A mostly-fixes release with two additions worth knowing. The new bits are a Dynamic workflow size control in /config that nudges how large Claude makes dynamic workflows (small / medium / large agent counts) — an advisory guideline, not an enforced cap — and two OpenTelemetry attributes, workflow.run_id and workflow.name, so a workflow run's activity can be reconstructed from your OTel data. The rest is bug fixes:
- Remote Control (mobile/web): commands sent into an interactive session no longer fail with "Unknown command"; images and files sent without a caption are no longer silently dropped; and
/remote-control sessions no longer show the wrong permission mode in the mobile and web apps.
- Workflows: scripts with unicode quote escapes are no longer corrupted before parsing, and a workflow parse error now points at the offending line instead of always blaming TypeScript.
- Sessions:
/rename on a background session is no longer reverted when the job restarts, and resuming a session by name is more reliable. Opening a chat from claude agents no longer trips a "currently running as a background agent" worker crash/respawn loop.
- Reliability: fixed transient mTLS handshake failures during in-place client-certificate rotation, an unbounded voice-dictation retry loop when the mic/recorder fails, a crash in the inline Ctrl+R history search when accepting or cancelling mid-scan, and a non-clickable
claude auth login sign-in URL when it wraps over SSH.
Who should care: anyone driving Claude Code from the Remote Control apps or leaning on dynamic workflows. Nothing here is a breaking change — it's an incremental reliability release on the 2.1 line.
Release notes →
Affects: /openclaw/, /openclaw/setup/, /openclaw/configuration/, /claude-cowork/
2026-07-07
NemoClaw
main · onboarding & inference sweep
A day of onboarding and local-inference reliability fixes on main. If you run NemoClaw with local GPUs or hit a snag partway through onboard, several of these will matter:
- Local vLLM tool calls fixed (#6326): the generic-Linux managed-vLLM default,
nvidia/NVIDIA-Nemotron-3-Nano-4B-FP8, launched without a tool-call parser, so agent requests using tool_choice: "auto" failed with HTTP 400 while plain completions worked. It now pins --enable-auto-tool-choice --tool-call-parser qwen3_coder, matching NVIDIA's model-card recipe. Reproduced on Ubuntu 24.04 / RTX 5070.
- Onboarding recovery is discoverable (#6008): when
nemoclaw onboard fails or is interrupted mid-step, it now prints that you can resume with nemoclaw onboard --resume (completed steps are skipped) instead of leaving you to assume a full reinstall.
- Spurious plugin warning removed (#6007): fresh sandboxes no longer print
plugins.entries.qqbot: plugin not installed on first launch — the unbundled qqbot entry was dropped from the default openclaw.json.
- arm64 NIM warning (#5868): Linux arm64 DGX Spark/Station users now get an advisory that some NIM images may not publish
linux/arm64 manifests, while the pull/start path still tries the selected image.
- Custom Anthropic endpoint streaming validation (#6297): custom Anthropic-compatible endpoints are now checked with a streaming probe at onboarding, so a malformed SSE layer is diagnosed up front instead of failing cryptically at runtime.
- Housekeeping:
nemoclaw gc now removes orphaned locally-prebuilt sandbox images (#6306); legacy pre-fingerprint sandboxes are rebuilt only after a full backup and exact-name confirmation (#6362); and nemohermes <sandbox> sessions now routes to the sandbox's own agent binary, fixing openclaw: command not found on Hermes sandboxes (#6249).
Who should care: local-GPU NemoClaw users (the vLLM tool-call fix in particular) and anyone who has been re-running onboard from scratch after a hiccup. These are commit-level fixes on main, not a tagged release — NemoClaw versions from the changelog, not version tags.
The bulk of NemoClaw's day — roughly a dozen commits (#6358, #6357, #6359, #6361, #6363, #6365, #6368, #6373, #6383, #5598, plus the #6385 docs-link fix) — was internal test-harness work: centralizing E2E fixtures, path helpers, and command/environment profiles, and cutting subprocess isolation to speed up integration suites. None of it changes command, config, or runtime behavior.
NemoClaw commits →
Affects: /nemoclaw/, /nemoclaw/setup/, /nemoclaw/local-gpu/, /nemoclaw/switching-providers/
OpenAI released GPT-Realtime-2.1 and a lower-cost mini for the realtime API. The updated realtime reasoning model brings improved alphanumeric recognition (think spelling out order numbers or codes over voice), better silence and noise handling, and improved interruption behavior. GPT-Realtime-2.1 mini is a faster, cheaper distilled reasoning variant for latency- or cost-sensitive realtime apps.
Who should care: developers building voice or low-latency streaming agents on OpenAI's v1/realtime endpoint. This is an API-model update, so it doesn't change the ChatGPT app experience — but if you price realtime workloads, the mini is worth a look. We'll fold pricing specifics into /chatgpt/pricing/ once OpenAI publishes rates.
OpenAI changelog →
Affects: /chatgpt/, /chatgpt/pricing/, /chatgpt/api-vs-chat/