Mario Martinez II retweeted
The initial response to this article has been incredible.
The S&P 500 is now <1% away from a record high but 206 of its components are down over -30%.
Never in history has this happened before.
It's the "Bear-Bull Market Paradox." Bookmark this.
Mario Martinez II retweeted
The world of personal agents keeps growing
Keep track of it here
Mario Martinez II retweeted
1/ muse spark open is the most capable US open-weight model to date.
816B total params, 42B active. 1M-token context. natively multimodal. built for coding and agentic work like long-horizon tasks, hard software engineering, multi-step reasoning.
america! 🇺🇲
Mario Martinez II retweeted
Absolutely incredible.
Just 10 stocks have driven 70% of the S&P 500’s gain since March 30th.
These stocks have added +16 percentage points to the index’s +23% gain over this period.
Nvidia, $NVDA, has contributed the most, accounting for 13% of the S&P 500's gain, or +3 percentage points.
This is followed by Micron, $MU, Apple, $AAPL, and Microsoft, $MSFT, each accounting for 9% of the S&P 500’s advance, adding +2 percentage points.
Return contributions of AMD, $AMD, and Alphabet, $GOOGL, have been 8% and 7%, respectively.
Meanwhile, the remaining 490 stocks have accounted for just 30% of the S&P 500's gain since March 30th, adding +7 percentage points.
Diversification is dead.
Mario Martinez II retweeted
Marc Andreessen reveals why housing, health care, and education keep getting more expensive while everything else gets cheaper
"There's basically three components to the American dream. A house near great job opportunities, great health care for you and your family, and sending your kids to good schools"
"If you disaggregate the economy over the last 25 years, everything that technology touches basically collapses in price. Television sets, computer games, anything with electronics in it. But if you chart housing, education, and health care, the prices of those things are skyrocketing"
"The government restricts supply, zoning laws and university certifications. And then the government responds to the political pressure caused by the rising prices by subsidizing demand. And of course, if you subsidize demand and a fixed supply, you get skyrocketing prices"
Mario Martinez II retweeted
Morgan Stanley's Adam Jonas in new @SpaceX note:
"The part of the SpaceX story that we believe holds the greatest growth potential, and that is being largely ignored/not contemplated by investors, is the Distributed Inference Cloud. Investors are already witnessing some degree of corporate/regulatory/social angst toward data centers. We see SPCX's efforts as instrumental in helping to get the Super Intelligence Economy and the Physical Intelligence Economy on the next leg of growth and economic development. As the 'density' of physical agents increases, we expect incrementally greater intelligence capability of the collective mobile swarm.
We ask investors to be prepared to hear about the Distributed Inference Cloud and physical AI swarms in the coming year. From our seat, we think this subject is very close to breaking out in the investor discussion. Gradually… then suddenly."
Mario Martinez II retweeted
“Every company is an API company now, whether they want to be or not.” (Sam Altman)
The @readme team, which built Restless, says ~72% of its own docs traffic is already agents. If your product can’t be reached that way, the agent picks a tool that can.
Their idea is to treat agents as your main users.
Restless is a tool you add to your API by running one command, `npx restless init`.
It records every request your users’ AI agents send, and when a call fails, the error response itself tells the agent how to fix it.
🤝 Paid partnership
Mario Martinez II retweeted
Gavin Baker on the AI race as the highest-stakes game of corporate chess ever played:
"You've got to give Meta a lot of credit. They were out of the game, and they got back in the game. It's just kind of amazing."
"Who could have imagined a year ago, when Gemini was ascendant, that Gemini wouldn't even be in the conversation, and Muse and Meta would be significantly ahead of them from a capability perspective."
"Some people have made bad moves, they've made good moves. You've seen some people come out of the game, others come back in."
"I don't know if we're going to call it multi-model, a hybrid model. I don't know what terminology the world is going to settle on, but I think that's the future."
