Target: NVIDIA/NemoClaw
Field reports: #44 and #45
Pull requests:
#7254 #7406
Result: both upstream accepted and merged
This field note covers two Scarab-originated patches accepted upstream in NVIDIA NemoClaw.
There were three NemoClaw PRs in this field run.
Two merged.
One was closed without merge because upstream work superseded the patch.
Zero were left dangling.
That matters.
PR accounting should distinguish between rejected work, accepted work, superseded work, and stale work left for maintainers to deal with. A superseded PR is not a failed repair. It is an upstream condition resolving before the proposed patch lands.
The point is not to keep every PR alive as a trophy.
The point is to help the repository get to the right state.
These two merged patches show two different forms of repair:
- a test-contract repair around onboarding policy resume behavior;
- a runtime diagnostic repair around sandbox lifecycle readiness.
One was small.
One was substantial.
Both were bounded.
Both were accepted upstream.
Field Lab record: NemoClaw #6042
This field test started with a macOS onboarding regression report.
The reported symptom was that the interactive onboard wizard skipped the Policy Presets TUI step after Sandbox Name. The issue described NemoClaw v0.0.70 on macOS, with repeated failures showing missing policy preset and sandbox creation steps.
On the surface, that looks like a wizard-flow bug.
The accepted patch was narrower.
It targeted the policy resume contract.
The useful boundary was:
recorded policy selection -> required preset reconciliation -> resume decision
The important question was not simply:
“Did the wizard render a screen?”
It was:
“Which policy preset state is the onboarding flow allowed to treat as already applied?”
An empty preset selection should not be treated as an applied policy preset.
And if Slack later becomes required after an empty recorded selection, the system needs to add its required preset and request reconciliation.
The patch added regression coverage for that existing behavior.
It verified that:
- an empty preset selection is not treated as applied;
- enabling Slack adds its required preset and requests reconciliation;
- tests based on impossible production states are removed.
Production behavior was unchanged.
That was deliberate.
This was not a broad onboarding rewrite. It was not a claim to resolve every macOS onboarding condition. It was a test-contract repair around one boundary involved in the larger issue.
Tests are part of the repository’s truth surface.
If tests encode impossible states, they can make the system appear to support behavior it does not actually produce. Repairing that test boundary is still repair.
NVIDIA/NemoClaw#7254 was merged on July 25, 2026.
The PR title was:
test(onboard): cover empty policy resume contracts
This field report claims upstream acceptance of the test-contract repair.
It does not claim that the full onboarding issue is closed, that every Policy Presets TUI path is fixed, or that NVIDIA maintainers endorsed Scarab.
The claim is narrower:
Scarab identified a policy-resume selection boundary in NVIDIA NemoClaw onboarding where tests needed to reflect the real production contract for empty selections and required preset reconciliation, and a human-submitted regression coverage patch was accepted upstream.
Field Lab record: NemoClaw #7387
This field test started with a lifecycle diagnostics problem in NemoClaw’s Brev sandbox path.
The visible symptom was subtle:
a sandbox could look healthy right now while still missing durable lifecycle metadata required later.
A runtime could be live.
A gateway could respond.
A sandbox could appear usable.
And the registered lifecycle contract could still be incomplete.
That is the boundary.
The broken assumption was:
runtime healthy = lifecycle ready
That assumption was too broad.
Runtime health and lifecycle readiness are different claims.
Runtime health asks:
“Can the system run now?”
Lifecycle readiness asks:
“Does the registered system contain the durable metadata, provenance, driver, image, dashboard, gateway, and recovery state needed for later lifecycle operations?”
A system can answer yes to the first question and no to the second.
Before this repair, current diagnostics could attest the live runtime without making the missing lifecycle-registration metadata visible early enough.
That meant the operator might only discover the problem after starting snapshot, rebuild, recovery, or another lifecycle operation.
The patch added a read-only doctor check for registered sandbox lifecycle metadata.
The diagnostic warns when a registry entry is incomplete even if runtime health probes are otherwise readable.
It reports missing or invalid metadata field names.
It does not return registry values.
It does not print captured runtime output.
It does not expose credentials or sensitive values.
It keeps runtime-health evidence separate from lifecycle-readiness evidence.
That separation is the repair.
The patch did not try to eagerly recover, rebuild, mutate, or guess.
It made the incomplete lifecycle contract visible before a later operation depended on it.
That matters because health checks can become false reassurance.
“Everything is reachable” is not the same as “the system is ready for lifecycle operations.”
A diagnostic command should not collapse those into one green result.
NVIDIA/NemoClaw#7406 was merged on July 25, 2026.
The PR title was:
fix(doctor): warn on incomplete lifecycle registration
The patch included lifecycle-registration diagnostics, redacted warning behavior, docs, and focused tests.
This field report claims upstream acceptance of the lifecycle-registration diagnostic repair.
It does not claim that every Brev lifecycle issue is fixed, that every fast-path registration path is corrected, or that NVIDIA maintainers endorsed Scarab.
The claim is narrower:
Scarab identified a lifecycle-registration readiness boundary in NVIDIA NemoClaw where current runtime health could obscure missing durable metadata required by later sandbox lifecycle operations, and a human-submitted repair was accepted upstream to make that condition visible through read-only doctor diagnostics.
Combined field result
Together, these two field reports show two different repair shapes.
The first repaired a test-contract boundary.
The second repaired a runtime diagnostic boundary.
One was small.
One was substantial.
Both were accepted upstream.
That is useful because software repair is not one shape.
Sometimes the right patch is production code.
Sometimes it is regression coverage.
Sometimes it is diagnostics.
Sometimes it is documentation.
Sometimes it is closing a PR because upstream already resolved the condition.
The common thread is not patch size.
The common thread is boundary discipline.
A PR should not be the search process.
A PR should be the receipt.
Scarab diagnosed the boundaries.
Codex prepared the patches.
NemoClaw accepted the changes.
Disclosure
This field report was prepared with AI-assisted editing from public field-test notes, public issue and PR records, and the public Field Lab record. The diagnostic claims, repair boundaries, and final wording were human reviewed.
Scarab Diagnostic Suite is proprietary. The Field Lab publishes public case records, issue links, validation summaries, and claim boundaries only.
SDS finds evidence. People make claims. Maintainers decide.
Top comments (0)