[nvidia-nemopatch-writer]

Outcome

Not ready for approval or merge. Upgrade NemoClaw to OpenShell 0.1.5 while preserving supported inference, lifecycle, credentials, ownership, recovery, legacy data, and GPU behavior. Full qualification is incomplete.

Reason

OpenShell 0.1.5 is the interim stable baseline for the runtime and native inference migration. Maintainer direction preserves GPU support and holds qualification until compatible runtime paths exist. Preserve legacy volumes without automatic copying; defer independent improvements.

Accepted dependency-update scope

The maintainer decision for #12603 authorizes the OpenShell v0.1.5 dependency update and continued qualification. #12558 owns the inference pre-work and excludes the dependency update; #12603 owns this separate update.

• Update the CLI, SDK, gateway, supervisor, images, and hashes required for v0.1.5.
• Preserve existing agents, providers, platforms, GPU support, credential custody, isolation, resource ownership, and rollback requirements.
• Preserve legacy volumes unchanged. Stop before mutation when ownership is unproven; do not add automatic copying.
• Keep repository budgets and meaningful coverage. Defer unrelated work and new integrations.
• Run full default PR E2E on the final candidate after its required images qualify. Attribute failures to main only with matching evidence.

This decision authorizes qualification, not merge or release. The released-runtime TLS, Personal web-access, and GPU qualification blockers remain unresolved.

Related issues

Refs #12558 and #13338. Registered-provider selection follows merged #12333. Controller #13486 is merged and consumed.

Changes

• Adapt SDK requests, workspace selection, pagination, authentication, policies, and forwarding. Reject incomplete inventories and ambiguous resource identities.

• Use sandbox-attached native inference providers and preserve operator-owned providers and credentials.

• Preserve retained-state refusals, legacy volumes, bounded Hermes recovery, and ownership checks during cleanup.

• Update the owning inference and credential documentation.

• Retry NVIDIA Endpoints HTTP 500 within existing request limits and retain the correct probe name in errors.

• Request existing gateway cleanup while Pi qualification still has a registered sandbox.

• Consume main's native-provider recovery and npm/image repairs; preserve registered-provider ownership, retained-volume handling, and credential isolation.

• Separate host credential validation from gateway-provider checks to keep the rebuild dependency budget unchanged. Existing tests retain invalid-credential, endpoint pinning, and pre-delete refusal coverage.

• Require HTTPS before using registered credential-bearing providers. Retained HTTP records remain readable and safely detachable; local-runtime HTTP contracts remain separate.

Verification

Current candidate

Candidate 675c1e10e31c4642d4accd8ba2718312255fffd1 rechecks sandbox identity immediately before native provider configuration writes. It retains main 82a1a8f9cb263b6fecdb508e407c5f84f72fd685 and OpenShell 0.1.5.

• Reject identity drift after asynchronous verification, before writing credentials or reconciling the native timeout. Preserve retained resources, receipts and existing refusal behavior.
• All 146 focused tests passed. CLI type-checking passed with the repository's existing 8 GB Node.js heap. Documentation built with zero errors and two existing warnings. Normal commit hooks and independent source, security and documentation review passed.
• This repair addresses the previous Advisor security finding. Fresh CI, automated review, managed images and full PR E2E remain pending. The recurring monitor remains stopped; runtime, merge and release holds remain.

Previous candidate 4c6560

Code CI and CodeQL passed. Advisor reported the configuration-write identity race repaired above. Seven of eight managed-image contracts were verified. Hermes ARM64 terminated with a native segmentation fault after the Switchyard probe printed its success message. The native crash cause remains unresolved; image activations were skipped. No full E2E was dispatched for that candidate.

Candidate 4c6560bca8f7a4dbd6daaf5935c77a2114fe015b adds the missing forward-port release classification after the rootless Podman activation failure on be981db9. It retains main 82a1a8f9cb263b6fecdb508e407c5f84f72fd685 and OpenShell 0.1.5.

• Report bound or indeterminate/<fixed-kind> without adding raw errors, command output or credentials. Update stop, destroy and rebuild messages to describe unverified release accurately.
• Preserve the five-second verification deadline, failure result, registry retention, process ownership and cleanup behavior. This diagnostic does not establish or repair the original listener failure.
• Focused tests passed: 111 tests in the initial run, 95 in the caller update run, and 38 after the typed fixture update. These runs overlap and are not additive. Documentation built with zero errors and two existing warnings. The first push was rejected by CLI type-checking in the new test fixture. The corrected fixture passed CLI type-checking and all 38 focused tests. Normal commit hooks and independent source, security and documentation review passed.
• The 325-predicate source review and complete 4,310-line failed-job log review preserve the remaining runtime evidence gaps. Those local results do not qualify the new candidate.

