Skip to content

📊 AI CLI Tools Digest 2026-07-23 #17

Open
Open
📊 AI CLI Tools Digest 2026-07-23#17
Labels
digest-en

Description

@bkutasi
bkutasi
opened on Jul 23, 2026

AI CLI Tools Community Digest 2026-07-23

Generated: 2026-07-23 09:29 UTC | Tools covered: 20

Cross-Tool Comparison

AI CLI 工具生态跨工具比较报告 - 2026-07-23

1. 生态系统概述

2026年中，AI CLI 工具生态已从早期的“代码生成器”演变为复杂的“多模态代理与推理平台”。竞争格局呈现三大梯队：以 Claude Code、GitHub Copilot CLI 和 OpenAI Codex 为代表的商业集成巨头；以 Ollama、llama.cpp 为核心的高性能开源推理引擎；以及 Gemini CLI、SGLang 等专注于特定场景（如智能体、分布式推理）的创新项目。硬件支持（尤其是对 AMD ROCm、Intel XPU 和 Apple Silicon 的争夺）与企业级功能（MCP 集成、安全管控、多账户管理）已成为新的差异化战场。所有主流项目都面临稳定性和跨平台一致性的严峻考验，反映出在快速迭代中平衡创新与可靠性的行业共同挑战。

2. 活动比较 (2026-07-23)

工具名称 Issues 活跃数 (Hot Issues) 活跃 PR 数 发布状态
Claude Code 10 10 有发布 (v2.1.218)
OpenAI Codex 10 10 有预发布 (v0.146.0-alpha.3/4)
Gemini CLI 10 10 有发布 (v0.52.0 稳定版，v0.53.0-preview)
GitHub Copilot CLI 10 2 有发布 (v1.0.74-x)
Kimi Code CLI 3 10 无发布
OpenCode 10 10 无发布
Qwen Code 10 10 无发布
llama.cpp 10 10 有发布 (b10085-b10092)
Hermes Agent 10 10 无发布
vLLM 10 10 无发布
SGLang 10 10 无发布
BeeLlama.cpp 4 2 无发布
buun-llama-cpp 1 1 无发布
Lucebox 0 4 无发布
Ollama 10 10 有发布 (v0.32.3-rc0)
Unsloth 4 (部分截断) (未列出) 无发布
bitsandbytes 1 (活跃讨论) 0 无发布
AutoGPTQ 0 0 无发布
MLC LLM 2 4 无发布
KoboldCpp 1 6 无发布

注：活跃数基于摘要中“Hot Issues”和“Key PR Progress”栏目列出的数量。发布状态指过去24小时内的新版本、预发布或候选版本。

3. 共享功能方向

多个工具社区表现出对以下需求的强烈共识：

  • 跨平台稳定性与特定平台修复：几乎所有工具都面临 Windows 和 Linux/macOS 的特异性 bug（如 Claude Code 的 Task 工具丢失、Copilot CLI 的 WSL2 剪贴板错误、llama.cpp 的 Vulkan 编译问题）。对 WSL2、ARM64 等子平台的支持是共同痛点。
  • 高级硬件支持：对 AMD ROCm（如 vLLM, SGLang, Ollama, Unsloth）、Intel GPU（如 vLLM, bitsandbytes）和 Apple Silicon（如 Ollama, bitsandbytes, SGLang）的优化和兼容性支持是核心诉求。新一代硬件（如 NVIDIA RTX 50/Blackwell）的支持快速跟进。
  • 企业级集成与安全性：MCP 协议的可靠实现与工具集成（Claude Code, OpenCode, MLC LLM）是基础。高级需求包括：BYOK（Bring Your Own Key）管理（Copilot CLI）、多账户/配置文件隔离（Claude Code, Hermes Agent）、安全供应链（Hermes Agent）和审计/可观测性（Ollama 添加 /metrics）。
  • 开发者体验优化：普遍关注 上下文管理（自动压缩、会话状态）、工具调用可靠性（尤其是结构化输出/JSON模式的解析，如 Ollama, Hermes Agent），以及 IDE 深度集成（如 Claude Code, Copilot CLI）。
  • 多模态与长上下文支持：对视觉模型的稳定支持（如 Ollama 的 JPEG 处理、llama.cpp 的图像合并修复）、视频输入（KoboldCpp），以及长上下文推理的优化（如 SGLang 的分布式 KV 缓存、llama.cpp 的分页注意力）是共同技术前沿。

4. 差异化分析

工具 核心定位与目标用户 技术与架构重点
Claude Code / Copilot CLI 专业开发工作流集成者。深度嵌入 IDE (VS Code)，面向付费订阅的企业与个人开发者，强调模型能力与工作流自动化。 依赖后端服务（Anthropic API / GitHub API），侧重会话管理、权限控制和工具（如 Task, Todo）的丰富性。
Kimi Code / Qwen Code 企业级多模型集成平台。强调对第三方模型提供商（包括私有部署）的兼容性与多账户管理，服务于需要灵活模型选择的企业用户。 提供抽象层（如 Kimi 的 --channel），支持混合云/私有部署，注重会话隔离和插件化扩展（MCP）。
Ollama 本地化部署与开发者友好的模型服务。目标是成为本地运行开源模型的“标准 Docker”，强调一键安装、跨平台和易用性。 聚焦于模型格式（GGUF）、加载性能和简单的 API，近期重点在工具调用修复和云服务拓展。
llama.cpp / SGLang / vLLM 高性能推理引擎。面向研究者、推理工程师和构建自己服务栈的开发者，专注于吞吐量、延迟和硬件利用率。 底层优化为核心（CUDA/SYCL/HIP 核算、量化、分页注意力、连续批处理），提供服务器模式，技术门槛高。
Gemini CLI 智能体与自主任务执行框架。旨在构建能执行复杂多步任务的可靠代理，强调工具使用、子代理编排和错误恢复。 架构聚焦于代理循环、工具安全（沙箱）、上下文压缩和 A2A 协议，有独特的“Caretaker Triage”架构。
Hermes Agent 多平台通信网关与自动化代理。作为连接 Telegram、Slack、Matrix 等消息平台的智能代理后端，强调跨平台集成和 Webhook 事件处理。 核心是消息路由、适配器管理和插件钩子，注重网络鲁棒性（如 SSE 重连）和进程守护。
KoboldCpp / BeeLlama.cpp 轻量级多模态应用服务器。面向 AI 创作者（故事、图像、音频生成），将语言、视觉、音频生成集成于一个进程。 基于 ggml 的轻量推理，集成 Stable Diffusion、Whisper 等，强调易用性和功能集成（如视频输入）。