@GavinSBaker @DavidGeorge83
Mario Martinez II retweeted
Next 18 months is the Kitty Hawk moment for RSI
Evals will be how you can tell
Mario Martinez II retweeted
SpaceXSI has officially published a complete guide explaining how to unify your 𝕏, Grok and Cursor accounts under the new 𝕏 Subscription
If you're curious about the new plans, shared usage limits, existing subscriptions, refunds or how to switch, here's everything you need to know
1. What's changing?
Instead of managing separate Grok, Cursor and 𝕏 Premium subscriptions, you can bring eligible accounts together under one subscription
Grok and Cursor get one unified workspace, one plan, one bill and a shared weekly SI usage pool. Connecting your personal 𝕏 account adds 𝕏 benefits based on your plan
Supported products include Grok Chat, Imagine, Voice, Build, Grok Bot, Grok API and Cursor, depending on your subscription tier
2. How much does it cost?
• Free: $0/month — Limited usage
• Plus: $8/month — Base usage
• Premium: $30/month — 4× Plus usage
• Super: $100/month — 15× Plus usage
• Ultra: $200/month — 50× Plus usage
Discounted annual plans:
• Plus: $80/year
• Premium: $300/year
• Super: $1,000/year
• Ultra: $2,000/year
Prices exclude applicable taxes
3. What happens to 𝕏 Premium?
Connect your personal 𝕏 account to receive benefits under the same subscription
• Free includes standard 𝕏
• Plus includes a checkmark, fewer ads and enhanced creator tools
• Premium, Super and Ultra include full 𝕏 Premium+ benefits
Grok Bot is included starting with Premium, NOT Plus
If you currently have X Premium+, choose Premium or higher to retain those benefits after switching
4. How does the shared usage pool work?
Eligible first-party Grok and Cursor activity draws from ONE weekly pool instead of separate product allowances
Different activities consume different amounts, and the included usage resets every week
IMPORTANT: Third-party models such as Claude Cursor are NOT included in the weekly allowance
They're charged separately at API list prices from the first use, using prepaid credits
If you run out of included usage, you can wait for the weekly reset or purchase additional usage credits
Auto Top-up is optional, and you can set spending caps
5. Do existing subscribers have to switch?
NO. Switching is completely optional
If you're already subscribed to Cursor, SuperGrok or 𝕏 Premium, your existing plan, pricing and features continue unless you choose to change or cancel them
6. What happens to existing subscriptions when you switch?
Eligible subscriptions included in the switch are automatically cancelled, with prorated refunds for unused time
Your new billing cycle starts immediately, and refunds go separately to the original payment method
But there are important exceptions:
• Cursor through Apple's App Store: You must cancel manually through Apple and request any refund from Apple
• Grok, Grok Bot or 𝕏 Premium through Apple's App Store: The subscription is cancelled, and Apple decides the refund
• Google Play: Check your subscriptions after switching and cancel any that are still set to renew
Subscriptions on accounts NOT included in the switch remain unchanged
7. How do you unify your accounts?
If you're starting from Cursor:
cursor.com/unify
If you're starting from Grok:
accounts.x.ai/unify
If you're starting from 𝕏:
Open your 𝕏 subscription settings and select the upgrade option
IMPORTANT: Starting from 𝕏 doesn't include an existing Cursor account. If you have a paid Cursor subscription, START FROM CURSOR
Choose the accounts you want to combine, review your new plan and complete checkout
Your accounts do NOT need to use the same email address
Existing subscribers switch through the web
8. The most important thing to know
Once you complete account unification, it is PERMANENT
Even Support cannot reverse it
Make sure you're connecting the correct accounts before confirming
Only personal 𝕏 accounts can be connected for 𝕏 benefits. Team and Enterprise accounts have separate eligibility rules
After switching, you can manage your subscription, billing, weekly usage and top-ups at
grok.com/dashboard
SpaceXSI is bringing social media, AI assistants, coding, creative tools and autonomous agents under one connected subscription ecosystem
The biggest change is the shared usage pool, allowing you to use supported SI products through one subscription instead of managing separate allowances
One subscription, multiple SI products and one connected ecosystem
Full official guide:
docs.x.ai/x-subscription/uni…
It's now officially ONE subscription for the entire SpaceXSI ecosystem, with a shared usage pool across every plan
SpaceXSI just merged Grok, Cursor and 𝕏 Premium into one plan called The 𝕏 Subscription
Until now you paid separately for Grok, for Cursor and for 𝕏 Premium. Now a single subscription covers all three, and Grok, Grok Bot and Cursor draw from one shared weekly usage pool, so your compute goes wherever you need it, whether that's chatting, coding or letting Grok Bot do the work
There are five tiers:
• Free to get started
• Plus at $8 gives you base usage plus 𝕏 Premium benefits
• Premium at $30 gives you 4x the usage, Grok Bot, Cursor cloud agents and ad-free 𝕏 on most surfaces
• Super at $100 gives you 15x the usage
• Ultra at $200 gives you 50x the usage, for people who code and build all day
Nobody gets forced over. Existing Grok, Cursor and 𝕏 plans stay exactly as they are if you do nothing. If you do switch, your eligible old subscriptions cancel automatically at checkout, and the unused time is refunded on a prorated basis
One SI platform, one bill, and one usage pool for chat, coding, agents and 𝕏
Mario Martinez II retweeted
Founders who build $10B companies:
• started at 29, not 38
• are 4x as likely to be dropouts
• are less likely to have sold a company before
I had AI profile 3,125 founders behind 1,739 unicorns. Most of what VCs screen for doesn't separate them.
