[nvidia-nemopatch-writer]

Outcome

Upgrade managed OpenClaw and its official plugins from 2026.9.2 to 2026.10.5 while preserving state restoration, credentials, plugin provenance and sandbox isolation. Resolve conflicts with upstream main 009a4f1612, including the merged Model Router retirement, while retaining rejected-field diagnostics and bounded observation behavior. Prepared for independent review at the author’s request; full runtime qualification remains incomplete, and no merge is authorized.

Reason

Adopt OpenClaw 2026.10.5 and repair lifecycle, messaging and startup incompatibilities found during qualification.

Changes

• Update runtime/plugin locks, archive integrity, managed images, dependency remediation and WeChat 2.4.10 together. Preserve security pins and sender-denial controls.

• Preserve original backups and pinned 2026.9.5/2026.9.7 migration bridges outside the ordinary runtime path. Retain downgrade, mixed-store, archival, ownership, rollback and recovery checks.

• Repair native pairing, client-token publication and shutdown notification handling. Retain native ownership authority, bounded admission retries and startup holds on uncertain state. Keep original operation deadlines.

• Restore a private Node compile cache outside the native home archive and propagate it to workers. Verify plugin roots through guarded native filesystem APIs with pinned implementation/dependency identities. Existing migration, startup and source-root regression tests protect these paths.

• Preserve host-execution evidence and paired SQLite transcript checks for admitted plugin turns. Correct survival-session and llama.cpp fixture authority; keep restore/upgrade fast-test mappings and failure-preserving diagnostics.

• Restore complete JavaScript CodeQL publication after both complementary query partitions finish; the workflow contract tests cover completion states and invalid report refusal. Refresh genuine Pi AMD64/ARM64 receipts with source parity, retain standalone configuration generation, and resolve observer assets directly within the architecture budget.

• Merge upstream installer shim ownership, OpenShell runtime provenance and Model Router retirement. Resolve the ten integration conflicts while preserving upstream receipt authority and the approved assertion allocation. Aggregate assertion limits decrease with the retired router test; no new budget allocation or exception is added.

• Restore the dedicated two-minute native plugin-turn allowance in sandbox survival, with an explicit numeric timeout parameter. Keep the existing readiness-probe, overall test and workflow deadlines.

• Make the NIM clipped-diagnostic fixture refuse readiness and hold its simulated diagnostic through TERM until forced shutdown. Preserve the timeout, redaction and identity assertions.

• Refresh both Pi qualification receipts from the successful 4181fc354a image cohort and match their authority digests. Wait for accepted survival deletion to settle within its existing verification deadline; preserve cleanup ownership until absence is confirmed.

Verification

Latest commit: 3695725be64a4481327134fc05cd1e2f151c91dc, integrating upstream main 009a4f1612d376fad16324ea238e35010fc8bfe9.

