Changelog — July 2, 2026

Claude in Chrome ships GA, and background agents get louder and more autonomous. The headline is that Claude in Chrome is now generally available. Alongside it, the claude agents background-agent workflow gained two big quality-of-life changes:

• Background-agent notifications: sessions that need input or finish now fire the Notification hook with agent_needs_input / agent_completed events — so you can wire desktop or webhook alerts when a delegated agent stalls or completes.
• Auto-PR on completion: background agents launched from claude agents now commit, push, and open a draft PR when they finish code work in a worktree, instead of stopping to ask.
• New /dataviz skill for chart and dashboard design guidance, with a runnable color-palette validator.
• Smarter delegation: the built-in Explore agent now inherits the main session's model (capped at Opus) instead of running on Haiku, and subagents plus context compaction now inherit the session's extended-thinking configuration — better output on delegated tasks.
• Gateway: added Claude Platform on AWS (anthropicAws) as an upstream provider, and model-not-found responses now advance the failover chain instead of dead-ending.

Notable fixes: transient network drops mid-response (ECONNRESET and friends) now retry with backoff instead of aborting the turn; agent-team teammates that die on an API error now report "failed" to the lead and can be woken to retry; background tasks in web/desktop/VS Code no longer get stuck on "Running"; the /diff panel refreshes on branch switch or outside commits; and AWS/Mantle sessions auto-refresh expired STS tokens instead of demanding /login.

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

"The Judgment Release" — every P0 and P1 in the repo is closed. Hermes v0.18.0 (calendar tag v2026.7.1) lands ~1,720 commits and 998 merged PRs since v0.17.0. The headline is a priority-backlog clean sweep: 100% of open P0 and P1 issues and PRs are now closed — roughly 700 highest-priority items, part of ~1,950 total issues/PRs resolved this window — and the team says it intends to hold P0/P1 at zero going forward. Beyond the cleanup, the release is about how well the agent reasons and how it knows when a job is actually done:

• Mixture-of-Agents is now first-class: named ensembles of models you can pick like any other model, with every reference model's reasoning shown and the aggregator's answer streamed live.
• Evidence-based self-verification: the agent now checks its own work against evidence instead of vibes.
• /goal completion contracts plus /learn and /journey turn self-improvement into something you can see and steer.
• Gateway at scale: scale-to-zero and drain coordination make it genuinely deployable in production.
• Desktop: first-class coding projects and a playable memory graph; subagents can now fan out in the background.

This is a major version step (0.17.0 → 0.18.0). If you self-host Hermes, review the gateway drain/scale-to-zero changes before upgrading, and expect new model entries in your picker for the Mixture-of-Agents ensembles.

Release notes → Affects: /hermes/, /hermes/setup/, /hermes/tasks/, /hermes/memory/, /hermes/mcp-tools/, /hermes/vps-install/

A security-heavy day, tagged as v0.0.72. NemoClaw landed a batch of SSRF and credential-hygiene fixes on top of a new channels status command. None of the security gaps were exploitable through the sandboxed agent — most required user-level CLI access — but they close real credential-boundary holes:

• Fail closed on DNS-backed HTTPS endpoints (#6139): DNS-backed HTTPS endpoints now fail closed before NemoClaw hands them to a downstream OpenShell/provider runtime, closing an HTTPS DNS-rebinding gap.
• Validate Hermes inference endpoint URLs + reject allowed_ips in user presets (#6087): the user-supplied Hermes endpointUrl is now SSRF-checked before it's persisted and passed to OpenShell, and user preset files can no longer expand the private-IP allowlist.
• Keep API keys out of curl argv (#5975): probe-time curl now routes credentials through a 0600 --config tmpfile so the key never lands in ps auxww on runtimes that share /proc with the host.
• Disable EC2 metadata discovery in OpenClaw (#6096): forces AWS_EC2_METADATA_DISABLED=true across OpenClaw image builds and gateway startup, blocking an impossible credential-discovery path.

On the feature side, a new channels status config summary (#6044) compares your sandbox registry inputs against the rendered channel config for Telegram, Teams, Slack, Discord, WeChat, and WhatsApp, and exec now rejects multi-line arguments with actionable recovery guidance instead of a cryptic OpenShell InvalidArgument error (#5980). The installer now recovers validated sandbox backups before generic onboarding (#6132). The remaining commits were docs and CI maintenance.

Commit log → Affects: /nemoclaw/, /nemoclaw/policy/, /nemoclaw/switching-providers/, /security/

Enterprise gets model entitlements; Fable 5 and Mythos 5 are back. Two updates landed in the Claude apps and API release notes dated July 1:

• Enterprise model entitlements: admins on Enterprise plans can now control which models their users have access to and what effort-level settings they can use — useful if you run Claude Cowork projects under a managed org and want to standardize on a specific model/effort tier.
• Fable 5 and Mythos 5 restored: access to Claude Fable 5 and Claude Mythos 5 was reinstated after a suspension (see Anthropic's statement).

Claude apps release notes → Affects: /claude-cowork/, /claude-cowork/pricing/, /claude-cowork/system-prompts/