openhands-sessions

name	openhands_sessions
description	Parse OpenHands session-event trajectories at the CodeTraceBench annotation granularity: one step per action event with a cause-paired observation (run, run_ipython, read, edit, think, recall). Outranks the seed openhands skill, which counts only run/run_ipython and undercounts.
fingerprints	["sessions/sessions","sessions/*.json"]
priority	45
metadata	{"version":"1.0","source":"agent-eval codetracebench-oht2-20260801"}

OpenHands Session-Event Parser

Directory Layout

trial_dir/
  results.json
  sessions/
    <session-id>.json            # flat event export (preferred view)
    sessions/
      <session-id>/
        events/0.json, 1.json    # per-event files (second view)
        event_cache/0-25.json    # batched shards (last view)
    agent.cast  agent.log        # recordings (not steps)

Step Convention

• Events come from the first available view: flat export, then events/, then event_cache; events sort by id.
• One step per action event whose id some observation event names as its cause — every action type counts, including read, edit, think, and recall. Actions without a paired observation (system, user message, finish, cut-off trailing actions) are not steps.
• run/run_ipython actions render as the raw command or code; think renders as its thought; other actions render as action(args) without the thought field. Blank commands render as their JSON literal so the step survives non-empty action gates.
• The action's args.thought becomes thinking; the observation's content is the observation.
• The first user message event is the task description, falling back to results.json instruction.

This enumeration reproduces the CodeTraceBench annotation step_count on 183 of 199 verified terminal-bench OpenHands rows at dataset revision aa213b84ffb6690fc37ca15766d6ca174ec36d4d. The other 16 fail loudly on the step-count gate: their published views (condensed or partial session streams) disagree with the annotated count.