• Resolve the single source-shape contract-list conflict by retaining both existing security entries. The numeric source-shape limit remains zero. No check, deadline or coverage requirement was weakened.
• All 173 focused admission, SDK and protected-workspace tests passed in isolation. Source-shape, repository and growth checks passed. The signed DCO merge commit passed normal isolated pre-commit, commit-message and pre-push hooks, including publication validation and CLI TypeScript. GitHub reports Verified, and the PR is mergeable again.
• Source CI passed all 24 applicable jobs, including 12 CLI shards and coverage aggregation. One documentation-only job skipped. Installer hashes and candidate admission also passed.
• CodeQL passed all six jobs. The service results retain 62 JavaScript and 16 Python findings, with zero added or removed results against the preceding candidate. All 201 JavaScript and 172 Python query definitions remain present; ShellCheck reported zero. This does not mean zero security findings.
• Advisor completed all nine specialists. All 27 reports were read. Its blocking delivery finding correctly identifies that JavaScript completion depends on unrelated Python analysis. A two-file repair is prepared locally; 30 existing workflow tests pass. It remains uncommitted while this qualification cycle runs. Advisor approval is not claimed.
• Managed images passed. All eight producer contracts passed the production parser. AMD64, ARM64 and rootless Podman activations each passed 13 phases and 13 cleanup callbacks, with 98, 98 and 100 matching command results. No local image builds occurred. The five named test sandboxes and containers were confirmed absent.
• Image activation uses synthetic inference. The Podman evidence retains nine forced 45-second stops; runner cleanup removed one catatonit and eight conmon processes. Graceful stopping and global process absence remain unproved. Native tool receipts retain replayInvalid=true. These checks do not qualify replay validity or the legacy installer deadline. Deep Agents Code remains NOT hardened.
• At the author’s renewed request, one full default E2E run is running on 3695725be6, using trusted controller 009a4f1612. Its immutable receipt verifies empty selectors. The trusted plan has 78 coverage rows; all 60 catalogue jobs were generated. The installer test remains pending. Staging Brev and explicit-only hardware jobs are excluded.
• That run uses the supported managed_image_revision=4181fc354a input for the previously qualified ancestor images. The CLI and tests use 3695725be6; image contracts retain their original revision and cohort. The separate current-image qualification passed as recorded above.
• Pi passed nine phases and five cleanup callbacks. Three hosted read tasks passed the native event verifier on their first attempts. Fresh and rebuilt native-local turns were authenticated, and no local image builds occurred.
• Sandbox survival passed 11 phases and three cleanup callbacks. Eight marker pairs, native plugin invocation, cold-recovery identities and health checks passed. Final deletion confirmed absence on the first poll.
• The Advisor-required backup/restore test passed both scenarios, 16 phases and two cleanup callbacks. Schema 20, replay and sticker state survived the 9.5/9.7 bridges. Retired Telegram state refused startup and retained the original backup and recovery transaction. Its expected negative rebuild retains an unverified maintenance-abort warning; graceful stopping is not claimed.
• The previous installer failure and all retry history remain recorded below. This renewed request authorized one additional full run; it does not authorize automatic unchanged repeats. Full-suite qualification, publication of the Advisor repair, independent approval and the older disputed P1 adjudication remain incomplete.

Previous candidate ab755477421a45b3689ebfa0133b63174dc13009, based on upstream main 1d518c967cbce498fa93d4771964534bc25f2d66:

