On This Page
Terminology
| Term | Meaning |
|---|---|
| Assistant | opencode, run with your assistant key, for objective facts and tests only; never for judgement. |
| Assistant key | The second key per task, used only by the assistant; separate from the task key. |
| Arm A, Arm B | The two runs of the same task. Arm A is hs arm-a, Arm B is hs arm-b, always in that order. You are not told how they differ. |
| hs | The command line tool. It runs the arms, exports sessions, opens the grading form, validates, and pushes your submission. |
| Task repo | Your GitHub repo for one task, named like boson-task-3. Holds the task and receives your submission. |
| prompt_id | The task id, like CCP11036. Names your pull request. |
| Opening prompt | The first message to the agent. The tool sends it in both arms; you never type it. |
| Interaction mode | Interactive (collaborative) or Autonomous (delegated). Shown on HAI and in task.md. |
| Follow-up | A message you send after the opening prompt in an Interactive task. Same follow-ups at the same points in both arms. |
| Unblocking input | A narrow answer to an agent question. Whether this is allowed in Autonomous tasks is awaiting Ops clarification; retain the no-follow-up rule until confirmed. |
| Session | One run of the agent in Docker. Ends when you press Ctrl-C or after 2 hours 30 minutes. |
| Workspace | The folder the agent works in. hs open <arm> shows it, hs diff <arm> shows what changed. |
| Pinned commit | The exact commit both arms start from. hs claim checks it out. |
| Grading form | The local page at http://127.0.0.1:8765 where you enter ratings, behavioral issues, and the comparison. |
| Ratings | Six 1 to 5 scores per arm, each with a rationale. |
| Behavioral issue | A tagged problem in a run: code, severity, description, evidence. |
| Severity | 0 Observation, 1 Minor, 2 Major, 3 Blocking. |
| Comparison | Four 0 to 7 scores and three comments, entered once both arms are graded. |
| Tie | 3.5 on a comparison question. Only when the arms were truly equal on that question. |
| Verification | The checks you ran yourself, saved with their raw output in verification/armA/ and verification/armB/. |
| Submission | The submissions/ folder in your task repo, pushed to your fork by hs submit --push --pr and sent as a pull request. |
| QC | Quality control: static validation, model-based arm reviews, consistency and evidence checks, and human review of your submission. A merge means accepted. |
| #boson-tasking | The Slack channel for questions. |