@tangle-network/agent-eval 0.93.0 → 0.94.0

package/dist/{chunk-CY6U5S3X.js → chunk-URWVKRCS.js}

@@ -5,7 +5,7 @@ import {
   pairedBootstrap,
   pairedMde,
   wilcoxonSignedRank
-} from "./chunk-LMZQ2Z4U.js";
+} from "./chunk-56GC6VYO.js";
 import {
   canonicalize,
   hashJson
@@ -877,4 +877,4 @@ export {
   RESEARCH_REPORT_HARD_PAIR_FLOOR,
   researchReport
 };
-//# sourceMappingURL=chunk-CY6U5S3X.js.map
+//# sourceMappingURL=chunk-URWVKRCS.js.map

package/dist/{chunk-T375SUOZ.js → chunk-ZOPLCJSY.js}

@@ -67,48 +67,101 @@ var SubprocessSandboxDriver = class {
     const start = Date.now();
     const effectiveCwd = config.cwd ?? this.defaultCwd;
     const effectiveEnv = { ...process.env, ...this.defaultEnv ?? {}, ...config.env ?? {} };
+    const maxOutputBytes = config.maxOutputBytes ?? 16 * 1024 * 1024;
     return await new Promise((resolve) => {
       const child = spawn(command, {
         shell: true,
         cwd: effectiveCwd,
-        env: effectiveEnv
+        env: effectiveEnv,
+        // Own process group so a timeout can SIGKILL the whole tree, not
+        // just the shell — otherwise a runaway grandchild keeps the pipes
+        // open and `close` never fires (the timeout silently does nothing).
+        detached: process.platform !== "win32"
       });
       let stdout = "";
       let stderr = "";
-      child.stdout?.on("data", (d) => {
-        stdout += String(d);
-      });
-      child.stderr?.on("data", (d) => {
-        stderr += String(d);
-      });
-      const timeout = setTimeout(
-        () => {
-          try {
-            child.kill("SIGKILL");
-          } catch (err) {
-            console.warn("[sandbox-harness] SIGKILL on timeout failed:", err);
-          }
-        },
-        config.timeoutMs ?? 10 * 6e4
-      );
-      child.on("close", (code) => {
+      let outputBytes = 0;
+      let outputTruncated = false;
+      let killedByTimeout = false;
+      let settled = false;
+      const onStdout = (d) => {
+        const chunk = capture(String(d));
+        if (chunk) stdout += chunk;
+      };
+      const onStderr = (d) => {
+        const chunk = capture(String(d));
+        if (chunk) stderr += chunk;
+      };
+      function capture(s) {
+        if (outputBytes >= maxOutputBytes) {
+          outputTruncated = true;
+          return "";
+        }
+        const room = maxOutputBytes - outputBytes;
+        if (s.length > room) {
+          outputBytes = maxOutputBytes;
+          outputTruncated = true;
+          return s.slice(0, room);
+        }
+        outputBytes += s.length;
+        return s;
+      }
+      child.stdout?.on("data", onStdout);
+      child.stderr?.on("data", onStderr);
+      const cleanup = () => {
         clearTimeout(timeout);
+        if (forceResolve) clearTimeout(forceResolve);
+        child.stdout?.off("data", onStdout);
+        child.stderr?.off("data", onStderr);
+        child.removeAllListeners("close");
+        child.removeAllListeners("error");
+      };
+      const finish = (outcome) => {
+        if (settled) return;
+        settled = true;
+        cleanup();
         const wallMs = Date.now() - start;
-        const parsed = phase === "test" && config.testParser ? config.testParser.parse(stdout, stderr, code ?? 1) : void 0;
+        const exitCode = killedByTimeout ? outcome.code && outcome.code !== 0 ? outcome.code : 124 : outcome.code ?? 1;
+        const parsed = !killedByTimeout && phase === "test" && config.testParser ? config.testParser.parse(stdout, stderr, exitCode) : void 0;
         resolve({
           phase,
-          exitCode: code ?? 1,
+          exitCode,
           stdout,
-          stderr,
+          stderr: outcome.runnerError ? stderr + outcome.runnerError : stderr,
           wallMs,
           testsTotal: parsed?.testsTotal,
-          testsPassed: parsed?.testsPassed
+          testsPassed: parsed?.testsPassed,
+          killedByTimeout: killedByTimeout || void 0,
+          outputTruncated: outputTruncated || void 0
         });
+      };
+      const killTree = () => {
+        try {
+          if (process.platform !== "win32" && typeof child.pid === "number") {
+            process.kill(-child.pid, "SIGKILL");
+          } else {
+            child.kill("SIGKILL");
+          }
+        } catch (err) {
+          console.warn("[sandbox-harness] SIGKILL on timeout failed:", err);
+        }
+      };
+      let forceResolve;
+      const timeout = setTimeout(
+        () => {
+          killedByTimeout = true;
+          killTree();
+          forceResolve = setTimeout(() => finish({ code: null }), 2e3);
+        },
+        config.timeoutMs ?? 10 * 6e4
+      );
+      child.on("close", (code) => {
+        if (forceResolve) clearTimeout(forceResolve);
+        finish({ code });
       });
       child.on("error", (err) => {
-        clearTimeout(timeout);
-        const wallMs = Date.now() - start;
-        resolve({ phase, exitCode: 127, stdout, stderr: stderr + String(err), wallMs });
+        if (forceResolve) clearTimeout(forceResolve);
+        finish({ code: 127, runnerError: String(err) });
       });
     });
   }
@@ -143,8 +196,9 @@ var SandboxHarness = class {
       if (config.setupCommand) {
         result.setup = await this.driver.exec("setup", config.setupCommand, config);
         result.totalWallMs += result.setup.wallMs;
-        if (result.setup.exitCode !== 0) {
-          await handle.fail(`setup failed (exit ${result.setup.exitCode})`, {
+        if (result.setup.killedByTimeout || result.setup.exitCode !== 0) {
+          const reason = result.setup.killedByTimeout ? `setup timed out after ${result.setup.wallMs}ms` : `setup failed (exit ${result.setup.exitCode})`;
+          await handle.fail(reason, {
             exitCode: result.setup.exitCode,
             wallMs: result.totalWallMs
           });
@@ -154,8 +208,9 @@ var SandboxHarness = class {
       if (config.runCommand) {
         result.run = await this.driver.exec("run", config.runCommand, config);
         result.totalWallMs += result.run.wallMs;
-        if (result.run.exitCode !== 0) {
-          await handle.fail(`run failed (exit ${result.run.exitCode})`, {
+        if (result.run.killedByTimeout || result.run.exitCode !== 0) {
+          const reason = result.run.killedByTimeout ? `run timed out after ${result.run.wallMs}ms` : `run failed (exit ${result.run.exitCode})`;
+          await handle.fail(reason, {
             exitCode: result.run.exitCode,
             wallMs: result.totalWallMs
           });
@@ -165,7 +220,7 @@ var SandboxHarness = class {
       if (config.testCommand) {
         result.test = await this.driver.exec("test", config.testCommand, config);
         result.totalWallMs += result.test.wallMs;
-        const passed = result.test.exitCode === 0;
+        const passed = !result.test.killedByTimeout && result.test.exitCode === 0;
         result.passed = passed;
         if (result.test.testsTotal !== void 0 && result.test.testsTotal > 0) {
           result.score = (result.test.testsPassed ?? 0) / result.test.testsTotal;
@@ -248,4 +303,4 @@ export {
   SandboxHarness,
   runTestGradedScenario
 };
-//# sourceMappingURL=chunk-T375SUOZ.js.map
+//# sourceMappingURL=chunk-ZOPLCJSY.js.map