🧵
Mario Martinez II retweeted
My bet would be Muse is going to win
Mario Martinez II retweeted
Grok Build is getting much better at running and tracking multiple live agent sessions at once
Rex terminal support now lets tabs show whether each Grok session is working, waiting for you, done or failed, while tool-call groups stay expanded as new calls arrive and Ctrl+E makes navigating long reasoning transcripts faster
Project hooks now load properly from repository subfolders, localhost HTTP hooks are supported, and status-line scripts can see the active turn’s prompt ID + start time
There’s also a pretty serious reliability pass underneath it all: context overflows can trigger automatic compaction, model switches persist correctly, background shell commands stop when Grok exits, approval prompts identify which session needs attention, worktree deletion is much safer, and Auto mode gets tighter permission enforcement
A really useful release for anyone running multiple agents, long sessions and heavier automated workflows
Release Notes: v1.0.52
Features
• Rex terminal support: Grok now reports its status (working, waiting for you, done, or failed) to Rex through OSC 7501, so Rex tabs show what each session is doing.
• Tool call density improvements: expanded tool call groups now stay expanded as new calls join them, and Ctrl+E expands or collapses thinking faster in long transcripts.
• Project hooks placed in a repository subfolder are now loaded when you launch inside that folder.
• HTTP hooks to localhost now work, and bad hook URLs are reported at load time.
• Command status-line scripts now see the current turn's prompt ID and start time while the turn is active.
Bug Fixes
• Grep searches that time out or fail now show an error instead of appearing as empty results.
• /model switches that fail no longer change the footer or saved default model.
• Export and copy commands no longer overwrite an existing file; they refuse with a clear message.
• Shell commands no longer keep running after you quit grok or it receives a signal.
• Approval cards for background subagents no longer disappear when the parent turn ends.
• grok worktree rm and gc no longer delete a worktree that an open session is using.
• grok inspect no longer warns about valid pins that only requirements.toml accepts.
• Pasted images that exceed the 50 MB send limit are now rejected with a clear message instead of silently dropped.
• grok -p --effort now prints a warning when the model ignores the flag and refuses invalid effort levels.
• Permission prompts now show which session needs approval.
• grok worktree rm now refuses to delete git repositories, .git folders, or any directory grok did not create.
• Grep and list_dir now report a clear error when the search root or directory cannot be read instead of silently returning no results.
• Cancelled terminal commands now show the actual command that was stopped.
• Wrapped URLs in prompts and quotes now link to the complete address instead of being cut at the wrap.
• @mentions and image placeholders now respect Read deny rules and are omitted from the prompt.
• Restarting after a model switch no longer reverts to the previous model.
• Long tool turns that include PostToolUse hook output now compact correctly instead of hitting the prompt length limit.
• Headless prompts that start with a byte-order mark no longer send the mark to the model or break JSON prompt files.
• When a permission rule blocks a tool call, any context returned by a PreToolUse hook still reaches the model.
• An invalid shell_environment_policy table in config.toml now causes grok to exit with a clear error instead of silently using the full environment.
• MCP tool replies containing invalid characters no longer cause the call to time out.
• Context overflow errors now trigger automatic compaction instead of failing the turn.
• Inline tags in agent replies no longer break line layout, and /workflow stop now picks the correct live run.
• Auto-compaction now stops after three context refills in a row instead of looping forever.
• Minimal mode now shows the permission prompt instead of a hidden plan or cancel panel when both are present.
• Project rules files that resolve outside the git root or workspace folder are now skipped.
• [permission] rules now load alongside compact deny/allow/ask lists in config.toml.
• Unrecognized permission keys or [permissions] sections now appear as skipped entries in grok inspect.
