Files
cline/evals/diff_editing/TestRunner.ts
T
2025-06-10 17:11:06 -07:00

392 lines
13 KiB
TypeScript

import { runSingleEvaluation, TestInput, TestResult } from "./ClineWrapper"
import { basicSystemPrompt } from "./prompts/basicSystemPrompt-06-06-25"
import { claude4SystemPrompt } from "./prompts/claude4SystemPrompt-06-06-25"
import { formatResponse } from "./helpers"
import { Anthropic } from "@anthropic-ai/sdk"
import * as fs from "fs"
import * as path from "path"
import { Command } from "commander"
import { InputMessage, ProcessedTestCase, TestCase, TestConfig, SystemPromptDetails, ConstructSystemPromptFn } from "./types"
function log(isVerbose: boolean, message: string) {
if (isVerbose) {
console.log(message)
}
}
const systemPromptGeneratorLookup: Record<string, ConstructSystemPromptFn> = {
basicSystemPrompt: basicSystemPrompt,
claude4SystemPrompt: claude4SystemPrompt,
}
type TestResultSet = { [test_id: string]: (TestResult & { test_id?: string })[] }
class NodeTestRunner {
private apiKey: string | undefined
constructor(isReplay: boolean) {
if (!isReplay) {
this.apiKey = process.env.OPENROUTER_API_KEY
if (!this.apiKey) {
throw new Error("OPENROUTER_API_KEY environment variable not set for a non-replay run.")
}
}
}
/**
* convert our messages array into a properly formatted Anthropic messages array
*/
transformMessages(messages: InputMessage[]): Anthropic.Messages.MessageParam[] {
return messages.map((msg) => {
// Use TextBlockParam here for constructing the input message
const content: (Anthropic.TextBlockParam | Anthropic.ImageBlockParam)[] = []
if (msg.text) {
// This object now correctly matches the TextBlockParam type
content.push({ type: "text", text: msg.text })
}
if (msg.images && Array.isArray(msg.images)) {
const imageBlocks = formatResponse.imageBlocks(msg.images)
content.push(...imageBlocks)
}
return {
role: msg.role,
content: content,
}
})
}
/**
* Generate the system prompt on the fly
*/
constructSystemPrompt(systemPromptDetails: SystemPromptDetails, systemPromptName: string) {
const systemPromptGenerator = systemPromptGeneratorLookup[systemPromptName]
const { cwd_value, browser_use, width, height, os_value, shell_value, home_value, mcp_string, user_custom_instructions } =
systemPromptDetails
const systemPrompt = systemPromptGenerator(
cwd_value,
browser_use,
width,
height,
os_value,
shell_value,
home_value,
mcp_string,
user_custom_instructions,
)
return systemPrompt
}
/**
* Loads our test cases from a directory of json files
*/
loadTestCases(testDirectoryPath: string): TestCase[] {
const testCasesArray: TestCase[] = []
const dirents = fs.readdirSync(testDirectoryPath, { withFileTypes: true })
for (const dirent of dirents) {
if (dirent.isFile() && dirent.name.endsWith(".json")) {
const testFilePath = path.join(testDirectoryPath, dirent.name)
const fileContent = fs.readFileSync(testFilePath, "utf8")
const testCase: TestCase = JSON.parse(fileContent)
// Use the filename (without extension) as the test_id if not provided
if (!testCase.test_id) {
testCase.test_id = path.parse(dirent.name).name
}
testCasesArray.push(testCase)
}
}
return testCasesArray
}
/**
* Saves the test results to the specified output directory.