• Normal isolated pre-commit, commit-message and pre-push hooks passed, including repository checks, growth budgets, publication validation and CLI TypeScript. The worktree is clean, DCO is present, and GitHub reports the commit as Verified.
• Both Pi receipts come from the successful image run for 4181fc354a. The strict production receipt checker returns qualified; image-source parity and the exact receipt digests remain enforced.
• Survival now observes only the named sandbox's Deleting phase within the existing 60-second verification deadline. It retains the final absence assertion and keeps cleanup armed until that assertion succeeds. Five controlled scenarios using the actual changed source passed, including command failure and deadline exhaustion. No timeout or assertion limit increased.
• Checked-JavaScript TypeScript passed. Selected polling, timeout and cleanup support checks reported 31 passed and one failed: isolation denied an unrelated installer test's account-home temporary directory. That result is retained; no isolation control or test was weakened.
• Source CI passed: 24 successful jobs and one docs-only skip, including all 12 CLI shards and the coverage aggregate.
• Advisor passed all 15 jobs. All nine specialists and 27 reports were read, with zero code findings. The downloaded context and every specialist receipt bind this commit and base. The verification specialist requires the existing state-backup-restore owner; it passed both scenarios in the completed full default suite, together with Pi qualification and sandbox survival. This is source-review evidence, not independent approval or a waiver of the deterministic E2E floor.
• Managed-image qualification passed: 14 successful jobs and four intended skips. All eight current producer contracts passed the production parser. Current AMD64, ARM64 and rootless Podman activations each passed 13 phases and 13 cleanup callbacks, with 98, 98 and 100 matching command results. All five named sandboxes and containers were confirmed absent; no local image builds occurred.
• Current image activation used synthetic inference. Podman retained nine 45-second forced stops, and runner cleanup removed one catatonit and 12 conmon processes. The host-release warning remains recorded. Graceful stopping and global process absence are unproved. Native tool receipts retain replayInvalid=true; these checks do not qualify replay validity, the original installer deadline, or full-suite completion. Deep Agents Code remains NOT hardened.
• At the author’s explicit request, one additional full default manual E2E run completed on ab75547742 with 80 successful jobs, 19 intended skips and two failures: the installer target and the derivative Relevant E2E aggregate. Its immutable dispatch receipt verifies the latest PR commit, trusted 1d518c967c controller and empty selectors. The trusted plan contains 78 coverage rows; all 60 catalogue jobs were generated. The v0.0.123 installer job failed its unchanged 35-minute deadline. Protected runtime passed. The final aggregate reports catalogue-github-read as the only unsuccessful selected workflow group, matching the installer failure.
• This run uses the workflow’s supported managed_image_revision input for qualified ancestor 4181fc354a. All image build inputs are unchanged. The CLI and tests use ab75547742; the image contracts retain their original revision and cohort. This does not qualify the separate current image build. Brev staging and opt-in hardware jobs are excluded.
• Current scoped runtime evidence: Pi qualification passed all nine phases and five cleanup callbacks. Three hosted read tasks passed the existing event verifier on their first attempts; native-local fresh and rebuilt turns were authenticated. No local image build occurred.
• Sandbox survival passed all 11 phases and three cleanup callbacks. Eight state markers, native plugin execution and cold-recovery identities were preserved. Final deletion confirmed absence on the first poll. The live run did not need the Deleting settlement branch; five controlled source scenarios cover that branch.
• The Advisor-required backup/restore owner passed both scenarios, all 16 phases and both cleanup callbacks. Legacy state migrated through 9.5/9.7 with schema 20, replay and sticker data preserved. Retired Telegram state refused startup and retained its original recovery backup. The expected negative rebuild reported an unverified maintenance abort before terminal refusal; graceful stopping and global container absence are not claimed.
• Protected GPU and local inference passed all five phases and four cleanup callbacks. All 39 command results match their stream files. OpenClaw, Hermes and Deep Agents Code each passed GPU-backed Ollama, vLLM and NIM completion, plus bootstrap rollback. The final owned-container and network inventory was clean. The trusted controller verified the candidate OpenShell projection before consumption and recorded its restoration. These protected localhost images are separate from shipped-image activation evidence; Deep Agents Code remains NOT hardened.
• Current installer failure evidence: the first two sandbox rebuilds completed. The third completed both 9.5 and 9.7 bridge passes and its first 10.5 Doctor pass. Its second 10.5 Doctor began and last printed a retained pre-repair capture manifest. The installer was killed at its 35-minute cap; the pending native operation remains unknown. The later preservation and upgraded-credential checks were not reached. All seven registered cleanup callbacks passed.
• The complete 4,071-line failed-job log and 978-line native startup log were reviewed. All 26 source bindings for the 142-predicate v0.0.123 assertion ledger match the prior audited source. No source-supported correction or transient service failure is established. No unchanged installer or full-suite replay is scheduled.
• Previous full-suite failures and retry history remain recorded. The author authorized this additional run; no further unchanged full-suite or installer repeat is scheduled. No current full-suite pass is claimed.

Previous candidate 4181fc354a:

• Source CI failed the stale Pi receipt check. The new commit refreshes both genuine receipts and their authority digests.
• Managed images passed all eight producer contracts and all three AMD64, ARM64 and Podman activation checks. Those results remain evidence for that candidate.
• The full default manual E2E run completed with 78 successful jobs, 19 skipped jobs and three failed targets plus the aggregate. Pi failed source parity; survival observed Deleting after accepted deletion; the v0.0.123 installer exceeded its original 35-minute deadline in the third restored sandbox's 10.5 Doctor. The installer cause remains unresolved. No unchanged exhausted installer rerun was launched.
• Earlier focused checks reported 104 passes, plus two clipped-diagnostic passes. The assertion allocation matches the existing maintainer-approved transition, digest 28b9c8f74a6778ed56e5d4f94474007fb972abef288656b0959aa3fd374bff04. The repair adds no allocation or exception.

