feat: transcript parsing based on vercel-labs/agent-eval

This commit is contained in:
Matt Rossman
2026-03-03 17:43:26 -05:00
parent 18f58ffc7c
commit cdd65c5256
3 changed files with 390 additions and 84 deletions

1
.gitignore vendored
View File

@@ -1,6 +1,7 @@
node_modules/
dist/
*.log
evals/output/
.DS_Store
.env

View File

@@ -1,63 +1,7 @@
import { spawnSync } from "node:child_process";
import { mkdtempSync, readdirSync, readFileSync } from "node:fs";
import os from "node:os";
import { mkdirSync, readdirSync, readFileSync } from "node:fs";
import path from "node:path";
import { z } from "zod";
/**
* Transcript schemas
*/
const TextBlock = z.object({ type: z.literal("text"), text: z.string() });
const ThinkingBlock = z.object({
type: z.literal("thinking"),
thinking: z.string(),
});
const ToolUseBlock = z.object({
type: z.literal("tool_use"),
id: z.string(),
name: z.string(),
input: z.record(z.string(), z.unknown()),
});
const ToolResultBlock = z.object({
type: z.literal("tool_result"),
tool_use_id: z.string(),
content: z.unknown(),
is_error: z.boolean().optional(),
});
const ContentBlock = z.union([
ToolUseBlock,
ThinkingBlock,
TextBlock,
ToolResultBlock,
z.looseObject({ type: z.string() }), // catch-all for unknown block types
]);
const AssistantEntry = z.object({
type: z.literal("assistant"),
sessionId: z.string(),
timestamp: z.string(),
uuid: z.string(),
message: z.object({
role: z.literal("assistant"),
content: z.array(ContentBlock),
stop_reason: z.string().nullable().optional(),
}),
});
// Catch-all — user messages, queue-operations, etc.
const TranscriptLine = z.union([
AssistantEntry,
z.looseObject({ type: z.string() }),
]);
/**
* Config
*/
import { parseClaudeCodeTranscript } from "./parse-transcript";
const apiKey = process.env.ANTHROPIC_API_KEY;
if (!apiKey) throw new Error("ANTHROPIC_API_KEY required");
@@ -75,23 +19,23 @@ SELECT * FROM orders WHERE user_id = 123 AND status = 'pending';
What indexes should I add and why?`;
/**
* Run the eval
*/
// Mount ~/.claude/projects to capture the built-in session transcript
const projectsDir = mkdtempSync(path.join(os.tmpdir(), "eval-projects-"));
// Mount ~/.claude/projects to capture the built-in session transcript.
// Written to a fixed path so you can inspect it after the run (e.g. in VSCode).
const projectsDir = path.join(repoRoot, "evals", "output");
mkdirSync(projectsDir, { recursive: true });
const result = spawnSync(
"docker",
[
"run",
"--rm",
"-e",
"--workdir",
"/eval",
"--env",
`ANTHROPIC_API_KEY=${apiKey}`,
"-v",
"--volume",
`${skillPath}:/home/claude/.claude/skills/supabase-postgres-best-practices:ro`, // :ro = read-only snapshot
"-v",
"--volume",
`${projectsDir}:/home/claude/.claude/projects`,
"evals-claude",
"claude",
@@ -106,26 +50,14 @@ if (result.status !== 0) {
throw new Error(result.stderr || `Exit code ${result.status}`);
}
/**
* Parse the transcript
*/
// Container's working dir is /, which becomes `-` in the projects path
const transcriptDir = path.join(projectsDir, "-");
// Container's working dir is /eval, which becomes `eval` in the projects path
const transcriptDir = path.join(projectsDir, "eval");
const [transcriptFile] = readdirSync(transcriptDir).filter((f) =>
f.endsWith(".jsonl"),
);
// Single typed array — all transcript entries parsed and validated
const transcript = readFileSync(
path.join(transcriptDir, transcriptFile),
"utf-8",
)
.split("\n")
.filter(Boolean)
.flatMap((l) => {
const parsed = TranscriptLine.safeParse(JSON.parse(l));
return parsed.success ? [parsed.data] : [];
});
const transcript = parseClaudeCodeTranscript(
readFileSync(path.join(transcriptDir, transcriptFile), "utf-8"),
);
console.log(JSON.stringify(transcript, null, 2));