Previous candidate be981db9

The runtime identity timing diagnostic passed code CI and all nine Advisor specialists. All eight image contracts and native AMD64 and ARM64 activations were verified. Rootless Podman activation failed because external OpenClaw destroy returned exit 1 after host-port release could not be proved. The old message omitted whether the result was bound or indeterminate.

Final sandbox/container absence and registered cleanup were verified. Gateway recreation occurred before later successful cleanup, so that later success does not prove the original port released on its own. OpenClaw receipt and disclosure predicates after the failed destroy check were not evaluated; additional false predicates are not excluded. No full E2E was dispatched for this candidate.

Previous candidate 9902508

Candidate 9902508b3942114ed3e4fd239bd064adfa163bca added bounded categories for failed OpenClaw pairing observations. Code CI, all nine Advisor specialists, image contracts and all three activation platforms passed. JavaScript/TypeScript CodeQL was cancelled at its existing timeout; it was not retried. Its full E2E result is recorded below. Historical approval and qualification evidence do not apply to a later commit.

Previous candidate e977a2f

Candidate e977a2f0eec46e88c6ed6f07b0947b01800b57d6 adds bounded evidence for the OpenClaw restart readiness failure observed on 3a2b73a. It retains integrated main 82a1a8f9cb263b6fecdb508e407c5f84f72fd685, OpenShell 0.1.5, the receipt correction, #13997 diagnostics, #14027 DRM policy, and the explicitly approved upstream timeouts.

• On OpenClaw readiness failure, retain the last 32 probe observations, total and dropped counts, native restart status, command/transport status, HTTP status, JSON parse result and Boolean readiness. Exclude raw bodies, stderr, URLs, configuration and credentials.
• Capture the existing bounded post-failure readiness artifact after native-model restoration restart. Preserve the original failure, cleanup, commands, retries, deadlines and readiness acceptance rule.
• All 81 focused CLI and E2E-support tests passed. Documentation validation passed: 125 routes, zero errors and two existing warnings. Independent correctness, security and documentation writer review passed for all five changed files. Standalone CLI type-checking exhausted the default Node heap. The actual isolated pre-push hook subsequently passed publication validation and CLI TypeScript checking with its existing configured heap. Normal commit hooks also passed.
• This change adds diagnostic evidence. It does not establish the historical restart failure's root cause. New candidate CI, reviews, image activation and runtime qualification remain pending. Runtime, release and merge holds remain.

Previous candidate 3a2b73a

Candidate 3a2b73af55138dc8bb0d77af2c513e4cc216f241 integrates main 82a1a8f9cb263b6fecdb508e407c5f84f72fd685 to resolve the Tavily test conflict and repairs plugin coverage. OpenShell remains pinned to 0.1.5. The receipt correction, #13997 diagnostics, #14027 DRM policy and explicitly authorized upstream timeouts remain included.

• Retain test(e2e): keep the Tavily check credential-free in hosted CI (#13082) #14025's credential-free hosted Tavily check. Hosted CI forwards no Tavily key or authenticated-mode flag and runs check 09 whether a key is missing, empty or present. Preserve failure and skip rejection, manual authenticated checks, audit evidence and cleanup tests.
• Keep Tavily tests in their existing extracted file while applying upstream behavior changes. Clarify the hosted and manual qualification paths in the E2E README.
• Move 18 existing pure host-pattern tests from the CLI caller suite to the shared plugin suite that owns the matcher. Add nine matcher cases and six endpoint-pin admission/refusal cases. Keep registered-provider security, ownership and transport tests, runtime source and coverage thresholds unchanged.
• All 160 focused CLI and E2E-support tests passed. All 1,038 plugin tests passed. Plugin coverage passed the unchanged ratchet: statements 96.10%, branches 89.13%, functions 99.55% and lines 96.95%. The live census remains 1,260 direct assertions across 77 files. Documentation validation passed with 125 routes, zero errors and two existing warnings. Normal commit hooks passed. Independent correctness, security and documentation writer review passed for all eight changed paths at this commit. Gitleaks and private-key checks passed; the diff contains no secrets, API keys or credentials.
• Code CI, CodeQL, all nine Advisor specialists, eight image contracts and AMD64/ARM64/rootless Podman activation passed. The full default PR E2E result below passed protected runtime but failed eight other scenarios. These results do not qualify a later candidate.

Previous candidate 1892a2b

Candidate 1892a2bf42a016c3247fdf509b7a94ab524da350 integrated main 444ce08acf2aa036060448357a0e36cfc6eb7b23, repaired the image-trigger and registered-provider transport findings, and retained custom-endpoint IP-pin conflict checks.