*/
saveTestResults(results: TestResultSet, outputPath: string) {
// Ensure output directory exists
if (!fs.existsSync(outputPath)) {
fs.mkdirSync(outputPath, { recursive: true })
}
// Write each test result to its own file
for (const testId in results) {
const outputFilePath = path.join(outputPath, `${testId}.json`)
const testResult = results[testId]
fs.writeFileSync(outputFilePath, JSON.stringify(testResult, null, 2))
}
}
/**
* Run a single test example
*/
async runSingleTest(testCase: ProcessedTestCase, testConfig: TestConfig): Promise<TestResult> {
if (testConfig.replay && !testCase.original_diff_edit_tool_call_message) {
return {
success: false,
error: "missing_original_diff_edit_tool_call_message",
errorString: `Test case ${testCase.test_id} is missing 'original_diff_edit_tool_call_message' for replay.`,
}
}
const customSystemPrompt = this.constructSystemPrompt(testCase.system_prompt_details, testConfig.system_prompt_name)
// messages don't include system prompt and are everything up to the first replace_in_file tool call which results in a diff edit error
const input: TestInput = {
apiKey: this.apiKey,
systemPrompt: customSystemPrompt,
messages: testCase.messages,
modelId: testConfig.model_id,
originalFile: testCase.file_contents,
originalFilePath: testCase.file_path,
parsingFunction: testConfig.parsing_function,
diffEditFunction: testConfig.diff_edit_function,
thinkingBudgetTokens: testConfig.thinking_tokens_budget,
originalDiffEditToolCallMessage: testConfig.replay ? testCase.original_diff_edit_tool_call_message : undefined,
}
return await runSingleEvaluation(input)
}
/**
* Runs all the text examples synchonously
*/
async runAllTests(testCases: ProcessedTestCase[], testConfig: TestConfig, isVerbose: boolean): Promise<TestResultSet> {
const results: TestResultSet = {}
for (const testCase of testCases) {
results[testCase.test_id] = []
log(isVerbose, `-Running test: ${testCase.test_id}`)
for (let i = 0; i < testConfig.number_of_runs; i++) {
const result = await this.runSingleTest(testCase, testConfig)
results[testCase.test_id].push(result)
}
}
return results
}
/**
* Runs all of the text examples asynchronously, with concurrency limit
*/
async runAllTestsParallel(
testCases: ProcessedTestCase[],
testConfig: TestConfig,
isVerbose: boolean,
maxConcurrency: number = 20,
): Promise<TestResultSet> {
const results: TestResultSet = {}
testCases.forEach((tc) => {
results[tc.test_id] = []
})
// Create a flat list of all individual runs we need to execute
const allRuns = testCases.flatMap((testCase) =>
Array(testConfig.number_of_runs)
.fill(null)
.map(() => testCase),
)
for (let i = 0; i < allRuns.length; i += maxConcurrency) {
const batch = allRuns.slice(i, i + maxConcurrency)
const batchPromises = batch.map((testCase) =>
this.runSingleTest(testCase, testConfig).then((result) => ({
...result,
test_id: testCase.test_id,
})),
)
const batchResults = await Promise.all(batchPromises)
// Calculate the total cost for this batch
const batchCost = batchResults.reduce((total, result) => {
return total + (result.streamResult?.usage?.totalCost || 0)
}, 0)
// Populate the results dictionary
for (const result of batchResults) {
if (result.test_id) {
results[result.test_id].push(result)
}
}
const batchNumber = i / maxConcurrency + 1
const totalBatches = Math.ceil(allRuns.length / maxConcurrency)
log(isVerbose, `-Completed batch ${batchNumber} of ${totalBatches}... (Batch Cost: $${batchCost.toFixed(6)})`)
}
return results
}
/**
* Print output of the tests
*/
printSummary(results: TestResultSet, isVerbose: boolean) {
let totalRuns = 0
let totalPasses = 0
let totalInputTokens = 0
let totalOutputTokens = 0
let totalCost = 0
let runsWithUsageData = 0
let totalDiffEditSuccesses = 0
let totalRunsWithToolCalls = 0
const testCaseIds = Object.keys(results)
log(isVerbose, "\n=== TEST SUMMARY ===")
for (const testId of testCaseIds) {
const testResults = results[testId]
const passedCount = testResults.filter((r) => r.success && r.diffEditSuccess).length
const runCount = testResults.length
totalRuns += runCount
totalPasses += passedCount
const runsWithToolCalls = testResults.filter((r) => r.success === true).length
const diffEditSuccesses = passedCount
totalRunsWithToolCalls += runsWithToolCalls
totalDiffEditSuccesses += diffEditSuccesses
// Accumulate token and cost data
for (const result of testResults) {
if (result.streamResult?.usage) {
totalInputTokens += result.streamResult.usage.inputTokens
totalOutputTokens += result.streamResult.usage.outputTokens
totalCost += result.streamResult.usage.totalCost
runsWithUsageData++
}
}
log(isVerbose, `\n--- Test Case: ${testId} ---`)
log(isVerbose, ` Runs: ${runCount}`)
log(isVerbose, ` Passed: ${passedCount}`)
log(isVerbose, ` Success Rate: ${runCount > 0 ? ((passedCount / runCount) * 100).toFixed(1) : "N/A"}%`)
}
log(isVerbose, "\n\n=== OVERALL SUMMARY ===")
log(isVerbose, `Total Test Cases: ${testCaseIds.length}`)
log(isVerbose, `Total Runs Executed: ${totalRuns}`)
log(isVerbose, `Overall Passed: ${totalPasses}`)
log(isVerbose, `Overall Failed: ${totalRuns - totalPasses}`)
log(isVerbose, `Overall Success Rate: ${totalRuns > 0 ? ((totalPasses / totalRuns) * 100).toFixed(1) : "N/A"}%`)
log(isVerbose, "\n\n=== OVERALL DIFF EDIT SUCCESS RATE ===")
if (totalRunsWithToolCalls > 0) {
const diffSuccessRate = (totalDiffEditSuccesses / totalRunsWithToolCalls) * 100
log(isVerbose, `Total Runs with Successful Tool Calls: ${totalRunsWithToolCalls}`)
log(isVerbose, `Total Runs with Successful Diff Edits: ${totalDiffEditSuccesses}`)
log(isVerbose, `Diff Edit Success Rate: ${diffSuccessRate.toFixed(1)}%`)
} else {
log(isVerbose, "No successful tool calls to analyze for diff edit success.")