package/dist/chunk-ZOPLCJSY.js.map

@@ -0,0 +1 @@
+{"version":3,"sources":["../src/sandbox-harness.ts","../src/test-graded-scenario.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` — spawn in a local cwd with env vars.\n *     Fast, no dependencies, fine for unit tests and most 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 { ConfigError } from './errors'\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\nexport interface 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    const { spawn } = await import('node:child_process')\n    const start = Date.now()\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    const effectiveEnv = { ...process.env, ...(this.defaultEnv ?? {}), ...(config.env ?? {}) }\n    const maxOutputBytes = config.maxOutputBytes ?? 16 * 1024 * 1024\n    return await new Promise<SandboxResult>((resolve) => {\n      const child = spawn(command, {\n        shell: true,\n        cwd: effectiveCwd,\n        env: effectiveEnv,\n        // Own process group so a timeout can SIGKILL the whole tree, not\n        // just the shell — otherwise a runaway grandchild keeps the pipes\n        // open and `close` never fires (the timeout silently does nothing).\n        detached: process.platform !== 'win32',\n      })\n      let stdout = ''\n      let stderr = ''\n      let outputBytes = 0\n      let outputTruncated = false\n      let killedByTimeout = false\n      let settled = false\n\n      const onStdout = (d: unknown) => {\n        const chunk = capture(String(d))\n        if (chunk) stdout += chunk\n      }\n      const onStderr = (d: unknown) => {\n        const chunk = capture(String(d))\n        if (chunk) stderr += chunk\n      }\n      // Bound the in-memory buffer so a runaway process can't OOM the\n      // harness. Once the cap is hit we keep draining the streams (to let\n      // the child reach `close`) but discard the bytes.\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      child.stdout?.on('data', onStdout)\n      child.stderr?.on('data', onStderr)\n\n      const cleanup = () => {\n        clearTimeout(timeout)\n        if (forceResolve) clearTimeout(forceResolve)\n        child.stdout?.off('data', onStdout)\n        child.stderr?.off('data', onStderr)\n        child.removeAllListeners('close')\n        child.removeAllListeners('error')\n      }\n\n      // Resolve exactly once. `code`/`signal` come from `close`; on the\n      // timeout path we force a non-zero exit so a SIGKILLed child that\n      // reports exit 0 can never read as a clean pass downstream.\n      const finish = (outcome: { code: number | null; runnerError?: string }) => {\n        if (settled) return\n        settled = true\n        cleanup()\n        const wallMs = Date.now() - start\n        const exitCode = killedByTimeout\n          ? outcome.code && outcome.code !== 0\n            ? outcome.code\n            : 124\n          : (outcome.code ?? 1)\n        const parsed =\n          !killedByTimeout && phase === 'test' && config.testParser\n            ? config.testParser.parse(stdout, stderr, exitCode)\n            : undefined\n        resolve({\n          phase,\n          exitCode,\n          stdout,\n          stderr: outcome.runnerError ? stderr + outcome.runnerError : stderr,\n          wallMs,\n          testsTotal: parsed?.testsTotal,\n          testsPassed: parsed?.testsPassed,\n          killedByTimeout: killedByTimeout || undefined,\n          outputTruncated: outputTruncated || undefined,\n        })\n      }\n\n      // Kill the whole process group on win32-less platforms (we spawned\n      // detached). On a clean close this is a no-op.