The remaining verification results below are historical evidence for 57cabe80bcad53ff885493dce528acffd3186ce4, integrating main 4d5d613f73. They do not qualify the latest commit.

• Combined readiness tests: 164 passed, seven existing skips across three files.
• Runtime-provenance support tests: 58 passed.
• Installer ACP contracts: 31 passed with system Git inside the credential-denying sandbox. The first harness run had four global-config fixture failures and one timeout; its log is retained. No repository timeout or check changed.
• Normal isolated commit and publication hooks: passed. DCO and commit verification: passed; GitHub Verified for the merge and all three included upstream commits.
• The reviewed diff contains no secrets, API keys or credentials. Original coverage, repository budgets and retry counts remain unchanged.

Historical candidate qualification for 57cabe80bc:

• Source CI passed: 51,447 passed, 85 unchanged skips, all 12 shard digests reconciled and CLI coverage ratchet passed.
• CodeQL passed both JavaScript partitions and service publication. All 201 query definitions were retained; 63 JavaScript and 16 Python findings are unchanged, with zero new or missing findings. ShellCheck reported zero. This does not mean zero security findings.
• Podman CPU, portable rootless lifecycle and native Windows tunnel passed their scoped contracts. Windows reported three passes and three native-platform-specific skips.
• Advisor passed all 15 jobs. All nine specialists and 27 reports were read, with zero reported findings. Its base-run context explicitly binds this head and base. The coordinator shadow was read-only and made zero approval writes; it supplies no independent approval or waiver.
• Managed images: all eight immutable producer contracts bind this commit and cohort ghrun-42662064951-1. AMD64, ARM64 and rootless Podman published-image activations each passed 13 phases and 13 cleanup callbacks. Their evidence contains 98, 98 and 100 command results respectively, each with matching stream files, 38 authenticated completion POSTs and zero local image-build commands. Two external-agent configs on each platform match the published configs; managed-agent catalog receipts are asserted by the owning test.

Podman retained nine 45-second forced stops and runner cleanup of one catatonit and eight conmon processes. Final absence was verified for the five named test sandboxes after bounded settlement; graceful stop and global process absence remain unproved. The Ubuntu 26.04 host warning is retained. Image activation uses synthetic inference, and its external rebuild is 10.5 to 10.5. Native child receipts demonstrate successful tools but retain replayInvalid=true; they do not qualify replay validity. Same-agent output files are overwritten during the external-image phase. These passes do not replace the original legacy installer deadline or full DEFAULT79 qualification.

The five-row runtime diagnostic passed concurrency within its unchanged 60-second contract, channel stop/start with seven preserved projections, both storage tests and the v0.0.89 upgrade. The v0.0.123 installer failed its unchanged 35-minute cap. Post-upgrade restoration, stopped-sibling lifecycle, systemd replacement and upgraded credential custody remain unproved. All seven failure-cleanup callbacks passed. The remaining 73-case plan is unexecuted, and original DEFAULT79 obligations remain recorded. No unchanged exhausted retry was launched.

A separate native AMD64 confined timing profile completed all six Doctor passes and 17 bridge subprocess checks: restore took 71.8 seconds, and the workload took 585.9 seconds. Repeated plugin/config capture is expensive; cache flushes totalled 0.9 seconds. This diagnostic does not prove full installer completion or the exclusive cause of its timeout. Both copy comparisons passed 960 assertions, but batch copying was slower, so no copy-path change was made. Original archives and retained resources remain intact; owned diagnostic containers and sandboxes were removed.