• Local integration validation initially passed 433 tests and failed two fixture expectations because the fixture already installed a fake curl executable. Comparing the original directory inventory repaired those expectations, and all 37 affected tests passed. Later compiled/shared checks passed 213 tests; final hook repairs passed 162 policy/matcher tests and 76 action tests. These runs overlap and are not additive. CLI build, docs (125 routes, zero errors, two existing warnings), the 1,260-assertion census, independent review of all 45 integration paths, and normal commit/push hooks passed.
• All eight image contracts and AMD64, ARM64 and rootless Podman activations passed in image run 41139855238. Archive hashes, candidate/cohort/image identities, recorded cleanup, external-image retention and legacy-volume protection were verified.
• Code CI failed only plugin coverage; all 1,005 plugin tests passed. The unchanged-source reproduction identified matcher tests in the wrong coverage project. The current candidate repairs that defect. All CLI shards and the aggregate CLI coverage job passed.
• The code-CI-triggered Advisor skipped because CI failed; no specialist qualification was claimed. Portable and Podman CPU checks passed with verified artifacts and their recorded cleanup limits. Hermes image entrypoint checks passed on both architectures; they do not establish live GPU enforcement or the historical ARM64 failure cause.
• CodeQL exceeded its unchanged 15-minute JavaScript limit during SARIF export after all 203 queries completed. The recorded base passed with the same tool versions and limit. The slowdown cause remains unresolved; no unchanged retry, timeout increase or main-failure exclusion was used.
• Candidate/base E2E admission passed. No full PR E2E was dispatched. Complete paginated feedback added no further source finding. These results do not qualify the current candidate.

Previous candidate a98ef85d

Candidate a98ef85d3749c69e676e3a82708dc4bef07e09fa refreshed Pi image receipts and integrated main 85a3a8b29146e1e2e03a76a57ede184aa8231aa8, including #13515. The explicitly authorized WSL shard 1 limit of 300 minutes and expanded dashboard limit of 115 minutes remain included.

• AMD64 and ARM64 Pi receipts came from successful image run 41118794250 on 43c50f4, with verified immutable references, source identity and receipt authority. All 69 receipt tests and 645 distinct integration tests across 21 files passed. Nine initial failures were local loopback denials; all 45 tests in the two affected files passed after localhost fixture traffic was allowed, with assertions and limits unchanged. Documentation validation passed with 125 routes, zero errors and two warnings; independent review covered all 54 changed files. Normal commit and push hooks passed.
• Code CI passed. All eight image contracts and AMD64, ARM64 and rootless Podman activation passed, with verified identities, archive hashes and recorded cleanup.
• All nine Advisor specialists completed. The catalog-trigger and registered-transport findings were valid and are repaired in the current candidate. The blueprint-use finding contradicted production runner calls; the scope finding repeated the accepted maintainer decision linked above. The hosted blocker remained failed.
• CodeQL exceeded its 15-minute JavaScript/TypeScript limit after query execution. Matching main completed under the same limit; the slowdown cause remains unproven. No retry, limit increase or main-failure exclusion was used.
• Candidate/base E2E admission passed. No full PR E2E was dispatched for this candidate. Its historical results do not qualify the current commit.

Previous candidate 43c50f4

Candidate 43c50f4ac677ba36ac8edbddcc4c9db015a6b824 integrated main 81fe4746e00d66c4ffc28dccda1d9931458d536b, including #14027's DRM read grant. Its 84 policy tests and documentation validation passed.

• All eight image contracts and three activation jobs passed. Verified AMD64, ARM64 and rootless Podman artifacts cover startup, retained images, identity drift, legacy-volume protection and recorded cleanup.
• Code CI failed the strict Pi source-parity check because historical receipts did not match the changed image inputs. The current candidate refreshes those receipts. The other code CI jobs passed; the code-triggered Advisor skipped without specialist qualification.
• CodeQL's JavaScript/TypeScript job exceeded its 15-minute limit. Python and ShellCheck passed. No unchanged retry or timeout increase was performed.
• Candidate/base E2E admission passed. No full PR E2E was dispatched for this candidate. These historical results do not qualify later candidates.

Previous candidate 6d4b87b

Candidate 6d4b87b6412321b0f91f0cae8abc58ac0dc78e2c integrates main f0be87d73f43338b99e12b1530c9317cf407d534, including Discord PR #13873 and fetch/restart diagnostics PR #13997. OpenShell remains pinned to 0.1.5. The receipt correction, answer diagnostic and nested fetch evidence repair remain included.

