Changelog — July 23, 2026

A broad fix-and-polish release. The three that change how the tool feels day to day:

• /code-review now runs as a background subagent. Review work no longer fills your conversation, and it keeps stacked slash commands as its review target. If you review often, your main transcript stays readable.
• Windows \u-path corruption fixed. Paths with a \u-prefixed segment (like C:\Users\unicorn) were being mangled into CJK characters in tool inputs, making those files inaccessible. See the callout above — this one bit a lot of Windows users silently.
• The left arrow key no longer discards your conversation with no undo. A press right after editing now asks to confirm, and Esc in the agent view returns you to the conversation it backgrounded.

The rest of the batch:

• MCP connection errors are visible now. claude mcp list and /mcp show the HTTP status and error text when a server fails to connect, plus a warning when an MCP config value has hidden leading or trailing whitespace — a classic silent-failure cause.
• /ultrareview accepts descriptive arguments. Passing something like "review my auth changes" no longer fails; it runs a review of your current branch with the text applied as a note to the findings.
• /code-review ultra launches the cloud review in non-interactive sessions instead of silently running a local review — so scripted/CI invocations actually get the billed cloud agents you asked for.
• Multi-line paste stops collapsing into one line in terminals that encode pasted newlines as Ctrl+J (previously you'd get a j where each newline should be).
• /context reports accurate token usage after compaction — it was showing stale pre-compact numbers after you compacted from the message picker.
• Gateway spend metering prices mapped model IDs correctly. Bedrock application-inference-profile ARNs and other config-mapped upstream model IDs are now billed at the configured model's rates.
• Accessibility: in --ax-screen-reader mode, word and line deletions (Option+Delete, Ctrl+W, Cmd+Backspace, Ctrl+U, Ctrl+K) now announce the deleted text.
• Assorted: fixed mojibake when a long IDE selection was truncated mid-emoji, a case where a tool-executor error could be silently dropped, and an engine teardown race.

Who should care: Windows users should upgrade for the path fix alone. Anyone who leans on /code-review or runs it in CI benefits from the backgrounding and the non-interactive ultra fix.

Claude Code v2.1.218 release notes → Affects: /openclaw/, /openclaw/setup/, /openclaw/configuration/, /openclaw/cost-optimisation/, /claude-cowork/, /claude-cowork/setup/

A documentation and CI day — the user-facing security substance shipped yesterday. The v0.0.92 changelog is now written (#7390), rolling up the ten PRs merged after v0.0.91: the OpenClaw 2026.7.1 + Node.js 22.23.1 security update (covered in yesterday's entry), the Jaeger header remediation regression coverage, WSL and macOS test stabilization, and the deployment-doc work below. The missing v0.0.91 changelog was also backfilled (#7376).

The change most likely to affect you if you run NemoClaw remotely: remote deployment is consolidated onto one provider-neutral headless Linux server guide over SSH (#7346). It covers OpenClaw, Hermes, and Deep Agents Code with unattended installation, loopback-only access, readiness checks, and manual recovery after a host reboot. The Brev deployment and web-UI pages are retired, with their URLs redirected to the shared guide — Brev is now described as just one way to provision a headless server, with NemoClaw setup depending on neither Brev nor its web UI. The three Quickstarts (OpenClaw, Hermes, Deep Agents) were also aligned so each keeps first-run Docker/Windows/DGX Station choices while linking durable detail to canonical owners (#7333).

The rebuild --force data-loss and base-image override constraints are now documented (#7389). Everything else is internal: a documentation-writer receipt CI pilot now extends to docs-only PRs (#7400), plus E2E gating, protected-environment approval delegation, and an existing-Station vLLM ownership handoff that never stops or reuses your running workload (#7285).

Who should care: if you were following the Brev remote-GPU deployment path, note it's retired — the new headless-server guide is the supported route. Otherwise nothing here requires action.

NemoClaw commits → Affects: /nemoclaw/, /nemoclaw/setup/, /nemoclaw/local-gpu/, /nemoclaw/switching-providers/

The API side is where the substance is. Claude's Managed Agents gained four things on July 22:

• An effort level on the agent's model configuration. Pass effort inside the model object when you create the agent to tune how hard the model works per turn — the same effort control Claude Code exposes, now on server-hosted agents.
• Environment and memory-store lifecycle webhooks. Four environment.* and three memory_store.* event types let you react to those lifecycle changes without polling.
• Session seeding. POST /v1/sessions now accepts initial_events — up to 50 user.message and user.define_outcome events. A non-empty list starts the agent loop in the same call, so you skip the separate send-events request.
• Optional version on agent update. Supply it for optimistic concurrency (a mismatch returns 409), or omit it to apply the update unconditionally.

On the Claude apps side, the one new item since our last check: HIPAA configuration for your organization is now self-serve (July 14) — admins can manage HIPAA readiness independently instead of routing through support.

Who should care: anyone building on Claude's Managed Agents API — effort levels and lifecycle webhooks are both real capability additions, not fixes. The HIPAA change matters only to org admins with a compliance requirement.

Claude API release notes → Affects: /claude-cowork/, /claude-cowork/vs-api/, /claude-cowork/pricing/

OpenAI added hard spend limits for organizations and projects on the API platform. Set a monthly cap and affected API requests return an HTTP 429 once tracked spend reaches the limit — a real ceiling, not just a notification. Spend alerts remain available to warn you before traffic is interrupted, so the recommended setup is an alert below the hard cap.

Who should care: anyone running production ChatGPT/OpenAI API workloads who has wanted a guaranteed budget ceiling rather than after-the-fact billing. Note the failure mode: at the cap, requests fail — size the limit with headroom, and pair it with an alert.

OpenAI platform changelog → Affects: /chatgpt/, /chatgpt/pricing/, /chatgpt/api-vs-chat/