373
evals/parse-transcript.ts Normal file
View File

@@ -0,0 +1,373 @@
/**
* Parser for Claude Code transcript format.
* Claude Code stores transcripts as JSONL at ~/.claude/projects/{path}/{session}.jsonl
*
* Format reference (based on Claude Code CLI output):
* - Messages have type: "user" | "assistant"
* - Tool use appears in assistant messages with tool_use blocks
* - Tool results appear as separate messages with type: "tool_result"
*
* Adapted from https://github.com/vercel-labs/agent-eval
*/
/** Canonical tool names. */
export type ToolName =
| "file_read"
| "file_write"
| "file_edit"
| "shell"
| "web_fetch"
| "web_search"
| "glob"
| "grep"
| "list_dir"
| "agent_task"
| "skill"
| "unknown";
/** A single event in the parsed transcript. */
export interface TranscriptEvent {
/** ISO timestamp of the event */
timestamp?: string;
/** Event type */
type: "message" | "tool_call" | "tool_result" | "thinking" | "error";
/** For message events: the role */
role?: "user" | "assistant" | "system";
/** Text content (for messages, thinking, errors) */
content?: string;
/** For tool_call and tool_result events */
tool?: {
name: ToolName;
originalName: string;
args?: Record<string, unknown>;
result?: unknown;
durationMs?: number;
success?: boolean;
};
/** Raw event data from the agent (for debugging) */
raw?: unknown;
}
function normalizeToolName(name: string): ToolName {
const toolMap: Record<string, ToolName> = {
// File operations
Read: "file_read",
read_file: "file_read",
ReadFile: "file_read",
Write: "file_write",
write_file: "file_write",
WriteFile: "file_write",
write_to_file: "file_write",
Edit: "file_edit",
edit_file: "file_edit",
EditFile: "file_edit",
str_replace_editor: "file_edit",
StrReplace: "file_edit",
// Shell
Bash: "shell",
bash: "shell",
Shell: "shell",
shell: "shell",
execute_command: "shell",
run_command: "shell",
// Web
WebFetch: "web_fetch",
web_fetch: "web_fetch",
fetch_url: "web_fetch",
mcp__fetch__fetch: "web_fetch",
WebSearch: "web_search",
web_search: "web_search",
// Search/navigation
Glob: "glob",
glob: "glob",
list_files: "glob",
Grep: "grep",
grep: "grep",
search_files: "grep",
LS: "list_dir",
list_dir: "list_dir",
ListDir: "list_dir",
// Agent/subagent tools
Task: "agent_task",
task: "agent_task",
// Skills
Skill: "skill",
skill: "skill",
};
return toolMap[name] || "unknown";
}
function extractFilePath(args: Record<string, unknown>): string | undefined {
return (args.path || args.file_path || args.filename || args.file) as
| string
| undefined;
}
function extractUrl(args: Record<string, unknown>): string | undefined {
return (args.url || args.uri || args.href) as string | undefined;
}
function extractCommand(args: Record<string, unknown>): string | undefined {
if (typeof args.command === "string") return args.command;
if (Array.isArray(args.command)) return args.command.join(" ");
if (typeof args.cmd === "string") return args.cmd;
return undefined;
}
/**
* Get the content array from data, handling nested message format.
* Claude Code wraps messages: { type: "assistant", message: { content: [...] } }
*/
function getContentArray(data: Record<string, unknown>): unknown[] | undefined {
if (Array.isArray(data.content)) return data.content;
const message = data.message as Record<string, unknown> | undefined;
if (message && Array.isArray(message.content)) return message.content;
return undefined;
}
function getStringContent(data: Record<string, unknown>): string | undefined {
if (typeof data.content === "string") return data.content;
const message = data.message as Record<string, unknown> | undefined;
if (message && typeof message.content === "string") return message.content;
return undefined;
}
function extractContent(data: Record<string, unknown>): string | undefined {
const stringContent = getStringContent(data);
if (stringContent) return stringContent;
const contentArray = getContentArray(data);
if (contentArray) {
const textBlocks = contentArray.filter(
(block: unknown) =>
(block as Record<string, unknown>).type === "text",
);
if (textBlocks.length > 0) {
return textBlocks
.map((b: unknown) => (b as Record<string, unknown>).text)
.join("\n");
}
}
if (typeof data.text === "string") return data.text;
return undefined;
}
function extractToolUses(
data: Record<string, unknown>,
): Array<{ name: string; input?: Record<string, unknown>; args?: Record<string, unknown> }> {
const toolUses: Array<{
name: string;
input?: Record<string, unknown>;
args?: Record<string, unknown>;
}> = [];
const contentArray = getContentArray(data);
if (contentArray) {