}
log(isVerbose, "\n\n=== TOKEN & COST ANALYSIS ===")
if (runsWithUsageData > 0) {
log(isVerbose, `Total Input Tokens: ${totalInputTokens.toLocaleString()}`)
log(isVerbose, `Total Output Tokens: ${totalOutputTokens.toLocaleString()}`)
log(isVerbose, `Total Cost: $${totalCost.toFixed(6)}`)
log(isVerbose, "---")
log(
isVerbose,
`Avg Input Tokens / Run: ${(totalInputTokens / runsWithUsageData).toLocaleString(undefined, {
maximumFractionDigits: 0,
})}`,
)
log(
isVerbose,
`Avg Output Tokens / Run: ${(totalOutputTokens / runsWithUsageData).toLocaleString(undefined, {
maximumFractionDigits: 0,
})}`,
)
log(isVerbose, `Avg Cost / Run: $${(totalCost / runsWithUsageData).toFixed(6)}`)
} else {
log(isVerbose, "No usage data available to analyze.")
}
}
}
async function main() {
const program = new Command()
const defaultTestPath = path.join(__dirname, "test_cases")
const defaultOutputPath = path.join(__dirname, "test_outputs")
program
.name("TestRunner")
.description("Run evaluation tests for diff editing")
.version("1.0.0")
.option("--test-path <path>", "Path to the directory containing test case JSON files", defaultTestPath)
.option("--output-path <path>", "Path to the directory to save the test output JSON files", defaultOutputPath)
.option("--model-id <model_id>", "The model ID to use for the test")
.option("--system-prompt-name <name>", "The name of the system prompt to use", "basicSystemPrompt")
.option("-n, --number-of-runs <number>", "Number of times to run each test case", "1")
.option("--parsing-function <name>", "The parsing function to use", "parseAssistantMessageV2")
.option("--diff-edit-function <name>", "The diff editing function to use", "constructNewFileContentV2")
.option("--thinking-budget <tokens>", "Set the thinking tokens budget", "0")
.option("--parallel", "Run tests in parallel", false)
.option("--replay", "Run evaluation from a pre-recorded LLM output, skipping the API call", false)
.option("-v, --verbose", "Enable verbose logging", false)
program.parse(process.argv)
const options = program.opts()
const isVerbose = options.verbose
const testPath = options.testPath
const outputPath = options.outputPath
const testConfig: TestConfig = {
model_id: options.modelId,
system_prompt_name: options.systemPromptName,
number_of_runs: parseInt(options.numberOfRuns, 10),
parsing_function: options.parsingFunction,
diff_edit_function: options.diffEditFunction,
thinking_tokens_budget: parseInt(options.thinkingBudget, 10),
replay: options.replay,
}
try {
const startTime = Date.now()
const runner = new NodeTestRunner(testConfig.replay)
const testCases = runner.loadTestCases(testPath)
const processedTestCases: ProcessedTestCase[] = testCases.map((tc) => ({
...tc,
messages: runner.transformMessages(tc.messages),
}))
log(isVerbose, `-Loaded ${testCases.length} test cases.`)
log(isVerbose, `-Executing ${testConfig.number_of_runs} run(s) per test case.`)
if (testConfig.replay) {
log(isVerbose, `-Running in REPLAY mode. No API calls will be made.`)
}
log(isVerbose, "Starting tests...\n")
const results = options.parallel
? await runner.runAllTestsParallel(processedTestCases, testConfig, isVerbose)
: await runner.runAllTests(processedTestCases, testConfig, isVerbose)
runner.printSummary(results, isVerbose)
const endTime = Date.now()
const durationSeconds = ((endTime - startTime) / 1000).toFixed(2)
log(isVerbose, `\n-Total execution time: ${durationSeconds} seconds`)
runner.saveTestResults(results, outputPath)
} catch (error) {
console.error("\nError running tests:", error)
process.exit(1)
}
}
if (require.main === module) {
main()
}