Review notes

Independent approval and the older disputed Advisor router P1 adjudication remain required. For the historical 57cabe80bc candidate, four trusted-base workflows initially failed before creating jobs, and the original failures remain recorded. After the qualification-description update, growth enforcement, upgrade admission, maintainer edits and issue assignment completed successfully for the unchanged candidate. Growth ran all seven checks against the recorded head/base. The earlier startup cause remains unexplained; no retry, source change or bypass was used to obtain these passes. The historical full DEFAULT run failed; later scoped passes do not replace full runtime qualification. Staging qualification is outside the authorized DEFAULT79 scope. CodeRabbit is user-excluded; its skipped review is not approval and no check configuration changed. Scheduled CI monitoring is active.

Sensitive paths include installer/runtime scripts, images, inference, onboarding, messaging, policy, workflows and E2E tooling. Retained source review and isolated hooks cover ab75547742; current conflict checks are clear. Previous source CI and nine-specialist Advisor reports cover ab75547742; its image qualification passed. For 3695725be6, source CI, CodeQL and image qualification passed; Advisor requires the prepared delivery repair. The requested full E2E run is pending, and the previous installer failure remains unresolved. Historical evidence for earlier commits remains identified above. Independent maintainer acceptance of these sensitive changes remains outstanding. San owns landing. Deep Agents Code remains NOT hardened.

Inherited host installer audit warnings remain disclosed: root 23 affected packages (two critical), plugin eight (two critical); locked affected entries match the previously audited trusted base. Production-only results were three moderate root reports and zero plugin reports. These historical counts do not waive full host audit warnings, and no automatic dependency fix or suppression was applied.

Detailed prior evidence and consumed retries remain in the qualification status record. Original authoring-session reports remain unchanged.

---

Signed-off-by: Prekshi Vyas prekshiv@nvidia.com

[copy-pr-bot]

This pull request requires additional validation before any workflows can run on NVIDIA's runners.

Pull request vetters can view their responsibilities here.

Contributors can view more details about this message here.

[coderabbitai]

Important

Review skipped

Bot user detected.

To trigger a single review, invoke the @coderabbitai review command.

Run configuration

• Source mode: Auto
• Repository: NVIDIA/NemoClaw
• Branch: main
• File: .coderabbit.yaml
• Commit: 009a4f1
• Layers: Repository YAML
• Review profile: CHILL
• Plan: Advanced
• Run ID: 40ac7a56-bfc5-4d7f-818d-4fecb605f40b

You can disable this status message by setting the reviews.review_status to false in the CodeRabbit configuration file.

Use the checkbox below for a quick retry:

• Trigger review

• Autofix · Keep fixing CodeRabbit findings and required CI, and resolving merge conflicts

[github-code-quality]

Code coverage overview

Languages: TypeScript

TypeScript / code-coverage/plugin

main 63002cd	nemopatch/openclaw-u... 3695725	+/-
96%	97%	+1%

View coverage of the 10 most impacted files.

File	main 63002cd	nemopatch/openclaw-u... 3695725	+/-
nemoclaw/src/sh...cy-boundary.cts	100%	97%	-3%
nemoclaw/src/onboard/config.ts	98%	96%	-2%
nemoclaw/src/ru...time-context.ts	100%	99%	-1%
nemoclaw/src/bl...print/runner.ts	95%	95%	0%
nemoclaw/src/bl...ime-identity.ts	97%	97%	0%
nemoclaw/src/bl...t-management.ts	100%	100%	0%
nemoclaw/src/index.ts	94%	99%	+5%
nemoclaw/src/de...search-state.ts	0%	100%	+100%
nemoclaw/src/on...native-route.ts	0%	100%	+100%
nemoclaw/src/search-config.ts	0%	100%	+100%

TypeScript / code-coverage/cli

main 63002cd	nemopatch/openclaw-u... 3695725	+/-
84%	86%	+2%

View coverage of the 10 most impacted files.

