Function runAgentOnCommit

evals/buffbench/agent-runner.ts:22–160

({
  client,
  agentId,
  commit,
  repoUrl,
  initCommand,
  env,
  localAgentDefinitions,
  printEvents,
  finalCheckCommands,
  externalAgentType,
}: {
  client: CodebuffClient
  agentId: string
  commit: EvalCommitV2
  repoUrl: string
  initCommand?: string
  env?: Record<string, string>
  localAgentDefinitions: any[]
  printEvents: boolean
  finalCheckCommands?: string[]
  externalAgentType?: ExternalAgentType
})

export type ExternalAgentType = 'claude' | 'codex' | 'opencode'

export async function runAgentOnCommit({
  client,
  agentId,
  commit,
  repoUrl,
  initCommand,
  env,
  localAgentDefinitions,
  printEvents,
  finalCheckCommands,
  externalAgentType,
}: {
  client: CodebuffClient
  agentId: string
  commit: EvalCommitV2
  repoUrl: string
  initCommand?: string
  env?: Record<string, string>
  localAgentDefinitions: any[]
  printEvents: boolean
  finalCheckCommands?: string[]
  externalAgentType?: ExternalAgentType
}): Promise<{
  diff: string
  contextFiles: Record<string, string>
  durationMs: number
  cost: number
  error?: string
  trace: AgentStep[]
  finalCheckOutputs?: FinalCheckOutput[]
}> {
  console.log(`[${commit.id}] Running agent ${agentId}...`)
  const startTime = Date.now()
  let diff = ''
  let contextFiles: Record<string, string> = {}
  let error: string | undefined
  let cost = 0
  const trace: AgentStep[] = []
  let finalCheckOutputs: FinalCheckOutput[] | undefined

  try {
    const timeoutMs = 60 * 60 * 1000 // 60 minutes
    await withTimeout(
      withTestRepo(
        {
          repoUrl,
          parentSha: commit.parentSha,
          initCommand,
          env,
        },
        async (repoDir) => {
          // Select the appropriate runner
          let runner: Runner
          if (externalAgentType === 'claude') {
            runner = new ClaudeRunner(repoDir, env)
          } else if (externalAgentType === 'codex') {
            runner = new CodexRunner(repoDir, env)
          } else if (externalAgentType === 'opencode') {