Sandbox credential blocking, org-level model restrictions, and a batch of connection-reliability fixes. The changes most likely to affect your setup:
- New
sandbox.credentials setting (security): block sandboxed commands from reading credential files and secret environment variables — a useful guardrail when you run untrusted commands or agents in the sandbox.
- Org-configured model restrictions: the model picker,
--model, /model, and ANTHROPIC_MODEL now honor org policy, showing a "restricted by your organization's settings" message when a blocked model is selected.
- Remote MCP tool calls no longer hang silently: calls that previously blocked with no response for 5 minutes now abort with an error. Tune the threshold with
CLAUDE_CODE_MCP_TOOL_IDLE_TIMEOUT.
- Structured-output reliability: with
--json-schema and workflow agent({schema}), the model can no longer re-call StructuredOutput indefinitely after a successful call, and follow-up turns now reliably return structured output. Background jobs that ended a turn without structured output no longer get stuck in "working" forever.
- Other fixes: mouse-click support in fullscreen select menus (permission prompts,
/model, /config); --resume no longer fails with "No conversation found" when the original -p run produced no model turns; pasted Korean/CJK text no longer turns into mojibake; channel connections no longer drop after navigating to the agents view or running /bg, /tui, or /update; and agent stop notifications now correctly attribute who stopped the agent ("finished"/"stopped").
Release notes →
Affects: /openclaw/, /openclaw/security/, /openclaw/configuration/, /claude-cowork/, /claude-cowork/setup/
A new release, plus a meaningful change for agents that read NemoClaw's docs. v0.0.66 release notes landed (#5679). The standout for our readers:
- Skill catalog consolidation (#5699): NemoClaw retired the generated
docs-to-skills pipeline and the expanded per-topic nemoclaw-user-* skill trees, replacing them with a single compact nemoclaw-user-guide skill. Instead of bundling docs into many generated skills, it routes AI agents to canonical NemoClaw documentation through the docs MCP server, llms.txt, and variant-specific Markdown pages. Note: the old llms-full.txt path is deprecated in favor of specific .md routes. If you were pinning the old generated user skills, switch to the single guide skill.
- Discord proxy fix (#5544): NemoClaw stopped emitting a non-loopback per-account Discord proxy that OpenClaw's Discord plugin rejects by design (it requires a loopback host). Discord egress now relies on the top-level managed proxy again, so the gateway WebSocket can reach READY. Telegram keeps its per-account proxy.
- LangChain Deep Code routing label (#5694): adds
integration: dcode as the canonical triage label for LangChain Deep Code work — a small follow-on to the experimental Deep Agents Code harness from yesterday.
- CLI fix: dropped the unsupported
--non-interactive flag from openclaw agents delete (#5656). The rest of the day was internal onboarding-runtime refactors and E2E test hardening.
Commit log →
Affects: /nemoclaw/, /nemoclaw/skills/, /nemoclaw/policy/