• Resolve the Discord integration conflicts while preserving native inference, ownership, Pi receipt authority, credential checks and all existing assertions. Regenerate the direct runtime bundle and verify its recorded hash.
• Make the local tar fixture omit macOS AppleDouble metadata, matching the Linux sandbox archive producer. Production archive inventory validation remains unchanged.
• Add bounded HTTP failure evidence and a separate HTTPS diagnostic probe. The later probe cannot prove the original response or change the failed assertion into success.
• Preserve raw OpenShell stop/start and original listener observations. Follow the accepted restart contract with explicit NemoClaw recovery and dashboard checks under both destruction policies. Require cleanup of the separate observer process.
• Preserve nested fetch run/parent correlation, strict success evidence, retry classification and all existing limits. Extract diagnostic status parsing within the serialized reducer to satisfy the unchanged complexity limit.
• The Discord integration passed 775 distinct targeted tests. Its managed-Hermes-only package test was skipped on this host. The diagnostics integration passed 279 targeted tests; all 95 affected tests passed again after the helper extraction. These sets overlap and are not a combined distinct count.
• CLI build, exact test-project membership and assertion census passed. Documentation validation passed with 125 routes, zero errors and two warnings. Independent correctness, security and documentation writer review passed. Normal commit and push hooks passed.
• At the next required dependency integration, hosted CI, reviews and image qualification for this candidate were still incomplete. Its early admission and installer-hash checks required a fresh candidate/base approval. Static checks then rejected the historical Pi receipts because their image inputs no longer matched. The documented source-first publication permits collecting new image receipts afterward; this is not a CI pass. No approval or full E2E was dispatched for this candidate. No historical Wikimedia cause, original forward exit cause or current runtime pass is claimed. Runtime, release and merge holds remain.

Previous candidate 8fef778

Candidate 8fef778a09bc149669d7b3bc17f712e1877890a3 integrated Tavily revocation PR #13662. Its 566 distinct targeted tests, later 164-test rerun, CLI build, docs and normal hooks passed. CodeQL passed. New main conflicts prevented code CI, images and Advisor from starting; admission required a new candidate-bound approval. Complete paginated feedback contained no new actionable source finding. This candidate did not run full E2E.

Previous candidate 893bef

Candidate 893bef0ac6c9f3ab99cd5fa504c1097b409477e5 adds evidence parsing for nested web_fetch results from pinned OpenClaw 2026.9.2. It retains the main integration, receipt repair and runtime diagnostics in 62b22a9b9d2618b19cc4b30acb0d3107d18b1007.

• Accept only bounded nested records tied to the selected completed run and an allowed wrapper in both the session and selected trajectory snapshot. Wrong-parent, wrong-run, malformed and error records cannot establish a successful fetch.
• Preserve existing URL, tool, response and credential-isolation assertions. Reduce raw tool data inside the sandbox; retain only the existing counts, flags and host/protocol fields in evidence.
• All 71 tests across two files passed after the final edit, including the generated SQLite reducer subprocess test. Independent correctness, security and documentation writer review passed for all four committed file hashes.
• This repairs a deterministic downstream evidence defect. Personal onboarding still refuses overlapping provider/web policies before this parser runs; no live Personal pass is claimed.
• CodeQL passed. Merge conflicts prevented code CI, image qualification and Advisor from starting. Admission required a new commit-bound E2E approval. Complete paginated feedback contained no new actionable source findings; this candidate did not run full E2E.

Previous candidate 62b22

Candidate 62b22a9b9d2618b19cc4b30acb0d3107d18b1007 integrates main 4ec4ec634e5e9cf5f00c6bb2d2786a5b5b73eef9 to resolve the merge conflict. OpenShell remains pinned to 0.1.5; the receipt correction, answer diagnostic and ba9214d fixture repairs remain included.

• Retain fix(onboard): verify dual-stack inference and accept maintainer reports #13767's direct-inference behavior with explicit user authorization: check every resolved IPv4/IPv6 address and accept connection refusal or network/host unreachable as unavailable access. A reachable response must satisfy the policy-denial contract, including the existing native typed fields. Transport failure does not establish policy enforcement.
• Preserve timeout and byte limits, model/tool checks, credential isolation and ownership checks. No additional retries or timeout increases were introduced.
• Add four native IPv6-to-IPv4 regression cases. Move the existing shell smoke cases without changing their assertions to keep both suites below the file-size limit.
• All 172 targeted tests passed: 126 inference smoke tests and 46 backup tests. Test-project membership passed. Documentation build passed with zero errors and two existing warnings.
• Independent correctness, security and documentation writer review passed for all 14 integrated files. Normal pre-commit, commit-msg and pre-push hooks passed. All 290 PR commits are Verified; local, SSH and PR identities agree.
• Code CI, CodeQL, documentation, growth and managed images passed. All eight image contracts have verified archive hashes, candidate/cohort identities and immutable image references.
• AMD64, ARM64 and rootless Podman activation each passed two scenarios. Verified artifacts record 13 activation phases, 13 cleanup actions, five final resource-absence pairs, external-image retention and identity-drift rejection, plus legacy-volume preservation and fixture cleanup. These results do not qualify protected GPU inference or the full E2E suite.
• All nine Advisor specialists completed. Eight are clear. Customer value repeats the scope objection addressed by the accepted maintainer decision linked above; the hosted blocker remains failed. CodeRabbit skipped the bot-authored PR.
• Admission initially lacked this candidate’s approval. The authorized candidate/base approval was published and verified before the full E2E run below.
• Hermes ARM64 image qualification passed this run. The native-exit diagnostic archives are retained, but the prior crash cause remains unproven. Full default PR E2E subsequently failed as detailed below. Runtime, release and merge holds remain.

