[Ubuntu 26.04][Agent&Skills] NemoHermes smoke reviewer journey exits without writing review.json
- Dominant language
- TypeScript
- Stars
- 22.5k
- Forks
- 3.1k
- Avg merge
- 1d 1h
- Merged PRs (30d)
- 715
Description
## Description
NemoHermes completed the reviewer journey command path but did not create the required review output file. The workflow then failed when downloading `/sandbox/work/hermes-review/review.json` from the sandbox because the file did not exist.
## Platform Scope
Reproduced on Ubuntu 26.04 GPU. A prior smoke run also showed the same missing `review.json` symptom on WSL2 x86; other platforms are not consistently reproduced for this exact symptom.
## Regression
Unknown — earlier versions were not tested for this exact smoke journey.
## Environment
- Product/version under test: NemoClaw v0.0.113
- Platform: Ubuntu 26.04 GPU
- Agent: NemoHermes
- Sandbox: `smoke-hermes-review`
- Cleanup status: preflight cleanup completed successfully before the smoke journey and reported the environment clean and ready.
## Steps To Reproduce
1. Run the NemoHermes reviewer smoke journey on Ubuntu 26.04 GPU with NemoClaw v0.0.113.
2. Onboard the NemoHermes smoke sandbox `smoke-hermes-review`.
3. Create `/sandbox/work/hermes-review` in the sandbox.
4. Upload `fixtures/user-journeys/hermes-review/checkout_retry.diff` and `fixtures/user-journeys/hermes-review/task.md` to `/sandbox/work/hermes-review/`.
5. Execute the Hermes task:
```bash
hermes -z "Read /sandbox/work/hermes-review/task.md and complete it exactly. Write exactly five findings, and do not combine findings. The five separate findings must cover: missing idempotency protection, transaction spanning the retry loop, charge call has no request timeout/deadline because it uses context.Background(), final payment error is swallowed by return nil after retries, and retry policy has no backoff or error filter."
```
6. Download `/sandbox/work/hermes-review/review.json` from the sandbox.
## Expected Result
- NemoHermes writes `/sandbox/work/hermes-review/review.json`.
- The file is non-empty valid JSON with `decision: "REQUEST_CHANGES"`.
- The deterministic verifier exits successfully and prints `PASS`.
## Actual Result
- The Hermes command path returned successfully enough for the workflow to continue to the download step.
- `/sandbox/work/hermes-review/review.json` was missing.
- Download failed with:
```text
Cannot download '/sandbox/work/hermes-review/review.json' from sandbox 'smoke-hermes-review': no such path in the sandbox.
```
## Why This Looks Like A Product Bug
- The fixture task explicitly instructs NemoHermes to write the review to `/sandbox/work/hermes-review/review.json`.
- The documented product expectation requires NemoHermes to write `review.json` with `REQUEST_CHANGES` and no inference or tool error.
- The smoke implementation follows that product-facing workflow and only fails after the expected artifact is absent.
- Cleanup completed successfully before the smoke journey, so the available evidence does not point to dirty runner state.
Contributor guide
Research direction
Start with the NemoHermes reviewer smoke journey and the fixtures at fixtures/user-journeys/hermes-review/checkout_retry.diff and task.md, using the hermes command from the reproduction steps on Ubuntu 26.04. Trace where the workflow should create and download review.json. Done means the file is non-empty valid JSON with decision "REQUEST_CHANGES" and the deterministic verifier prints PASS.
Written by the indexing model from the issue text.
Assessment
- Tech stack
- typescript, ubuntu
- Domain
- testing-qa
- Issue type
- Bug
- Difficulty
- 4/5
- Estimated time
- 3-5 days
- Activity status
- Active
- Clarity
- Mostly clear
- Newbie friendliness
- 52/100