# NemoClaw v0.0.53 is out

[cv]

NemoClaw v0.0.53 is about surviving the awkward middle of real machines. It does not introduce one grand surface area so much as it tightens the agreements between what the host says, what the gateway remembers, and what the sandbox can safely do.

Onboarding is the center of gravity. Local Ollama gets runtime context-window detection, a real no-autostart path, consistent no-tools overrides, healthier /api/tags probes, and better failure messages for non-interactive runs. Brave Search and OpenClaw pricing now use policy/provider paths that match how the sandbox actually rewrites outbound requests. WSL Docker Desktop, macOS Docker-driver sandboxes, Jetson bridge failures, and under-provisioned hosts all get sharper first-contact behavior.

The other theme is preservation. A stopped gateway no longer means lost PVC data; recreate now backs up workspace state before delete; backup and upgrade skip non-Ready sandboxes instead of aborting on them; gateway recreation moves out of read-only preflight; and shields, status, and config paths cross-check the filesystem they claim to protect.

Finally, release machinery grows teeth. Skills get a signed catalog export and the workflow fixes needed to actually generate, flatten, sign, and publish it. Scenario E2E gets isolated from legacy shell tests, profiling traces become CI artifacts, and the PR advisor becomes a sequence of focused turns. In a release with 44 merged PRs, the pattern is not more motion for its own sake. It is fewer places where the tool says "yes" while the machine means "not yet."

Onboarding, Inference, and Host Reality

• fix(inference): auto-detect Ollama context window during onboard #4253 auto-detects the loaded Ollama model context window during onboarding, while fix(onboard): gate Ollama auto-start behind --no-ollama-autostart #4212 adds --no-ollama-autostart / NEMOCLAW_OLLAMA_NO_AUTOSTART=1 so unreachable-Ollama paths can stay honest instead of silently spawning the daemon.
• fix(onboard): honor Ollama no-tools override during validation #4250 and fix(onboard): honor accepted no-tools Ollama override end-to-end (#4241) #4289 carry the accepted Ollama no-tools override through every validation boundary, so users who explicitly accept a tools-incompatible model are not looped back later by stricter probes.
• fix(onboard): non-fatal inference set for local Ollama and vLLM (#4257) #4291 makes failed openshell inference set calls recoverable for Local Ollama and vLLM, and fix(onboard): surface non-interactive provider failures with explicit exit messages #4219 adds explicit non-interactive provider failure messages where install logs previously ended with a bare exit.
• fix(inference): validate Ollama /api/tags response body in health probes #4295 validates the shape of Ollama /api/tags responses so a proxy, stale listener, or HTML stub cannot masquerade as a healthy backend.
• fix(onboard): drive Brave Search through OpenShell provider rewrite (#3626) #4225 routes Brave Search through an OpenShell provider rewrite for X-Subscription-Token, and fix(policy): add openclaw-pricing preset and auto-suggest for OpenClaw #4220 adds an openclaw-pricing policy preset so OpenClaw can fetch model pricing without noisy gateway warnings or zeroed cost records.
• fix(onboard): make timeout/killed DNS probes fatal and add Jetson bridge preflight #4227 makes timeout/killed DNS probes fatal and adds Jetson bridge preflight coverage; fix(onboard): use Docker --gpus instead of CDI repair on WSL Docker Desktop #4298 keeps WSL Docker Desktop on the Docker --gpus path instead of asking users to repair CDI inside the distro; fix(onboard): record Docker driver for macOS docker-driver sandboxes (#3728) #4221 records Docker-driver metadata correctly on macOS.
• fix(onboard): default to abort for under-provisioned resource warning #4273 changes the under-provisioned resource warning default to abort, and docs(inference): correct validation API paths #4249 corrects inference docs so provider validation paths match the implementation.