Historical repair candidate ba9214d

Candidate ba9214d17fdca2f0932d3ece814ae9f75027224a adds two test-fixture repairs and a bounded Hermes native-exit diagnostic on top of 29bde. OpenShell remains pinned to 0.1.5.

• Mock the gateway failure classifier in status rendering tests so they do not start real Docker or network probes. Assert the expected classification and classifier call.
• Preserve descriptor 3 through the snapshot test's tar wrapper. Production restore reads the validated, unlinked archive from that descriptor.
• Enable Python faulthandler in the Hermes Switchyard probe and record whether the library remains mapped after shutdown. Preserve all 22 Python assertions, the required zero exit status, timeouts and cleanup.
• All 74 focused status/snapshot tests and 64 managed-image workflow/installer tests passed. Syntax and assertion-preservation checks passed. Independent review passed for all four changed files; no public documentation change is required.
• Complete review of the 18,328-line Hermes failed-job log and 340 source predicates confirms the failure occurs after the Python assertions. The native fault frame remains unknown. This diagnostic is not a crash fix or runtime qualification.
• Normal pre-commit, commit-msg and pre-push hooks passed. All 289 commits were Verified. CodeQL and growth checks passed; admission required a new candidate approval. A merge conflict prevented code CI and image qualification. Full default PR E2E was not dispatched.

Main integration candidate 29bde

Candidate 29bdeabb1f1ce3247d5bd3f921dccaae0ab8fb10 integrates main de00977585d04f3bac7b090a0870b0d2dca4972d to resolve merge conflicts. It retains the host receipt correction from #13887 and the bounded answer diagnostic from #13932. The OpenShell CLI and SDK remain pinned to 0.1.5.

• Preserve detached sandbox creation, mTLS Health probes, native provider policies, credential isolation, ownership and retained data during integration.
• Accept complete OpenShell 0.1.x sandbox inventories during portable transaction cleanup. Reject partial pages, malformed rows, duplicate names and transport failures.
• Keep the providers_v2 refusal regression on its legacy runtime branch; retain the separate native projection tests.
• Retain ci(e2e): retain portable Hermes failure diagnostics and cleanup #13686's upstream portable rootless timeouts with explicit user authorization: 46 minutes 15 seconds for the test and 85 minutes for its job. No further increases were made.
• Initial focused validation: 356 passed, 1 skipped and 1 failure in the legacy refusal fixture. After repair, all 103 targeted tests passed across five files, including 13 cleanup inventory cases. These runs overlap.
• Test project membership passed for 2,923 candidate files across seven disjoint projects. Documentation build passed with zero errors and two existing warnings.
• Independent review passed for all conflict resolutions and affected integration overlaps. Normal pre-commit, commit-msg and pre-push hooks passed, including publication validation and CLI type checks. All 288 commits in the PR are Verified on GitHub.
• Code CI failed in the status-flow and native snapshot test fixtures. Hermes ARM64 image qualification failed after the Switchyard probe printed success and exited 139; five other root-entrypoint scenarios passed. Seven other image contracts passed, but dependent activation was skipped. Advisor was skipped after CI failed. Full default PR E2E was not dispatched.

Historical candidate validation

Preceding candidate bfa084f396c11ae1aa504f286aaae88d56909737 integrated main b10117ecdf3318da48e882362d971dce6879750c. Its 566 focused tests, later 185-test overlap run, documentation build and publication hooks passed. CodeQL and growth checks passed. E2E admission required a new candidate approval, and subsequent merge conflicts prevented normal PR code CI and image qualification.

Earlier candidate 5dfa15c completed code CI, CodeQL, admission and image qualification, including eight image contracts and three activation platforms. Its registered-provider HTTPS correction remains included. Nine Advisor specialists completed; the hosted scope blocker remained red despite the documented accepted scope decision. Historical results do not qualify the current candidate or bypass a current gate.

