({
taskId,
jobToken,
logger,
}: {
taskId: number
jobToken: string | null
logger?: Logger
})
| 11 | import { runTaskInVscode } from "./runTaskInVscode" |
| 12 | |
| 13 | export const processTask = async ({ |
| 14 | taskId, |
| 15 | jobToken, |
| 16 | logger, |
| 17 | }: { |
| 18 | taskId: number |
| 19 | jobToken: string | null |
| 20 | logger?: Logger |
| 21 | }) => { |
| 22 | const task = await findTask(taskId) |
| 23 | const { language, exercise } = task |
| 24 | const run = await findRun(task.runId) |
| 25 | await registerRunner({ runId: run.id, taskId, timeoutSeconds: (run.timeout || 5) * 60 }) |
| 26 | |
| 27 | const containerized = isDockerContainer() |
| 28 | |
| 29 | logger = |
| 30 | logger || |
| 31 | new Logger({ |
| 32 | logDir: containerized ? `/var/log/evals/runs/${run.id}` : `/tmp/evals/runs/${run.id}`, |
| 33 | filename: `${language}-${exercise}.log`, |
| 34 | tag: getTag("runTask", { run, task }), |
| 35 | }) |
| 36 | |
| 37 | try { |
| 38 | const publish = async (e: TaskEvent) => { |
| 39 | const redis = await redisClient() |
| 40 | await redis.publish(getPubSubKey(run.id), JSON.stringify(e)) |
| 41 | } |
| 42 | |
| 43 | const executionMethod = run.executionMethod || "vscode" |
| 44 | logger.info(`running task ${task.id} (${language}/${exercise}) via ${executionMethod}...`) |
| 45 | |
| 46 | if (executionMethod === "cli") { |
| 47 | await runTaskWithCli({ run, task, jobToken, publish, logger }) |
| 48 | } else { |
| 49 | await runTaskInVscode({ run, task, jobToken, publish, logger }) |
| 50 | } |
| 51 | |
| 52 | logger.info(`testing task ${task.id} (${language}/${exercise})...`) |
| 53 | const passed = await runUnitTest({ task, logger }) |
| 54 | |
| 55 | logger.info(`task ${task.id} (${language}/${exercise}) -> ${passed}`) |
| 56 | await updateTask(task.id, { passed }) |
| 57 | |
| 58 | await publish({ |
| 59 | eventName: passed ? RooCodeEventName.EvalPass : RooCodeEventName.EvalFail, |
| 60 | taskId: task.id, |
| 61 | }) |
| 62 | } finally { |
| 63 | await deregisterRunner({ runId: run.id, taskId }) |
| 64 | } |
| 65 | } |
| 66 | |
| 67 | export const processTaskInContainer = async ({ |
| 68 | taskId, |
no test coverage detected