\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          console.warn('[sandbox-harness] SIGKILL on timeout failed:', err)\n        }\n      }\n\n      let forceResolve: ReturnType<typeof setTimeout> | undefined\n      const timeout = setTimeout(\n        () => {\n          killedByTimeout = true\n          killTree()\n          // Give `close` a brief window to fire (flushing any final output)\n          // after the group dies; if an orphaned grandchild keeps the pipes\n          // open, force-resolve so the harness can't hang on a runaway.\n          forceResolve = setTimeout(() => finish({ code: null }), 2000)\n        },\n        config.timeoutMs ?? 10 * 60_000,\n      )\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        finish({ code: 127, runnerError: String(err) })\n      })\n    })\n  }\n}\n\nexport class DockerSandboxDriver implements SandboxDriver {\n  id = 'docker'\n\n  async exec(\n    phase: SandboxResult['phase'],\n    command: string,\n    config: HarnessConfig,\n  ): Promise<SandboxResult> {\n    if (!config.image) throw new ConfigError('DockerSandboxDriver requires config.image')\n    const sub = new SubprocessSandboxDriver()\n    const envArgs = Object.entries(config.env ?? {})\n      .map(([k, v]) => `-e ${shellQuote(k)}=${shellQuote(v)}`)\n      .join(' ')\n    const wrapped = `docker run --rm ${envArgs} ${shellQuote(config.image)} sh -c ${shellQuote(command)}`\n    return sub.exec(phase, wrapped, { ...config, env: undefined })\n  }\n}\n\nfunction shellQuote(v: string): string {\n  if (/^[A-Za-z0-9_\\-/.@:=]+$/.test(v)) return v\n  return `'${v.replace(/'/g, `'\\\\''`)}'`\n}\n\n// ── Harness orchestration ────────────────────────────────────────────\n\nexport interface SandboxHarnessResult {\n  passed: boolean\n  setup?: SandboxResult\n  run?: SandboxResult\n  test?: SandboxResult\n  totalWallMs: number\n  /** Final score — 0 when no tests; otherwise testsPassed/testsTotal. */\n  score: number\n}\n\nexport class SandboxHarness {\n  private driver: SandboxDriver\n  constructor(driver: SandboxDriver = new SubprocessSandboxDriver()) {\n    this.driver = driver\n  }\n\n  async run(config: HarnessConfig, emitter: TraceEmitter): Promise<SandboxHarnessResult> {\n    const handle = await emitter.sandbox({\n      name: `sandbox(${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\nexport interface 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"],"mappings":";;;;;;;;AAuFO,IAAM,mBAAqC;AAAA,EAChD,IAAI;AAAA,EACJ,MAAM,QAAQ;AACZ,UAAM,IAAI,OAAO,MAAM,sEAAsE;AAC7F,QAAI,CAAC,EAAG,QAAO;AACf,QAAI,SAAS;AACb,QAAI,SAAS;AACb,UAAM,IAAI,SAAS,EAAE,CAAC,GAAI,EAAE;AAC5B,UAAM,SAAS,EAAE,CAAC,EAAG,YAAY;AACjC,QAAI,WAAW,SAAU,WAAU;AAAA,QAC9B,WAAU;AACf,QAAI,EAAE,CAAC,KAAK,EAAE,CAAC,GAAG;AAChB,YAAM,IAAI,SAAS,EAAE,CAAC,GAAG,EAAE;AAC3B,UAAI,EAAE,CAAC,EAAE,YAAY,MAAM,SAAU,WAAU;AAAA,UAC1C,WAAU;AAAA,IACjB;AACA,WAAO,EAAE,YAAY,SAAS,QAAQ,aAAa,OAAO;AAAA,EAC5D;AACF;AAGO,IAAM,mBAAqC;AAAA,EAChD,IAAI;AAAA,EACJ,MAAM,QAAQ;AACZ,UAAM,QAAQ,OAAO,MAAM,6BAA6B;AACxD,UAAM,SAAS,OAAO,MAAM,gBAAgB;AAC5C,QAAI,CAAC,SAAS,CAAC,OAAQ,QAAO;AAC9B,WAAO,EAAE,YAAY,SAAS,MAAM,CAAC,GAAI,EAAE,GAAG,aAAa,SAAS,OAAO,CAAC,GAAI,EAAE,EAAE;AAAA,EACtF;AACF;AAGO,IAAM,iBAAmC;AAAA,EAC9C,IAAI;AAAA,EACJ,MAAM,QAAQ;AACZ,UAAM,IAAI,OAAO,MAAM,uEAAuE;AAC9F,QAAI,CAAC,EAAG,QAAO;AACf,WAAO,EAAE,YAAY,SAAS,EAAE,CAAC,GAAI,EAAE,GAAG,aAAa,SAAS,EAAE,CAAC,GAAI,EAAE,EAAE;AAAA,EAC7E;AACF;AAGO,SAAS,kBAAkB,SAA+C;AAC/E,SAAO;AAAA,IACL,IAAI,QAAQ,IAAI,CAAC,MAAM,EAAE,EAAE,EAAE,KAAK,GAAG;AAAA,IACrC,MAAM,QAAQ,QAAQ,UAAU;AAC9B,iBAAW,KAAK,SAAS;AACvB,cAAM,MAAM,EAAE,MAAM,QAAQ,QAAQ,QAAQ;AAC5C,YAAI,IAAK,QAAO;AAAA,MAClB;AACA,aAAO;AAAA,IACT;AAAA,EACF;AACF;AAqBO,IAAM,0BAAN,MAAuD;AAAA,EAC5D,KAAK;AAAA,EACG;AAAA,EACA;AAAA,EAER,YAAY,UAA0C,CAAC,GAAG;AACxD,SAAK,aAAa,QAAQ;AAC1B,SAAK,aAAa,QAAQ;AAAA,EAC5B;AAAA,EAEA,MAAM,KACJ,OACA,SACA,QACwB;AACxB,UAAM,EAAE,MAAM,IAAI,MAAM,OAAO,eAAoB;AACnD,UAAM,QAAQ,KAAK,IAAI;AAIvB,UAAM,eAAe,OAAO,OAAO,KAAK;AACxC,UAAM,eAAe,EAAE,GAAG,QAAQ,KAAK,GAAI,KAAK,cAAc,CAAC,GAAI,GAAI,OAAO,OAAO,CAAC,EAAG;AACzF,UAAM,iBAAiB,OAAO,kBAAkB,KAAK,OAAO;AAC5D,WAAO,MAAM,IAAI,QAAuB,CAAC,YAAY;AACnD,YAAM,QAAQ,MAAM,SAAS;AAAA,QAC3B,OAAO;AAAA,QACP,KAAK;AAAA,QACL,KAAK;AAAA;AAAA;AAAA;AAAA,QAIL,UAAU,QAAQ,aAAa;AAAA,MACjC,CAAC;AACD,UAAI,SAAS;AACb,UAAI,SAAS;AACb,UAAI,cAAc;AAClB,UAAI,kBAAkB;AACtB,UAAI,kBAAkB;AACtB,UAAI,UAAU;AAEd,YAAM,WAAW,CAAC,MAAe;AAC/B,cAAM,QAAQ,QAAQ,OAAO,CAAC,CAAC;AAC/B,YAAI,MAAO,WAAU;AAAA,MACvB;AACA,YAAM,WAAW,CAAC,MAAe;AAC/B,cAAM,QAAQ,QAAQ,OAAO,CAAC,CAAC;AAC/B,YAAI,MAAO,WAAU;AAAA,MACvB;AAIA,eAAS,QAAQ,GAAmB;AAClC,YAAI,eAAe,gBAAgB;AACjC,4BAAkB;AAClB,iBAAO;AAAA,QACT;AACA,cAAM,OAAO,iBAAiB;AAC9B,YAAI,EAAE,SAAS,MAAM;AACnB,wBAAc;AACd,4BAAkB;AAClB,iBAAO,EAAE,MAAM,GAAG,IAAI;AAAA,QACxB;AACA,uBAAe,EAAE;AACjB,eAAO;AAAA,MACT;AAEA,YAAM,QAAQ,GAAG,QAAQ,QAAQ;AACjC,YAAM,QAAQ,GAAG,QAAQ,QAAQ;AAEjC,YAAM,UAAU,MAAM;AACpB,qBAAa,OAAO;AACpB,YAAI,aAAc,cAAa,YAAY;AAC3C,cAAM,QAAQ,IAAI,QAAQ,QAAQ;AAClC,cAAM,QAAQ,IAAI,QAAQ,QAAQ;AAClC,cAAM,mBAAmB,OAAO;AAChC,cAAM,mBAAmB,OAAO;AAAA,MAClC;AAKA,YAAM,SAAS,CAAC,YAA2D;AACzE,YAAI,QAAS;AACb,kBAAU;AACV,gBAAQ;AACR,cAAM,SAAS,KAAK,IAAI,IAAI;AAC5B,cAAM,WAAW,kBACb,QAAQ,QAAQ,QAAQ,SAAS,IAC/B,QAAQ,OACR,MACD,QAAQ,QAAQ;AACrB,cAAM,SACJ,CAAC,mBAAmB,UAAU,UAAU,OAAO,aAC3C,OAAO,WAAW,MAAM,QAAQ,QAAQ,QAAQ,IAChD;AACN,gBAAQ;AAAA,UACN;AAAA,UACA;AAAA,UACA;AAAA,UACA,QAAQ,QAAQ,cAAc,SAAS,QAAQ,cAAc;AAAA,UAC7D;AAAA,UACA,YAAY,QAAQ;AAAA,UACpB,aAAa,QAAQ;AAAA,UACrB,iBAAiB,mBAAmB;AAAA,UACpC,iBAAiB,mBAAmB;AAAA,QACtC,CAAC;AAAA,MACH;AAIA,YAAM,WAAW,MAAM;AACrB,YAAI;AACF,cAAI,QAAQ,aAAa,WAAW,OAAO,MAAM,QAAQ,UAAU;AACjE,oBAAQ,KAAK,CAAC,MAAM,KAAK,SAAS;AAAA,UACpC,OAAO;AACL,kBAAM,KAAK,SAAS;AAAA,UACtB;AAAA,QACF,SAAS,KAAK;AACZ,kBAAQ,KAAK,gDAAgD,GAAG;AAAA,QAClE;AAAA,MACF;AAEA,UAAI;AACJ,YAAM,UAAU;AAAA,QACd,MAAM;AACJ,4BAAkB;AAClB,mBAAS;AAIT,yBAAe,WAAW,MAAM,OAAO,EAAE,MAAM,KAAK,CAAC,GAAG,GAAI;AAAA,QAC9D;AAAA,QACA,OAAO,aAAa,KAAK;AAAA,MAC3B;AAEA,YAAM,GAAG,SAAS,CAAC,SAAS;AAC1B,YAAI,aAAc,cAAa,YAAY;AAC3C,eAAO,EAAE,KAAK,CAAC;AAAA,MACjB,CAAC;AACD,YAAM,GAAG,SAAS,CAAC,QAAQ;AACzB,YAAI,aAAc,cAAa,YAAY;AAC3C,eAAO,EAAE,MAAM,KAAK,aAAa,OAAO,GAAG,EAAE,CAAC;AAAA,MAChD,CAAC;AAAA,IACH,CAAC;AAAA,EACH;AACF;AAEO,IAAM,sBAAN,MAAmD;AAAA,EACxD,KAAK;AAAA,EAEL,MAAM,KACJ,OACA,SACA,QACwB;AACxB,QAAI,CAAC,OAAO,MAAO,OAAM,IAAI,YAAY,2CAA2C;AACpF,UAAM,MAAM,IAAI,wBAAwB;AACxC,UAAM,UAAU,OAAO,QAAQ,OAAO,OAAO,CAAC,CAAC,EAC5C,IAAI,CAAC,CAAC,GAAG,CAAC,MAAM,MAAM,WAAW,CAAC,CAAC,IAAI,WAAW,CAAC,CAAC,EAAE,EACtD,KAAK,GAAG;AACX,UAAM,UAAU,mBAAmB,OAAO,IAAI,WAAW,OAAO,KAAK,CAAC,UAAU,WAAW,OAAO,CAAC;AACnG,WAAO,IAAI,KAAK,OAAO,SAAS,EAAE,GAAG,QAAQ,KAAK,OAAU,CAAC;AAAA,EAC/D;AACF;AAEA,SAAS,WAAW,GAAmB;AACrC,MAAI,yBAAyB,KAAK,CAAC,EAAG,QAAO;AAC7C,SAAO,IAAI,EAAE,QAAQ,MAAM,OAAO,CAAC;AACrC;AAcO,IAAM,iBAAN,MAAqB;AAAA,EAClB;AAAA,EACR,YAAY,SAAwB,IAAI,wBAAwB,GAAG;AACjE,SAAK,SAAS;AAAA,EAChB;AAAA,EAEA,MAAM,IAAI,QAAuB,SAAsD;AACrF,UAAM,SAAS,MAAM,QAAQ,QAAQ;AAAA,MACnC,MAAM,WAAW,KAAK,OAAO,EAAE;AAAA,MAC/B,OAAO,OAAO;AAAA,MACd,SAAS,CAAC,OAAO,cAAc,OAAO,YAAY,OAAO,WAAW,EACjE,OAAO,OAAO,EACd,KAAK,MAAM;AAAA,IAChB,CAAC;AACD,UAAM,SAA+B,EAAE,QAAQ,OAAO,aAAa,GAAG,OAAO,EAAE;AAC/E,QAAI;AACF,UAAI,OAAO,cAAc;AACvB,eAAO,QAAQ,MAAM,KAAK,OAAO,KAAK,SAAS,OAAO,cAAc,MAAM;AAC1E,eAAO,eAAe,OAAO,MAAM;AACnC,YAAI,OAAO,MAAM,mBAAmB,OAAO,MAAM,aAAa,GAAG;AAC/D,gBAAM,SAAS,OAAO,MAAM,kBACxB,yBAAyB,OAAO,MAAM,MAAM,OAC5C,sBAAsB,OAAO,MAAM,QAAQ;AAC/C,gBAAM,OAAO,KAAK,QAAQ;AAAA,YACxB,UAAU,OAAO,MAAM;AAAA,YACvB,QAAQ,OAAO;AAAA,UACjB,CAAyB;AACzB,iBAAO;AAAA,QACT;AAAA,MACF;AACA,UAAI,OAAO,YAAY;AACrB,eAAO,MAAM,MAAM,KAAK,OAAO,KAAK,OAAO,OAAO,YAAY,MAAM;AACpE,eAAO,eAAe,OAAO,IAAI;AACjC,YAAI,OAAO,IAAI,mBAAmB,OAAO,IAAI,aAAa,GAAG;AAC3D,gBAAM,SAAS,OAAO,IAAI,kBACtB,uBAAuB,OAAO,IAAI,MAAM,OACxC,oBAAoB,OAAO,IAAI,QAAQ;AAC3C,gBAAM,OAAO,KAAK,QAAQ;AAAA,YACxB,UAAU,OAAO,IAAI;AAAA,YACrB,QAAQ,OAAO;AAAA,UACjB,CAAyB;AACzB,iBAAO;AAAA,QACT;AAAA,MACF;AACA,UAAI,OAAO,aAAa;AACtB,eAAO,OAAO,MAAM,KAAK,OAAO,KAAK,QAAQ,OAAO,aAAa,MAAM;AACvE,eAAO,eAAe,OAAO,KAAK;AAGlC,cAAM,SAAS,CAAC,OAAO,KAAK,mBAAmB,OAAO,KAAK,aAAa;AACxE,eAAO,SAAS;AAChB,YAAI,OAAO,KAAK,eAAe,UAAa,OAAO,KAAK,aAAa,GAAG;AACtE,iBAAO,SAAS,OAAO,KAAK,eAAe,KAAK,OAAO,KAAK;AAAA,QAC9D,OAAO;AACL,iBAAO,QAAQ,SAAS,IAAI;AAAA,QAC9B;AACA,cAAM,OAAO,IAAI;AAAA,UACf,UAAU,OAAO,KAAK;AAAA,UACtB,YAAY,OAAO,KAAK;AAAA,UACxB,aAAa,OAAO,KAAK;AAAA,UACzB,QAAQ,OAAO;AAAA,UACf,QAAQ,SAAS,OAAO;AAAA,QAC1B,CAAyB;AAAA,MAC3B,OAAO;AACL,eAAO,SAAS;AAChB,eAAO,QAAQ;AACf,cAAM,OAAO,IAAI,EAAE,QAAQ,OAAO,YAAY,CAAyB;AAAA,MACzE;AAAA,IACF,SAAS,KAAK;AACZ,YAAM,OAAO,KAAK,eAAe,QAAQ,MAAM,OAAO,GAAG,CAAC;AAC1D,YAAM;AAAA,IACR;AACA,WAAO;AAAA,EACT;AACF;;;AClXA,eAAsB,sBACpB,UACA,OACA,UAAgC,CAAC,GACH;AAC9B,QAAM,UAAU,IAAI,aAAa,KAAK;AACtC,QAAM,QAAQ,SAAS;AAAA,IACrB,YAAY,SAAS;AAAA,IACrB,WAAW,QAAQ;AAAA,IACnB,gBAAgB,SAAS;AAAA,IACzB,MAAM,SAAS;AAAA,IACf,GAAG,QAAQ;AAAA,EACb,CAAC;AACD,QAAM,UAAU,IAAI,eAAe,QAAQ,MAAM;AACjD,QAAM,SAAS,MAAM,QAAQ,IAAI,SAAS,SAAS,OAAO;AAC1D,QAAM,YAAY,SAAS,iBAAiB;AAC5C,QAAM,OAAO,OAAO,UAAU,OAAO,SAAS;AAC9C,QAAM,cAAc,OAAO,UAAU,UAAa,OAAO,MAAM,aAAa;AAC5E,QAAM,YAAY,OAAO,QAAQ,UAAa,OAAO,IAAI,aAAa;AACtE,QAAM,aAAa,OAAO,SAAS,UAAa,OAAO,KAAK,aAAa;AACzE,QAAM,eAAyC,OAC3C,YACA,eAAe,YACb,oBACA,aACE,iBACA;AACR,QAAM,QAAQ,OAAO;AAAA,IACnB;AAAA,IACA,OAAO,OAAO;AAAA,IACd;AAAA,IACA,OAAO,OAAO,SAAY,iBAAiB,MAAM;AAAA,EACnD,CAAC;AACD,SAAO;AAAA,IACL,OAAO,QAAQ;AAAA,IACf;AAAA,IACA,SAAS;AAAA,IACT;AAAA,IACA,OAAO,OAAO;AAAA,IACd;AAAA,EACF;AACF;AAEA,SAAS,iBAAiB,QAAsC;AAC9D,MAAI,OAAO,SAAS,OAAO,MAAM,aAAa;AAC5C,WAAO,sBAAsB,OAAO,MAAM,QAAQ;AACpD,MAAI,OAAO,OAAO,OAAO,IAAI,aAAa,EAAG,QAAO,oBAAoB,OAAO,IAAI,QAAQ;AAC3F,MAAI,OAAO,MAAM;AACf,QAAI,OAAO,KAAK,eAAe,QAAW;AACxC,aAAO,UAAU,OAAO,KAAK,eAAe,CAAC,IAAI,OAAO,KAAK,UAAU;AAAA,IACzE;AACA,WAAO,aAAa,OAAO,KAAK,QAAQ;AAAA,EAC1C;AACA,SAAO;AACT;","names":[]}

