NVIDIA-NeMo / NVIDIA-NeMo/Gym

[checkpoint] Restore Harbor agent jobs

Open
#3,280 0 comments 0 reactions 0 assignees View on GitHub

Nobody has claimed this yet.

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

Description

Parent

Child of #3222 and #3024. Depends on #3275.

Runtime state

responses_api_agents/harbor_agent/ and responses_api_agents/harbor_agent_general/ owns Ray task references, Harbor job/environment state, trial directory, stable job identity, and terminal result artifacts.

Adapter requirements

Persist the logical execution-to-runtime mapping, provider/version and physical generation, sandbox snapshot/reference, image/spec and mount fingerprints, agent or verifier cursor, and result-manifest identity. Rebind a replacement runtime atomically and fence the old handle. Never claim that a filesystem-only copy restores unsupported live process state.

Acceptance criteria

  • Restore reattaches or consumes the existing trial without launching a second job; running, partial, and terminal phases are distinct.
  • Mutate runtime state, commit, replace both application and runtime process where supported, restore, and compare the next operation/result.
  • Stale handles and delayed old-runtime results are rejected.
  • Partial restore activates no execution mapping.
  • Duplicate retirement and failed-provider cleanup release the restored runtime exactly once.

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/harbor_agent/ and responses_api_agents/harbor_agent_general/, then read the parent issues #3222 and #3024 and dependency #3275. Trace how Ray task references, Harbor state, trial directories, runtime identity, and terminal artifacts are currently managed. Done means restore distinguishes running, partial, and terminal phases, rejects stale results, and releases restored runtimes exactly once without launching duplicate jobs.

Written by the indexing model from the issue text.

Assessment

Tech stack
python
Domain
backend, distributed-systems
Issue type
Feature
Difficulty
5/5
Estimated time
Over a week
Activity status
Active
Clarity
Mostly clear
Newbie friendliness
35/100

Get new issues in your inbox

A short digest of beginner-friendly GitHub issues.