Repository navigation
fix(examples): retain attributed Arena failures with zero reward - #1724
Merged
Merged
Conversation
Le8r0nJames
requested review from
CormickKneey,
HwVanICI,
PrometheusComing,
garrett4wade and
sitabulaixizawaluduo
as code owners
September 17, 2026 07:12
Integrate current main without rewriting published history.
Collaborator
|
LGTM |
sitabulaixizawaluduo
approved these changes
Sep 18, 2026
This was referenced Sep 22, 2026
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Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be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.
Description
Follow-up to #1702 (Arena single- and multi-stream rollouts).
Recognize structured GameAgent harness outcome codes when classifying failed
Arena episodes. Allowlisted HARNESS_FAILED outcomes with existing interactions
can use the existing zero-reward recovery/export path instead of being rejected
as unknown failures. Keep the existing interaction-count and proxy-system-error
guards, legacy failure handling, and unrelated-status rejection.
Also provide a validated identity_reward transform for configurations that need
the original Arena score without threshold scaling. This does not change the
default reward transform or fix nullable rewards in export statistics.
Review follow-up (2026-09-18)
One precedence boundary remains: an explicit structured non-model outcome can still match the legacy Claude error envelope and be classified as a model failure. This is inherited legacy fallback behavior interacting with the new structured-code path; its policy needs clarification before treating structured codes as authoritative across all fallbacks.
Related Issue
Follow-up to #1702.
No separate issue.
Type of Change
Checklist
pre-commit run --all-files)./docs/build_all.sh)main/review-prcommand/create-prBreaking Change Details (if applicable): None.
Additional Context
This is an independent change based on main, not the experimental runtime stack.
It reuses a minimal subset of previously developed coverage; no new test scenarios
were introduced in this update.
Validation:
pytest -q tests/test_swe_arena_multi_stream.py tests/experimental/openai/test_context_overflow_recovery.py— 75 passed in the target image. Redundant identity/export coverage was removed; the new classification coverage is reduced to 69 lines. Production code is unchanged.Changed-file pre-commit
hooks pass. Full-tree hooks were attempted but hit local Python 3.10 environment
limitations (missing tomllib and the installed Torch lacking CPUOffloadPolicy).
No full distributed end-to-end regression is claimed for this rebased branch.