5. 社区活力与成熟度

  • 最活跃与快速迭代：Claude Code、GitHub Copilot CLI、llama.cpp 和 vLLM 今日展现了最高的综合活动度，结合了高数量的活跃 Issues/PRs 和频繁的版本发布（包括预发布）。这表明其开发节奏快，用户反馈循环紧密。
  • 活跃但聚焦特定领域：SGLang、Gemini CLI、Kimi Code CLI 和 Ollama 社区在特定方向（如分布式推理、代理框架、本地部署）上非常活跃，拥有深度的技术讨论和贡献。
  • 稳定维护型：MLC LLM、Unsloth、bitsandbytes 和 KoboldCpp 显示出持续的维护活动，主要集中在 bug 修复、依赖更新和特定功能增强上。
  • 相对沉寂：AutoGPTQ 在本报告周期内无活动，可能进入维护模式或开发重点转移。部分工具（如 OpenCode、Qwen Code）虽无新发布，但仍有持续的 PR 工作，表明处于功能开发阶段。
  • 成熟度指标：项目成熟度不仅体现在代码量，也体现在对复杂问题（如安全、多账户、分布式状态）的解决深度。Claude Code、Copilot CLI 和 Ollama 在处理企业级痛点（会话限制、BYOK、云集成）上显示出更高的成熟度。llama.cpp 等在硬件兼容性和性能调优上体现出底层工程的成熟。

6. 趋势信号 (对开发者的参考价值)

  1. “工具即代理”深化：CLI 工具正从“回答问题的聊天机器人”转向“能使用工具、管理状态、执行复杂任务的代理”。开发者在选择时，需评估其工具生态集成能力（MCP 等）和多步骤任务可靠性。
  2. 硬件多元化加速：NVIDIA GPU 不再是唯一选择。AMD ROCm 生态（vLLM, SGLang）日趋成熟，Intel 和 Apple Silicon 的支持成为新战场。开发者需根据硬件环境评估工具的硬件适配成熟度。
  3. “本地优先”与云服务并行：Ollama 等强化本地体验，而 Copilot、Codex、Kimi 则深化云服务。选择取决于对延迟、隐私、数据主权和模型获取便利性的权衡。
  4. 企业需求专业化：对安全审计、成本控制、多团队隔离、私有化部署的需求已从边缘变为关键。在评估企业级工具时，需重点考察这些“非功能性”但至关重要的能力。
  5. 稳定性成为核心竞争力：频繁的回归 bug 和平台特定问题表明，在快速迭代中保障跨平台、跨版本的稳定性是赢得开发者信任的关键。新工具的早期用户应优先考虑社区活跃度和修复响应速度。

Per-Tool Reports

Claude Code — anthropics/claude-code

Claude Code Skills Highlights

Source: anthropics/skills

Claude Code Skills Community Highlights Report

Data as of 2026-07-23

1. Top Skills Ranking (Most-Discussed PRs)

  1. skill-creator Evaluation Toolchain Fixes (PR #1298)

    • Functionality: Addresses critical bugs in run_eval.py causing 0% recall reporting, breaking the description-optimization loop. Fixes Windows compatibility and trigger detection.
    • Discussion Focus: Foundational tooling issue blocking the improvement of all skills. Multiple PRs (#1099, #1050, #1323) and issues (#556, #1169, #1061) address related Windows and logic bugs.
    • Status: [OPEN]

  2. self-audit Reasoning Quality Gate (PR #1367)

    • Functionality: A meta-skill that audits AI output through mechanical file verification followed by a four-dimension reasoning quality audit. Aims to be a universal delivery quality check.
    • Discussion Focus: Represents a sophisticated approach to AI output verification and governance, reflecting demand for higher reliability.
    • Status: [OPEN]

  3. document-typography Quality Control (PR #514)

    • Functionality: Prevents common typographic flaws in AI-generated documents, such as orphan words and widow paragraphs.
    • Discussion Focus: Highlights a niche but high-quality demand for professional document production beyond basic content generation.
    • Status: [OPEN]

  4. testing-patterns Comprehensive Guide (PR #723)

    • Functionality: A broad skill covering testing philosophy, unit tests, React component testing, and integration patterns using the "Testing Trophy" model.
    • Discussion Focus: Addresses a core development workflow need with a structured, opinionated approach.
    • Status: [OPEN]

  5. color-expert Design Knowledge (PR #1302)

    • Functionality: Provides expert knowledge on color naming systems, spaces (OKLCH, CAM16), and practical design theory.
    • Discussion Focus: Targets the specialized design and creative workflow of developers.
    • Status: [OPEN]

  6. frontend-design Clarity Improvement (PR 🦞 OpenClaw 生态日报 2026-04-19 gsscsd/big_model_radar#210)

    • Functionality: Revises the existing frontend-design skill to make its instructions more specific, actionable, and coherent for Claude to follow.
    • Discussion Focus: Focuses on improving skill quality and effectiveness through better prompt engineering.
    • Status: [OPEN]

  7. skill-quality-analyzer and skill-security-analyzer (PR 📈 AI 开源趋势日报 2026-03-23 gsscsd/big_model_radar#83)

    • Functionality: Two meta-skills to evaluate the quality (structure, docs, etc.) and security of other skills.
    • Discussion Focus: Signals a need for governance and standardization as the skills library grows.
    • Status: [OPEN]

2. Community Demand Trends (from Issues)

3. High-Potential Pending Skills

4. Skills Ecosystem Insight

The community's most concentrated demand is for professional-grade reliability, governance, and cross-platform robustness of the core skills tooling and distribution model, which is currently perceived as fragile and centralized.

Claude Code Community Digest – 2026-07-23

1. Today's Highlights

The primary focus today is the fallout from the v2.1.218 release, which appears to have caused a regression disabling task-tracking tools (TaskCreate, TodoWrite) for many users, particularly on the Opus model. This is dominating the issue tracker. Additionally, long-standing concerns about Max plan session limits and platform-specific bugs continue to garner significant community engagement and frustration.

2. Releases

v2.1.218 (released within last 24h)

  • Background Code Reviews: The /code-review command now runs as a background subagent, preventing review output from cluttering the main conversation.
  • Accessibility Improvements: Added screen-reader announcements for word and line deletions.