File	main 63002cd	nemopatch/openclaw-u... 3695725	+/-
src/lib/onboard.ts	62%	50%	-12%
src/lib/actions...ess-recovery.ts	65%	83%	+18%
src/lib/inferen...ollama/proxy.ts	41%	61%	+20%
src/lib/onboard.../application.ts	55%	84%	+29%
src/lib/domain/...y/dimensions.ts	0%	73%	+73%
src/lib/inferen...er/lifecycle.ts	0%	75%	+75%
src/lib/actions...ry/operation.ts	0%	76%	+76%
src/lib/actions...try/snapshot.ts	0%	83%	+83%
src/lib/inferen...ovider/index.ts	0%	89%	+89%
src/lib/actions...ned-recovery.ts	0%	93%	+93%

ℹ️ Calculations represent line coverage. Results may change as calculations are updated.

---

_{Updated November 10, 2026 14:33 UTC}

[prekshivyas]

@coderabbitai review

[prekshivyas]

@coderabbitai review

[prekshivyas]

@coderabbitai review

[coderabbitai]

⚠️ Action not completed

Review skipped.

Note: CodeRabbit is an incremental review system and does not re-review already reviewed commits. This command is applicable only when automatic reviews are paused.

[prekshivyas]

Resolved the four readiness conflicts against main 4d5d613f73 and published Verified, DCO-signed commit 57cabe80bcad53ff885493dce528acffd3186ce4. GitHub reports no merge conflicts. The author requested ready-for-review status; independent approval and merge readiness remain outstanding. No maintainer review was requested and no merge is authorized.

Local checks: 164 readiness tests passed with seven existing skips; 58 runtime-provenance tests and 31 installer ACP contracts passed. Normal isolated commit/pre-push hooks, repository budgets and CLI type checking passed. The original installer harness failures are retained; system Git inside the same credential-denying sandbox corrected fixture-global-config interference without changing repository checks or timeouts.

New source CI, image, CodeQL and scoped runtime checks are running. The latest Advisor evaluation and every scheduled specialist report still require collection. Prior candidate 9203b203a results do not qualify this merge. All feedback pages were collected; all 25 threads remain resolved. CodeRabbit remains user-excluded, with no review request or check mutation.

Four trusted-base workflows failed before creating any jobs: growth enforcement, upgrade admission, maintainer edits and issue assignment. Their files match main, and unrelated PRs show the same startup failure. GitHub reports a likely workflow-file issue, but the retrieved API supplies no detailed root cause. This requires maintainer investigation of trusted-base workflow/Actions configuration. No deterministic rerun or bypass was attempted.

The original v0.0.123 installer still fails its unchanged 35-minute cap. Later storage, concurrency, channel and v0.0.89 passes remain scoped evidence; full DEFAULT79 qualification is incomplete. The older disputed Advisor P1 still needs independent adjudication. Scheduled monitoring remains paused, original retry budgets and backups remain intact, and Deep Agents Code remains NOT hardened.

[prekshivyas]

Independent review requested for 57cabe80bcad53ff885493dce528acffd3186ce4, integrating main 4d5d613f73. The author has explicitly asked to obtain approval; this request does not authorize merge or waive unresolved qualification.

Please adjudicate the older Advisor P1 F-customer-value-behavior-0bb4d28d2e6ea69be779: “Failed Model Router startup can leave a router that blocks retries” (original Advisor run). It concerns startup-health failure plus rejected identity-bound termination. Prior investigation disputed the finding because public reconciliation discovers orphan ownership before another start, with mocked diagnostic coverage; that is not an independent adjudication. Please record the disposition against the current router lifecycle source and process ownership source.

Current source CI, image qualification, and all nine Advisor specialists passed. Their scoped results and limitations are in the PR description.

The original runtime diagnostic still failed the v0.0.123 installer’s unchanged 35-minute cap. The remaining 73-case plan and full DEFAULT79 qualification remain incomplete. This review request does not claim those failures are fixed or seek a check waiver. Deep Agents Code remains NOT hardened. No merge is authorized.

[github-actions]

PR Review Advisor finished for commit 3695725. Include the Advisor findings in the complete PR feedback collection. Verify and group valid findings before repair.

Request review only when Require no Advisor blockers is green.

All previous runs

[prekshivyas]

https://github.com/NVIDIA/NemoClaw/actions/runs/42847156265 e2e evidence