Latest completed full PR E2E

Run 41189216914, attempt 1 tested 9902508b3942114ed3e4fd239bd064adfa163bca, starting October 31 at 09:34:02Z (2:34 AM Pacific). Empty job and target selectors selected the full default PR suite. Base and executing workflow: 82a1a8f9cb263b6fecdb508e407c5f84f72fd685.

Result: 73 successful jobs, 19 skipped, nine failed scenarios and the failed Relevant E2E summary.

Failed scenarios	Count	Established result
Private MCP TLS	3	Private TLS servers recorded unknown-CA alerts before HTTP. The upstream CA feature is absent from the pinned release.
Personal web access and policy ordering	2	The Personal compatibility guard refused after onboarding. The specific overlap or unverifiable-endpoint branch was not captured.
Hermes GPU compatibility and fallback	2	The selector rejected the supervisor/workload pair. A selector-only repair cannot provide the required OpenShell-owned bootstrap and recreation contract.
Runtime identity	1	Two OAuth issuances preceded the first resource assertion. All 17 probes authenticated with version 2, while the assertion required version 1. Exact initiating callers remain unproved.
Protected managed-image runtime	1	All three agents passed Ollama and vLLM. NIM image preparation exhausted the shared deadline before NIM agent checks or rollback. Buffered child output does not distinguish pull from tag at interruption.

Both public-reference fetch scenarios and the concurrent dashboard lifecycle passed. These results do not establish the historical failures' causes. Both native startup architectures passed with verified artifacts. Failed-job content, candidate/workflow provenance, result receipt bindings and the retained protected-runtime cleanup evidence were reviewed.

Registered cleanup callbacks passed within their recorded limits. Full protected-runtime resource absence remains unverified. The failed identity scenario did not reach its provider/profile deletion commands, which are not registered failure cleanup callbacks. No resource leak is confirmed. No unchanged retry, relaxed assertion or additional runtime limit was used. This run does not qualify the new diagnostic commit.

Historical full PR E2E on 3a2b73a

Run 41151876561, attempt 1 tested 3a2b73af55138dc8bb0d77af2c513e4cc216f241, starting October 31 at 00:45:23Z (October 30 at 5:45 PM Pacific). Empty job and target selectors selected the full default PR suite. Base and executing workflow: 82a1a8f9cb263b6fecdb508e407c5f84f72fd685.

Result: 74 successful jobs, 19 skipped, eight failed scenarios and the failed Relevant E2E summary. Release qualification was skipped as expected for manual PR mode.

Failed scenarios	Count	Established result and remaining work
Private MCP: OpenClaw, Hermes, Deep Agents Code	3	Private discovery failed with unknown-CA TLS alerts. OpenShell #3292 merged, but its fix is absent from pinned 0.1.5 and latest stable 0.1.6. A released runtime containing the fix and gateway CA integration need qualification.
Hermes GPU compatibility-only and fallback	2	The legacy container guard conflicts with the native workload/supervisor pair. A selector-only change would not preserve runtime-owned bootstrap, identity and network authority. OpenShell #4837 remains open; its WSL prototype is unpublished and does not qualify these scenarios.
Personal public fetch and policy preset ordering	2	NemoClaw refused overlapping Personal/custom-provider policy before later assertions. OpenShell #4662 still requires an accepted policy-composition design.
OpenClaw security posture	1	Native model invocation returned PONG; model restoration and validation passed. The subsequent restart failed its readiness wait. The retained evidence lacks the HTTP/readiness result; this revision adds bounded diagnostics.

Protected managed-image runtime passed all nine agent/provider combinations across OpenClaw, Hermes and Deep Agents Code with Ollama, vLLM and NIM. All three bootstrap rollback cases, five phases, four registered cleanup actions, final owned provider/sandbox/network inventory, child cleanup guards and workflow registry/auth cleanup passed. Archive bytes and SHA-256, candidate/workflow identity, seven projected/restored sources and all 39 command artifacts were verified. The executing harness belongs to trusted workflow 82a1a8f; this does not claim execution of the differing candidate helper. NIM's recorded image ID binds its mutable requested tag.

Both public-reference jobs and concurrent dashboard lifecycle passed. These passes do not establish the causes of the historical Wikimedia or forward-lifetime failures. Both native startup architectures also passed with verified evidence. All eight failed scenario logs and the source/runtime reconciliation are complete. No failure was excluded as inherited without evidence, and no unchanged retry, relaxed assertion or increased runtime limit was used. This run does not qualify the next commit or clear full runtime, release or merge holds.