3. Hot Issues

  1. #38335 - Max Plan Session Limits Exhausted Abnormally Fast (800 comments, 👍: 470)
    Why it matters: A persistent, high-frustration issue for paying subscribers, indicating a potential systemic problem with usage tracking or plan enforcement. The massive comment count shows unresolved pain.
  2. #80467 / #80487 - Task/Todo Tools Missing in v2.1.218 (New, active)
    Why it matters: A recent regression breaking core workflow functionality for many users. The linked issues point to a likely server-side model gate as the cause, affecting planning and task management.
  3. #80493 - Task Tools Missing on Opus but Present in Headless Mode (New)
    Why it matters: Highlights the inconsistency of the bug, suggesting complex interactions between model selection, session type, and client configuration.
  4. #80210 - Task/Todo Tools Not Exposed Despite Feature Flag (Open, 👍: 3)
    Why it matters: Precedes the latest version, indicating the tool disappearance issue began earlier and may involve account or feature gating.
  5. #24726 - VS Code: Disable Auto-attach of File/Selection (61 comments, 👍: 189)
    Why it matters: A long-running feature request for IDE integration, showing a desire for more control over the default assistant behavior to reduce unwanted context.
  6. #62699 - Linux: Cannot Copy Text from Output (34 comments, 👍: 51)
    Why it matters: A critical usability bug for Linux users, impacting the core function of interacting with generated code and output.
  7. #80139 - Windows Desktop: Filesystem Extension Not Dispatching (New, regression)
    Why it matters: A severe regression in a recent build that breaks file system access, a fundamental tool for coding agents.
  8. #57371 - Windows: Disable Cowork Background Service (13 comments, 👍: 37)
    Why it matters: Reflects a user preference for minimal installations and control over background processes, especially for those not using the Cowork feature.
  9. #16037 - Bash Permission Check Fails on Multi-line Scripts (Open, has repro)
    Why it matters: A long-standing bug (since Jan 2026) affecting the security plugin, now with an open PR fix (#80495), demonstrating community-driven patching.
  10. #77144 - Prompt Suggestions Stopped Appearing (Windows) (Open)
    Why it matters: A regression affecting a key productivity feature (Tab-completion suggestions), with no apparent fix after multiple weeks.

4. Key PR Progress

  1. #80495 - Fix(ralph-wiggum): Stop parsing prompt as shell code (Open)
    Fixes the long-standing #16037. Addresses the core issue of unsafe prompt text substitution in the /ralph-loop command.
  2. #80353 - docs(gcp): Stop on checksum mismatch (Open)
    Improves deployment security by halting the GCP gateway setup if the downloaded binary fails verification.
  3. #80326 - Add account profiles plugin (Open)
    A community contribution to manage isolated CLAUDE_CONFIG_DIR environments, aiding users with multiple (personal/work) accounts.
  4. #80112 - Make devcontainer firewall init resilient (Open)
    Hardens development environment setup by preventing a single DNS failure from aborting the entire firewall initialization.
  5. #80294 / #80229 - docs: fix broken links via archive.org (Open)
    Automated or community-driven maintenance to fix broken external links in documentation.
  6. #80241 - fix: Console scrolling to top of history (Open)
    Addresses a UI bug where console output resets the scroll position, disrupting reading flow.
  7. #80196 - fix: Auto-compact never triggering (Open)
    Aims to fix an issue where context compaction fails to activate, leading to potential context window overflows.
  8. #80195 - fix: Instantly hitting usage limits with Max (Open)
    A community attempt to address the client-side aspects of the widely-reported #38335 session limit issue.
  9. #18217 - feat: add /planwith command (Closed)
    Though closed, this PR discussed a popular UX improvement: allowing inline arguments for plan mode activation to reduce friction.
  10. #80094 / PR for #16037
    The closed issue #78315 on browser tool permissions also highlights ongoing refinements to the security and permission model for integrated tools.

5. Feature Request Trends

The community is clearly pushing for enhancements in three main areas:

  1. IDE Integration Refinement: Requests to control auto-context attachment (#24726) and add "selection-as-context" (#72301) aim to make the VS Code extension more configurable and interactive.
  2. Desktop Application Control & Polish: Users want more granular control over background services (#57371) and missing UI elements (#78754, #79404), emphasizing a desire for a stable and predictable desktop experience.
  3. Cross-Platform Feature Parity & Bug Fixes: Persistent issues on Linux (#62699) and Windows (#77144, #77754) indicate a demand for better testing and support for core functionality (copy, prompts, connections) across all platforms.

6. Developer Pain Points

  1. Regressions & Feature Instability: The sudden loss of task-tracking tools post-update is the primary source of immediate frustration, eroding trust in new releases.
  2. Platform-Specific Bugs: Long-standing, unresolved bugs on Linux (text copy) and Windows (prompt suggestions, daemon hangs, filesystem tool dispatch) hinder adoption and daily use.
  3. Uncertain Limits & Billing: The ongoing saga of Max plan session limits (#38335) creates financial and workflow uncertainty, damaging the value perception of paid plans.
  4. UI Inconsistencies & Disappearing Features: Reports of filters vanishing (#78754, #79404) and inconsistent tool availability (#80493) point to challenges in maintaining a stable, predictable user interface across updates and models.
  5. Need for Greater Customization: Requests to disable auto-attach, background services, and improve hooks suggest power users are bumping up against default behaviors and desire more control.
OpenAI Codex — openai/codex

OpenAI Codex Community Digest | 2026-07-23

1. Today's Highlights

Two new alpha releases of the Rust-based Codex core (v0.146.0-alpha.3 & alpha.4) were pushed today, signaling ongoing rapid development. Community activity is heavily focused on cross-platform stability, with a surge of detailed bug reports from Windows and macOS users highlighting performance regressions, authentication issues, and UI glitches following recent updates.

2. Releases

  • rust-v0.146.0-alpha.4 & rust-v0.146.0-alpha.3: Both are pre-release versions indicating work on the core Rust components of Codex. Specific changelogs are not provided in the data, but their rapid succession points to active bug-fixing or feature iteration on the backend/CLI layer.

