic ? 5 : opts.maxItems;\n  if (effectiveMax !== null && effectiveMax !== undefined) {\n    evalArgs.push(\"--max-items\", String(effectiveMax));\n  }\n\n  const stdoutLog = createWriteStream(stdoutPath);\n  const stderrLog = createWriteStream(stderrPath);\n  let stderrTail = \"\";\n\n  const child: ChildProcess = spawn(interp.cmd, evalArgs, {\n    cwd: opts.workspace,\n    stdio: [\"ignore\", \"pipe\", \"pipe\", \"pipe\"],\n    env: process.env,\n  });\n\n  // Tee stdout/stderr to log files. Capture last ~4KB of stderr for crash diagnostics.\n  child.stdout!.on(\"data\", (chunk: Buffer) => stdoutLog.write(chunk));\n  child.stderr!.on(\"data\", (chunk: Buffer) => {\n    stderrLog.write(chunk);\n    stderrTail = (stderrTail + chunk.toString(\"utf-8\")).slice(-4096);\n  });\n\n  // Event stream on fd 3. spawn(['ignore','pipe','pipe','pipe']) makes child.stdio[3] a Readable.\n  const eventChan = child.stdio[3] as Readable;\n  const itemEvents: ItemEvent[] = [];\n  let sawComplete: CompleteEvent | null = null;\n  let parseErrors = 0;\n  let contractErrors = 0;\n\n  const reader = new NdjsonReader(\n    (e) => onEvent(e),\n    (line, err) => {\n      parseErrors++;\n      stderrLog.write(\n        `[runner] could not parse event: ${line.slice(0, 200)}\\n`,\n      );\n      stderrLog.write(\n        `[runner] ${err instanceof Error ? err.message : String(err)}\\n`,\n      );\n    },\n  );\n  eventChan.setEncoding?.(\"utf-8\");\n  eventChan.on(\"data\", (chunk) => reader.push(chunk as string | Buffer));\n  // The event channel may close after the child's `exit` fires (Node docs\n  // explicitly note stdio streams can outlive `exit`). We must drain it\n  // before deciding terminal state — otherwise a successful run that\n  // emitted a `complete` event still in the pipe buffer gets misclassified\n  // as \"crashed\" because `sawComplete` is observed null.\n  const eventChanEnded = new Promise<void>((resolve) => {\n    eventChan.on(\"end\", () => {\n      reader.end();\n      resolve();\n    });\n    eventChan.on(\"close\", () => {\n      reader.end();\n      resolve();\n    });\n  });\n\n  function onEvent(e: Event): void {\n    appendNdjsonLine(eventsPath, e);\n    if (e.type === \"start\") {\n      const s = e;\n      if (s.eval_version !== system.frontmatter.eval_version) {\n        contractErrors++;\n        stderrLog.write(\n          `[runner] start.eval_version ${s.eval_version} does not match system eval_version ${system.frontmatter.eval_version}\\n`,\n        );\n      }\n      if (s.dataset_version !== system.frontmatter.dataset_version) {\n        contractErrors++;\n        stderrLog.write(\n          `[runner] start.dataset_version ${s.dataset_version} does not match system dataset_version ${system.frontmatter.dataset_version}\\n`,\n        );\n      }\n      state.n_total = s.n;\n      state.updated_at = new Date().toISOString();\n      writeJsonAtomic(join(dir, \"state.json\"), state);\n    } else if (e.type === \"item\") {\n      const it = e;\n      itemEvents.push(it);\n      state.n_done = (state.n_done ?? 