Changelog — June 9, 2026

Claude Fable 5 available now. OpenClaw v2.1.170 introduces Claude Fable 5 — a Mythos-class model described by Anthropic as the most capable model they have ever made generally available. Update to v2.1.170 (npm update -g @anthropic-ai/claude-code) to unlock access. See Anthropic's announcement for full capability details.

Transcript save fix for VS Code users. Sessions launched from the VS Code integrated terminal — or any shell that inherited Claude Code environment variables — were silently failing to save transcripts, which meant those sessions did not appear in --resume. This regression is fixed in v2.1.170.

Release notes → Affects: /openclaw/, /openclaw/setup/, /openclaw/configuration/, /claude-cowork/

Safe-mode troubleshooting flag. --safe-mode (or CLAUDE_CODE_SAFE_MODE=1) starts Claude Code with all customizations disabled — CLAUDE.md files, plugins, skills, hooks, and MCP servers — making it easy to isolate whether a problem is caused by your configuration or a core bug.

New /cd slash command. Move your session to a different working directory mid-session without breaking the prompt cache. Previously you had to end and restart the session, losing cache warm-up.

disableBundledSkills setting. Operators and enterprises can now hide all bundled skills, workflows, and built-in slash commands from the model via the new disableBundledSkills setting or CLAUDE_CODE_DISABLE_BUNDLED_SKILLS env var — useful for locked-down deployments.

Bug fixes:

• Up/Down arrows in long inputs — now step through visual rows of a wrapped multi-line input first before jumping to command history, matching expected terminal behavior.
• Enterprise MCP policy enforcement — allowedMcpServers/deniedMcpServers policies were not being enforced on reconnect, IDE-typed configs, --mcp-config servers during first install, or before remote settings loaded. Fixed. Cold starts for orgs without remote settings are also faster.
• macOS 30–50ms UI stall — a per-turn stall for macOS users logged in with claude.ai credentials is resolved.
• Windows claude -p slowness — claude -p was slow or appeared to hang on Windows while waiting for the slash-command/skill scan (regression in 2.1.161). Fixed.
• Remote Control OAuth reconnect — Remote Control getting stuck on "reconnecting" after a session resume when an OAuth token refresh coincided with reconnect is fixed.
• Windows Git Credential Manager popup — the "Connect to GitHub" popup appearing on Windows at startup when background git commands ran without cached credentials is suppressed.
• Custom statusline footer hints — hints like "esc to interrupt" were not showing for users with a custom statusline. Fixed.
• Stale remote session prompts — stale permission and dialog prompts no longer reappear on every re-attach to a remote session whose worker died while waiting on them.
• claude agents --json — was omitting blocked and just-dispatched agents. Fixed.

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

Terminal Bench scores in model details. Kilo Code v7.3.41 now shows Terminal Bench completion scores and per-attempt costs directly in the model detail pane for supported models — making it easier to compare model efficiency without leaving the IDE.

Cloud session import fix. Filesystem changes from synced session diffs are now correctly restored when importing forked sessions, including changes inherited across a chain of session forks. This closes a gap where cloud session imports could miss earlier file edits.

Patch fixes: Agent-manager model sync on config change is fixed (#10094). A pointer cursor is now shown for clickable controls and links inside Kilo webviews.

Release notes → Affects: /kilocode/, /kilocode/setup/, /kilocode/models/, /kilocode/orchestrator/

NemoClaw merged 16 commits on June 9 — one new CLI command and 15 onboarding and sandbox reliability fixes addressing real-world failure reports.

New: nemoclaw <name> agents list (#5043): The host-side agent lifecycle CLI is now complete. agents add and agents delete shipped earlier; agents list passes through to openclaw agents list via openshell sandbox exec and returns the agent listing in --json format. Use it to confirm which agents are active inside a named sandbox without connecting interactively.

