agentv 4.10.0 → 4.11.2-next.1

package/dist/{chunk-XOSNETAV.js → chunk-BAUNAXHT.js}

@@ -562,4 +562,4 @@ export {
   serviceInstanceIdDetector,
   init_esm4 as init_esm
 };
-//# sourceMappingURL=chunk-XOSNETAV.js.map
+//# sourceMappingURL=chunk-BAUNAXHT.js.map

package/dist/chunk-BPGJ4HBU.js

@@ -0,0 +1,183 @@
+import { createRequire } from 'node:module'; const require = createRequire(import.meta.url);
+
+// ../../packages/core/dist/chunk-SDIANPEY.js
+var DefaultCommandExecutor = class {
+  async exec(argv, options = {}) {
+    const { execFileWithStdin } = await import("./exec-AR6JUUN5-6MBPURPR.js");
+    return execFileWithStdin(argv, options.stdin ?? "", {
+      timeoutMs: options.timeoutMs
+    });
+  }
+};
+var DEFAULT_TIMEOUT_S = 1800;
+function buildGitCommand(target, args) {
+  if (!target?.path) {
+    return ["git", ...args];
+  }
+  return ["git", "-C", target.path, ...args];
+}
+var DockerWorkspaceProvider = class {
+  config;
+  executor;
+  timeoutMs;
+  constructor(config, executor) {
+    this.config = config;
+    this.executor = executor ?? new DefaultCommandExecutor();
+    this.timeoutMs = (config.timeout ?? DEFAULT_TIMEOUT_S) * 1e3;
+  }
+  /** Check whether the Docker CLI is available on the host. */
+  async isDockerAvailable() {
+    try {
+      const result = await this.executor.exec(
+        ["docker", "version", "--format", "{{.Server.Version}}"],
+        {
+          timeoutMs: 1e4
+        }
+      );
+      return result.exitCode === 0;
+    } catch {
+      return false;
+    }
+  }
+  /** Pull the configured Docker image. No-op if already cached locally. */
+  async pullImage() {
+    const inspectResult = await this.executor.exec(
+      ["docker", "image", "inspect", this.config.image],
+      {
+        timeoutMs: 1e4
+      }
+    );
+    if (inspectResult.exitCode === 0) {
+      return;
+    }
+    const result = await this.executor.exec(["docker", "pull", this.config.image], {
+      timeoutMs: this.timeoutMs
+    });
+    if (result.exitCode !== 0) {
+      throw new Error(`docker pull failed (exit ${result.exitCode}): ${result.stderr.trim()}`);
+    }
+  }
+  /** Create a stopped container from the configured image with resource limits. Returns container ID. */
+  async createContainer() {
+    const argv = ["docker", "create"];
+    if (this.config.memory) {
+      argv.push(`--memory=${this.config.memory}`);
+    }
+    if (this.config.cpus !== void 0) {
+      argv.push(`--cpus=${this.config.cpus}`);
+    }
+    argv.push(this.config.image, "sleep", "infinity");
+    const result = await this.executor.exec(argv, { timeoutMs: 3e4 });
+    if (result.exitCode !== 0) {
+      throw new Error(`docker create failed (exit ${result.exitCode}): ${result.stderr.trim()}`);
+    }
+    return result.stdout.trim();
+  }
+  /** Start a previously created container. */
+  async startContainer(containerId) {
+    const result = await this.executor.exec(["docker", "start", containerId], {
+      timeoutMs: 3e4
+    });
+    if (result.exitCode !== 0) {
+      throw new Error(`docker start failed (exit ${result.exitCode}): ${result.stderr.trim()}`);
+    }
+  }
+  /**
+   * Reset the container checkout to the specified target refs, if any.
+   * This is used for SWE-bench images where the repo state must match the
+   * dataset's base snapshot before grading begins.