package/dist/cli.js

@@ -5,8 +5,8 @@ import {
   runRpcBatch,
   runRpcOnce,
   startServer
-} from "./chunk-QS3RBQPI.js";
-import "./chunk-CVVHBFGN.js";
+} from "./chunk-D5KBVULY.js";
+import "./chunk-CWNP4DV4.js";
 import "./chunk-PC4UYEBM.js";
 import "./chunk-3BFEG2F6.js";
 import "./chunk-PZ5AY32C.js";

package/dist/contract/index.d.ts

@@ -1,34 +1,34 @@
 import { S as Scenario, M as MutableSurface, b as DispatchContext, a as JudgeConfig, I as ImprovementDriver, G as Gate, L as LabeledScenarioStore, c as GateDecision } from '../types-BU-7W85F.js';
 export { C as CampaignAggregates, d as CampaignArtifactWriter, e as CampaignCellResult, f as CampaignCostMeter, g as CampaignResult, h as CampaignTraceWriter, i as CodeSurface, D as Dispatch, j as GateContext, k as GateResult, l as GenerationCandidate, m as GenerationRecord, n as JudgeDimension, J as JudgeScore, o as Mutator, O as OptimizerConfig, p as SessionScript } from '../types-BU-7W85F.js';
-import { L as LoopProvenanceRecord } from '../provenance-LnqRT0sS.js';
-export { A as AxisEvidence, a as AxisVerdict, B as BuildEvidenceVectorOptions, D as DefaultProductionGateOptions, E as EvidenceVector, b as EvolutionaryDriverOptions, H as HeldOutGateOptions, O as ObjectiveSource, P as ParetoSignificanceGateOptions, c as PromotionObjective, d as PromotionPolicy, R as RunEvalOptions, e as buildEvidenceVector, f as composeGate, g as defaultProductionGate, h as evolutionaryDriver, i as heldOutGate, p as paretoPolicy, j as paretoSignificanceGate, r as runEval } from '../provenance-LnqRT0sS.js';
-import { C as CampaignStorage, R as RunOptimizationOptions, a as RunImprovementLoopResult } from '../run-improvement-loop-5z_l5zDz.js';
-export { G as GepaDriverOptions, b as RunCampaignOptions, c as RunImprovementLoopOptions, f as fsCampaignStorage, g as gepaDriver, i as inMemoryCampaignStorage, r as runCampaign, d as runImprovementLoop } from '../run-improvement-loop-5z_l5zDz.js';
+import { L as LoopProvenanceRecord } from '../provenance-P-bCL2Fo.js';
+export { A as AxisEvidence, a as AxisVerdict, B as BuildEvidenceVectorOptions, D as DefaultProductionGateOptions, E as EvidenceVector, b as EvolutionaryDriverOptions, H as HeldOutGateOptions, O as ObjectiveSource, P as ParetoSignificanceGateOptions, c as PromotionObjective, d as PromotionPolicy, R as RunEvalOptions, e as buildEvidenceVector, f as composeGate, g as defaultProductionGate, h as evolutionaryDriver, i as heldOutGate, p as paretoPolicy, j as paretoSignificanceGate, r as runEval } from '../provenance-P-bCL2Fo.js';
+import { C as CampaignStorage, R as RunOptimizationOptions, a as RunImprovementLoopResult } from '../run-improvement-loop-DBahB8Ax.js';
+export { G as GepaDriverOptions, b as RunCampaignOptions, c as RunImprovementLoopOptions, f as fsCampaignStorage, g as gepaDriver, i as inMemoryCampaignStorage, r as runCampaign, d as runImprovementLoop } from '../run-improvement-loop-DBahB8Ax.js';
 export { D as DeploymentOutcome, F as FileSystemOutcomeStore, a as FileSystemOutcomeStoreOptions, I as InMemoryOutcomeStore, b as OutcomeStore } from '../outcome-store-rnXLEqSn.js';
 import { HostedTenant, EvalRunCellScore, EvalRunGenerationSnapshot, EvalRunEvent, TraceSpanEvent } from '../hosted/index.js';
 import { R as RunRecord, a as RunSplitTag } from '../run-record-e7vj1uZQ.js';
-import { I as InsightReport } from '../insight-report-BBwvOh6x.js';
-export { F as FailureClusterInsight, a as InterRaterInsight, J as JudgeInsight, L as LiftInsight, O as OutcomeCorrelationInsight, R as Recommendation, b as ReleaseSummary, S as ScalarDistribution } from '../insight-report-BBwvOh6x.js';
-export { D as DefaultAnalystRegistryOptions, b as buildDefaultAnalystRegistry } from '../default-registry-zoGHUQEH.js';
-export { c as AnalystFinding } from '../types-2VVIL04s.js';
-export { A as AnalyzeRunsOptions, a as analyzeRuns } from '../analyze-runs-DwCEkpO_.js';
+import { I as InsightReport } from '../insight-report-DWl3z9tl.js';
+export { F as FailureClusterInsight, a as InterRaterInsight, J as JudgeInsight, L as LiftInsight, O as OutcomeCorrelationInsight, R as Recommendation, b as ReleaseSummary, S as ScalarDistribution } from '../insight-report-DWl3z9tl.js';
+export { D as DefaultAnalystRegistryOptions, b as buildDefaultAnalystRegistry } from '../default-registry-6dhErQbs.js';
+export { c as AnalystFinding } from '../types-Ce17tDlG.js';
+export { A as AnalyzeRunsOptions, a as analyzeRuns } from '../analyze-runs-B6Ljo_dI.js';
 export { C as CodeAgentSessionDiagnostic, a as CodeAgentSessionIntakeOptions, b as CodeAgentSessionIntakeResult, c as CodeAgentSessionMetrics, d as CodeAgentSessionSource, P as ParsedCodeAgentJsonl, f as fromClaudeCodeSession, e as fromCodexSession, g as fromKimiCodeSession, h as fromOpenCodeSession, i as fromPiSession, j as fromPigraphSession, p as parseCodeAgentJsonl } from '../code-agent-session-BO8nCnv3.js';
-import '../red-team-BXHil6c8.js';
+import '../red-team-BWdoyleI.js';
 import '../dataset-BbGkaN2I.js';
 import '../errors-CzMUYo7b.js';
-import '../store-CKUAgsJz.js';
+import '../store-BcFXE6LG.js';
 import '../schema-m0gsnbt3.js';
 import '../pareto-E-pembql.js';
-import '../statistics-C7PozGrZ.js';
+import '../statistics-CCJpTGOS.js';
 import '../judge-calibration-0p2QcWNE.js';
-import '../types-Croy5h7V.js';
+import '../types-C7DGg5ex.js';
 import '@tangle-network/tcloud';
-import '../llm-client-BeEcAokY.js';
+import '../llm-client-Bj7g0rqu.js';
 import '../raw-provider-sink-C46HDghv.js';
-import '../summary-report-DGmUucwQ.js';
-import '../failure-cluster-CL7IVgkJ.js';
+import '../summary-report-BDOFevaT.js';
+import '../failure-cluster-DH9Flgcf.js';
 import '@ax-llm/ax';
-import '../kind-factory-5b7xXXOr.js';
+import '../kind-factory-0BhLSI27.js';
 import '../store-C1YxJDEK.js';
 import 'zod';
 

package/dist/contract/index.js

@@ -18,40 +18,40 @@ import {
   paretoPolicy,
   paretoSignificanceGate,
   runEval
-} from "../chunk-6SOJM3VR.js";
+} from "../chunk-SD2YFWQQ.js";
 import {
   analyzeRuns
-} from "../chunk-Y47J2LJ3.js";
+} from "../chunk-2KNZHH3P.js";
 import {
   emitLoopProvenance,
   gepaDriver,
   heldOutGate,
   runImprovementLoop,
   surfaceContentHash
-} from "../chunk-ZFIBGEOL.js";
+} from "../chunk-EGPMSBEZ.js";
 import {
   fsCampaignStorage,
   inMemoryCampaignStorage,
   runCampaign
-} from "../chunk-TWS7AZEY.js";
+} from "../chunk-2K6UUZ7P.js";
 import "../chunk-VI2UW6B6.js";
 import {
   buildDefaultAnalystRegistry
-} from "../chunk-UMMZHCPB.js";
-import "../chunk-D3V5B42D.js";
+} from "../chunk-MIFZUPEK.js";
+import "../chunk-Q5LIB7BC.js";
 import "../chunk-5HRORJQY.js";
 import {
   FileSystemOutcomeStore,
   InMemoryOutcomeStore
 } from "../chunk-3RF76KTD.js";
-import "../chunk-CVVHBFGN.js";
-import "../chunk-CY6U5S3X.js";
+import "../chunk-CWNP4DV4.js";
+import "../chunk-URWVKRCS.js";
 import "../chunk-45EEMHTC.js";
-import "../chunk-LMZQ2Z4U.js";
+import "../chunk-56GC6VYO.js";
 import "../chunk-GGE4NNQT.js";
-import "../chunk-47X6LRCE.js";
+import "../chunk-UFSG7ACU.js";
 import "../chunk-5BKGXME7.js";
-import "../chunk-QAY5UIJO.js";
+import "../chunk-QMUEXQJS.js";
 import "../chunk-PC4UYEBM.js";
 import "../chunk-VSMTAMNK.js";
 import "../chunk-3BFEG2F6.js";

