An all-bugfix release centered on background agents, subagents, and streaming reliability. If you delegate work to subagents or run background agents, this is worth taking — several fixes stop failures from being silently swallowed:
- Subagents no longer fail silently: a subagent cut off by a rate limit or server error now returns its partial work to the parent instead of vanishing, and a subagent that hit an API error (e.g. usage limit reached) is now reported as an error to the parent rather than as a successful result.
- Partial streams are kept: when the API emits a mid-stream overloaded/server error after partial output, the partial response is retained with an incomplete-response notice instead of being discarded.
- Stacked slash-skills load: invocations like
/skill-a /skill-b do XYZ now load all leading skills (up to 5), not just the first.
- SSL errors fail fast: TLS-inspecting proxies, a missing
NODE_EXTRA_CA_CERTS, or expired certs now fail immediately with an actionable fix hint instead of burning through retries first.
- Linux daemon self-kill loop fixed: the background-agent daemon no longer kills itself and every running agent every ~50 seconds after an unclean shutdown left a corrupted worker record.
More fixes: background agents failing to cold-start over SSH on macOS ("Could not switch to audit session," a 2.1.196 regression); claude stop being silently undone when it raced a background-agent respawn; stalled background-job progress indicators; a clearer low-memory error on starved machines; remote sessions flapping between Working and Idle; idle subagents vanishing from the agent panel (surplus idle agents now collapse into an expandable row); and typing /model or /fast while viewing a subagent silently opening the lead's model picker.
Release notes →
Affects: /openclaw/, /openclaw/skills-guide/, /openclaw/configuration/, /claude-cowork/, /claude-cowork/skills-guide/
Auto-named Agent Manager branches, plus session-cleanup fixes. The one minor feature: Agent Manager branches now automatically get task-focused names once the conversation establishes a clear workstream, with optional prefixes and a setting to disable automatic naming — so parallel agent branches stop being a wall of generic names. The patch fixes are all about session state:
- Webview memory freed for deleted sessions: unsent prompt text, review comments, and pending image attachments retained in the per-session draft cache are now cleared when a VS Code session is deleted.
- In-flight drafts survive a delete: a failed draft that was mid-send is restored into the live prompt after a session is deleted (whether user-initiated or via an external CLI/TUI/cascade delete) — but it is never rehydrated into a prompt you explicitly cleared.
- Chat stays in sync: when the backend connection is briefly unavailable, switching Agent Manager sessions now updates the conversation too, not just the side diff.
Release notes →
Affects: /kilocode/, /kilocode/orchestrator/, /kilocode/setup/
A batch of user-facing CLI and messaging fixes landed on main. Two are worth upgrading for on their own — a tunnel gap that broke the Web UI, and a rebuild path that could permanently lose a sandbox:
- Tunnel Web UI now works out of the box (#6235):
nemoclaw tunnel start printed the public *.trycloudflare.com URL but never registered that origin in the gateway allowlist, so every Web UI WebSocket through the tunnel was rejected with "origin not allowed." The tunnel origin is now auto-registered (exact origin, never a wildcard) and the gateway restarted — no manual config.
- Rebuild no longer destroys a sandbox on credential conflict (#5954): a
channels start/rebuild that shared a messaging credential with another sandbox used to surface the conflict only after the original had been backed up and destroyed — so the rebuild aborted with the sandbox permanently lost. The conflict check moved into the rebuild Step-0 preflight, so it now aborts before any destructive backup/delete.
- Enabled channel policies now persist (#5967): Discord, Telegram, WhatsApp, Teams, and WeChat showed as not-applied (
○) in policy-list after onboard because finalization only re-merged Slack's create-time preset. Every enabled channel's policy preset is now applied to the live gateway and recorded in the registry.
destroy --force cleans up when the gateway is down (#6046): with the OpenShell gateway unreachable, --force now falls back to local cleanup (registry entry + local artifacts) instead of dead-ending on connection-refused.
Also on main: update --fresh now reinstalls even when already up to date (#5960); the WSL2 dashboard URL keeps 127.0.0.1 as primary with the WSL host IP as an explicit fallback (#6181); channels status surfaces Telegram's group mention mode (#6220); uninstall now removes the nemohermes/nemo-deepagents alias shims (#6098); user-error CLI surfaces now exit non-zero for scriptability (#5974); and messaging policy data moved into the channel package tree (#6129). The remaining commits were onboarding DNS-message clarity, a dcode snapshot-probe hardening, and CI/E2E maintenance.
Commit log →
Affects: /nemoclaw/, /nemoclaw/setup/, /nemoclaw/policy/, /nemoclaw/switching-providers/