0) + 1;\n      state.updated_at = new Date().toISOString();\n      writeJsonAtomic(join(dir, \"state.json\"), state);\n    } else if (e.type === \"progress\") {\n      // n_done is authoritative from item events; progress is informational.\n      state.updated_at = new Date().toISOString();\n      writeJsonAtomic(join(dir, \"state.json\"), state);\n    } else if (e.type === \"complete\") {\n      sawComplete = e;\n      if (state.n_total !== null && sawComplete.n !== state.n_total) {\n        contractErrors++;\n        stderrLog.write(\n          `[runner] complete.n ${sawComplete.n} does not match start.n ${state.n_total}\\n`,\n        );\n      }\n      state.score = sawComplete.score;\n      state.updated_at = new Date().toISOString();\n      writeJsonAtomic(join(dir, \"state.json\"), state);\n    }\n    // error/promotion/etc. — no state mutation here\n  }\n\n  // 5. Wait for child to exit AND for the event channel to be fully drained.\n  // Both must complete before we decide terminal state — the child's exit\n  // doesn't guarantee its stdio buffers have been consumed by us yet.\n  const { code, signal } = await waitForExit(child);\n  await eventChanEnded;\n\n  // Flush logs.\n  stdoutLog.end();\n  stderrLog.end();\n  await Promise.all([\n    new Promise<void>((r) => stdoutLog.on(\"close\", () => r())),\n    new Promise<void>((r) => stderrLog.on(\"close\", () => r())),\n  ]);\n\n  // 6. Decide terminal state.\n  const endedAt = new Date().toISOString();\n  if (\n    code === 0 &&\n    sawComplete !== null &&\n    parseErrors === 0 &&\n    contractErrors === 0\n  ) {\n    state.status = \"complete\";\n    state.score = (sawComplete as CompleteEvent).score;\n    // The complete event is already in events.jsonl from onEvent.\n  } else {\n    const crashed: CrashedEvent = {\n      type: \"crashed\",\n      exit_code: code,\n      signal: signal ?? null,\n      stderr_tail: stderrTail.slice(-2000),\n      reason:\n        code !== 0 && code !== null\n          ? `eval exited with code ${code}`\n          : signal\n            ? `eval terminated by signal ${signal}`\n            : parseErrors > 0\n              ? `eval emitted ${parseErrors} malformed event(s)`\n              : contractErrors > 0\n                ? `eval emitted ${contractErrors} contract-violating event(s)`\n                : sawComplete !== null\n                  ? \"eval emitted complete but exited non-zero\"\n                  : \"eval exited without emitting complete event\",\n      ts: Date.now() / 1000,\n    };\n    appendNdjsonLine(eventsPath, crashed);\n    state.status = \"crashed\";\n  }\n  state.ended_at = endedAt;\n  state.updated_at = endedAt;\n\n  // 7. Write failures.jsonl (worst-K).\n  if (itemEvents.length > 0) {\n    const sorted = [...itemEvents].sort((a, b) => a.score - b.score);\n    const worst = sorted.slice(0, Math.min(FAILURE_K, sorted.length));\n    const text =\n      worst.map((it) => JSON.stringify(it)).join(\"\\n\") +\n      (worst.length > 0 ? \"\\n\" : \"\");\n    writeFileSync(join(dir, \"failures.jsonl\"), text);\n  }\n\n  // 8. Promotion (only if completed cleanly).