Previously, the AI assistant would automatically run a compact check right after saving a design or plan doc, effectively bypassing the review phase and auto-proceeding. The prompt definitions for brainstorming and writing plans have been refactored to explicitly wait for human approval signals (like "LGTM" or "yes proceed") before checking context limits and moving on. This puts the user back in control and guarantees your changes are actually reviewed before execution begins.

Previously, the system blocked message processing for up to 60 seconds whenever it spawned a background task, causing unresponsiveness. This update makes background executions completely non-blocking so the main message loop immediately returns to action. We've also introduced phase completion markers, an /end-session command, and state tracking to handle transitions more gracefully. Your bot will now feel significantly snappier while juggling heavy asynchronous tasks.

This update introduces a comprehensive retry tracking and escalation workflow to prevent agents from getting stuck in infinite failure loops. By enforcing a 3-retry limit and explicitly seeking confirmation before escalation, the workflow ensures persistent failures are addressed properly. Additionally, an extensive refactoring of eight core skills offloaded heavy context into reference files, resulting in an estimated 65% reduction in token usage per message.

The Commander skill documentation has been updated to include usage instructions for the send-file.sh helper script, allowing the bot to send file attachments like markdown files or documents directly to Telegram. The subagent workflow has also been revised to utilize the background-tasks skill, which keeps the bot responsive to new messages instead of blocking during long operations. This properly offloads heavy implementations and improves overall bot responsiveness.

We've revamped how Commander handles its context to drastically reduce startup token usage from ~3500 to ~400 tokens. Memory files are now lazy-loaded on demand via a new /memory command, rather than being dumped in at the start of every session. We also added an identity.md file that is read on every message to ensure core identity persists through context compactions. This makes starting fresh sessions significantly cheaper and keeps the context window focused.

Previously, the /stop command in the gateway was queued with a standard delay alongside other messages, meaning it wouldn't halt operations right away. The code now sends an Escape keypress directly to tmux to interrupt any active pending inputs, and immediately follows up with the /stop command bypassing the delay queue. This results in a much snappier and more reliable experience when trying to interrupt ongoing processes.

We've introduced a new background-tasks skill to handle long-running operations without blocking the main event flow, keeping the system responsive to incoming Telegram messages while work continues in the background. Alongside this, the brainstorming skill was updated to format questions with easily selectable lettered options and route queries properly through Telegram replies rather than UI prompts. A new /typing endpoint provides real-time typing indicators to users, resulting in a much smoother, async-friendly chat experience.

We've introduced a new set of workflow skills designed to enable structured, multi-step AI development. This update rolls out capabilities for requirement brainstorming, bite-sized plan creation, subagent-driven implementation, and rigorous verification checks prior to completion. Moving forward, the commander agent will utilize these workflow tools to drive more autonomous, robust, and well-vetted coding sessions.

To help users easily expose their local instances, the quickstart and user guides have been updated with detailed instructions for configuring tunnels. The new documentation covers using ngrok for rapid local testing as well as setting up Cloudflare Tunnels for a more robust, production-ready environment. Complete with example setups and a systemd service configuration for auto-starting, this addition significantly streamlines the deployment process.

The Dev Workspace gateway no longer relies on Redis to run, officially removing ioredis from the project. Previously, the database was used to store basic session context, but it has now been replaced entirely by file-based memory persistence. This makes the local environment much easier to launch without needing to manage a background database service.

The project README has been completely overhauled to introduce "Dev Workspace 2.0," repositioning the platform as a personal, remote-controllable development assistant. The updated documentation clearly defines the architecture and highlights use cases like maintaining persistent context across sessions, delegating tasks to AI subagents, and managing development from anywhere via Telegram. This update provides a much clearer vision of the project's goals and improves the onboarding experience for new users.

Pichu is a new multi-agent Telegram orchestration system powered by Claude Code. It features a persistent orchestrator ("Pichu") that runs in a tmux session, receiving Telegram messages via an Express gateway server. The orchestrator delegates implementation tasks to fresh, isolated sub-agents and includes support for file-based memory persistence, image and file analysis, file attachments, and reply threading. This provides a really cool way to interact with an AI coding orchestrator directly from Telegram!