for (const block of contentArray) {
const b = block as Record<string, unknown>;
if (b.type === "tool_use") {
toolUses.push({
name: b.name as string,
input: b.input as Record<string, unknown> | undefined,
});
}
}
}
// Also handle OpenAI-style tool_calls array
const toolCalls =
data.tool_calls ||
(data.message as Record<string, unknown>)?.tool_calls;
if (Array.isArray(toolCalls)) {
for (const call of toolCalls) {
const c = call as Record<string, unknown>;
const func = c.function as Record<string, unknown> | undefined;
toolUses.push({
name: (func?.name || c.name) as string,
args: func?.arguments
? JSON.parse(func.arguments as string)
: ((c.arguments || c.input) as
| Record<string, unknown>
| undefined),
});
}
}
return toolUses;
}
function extractThinking(data: Record<string, unknown>): string | undefined {
const contentArray = getContentArray(data);
if (contentArray) {
const thinkingBlocks = contentArray.filter(
(block: unknown) =>
(block as Record<string, unknown>).type === "thinking",
);
if (thinkingBlocks.length > 0) {
return thinkingBlocks
.map((b: unknown) => {
const block = b as Record<string, unknown>;
return block.thinking || block.text;
})
.join("\n");
}
}
return undefined;
}
function parseClaudeCodeLine(line: string): TranscriptEvent[] {
const events: TranscriptEvent[] = [];
try {
const data = JSON.parse(line);
if (data.type === "user" || data.role === "user") {
const contentArray = getContentArray(data);
const toolResults = contentArray?.filter(
(block: unknown) =>
(block as Record<string, unknown>).type === "tool_result",
);
if (toolResults && toolResults.length > 0) {
for (const result of toolResults) {
const r = result as Record<string, unknown>;
events.push({
timestamp: data.timestamp,
type: "tool_result",
tool: {
name: "unknown",
originalName: (r.tool_use_id || "unknown") as string,
result: r.content,
success: !r.is_error && !r.error,
},
raw: r,
});
}
} else {
events.push({
timestamp: data.timestamp,
type: "message",
role: "user",
content: extractContent(data),
raw: data,
});
}
} else if (data.type === "assistant" || data.role === "assistant") {
const content = extractContent(data);
if (content) {
events.push({
timestamp: data.timestamp,
type: "message",
role: "assistant",
content,
raw: data,
});
}
for (const toolUse of extractToolUses(data)) {
events.push({
timestamp: data.timestamp,
type: "tool_call",
tool: {
name: normalizeToolName(toolUse.name),
originalName: toolUse.name,
args: toolUse.input || toolUse.args || {},
},
raw: toolUse,
});
}
const thinking = extractThinking(data);
if (thinking) {
events.push({
timestamp: data.timestamp,
type: "thinking",
content: thinking,
raw: data,
});
}
} else if (
data.type === "tool_result" ||
data.type === "tool_response"
) {
events.push({
timestamp: data.timestamp,
type: "tool_result",
tool: {
name: "unknown",
originalName: data.tool_use_id || "unknown",
result: data.content || data.output || data.result,
success: !data.is_error && !data.error,
},
raw: data,
});
} else if (data.type === "system" || data.role === "system") {
events.push({
timestamp: data.timestamp,
type: "message",
role: "system",
content: extractContent(data),
raw: data,
});
} else if (data.type === "error" || data.error) {
events.push({
timestamp: data.timestamp,
type: "error",
content:
data.error?.message ||
data.message ||
JSON.stringify(data.error),
raw: data,
});
}
} catch {
// Skip unparseable lines
}
return events;
}
/** Parse a Claude Code JSONL transcript into a flat list of events. */
export function parseClaudeCodeTranscript(raw: string): {
events: TranscriptEvent[];
errors: string[];
} {
const events: TranscriptEvent[] = [];
const errors: string[] = [];
for (const line of raw.split("\n").filter((l) => l.trim())) {
try {
events.push(...parseClaudeCodeLine(line));
} catch (e) {
errors.push(
`Failed to parse line: ${e instanceof Error ? e.message : String(e)}`,
);
}
}
// Post-process: extract metadata from tool args
for (const event of events) {
if (event.type === "tool_call" && event.tool) {
const args = event.tool.args || {};
if (["file_read", "file_write", "file_edit"].includes(event.tool.name)) {
const filePath = extractFilePath(args);
if (filePath) event.tool.args = { ...args, _extractedPath: filePath };
}
if (event.tool.name === "web_fetch") {
const url = extractUrl(args);
if (url) event.tool.args = { ...args, _extractedUrl: url };
}
if (event.tool.name === "shell") {
const command = extractCommand(args);
if (command) event.tool.args = { ...args, _extractedCommand: command };
}
}
}
return { events, errors };
}