| 179 | } |
| 180 | |
| 181 | interface BenchmarkOptions { |
| 182 | task: string; |
| 183 | systemPrompt: string; |
| 184 | mcpServers: Record<string, { command: string; args: string[] }>; |
| 185 | disallowedTools?: string[]; |
| 186 | } |
| 187 | |
| 188 | async function runAgentBenchmark(options: BenchmarkOptions): Promise<BenchmarkResult> { |
| 189 | const toolUsage: ToolUsage = {}; |
nothing calls this directly
no outgoing calls
no test coverage detected