Historical full PR E2E on 62b22

Run 41087961129, attempt 1 tested 62b22a9b9d2618b19cc4b30acb0d3107d18b1007 on October 30, from 16:13:52Z to 17:32:21Z. Empty job and target selectors selected the full default suite. Base and trusted workflow were 4ec4ec634e5e9cf5f00c6bb2d2786a5b5b73eef9.

Result: 74 successful jobs, 19 skipped, eight failed scenarios and the failed summary job.

Failed scenario	Established result and remaining limit
OpenClaw private MCP	Private endpoint recorded unknown-CA TLS alerts and no secure connection. Private discovery failed.
Hermes private MCP	Same private TLS rejection; provider attachment and pins do not establish successful discovery.
Deep Agents Code MCP	Docker readiness exceeded its 15-second limit before sandbox creation. The Docker stall cause is absent from retained output.
Hermes GPU fallback	Compatibility cutover failed after the main process exited. Native workload/supervisor layout is incompatible with legacy recreation; the precise rejected guard predicate is not captured.
Hermes GPU compatibility-only	Same startup/cutover boundary. A selector-only repair would not establish runtime-owned bootstrap and network authority.
Personal public fetch	NemoClaw refused overlapping Personal/custom-provider policy. Later fetch and parser assertions were not reached.
Concurrent dashboard	Sibling destruction preserved the survivor. Dashboard probes refused connections after raw stop/start; the original forward exit branch was not captured.
Protected managed-image GPU	OpenClaw/Ollama passed. Hermes/Ollama failed startup; later agents, vLLM, NIM, rollback and final inventory were not reached. The narrow DRM policy prerequisite merged through #14027 and is integrated; a new runtime result is still required.

Both public-reference jobs passed in this run. This does not prove the cause of the historical Wikimedia failure. Native startup passed for all three agents on AMD64 and ARM64, with verified artifacts.

All failed scenario logs, the summary log and 1,806 source predicates are reviewed. Candidate/workflow provenance and downloaded artifact hashes are verified. Registered cleanup actions passed within their recorded contracts; full protected sandbox-record and volume absence is not established. No leak is confirmed. No unchanged retry, relaxed assertion or additional timeout increase was used.

Historical full PR E2E on 909db

Run 40630819405, attempt 1 tested preceding candidate 909db53d941a772a778940ce5033f32afd2f2781: 73 passed jobs, 19 skipped, nine failed scenarios plus the failed summary check. Empty job and target selectors selected the full default suite. Base 1864afdba and trusted workflow 3619e535e are verified. No failure is excluded as already present on main.

Failed scenarios	Count	Evidence and required work
Private MCP: Deep Agents, Hermes, OpenClaw	3	Private discovery fails with unknown-CA TLS alerts and no secure private connections after provider attachment and matching pins. The Docker supervisor CA gap remains unresolved.
Hermes GPU compatibility and fallback	2	Canonical startup exits before compatibility replacement. The rejected startup-guard predicate remains unproven. Preserve the guard and qualify runtime-owned GPU handling.
Personal public fetch	1	Onboarding explicitly rejects provider/Personal policy overlap. Safe coexistence must preserve provider restrictions.
OpenClaw public reference	1	The agent reports Wikimedia 503; no raw HTTP evidence confirms it. Two failed tools and no successful tool names leave an incomplete replay that the parser correctly rejects.
Concurrent gateway lifecycle	1	The survivor gateway, registry and Ready state remain intact, but dashboard port 18789 refuses all twelve probes after raw unscoped stop/start. OpenShell forward lifetime is a plausible cause, without captured process-exit proof.
Protected managed-image GPU/local inference	1	OpenClaw bootstrap, CUDA initialization and the direct provider probe pass. The host receipt check then lacks authority because it receives the filtered sandbox environment. The trusted harness correction merged through #13887 and is included in this candidate; full live qualification is pending.

All nine failure archives have verified identities and successful cleanup. Complete failed-job logs and the affected source assertion audit are reviewed. In that historical PR run, protected NIM/vLLM and later agent phases were unreached. Source diagnosis and local tests do not establish live success.

For tested candidate 909db, code CI, admission, installer hashes and managed images passed. All eight builds and AMD64, ARM64 and rootless Podman activation passed; their archives and cleanup are verified. CodeQL passed its one bounded cancellation retry. These results do not qualify the current candidate.

Review notes

Independent correctness, security and documentation review passed for the four-file timing diagnostic. The identity assertion review preserves the unresolved issuance and cleanup limits; it does not approve runtime qualification or a new E2E run.

The PR author, nvidia-nemopatch-writer[bot], posted the opening authoring-session report. The report marks the original session evidence unavailable and excludes later repair and qualification work. Exact bot identity, unchanged report text and complete comment pagination were verified; the retrieved report passes the repository validator.