\n  let promotionEvent: PromotionEvent | null = null;\n  if (state.status === \"complete\" && !opts.noAutoPromote) {\n    const allRuns = listRuns(stateDir, opts.systemId).filter(\n      (r) => r.run_id !== runId,\n    );\n    const baselineRun = currentBaseline(\n      allRuns,\n      system.frontmatter.eval_version,\n      system.frontmatter.dataset_version,\n    );\n    const baseline = baselineRun\n      ? {\n          run_id: baselineRun.run_id,\n          variant: baselineRun.variant,\n          score: baselineRun.score!,\n          events_path: baselineRun.events_path,\n        }\n      : null;\n    const decision = decidePromotion(\n      mapFromItems(itemEvents),\n      itemEvents,\n      baseline,\n      { subgroupKeys: subgroupKeys(system) },\n    );\n    promotionEvent = { type: \"promotion\", ts: Date.now() / 1000, ...decision };\n    appendNdjsonLine(eventsPath, promotionEvent);\n    state.promoted = decision.promoted;\n  } else if (opts.noAutoPromote) {\n    state.promoted = false;\n  }\n  writeJsonAtomic(join(dir, \"state.json\"), state);\n\n  // 9. Append hypothesis log line (every run, success or failure).\n  const hypoLine = {\n    run_id: runId,\n    system: opts.systemId,\n    variant: opts.variant,\n    parent_id: opts.parent,\n    hypothesis: opts.hypothesis,\n    status: state.status,\n    score: state.score,\n    promoted: state.promoted,\n    started_at: startedAt,\n    ended_at: endedAt,\n    eval_version: system.frontmatter.eval_version,\n    dataset_version: system.frontmatter.dataset_version,\n  };\n  appendNdjsonLine(hypothesesPath(stateDir, opts.systemId), hypoLine);\n\n  // 10. Clear lock.\n  clearLock(dir);\n\n  void parseErrors; // not surfaced yet; logged to stderr.log\n\n  return {\n    runId,\n    status: state.status as \"complete\" | \"crashed\" | \"aborted\",\n    score: state.score,\n    promoted: state.promoted ?? null,\n    exitCode: state.status === \"complete\" ? 0 : 1,\n  };\n}\n\n// --- helpers ---\n\ninterface InterpreterChoice {\n  cmd: string;\n  preArgs: string[];\n}\n\nfunction detectInterpreter(evalPath: string): InterpreterChoice {\n  const ext = extname(evalPath).toLowerCase();\n  if (ext === \".py\") return { cmd: \"python3\", preArgs: [] };\n  if (ext === \".js\" || ext === \".mjs\" || ext === \".cjs\")\n    return { cmd: \"node\", preArgs: [] };\n  if (ext === \".ts\" || ext === \".tsx\" || ext === \".mts\" || ext === \".cts\") {\n    return {\n      cmd: process.execPath,\n      preArgs: [\"--import\", require.resolve(\"tsx\")],\n    };\n  }\n  // Default: assume executable with shebang.\n  return { cmd: evalPath, preArgs: [] };\n}\n\nfunction waitForExit(\n  child: ChildProcess,\n): Promise<{ code: number | null; signal: NodeJS.Signals | null }> {\n  return new Promise((resolve) => {\n    child.on(\"exit\", (code: number | null, signal: NodeJS.Signals | null) =>\n      resolve({ code, signal }),\n    );\n  });\n}\n\nfunction gitSha(cwd: string): string | null {\n  try {\n    return (\n      execSync(\"git rev-parse HEAD\", {\n        cwd,\n        stdio: [\"ignore\", \"pipe\", \"ignore\"],\n      })\n        .toString()\n        .trim() || null\n    );\n  } catch {\n    return null;\n  }\n}\n\nfunction gitBranch(cwd: string): string | null {\n  try {\n    return (\n      execSync(\"git rev-parse --abbrev-ref HEAD\", {\n        cwd,\n        stdio: [\"ignore\", \"pipe\", \"ignore\"],\n      })\n        .toString()\n        .trim() || null\n    );\n  } catch {\n    return null;\n  }\n}\n\nfunction gitTopLevel(cwd: string): string | null {\n  try {\n    return (\n      execSync(\"git rev-parse --show-toplevel\", {\n        cwd,\n        stdio: [\"ignore\", \"pipe\", \"ignore\"],\n      })\n        .toString()\n        .trim() || null\n    );\n  } catch {\n    return null;\n  }\n}\n\nfunction gitCommonDir(cwd: string): string | null {\n  try {\n    return (\n      execSync(\"git rev-parse --git-common-dir\", {\n        cwd,\n        stdio: [\"ignore\", \"pipe\", \"ignore\"],\n      })\n        .toString()\n        .trim() || null\n    );\n  } catch {\n    return null;\n  }\n}\n\nfunction mapFromItems(items: ItemEvent[]): Map<string, number> {\n  const m = new Map<string, number>();\n  for (const it of items) m.set(it.id, it.score);\n  return m;\n}\n\nfunction subgroupKeys(system: SystemDef): string[] {\n  const sg = (\n    system.frontmatter as {\n      subgroups?: Array<{ field?: string; name?: string }>;\n    }\n  ).subgroups;\n  if (!Array.isArray(sg)) return [];\n  return sg.map((s) => s.field ?? s.name ?? \"\").filter(Boolean);\n}\n\n/**\n * Sweep stale runs whose lock points to a dead PID and which never recorded a terminal event.\n * Mark them `crashed` with reason \"supervisor died\" so the dashboard / log don't show ghosts.\n */\nexport function reapStaleRuns(stateDir: string, systemId: string): void {\n  const dir = systemRunsDir(stateDir, systemId);\n  if (!existsSync(dir)) return;\n  for (const run of listRuns(stateDir, systemId)) {\n    if (run.status !== \"running\") continue;\n    const lock = readLock(join(dir, run.run_id));\n    if (!lock) continue;\n    if (isPidAlive(lock.pid)) continue;\n    const reapedEvent: CrashedEvent = {\n      type: \"crashed\",\n      exit_code: null,\n      signal: null,\n      reason: `runner pid ${lock.pid} died without writing a terminal event; reaped`,\n      ts: Date.now() / 1000,\n    };\n    appendNdjsonLine(run.events_path, reapedEvent);\n\n    const state = readJsonIfExists<StateFile>(run.state_path);\n    if (state) {\n      state.status = \"crashed\";\n      state.ended_at = new Date().toISOString();\n      writeJsonAtomic(run.state_path, state);\n    }\n    clearLock(join(dir, run.run_id));\n  }\n}\n"],"mappings":";;;;;;;;;;;;AA4DA,MAAM,YAAY;AAClB,MAAM,UAAU,cAAc,OAAO,KAAK,IAAI;AAE9C,eAAsB,cAAc,MAAsC;CACxE,MAAM,WAAW,gBAAgB,KAAK,UAAU;AAGhD,eAAc,UAAU,KAAK,SAAS;CAGtC,MAAM,SAAS,WAAW,KAAK,WAAW,KAAK,SAAS;CACxD,MAAM,UAAU,gBAAgB,KAAK,WAAW,OAAO;AACvD,KAAI,CAAC,WAAW,QAAQ,CACtB,OAAM,IAAI,MACR,4BAA4B,QAAQ,0BAA0B,OAAO,YAAY,SAAS,qBAAqB,KAAK,SAAS,aAC9H;CAEH,MAAM,SAAS,kBAAkB,QAAQ;CAGzC,MAAM,QAAQ,eAAe;CAC7B,MAAM,MAAM,OAAO,UAAU,KAAK,UAAU,MAAM;AAClD,WAAU,IAAI;CACd,MAAM,6BAAY,IAAI,MAAM,EAAC,aAAa;CAC1C,MAAM,WAAyB