Sandbox State, Gateway, and Security Posture

• fix(onboard): preserve PVC across host stop/start by recovering stopped gateway (#4187) #4210 recovers a stopped gateway after host stop/start before any destructive cleanup, preserving PVC-backed workspace data.
• fix(onboard): back up workspace state before sandbox recreate #4197 backs up workspace state before sandbox recreate, even for not-ready sandboxes, and aborts on partial backup unless the user explicitly opts out.
• fix(maintenance): skip non-Ready sandboxes in pre-upgrade backup loop #4256 skips non-Ready sandboxes during backup-all and upgrade classification, removing a common upgrade blocker for stopped or errored sandboxes.
• fix(onboard): defer Docker-driver gateway recreation to step [2/8] #4254 defers Docker-driver gateway recreation from preflight to step [2/8], keeping preflight read-only and preserving the Ctrl+C escape window.
• fix(onboard): verify gateway container before reusing healthy state on port conflict #4306 verifies the gateway container still exists before reusing a healthy port-conflict state, so stale registry metadata is cleaned up instead of trusted.
• fix(shields,config): use direct privileged exec for VM-driver sandboxes #4287 routes shields and host-side config mutation through direct privileged exec for VM-driver sandboxes, while fix(shields): cross-check sandbox filesystem in shields status #4294 makes shields status compare declared state with the actual sandbox filesystem and re-lock drifted state correctly.
• fix(hermes): avoid false locked layout detection #4297 avoids false Hermes locked-layout detection when the root-owned directory is not actually protected by locked config files.
• fix(sandbox): refuse to start when bounding-set cap drop fails (#4264) #4266 briefly made residual dangerous Linux capabilities fail-closed, ci(e2e): opt functional jobs into residual caps on CI #4335 adapted CI around that posture, and fix(sandbox): restore residual capability warning behavior #4341 restores the release behavior to warning-only before the tag.
• fix(cli): preserve EROFS errno on share mount RO target #4315 preserves the EROFS signal for read-only share mount targets, fix(host-aliases): prepend docker exec subcommand for kubectl invocation #4319 fixes host alias commands by invoking docker exec correctly, and fix(cli): clarify live gateway drift in list #3907 clarifies when nemoclaw list is reporting live gateway drift.

OpenClaw, Messaging, and Day-Two Behavior

• fix(openclaw): repair Telegram placeholders and host-gateway web_fetch #4288 repairs Telegram placeholders and host-gateway web_fetch, addressing the OpenClaw integration path behind the Telegram silent-start candidate.
• fix(gateway): keep auto-pair watcher alive for late OpenClaw CLI scope upgrades (#4263) #4292 keeps the auto-pair watcher alive long enough for late OpenClaw CLI scope upgrades instead of leaving approval pending forever.
• fix(shields,config): use direct privileged exec for VM-driver sandboxes #4287 and fix(shields): cross-check sandbox filesystem in shields status #4294 also improve day-two config and shields operations by making the mutation path more explicit and the reported posture more verifiable.

Skills, Docs, E2E, and Release Confidence

• feat(skills): add signed catalog export #4284 adds a deterministic signed catalog export for NemoClaw skills; ci(skills): align catalog refresh with repo gh-CLI PR pattern #4334 moves the refresh workflow to the repo's gh CLI mutation pattern; fix(skills): flatten catalog export and detect untracked files in refresh diff #4342 detects untracked generated files and flattens the export layout; chore(ci): install NVSkills CI request listener for /nvskills-ci #4345 installs the /nvskills-ci listener so signing requests actually run.
• feat(skills): add nemoclaw-maintainer-verify-stale #3327 adds the nemoclaw-maintainer-verify-stale skill for reproducing old issues on fresh Brev hosts and tagging findings for maintainer review.
• feat(advisor): run PR review advisor as focused turns #4281 splits the PR Review Advisor into focused turns with inspectable prompt artifacts.
• test(e2e): isolate scenario suite assets #4283 isolates the hybrid scenario E2E framework under test/e2e-scenario, keeping legacy shell E2Es separate.
• perf(onboard): operationalize profiling traces #4094 operationalizes onboard profiling traces with opt-in trace files and CI artifact upload.
• fix(ci): remove unsupported Brev nightly inputs #4300 removes unsupported Brev nightly inputs, and test(e2e): update cloudflared tunnel pin #4321 refreshes the cloudflared tunnel pin used by tunnel validation.
• docs: refresh release notes and skills for v0.0.52 #4293 refreshes v0.0.52 release notes and generated skills, docs: remove prompt markers from Windows setup commands #3909 removes prompt markers from copyable Windows setup commands, and docs: clarify retired rcf patch validation #4202 clarifies that the old RCF patch validation script is intentionally retired.