Function summarizeCodexEvents

plugins/plugin-eval/src/core/benchmark-events.js:180–233

(events)

}

export function summarizeCodexEvents(events) {
  const errorMessages = [];
  const toolNames = [];
  const shellCommands = [];
  const failedShellCommands = [];
  let threadId = null;
  let turnFailed = false;

  for (const event of events) {
    if (!threadId && typeof event.thread_id === "string") {
      threadId = event.thread_id;
    }

    const errorMessage = collectErrorMessage(event);
    if (errorMessage) {
      errorMessages.push(errorMessage);
    }

    if (isToolEvent(event)) {
      toolNames.push(extractToolName(event) || event.type || "tool");
    }

    if (isShellEvent(event)) {
      const command = extractShellCommand(event) || event.type || "shell";
      shellCommands.push(command);
      if (isFailureEvent(event)) {
        failedShellCommands.push(command);
      }
    }

    if (event.type === "turn.failed") {
      turnFailed = true;
    }
  }

  const usage = extractLatestUsage(events);
  const completed = events.some((event) => event.type === "turn.completed");
  const failed = turnFailed || events.some((event) => event.type === "turn.cancelled");

  return {
    threadId,
    eventCount: events.length,
    ignoredLineCount: 0,
    finalStatus: failed ? "failed" : completed ? "completed" : "unknown",
    errorMessages,
    toolCallCount: toolNames.length,
    shellCommandCount: shellCommands.length,
    failedShellCommandCount: failedShellCommands.length,
    toolNames,
    shellCommands,
    usage,
    usageAvailability: usage ? "present" : "unavailable",
  };
}