3. Hot Issues

  1. Voice Transcription for IDE (#3000): The most popular enhancement request (196👍), asking for a push-to-talk button in the IDE panel to dictate prompts. The community is eager for more natural interaction modes.
  2. Windows Defender & WMI CPU Spike (#33875): A critical performance issue (16👍, 14 comments) where launching Codex on Windows 10 causes sustained high CPU usage from system processes. This is a major user experience and system resource concern.
  3. Admin Sandbox Setup Failure (#14808): A longstanding (since March) configuration issue preventing users on Windows with "ChatGPT Business" subscriptions from setting up the admin sandbox correctly, blocking core functionality.
  4. MCP File Descriptor Leak (#26988): A severe stability bug in the CLI where MCP stdio servers leak file descriptors and orphan processes, eventually leading to a fatal "Too many open files" error during long sessions.
  5. SSH Remote "No Chats" Regression (#30808): A regression in v0.142.5 causing remote projects opened via symlinks to show an empty chat history, disrupting remote development workflows.
  6. Business Subscription Auth Loop (#28672): A usability blocker for some "ChatGPT Business" users in the US, where Codex returns repeated 401 authentication errors after only a few messages, rendering the service unusable.
  7. macOS Sidebar Freeze (#34376): A UI performance bug on macOS where hovering/clicking the sidebar triggers a 3–10 second freeze, linked to recursive file system watcher teardown.
  8. Model Compatibility Error (#30422): Users report getting "This model is not supported" errors when attempting to use the latest models from the picker, indicating a potential backend/model compatibility issue.
  9. Computer Use Tools Missing on Windows (#33648): Despite enabling the Computer Use feature, the necessary runtime/control tools are not exposed in sessions on the Windows app, preventing its use.
  10. Desktop Automation Timezone Bug (#26633): Scheduled RRULE automations compute the next execution time as if it's in UTC, ignoring the specified TZID, causing unexpected timing for users in different time zones.

4. Key PR Progress

  1. Standalone Installer Improvement (#34910): Makes shell/PowerShell installers prioritize downloading from releases.openai.com with better fallback, improving installation reliability and speed.
  2. Multi-Agent Wait Tool Control (#34887): Adds a setting to disable the collaboration.wait_agent tool, giving developers more control over multi-agent interaction patterns.
  3. MCP HTTP User-Agent (#34883): Sets a default codex-mcp-client user-agent for MCP HTTP requests, improving telemetry and compatibility.
  4. Retry on Model Capacity Errors (#31176): Introduces automatic retry for "model capacity" errors to prevent goals from stopping unnecessarily, reducing user intervention.
  5. Permissions-Scoped Exec Rules (#29500): Makes command approval rules aware of the active permissions profile (e.g., managed, sandbox), enhancing security and context-awareness.
  6. Plugin Cache & Metadata Enhancements (#34849, #34851): Caches remote plugin catalogs and uses batch metadata loading, aiming to improve plugin list responsiveness and reliability.
  7. Web Search for Custom Providers (#34846): Allows custom model providers to opt into using the standalone web search tool, expanding its utility beyond first-party models.
  8. Thread Pinning (#34840): Adds server-side support for pinning threads, a feature to help users keep important conversations at the top of their list.
  9. Multi-Agent Mode State Tracking (#34845): Makes multi-agent mode instructions durable in "world state," preventing them from being lost or repeated after history changes.
  10. Image Gen for Free Plans (#34850): Explicitly disables the image_generation tool for Free-plan accounts, clarifying feature availability.

5. Feature Request Trends

  • Voice & Collaboration: The top request is for voice input (#3000), with a separate proposal for duplex voice collaboration ("Codex Live") to attach Live Mode to active sessions (#33966).
  • Enhanced IDE Integration: Requests focus on deeper integration, such as exposing the interactive user-question tool outside Plan Mode (#30150) and showing iOS Live Activities for tasks running on macOS (#34912).
  • Advanced Automation & Scheduling: Users want more reliable and feature-rich scheduling for desktop automations, beyond basic RRULE support.

6. Developer Pain Points

  • Windows Stability & Performance: A cluster of high-traffic issues (#33875, ##33940, #34107) point to critical system-level problems on Windows, including CPU spikes, process crashes, and high WMI host usage, significantly impacting the user experience.
  • macOS Performance Regressions: Similar to Windows, macOS users report new performance regressions, such as severe UI freezing (#34376) and issues with bundled assets causing disk leaks (#34567).
  • Authentication & Session Management: Persistent and disruptive bugs related to authentication loops for business users (#28672) and disappearing chat histories after updates (#30808, #24364) are major reliability concerns.
  • Cross-Platform Feature Parity: Features like Computer Use (#33648) and file handling in macOS apps (#32105) are reported as non-functional or inconsistent across platforms, creating a fragmented experience.
Gemini CLI — google-gemini/gemini-cli

Gemini CLI Community Digest - 2026-07-23

1. Today's Highlights

The Gemini CLI ecosystem saw a significant push in agent reliability and developer tooling, highlighted by the v0.53.0-preview.0 release which addresses critical A2A protocol bugs and introduces the foundational "Caretaker Triage" orchestration system. Core improvements continue to focus on authentication resilience and evaluation infrastructure.

2. Releases

  • v0.52.0 (Stable): This release includes workspace cleanup by excluding transient CI files and adds the foundational modules for the new caretaker-triage system. (PR #28216, PR #28301)
  • v0.53.0-preview.0 (Preview): A key bugfix release that resolves a 400 Bad Request error caused by incorrect handling of cancelled tool responses and role coalescing. It also implements the core "LLM triage orchestrator" for the caretaker-triage feature. (PR #28407, PR #28441)
  • v0.52.0-nightly.20260723.g9681621c6 (Nightly): The nightly build restores the GOOGLE_APPLICATION_CREDENTIALS fallback for credential verification and adds a new eval:coverage command for assessing tool test coverage. (PR #28472, PR #28169)

3. Hot Issues

  1. Subagent Recovery Misreported (#22323): A priority bug where a subagent hitting its turn limit falsely reports success as "GOAL". This hides operational failures. (👍 2, 12 comments)
  2. Generalist Agent Hangs (#21409): A persistent and highly-voted issue causing the generalist agent to hang indefinitely on simple tasks, significantly impacting core functionality. (👍 8, 8 comments)
  3. 400 Error with >128 Tools (#24246): A technical limit hit when a large number of tools are enabled, suggesting a need for better tool scoping and prioritization. (7 comments)
  4. Shell Command Execution Hang (#25166): CLI commands incorrectly report "Waiting input" after completion, blocking workflow. (👍 3, 4 comments)
  5. Browser Subagent Fails on Wayland (#21983): A platform-specific bug preventing the browser agent from functioning on Wayland display servers. (👍 1, 4 comments)
  6. Lack of Skill/Sub-Agent Utilization (#21968): Users report the model doesn't proactively leverage custom skills or sub-agents without explicit instruction. (6 comments)
  7. Auto Memory Indefinite Retry (#26522): The Auto Memory system retries low-signal sessions endlessly instead of quarantining them, impacting performance. (5 comments)
  8. Browser Agent Config Ignored (#22267): The browser agent does not respect settings.json overrides like maxTurns. (3 comments)
  9. Get-Shit-Done Output Crash (#22186): A reproducible crash occurs when the output hook finishes, disrupting long-form task execution. (👍 2, 3 comments)
  10. Destructive Agent Behavior (#22672): Concerns over the agent using risky commands (e.g., git reset --force) without safer alternatives. (👍 1, 3 comments)