+   */
+  async resetContainerCheckout(containerId, repoCheckouts) {
+    if (!repoCheckouts || repoCheckouts.length === 0) {
+      return;
+    }
+    for (const target of repoCheckouts) {
+      const resetResult = await this.execInContainer({
+        containerId,
+        command: buildGitCommand(target, ["reset", "--hard", target.ref])
+      });
+      if (resetResult.exitCode !== 0) {
+        throw new Error(
+          `docker git reset failed (exit ${resetResult.exitCode}): ${resetResult.stderr.trim()}`
+        );
+      }
+      const verifyResult = await this.execInContainer({
+        containerId,
+        command: buildGitCommand(target, ["rev-parse", "HEAD"]),
+        timeoutMs: 3e4
+      });
+      if (verifyResult.exitCode !== 0) {
+        throw new Error(
+          `docker checkout verification failed (exit ${verifyResult.exitCode}): ${verifyResult.stderr.trim()}`
+        );
+      }
+      const head = verifyResult.stdout.trim();
+      if (head !== target.ref) {
+        throw new Error(
+          `docker checkout verification failed: expected ${target.ref} but found ${head || "<empty>"}`
+        );
+      }
+    }
+  }
+  /** Copy a local file or directory into a running container. */
+  async copyToContainer(containerId, localPath, containerPath) {
+    const result = await this.executor.exec(
+      ["docker", "cp", localPath, `${containerId}:${containerPath}`],
+      { timeoutMs: 6e4 }
+    );
+    if (result.exitCode !== 0) {
+      throw new Error(`docker cp failed (exit ${result.exitCode}): ${result.stderr.trim()}`);
+    }
+  }
+  /**
+   * Execute a command inside a running container.
+   * If stdin is provided, it is piped via `docker exec -i`.
+   */
+  async execInContainer(options) {
+    const { containerId, command, timeoutMs, stdin } = options;
+    const argv = ["docker", "exec"];
+    if (stdin !== void 0) {
+      argv.push("-i");
+    }
+    argv.push(containerId, ...command);
+    return this.executor.exec(argv, {
+      timeoutMs: timeoutMs ?? this.timeoutMs,
+      stdin
+    });
+  }
+  /** Force-remove a container (always succeeds, even if container doesn't exist). */
+  async removeContainer(containerId) {
+    try {
+      await this.executor.exec(["docker", "rm", "-f", containerId], {
+        timeoutMs: 3e4
+      });
+    } catch {
+    }
+  }
+  /** Full lifecycle: create → start → exec → cleanup. Convenience for single-command grading. */
+  async runGraderInContainer(options) {
+    const containerId = await this.createContainer();
+    try {
+      await this.startContainer(containerId);
+      await this.resetContainerCheckout(containerId, options.repoCheckouts);
+      if (options.copyFiles) {
+        for (const file of options.copyFiles) {
+          await this.copyToContainer(containerId, file.localPath, file.containerPath);
+        }
+      }
+      return await this.execInContainer({
+        containerId,
+        command: options.command,
+        stdin: options.stdin
+      });
+    } finally {
+      await this.removeContainer(containerId);
+    }
+  }
+};
+
+export {
+  DefaultCommandExecutor,
+  DockerWorkspaceProvider
+};
+//# sourceMappingURL=chunk-BPGJ4HBU.js.map