Main integration includes the receipt-authority correction from #13887 and the bounded answer diagnostic from #13932. Main protected-runtime run 40877717150 passed all-agent Ollama/vLLM phases and the OpenClaw NIM command, then exhausted the shared 45-minute deadline during Hermes NIM. That historical main run did not qualify Hermes/Deep Agents Code NIM or rollback, and interrupted-child resource absence was not established. Historical PR run 41151876561 on 3a2b73a completed those phases. Neither run proves a permanent repair for the historical empty-answer behavior.

Controller repairs #13635 and #13763 and SDK prerequisite #13796 remain integrated. The preceding update integrated main 82a1a8f9cb263b6fecdb508e407c5f84f72fd685. This revision adds runtime identity timing diagnostics after the preceding candidate feedback and E2E evidence completed. Runtime qualification, release and merge holds remain.

The historical Podman restart race remains tracked by OpenShell #4932. Later activation passes do not prove that race fixed. No restart workaround, readiness relaxation or unsupported main-failure exclusion was introduced.

Sensitive paths in the overall PR include workflows, runtime scripts and images, policy, inference, messaging, onboarding and validation tooling. Independent review covers the recorded repair files; hosted review of prior candidate 62b22 retained the scope blocker described above. Hosted review of the new candidate remains pending. Full qualification, approval and merge remain blocked.

---

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: 82a1a8f
• Layers: Repository YAML
• Review profile: CHILL
• Plan: Advanced
• Run ID: c1f759a2-fb74-4dd9-832c-93df10c6080e

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-actions]

🌿 Preview your docs: https://nvidia-preview-pr-12603.docs.buildwithfern.com/nemoclaw

[github-code-quality]

Code coverage overview

Languages: TypeScript

TypeScript / code-coverage/plugin

main 63002cd	nemopatch/openshell-... 675c1e1	+/-
96%	97%	+1%

View coverage of the 10 most impacted files.

File	main 63002cd	nemopatch/openshell-... 675c1e1	+/-
nemoclaw/src/onboard/config.ts	98%	96%	-2%
nemoclaw/src/index.ts	94%	93%	-1%
nemoclaw/src/bl...print/runner.ts	95%	95%	0%
nemoclaw/src/sh...cy-boundary.cts	100%	100%	0%
nemoclaw/src/bl...ime-identity.ts	97%	97%	0%
nemoclaw/src/bl...t-management.ts	100%	100%	0%
nemoclaw/src/co.../config-show.ts	100%	100%	0%
nemoclaw/src/commands/slash.ts	100%	100%	0%
nemoclaw/src/bl...est-fixtures.ts	98%	99%	+1%
nemoclaw/src/on...native-route.ts	0%	100%	+100%

TypeScript / code-coverage/cli

main 63002cd	nemopatch/openshell-... 675c1e1	+/-
84%	86%	+2%

View coverage of the 10 most impacted files.

File	main 63002cd	nemopatch/openshell-... 675c1e1	+/-
src/lib/onboard.ts	62%	51%	-11%
src/lib/onboard...mage-runtime.ts	52%	71%	+19%
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/actions...ry/operation.ts	0%	76%	+76%
src/lib/actions...try/snapshot.ts	0%	83%	+83%
src/lib/inferen...er/lifecycle.ts	0%	87%	+87%
src/lib/inferen...r/registered.ts	0%	92%	+92%
src/lib/inferen...er/blueprint.ts	0%	93%	+93%

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

---

_{Updated October 31, 2026 14:28 UTC}

[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]

@coderabbitai full review

[coderabbitai]

⚠️ Action not completed

Review skipped.

[github-actions]

PR Review Advisor finished for commit a98ef85. 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]

NemoClaw-Upgrade-E2E: approve 1892a2b 444ce08

[prekshivyas]

NemoClaw-Upgrade-E2E: approve 3a2b73a 82a1a8f

[github-actions]

PR Review Advisor finished for commit 3a2b73a. 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]

NemoClaw-Upgrade-E2E: approve e977a2f 82a1a8f

[github-actions]

PR Review Advisor finished for commit e977a2f. 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]

NemoClaw-Upgrade-E2E: approve 9902508 82a1a8f

[github-actions]

PR Review Advisor finished for commit 9902508. 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]

NemoClaw-Upgrade-E2E: approve be981db 82a1a8f

[github-actions]

PR Review Advisor finished for commit be981db. 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]

NemoClaw-Upgrade-E2E: approve 4c6560b 82a1a8f

[github-actions]

PR Review Advisor finished for commit 4c6560b. 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]

NemoClaw-Upgrade-E2E: approve 675c1e1 82a1a8f

[github-actions]

PR Review Advisor finished for commit 675c1e1. 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