4. Key PR Progress

  1. Security: Block Variable Expansion Bypass (#28403): Critical security fix hardening the detection of shell substitution patterns ($VAR, ${VAR}). (Priority/P1)
  2. Core: Shorten MCP Discovery Timeout (#28410): Prevents the CLI from hanging for up to 10 minutes during startup if an MCP server is unresponsive. (Priority/P1)
  3. Core: Fix Model Availability for Sub-Agents (#28406): Ensures utility tools (like web-search) correctly resolve model IDs, fixing API errors for non-preview users. (Priority/P1)
  4. Core: Rotate Session ID on Fallback (#28469): Fixes a blocking API error when a model falls back to gemini-2.5-flash by rotating the session ID. (Priority/P2)
  5. CLI: Add gemini-3.5-flash to Model Selector (#28485): Makes the latest Flash models selectable in the UI for all users. (Priority/P2)
  6. CLI: Improve CJK Markdown Rendering (#28309): Solves text wrapping and list misinterpretation issues for CJK languages in the terminal. (Status/Need-Issue)
  7. Core: Filter Thought Parts from History (#28509): Prevents internal monologue/thinking parts from leaking into the chat history when context management is off. (Status/Need-Issue)
  8. CLI: Add Windows PowerShell Troubleshooting (#28447): Documentation improvement addressing a common installation issue on Windows. (Priority/P2)
  9. Core: Override google-auth-library Version (#28404): Addresses dependency version conflicts. (Size/XS)
  10. CLI: Propagate AbortSignal in /compress (#28506): Allows cancellation of the background chat compression command, preventing dangling requests.

5. Feature Request Trends

  • Advanced Sub-Agent Control & Observability: Strong demand for better control over agent permissions, visibility into sub-agent trajectories (#22598), and more accurate state reporting.
  • AST-Aware Code Navigation: Multiple open issues (#22745, #22746) investigate using AST tools for more precise code reading, searching, and mapping to improve agent efficiency.
  • Memory & Context System Improvements: A series of issues (#26516, #26523, #26525) focus on enhancing the reliability, security, and quality of the Auto Memory system.
  • Enhanced Tool & Agent Safety: Requests for sandboxing, discouragement of destructive actions, and better handling of interactive prompts (#22465, #22672, #19873).

6. Developer Pain Points

  • Agent Reliability & Hangs: The most critical pain point is agents hanging indefinitely (#21409) or failing silently, severely impacting trust and usability.
  • Shell & Editor Integration: Recurring issues with shell command execution state (#25166), external editor corruption (#24935), and handling of interactive prompts (#22465).
  • Configuration & Environment Issues: Challenges with model selection availability (#28483), platform-specific failures (Wayland #21983), and inconsistent application of settings (#22267).
  • Memory & State Management: Frustration with the Auto Memory system's retry loops (#26522) and lack of user visibility into its operations (#26523).
GitHub Copilot CLI — github/copilot-cli

GitHub Copilot CLI Community Digest (2026-07-23)

1. Today's Highlights

The latest releases (v1.0.74-x) introduce a first-run splash screen and add support for the gemini-3.6-flash model, alongside a fix for session multiplexing UI leaks. However, community discussion is dominated by significant ongoing issues around session history bloat from oversized attachments or binary diffs exceeding the CAPI 5MB limit, with multiple reports confirming recovery difficulties and regressions.

2. Releases

Three new patch versions were published in the last 24 hours. The primary feature update is in v1.0.74-1:

  • Added: A first-run splash screen for users to opt into the default sandbox.
  • Added: Support for the gemini-3.6-flash model.
  • Improved: Fixed a session multiplexing bug where an open dialog from one session would leak into another.
  • Note: The $ interactive shell shortcut change is listed but the description is truncated. The other two releases (-2, -3) are labeled as "Fixes and changes."

3. Hot Issues

  1. #3767 / #4097 - Oversized Attachment/Binary Diff Permanently Wedges Session: A critical usability blocker. Attachments or diffs (e.g., from apply_patch deleting a binary) that push the request over the 5MB CAPI limit cause unrecoverable errors, even with /compact. #3767 is closed but indicates the core limit persists; #4097 highlights a common trigger. (12 & 4 comments, 1 & 5 👍).
  2. #443 - Feature Request: Built-in PDF Reading Support: A long-standing, highly-voted (33 👍) request for native PDF handling, a key limitation for users working with documentation and research.
  3. #4016 - BYOK Rejected in --acp Mode (Regression): A regression affecting enterprise/power users. Custom providers (COPILOT_PROVIDER_*) work in -p mode but are rejected in the Agent Client Protocol (--acp --stdio) flow, requiring GitHub login unexpectedly. (5 comments, 4 👍).
  4. #3534 - WSL2 (ARM64): /copy fails with clip.exe error: A platform-specific bug preventing clipboard writes in WSL2 on ARM64 systems, caused by a cmd.exe quoting issue. (5 comments, 4 👍).
  5. #4206 - Environment Footer Stuck on "Loading:": A UI/UX bug under enterprise org MCP policies where the status footer never resolves, leaving users uncertain if their environment is ready. (3 comments, 2 👍).
  6. #4089 - Atlassian MCP Server: OAuth Succeeds but Zero Tools Exposed: A significant integration failure with a major vendor's MCP server, despite successful authentication. (4 comments, 0 👍).
  7. #4163 - copilot CLI Does Not Reap Child Processes (Zombies): A stability concern on Linux where shell tool usage leaks zombie processes, accumulating over time. (3 comments, 2 👍).
  8. #4161 - task_complete tool unavailable after autopilot mode switch: A regression where a key agent tool becomes unavailable after mode changes, impacting workflow. (2 comments, 1 👍).
  9. #1688 - Add Configurable Auto-compaction Threshold: A persistent feature request (5 👍) for performance tuning, especially for slower, high-capacity models where default compaction triggers too late.
  10. #4211 - Couldn't handle BigInt in structured MCP response: A new data serialization bug causing agent task failures with certain MCP servers.

4. Key PR Progress

Only two Pull Requests were active in the last 24 hours, limiting this section:

  • #4228 [CLOSED] Withdrawn: incorrect scope for #3534: The community contributor's fix for the WSL2 clipboard issue was withdrawn because it targeted documentation rather than the private runtime implementation.
  • #3163 [OPEN] ViewSonic monitor: An older, seemingly stale PR related to a monitor issue (#2591, #3561, #3559), with a mention of a GitHub Action.