package/dist/chunk-BPGJ4HBU.js.map

@@ -0,0 +1 @@
+{"version":3,"sources":["../../../packages/core/src/evaluation/workspace/docker-workspace.ts"],"sourcesContent":["/**\n * Docker workspace provider — manages Docker container lifecycle for eval grading.\n *\n * Flow: pull image → create container → copy files in → exec grader → parse output → destroy container.\n * All Docker commands use `execFile` (no shell) for security.\n *\n * To add a new Docker command: add a method that calls `this.exec(...)` with the appropriate argv.\n *\n * Design decisions:\n * - CommandExecutor interface for testability (mock `execFile` in tests)\n * - Always `docker rm -f` in cleanup, even on errors (try/finally)\n * - Lazy-loaded: non-Docker evals never import this module\n */\n\nimport type { DockerWorkspaceConfig } from '../types.js';\nimport type { RepoCheckoutTarget } from './repo-checkout.js';\n\n/** Result of a command execution */\nexport interface ExecResult {\n  readonly stdout: string;\n  readonly stderr: string;\n  readonly exitCode: number;\n}\n\n/** Abstraction over process execution for testability */\nexport interface CommandExecutor {\n  exec(\n    argv: readonly string[],\n    options?: { timeoutMs?: number; stdin?: string },\n  ): Promise<ExecResult>;\n}\n\n/**\n * Default command executor using Bun.spawn / Node child_process.\n * Mirrors the pattern in runtime/exec.ts.\n */\nexport class DefaultCommandExecutor implements CommandExecutor {\n  async exec(\n    argv: readonly string[],\n    options: { timeoutMs?: number; stdin?: string } = {},\n  ): Promise<ExecResult> {\n    const { execFileWithStdin } = await import('../../runtime/exec.js');\n    return execFileWithStdin(argv, options.stdin ?? '', {\n      timeoutMs: options.timeoutMs,\n    });\n  }\n}\n\n/** Options for creating a Docker container */\nexport interface CreateContainerOptions {\n  readonly image: string;\n  readonly memory?: string;\n  readonly cpus?: number;\n}\n\n/** Options for executing a command inside a container */\nexport interface ExecInContainerOptions {\n  readonly containerId: string;\n  readonly command: readonly string[];\n  readonly timeoutMs?: number;\n  readonly stdin?: string;\n}\n\nconst DEFAULT_TIMEOUT_S = 1800;\n\nfunction buildGitCommand(\n  target: RepoCheckoutTarget | undefined,\n  args: readonly string[],\n): string[] {\n  if (!target?.path) {\n    return ['git', ...args];\n  }\n  return ['git', '-C', target.path, ...args];\n}\n\n/**\n * Manages Docker container lifecycle for workspace-based evaluations.\n *\n * Usage:\n *   const docker = new DockerWorkspaceProvider(config);\n *   await docker.pullImage();\n *   const containerId = await docker.createContainer();\n *   try {\n *     await docker.copyToContainer(containerId, localPath, containerPath);\n *     const output = await docker.execInContainer({ containerId, command: [...] });\n *     // parse output...