package/dist/{control-_Qb7skHX.d.ts → control-D6qwHXIR.d.ts}

@@ -1,8 +1,8 @@
-import { b as FeedbackLabel, p as ProposedSideEffect } from './feedback-trajectory-D9OVLrg9.js';
-import { C as ControlEvalResult, a as ControlRunResult, h as ControlRuntimeConfig } from './control-runtime-DuFBYg7A.js';
-import { T as TraceEmitter } from './emitter-DEZwY14K.js';
+import { b as FeedbackLabel, p as ProposedSideEffect } from './feedback-trajectory-BxY0cKfs.js';
+import { C as ControlEvalResult, a as ControlRunResult, h as ControlRuntimeConfig } from './control-runtime-Acf9CGhw.js';
+import { T as TraceEmitter } from './emitter-C2rqGH_l.js';
 import { F as FailureClass } from './schema-m0gsnbt3.js';
-import { T as TraceStore } from './store-CKUAgsJz.js';
+import { T as TraceStore } from './store-BcFXE6LG.js';
 import { a as RunSplitTag, b as RunTokenUsage, R as RunRecord } from './run-record-e7vj1uZQ.js';
 
 interface ActionExecutionPolicy {

package/dist/{control-runtime-DuFBYg7A.d.ts → control-runtime-Acf9CGhw.d.ts}

@@ -1,6 +1,6 @@
-import { T as TraceEmitter } from './emitter-DEZwY14K.js';
+import { T as TraceEmitter } from './emitter-C2rqGH_l.js';
 import { F as FailureClass } from './schema-m0gsnbt3.js';
-import { T as TraceStore } from './store-CKUAgsJz.js';
+import { T as TraceStore } from './store-BcFXE6LG.js';
 
 /**
  * Policy-based agent control runtime.

package/dist/control.d.ts

@@ -1,9 +1,9 @@
-export { A as ActionExecutionPolicy, a as ActionPolicyDecision, C as ControlRunToRunRecordOptions, e as ProposeReviewConfig, f as ProposeReviewControlAction, g as ProposeReviewControlConfig, h as ProposeReviewControlResult, i as ProposeReviewControlState, j as ProposeReviewReport, p as RunEvidenceMetadata, s as controlRunToRunRecord, u as evaluateActionPolicy, x as runProposeReview, y as runProposeReviewAsControlLoop, z as scoreFromEvals } from './control-_Qb7skHX.js';
-export { c as ControlActionFailureMode, d as ControlActionOutcome, e as ControlBudget, f as ControlContext, g as ControlDecision, C as ControlEvalResult, a as ControlRunResult, h as ControlRuntimeConfig, i as ControlRuntimeError, j as ControlSeverity, b as ControlStep, k as ControlStopPolicies, S as StopDecision, l as allCriticalPassed, o as objectiveEval, r as runAgentControlLoop, s as stopOnNoProgress, m as stopOnRepeatedAction, n as subjectiveEval } from './control-runtime-DuFBYg7A.js';
-import './feedback-trajectory-D9OVLrg9.js';
+export { A as ActionExecutionPolicy, a as ActionPolicyDecision, C as ControlRunToRunRecordOptions, e as ProposeReviewConfig, f as ProposeReviewControlAction, g as ProposeReviewControlConfig, h as ProposeReviewControlResult, i as ProposeReviewControlState, j as ProposeReviewReport, p as RunEvidenceMetadata, s as controlRunToRunRecord, u as evaluateActionPolicy, x as runProposeReview, y as runProposeReviewAsControlLoop, z as scoreFromEvals } from './control-D6qwHXIR.js';
+export { c as ControlActionFailureMode, d as ControlActionOutcome, e as ControlBudget, f as ControlContext, g as ControlDecision, C as ControlEvalResult, a as ControlRunResult, h as ControlRuntimeConfig, i as ControlRuntimeError, j as ControlSeverity, b as ControlStep, k as ControlStopPolicies, S as StopDecision, l as allCriticalPassed, o as objectiveEval, r as runAgentControlLoop, s as stopOnNoProgress, m as stopOnRepeatedAction, n as subjectiveEval } from './control-runtime-Acf9CGhw.js';
+import './feedback-trajectory-BxY0cKfs.js';
 import './dataset-BbGkaN2I.js';
 import './errors-CzMUYo7b.js';
-import './emitter-DEZwY14K.js';
+import './emitter-C2rqGH_l.js';
 import './schema-m0gsnbt3.js';
-import './store-CKUAgsJz.js';
+import './store-BcFXE6LG.js';
 import './run-record-e7vj1uZQ.js';

package/dist/{corpus-BoR-041R.d.ts → corpus-B8A4BDR3.d.ts}

@@ -1,6 +1,6 @@
 import { R as RunRecord, a as RunSplitTag } from './run-record-e7vj1uZQ.js';
 import { S as Span } from './schema-m0gsnbt3.js';
-import { T as TraceStore } from './store-CKUAgsJz.js';
+import { T as TraceStore } from './store-BcFXE6LG.js';
 
 /**
  * Preference dataset extraction — bridge from `RunRecord[]` to RL training.

package/dist/{counterfactual-Dwibr5IW.d.ts → counterfactual-DlOz8PBx.d.ts}

@@ -1,6 +1,6 @@
-import { T as TraceEmitter } from './emitter-DEZwY14K.js';
-import { T as TraceStore } from './store-CKUAgsJz.js';
-import { T as Trajectory, a as TrajectoryStep } from './trajectory-GEdXJCL5.js';
+import { T as TraceEmitter } from './emitter-C2rqGH_l.js';
+import { T as TraceStore } from './store-BcFXE6LG.js';
+import { T as Trajectory, a as TrajectoryStep } from './trajectory-2TkpSEVh.js';
 
 /**
  * Counterfactual replay — "what would have happened if we'd changed

package/dist/{default-registry-zoGHUQEH.d.ts → default-registry-6dhErQbs.d.ts}

@@ -1,6 +1,6 @@
 import { AxAIService } from '@ax-llm/ax';
-import { T as TraceAnalystKindSpec } from './kind-factory-5b7xXXOr.js';
-import { A as Analyst, a as AnalystContext, b as AnalystRunSummary, c as AnalystFinding, d as AnalystRunResult, C as ChatClient, e as AnalystRunInputs, f as AnalystRunEvent } from './types-2VVIL04s.js';
+import { T as TraceAnalystKindSpec } from './kind-factory-0BhLSI27.js';
+import { A as Analyst, a as AnalystContext, b as AnalystRunSummary, c as AnalystFinding, d as AnalystRunResult, C as ChatClient, e as AnalystRunInputs, f as AnalystRunEvent } from './types-Ce17tDlG.js';
 