5. Feature Request Trends

Community demand coalesces around several key directions:

  • Enhanced File & Data Processing: Native PDF support (#443) and improved handling of large attachments/diffs to prevent session breaks.
  • Advanced Context & Session Management: Configurable auto-compaction thresholds (#1688) and mechanisms to gracefully handle context limits.
  • Deeper MCP & Tool Integration: Resource subscriptions (#3073), inheriting tools from connected VS Code instances (#4143), and resolving tool exposure bugs (#4089).
  • Cross-Platform & Environment Robustness: Fixes for platform-specific issues on Windows/WSL2 (#3534) and Linux (#4163), alongside better shell integration (#3428).

6. Developer Pain Points

Recurring frustrations center on stability and predictability:

  • Unrecoverable Session States: The 5MB CAPI limit causing permanent session wedges is a major pain point (#3767, #4097).
  • Regression & Stability Concerns: Reports of regressions in authentication (#4016), tool availability (#4161), and terminal rendering (#4222) erode confidence.
  • Platform-Specific Quirks: Persistent issues on specific combinations like WSL2/ARM64 and Windows native terminal remain unresolved and block core workflows.
  • Permission System False Positives: Overzealous or misinformed permission prompts for read-only operations (#4220, #4221) interrupt agent planning and execution.
  • MCP Integration Gaps: Basic features like stable tool exposure and resource notifications are not yet reliable, hindering the ecosystem.
Kimi Code CLI — MoonshotAI/kimi-cli

Kimi Code CLI Community Digest (2026-07-23)

Today's Highlights

A significant maintenance push is underway, led by developer @lihailong00, addressing cross-platform stability (Windows logging, process termination) and robustness for the MCP plugin ecosystem. A critical bug causing session blocks in the kimi-datasource plugin was reported, alongside a new community feature request for cost-optimized multi-agent workflows.

Releases

No new releases in the last 24 hours.

Hot Issues

  1. #2538 [Bug] kimi-datasource plugin worker pool blocks all sessions on timeout

    • Why it matters: This is a critical multi-session blocking bug. When using the yahoo_finance API through the datasource plugin, a timeout in one session can deadlock all others that share the worker pool, crippling concurrent workflows.
    • Community Reaction: 0 comments, but the impact is high for power users running parallel data-fetching tasks.

  2. #2534 [Bug] Model API error 400 Validation: Unsupported parameter prompt_cache_key

    • Why it matters: Breaks compatibility with third-party Kimi-compatible APIs (e.g., Nvidia NIM models) after a recent update. The CLI is sending a Moonshot-specific parameter (prompt_cache_key) that external providers don't understand.
    • Community Reaction: Reported immediately after the update, indicating a regression affecting a segment of the user base.

  3. #2533 Feature Request: Per-agent model selection for sub-agents

    • Why it matters: Proposes a key architectural enhancement for cost and performance optimization. It would allow sub-agents (e.g., for simple data formatting) to use cheaper models, while complex coding agents use more capable ones.
    • Community Reaction: This is the first feature request of the day, showing a forward-looking interest in optimizing agent economics.

Key PR Progress

  1. #2543 fix(hooks): notify on permission prompts
    Improves the hook system to emit notifications when manual approval is required, enabling better workflow automation.
  2. #2541 fix(mcp): continue after deferred startup failure
    Enhances resilience by preventing a background MCP server startup failure from crashing an interactive session.
  3. #2542 fix(logging): isolate Windows process log files
    Fixes a Windows-specific bug where concurrent CLI instances would interfere with each other's log files.
  4. #2544 fix(kaos): terminate local process trees
    Improves process management on Windows by correctly terminating entire local process groups on cancellation or timeout.
  5. #2539 fix(mcp): normalize tools for Moonshot API
    Ensures tools from MCP servers have stable aliases compatible with the Moonshot API while preserving the original names for internal routing.
  6. #2537 fix(shell): support numeric keypad input
    Adds support for numeric keypad digits from Windows Terminal, fixing a specific UI input issue.
  7. #2536 fix(web): make server banner encoding-safe
    Fixes a display bug with the server banner on non-UTF-8 terminals (e.g., GBK encoding on Windows).
  8. #2535 fix(llm): scope prompt cache keys to Moonshot APIs
    Directly addresses issue #2534 by restricting the prompt_cache_key parameter to only official Moonshot/Kimi APIs.
  9. #2524 fix(tools): count StrReplaceFile replacements correctly
    Fixes a counting bug in sequential file edits where replacements based on prior edits were incorrectly tallied against the original file content.
  10. #2530 fix(shell): stop blocking until timeout for detached children
    Fixes a shell command hang where the CLI would block indefinitely waiting for stdout/stderr EOF from a detached child process before checking its exit code.

Feature Request Trends

  • Multi-Agent Optimization & Cost Control: The request for per-agent model selection (#2533) is the primary trend, indicating a desire for finer-grained control over resource allocation in complex, multi-step agent pipelines.

Developer Pain Points

  1. Session Stability & Concurrency: A major frustration point is session blocking and crashes. The worker pool deadlock in the datasource plugin (#2538) and the need for better process group termination (#2544) highlight instability in concurrent and multi-session scenarios.
  2. Cross-Platform Reliability (Windows): Multiple PRs (#2542, #2544, #2537, #2536) target Windows-specific bugs, indicating ongoing friction for developers on that platform regarding logging, input handling, and process management.
  3. Third-Party API Compatibility: The regression in #2534 shows that updates can inadvertently break compatibility with non-Moonshot backends, a pain point for users leveraging the CLI with different model providers.
OpenCode — anomalyco/opencode

OpenCode Community Digest (2026-07-23)

1. Today's Highlights

No new official releases were published in the last 24 hours. Development activity is heavily focused on core infrastructure migration for the "v2" application layer, with numerous PRs from a single contributor working to modernize the app's data and communication protocols. A critical authentication bug continues to disrupt access to premium "OpenCode Go" subscription models for several users.

2. Releases

There were no new releases or version updates for OpenCode in the last 24 hours.

3. Hot Issues

Here are 10 noteworthy issues with significant community engagement:

  1. #38195 - 401 AuthError: Request blocked by upstream provider (👍: 15)

    • Why it matters: Users with active "OpenCode Go" subscriptions are unable to use paid models, facing a 401 error. This is a service-blocking issue affecting paying customers.
    • Community reaction: High concern from multiple users reporting the same issue across different platforms (Desktop, Hermes, Windows/macOS).

