Alex Baron

agent-relay-protocol.github.io/agrp-rfc/ Agent Relay Protocol (AGRP). Switching Fabric and Workspace Infrastructure for Agent Meshes. Draft.

When you're on track, stay on track. 10 days. 237 commits. 305,716 lines of code. The gap between what's possible and what most teams settle for has never been wider. --- On a media day Aleksei delivers what would take a typical senior dev ~47 days. On peak days (Mar 18) that stretches to ~712 human/days in a single calendar day. Sustained over 10 days at 43.9x his own median, that's roughly 5.6 years of typical industry output compressed into a week and a half.

CCBox remote — coming soon. Secure access to your coding sessions from anywhere. CCBox remote terminal is on the way. You'll be able to connect to your ccbox instance securely in two ways: - direct connection: point-to-point TLS tunnel between your client and the ccbox server, ideal for local network or VPN setups. - relay connection: connect through an authenticated relay server when direct access isn't possible (NAT, firewalls, mobile networks). End-to-end encrypted: the relay sees nothing. Once connected, the full ccbox experience is available remotely: - Spawn agents on your project from any device. - Plan and track tasks across sessions. - Watch project traces: session timelines, token usage, skill spans in real time. - Browse and resume sessions without SSH or port forwarding. No exposed ports. No credentials in plaintext. Just a secure channel to your running ccbox instance. Stay tuned.

#opensource diskd-ai/skills — a collection of skills for coding agents. When you work with AI agents every day, you build up experience: how to correctly call an API, structure code reviews, and avoid repeating mistakes. We've packaged this into open-source skills that can be installed with a single command. Compatibility: Claude Code, Cursor, Cline, Copilot, Codex, Windsurf, Gemini, and over 10 other agents through skills.sh. To install any skill: npx skills add diskd-ai/ --skill --global --yes API Integrations — ready-made templates for working with providers: - claude-api — Anthropic Messages API, tool use, vision, streaming, prompt caching, extended thinking - openai-api — Chat Completions, function calling, DALL-E, Whisper, Assistants API, fine-tuning - gemini-api — multimodal input (images, video, audio, PDF), thinking, structured outputs, Veo/Lyria - groq-api — ultra-fast inference, Whisper, Llama Guard, tool use - together-api — open-source models, FLUX, Stable Diffusion, Whisper, fine-tuning, reranking - cerebras-api — fast inference, reasoning models with thinking tokens, structured outputs - perplexity-api — Sonar models with real-time web search and citations, deep research - nebius-api — Token Factory, OpenAI-compatible API, batches, fine-tuning, custom models Development Tools: - code-review — structured review with importance tags, file+line, verdict. Works with GitHub PRs (gh) and GitLab MRs (glab) - design-doc — generation of TRD/RFC/ADR/specs upon request - prompting — prompt engineering: few-shot, chain-of-thought, role prompting, structured outputs - redmine-cli — manage Redmine tasks from an agent: create, edit, comment, link - tilt-config — Tiltfile, docker_build, live_update, k8s_resource, Helm/Compose workflow - assemblyai-cli — transcribe audio/video to text/srt/vtt with speaker diarization Other: - clarify — decompose and clarify tasks from the tracker (uses redmine-cli, design-doc) before starting work on the task - human-interface-guidelines-1992 — audit desktop UIs according to classic Apple 1992 guidelines (menus, dialogs, Balloon Help, accessibility) --- ccbox + ccbox-insights — a separate story ccbox is a TUI for viewing coding agent sessions (Claude, Codex, Gemini, OpenCode). It shows projects, timelines, and tool call histories. ccbox-insights, however, is the real reason this was created, but it evolved beyond that. Did you know that when developing with agents, up to 30% of time can be wasted on unnecessary work? An agent tries a path, makes mistakes, tries again, and the user corrects it — and the cycle repeats. ccbox-insights analyzes unsuccessful sessions and extracts lessons: - Finds recurring error patterns (invalid arguments, missing files, rejected actions, timeouts) - Classifies failures into a taxonomy of 12 categories (invalid_input, path_not_found, permission_denied, wrong_tool_scope...) - Evaluates the session's outcome and user satisfaction signals - Generates ready-to-insert instructions for AGENTS.md (project-level) and global agent settings In essence, it's a post-mortem for AI agents. Run it at the end of the workday to get specific rules to insert into the configuration — and the agent won’t repeat yesterday's mistakes tomorrow. --- Installation: npx skills add diskd-ai/ccbox --skill ccbox-insights --global --yes GitHub: github.com/diskd-ai/ccbox Catalog of all skills: skills.sh/diskd-ai 200+ installations, 16 skills, all MIT.

Did you know that when developing with agents, you can waste up to 30% of your time on unnecessary work? Want to significantly increase the success rate and speed of agents? Simply remove incorrect steps to ensure that actions consistently hit the target. Install the skill code-insights: npx skills add diskd-ai/ccbox --skill ccbox-insights --global --yes This tool analyzes unsuccessful project calls and suggests improvements in agent instructions, helping you save considerable time and resources. github.com/diskd-ai/ccbox

Here is a repo: github.com/diskd-ai/ccbox

The core idea of the ccbox project is to build an OpenClaw-like system for managing coding agents, specifically focused on software development workflows.

ccbox is a Rust TUI "box" for managing coding-agent sessions (Codex + Claude / comming soon): browse local session logs now, and later reconstruct exactly what the agent did (files, tools, tokens). github.com/diskd-ai/ccbox Roadmap: - Manage spawned sessions (processes) and their lifecycle. - Manage a tasks queue and assign tasks to an engine (Codex/Claude/etc). - Support multi-session workflows (handoffs, context carry, session grouping). - Manage Ralph loops. - Loops planning and task decomposition routines. - Remote control via channels: Telegram, WhatsApp, Signal, E-mail

Agent Skill for AssemblyAI that transcribes audio and video. Supports local files and automatically pulls content from YouTube using yt-dlp. Designed to integrate with agents: Codex, Claude, Pi, OpenClaw). github.com/diskd-ai/assem…

Skill to use Redmine with Codex, Claude, OpenClaw, Antigravity, etc.: github.com/diskd-ai/redmi…

New skill in my collection (skills.sh/diskd-ai) for code review: skills.sh/diskd-ai/code-…

Design-Doc: Technical Design Document Skill Create structured technical design documents that communicate system behavior, implementation approach, and acceptance criteria. skills.sh/diskd-ai

attention is not all you need

Chat, here is a PRD / TRD management LLM app that keeps requirements granular and cares about the architecture. In couple of weeks you will be able to connect it to your coding agents. So your novel codebase going to look less surprising and more familiar to you. Video is a quick tutorial for enterprise customers.

Buy The RAG Architect's Playbook: 15 Advanced Design Patterns for Building AI Systems That Actually Work on @Gumroad 128976649140.gumroad.com/l/rag-design-p…

@raydistributed Hey any best practices or workarounds for improving Python type safety in Pylance when working with Ray? Would love to see better type inference and integration for distributed workloads! Can you recommend some best practices?

The true intelligence of AI systems lies not in inference models, but in knowledge graphs - the "brain" that stores, contextualizes, and evolves information.

A Python script that creates an SSH tunnel to a Docker container running on a remote host, allowing secure port forwarding to access container services (works similar to port-forward but for docker) gist.github.com/it-baron/ef7c0…