에이전트·개발 도구 중심의 안정성·입력 파이프라인 업데이트

DogeDesigner

@cb_doge

NEW GROK BUILD UPDATE v0.2.121 — 2026-08-05 Features: • Dashboard rows now show a short summary of what the agent did in the previous turn. • The Extensions modal now shows items grouped and sorted alphabetically with collapsible sections for Skills. • When a subagent runs in the background the parent agent is now reminded to keep working on the original task. • Clients can now reattach to a running session without replaying the transcript and explicitly close sessions. • Grok no longer asks which project directory to use when launched from your home folder or other non-project directories. • /feedback now opens a dedicated report box instead of entering prompt mode. • Auto theme detection now works over SSH and inside tmux sessions. • Voice and Finance tool cards are now rendered with icons and localized labels. • Markdown tables now reflow inside cells on narrow panes instead of clipping the right border. • Permission prompts now show the complete script clearly instead of dimmed fragments. • Long bash commands in permission prompts can now be expanded with Ctrl-F. Bug Fixes: • MCP tools that return images no longer drop or corrupt screenshots when output is large. • Resuming a restored child session after a remote parent restore no longer fails with 404. • The default branch is now correctly detected for hand-initialized repos that lack origin/HEAD. • Disabled MCP servers that can still be re-enabled now remain visible in the list instead of disappearing. • Rapid send-now presses or messages sent while waiting on subagents no longer lose earlier queued messages. • Pressing Esc or the stop button now reliably prevents background tasks from restarting the model after a cancel. • Login no longer skips when an invalid first-party API key is present in the environment. • Model picker and command palette now work while reviewing a plan before approving. • Workflow parallel() panels no longer launch hundreds of children at once and stall. • Dashboard overlay no longer shows useless prev/next shortcuts when only one agent is present. • Pinned prompt headers can now be selected and copied with the mouse. • Tab and Esc now behave consistently on every blocking card (question, permission, cancel-turn). • Dashboard navigation after /new now correctly exits back to the dashboard from an empty prompt. • Codebase restore no longer hangs on large or shallow git repositories. • Remote session resume restores conversation only unless --restore-code is explicitly used. • Copying CJK text with the mouse now includes every character at the selection edges. • Resume search now finds sessions by UUID even when they live in other directories. • API errors now appear as clean banners instead of raw JSON dumps in the TUI. • Typing exit or quit in the dashboard now exits the CLI instead of starting a new session. • Mode indicator (plan/agent/ask) now correctly reflects the session's actual mode after resume and transitions. • /delete now returns to the dashboard when you delete a session you opened from it. • Enter in the slash command menu now runs the highlighted command. • Grok now retries more server errors during outages for better reliability. • Syntax highlighting in long diff lines now stays correct when wrapped. • Slash commands that require a session now show a helpful message when used from the dashboard. • Exiting the CLI now properly resets terminal modes even in minimal mode. • Queued prompts now stay visible and reachable while waiting on subagents. • Auto recaps no longer appear in the middle of new turns or while busy. • Error messages in /btw side questions now show fully wrapped. • Queued slash commands and images can now be reordered in the queue pane. • Error messages in /btw side questions now show fully wrapped. • /feedback no longer resets the composer input mode after the pane closes.

Claude Code 2.1.223 has been released. 19 CLI changes Highlights: • owner/* wildcard added to marketplace settings to allow or block all repos in a GitHub org for org policies • Permission prompts block commands padded with tabs or invisible Unicode so dialogs display the actual command • New tool launches agents to coordinate complex multi-step tasks, enabling automated multi-step workflows Full details are in thread ↓

