Repository navigation
fix(miniswe): restore bounded observations and length-aware recovery - #3624
Merged
Merged
Conversation
lafrancef
added
bug
Contributor
Author
|
/ok to test 28920b8 |
Use the pinned mini-SWE observation template and preserve timeout feedback. Forward output limits to recovery prompts and distinguish exhausted context from malformed tool calls while retaining partial work for verification. Signed-off-by: Frederic Lafrance <flafrance@nvidia.com>
lafrancef
force-pushed
the
flafrance/miniswe-bounded-observations
branch
from
September 22, 2026 18:27
28920b8 to
40c9715
Compare
Contributor
Author
|
/ok to test 40c9715 |
bxyu-nvidia
approved these changes
Sep 22, 2026
bxyu-nvidia
marked this pull request as ready for review
September 22, 2026 18:34
lafrancef
enabled auto-merge
September 22, 2026 18:35
Contributor
|
🔄 Merge queue validation started! You can track the progress here: https://github.com/NVIDIA-NeMo/Gym/actions/runs/35769065832 |
This branch was previously deployed
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
Large shell outputs were replayed without mini-SWE’s default cap, and context/output limits could appear as repeated tool-call format failures. Restore the pinned mini-SWE observation template (5,000-character head and tail), retain full output in trajectories, and include command-timeout feedback.
Pass output truncation into mini-SWE’s recovery prompt and distinguish
ContextWindowExceeded/OutputTokenLimitExceededfrom malformed calls, allowing partial work to reach verification. To enable explicit context-overflow handling with the vLLM adapter, setpolicy_model.responses_api_models.vllm_model.propagate_context_overflow_errors=trueas documented in the agent README. This changes model-visible observations and may affect benchmark results.Real Qwen3.8 Flash Next TB4 rollouts at the pre-rebase commit
28920b828c8e56f0be8db693ef2fcae05eadc774confirm bounded observations, agent-to-verifier completion, length-aware recovery, and three propagated context-overflow terminations followed by completed verification. The full 66-task comparison is still running; keep this PR as a draft until the results are reviewed. RepeatedOutputTokenLimitExceededtermination has not yet been observed in the live run.Local check limitation:
pre-commit run --all-filescould not start because this Jujutsu workspace has no.gitdirectory; the full hook suite remains pending. The commit includes DCO sign-off.Rebased onto
mainate86746aa. The overlapping length-aware recovery change keeps upstream’s"stop"finish reason for non-truncated responses. The ongoing rollout retains its original commit pin; no fresh full benchmark has been run on the rebased head.No separate issue is needed for this focused correction to the existing mini-SWE integration.