  2. #10119 - VSCode extension "OpenCode Beta" - "There is no data provider registered" (👍: 14)

    • Why it matters: A long-standing bug (created Jan 2026) in the VSCode extension renders it non-functional for many users, showing a blank panel with an error.
    • Community reaction: Persistent frustration, as this limits the utility of the official IDE integration.

  3. #37716 - Internal Server Error (👍: 5)

    • Why it matters: A generic but frequently reported error indicating backend instability or unhandled exceptions in the core service.
    • Community reaction: Users are seeking clarity on the root cause, as the error message provides little diagnostic information.

  4. #14292 - [FEATURE]: Save conversations and session data to project folder (👍: 20)

    • Why it matters: A highly requested feature for better project portability and session data management. Storing data globally (~/.opencode) is inconvenient for project-scoped work.
    • Community reaction: Strong agreement on the need for project-local data storage, with 20 upvotes.

  5. #31932 - [FEATURE]: Cross-project session list / picker for TUI (👍: 4)

    • Why it matters: Developers working across multiple repositories need a way to manage sessions globally, not just within a single project scope.
    • Community reaction: Positive reception from developers managing multiple codebases.

  6. #29757 - Opencode + Qwen gives JSON output instead of writing to disk (👍: 2)

    • Why it matters: Indicates a breakdown in the tool-use (function calling) mechanism for certain local LLMs, preventing automated file operations.
    • Community reaction: Concerns about reliability and compatibility with popular open-source models.

  7. #21277 - [Windows] Terminal left with raw ANSI escape codes after crash (👍: 0)

    • Why it matters: A usability and stability issue specific to Windows users, leaving their terminal in a broken state after an OpenCode crash.
    • Community reaction: Ongoing bug report seeking a fix for terminal state cleanup.

  8. #36424 - [2.0] v2: shell: no handle at launch to stop/cancel running background commands (👍: 2)

    • Why it matters: A regression or missing feature in v2 affecting user control over long-running processes, reducing interactivity.
    • Community reaction: Highlighted as a regression from v1 functionality.

  9. #38470 - MCP tools missing from sessions created via POST /api/session on opencode serve (👍: 0)

    • Why it matters: A critical bug for programmatic/headless usage (opencode serve), where the API fails to initialize MCP server tools, limiting automation.
    • Community reaction: Affects developers integrating OpenCode into custom workflows or CI/CD pipelines.

  10. #38471 - Desktop migrates shared DB; older CLI on PATH fails (👍: 0)

    • Why it matters: A data migration conflict between Desktop and CLI versions can break workflows for users with both installed, causing schema mismatch errors.
    • Community reaction: Reports of unexpected errors after a Desktop update.

4. Key PR Progress

Here are 10 important Pull Requests shaping the codebase:

  1. #38459 - feat(app): project current server state by @Brendonovich

    • Description: A major piece of the v2 migration. It normalizes core data models (sessions, messages, projects, providers) to work with a new, protocol-agnostic API facade.

  2. #38460 - feat(app): support current review data by @Brendonovich

    • Description: Migrates the code review UI components to accept data from the new "current" API response format, ensuring feature parity during the transition.

  3. #38461 - feat(app): migrate session interactions by @Brendonovich

    • Description: Rewrites the session interaction layer (prompts, commands, archival, permissions) to use the new v2 API contracts.

  4. #38465 - feat(app): migrate discovery workflows by @Brendonovich

    • Description: Updates provider, project, directory, and MCP discovery logic to function with both v1 and v2 server protocols.

  5. #38466 - feat(app): render current session timeline by @Brendonovich

    • Description: Adapts the session timeline renderer to project and display messages from the new v2 data format.

  6. #38463 - feat(app): support current pty transport by @Brendonovich

    • Description: Migrates PTY (pseudo-terminal) lifecycle management and WebSocket routing to the compatible v2 API.

  7. #12679 - feat(tui): vim motions in prompt input by @leohenon

    • Description: (Closed) Adds optional Vim-style keybindings for power users in the terminal prompt, enhancing TUI navigation efficiency.

  8. #38298 - feat(session-ui): collapsible JSON tree view by @jiong-nba

    • Description: Enhances the context panel by replacing flat JSON blocks with an interactive, collapsible tree view for easier inspection of raw message data.

  9. #33656 - fix(glob): time out slow searches by @lin-hongkuan

    • Description: Adds a 30-second timeout to the file globbing tool to prevent hangs and provide clear failure states on large or slow repositories.

  10. #36796 - fix(opencode): wait for shell output capture by @MaartenBaert

    • Description: Fixes a race condition where shell output could be lost if the process closed before its output capture fiber completed.

5. Feature Request Trends

The most requested feature directions from the community are:

  1. Data Portability & Project Scope: A strong trend towards storing session and conversation data locally within the project folder instead of a global location, for better version control, sharing, and isolation.
  2. Enhanced Session Management: Users want better tools for navigating and managing sessions across multiple projects, including a cross-project session picker and improved UI for session metadata (e.g., using description fields as titles).
  3. Improved UI/UX & Compatibility: Requests for a more standard, accessible UI (e.g., VSCode-like), better rendering of markdown/code blocks in the TUI, and fixes for platform-specific issues (Windows terminal state, file attachment paths).
  4. Plugin & Extension Ecosystem: Interest in more powerful plugin capabilities, such as exposing global broadcast events for cross-plugin communication and better MCP tool integration.

6. Developer Pain Points

Recurring frustrations and high-frequency requests include:

  1. Authentication & Subscription Reliability: The most acute pain point is the inability to use paid "OpenCode Go" models, breaking core value for subscribers. Generic "Internal Server Errors" also undermine trust in platform stability.
  2. TUI Rendering & State Issues: Multiple reports of broken rendering (ANSI codes after crash, incorrect markdown/JSON display) and state management problems (/undo during generation) harm the terminal user experience.
  3. IDE Integration Gaps: The VSCode extension's long-standing non-functional state is a major barrier for developers who prefer integrated editor workflows.
  4. Cross-Platform Inconsistencies: Windows users face specific stability issues (terminal corruption, path handling), indicating a need for better platform-specific QA and testing.
  5. API & Programmatic Use Friction: Bugs affecting headless usage (opencode serve missing MCP tools) and version conflicts (Desktop/CLI DB migrations) create obstacles for developers building on top of or automating OpenCode.
Qwen Code — QwenLM/qwen-code

Qwen Code Community Digest - 2026-07-23

1. Today's Highlights
The community's focus today is on expanding integration capabilities and fixing critical compatibility issues. A major proposal for an enterprise external-memory integration profile (#7449) aims to standardize how Qwen Code connects with external knowledge bases. Simultaneously, significant effort is being directed at resolving npm update check failures on newer Node.js/npm versions (#7520, #7543), which is a critical toolchain blocker for many users.