\n *   } finally {\n *     await docker.removeContainer(containerId);\n *   }\n */\nexport class DockerWorkspaceProvider {\n  private readonly config: DockerWorkspaceConfig;\n  private readonly executor: CommandExecutor;\n  private readonly timeoutMs: number;\n\n  constructor(config: DockerWorkspaceConfig, executor?: CommandExecutor) {\n    this.config = config;\n    this.executor = executor ?? new DefaultCommandExecutor();\n    this.timeoutMs = (config.timeout ?? DEFAULT_TIMEOUT_S) * 1000;\n  }\n\n  /** Check whether the Docker CLI is available on the host. */\n  async isDockerAvailable(): Promise<boolean> {\n    try {\n      const result = await this.executor.exec(\n        ['docker', 'version', '--format', '{{.Server.Version}}'],\n        {\n          timeoutMs: 10_000,\n        },\n      );\n      return result.exitCode === 0;\n    } catch {\n      return false;\n    }\n  }\n\n  /** Pull the configured Docker image. No-op if already cached locally. */\n  async pullImage(): Promise<void> {\n    // Skip pull if image already exists locally (e.g. locally-built images)\n    const inspectResult = await this.executor.exec(\n      ['docker', 'image', 'inspect', this.config.image],\n      {\n        timeoutMs: 10_000,\n      },\n    );\n    if (inspectResult.exitCode === 0) {\n      return; // Image exists locally, no pull needed\n    }\n\n    const result = await this.executor.exec(['docker', 'pull', this.config.image], {\n      timeoutMs: this.timeoutMs,\n    });\n    if (result.exitCode !== 0) {\n      throw new Error(`docker pull failed (exit ${result.exitCode}): ${result.stderr.trim()}`);\n    }\n  }\n\n  /** Create a stopped container from the configured image with resource limits. Returns container ID. */\n  async createContainer(): Promise<string> {\n    const argv: string[] = ['docker', 'create'];\n\n    if (this.config.memory) {\n      argv.push(`--memory=${this.config.memory}`);\n    }\n    if (this.config.cpus !== undefined) {\n      argv.push(`--cpus=${this.config.cpus}`);\n    }\n\n    // Keep the container alive with a long sleep so we can exec into it\n    argv.push(this.config.image, 'sleep', 'infinity');\n\n    const result = await this.executor.exec(argv, { timeoutMs: 30_000 });\n    if (result.exitCode !== 0) {\n      throw new Error(`docker create failed (exit ${result.exitCode}): ${result.stderr.trim()}`);\n    }\n    return result.stdout.trim();\n  }\n\n  /** Start a previously created container. */\n  async startContainer(containerId: string): Promise<void> {\n    const result = await this.executor.exec(['docker', 'start', containerId], {\n      timeoutMs: 30_000,\n    });\n    if (result.exitCode !== 0) {\n      throw new Error(`docker start failed (exit ${result.exitCode}): ${result.stderr.trim()}`);\n    }\n  }\n\n  /**\n   * Reset the container checkout to the specified target refs, if any.\n   * This is used for SWE-bench images where the repo state must match the\n   * dataset's base snapshot before grading begins.\n   */\n  async resetContainerCheckout(\n    containerId: string,\n    repoCheckouts?: readonly RepoCheckoutTarget[],\n  ): Promise<void> {\n    if (!repoCheckouts || repoCheckouts.length === 0) {\n      return;\n    }\n\n    for (const target of repoCheckouts) {\n      const resetResult = await this.execInContainer({\n        containerId,\n        command: buildGitCommand(target, ['reset', '--hard', target.ref]),\n      });\n      if (resetResult.exitCode !