Changelog — June 11, 2026

Fable 5 model-name normalization. Model names that include a [1m] suffix — used in some configurations to signal 1M context — are now automatically stripped during normalization. Fable 5 includes 1M context by default, so the suffix is redundant and was causing lookup failures when the raw string was passed to the API.

Windows sandbox warning squashed. A spurious "sandbox dependencies missing" startup warning was appearing on Windows even when sandbox was correctly enabled in settings. The false positive is now suppressed.

Release notes → Affects: /openclaw/, /openclaw/setup/, /openclaw/configuration/

Recursive sub-agent spawning (up to 5 levels). Sub-agents can now spin up their own sub-agents, and those sub-agents can do the same — up to 5 levels deep. This unlocks complex nested orchestration patterns where a top-level agent delegates to specialized workers, each of which can further delegate without workarounds.

Stuck 1M-context sessions now auto-compact. Sessions using 1M context without sufficient usage credits were getting permanently stuck with no way to recover. OpenClaw now automatically compacts the conversation back under the standard context limit so the session can continue.

Amazon Bedrock region auto-detection. When AWS_REGION is not set as an environment variable, OpenClaw now reads the region from ~/.aws/config following standard AWS SDK precedence. The /status command shows which source the region came from.

Plugin marketplace search bar. The /plugin marketplace browser now has a search bar, making it easier to find plugins in larger registries without scrolling through the full list.

Background agent fixes. Several background-agent bugs are resolved: images in multi-image conversations no longer trigger a repeating "image could not be processed" error; agents no longer stay stuck as "Working" for up to 30 seconds after completing; background agents no longer accidentally read another directory's project settings (.mcp.json approvals, trust) when dispatched onto a pre-warmed worker; background-session attach no longer fails with EAUTH for sessions started on an older version after a daemon auto-update; nested sub-agents stopped by the user no longer stay stuck as "active" in the agent panel.

Model picker fixes. /model suggestions in the agent dispatch input no longer render with a misleading slash prefix or show models disabled for your org. availableModels allowlist restrictions are now properly applied to sub-agent model overrides, the agent dispatch model picker, and the advisor model. Allowlists using version-specific IDs like claude-opus-4-8 no longer incorrectly hide the Opus and Sonnet 1M rows in the picker.

Release notes → Affects: /openclaw/, /openclaw/setup/, /openclaw/skills-guide/, /openclaw/configuration/, /openclaw/cost-optimisation/, /claude-cowork/, /claude-cowork/setup/, /claude-cowork/pricing/

NemoClaw merged a large batch of commits today. The majority are internal refactors — extracting onboarding logic (provider selection, sandbox registration, messaging token prep, dashboard port resolution, Docker gateway runtime helpers) into focused, testable modules as part of the ongoing onboard.ts cleanup. No user-visible behavior changes from the refactors. Four fixes and one docs correction are user-facing:

OpenClaw scope-upgrade approval recovery (#5210): When OpenClaw replaces a pending scope-upgrade approval with a same-device request that includes operator.admin, NemoClaw's approval wrapper was returning failure even though the new request was valid. The fix narrows the post-failure recovery path to detect and handle these same-device replacement requests correctly, merging scopes and approving only the expected operator scope closure.

Gateway healthcheck marker fix (#4710): Docker-driver sandbox containers were being marked (unhealthy) on every fresh onboard because the in-container gateway healthcheck marker was gated on an OPENSHELL_DRIVERS=docker env variable that OpenShell never exports into the sandbox. The marker is now written immediately before each gateway launch — true-by-construction — so containers show as healthy after onboarding as expected.

WSL and macOS CLI stability (#5204, #5209): Public CLI routes are now dispatched by registered oclif command ID instead of re-resolving through native argv, fixing CLI dispatch failures on WSL. Gateway process identity is now matched by command token or exact path instead of substring checks, resolving false positives on macOS and WSL.

Security docs correction (#5099): The NemoClaw security best-practices page incorrectly described four protection layers when the documentation actually covers five (network, filesystem, process, inference, and gateway authentication). The intro, Mermaid diagram, and at-a-glance table now consistently say five layers. The Sandbox Hardening link was also pointing at the wrong path and is corrected.

Commit log → Affects: /nemoclaw/, /nemoclaw/setup/, /nemoclaw/policy/, /nemoclaw/local-gpu/, /nemoclaw/switching-providers/, /nemoclaw/skills/

Anthropic apps changelog updated. Anthropic's Claude apps release notes page was updated today. The release monitor detected the change; check the official page for the full details of what was added.

Release notes → Affects: /claude-cowork/, /claude-cowork/projects/, /claude-cowork/skills-guide/, /claude-cowork/skills-database/