Repository navigation
fix(nooa): preserve native evidence and add conformance probes - #4813
Conversation
|
Auto-sync is disabled for draft pull requests in this repository. Workflows must be run manually. Contributors can view more details about this message here. |
init-nikhil
left a comment
There was a problem hiding this comment.
CODEX SELF REVIEW
Reviewed head: 441e779bc57f5c49558b8edd1378ffbfd0c92e79
Actual PR base: 634921ea21d8221c6e62461193841cccc37d1e7d
The official Codex review engine (codex-cli 0.154.0, fresh codex exec review --base process, read-only sandbox, inherited account/model configuration, GYM_DEV_REVIEW_CHILD=1) returned no actionable findings.
The engine ran 32 focused tests successfully: 26 in tests/unit_tests/harness_capabilities/test_behavior.py and 6 in scripts/harness_conformance/tests/test_ci_install.py. Full NOOA runtime validation was unavailable because the pinned dependencies were not installed in the read-only review environment. This result does not guarantee correctness or complete the remaining workflow validation.
Run evidence: gymcheck-nooa-conformance-20261020T225915Z; engine exit status 0; completed 2026-10-21 00:22:35 UTC. The checkout remained clean.
- Final-output SHA-256:
b87a7ff402fda7d3c88c3efb4250a587496cd95dcbdb63156f1048427f8a37ea - Raw engine-output SHA-256:
5dde3cb05bd04c025329c63ff4d496729a979858ad39c93c0fe9540d4839d247
|
/ok to test 4e9dd08 |
|
/ok to test f627fdb |
init-nikhil
left a comment
There was a problem hiding this comment.
CODEX SELF REVIEW
Reviewed head: f627fdbfce09d85a91c55378c4b45ad8157bd56c
Actual PR base: 634921ea21d8221c6e62461193841cccc37d1e7d
The official Codex review engine (codex-cli 0.154.0, fresh codex exec review --base process, read-only sandbox, inherited account/model configuration, GYM_DEV_REVIEW_CHILD=1) returned no actionable findings for the full 21-file PR.
All 32 focused behavior and installation tests passed. Full NOOA integration validation was unavailable because the installed NOOA version lacks required APIs. This result does not guarantee correctness or complete the remaining workflow validation.
Run evidence: gymcheck-nooa-conformance-20261020T225915Z, reviewer-repair2; engine exit status 0; completed 2026-10-21 01:45:38 UTC. The checkout remained clean.
- Final-output SHA-256:
767587babe54e152320cc47479761d90169dc990ce7993d5bf31db1abce0bb37 - Raw engine-output SHA-256:
4720ebb8e0b7f3125bfcb2fd143f70e1b08adbc7a82045bf15795807fe64f856
|
/ok to test 2b7f9b4 |
init-nikhil
left a comment
There was a problem hiding this comment.
CODEX SELF REVIEW
Reviewed head 2b7f9b446f6679df2ce97379511548965a625bf8 against actual stacked base 4c632f32c37e52485dfc002a985823ed144b16fe with the official Codex CLI 0.154.0 (gpt-6-astra, fresh read-only review process). No actionable findings were returned.
The engine performed static source inspection; it did not execute tests. Local, live-model, and CI validation are tracked separately. This is not a readiness or correctness guarantee.
|
/ok to test 46d9db6 |
init-nikhil
left a comment
There was a problem hiding this comment.
CODEX SELF REVIEW
Reviewed head 46d9db6a3049710394741593c87013751ebd839e against actual stacked base c5e26e893cec11c6e9709832e1e6b52cc2533585 with official Codex CLI 0.154.0 (gpt-6-astra, fresh read-only review process).
- [P2] Preserve previously captured Python feedback across rerenders —
responses_api_agents/nooa_agent/observability.py:271-273. A differently truncated rerender can replace previously captured feedback even when the next model call is stopped by the policy-call budget. Projection can then discard that unsent replacement and lose feedback that reached an earlier request. Independent reproduction is in progress.
The engine performed static source inspection; it did not execute tests. Model, benchmark and CI validation are tracked separately.
|
/ok to test 5f3e148 |
|
/ok to test eba0827 |
eba0827 to
5efa36d
Compare
5efa36d to
231ff41
Compare
231ff41 to
d8b273e
Compare
|
/ok to test d8b273e |
|
@NVIDIA-NeMo/automation can someone review this? need codeowner approval for the workflow change that runs nooa conformance in a container. |
|
/ok to test bca323f |
|
/ok to test e69e63a |
|
/ok to test 1975665 |
Signed-off-by: Nikhil Deshmudre <nikhil.deshmudre@gmail.com>
Signed-off-by: Nikhil Deshmudre <nikhil.deshmudre@gmail.com>
Signed-off-by: Nikhil Deshmudre <nikhil.deshmudre@gmail.com>
|
/ok to test 13f0f9f |
|
🔄 Merge queue validation started! You can track the progress here: https://github.com/NVIDIA-NeMo/Gym/actions/runs/42134907028 |
NOOA now retains native tool identities, authoritative history, model-visible feedback, and terminal provider outcomes through its sandboxed worker. Shared harness conformance exercises its production borrowed-sandbox path and saves independent execution evidence.
Terminal model rejections retain their numeric HTTP status and follow the existing default/fail/retry policy. Worker, transport, resources, and cleanup failures remain ungradable. Shared grading policy is unchanged.
The conformance workflow runs NOOA in a disposable Linux Python container. Runtime installation follows main's locked
uv syncsetup. Runtime provenance hashespyproject.tomlanduv.lock; it no longer reads the removedrequirements.txtthat previously prevented all NOOA scenarios from starting.Validation on final head
13f0f9f2d9563b1c07eee683877cae60f061a912:harness-conformance-nooa-42118673574-1artifact contains the report, saved model calls, and independent witnesses. This is not a claim of complete P1 conformance.fd1fa273, including the dependency-fingerprint regression and current session lifecycle naming.Main supplied the OpenClaw integration-test cleanup-budget fix that addressed the repeated CI timing failure; the duplicate local timing change was dropped. One added NOOA test assertion follows main's renamed fenced-session collection. No production timeout was changed by this PR.
The five lower stack PRs have merged, and this PR is based on main. The workflow change is unchanged from the version previously reviewed by automation. Fresh current-head code-owner approval remains the landing gate. Earlier benchmark evidence is historical and is not a readiness claim for this head.