== 0) {\n        throw new Error(\n          `docker git reset failed (exit ${resetResult.exitCode}): ${resetResult.stderr.trim()}`,\n        );\n      }\n\n      const verifyResult = await this.execInContainer({\n        containerId,\n        command: buildGitCommand(target, ['rev-parse', 'HEAD']),\n        timeoutMs: 30_000,\n      });\n      if (verifyResult.exitCode !== 0) {\n        throw new Error(\n          `docker checkout verification failed (exit ${verifyResult.exitCode}): ${verifyResult.stderr.trim()}`,\n        );\n      }\n\n      const head = verifyResult.stdout.trim();\n      if (head !== target.ref) {\n        throw new Error(\n          `docker checkout verification failed: expected ${target.ref} but found ${head || '<empty>'}`,\n        );\n      }\n    }\n  }\n\n  /** Copy a local file or directory into a running container. */\n  async copyToContainer(\n    containerId: string,\n    localPath: string,\n    containerPath: string,\n  ): Promise<void> {\n    const result = await this.executor.exec(\n      ['docker', 'cp', localPath, `${containerId}:${containerPath}`],\n      { timeoutMs: 60_000 },\n    );\n    if (result.exitCode !== 0) {\n      throw new Error(`docker cp failed (exit ${result.exitCode}): ${result.stderr.trim()}`);\n    }\n  }\n\n  /**\n   * Execute a command inside a running container.\n   * If stdin is provided, it is piped via `docker exec -i`.\n   */\n  async execInContainer(options: ExecInContainerOptions): Promise<ExecResult> {\n    const { containerId, command, timeoutMs, stdin } = options;\n    const argv: string[] = ['docker', 'exec'];\n\n    if (stdin !== undefined) {\n      argv.push('-i');\n    }\n\n    argv.push(containerId, ...command);\n\n    return this.executor.exec(argv, {\n      timeoutMs: timeoutMs ?? this.timeoutMs,\n      stdin,\n    });\n  }\n\n  /** Force-remove a container (always succeeds, even if container doesn't exist). */\n  async removeContainer(containerId: string): Promise<void> {\n    try {\n      await this.executor.exec(['docker', 'rm', '-f', containerId], {\n        timeoutMs: 30_000,\n      });\n    } catch {\n      // Best-effort cleanup — don't throw on removal failure\n    }\n  }\n\n  /** Full lifecycle: create → start → exec → cleanup. Convenience for single-command grading. */\n  async runGraderInContainer(options: {\n    readonly command: readonly string[];\n    readonly stdin?: string;\n    readonly copyFiles?: ReadonlyArray<{ localPath: string; containerPath: string }>;\n    readonly repoCheckouts?: readonly RepoCheckoutTarget[];\n  }): Promise<ExecResult> {\n    const containerId = await this.createContainer();\n    try {\n      await this.startContainer(containerId);\n      await this.resetContainerCheckout(containerId, options.repoCheckouts);\n\n      if (options.copyFiles) {\n        for (const file of options.copyFiles) {\n          await this.copyToContainer(containerId, file.localPath, file.containerPath);\n        }\n      }\n\n      return await this.execInContainer({\n        containerId,\n        command: options.command,\n        stdin: options.stdin,\n      });\n    } finally {\n      await this.removeContainer(containerId);\n    }\n  }\n}\n"],"mappings":";;;AAoCO,IAAM,yBAAN,MAAwD;EAC7D,MAAM,KACJ,MACA,UAAkD,CAAC,GAC9B;AACrB,UAAM,EAAE,kBAAkB,IAAI,MAAM,OAAO,6BAAuB;AAClE,WAAO,kBAAkB,MAAM,QAAQ,SAAS,IAAI;MAClD,WAAW,QAAQ;IACrB,CAAC;EACH;AACF;AAiBA,IAAM,oBAAoB;AAE1B,SAAS,gBACP,QACA,MACU;AACV,MAAI,CAAC,QAAQ,MAAM;AACjB,WAAO,CAAC,OAAO,GAAG,IAAI;EACxB;AACA,SAAO,CAAC,OAAO,MAAM,OAAO,MAAM,GAAG,IAAI;AAC3C;AAiBO,IAAM,0BAAN,MAA8B;EAClB;EACA;EACA;EAEjB,YAAY,QAA+B,UAA4B;AACrE,SAAK,SAAS;AACd,SAAK,WAAW,YAAY,IAAI,uBAAuB;AACvD,SAAK,aAAa,OAAO,WAAW,qBAAqB;EAC3D;;EAGA,MAAM,oBAAsC;AAC1C,QAAI;AACF,YAAM,SAAS,MAAM,KAAK,SAAS;QACjC,CAAC,UAAU,WAAW,YAAY,qBAAqB;QACvD;UACE,WAAW;QACb;MACF;AACA,aAAO,OAAO,aAAa;IAC7B,QAAQ;AACN,aAAO;IACT;EACF;;EAGA,MAAM,YAA2B;AAE/B,UAAM,gBAAgB,MAAM,KAAK,SAAS;MACxC,CAAC,UAAU,SAAS,WAAW,KAAK,OAAO,KAAK;MAChD;QACE,WAAW;MACb;IACF;AACA,QAAI,cAAc,aAAa,GAAG;AAChC;IACF;AAEA,UAAM,SAAS,MAAM,KAAK,SAAS,KAAK,CAAC,UAAU,QAAQ,KAAK,OAAO,KAAK,GAAG;MAC7E,WAAW,KAAK;IAClB,CAAC;AACD,QAAI,OAAO,aAAa,GAAG;AACzB,YAAM,IAAI,MAAM,4BAA4B,OAAO,QAAQ,MAAM,OAAO,OAAO,KAAK,CAAC,EAAE;IACzF;EACF;;EAGA,MAAM,kBAAmC;AACvC,UAAM,OAAiB,CAAC,UAAU,QAAQ;AAE1C,QAAI,KAAK,OAAO,QAAQ;AACtB,WAAK,KAAK,YAAY,KAAK,OAAO,MAAM,EAAE;IAC5C;AACA,QAAI,KAAK,OAAO,SAAS,QAAW;AAClC,WAAK,KAAK,UAAU,KAAK,OAAO,IAAI,EAAE;IACxC;AAGA,SAAK,KAAK,KAAK,OAAO,OAAO,SAAS,UAAU;AAEhD,UAAM,SAAS,MAAM,KAAK,SAAS,KAAK,MAAM,EAAE,WAAW,IAAO,CAAC;AACnE,QAAI,OAAO,aAAa,GAAG;AACzB,YAAM,IAAI,MAAM,8BAA8B,OAAO,QAAQ,MAAM,OAAO,OAAO,KAAK,CAAC,EAAE;IAC3F;AACA,WAAO,OAAO,OAAO,KAAK;EAC5B;;EAGA,MAAM,eAAe,aAAoC;AACvD,UAAM,SAAS,MAAM,KAAK,SAAS,KAAK,CAAC,UAAU,SAAS,WAAW,GAAG;MACxE,WAAW;IACb,CAAC;AACD,QAAI,OAAO,aAAa,GAAG;AACzB,YAAM,IAAI,MAAM,6BAA6B,OAAO,QAAQ,MAAM,OAAO,OAAO,KAAK,CAAC,EAAE;IAC1F;EACF;;;;;;EAOA,MAAM,uBACJ,aACA,eACe;AACf,QAAI,CAAC,iBAAiB,cAAc,WAAW,GAAG;AAChD;IACF;AAEA,eAAW,UAAU,eAAe;AAClC,YAAM,cAAc,MAAM,KAAK,gBAAgB;QAC7C;QACA,SAAS,gBAAgB,QAAQ,CAAC,SAAS,UAAU,OAAO,GAAG,CAAC;MAClE,CAAC;AACD,UAAI,YAAY,aAAa,GAAG;AAC9B,cAAM,IAAI;UACR,iCAAiC,YAAY,QAAQ,MAAM,YAAY,OAAO,KAAK,CAAC;QACtF;MACF;AAEA,YAAM,eAAe,MAAM,KAAK,gBAAgB;QAC9C;QACA,SAAS,gBAAgB,QAAQ,CAAC,aAAa,MAAM,CAAC;QACtD,WAAW;MACb,CAAC;AACD,UAAI,aAAa,aAAa,GAAG;AAC/B,cAAM,IAAI;UACR,6CAA6C,aAAa,QAAQ,MAAM,aAAa,OAAO,KAAK,CAAC;QACpG;MACF;AAEA,YAAM,OAAO,aAAa,OAAO,KAAK;AACtC,UAAI,SAAS,OAAO,KAAK;AACvB,cAAM,IAAI;UACR,iDAAiD,OAAO,GAAG,cAAc,QAAQ,SAAS;QAC5F;MACF;IACF;EACF;;EAGA,MAAM,gBACJ,aACA,WACA,eACe;AACf,UAAM,SAAS,MAAM,KAAK,SAAS;MACjC,CAAC,UAAU,MAAM,WAAW,GAAG,WAAW,IAAI,aAAa,EAAE;MAC7D,EAAE,WAAW,IAAO;IACtB;AACA,QAAI,OAAO,aAAa,GAAG;AACzB,YAAM,IAAI,MAAM,0BAA0B,OAAO,QAAQ,MAAM,OAAO,OAAO,KAAK,CAAC,EAAE;IACvF;EACF;;;;;EAMA,MAAM,gBAAgB,SAAsD;AAC1E,UAAM,EAAE,aAAa,SAAS,WAAW,MAAM,IAAI;AACnD,UAAM,OAAiB,CAAC,UAAU,MAAM;AAExC,QAAI,UAAU,QAAW;AACvB,WAAK,KAAK,IAAI;IAChB;AAEA,SAAK,KAAK,aAAa,GAAG,OAAO;AAEjC,WAAO,KAAK,SAAS,KAAK,MAAM;MAC9B,WAAW,aAAa,KAAK;MAC7B;IACF,CAAC;EACH;;EAGA,MAAM,gBAAgB,aAAoC;AACxD,QAAI;AACF,YAAM,KAAK,SAAS,KAAK,CAAC,UAAU,MAAM,MAAM,WAAW,GAAG;QAC5D,WAAW;MACb,CAAC;IACH,QAAQ;IAER;EACF;;EAGA,MAAM,qBAAqB,SAKH;AACtB,UAAM,cAAc,MAAM,KAAK,gBAAgB;AAC/C,QAAI;AACF,YAAM,KAAK,eAAe,WAAW;AACrC,YAAM,KAAK,uBAAuB,aAAa,QAAQ,aAAa;AAEpE,UAAI,QAAQ,WAAW;AACrB,mBAAW,QAAQ,QAAQ,WAAW;AACpC,gBAAM,KAAK,gBAAgB,aAAa,KAAK,WAAW,KAAK,aAAa;QAC5E;MACF;AAEA,aAAO,MAAM,KAAK,gBAAgB;QAChC;QACA,SAAS,QAAQ;QACjB,OAAO,QAAQ;MACjB,CAAC;IACH,UAAA;AACE,YAAM,KAAK,gBAAgB,WAAW;IACxC;EACF;AACF;","names":[]}