;EAC7B,QAAQ;EACR,QAAQ,KAAK;EACb,SAAS,KAAK;EACd,WAAW,KAAK;EAChB,YAAY,KAAK;EACjB,SAAS,OAAO,KAAK,UAAU;EAC/B,YAAY,UAAU,KAAK,UAAU;EACrC,eAAe,YAAY,KAAK,UAAU;EAC1C,gBAAgB,aAAa,KAAK,UAAU;EAC5C,cAAc,OAAO,YAAY;EACjC,iBAAiB,OAAO,YAAY;EACpC,YAAY;EACZ,MAAM,GAAG,UAAU,CAAC,GAAG,MAAM,CAAC,IAAI,UAAU,CAAC;EAC7C,gBAAgB,KAAK;EACrB,WAAW;EACX,YAAY,KAAK;EACjB,iBAAiB,KAAK,aAAa,MAAM;EACzC,kBAAkB,KAAK,aAAa,OAAO;EAC5C;AACD,iBAAgB,KAAK,KAAK,gBAAgB,EAAE,SAAS;AACrD,WAAU,IAAI;CAEd,MAAM,aAAa,KAAK,KAAK,eAAe;CAC5C,MAAM,aAAa,KAAK,KAAK,aAAa;CAC1C,MAAM,aAAa,KAAK,KAAK,aAAa;CAG1C,MAAM,QAQF;EACF,QAAQ;EACR,QAAQ,KAAK;EACb,SAAS,KAAK;EACd,QAAQ;EACR,OAAO;EACP,SAAS;EACT,QAAQ;EACR,UAAU;EACV,YAAY;EACZ,YAAY;EACZ,UAAU;EACV,cAAc,OAAO,YAAY;EACjC,iBAAiB,OAAO,YAAY;EACrC;AACD,iBAAgB,KAAK,KAAK,aAAa,EAAE,MAAM;CAM/C,MAAM,WAAW;EACf,GAAG,OAAO;EACV,GAAI,OAAO,QAAQ,UAAU,EAAE,GAAG,CAAC,QAAQ;EAC3C;EACA,KAAK;EACL;EACA,OAAO,YAAY;EACnB;EACA;EACD;CACD,MAAM,eAAe,KAAK,aAAa,IAAI,KAAK;AAChD,KAAI,iBAAiB,QAAQ,iBAAiB,KAAA,EAC5C,UAAS,KAAK,eAAe,OAAO,aAAa,CAAC;CAGpD,MAAM,YAAY,kBAAkB,WAAW;CAC/C,MAAM,YAAY,kBAAkB,WAAW;CAC/C,IAAI,aAAa;CAEjB,MAAM,QAAsB,MAAM,OAAO,KAAK,UAAU;EACtD,KAAK,KAAK;EACV,OAAO;GAAC;GAAU;GAAQ;GAAQ;GAAO;EACzC,KAAK,QAAQ;EACd,CAAC;AAGF,OAAM,OAAQ,GAAG,SAAS,UAAkB,UAAU,MAAM,MAAM,CAAC;AACnE,OAAM,OAAQ,GAAG,SAAS,UAAkB;AAC1C,YAAU,MAAM,MAAM;AACtB,gBAAc,aAAa,MAAM,SAAS,QAAQ,EAAE,MAAM,MAAM;GAChE;CAGF,MAAM,YAAY,MAAM,MAAM;CAC9B,MAAM,aAA0B,EAAE;CAClC,IAAI,cAAoC;CACxC,IAAI,cAAc;CAClB,IAAI,iBAAiB;CAErB,MAAM,SAAS,IAAI,cAChB,MAAM,QAAQ,EAAE,GAChB,MAAM,QAAQ;AACb;AACA,YAAU,MACR,mCAAmC,KAAK,MAAM,GAAG,IAAI,CAAC,IACvD;AACD,YAAU,MACR,YAAY,eAAe,QAAQ,IAAI,UAAU,OAAO,IAAI,CAAC,IAC9D;GAEJ;AACD,WAAU,cAAc,QAAQ;AAChC,WAAU,GAAG,SAAS,UAAU,OAAO,KAAK,MAAyB,CAAC;CAMtE,MAAM,iBAAiB,IAAI,SAAe,YAAY;AACpD,YAAU,GAAG,aAAa;AACxB,UAAO,KAAK;AACZ,YAAS;IACT;AACF,YAAU,GAAG,eAAe;AAC1B,UAAO,KAAK;AACZ,YAAS;IACT;GACF;CAEF,SAAS,QAAQ,GAAgB;AAC/B,mBAAiB,YAAY,EAAE;AAC/B,MAAI,EAAE,SAAS,SAAS;GACtB,MAAM,IAAI;AACV,OAAI,EAAE,iBAAiB,OAAO,YAAY,cAAc;AACtD;AACA,cAAU,MACR,+BAA+B,EAAE,aAAa,sCAAsC,OAAO,YAAY,aAAa,IACrH;;AAEH,OAAI,EAAE,oBAAoB,OAAO,YAAY,iBAAiB;AAC5D;AACA,cAAU,MACR,kCAAkC,EAAE,gBAAgB,yCAAyC,OAAO,YAAY,gBAAgB,IACjI;;AAEH,SAAM,UAAU,EAAE;AAClB,SAAM,8BAAa,IAAI,MAAM,EAAC,aAAa;AAC3C,mBAAgB,KAAK,KAAK,aAAa,EAAE,MAAM;aACtC,EAAE,SAAS,QAAQ;GAC5B,MAAM,KAAK;AACX,cAAW,KAAK,GAAG;AACnB,SAAM,UAAU,MAAM,UAAU,KAAK;AACrC,SAAM,8BAAa,IAAI,MAAM,EAAC,aAAa;AAC3C,mBAAgB,KAAK,KAAK,aAAa,EAAE,MAAM;aACtC,EAAE,SAAS,YAAY;AAEhC,SAAM,8BAAa,IAAI,MAAM,EAAC,aAAa;AAC3C,mBAAgB,KAAK,KAAK,aAAa,EAAE,MAAM;aACtC,EAAE,SAAS,YAAY;AAChC,iBAAc;AACd,OAAI,MAAM,YAAY,QAAQ,YAAY,MAAM,MAAM,SAAS;AAC7D;AACA,cAAU,MACR,uBAAuB,YAAY,EAAE,0BAA0B,MAAM,QAAQ,IAC9E;;AAEH,SAAM,QAAQ,YAAY;AAC1B,SAAM,8BAAa,IAAI,MAAM,EAAC,aAAa;AAC3C,mBAAgB,KAAK,KAAK,aAAa,EAAE,MAAM;;;CAQnD,MAAM,EAAE,MAAM,WAAW,MAAM,YAAY,MAAM;AACjD,OAAM;AAGN,WAAU,KAAK;AACf,WAAU,KAAK;AACf,OAAM,QAAQ,IAAI,CAChB,IAAI,SAAe,MAAM,UAAU,GAAG,eAAe,GAAG,CAAC,CAAC,EAC1D,IAAI,SAAe,MAAM,UAAU,GAAG,eAAe,GAAG,CAAC,CAAC,CAC3D,CAAC;CAGF,MAAM,2BAAU,IAAI,MAAM,EAAC,aAAa;AACxC,KACE,SAAS,KACT,gBAAgB,QAChB,gBAAgB,KAChB,mBAAmB,GACnB;AACA,QAAM,SAAS;AACf,QAAM,QAAS,YAA8B;QAExC;AAoBL,mBAAiB,YAAY;GAlB3B,MAAM;GACN,WAAW;GACX,QAAQ,UAAU;GAClB,aAAa,WAAW,MAAM,KAAM;GACpC,QACE,SAAS,KAAK,SAAS,OACnB,yBAAyB,SACzB,SACE,6BAA6B,WAC7B,cAAc,IACZ,gBAAgB,YAAY,uBAC5B,iBAAiB,IACf,gBAAgB,eAAe,gCAC/B,gBAAgB,OACd,8CACA;GACd,IAAI,KAAK,KAAK,GAAG;GAEiB,CAAC;AACrC,QAAM,SAAS;;AAEjB,OAAM,WAAW;AACjB,OAAM,aAAa;AAGnB,KAAI,WAAW,SAAS,GAAG;EACzB,MAAM,SAAS,CAAC,GAAG,WAAW,CAAC,MAAM,GAAG,MAAM,EAAE,QAAQ,EAAE,MAAM;EAChE,MAAM,QAAQ,OAAO,MAAM,GAAG,KAAK,IAAI,WAAW,OAAO,OAAO,CAAC;EACjE,MAAM,OACJ,MAAM,KAAK,OAAO,KAAK,UAAU,GAAG,CAAC,CAAC,KAAK,KAAK,IAC/C,MAAM,SAAS,IAAI,OAAO;AAC7B,gBAAc,KAAK,KAAK,iBAAiB,EAAE,KAAK;;CAIlD,IAAI,iBAAwC;AAC5C,KAAI,MAAM,WAAW,cAAc,CAAC,KAAK,eAAe;EAItD,MAAM,cAAc,gBAHJ,SAAS,UAAU,KAAK,SAAS,CAAC,QAC/C,MAAM,EAAE,WAAW,MAGb,EACP,OAAO,YAAY,cACnB,OAAO,YAAY,gBACpB;EACD,MAAM,WAAW,cACb;GACE,QAAQ,YAAY;GACpB,SAAS,YAAY;GACrB,OAAO,YAAY;GACnB,aAAa,YAAY;GAC1B,GACD;EACJ,MAAM,WAAW,gBACf,aAAa,WAAW,EACxB,YACA,UACA,EAAE,cAAc,aAAa,OAAO,EAAE,CACvC;AACD,mBAAiB;GAAE,MAAM;GAAa,IAAI,KAAK,KAAK,GAAG;GAAM,GAAG;GAAU;AAC1E,mBAAiB,YAAY,eAAe;AAC5C,QAAM,WAAW,SAAS;YACjB,KAAK,cACd,OAAM,WAAW;AAEnB,iBAAgB,KAAK,KAAK,aAAa,EAAE,MAAM;CAG/C,MAAM,WAAW;EACf,QAAQ;EACR,QAAQ,KAAK;EACb,SAAS,KAAK;EACd,WAAW,KAAK;EAChB,YAAY,KAAK;EACjB,QAAQ,MAAM;EACd,OAAO,MAAM;EACb,UAAU,MAAM;EAChB,YAAY;EACZ,UAAU;EACV,cAAc,OAAO,YAAY;EACjC,iBAAiB,OAAO,YAAY;EACrC;AACD,kBAAiB,eAAe,UAAU,KAAK,SAAS,EAAE,SAAS;AAGnE,WAAU,IAAI;AAId,QAAO;EACL;EACA,QAAQ,MAAM;EACd,OAAO,MAAM;EACb,UAAU,MAAM,YAAY;EAC5B,UAAU,MAAM,WAAW,aAAa,IAAI;EAC7C;;AAUH,SAAS,kBAAkB,UAAqC;CAC9D,MAAM,MAAM,QAAQ,SAAS,CAAC,aAAa;AAC3C,KAAI,QAAQ,MAAO,QAAO;EAAE,KAAK;EAAW,SAAS,EAAE;EAAE;AACzD,KAAI,QAAQ,SAAS,QAAQ,UAAU,QAAQ,OAC7C,QAAO;EAAE,KAAK;EAAQ,SAAS,EAAE;EAAE;AACrC,KAAI,QAAQ,SAAS,QAAQ,UAAU,QAAQ,UAAU,QAAQ,OAC/D,QAAO;EACL,KAAK,QAAQ;EACb,SAAS,CAAC,YAAY,QAAQ,QAAQ,MAAM,CAAC;EAC9C;AAGH,QAAO;EAAE,KAAK;EAAU,SAAS,EAAE;EAAE;;AAGvC,SAAS,YACP,OACiE;AACjE,QAAO,IAAI,SAAS,YAAY;AAC9B,QAAM,GAAG,SAAS,MAAqB,WACrC,QAAQ;GAAE;GAAM;GAAQ,CAAC,CAC1B;GACD;;AAGJ,SAAS,OAAO,KAA4B;AAC1C,KAAI;AACF,SACE,SAAS,sBAAsB;GAC7B;GACA,OAAO;IAAC;IAAU;IAAQ;IAAS;GACpC,CAAC,CACC,UAAU,CACV,MAAM,IAAI;SAET;AACN,SAAO;;;AAIX,SAAS,UAAU,KAA4B;AAC7C,KAAI;AACF,SACE,SAAS,mCAAmC;GAC1C;GACA,OAAO;IAAC;IAAU;IAAQ;IAAS;GACpC,CAAC,CACC,UAAU,CACV,MAAM,IAAI;SAET;AACN,SAAO;;;AAIX,SAAS,YAAY,KAA4B;AAC/C,KAAI;AACF,SACE,SAAS,iCAAiC;GACxC;GACA,OAAO;IAAC;IAAU;IAAQ;IAAS;GACpC,CAAC,CACC,UAAU,CACV,MAAM,IAAI;SAET;AACN,SAAO;;;AAIX,SAAS,aAAa,KAA4B;AAChD,KAAI;AACF,SACE,SAAS,kCAAkC;GACzC;GACA,OAAO;IAAC;IAAU;IAAQ;IAAS;GACpC,CAAC,CACC,UAAU,CACV,MAAM,IAAI;SAET;AACN,SAAO;;;AAIX,SAAS,aAAa,OAAyC;CAC7D,MAAM,oBAAI,IAAI,KAAqB;AACnC,MAAK,MAAM,MAAM,MAAO,GAAE,IAAI,GAAG,IAAI,GAAG,MAAM;AAC9C,QAAO;;AAGT,SAAS,aAAa,QAA6B;CACjD,MAAM,KACJ,OAAO,YAGP;AACF,KAAI,CAAC,MAAM,QAAQ,GAAG,CAAE,QAAO,EAAE;AACjC,QAAO,GAAG,KAAK,MAAM,EAAE,SAAS,EAAE,QAAQ,GAAG,CAAC,OAAO,QAAQ;;;;;;AAO/D,SAAgB,cAAc,UAAkB,UAAwB;CACtE,MAAM,MAAM,cAAc,UAAU,SAAS;AAC7C,KAAI,CAAC,WAAW,IAAI,CAAE;AACtB,MAAK,MAAM,OAAO,SAAS,UAAU,SAAS,EAAE;AAC9C,MAAI,IAAI,WAAW,UAAW;EAC9B,MAAM,OAAO,SAAS,KAAK,KAAK,IAAI,OAAO,CAAC;AAC5C,MAAI,CAAC,KAAM;AACX,MAAI,WAAW,KAAK,IAAI,CAAE;EAC1B,MAAM,cAA4B;GAChC,MAAM;GACN,WAAW;GACX,QAAQ;GACR,QAAQ,cAAc,KAAK,IAAI;GAC/B,IAAI,KAAK,KAAK,GAAG;GAClB;AACD,mBAAiB,IAAI,aAAa,YAAY;EAE9C,MAAM,QAAQ,iBAA4B,IAAI,WAAW;AACzD,MAAI,OAAO;AACT,SAAM,SAAS;AACf,SAAM,4BAAW,IAAI,MAAM,EAAC,aAAa;AACzC,mBAAgB,IAAI,YAAY,MAAM;;AAExC,YAAU,KAAK,KAAK,IAAI,OAAO,CAAC"}

