| name | terminus2 |
| description | Parse Terminus2 agent trajectories. Terminus2 stores each episode as agent-logs/episode-N/response.txt (JSON with commands/analysis/plan) and the next episode's prompt.txt contains the observation (New Terminal Output section).
|
| fingerprints | ["agent-logs/episode-*/response.txt"] |
| priority | 50 |
| metadata | {"version":"1.0","source":"seed"} |
Terminus2 Trajectory Parser
Directory Layout
run_dir/
results.json # Task metadata
agent-logs/
episode-0/response.txt # JSON: {commands, analysis, plan}
episode-1/prompt.txt # Contains "New Terminal Output:" section
episode-1/response.txt
...
Step Extraction Logic
- Each episode-N/response.txt is one action (step_id = N+1)
- Observation comes from episode-(N+1)/prompt.txt
- Commands field contains keystrokes list
- If response.txt is valid JSON, extract commands; otherwise use raw text
Format Fingerprint
Presence of agent-logs/episode-*/response.txt files.