Claude Code 2.1.223: additional changes CLI surface: Added: • models: claude-chrome-screenshots- Removed: • env vars: CLAUDE_CODE_INVESTIGATE_FIRST • models: claude-chrome-screenshots-8 File: https:// github.com/marckrenn/clau de-code-changelog/commit/169cfdd71add61d02e362934c5948ccbc92d935d/meta/cli-surface.md#diff-662031a066e433468319e799350331e143e4635468b9c2924019d16654027e31L7-R7 … Metadata: • Time since 2.1.222 release: 1d 16h 48m 26s • Bundle file size: +132.5 kB (+0.3%) • Prompt files: +3 (+12.0%) • Prompt tokens: +9,566 (+14.3%) Prompt token mix by kind: • system: 73.9%→74.3% • tools: 26.1%→25.7% Files: • metadata: https:// github.com/marckrenn/clau de-code-changelog/commit/169cfdd71add61d02e362934c5948ccbc92d935d/meta/metadata.md#diff-a08dfa163baef97ea4f56d29009ac65f5f7dbc46e3a58e80d7aa06a2a2520899L10-R10 … • prompt stats: https:// github.com/marckrenn/clau de-code-changelog/commit/169cfdd71add61d02e362934c5948ccbc92d935d/meta/prompt-stats.md#diff-fa8e31336ece71d159f37b6fd4ed0c61939849140bd5c08c7ca793b5b2ae7f58L1-R1 … Other prompt/string highlights: 1) Clarified the Grep tool description and usage—now explicitly states it's ripgrep-based, recommends using Grep for search tasks, and details output modes, head_limit, and offset behavior. 2) Added a tool description to launch a new agent for handling complex, multi-step tasks. Links: 1) https:// github.com/marckrenn/clau de-code-changelog/commit/8b4ec670f23ea4652c45da0f0deac2d7ca25ff7e/system-prompts/tool-description-grep.md#diff-98f084cc9ae739dcef3d2c75ea5cf55429dbb647973737f30693e59689d25d58R3 … 2) https:// github.com/marckrenn/clau de-code-changelog/commit/15d3b004e6490fe05b9b74e049635e32222b5e40/system-prompts/tool-description-launch-new-handle-complex-multi.md#diff-f9c65488006fcb18d293ebca17f2420fa5b869c7f4fca44e5a4b8ec47d6cad2fR26 … More details: https:// github.com/marckrenn/clau de-code-changelog/releases/tag/v2.1.223 …

tonbi

@tonbistudio

This new Hermes Desktop browser isn't just an extra window with a browser bolted on, the agent can operate, see, and analyze what it sees along with you I made a short video demonstrating this, but here are just a few things you can do with it: You can scroll X, then ask for a summary of your feed or what the hot topics are You can watch a tutorial video from your favorite YouTuber **cough cough* and then ask your agent to pull transcripts, explain certain concepts from the video, or even implement them directly! You can ask for research on some open source memory solutions for agents, open those links right in the browser, then review them together

Teknium

@Teknium

Hermes Agent can now read anything you throw at it. PDF, Word, PowerPoint, Excel, OpenDocument, RTF, EPUB — all auto-converted to clean Markdown the moment the agent reads the file, all locally! Powered by @firecrawl_dev's new open-source anydoc (Rust). Zero setup, installs itself on first use.

I don't really think these legends needed a pitchdeck. Congrats on taking a crazy, but inspiring leap. I recommend you explore /goal in Codex, it's a pretty powerful loop with GPT-5.6 Sol.

Cline

@cline

The harness is half the battle. Meta trained Muse Spark on its harness, and our benchmarks agree with this approach. We let Kimi K3 tune the Cline harness for its RL trained tendencies, and saw an +11.3% bump on Terminal-Bench w/ 37% lower cost. Same model, different harnesses.

We had Kimi K3 recursively self-improve the Cline harness to improve its own performance. 17 hours later, it went from 77.5% to 88.8% on Terminal Bench, and cut run cost from $79 to $49.8.

(1/6) Qdrant 1.19 is now here! Below are the main updates Full release blog:

(6/6) One upgrade note: the legacy /search, /recommend, and /discover endpoints are gone. Migrate to /query before you upgrade. Full release blog here:

Meet Alibaba Cloud SecOps Agent! Native AI in your console. Just type your intent & automate SecOps. 600+ Skills: Covers 20+ security domains. Rapid Response: Cuts MTTR from 30 to 3 mins. Safe & Auditable: High-risk actions require "Policy Gateway Confirmation." Public Preview is Open!  https:// click.alibabacloud.com/m/20000001252/ #CyberSecurity #AlibabaCloud #AI #SecOps #CloudSecurity