Ollama onboarding on Windows and minimal Linux (#5012): waitForPort() previously shelled out to nc -z to probe TCP ports. On hosts where netcat is not installed — minimal Linux distros (CachyOS and similar) and Windows — every probe read as "port not ready," causing Ollama onboarding to abort with a misleading "proxy did not become ready" error even when the proxy started fine. The fix falls back to a short-lived Node.js subprocess to probe the TCP connection, which is always available on any NemoClaw host. Behavior on hosts with nc is unchanged.

WhatsApp QR code compact rendering (#4522): The WhatsApp channel pairing QR was still rendering at full size (~56 rows), overflowing the terminal and making it impossible to scan. The root cause was that the bundled @openclaw/whatsapp plugin calls the qrcode package's toString — not qrcode-terminal, which an earlier fix had patched. The preload script now patches the correct renderer, forcing small: true for terminal renders regardless of what plugin version is installed.

CUDA on Jetson Tegra (#4231): On Jetson Orin, the sandbox could see GPU devices but CUDA failed with cuInit(0)=999 because the sandbox user was not a member of the video group that owns /dev/nvmap. NemoClaw now detects Tegra device node GIDs on the host and passes them as --group-add flags during sandbox recreate, so CUDA actually initializes. nemoclaw status now reports "(CUDA verified)" instead of a misleading bare "enabled".

tmux PTY allocation in OpenClaw sandbox (#4513): OpenClaw's tmux-session flow failed with "create window failed: fork failed: Permission denied" because the OpenShell sandbox landlock policy granted /dev/null and /dev/urandom but not /dev/pts. forkpty() opening /dev/ptmx was denied. The policy now grants /dev/pts, PTY allocation works, and the E2E test is restored to a hard assertion.

Security: dashboard port 8642 rejected at host level (#4984): NemoHermes onboarding previously accepted NEMOCLAW_DASHBOARD_PORT=8642 (the port reserved for the Hermes OpenAI-compatible API) and proceeded to build a sandbox. A host-side guard now rejects that port before any sandbox is created, emitting a clear error message matching the in-sandbox guard.

GPU sandbox local inference routing (#4509): On GPU host-network setups (NEMOCLAW_DOCKER_GPU_PATCH_NETWORK=host), onboarding reported "local inference reachable" but the agent then failed with ECONNREFUSED. The probe was running against a recreated container whose network namespace was the host's — but OpenClaw runs in OpenShell's isolated sandbox network and cannot reach the host loopback. The fix downgrades the host-network GPU patch to the OpenShell bridge so inference routes through inference.local, and the post-ready reachability gate now probes via openshell sandbox exec — the exact path the agent uses.

OpenClaw config permissions after doctor --fix (#4538): Running openclaw doctor --fix directly inside a mutable sandbox was tightening /sandbox/.openclaw permissions back to OpenClaw's single-user 700/600 layout, which locked the gateway UID out of persisting config. The always-on in-sandbox openclaw() guard function now re-asserts the NemoClaw mutable contract (2770 setgid dir, 660 config) after every routed openclaw command, even when doctor exits nonzero.

Hermes provider visibility in config.yaml (#4972 / #4973): The Hermes sandbox config.yaml now includes a _nemoclaw_upstream block recording the actual upstream provider name and model, alongside a human-readable YAML comment header. Previously, only Hermes' proxy-routing provider: custom tag appeared, making it hard to see which upstream route was in use. The onboard menu header is also renamed to "Select your inference provider:" and the Hermes Provider entry now lists the Nous portal model families.

Cleaner sandbox creation output (#5030): Fresh sandbox creation no longer prints repeated sandbox not found errors before the real create step. Provider cleanup probes during onboarding were not suppressing expected missing-sandbox output; this is now fixed.

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

Web search returns image results. The v1/responses web search tool can now return image results alongside regular text results. Use image search when your application needs current or visually-grounded content such as product photos, landmarks, places, events, or visual references. See the OpenAI web search guide for usage details.

OpenAI changelog → Affects: /chatgpt/, /chatgpt/tips/, /chatgpt/custom-gpts/