Changelog — June 5, 2026

v2.1.163 is a significant feature release. Org administrators can now enforce a version range via requiredMinimumVersion and requiredMaximumVersion managed settings — Claude Code refuses to start if its installed version falls outside the allowed range and directs users to an approved version. This gives fleet managers a hard backstop against both regressions and unapproved upgrades.

Two new quality-of-life additions: /plugin list shows all installed plugins with optional --enabled/--disabled filters, making plugin inventory visible without digging through config files. The /btw answer panel now has a "c to copy" shortcut that copies the raw markdown to the clipboard, preserving formatting when pasted elsewhere.

Hooks gain meaningful power: Stop and SubagentStop hooks can now return hookSpecificOutput.additionalContext to feed Claude feedback and keep the turn alive without the response being flagged as a hook error. Skills pick up a \$ escape syntax for including a literal $ before a digit in command bodies. stdio MCP servers now receive the same CLAUDE_CODE_SESSION_ID as hooks and Bash when the session is resumed with --resume.

Key bug fixes: claude -p no longer hangs after its final result when a backgrounded command never exits — background shells are stopped ~5 s after the result once stdin closes. claude -p no longer fails with "ANTHROPIC_API_KEY required" on Bedrock/Vertex/Foundry in CI when CI=true and no Anthropic key is set. A regression introduced in v2.1.154 that overrode $TMPDIR for all commands (not just sandboxed ones) is fixed, restoring compatibility with bazel and EDR-protected Go workflows. Windows users on OneDrive or read-only paths no longer see "EEXIST: file already exists" errors on the session-env directory. Org-managed permission rules now apply correctly for the full session even when the managed settings fetch completes during startup on a fresh config directory. Background sessions in claude agents no longer lose their running background tasks on reconnect.

v2.1.165, released the same day, ships additional bug fixes and reliability improvements.

v2.1.163 release notes →  ·  v2.1.165 release notes → Affects: /openclaw/, /openclaw/setup/, /openclaw/skills-guide/, /openclaw/configuration/, /openclaw/security/, /openclaw/cost-optimisation/, /claude-cowork/, /claude-cowork/setup/, /claude-cowork/skills-guide/, /claude-cowork/system-prompts/, /claude-cowork/pricing/

macOS Apple Silicon users who found the Kilo Code CLI failing to start should update to v7.3.33, which fixes a malformed bundled exports issue that caused the CLI to crash on M-series Macs. Versions v7.3.30 through v7.3.32 contained no user-facing changes and were internal build-process iterations.

Release notes → Affects: /kilocode/, /kilocode/setup/, /kilocode/models/, /kilocode/orchestrator/, /kilocode/security/, /kilocode/vs-claude-code/

NemoClaw's v0.0.59 release notes are now published, covering CLI command updates, inference documentation, and credential storage changes. Two new read-only network egress presets ship with this wave: weather (public geocoding and weather APIs) is added to the balanced tier by default, and public-reference (curated reference APIs) is included in the open tier. Hermes open policy also now auto-includes all Hermes Nous managed-tool presets while keeping OpenClaw open defaults agent-specific — so agent-appropriate egress is configured without manual preset selection.

Inference routing: nemoclaw inference set and nemohermes inference set now correctly resolve the managed inference API family before patching in-sandbox config. This fixes a regression where switching to an Anthropic Messages–compatible provider left the agent calling the wrong route and hitting a 403 connection not allowed by policy. The fix covers both Hermes and OpenClaw runtime-switch paths.

Runtime display fix: NemoClaw's internal runtime instructions (<nemoclaw-runtime> block containing sandbox policy) were leaking into the visible chat UI on the third message turn with models like Nemotron 3 Super 120B via Ollama. The fix switches from prependContext to prependSystemContext so these instructions inject into the system prompt, invisible to users.

GPU patch hardening: The Docker GPU patch reconnect window is extended from ~10 s to ~30 s of sustained error tolerance before failing. If reconnect still fails, a new rollback helper stops the new GPU container and restarts the backup — previously a failed reconnect left users with no running container at all. The non-root rc shim cleanup also exits cleanly when it cannot rewrite a root-owned rc file instead of crashing the container.

DGX Spark vLLM: The DGX Spark profile now uses the stable NGC release nvcr.io/nvidia/vllm:26.05.post1-py3 instead of the upstream nightly build, reducing the risk of unexpected breakage from nightly regressions. Nemotron 3 Ultra 550B-A55B (550B total / 55B active parameters, up to 1M context) is now available as a second curated NVIDIA Endpoints model option alongside the default Nemotron 3 Super 120B.

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

OpenAI's omni-moderation-latest model now surfaces moderation scores directly inside Responses API and Chat Completions API responses. Pass a moderation object in a generation request to receive moderation results for both the model input and generated output in the same response — no separate moderation API call required. This simplifies content-safety workflows by collapsing two round-trips into one.

Platform changelog → Affects: /chatgpt/, /chatgpt/pricing/, /chatgpt/custom-gpts/, /chatgpt/tips/