• Approval cards now close the block viewer so the card is visible.
• Shell redirects to wildcard or variable paths now correctly respect deny rules in Auto mode.
• The in-session resume picker now matches the ungrouped layout shown on the welcome screen when configured.
• HTTP MCP tools that return a JSON-RPC error are now executed only once.
• StopFailure hook matchers containing commas or spaces are now treated as regex patterns.
Performance
• Multiple Notification hook chains now start in parallel instead of waiting for the previous chain to finish.
Grok Build just got a monster update......this one pushes long-running agents, browser workflows, worktrees, hooks, permissions, and session recovery much further
Background workflows can now be listed, stopped and resumed directly with /workflow, headless Grok waits for background goals and reports their outcome, Chrome profiles can persist across restarts, browser tabs can restore on demand, and workspace sessions can share browser cookies + storage
There’s also a new /density control for tool-call detail, persistent domain approvals for web fetches, live slash-command updates when skills change, and secure request forms that can collect sensitive details like passwords or card numbers without sending those values to the model
Underneath that is a massive hardening pass: cancel/edit-resend no longer unexpectedly reverts workspace files, long prompts are offloaded safely, worktrees recover and clean up more reliably, hooks and approval flows are tighter, browser state is more stable, background agents respect saved approvals, and resume/worktree performance is much faster
Release Notes: v1.0.51
Breaking Changes
• Cancel and edit-resend on long sessions no longer unexpectedly revert workspace files.
• Memory settings simplified: rollout and file_writes_enabled options removed; memory is on or off.
• Removed the /share command and related session sharing features from the CLI and SDK.
Features
• Inline viewports can now be restored at a previously saved screen area without extra cursor queries.
• Headless grok -p now waits for background goals to finish and reports their outcome.
• /imagine and similar commands now run the matching bundled skill when the required tools are unavailable.
• New /density command and Tool call density setting let you choose how much detail tool calls show in the transcript.
• Web fetch approval cards now include "always allow" and "never allow" domain options that persist for the folder.
• Slash command completion now updates live when skills are added or changed during a session.
• Request form now securely collects and fills sensitive details like passwords or card numbers without sending values to the model.
• Chrome profiles can now be persisted across restarts using durable state directories.
• Workspace server sessions now share one Chrome instance for cookies and storage.
• /feedback Drafts tab now supports G / Home and End to jump to the first or last draft.
• Pinned sessions on the dashboard now stay in the order you pinned or reordered them.
• Shift+Enter, Ctrl+., Ctrl+/ and Ctrl+[ / ] now work inside tmux when extended-keys are enabled.
• grok -p --output-format json now reports the exact worktree folder used (useful after -w name collisions).
• Bare hook commands such as my-guard now run from PATH when not found next to the hook file.
• Headless JSON error output now includes the worktree folder when a session was started.
• grok sessions list and /resume now show worktree sessions started from the same subfolder.
• /exit (and aliases) in a dashboard details view now closes only that session instead of quitting Grok.
• Editors offering file access now receive a one-time transcript notice that the agent reads and writes disk files.
• --background-wait-timeout now also stops an active headless goal after the timeout and exits with an error.
• Interactive sessions now include the Grok Build user guide location so the agent can answer usage questions.
• New /workflow commands let you list, stop, and resume background workflow runs without sending them to the model.
• Background workflow runs now appear in the tasks pane with a stop button.
• Hooks that ask about web searches or file deletes now show the hook name on the approval card.
• config.toml now accepts an [update] section without breaking other commands.
• Hook load errors listed in /hooks can now be expanded to see the full message.
• grok worktree commands now work correctly inside Jujutsu repositories.
• Images pasted with skill commands like /pr-workflow are now sent to the model.
• Pressing Ctrl+C on a goal now pauses the goal instead of letting it continue.
• grok -p now runs auth_provider_command when no login is stored.
• Reopening a chat with an active goal now continues the goal automatically.
• Browser tabs can now be restored on demand instead of all at once when a session resumes.
• grok agent stdio now supports listing sessions via ACP and returns consistent reply envelopes.
• Answering an approval with a message now tells the model it is a follow-up.
• --allowed-tools is now accepted as an alias for --allow in headless mode.
Bug Fixes
• Slash command suggestions no longer make the text after the cursor disappear while typing.
• Running workflows in the dock now show the same pulsing indicator as in the main scrollback.
