NemoClaw v0.0.79 focuses on reliability at the edges where agents meet real machines: provider selection, local inference startup, gateway routing, sandbox cleanup, and protected runtime state. The release adds first-class OpenRouter onboarding, safer managed vLLM startup, clearer route-drift reporting, and more predictable recovery when onboarding or uninstall flows encounter partially preserved state.

This release also tightens the sandbox boundary. Deep Agents Code now runs with enforced process and file-descriptor limits, credential-shaped OTLP URLs avoid false-positive blocking without weakening secret checks, proxy environment setup stays POSIX-compatible, and TOML config handling fails closed when input is ambiguous. Network policy docs and tests now make GitHub, Gmail, and custom MCP egress expectations more explicit.

The release train itself is more auditable. PR risk planning, post-merge E2E shadow evidence, advisor ledger handling, generated docs variants, and release-note preparation all received targeted hardening so maintainers have better evidence before deciding whether a tag is ready.

Inference, Providers, and Local Models

• #6461 adds OpenRouter as a first-class onboarding provider for OpenClaw, Hermes, and LangChain Deep Agents Code. This gives users a supported hosted inference path with OpenRouter credential validation, model catalog handling, and routing through NemoClaw's managed OpenAI-compatible gateway.
• #6523 runs managed vLLM Docker containers with validated argument vectors instead of shell-composed strings. This reduces quoting risk, keeps sensitive token values in the subprocess environment, and makes managed local inference startup more reliable.
• #6551 serves the DGX Spark Qwen3.6 managed vLLM profile with the qwen3_coder tool parser. This aligns the Spark recipe with the model's tool-calling expectations so local inference can handle agent tool calls correctly.
• #6484 bounds managed OpenClaw inference compaction. This keeps oversized context handling from growing without limit and makes long-running managed inference behavior more predictable.
• #6488 calibrates provider validation timeouts during onboarding. This prevents slow but valid providers from failing too early while still bounding waits for unreachable endpoints.

CLI, Gateway, and Install Diagnostics

• #6271 adds $$nemoclaw completion for Bash, Zsh, and Fish. This improves day-to-day operator ergonomics by exposing shell completions for installed commands, flags, and local sandbox names.
• #6272 adds structured logging through --debug and NEMOCLAW_LOG_LEVEL. This gives users and support engineers a consistent way to collect more detailed diagnostics without changing command behavior.
• #6465 warns when the live OpenShell gateway route differs from the sandbox route recorded by NemoClaw. This helps operators detect out-of-band route changes and realign inference state through supported commands.
• #6570 recommends and runs openshell gateway remove consistently during gateway cleanup. This makes teardown guidance match the supported OpenShell command across platforms.
• #6539 reports orphaned sandboxes honestly across uninstall and reinstall flows. This reduces confusion when user data is preserved and a sandbox still exists outside the active registry path.
• #6528 uses my-assistant as the DGX Spark express-install sandbox name. This keeps the express path aligned with the documented first-run experience.

Sandbox Runtime, Configuration, and Recovery

• #6559 enforces process and file-descriptor limits for managed LangChain Deep Agents Code sandboxes. This narrows resource-exhaustion risk and brings the Deep Agents runtime in line with the sandbox hardening posture.
• #6538 allows plain OTLP endpoint URLs past the Deep Agents secret guard. This avoids blocking legitimate observability configuration while preserving credential-shaped value checks.
• #6560 keeps proxy environment exports POSIX-compatible. This prevents shell portability problems in sandbox startup and messaging flows that depend on managed proxy variables.
• #6568 supports nested OpenShell PID namespaces in config guards. This prevents supervised runtime checks from misclassifying valid nested OpenShell environments.
• #6552 parses Deep Agents TOML config through sandbox tomllib for config get. This makes config reads match the runtime parser instead of trusting ad hoc host-side interpretation.
• #6567 hardens TOML config handling. This improves fail-closed behavior when configuration content is malformed, ambiguous, or unsafe to apply.
• #6587 reconciles registered extra providers during onboarding. This removes stale provider state so the sandbox registry reflects the currently selected provider set.

Onboarding, Policy, and Messaging

• #6253 makes onboarding finite-state-machine resume recovery a single explicit path. This reduces divergent recovery behavior after interrupted onboarding or rebuild sessions.
• #6572 preserves pending route reservations across onboarding resume. This keeps in-flight route choices from being lost when a user continues a partially completed flow.
• #6444 extracts sandbox create-failure reporting and create-step handling into focused modules. This makes onboarding failure recovery easier to reason about and test.
• #6536 avoids forcing plain BuildKit progress during onboarding. This keeps build output compatible with environments that prefer native BuildKit progress behavior.
• #5860 covers null-name resume sessions. This prevents malformed or canceled session state from sending users down the wrong recovery path.
• #6541 requires the GitHub policy preset for Homebrew Git egress. This keeps Git network access inside an explicit policy boundary instead of relying on broader package-manager allowance.
• #5415 adds a Gmail network-policy preset and integration example. This gives users a documented, narrower egress shape for Gmail-backed workflows.
• #6246 documents custom URL-based MCP server allowlists. This helps operators configure MCP egress explicitly instead of widening policy more than needed.
• #6496 reconciles WhatsApp post-pair gateway restarts without requiring operator.admin. This keeps channel activation within the intended runtime privilege boundary.

Docs, Advisor, and Release Confidence

• #6508 defines extension taxonomy and SDK readiness. This clarifies what counts as lifecycle contributions, managed packages, agent-native plugins, and future public SDK readiness gates.
• #6527 streamlines the OpenClaw, Hermes, and Deep Agents quickstarts. This reduces setup-page duplication and makes first-run guidance easier to follow.
• #6573 gates messaging-channel overview text to OpenClaw and Hermes variants. This prevents Deep Agents readers from seeing unsupported channel setup guidance.
• #6588 preserves list spacing in generated agent-variant docs. This keeps generated OpenClaw, Hermes, and Deep Agents pages readable after variant expansion.
• #5506 clarifies legacy k3s sandbox resources. This helps users distinguish current Docker-based behavior from older preserved documentation.
• #6446 adds deterministic PR risk plans. This gives maintainers a stable, commit-bound view of required E2E evidence for high-risk surfaces.
• #6447 adds post-merge E2E risk shadowing. This lets the trusted branch gather selective live evidence after merge without exposing secrets to unmerged code.
• #6547, #6566, #6591, and #6592 harden advisor session, retry, and ledger behavior. This makes automated review output more reliable and easier for maintainers to audit.
• #6582 guards queued Jetson E2E dispatches. This prevents selected Jetson runs from sitting in a queue where job timeouts do not apply before runner assignment.
• #6296 adds coverage for the OpenClaw TUI idle-timeout regression. This protects a real interactive chat path from regressing silently.
• #6586 enforces the Homebrew Git policy boundary in E2E. This verifies the policy tightening for Git egress in an end-to-end path.
• #6594 prepares the v0.0.79 release notes. This keeps the public documentation synchronized with the release tag.

Compare

Full diff: v0.0.78...v0.0.79

Thank you

Thank you to external contributors @HOYALIM, @Hokonoken, @Kunal-Somani, @atulya-singh, @deepujain, @fallintoplace, @glenn-agent, @harjothkhara, @kagura-agent, @sauravdev, and @tree586 for contributions included in this release.