Download scripts/show_stream.py from ZJUSCL/TwinLab-Bench: direct link, hf CLI and curl.
- Browser
- Download file 3.2 kB
-
https://huggingface.co/datasets/ZJUSCL/TwinLab-Bench/resolve/main/scripts/show_stream.py
- Command line
-
hf download hf://datasets/ZJUSCL/TwinLab-Bench/scripts/show_stream.py
-
curl -L -o show_stream.py https://huggingface.co/datasets/ZJUSCL/TwinLab-Bench/resolve/main/scripts/show_stream.py
3.2 kB
| """Print a compact view of one actor stream: tool calls, tool results (truncated), and text. | |
| python scripts/show_stream.py runs/.../run_0/actor_stream_r0_0.jsonl [max_chars] | |
| The stream format (Claude Code or Codex) is taken from the run's run_config.json. | |
| """ | |
| import json | |
| import sys | |
| from pathlib import Path | |
| DEFAULT_MAX_CHARS = 600 | |
| def show_claude_code(path: Path, n: int): | |
| for line in path.read_text().splitlines(): | |
| m = json.loads(line) | |
| t = m["type"] | |
| if t == "system" and m.get("subtype") == "init": | |
| print(f"[init] model={m.get('model')} tools={m.get('tools')} mcp={m.get('mcp_servers')} " | |
| f"skills={m.get('skills')} slash={m.get('slash_commands')}") | |
| elif t in ("assistant", "user"): | |
| content = m["message"]["content"] | |
| if isinstance(content, str): | |
| print(f"[{t} text] {content[:n]}") | |
| continue | |
| for c in content: | |
| if c["type"] == "text": | |
| print(f"[{t} text] {c['text'][:n]}") | |
| elif c["type"] == "tool_use": | |
| print(f"[tool_use {c['name']}] {json.dumps(c['input'])[:n]}") | |
| elif c["type"] == "tool_result": | |
| cc = c["content"] | |
| if isinstance(cc, list): | |
| cc = " ".join(x.get("text", f"<{x['type']}>") for x in cc) | |
| print(f"[tool_result err={c.get('is_error')}] {str(cc)[:n]}") | |
| elif t == "result": | |
| print(f"[result] turns={m.get('num_turns')} cost={m.get('total_cost_usd')} err={m.get('is_error')}") | |
| def show_codex(path: Path, n: int): | |
| for line in path.read_text().splitlines(): | |
| m = json.loads(line) | |
| t = m["type"] | |
| if t == "thread.started": | |
| print(f"[thread] {m['thread_id']}") | |
| elif t == "turn.completed": | |
| print(f"[turn completed] usage={json.dumps(m['usage'])}") | |
| elif t == "item.completed": | |
| it = m["item"] | |
| if it["type"] in ("agent_message", "reasoning"): | |
| print(f"[{it['type']}] {it.get('text', '')[:n]}") | |
| elif it["type"] == "command_execution": | |
| print(f"[shell rc={it['exit_code']}] {it['command'][:n]}\n -> {it['aggregated_output'][:n]}") | |
| elif it["type"] == "mcp_tool_call": | |
| res = it["error"] if it["error"] is not None else " ".join( | |
| x.get("text", f"<{x['type']}>") for x in it["result"]["content"]) | |
| print(f"[mcp {it['server']}.{it['tool']}] {json.dumps(it['arguments'])[:n]}\n -> {str(res)[:n]}") | |
| else: | |
| print(f"[{it['type']}] {json.dumps(it)[:n]}") | |
| elif t in ("error", "turn.failed"): | |
| print(f"[{t}] {json.dumps(m)[:n]}") | |
| def main(): | |
| path = Path(sys.argv[1]) | |
| n = int(sys.argv[2]) if len(sys.argv) > 2 else DEFAULT_MAX_CHARS | |
| cfg = path.parent / "run_config.json" | |
| assert cfg.is_file(), f"no run_config.json next to {path}; cannot tell the stream format" | |
| agent = json.loads(cfg.read_text())["actor_agent"] | |
| {"claude-code": show_claude_code, "codex": show_codex}[agent](path, n) | |
| if __name__ == "__main__": | |
| main() | |