| name | terminus2_commands |
| description | Parse Terminus2 trajectories at command granularity: one step per entry of each episode response's commands array, matching the CodeTraceBench annotation convention. Outranks the seed terminus2 skill, whose episode-level steps do not align with the annotations.
|
| fingerprints | ["agent-logs/episode-*/response.txt"] |
| priority | 60 |
| metadata | {"version":"1.0","source":"agent-eval codetracebench-oht2-20260801"} |
Terminus2 Command-Level Parser
Directory Layout
trial_dir/
results.json # instruction, verifier outcome
commands.txt
agent-logs/
episode-0/
prompt.txt # context; holds "New Terminal Output:" tail
response.txt # JSON: analysis/plan + commands[]
debug.json
episode-1/ ...
sessions/ panes/ # recordings and final panes (not steps)
Step Convention
- One step per
commands[] entry across episodes, in episode order then
array order. Episodes with unparsable response.txt or no commands
contribute no steps.
- The action is the command's raw
keystrokes; keystrokes that are blank
after trimming (bare Enter, empty send) render as their JSON literal so
the step survives non-empty action gates losslessly.
- Terminal output exists per episode, not per command: the next episode's
prompt tail after "New Terminal Output:" attaches to the episode's final
command; earlier commands carry a null observation.
- All steps of one episode share
parallel_group = episode number.
- The episode's
analysis/plan text becomes thinking on its first step.
This enumeration reproduces the CodeTraceBench annotation step_count on
220 of 222 verified Terminus2 rows at dataset revision
aa213b84ffb6690fc37ca15766d6ca174ec36d4d; the other 2 archives ship empty
trial directories (no agent-logs anywhere).