Patch release: the session picker moves to Ctrl+R, organization policy for MCP servers and marketplaces is substantially expanded, and you get local feedback drafting plus new rate-limit and mTLS model settings.
Changelog
[new]Local feedback drafting. Grok can save or update feedback drafts with the send_feedback tool (title, details, type, optional area/task category/failure mode). Drafts stay local until you review and send them from the /feedback modal; the tool does not open UI or end the turn.
[changed]Session picker shortcut is Ctrl+R. Opening the session picker (same as /resume) is bound to Ctrl+R instead of F3 on the agent screen and welcome screen. With ui.mouse_reporting_toggle enabled, Ctrl+R still toggles mouse capture when the scrollback pane is focused; from the prompt it opens the picker.
[changed]Organization policy for plugins, marketplaces, and MCP. Admins can enforce allow/deny lists and pins through native TOML (managed_config.toml / requirements.toml, plus macOS MDM) as well as optional Claude managed-settings.json (advisory, foreign-defined sources only). Layers combine strictest-wins; CamelCase and snake_case keys are both accepted.
[new]MCP server allow/deny and lockdown. Policy can set allowedMcpServers / deniedMcpServers (by URL, command, full argv, or server name), allowManagedMcpServersOnly, and enableAllProjectMcpServers. Non-matching servers are dropped at merge time with MCP server blocked by managed settings policy. grok inspect shows the merged lists, lockdown scope, and remaining servers.
[new]Marketplace source restrictions and pins.strictKnownMarketplaces limits which git/GitHub sources can be added; local-path adds fail closed unless an admin extraKnownMarketplaces pin names that path. Extra pins can provision sources and turn off session-start plugin auto-update (plugin_auto_update / pluginAutoUpdate, tighten-only).
[new]marketplace.require_sha. Tighten-only gate (also GROK_MARKETPLACE_REQUIRE_SHA) so remote plugin installs and updates must pin a full commit SHA; neither the key nor the env var can turn the gate back off.
[new]Rate-limit retry controls.rate_limit_retry_threshold (global under [models] or per model.<id>) sets a total-attempt ceiling for 429s (capped by max_retries). When set, the sampler owns those retries and the separate subagent 429 wait loop is disabled. subagent_rate_limit_max_attempts still applies only when the sampler threshold is unset (default 8, max 32, 0 disables).
[new]mTLS for custom model endpoints.model.<id>.mtls_cert_dir points at a directory with client.crt/client.key or tls.crt/tls.key. Config is rejected unless that model has a single HTTPS base_url, no api_base_url, and requests do not follow redirects.
Under the hood
Feedback is a first-class subsystem (draft store, taxonomy, feedback modal UI) wired through a grok_build/send_feedback tool path.
OpenTelemetry export is reworked into a dedicated otel stack (provider, OTLP, redaction, trace context) separate from the older telemetry layer layout.
Plugin loading is split into a clearer acquire/sources module layout.
Subagent rate-limit wait handling and session turn-end / create-prefetch paths are tightened to match the new retry knobs.
Workspace pager layout/sync and block-viewer selection support continue the multi-agent workspace work.
Upgrade notes
Relearn the session picker chord: use Ctrl+R, not F3.
If you rely on org-managed MCP or marketplaces, review native requirements.toml / managed_config.toml (and any Claude managed-settings.json) so allow/deny lists and lockdown flags match what you intend; blocked servers no longer start.
Optional: set models.rate_limit_retry_threshold or per-model overrides if you want sampler-owned 429 retries instead of the subagent wait loop; configure mtls_cert_dir only for HTTPS model endpoints that need client certificates.