• Mode switch tool calls now appear in the transcript with the target mode name.
• Always-approve mode now automatically approves web requests without extra prompts.
• The follow indicator no longer appears when only blank padding sits below the viewport.
• Search failures now keep the original pattern instead of showing an empty Search "".
• Sessions that open without a model list now use the public Auto id and succeed on every account.
• Background wake turns that request plan changes or mode switches are now handled correctly.
• Model-call retries now show progress status instead of a generic waiting message.
• Foreground subagent activity now appears correctly after resume.
• Very large pasted prompts are now offloaded to a file instead of being sent inline.
• Sign-in failures now show a clear plain-English message with the next step instead of raw errors.
• Logging out now stops any running turns on local sessions instead of leaving them active.
• The send-now shortcut no longer accidentally enables always-approve mode while a permission card is open.
• /find now keeps match counts and navigation accurate while replies or command output stream in.
• Images from tools are now validated so bad files no longer break the current turn or future replays.
• Screenshots and other image artifacts saved by the browser tool are now stored as real image files.
• Rate-limit errors now show the expected message; very small images are rejected early.
• Send-now (Ctrl+Enter) no longer accidentally answers an open question card.
• Opening the queue with Up now scrolls so the selected follow-up is visible.
• Command output that contains clear-screen codes is no longer discarded.
• Dashboard cursor and overlay counter now correctly track the active grouping view.
• Headless JSON output now reports the correct stop reason when the turn limit is reached and removes duplicate error prefixes.
• Session picker footer no longer lists the same shortcut twice.
• Dragging images from a file manager onto terminals like GNOME Terminal or xfce4-terminal now attaches them correctly.
• Plan comments added while a revision runs are now preserved and sent when the plan is approved.
• Permission prompts that fail now show a clearer message without blaming the interface.
• --disallowed-tools now accepts the exact names shown to the model (run_terminal_command, spawn_subagent, etc.).
• Long pastes without bracketed-paste markers are now read completely instead of stalling after ~1 KB.
• A feedback send now shows "Sending feedback…" immediately; failed sends no longer repeat the lead phrase.
• Typing, arrows or Enter while an approval card appears no longer accidentally approve the action or enable always-approve mode.
• Error hints for legacy session import now appear in full instead of being cut off by the pager.
• Skills in the folder where grok started now correctly show as (local).
• Worktree creation no longer hangs when the repo contains named pipes or sockets.
• --no-subagents flag now works correctly in headless / -p sessions.
• Resumed sessions now correctly show aborted subagent tasks as cancelled instead of failed.
• Large parallel shell batches no longer drop calls with "hook state lost" errors.
• Write approval cards now show the existing file contents that will be replaced.
• Slash command completion now shows the correct name for files under commands/ instead of always showing /commands.
• PreToolUse hooks now validate rewrites against the tool schema and deny invalid ones before running the tool.
• Ctrl+C during a turn now cancels the turn even when a modal such as /mcps or the block viewer is open.
• grok -w now opens the session at the worktree root (with a note) when the original launch folder does not exist in the new worktree.
• /cd and the location picker now accept quoted or escaped folder paths.
• Web search hooks now fire under the correct web_search name and carry query.
• Large file writes or many screenshots no longer end the session.
• Glob tool hooks now match under list_dir and carry the correct directory.
• /find now scrolls to the exact wrapped row containing the current match.
• Downloads via FetchMcpResource are now refused in plan mode and resolve paths consistently.
• Pinned dashboard sessions now survive restarts and layout changes instead of disappearing from the saved list.
• Slash commands like /imagine now work regardless of capitalization and correctly substitute skills when tools are missing.
• Turning off bypass mode without naming a session now correctly restores prompts in all open sessions.
• /delete with extra text now shows an error instead of deleting the current session.
• Failed /feedback status messages now wrap inside the dialog instead of clipping at the border.
• Number-key picks in the dashboard peek no longer leak to the wrong agent after Up/Down navigation.
• Worktrees created with -w are now correctly tracked and resumable even when GROK_HOME is reached through a symlink.
• grok worktree rm and gc now clean up names after manually deleting a worktree folder so -w NAME works again.
• Hook scripts that read stdin with read now receive a complete line.
• A policy that disables non-managed hooks now applies to the next session instead of only after restart.
• /resume and grok sessions list no longer show worktree sessions from a sibling submodule or separate-git-dir repo.
