Changelog — July 14, 2026

Claude Code shipped v2.1.208, a feature-plus-fix batch on top of v2.1.207, followed hours later by the v2.1.209 patch. The additions:

• Screen-reader mode: an opt-in plain-text rendering for screen-reader users. Turn it on with claude --ax-screen-reader, CLAUDE_AX_SCREEN_READER=1, or "axScreenReader": true in settings.
• vimInsertModeRemaps setting: map two-key insert-mode sequences like jj to Esc when using vim mode — the long-requested "jj to escape" ergonomics.
• CLAUDE_CODE_PROCESS_WRAPPER: the agent view and background service now honor a corporate launcher by running every Claude Code self-spawn through a required wrapper executable — for locked-down enterprise environments that must interpose on child processes.
• Mouse clicks in fullscreen menus: multi-select menus and the "Other" input rows are now clickable in fullscreen mode, not keyboard-only.

The reliability fixes in v2.1.208 land mostly around background and long-running sessions:

• Lost background-agent replies recovered: a reply typed to a background agent whose delivery fails is now saved and delivered when the session restarts, instead of vanishing.
• Attach-after-update repaired: attaching to a background session no longer fails permanently ("Couldn't start the background daemon") after an update replaced the binary a running claude agents process was launched from.
• No more false "100% context used": the context window (and auto-compact indicator) no longer briefly resets to 200k after the CLI auto-updates, which had thrown a bogus full-context warning when resuming long-context sessions.
• Fewer crashes and truncations: fixes for supervised/background sessions crashing on an HTTP/2 GOAWAY mid-request, truncated stream-json/JSON output when piping large responses from claude -p, CLAUDE_CODE_MAX_OUTPUT_TOKENS silently reading the mantissa of scientific-notation values (1e6 → 1), fast mode staying off after switching back to a supporting model, and very large markdown tables stalling rendering (tables over 200 rows now show the first 200 with a "… N more rows" notice).
• v2.1.209 patch: restores /model and other dialogs that had been blocked inside claude agents background sessions — a revert of an overly broad guard shipped just before.

Who should care: Claude Code and OpenClaw CLI users. The screen-reader mode is a genuine accessibility milestone; vimInsertModeRemaps and clickable multi-selects are everyday ergonomics; and CLAUDE_CODE_PROCESS_WRAPPER is the one to note for enterprise/managed deployments. If you run background agent sessions, v2.1.208 + v2.1.209 together fix several ways those sessions could silently lose input, fail to reattach, or block dialogs — worth updating for.

Claude Code v2.1.208 release notes → Affects: /openclaw/, /openclaw/setup/, /openclaw/configuration/, /openclaw/security/, /claude-cowork/, /claude-cowork/setup/

v0.0.82 is now tagged (release notes #6824), summarizing the user-facing changes merged on main since v0.0.81. The changes that matter to users:

• Snapshot clones get their own dashboard port (#6749): snapshot restore --to <dst> auto-creates the destination by spreading the source's registry entry, which meant the clone inherited the source's dashboardPort — so the clone's dashboard-url pointed at the source and every rebuild of the clone was rejected ("Dashboard port NNNNN belongs to sandbox '<src>'"). The clone now gets a destination-owned dashboard port allocated at registration, before any destructive --force action, so dashboard-url reports the clone's own port and rebuild proceeds.
• Actionable onboarding guidance when Chat Completions is missing (#6769): onboarding an agent that requires OpenAI Chat Completions (such as Deep Agents) against an Anthropic-compatible endpoint that doesn't serve /v1/chat/completions now prints a concrete next step before exiting — use an OpenAI-compatible endpoint, or switch to OpenClaw — instead of a bare failure. A model-specific 404 still offers a model retry rather than wrongly claiming the route is absent.
• ssh mux-forward fix stops false sandbox teardown (#6833): when the host's ssh config applies ControlMaster auto, the spawned client delegates its -L forward to the mux daemon and exits, which OpenShell 0.0.72+ reports as "ssh exited before local forward listener opened." The onboard untracked-forward fallback didn't match that diagnostic, so the live-port probe never ran and a healthy sandbox was deleted after a false 180s forward timeout. The fallback now recognizes the new diagnostics and confirms via the live-port probe.
• Locked, offline WeChat runtime graph (#6739): a NemoClaw-owned production lock for the WeChat/OpenClaw runtime dependency graph, installed offline from a reviewed cache and verified against the reviewed lock — preventing dependency/lock drift during WeChat runtime installs.

Who should care: NemoClaw users cloning sandboxes via snapshot restore --to (rebuilds of clones now work), anyone onboarding an OpenAI-completions agent against an Anthropic-only endpoint (you now get told what to do), and hosts using ssh ControlMaster multiplexing (onboarding no longer tears down a working sandbox). Most other churn on main today was CI/test stabilization — macOS and WSL Vitest lanes and the PR E2E gate — with no user-facing effect.

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