Patch release focused on a more capable /memory modal, more reliable /learn runs, accurate headless MCP status, and a new Helena marketplace plugin.
Changelog
[changed]/memory works when memory is config-disabled. The modal is available whenever a memory store is configured for the session, including when [memory] enabled = false, so you can browse saved notes and turn memory on for the current session with t. It stays hidden only when no store is configured, or when --no-memory / GROK_MEMORY=0 force memory off for the whole process.
[changed]Session memory toggle is bidirectional. Press t in /memory to turn memory on or off for this session without editing config.toml. Turning on re-registers tools, restores memory instructions, and injects the index on the next turn (after any in-flight turn finishes). New sessions still follow config; the toggle cannot override process-wide disable flags.
[changed]Memory modal reading and filtering.Enter focuses the preview for keyboard scrolling; filter matches note names and contents (all words must match) and jumps the preview to the first hit; drag preview text to copy it; generated MEMORY.md indexes are not deletable. Narrow layouts switch to list-only under 64 columns, with full-width read via Enter.
[fix]/learn always shows the cost before running. Even when you say “just run it” / auto, the estimate line (sessions, ~tokens, minutes) is printed first. Auto scopes over ~10M tokens still get a one-line “this will use ~N tokens, go?” confirm.
[fix]/learn report stall handling. Default report-stall window is 45 minutes (was 20). Slow report workers—common on large Windows runs—are treated as still alive when the host shows an active run, with a longer wait instead of a false stall. Resume retries the report step alone when the run actually failed; completed runs are no longer offered a resume/pause path that cannot work.
[fix]/learn synthesis writes stay under API output limits. The reduce step writes the synthesis file in pieces and keeps single tool-call arguments under roughly 6,000 words so long reports are not cut off by max_tokens_truncation.
[changed]Headless streaming-messages-json MCP status.mcp_servers[].status is a live x.ai/mcp/list snapshot: connected, failed, needs-auth, pending, or disabled, instead of always reporting configured servers as connected. Other output formats omit the array.
[new]Helena marketplace plugin. AI marketer plugin from Enrich Labs (MCP server helena) for planning and optimizing campaigns across SEO, ads, social, and email — helena.
[new]Stripe plugin: stripe-pay skill. Helps send funds to another Stripe business, transfer to a Stripe Profile handle or network ID, and related payment flows.
[changed]Marketplace plugin updates. Source revisions refreshed for modern-web-guidance, netlify, pstack, and stripe.
[changed]Docs: skill path config.[paths] extra_skill_dirs is only a record written by /import-claude; put extra skill directories in [skills] paths if you want them loaded. Skills guide tool name example updated to run_terminal_cmd.
Under the hood
Bundled content bumped to public-2026-09-16-r1, including the learn skill/workflow and state helper defaults above.
Headless MCP init path refined so status reporting can reflect handshake progress without blocking the prompt toolset’s startup grace.
Cursor-worker extensibility groundwork (artifacts, frontmatter, skills, subagents) and shared glob helpers landed for upcoming packaging work.