 /**
  * AnalystRegistry — orchestrate N analysts against one run.

package/dist/diagnose.d.ts

@@ -1,16 +1,16 @@
-import { C as CounterfactualMutation, a as attributeCounterfactuals, b as CounterfactualRunner } from './counterfactual-Dwibr5IW.js';
-export { c as CounterfactualContext, d as CounterfactualResult } from './counterfactual-Dwibr5IW.js';
+import { C as CounterfactualMutation, a as attributeCounterfactuals, b as CounterfactualRunner } from './counterfactual-DlOz8PBx.js';
+export { c as CounterfactualContext, d as CounterfactualResult } from './counterfactual-DlOz8PBx.js';
 import { S as Span } from './schema-m0gsnbt3.js';
-import { T as TraceStore } from './store-CKUAgsJz.js';
-import { a as TrajectoryStep, T as Trajectory } from './trajectory-GEdXJCL5.js';
-import { h as AnalystSeverity, c as AnalystFinding } from './types-2VVIL04s.js';
-import { C as CorpusRecord } from './corpus-BoR-041R.js';
+import { T as TraceStore } from './store-BcFXE6LG.js';
+import { a as TrajectoryStep, T as Trajectory } from './trajectory-2TkpSEVh.js';
+import { h as AnalystSeverity, c as AnalystFinding } from './types-Ce17tDlG.js';
+import { C as CorpusRecord } from './corpus-B8A4BDR3.js';
 import { R as RunRecord } from './run-record-e7vj1uZQ.js';
-import './emitter-DEZwY14K.js';
+import './emitter-C2rqGH_l.js';
 import './store-C1YxJDEK.js';
-import './types-Croy5h7V.js';
+import './types-C7DGg5ex.js';
 import '@tangle-network/tcloud';
-import './llm-client-BeEcAokY.js';
+import './llm-client-Bj7g0rqu.js';
 import './errors-CzMUYo7b.js';
 import './raw-provider-sink-C46HDghv.js';
 

package/dist/diagnose.js

@@ -10,7 +10,7 @@ import {
 } from "./chunk-45EEMHTC.js";
 import {
   confidenceInterval
-} from "./chunk-LMZQ2Z4U.js";
+} from "./chunk-56GC6VYO.js";
 import {
   validateRunRecord
 } from "./chunk-KWRRMR3J.js";

package/dist/{emitter-DEZwY14K.d.ts → emitter-C2rqGH_l.d.ts}

@@ -1,5 +1,5 @@
 import { b as RunOutcome, R as Run, S as Span, c as SpanKind, L as LlmSpan, T as ToolSpan, d as RetrievalSpan, J as JudgeSpan, e as SandboxSpan, E as EventKind, a as TraceEvent, B as BudgetLedgerEntry, A as Artifact, M as Message } from './schema-m0gsnbt3.js';
-import { T as TraceStore } from './store-CKUAgsJz.js';
+import { T as TraceStore } from './store-BcFXE6LG.js';
 
 /**
  * TraceEmitter — hierarchical span builder that auto-parents using an

package/dist/{failure-cluster-CL7IVgkJ.d.ts → failure-cluster-DH9Flgcf.d.ts}

@@ -1,5 +1,5 @@
 import { R as Run, S as Span, a as TraceEvent, F as FailureClass } from './schema-m0gsnbt3.js';
-import { T as TraceStore } from './store-CKUAgsJz.js';
+import { T as TraceStore } from './store-BcFXE6LG.js';
 
 /**
  * Failure taxonomy — canonical classes + a default classifier.

package/dist/{feedback-trajectory-D9OVLrg9.d.ts → feedback-trajectory-BxY0cKfs.d.ts}

@@ -1,4 +1,4 @@
-import { C as ControlEvalResult, a as ControlRunResult, b as ControlStep } from './control-runtime-DuFBYg7A.js';
+import { C as ControlEvalResult, a as ControlRunResult, b as ControlStep } from './control-runtime-Acf9CGhw.js';
 import { D as DatasetSplit, a as DatasetScenario } from './dataset-BbGkaN2I.js';
 
 type FeedbackArtifactType = 'text' | 'code' | 'plan' | 'research' | 'action' | 'ui' | 'decision' | 'data' | 'other';

package/dist/governance/index.d.ts

@@ -1,8 +1,8 @@
 import { c as DatasetManifest } from '../dataset-BbGkaN2I.js';
 import { a as CalibrationResult } from '../judge-calibration-0p2QcWNE.js';
 import { b as OutcomeStore } from '../outcome-store-rnXLEqSn.js';
-import { d as RedTeamReport } from '../red-team-BXHil6c8.js';
-import { T as TraceStore } from '../store-CKUAgsJz.js';
+import { d as RedTeamReport } from '../red-team-BWdoyleI.js';
+import { T as TraceStore } from '../store-BcFXE6LG.js';
 import '../errors-CzMUYo7b.js';
 import '../schema-m0gsnbt3.js';
 

package/dist/{harness-optimizer-EnEnQPsr.d.ts → harness-optimizer-mOl9XX_O.d.ts}

@@ -1,5 +1,5 @@
 import { O as Objective, P as ParetoResult } from './pareto-E-pembql.js';
-import { R as RunScore, a as RunTrace, b as RunScoreWeights } from './run-critic-BAIjX99r.js';
+import { R as RunScore, a as RunTrace, b as RunScoreWeights } from './run-critic-CmMf05uV.js';
 
 interface SteeringRolePrompt {
     system?: string;

package/dist/hosted/index.d.ts

@@ -1,11 +1,11 @@
 import { M as MutableSurface, c as GateDecision } from '../types-BU-7W85F.js';
-import { I as InsightReport } from '../insight-report-BBwvOh6x.js';
+import { I as InsightReport } from '../insight-report-DWl3z9tl.js';
 import '../run-record-e7vj1uZQ.js';
 import '../errors-CzMUYo7b.js';
 import '../schema-m0gsnbt3.js';
-import '../summary-report-DGmUucwQ.js';
-import '../failure-cluster-CL7IVgkJ.js';
-import '../store-CKUAgsJz.js';
+import '../summary-report-BDOFevaT.js';
+import '../failure-cluster-DH9Flgcf.js';
+import '../store-BcFXE6LG.js';
 import '../judge-calibration-0p2QcWNE.js';
 
 /**