• A hook listed under two different matchers now runs only once per tool call.
• Typing a hidden command such as /cd in a session and pressing Enter now runs it instead of accepting a near-match skill.
• @-file completion no longer accepts a match from an earlier query when you type quickly.
• Tool hooks now run one at a time per session and stop promptly on cancel or exit.
• PostToolUse hooks now receive a correct whole-millisecond duration.
• PreToolUse hooks for shell commands now see the model's block_until_ms value.
• MCP tool calls now show hook permission prompts and cards name the hook that asked.
• Typing while a permission or ask card is visible no longer selects or answers the card.
• grok worktree list now shows unfinished copies left by an interrupted create.
• Auto mode now asks before deleting files that the classifier has not already allowed.
• Fixed intermittent List/Read/Open failures during concurrent v2 hardening migrations.
• grok -p -s with a reused session id now shows the plain refusal message instead of JSON-RPC noise.
• Settings saves no longer re-insert the built-in default fork model into config.toml.
• Fixed several pager rows that showed wrong paths, truncated names, or hid failure/cancel states.
• --system-prompt now correctly appends rules from --append-system-prompt instead of dropping them.
• /new now inherits always-approve mode from the tab it was run from instead of another tab.
• Multi-select questions now let you toggle several answers with number keys before submitting.
• Slash-command notices typed while a reply streams no longer hide the streaming reply.
• @ file search no longer accepts keys or moves the highlight while results are still updating.
• Worktree db rebuild now restores names so rm and show by name work again.
• A prompt blocked by a hook no longer titles the session.
• Background subagents now respect saved approvals and --always-approve after the parent turn finishes.
• Per-session system prompt overrides now correctly use only the session's own rules.
• Connection stalled errors now show a clearer message with network advice.
• Invalid model provider tables now produce a clear startup error instead of silent fallback.
• Narrow feedback dialog now shows full draft titles and readable confirm prompts.
• /workflow stop|resume now correctly targets the live run instead of an ended first run.
• Dialogs like /feedback no longer disappear on very short terminal windows.
• Delete operations now respect the same permission rules as shell rm commands.
• Hooks can no longer rewrite a web-search call to restrict domains.
• Workspace server now starts with per-session private Chrome by default.
• Notification hooks for permission prompts and task completion now run reliably before a headless session exits.
• Headless goal runs no longer print a spurious "Goal paused" message after completion.
• Legacy UI density settings now combine correctly across config tiers and produce expected folding.
• Approval answers sent after a turn ends no longer affect the turn.
• Background shell commands now fail immediately when the working directory does not exist.
• grok worktree rm without -f now safely keeps worktrees when the unsaved-work check fails or times out.
• The dashboard slash menu now shows skills from the home session even when opened early.
• grok worktree rm and gc now correctly remove worktrees whose only untracked content is empty folders.
• Editing a queued prompt now prevents accidental removal with a stale client view.
• The model is now told when memory is turned off or on for the session.
• Hook results in the UI now correctly distinguish between notes and tool outcomes.
• Permission cards for MCP tools can no longer be tricked into showing fake option rows.
• The agent stdio mode now returns clearer errors for malformed client requests and consistent recap replies.
• Scroll logging and acceleration calculations no longer produce backwards timestamps under load.
• Hooks and tool commands started from the TUI can no longer read or write the interactive terminal.
• Delete operations now correctly match Bash(rm …) ask/deny rules even when the path contains .. or the cwd is a symlink.
• Web search and image approval cards now only offer always-approve mode to clients that support it.
• Closing a session in the dashboard no longer leaves it listed under Inactive.
• Running /compact on an empty session no longer resets the reported context window.
• With goals disabled, a headless /goal line is now treated as a normal prompt instead of failing.
• Background shell commands no longer flip back to running when late output arrives.
• Only slash commands at the start of a message trigger skills.
• Browser tabs restore and close more reliably with fewer races.
• Fast model names in the footer no longer show double parentheses.
• SubagentStart and Notification hook activity is no longer reported to the UI.
• Voice error messages now display with correct punctuation.
• disable_zdr_incompatible_tools changes now apply to new turns without restarting the worker.
• Project skills are now labeled correctly for nested projects.
Performance
• Opening the resume picker is now much faster on machines with many worktrees.
• grok worktree commands start faster by skipping unnecessary model catalog work at boot.
