NVIDIA-NeMo / NVIDIA-NeMo/Gym

[checkpoint] Continue scicode_agent after verified sub-steps

Open
#3,258 0 comments 0 reactions 0 assignees View on GitHub

Nobody has claimed this yet.

area:agent feature
Dominant language
Python
Stars
1.2k
Forks
349
Avg merge
1d 21h
Merged PRs (30d)
318

Description

Parent

Child of #3220 and #3024. No additional resource child.

State owner

responses_api_agents/scicode_agent/app.py owns sub-step cursor, generated scripts/responses, verification feedback, model/resource cookies, usage, and revision.

Adapter requirements

Use the shared whitebox participant and versioned boundary records. Persist output items, usage, cookies, last model-call identity, resource revisions, remaining budgets, and typed agent-specific state. Commit only after all effects of an operation are durable. On attempt N+1, skip reseeding and do not repeat completed model, tool, verifier, or judge work.

Acceptance criteria

  • Restore after each verified sub-step continues with the next problem part without rerunning prior scripts or verification.
  • Prepare/commit/process replacement/restore/resume succeeds at every supported boundary.
  • Saved model responses, usage, cookies, and pending actions are applied exactly once.
  • Checkpoint refusal retries without advancing history or the operation cursor.
  • Resource revision mismatch, stale attempt, duplicate restore, and corrupt state fail closed.
  • Completed-unacknowledged results cannot also be published as resumable work.

Contributor guide

Open the contributing guide

First steps

  1. Read the whole issue, then the project's contributing guide.
  2. Comment on the issue to say you are picking it up — it saves two people doing the same work.
  3. Fork the repository and make your change on a branch.
  4. Open a pull request that references the issue number.

Research direction

Start with responses_api_agents/scicode_agent/app.py and the parent issues #3220 and #3024. Trace the sub-step cursor, generated scripts and responses, verification feedback, cookies, usage, revisions, and checkpoint boundaries described in the issue. Done means every listed acceptance criterion holds, including exact-once resume behavior and fail-closed handling of invalid state.

Written by the indexing model from the issue text.

Assessment

Tech stack
python
Domain
ai, backend
Issue type
Feature
Difficulty
5/5
Estimated time
Over a week
Activity status
Active
Clarity
Mostly clear
Newbie friendliness
25/100

Get new issues in your inbox

A short digest of beginner-friendly GitHub issues.