package/dist/lib/system.js

@@ -1,57 +0,0 @@
-import { kaizenSystemPath } from "../shared/workspace-paths.js";
-import "./paths.js";
-import { existsSync, readFileSync } from "node:fs";
-import { resolve } from "node:path";
-import { parse } from "yaml";
-//#region src/lib/system.ts
-function systemPath(workspaceRoot, id) {
-	return kaizenSystemPath(workspaceRoot, id);
-}
-function loadSystem(workspaceRoot, id) {
-	const path = systemPath(workspaceRoot, id);
-	if (!existsSync(path)) throw new Error(`system "${id}" not found at ${path}. run \`kaizen create system ${id}\` to scaffold one.`);
-	const { frontmatter, body } = splitFrontmatter(readFileSync(path, "utf-8"));
-	if (!frontmatter) throw new Error(`system ${id}: no YAML frontmatter found in ${path}`);
-	const fm = parse(frontmatter);
-	if (!fm || typeof fm !== "object") throw new Error(`system ${id}: frontmatter is not a YAML mapping`);
-	for (const k of [
-		"name",
-		"run_eval",
-		"eval_version",
-		"dataset_version",
-		"primary_metric"
-	]) if (fm[k] === void 0 || fm[k] === null || fm[k] === "") throw new Error(`system ${id}: frontmatter missing required field "${k}"`);
-	return {
-		id,
-		path,
-		frontmatter: fm,
-		body
-	};
-}
-function resolveEvalPath(workspaceRoot, system) {
-	return resolve(workspaceRoot, system.frontmatter.run_eval);
-}
-function splitFrontmatter(text) {
-	if (!text.startsWith("---")) return {
-		frontmatter: null,
-		body: text
-	};
-	const lines = text.split("\n");
-	let close = -1;
-	for (let i = 1; i < lines.length; i++) if (lines[i].trim() === "---") {
-		close = i;
-		break;
-	}
-	if (close === -1) return {
-		frontmatter: null,
-		body: text
-	};
-	return {
-		frontmatter: lines.slice(1, close).join("\n"),
-		body: lines.slice(close + 1).join("\n")
-	};
-}
-//#endregion
-export { loadSystem, resolveEvalPath };
-
-//# sourceMappingURL=system.js.map

package/dist/lib/system.js.map

@@ -1 +0,0 @@
-{"version":3,"file":"system.js","names":["parseYaml"],"sources":["../../src/lib/system.ts"],"sourcesContent":["import { existsSync, readFileSync } from \"node:fs\";\nimport { resolve } from \"node:path\";\nimport { parse as parseYaml } from \"yaml\";\nimport { kaizenSystemPath } from \"./paths.js\";\n\nexport type EvalStyle = \"ground-truth\" | \"llm-as-judge\" | \"hybrid\";\n\nexport interface SubgroupDef {\n  name: string;\n  field: string;\n}\n\nexport interface SystemFrontmatter {\n  name: string;\n  description?: string;\n  run_eval: string;\n  eval_version: number;\n  dataset_version: string;\n  eval_style?: EvalStyle;\n  primary_metric: string;\n  target?: number;\n  kaizen_version?: string;\n  subgroups?: SubgroupDef[];\n  [key: string]: unknown;\n}\n\nexport interface SystemDef {\n  id: string;\n  path: string;\n  frontmatter: SystemFrontmatter;\n  body: string;\n}\n\nfunction systemPath(workspaceRoot: string, id: string): string {\n  return kaizenSystemPath(workspaceRoot, id);\n}\n\nexport function loadSystem(workspaceRoot: string, id: string): SystemDef {\n  const path = systemPath(workspaceRoot, id);\n  if (!existsSync(path)) {\n    throw new Error(\n      `system \"${id}\" not found at ${path}. run \\`kaizen create system ${id}\\` to scaffold one.