2. Releases
No new stable or nightly releases were published in the last 24 hours.

3. Hot Issues

  • #7449 - Enterprise Memory Profile: Proposes a standard, provider-neutral profile for integrating external memory systems. Crucial for enterprise adoption and complex workflows. Awaiting core team feedback. (Link)
  • #6014 - Missing Agent File Read Info: Users report a UI regression where the agent only shows "read 1 file" without the filename. Impacts debugging and transparency. Status: needs information. (Link)
  • #6806 - Context Usage Not Updating: The token usage percentage in the status line doesn't refresh after /compress. Affects user awareness of context limits. Status: needs-triage. (Link)
  • #7520 & #7543 - npm 12 / Node 26 Update Check Failures: Two separate reports confirming that qwen update fails with "registry error" due to incompatibilities with npm 12's output format and mise's bash wrapper. A high-impact bug for users on latest toolchains. Both are welcome-pr. (#7520, #7543)
  • #7575 - User-level Skills Not Loaded in Serve Mode: Skills from ~/.qwen/skills/ are ignored when running qwen serve --channel, breaking expected precedence. A priority P2 bug. (Link)
  • #7566 - Monitor Stop Triggers Auto-Notification: Stopping a monitor incorrectly generates a user notification that triggers a model turn, creating unexpected behavior. A priority P2 core bug, now closed. (Link)
  • #7585 - Direct External Context Provider: A feature request for a new extension to search trusted repository corpora via Mem0 or a generic HTTP contract, complementing the memory profile proposal. (Link)
  • #7515 - Update Check Failure: Earlier report of the "registry error" during updates, which has been addressed by a subsequent PR (#7528). (Link)
  • #7555 - Nightly Release Workflow Failure: The automated release for a nightly build failed at the integration_docker job, halting deployment. (Link)
  • #7167 - Fleet Shepherd Dashboard: A maintenance dashboard issue updated by the bot, reflecting CI/CD pipeline activity. Status: need-information. (Link)

4. Key PR Progress

  • #7381 - Fix Queued Message Display: Corrects styling and ordering for user input that arrives while the model is responding, improving UX clarity. (Link)
  • #7557 - Persist Usage for Tool-only Rounds: Ensures token usage metadata is recorded for sub-agent rounds that contain only tool calls, improving accounting for trajectory exporters. (Link)
  • #7534 - Retry on Thinking Requirement: Introduces automatic retry logic when a provider returns HTTP 400 because enable_thinking must be true, improving resilience. (Link)
  • #7577 - Workspace-scoped Channel Lifecycle: Implements workspace-level management for named channels, allowing start/stop/restart within a specific workspace context. (Link)
  • #7586 - Direct External Context Provider: The implementing PR for the feature proposed in issue #7585, adding the private integration for searching a trusted repository corpus. (Link)
  • #7388 - Explicit Channel Delivery for Daemon: Adds a delivery contract for routing daemon notifications and task results to specific named channels and users. (Link)
  • #7554 - Autofix: Auto-update Red PRs: Enhances the autofix workflow to automatically merge main into a PR that is failing only due to a stale, previously broken base commit. (Link)
  • #7580 - Visualize Plan Execution: Connects the "Todo" plan system with live sub-agent execution for ordinary sessions, providing real-time progress tracking. (Link)
  • #7569 - Bilingual Review Comments: Makes the review-address report comment fully bilingual (English/Chinese), improving accessibility for the global community. (Link)
  • #7539 - Clean Orphaned NPM Update Artifacts: Adds a cleanup step before managed npm updates to remove stale files from interrupted previous updates, preventing potential update conflicts. (Link)

5. Feature Request Trends
The dominant trend is a strong push for deeper, standardized integration with external systems. This is exemplified by proposals for both an enterprise memory profile (#7449) and a direct external context provider (#7585). There is also continued interest in enhancing the CLI and Web Shell experience, with PRs focused on better git mode selection (#7471), advisor commands (#7567), and version upgrade notices (#7542).

6. Developer Pain Points
Recurring frustrations center on toolchain stability and UI consistency. The most severe current pain point is the npm update check failing on Node 26/npm 12, completely blocking the update mechanism for affected users. Other notable frustrations include UI regressions like missing file read information (#6014) and non-refreshing token usage (#6806), as well as broken functionality in specific modes, such as user-level skills not loading during serve mode (#7575). The community is actively providing patches, but these core stability issues are high priority.

llama.cpp — ggml-org/llama.cpp

llama.cpp Community Digest – 2026-07-23

1. Today's Highlights

The project saw a burst of releases focusing on backend expansion (PowerPC/AIX, SYCL, WebGPU, CUDA) and architecture support (Laguna XS.2/M.1, DeepSeek4). Community discussions remain heated around server stability in parallel and SWA models, while new PRs tackle critical performance gaps on integrated GPUs and MoE models.

2. Releases

  • b10092: Enabled PowerPC backend variants for AIX (#25983).
  • b10091: Fixed SYCL package shared library lookup (#25987).
  • b10090: Added CONV_2D_DW (depthwise conv2d) kernel for WebGPU backend (#25847).
  • b10089: Added k-quant support to GET_ROWS in CUDA backend (#25962).
  • b10088: Fixed DeepSeek4 APE tensor operation (#25945).
  • b10087: Added support for Laguna XS.2 & M.1 chips (#25165).
  • b10085: Fixed align_corners for Qwen3VL vision position embedding interpolation (#25781).

3. Hot Issues

  1. #20837 [Bug] Tool calls in XML when thinking enabled (57 comments). Critical for chat parser reliability, causing workflow interrupts.
  2. #21831 [Bug] Server forces full prompt re-processing (47 comments). Major performance regression for SWA/recurrent models, impacting all server users.
  3. #14909 [Enhancement] Implement missing ops from backends (47 comments). Foundational for feature parity across all backends.
  4. **[#25436](http

⚠️ Content exceeds GitHub Issue limit. See the committed Markdown file for the full report.

Metadata

Metadata

Assignees

No one assigned

    Labels

    digest-en

    Projects

    No projects

    Milestone

    No milestone

    Relationships

    None yet

    Development

    No branches or pull requests

    Issue actions