We resolved an issue where pressing Enter during skill loading would inadvertently insert a newline instead of executing the command. This was fixed by making the injection delay configurable and increasing the default from 300ms to 5 seconds to provide adequate time for skills to initialize. Both asynchronous and synchronous tmux execution methods were updated to support this new delay parameter, ensuring commands are safely fired off once the CLI is completely ready.

A bug was causing cross-bot pollution within the gateway, where bots could receive mixed inbox streams. This update resolves the issue by bumping the modules/bots submodule to a version that enforces bot-specific streams, properly isolating their messages. We also added a gateway.local.yaml configuration to ease the setup process for local gateway development. This ensures cleaner data separation between bots and a smoother testing experience for developers.

This massive update introduces a unified orchestrator to manage specialized multi-agent teams across backend, frontend, QA, and git operations. To boost collaboration, agents can now communicate directly via a Telegram gateway that supports N bots to N sessions with reply threading and Redis-backed message queues. In an exciting architectural twist, the system also embeds a new AI self-evolution protocol, where successful agent workflows are solidified as "genes" so future sessions can seamlessly learn and adapt from previous tasks. It's a huge step forward for building deeply autonomous, context-aware development pipelines!

To enforce a cleaner separation of concerns in the multi-agent pipeline, Git operations have been entirely stripped from the backend, frontend, and QA agents. The task-complete skill now focuses purely on updating task progress to trigger orchestrator handoffs without performing any source control actions. All commits, pushes, and pull requests are now managed exclusively by the review-git agent following successful QA approval. This centralized approach guarantees more reliable, context-aware commits and a smoother flow across automated development workflows. 🧹

A new set of bash scripts simplifies gateway orchestration across local, tmux, and systemd environments. The new start-gateway.sh handles pre-flight checks (Redis, Cloudflared tunnel status, bot tokens) and webhook registration before booting, while stop-gateway.sh ensures graceful shutdowns for tmux or orphan processes. Additionally, install-service.sh creates a smoother path to production by registering the gateway as a systemd service. These tooling additions take the guesswork out of deployments and bring much-needed stability to the bot lifecycle.

This major structural overhaul establishes specialized handler skills—like backend, frontend, QA, and git-decision—to power an end-to-end multi-agent orchestration pipeline. Tasks are now seamlessly routed from agent code implementation through automated independent QA, culminating in a git-decision step that explicitly asks for user approval before staging and committing. With new agent-identity-aware Telegram notifications and improved task routing in the core orchestrator, autonomous contributions are now tightly structured and rigorously verified.

The CLI now supports a --skills flag to dynamically inject a comma-separated list of skills when spawning an agent. In addition, the orchestrator (or pichu) now automatically receives the commander skill by default. This makes it much easier to test specific agent capabilities and ensures orchestration agents are properly equipped without extra boilerplate.

The gateway configuration has been updated to pass the bot name in wake commands, ensuring targeted agent activation. We've also drafted an implementation plan for a dedicated /ack endpoint to allow eager acknowledgment of polled Redis stream messages. This will prevent duplicate message re-processing in cases where an agent crashes between polling and replying.

This update overhauls how agents coordinate, introducing a structured dev-docs → dev-git → task-complete handoff loop that the orchestrator actively monitors. Skills are now decoupled into auto-loading layers—Persona, Role, and Utility—ensuring specialized agents like backend-developer or qa-developer spin up with exactly the right set of domain knowledge and lazy-loaded tools. As a bonus, we fixed an annoying bug where ! in command injections was accidentally triggering Bash history expansion in Tmux sessions. Handoffs in multi-bot pipelines will be significantly more reliable.

Developers can now directly specify an agent's specialized skill set and persona when launching them via the cc-orch CLI. By passing the --role (e.g., backend-developer) and --persona flags, you gain granular control over agent instantiation right from the terminal. This makes testing and deploying custom agents much faster and more flexible.

The orchestrator now uses agent configurations to spawn core agents dynamically rather than relying on hardcoded defaults. By fetching details like the role_skill, persona, and memory paths for each agent, the spawn process respects the specific configuration of each bot. This makes agent initialization much more flexible and centralized overall.

The spawnAgent function now supports an optional role parameter that loads a role-based skill directly. When a role skill is provided, it handles lazy-loading any referenced sub-skills during invocation rather than loading them all upfront. This provides a cleaner way to initialize specialized agents while deferring the loading of nested dependencies until they are actually needed.