`,\n    );\n  }\n  const text = readFileSync(path, \"utf-8\");\n  const { frontmatter, body } = splitFrontmatter(text);\n  if (!frontmatter) {\n    throw new Error(`system ${id}: no YAML frontmatter found in ${path}`);\n  }\n  const fm = parseYaml(frontmatter) as Partial<SystemFrontmatter> | null;\n  if (!fm || typeof fm !== \"object\") {\n    throw new Error(`system ${id}: frontmatter is not a YAML mapping`);\n  }\n  const required = [\n    \"name\",\n    \"run_eval\",\n    \"eval_version\",\n    \"dataset_version\",\n    \"primary_metric\",\n  ] as const;\n  for (const k of required) {\n    if (fm[k] === undefined || fm[k] === null || fm[k] === \"\") {\n      throw new Error(\n        `system ${id}: frontmatter missing required field \"${k}\"`,\n      );\n    }\n  }\n  return {\n    id,\n    path,\n    frontmatter: fm as SystemFrontmatter,\n    body,\n  };\n}\n\nexport function resolveEvalPath(\n  workspaceRoot: string,\n  system: SystemDef,\n): string {\n  return resolve(workspaceRoot, system.frontmatter.run_eval);\n}\n\nfunction splitFrontmatter(text: string): {\n  frontmatter: string | null;\n  body: string;\n} {\n  // Frontmatter starts on line 1 with ---, ends with --- on its own line.\n  if (!text.startsWith(\"---\")) return { frontmatter: null, body: text };\n  const lines = text.split(\"\\n\");\n  // First line is \"---\". Find the next \"---\".\n  let close = -1;\n  for (let i = 1; i < lines.length; i++) {\n    if (lines[i].trim() === \"---\") {\n      close = i;\n      break;\n    }\n  }\n  if (close === -1) return { frontmatter: null, body: text };\n  return {\n    frontmatter: lines.slice(1, close).join(\"\\n\"),\n    body: lines.slice(close + 1).join(\"\\n\"),\n  };\n}\n"],"mappings":";;;;;;AAiCA,SAAS,WAAW,eAAuB,IAAoB;AAC7D,QAAO,iBAAiB,eAAe,GAAG;;AAG5C,SAAgB,WAAW,eAAuB,IAAuB;CACvE,MAAM,OAAO,WAAW,eAAe,GAAG;AAC1C,KAAI,CAAC,WAAW,KAAK,CACnB,OAAM,IAAI,MACR,WAAW,GAAG,iBAAiB,KAAK,+BAA+B,GAAG,qBACvE;CAGH,MAAM,EAAE,aAAa,SAAS,iBADjB,aAAa,MAAM,QACmB,CAAC;AACpD,KAAI,CAAC,YACH,OAAM,IAAI,MAAM,UAAU,GAAG,iCAAiC,OAAO;CAEvE,MAAM,KAAKA,MAAU,YAAY;AACjC,KAAI,CAAC,MAAM,OAAO,OAAO,SACvB,OAAM,IAAI,MAAM,UAAU,GAAG,qCAAqC;AASpE,MAAK,MAAM,KAAK;EANd;EACA;EACA;EACA;EACA;EAEsB,CACtB,KAAI,GAAG,OAAO,KAAA,KAAa,GAAG,OAAO,QAAQ,GAAG,OAAO,GACrD,OAAM,IAAI,MACR,UAAU,GAAG,wCAAwC,EAAE,GACxD;AAGL,QAAO;EACL;EACA;EACA,aAAa;EACb;EACD;;AAGH,SAAgB,gBACd,eACA,QACQ;AACR,QAAO,QAAQ,eAAe,OAAO,YAAY,SAAS;;AAG5D,SAAS,iBAAiB,MAGxB;AAEA,KAAI,CAAC,KAAK,WAAW,MAAM,CAAE,QAAO;EAAE,aAAa;EAAM,MAAM;EAAM;CACrE,MAAM,QAAQ,KAAK,MAAM,KAAK;CAE9B,IAAI,QAAQ;AACZ,MAAK,IAAI,IAAI,GAAG,IAAI,MAAM,QAAQ,IAChC,KAAI,MAAM,GAAG,MAAM,KAAK,OAAO;AAC7B,UAAQ;AACR;;AAGJ,KAAI,UAAU,GAAI,QAAO;EAAE,aAAa;EAAM,MAAM;EAAM;AAC1D,QAAO;EACL,aAAa,MAAM,MAAM,GAAG,MAAM,CAAC,KAAK,KAAK;EAC7C,MAAM,MAAM,MAAM,QAAQ,EAAE,CAAC,KAAK,KAAK;EACxC"}