miniswe

name	miniswe
description	Parse MiniSWE agent trajectories. MiniSWE stores data in one of three formats: agent-logs/mini.traj.json (JSON with messages array), sessions/agent.log (structured text log), or commands.txt (keystroke lists).
fingerprints	["agent-logs/mini.traj.json","sessions/agent.log","commands.txt"]
priority	10
metadata	{"version":"1.0","source":"seed"}

MiniSWE Trajectory Parser

Directory Layout

Primary: traj.json

run_dir/
  results.json
  agent-logs/
    mini.traj.json     # {messages: [{role, content, timestamp}, ...]}

Alternative: agent.log

run_dir/
  sessions/
    agent.log          # Structured text with ```bash blocks and <returncode>

Fallback: commands.txt

run_dir/
  commands.txt         # Each line is a Python list literal: ["cmd", "Enter"]

Step Extraction Logic

• traj.json: Assistant messages with ```bash blocks are actions; next user message with <returncode> is observation
• agent.log: Scan for ```bash fences, pair with following <returncode> lines
• commands.txt: Each line is ast.literal_eval to get keystroke list, last element must be "Enter"

Format Fingerprint

Presence of agent-logs/mini.traj.json, sessions/agent.log, or commands.txt.