Chorus v0.12.0: Wake a Remote Claude Code, Hit Exactly the Right One
One daemon watches one directory. But your code lives in three repos across two machines. When a task lands, you don't get to say where it runs.
Deep dives into plugin architecture, multi-agent orchestration, and the infrastructure that makes AI-human collaboration work.
A few ideas, each with proposals, documents, and tasks hanging off it, all tangled in dependencies. Can you still say what the project looks like right now?
One daemon watches one directory. But your code lives in three repos across two machines. When a task lands, you don't get to say where it runs.
You give an agent permissions, hand it a task, and it sits in a database row waiting for you to open a terminal and wake it. That's your always-on teammate?
You work on one idea and it spawns three more. The line between them only ever lived in your head.
OpenClaw 2026.4.27 introduced a Plugin SDK. Codex reworked how hooks load. Both Chorus client plugins were rewritten to keep up — and the main repo got a month's worth of rough edges sanded down along the way.
Pinned down on technical decisions before the idea is even shaped, then disbanded the moment it ships. This release patches both ends.
Writing specs as local files is great. So how does the collaboration platform keep up? This release wires OpenSpec's local file flow into Chorus's AI-DLC.
Built one plugin for Claude Code, one for Codex CLI. Same logic. Wildly different experience.
Creating an agent used to be a three-way choice: PM, Developer, or Admin. Now you pick by resource and action.
The Chorus plugin was Claude-Code-only. Now Codex CLI gets it too, one command away.
No Docker, no database, no config files. Just npx and a browser. Plus: export your docs.
Setting up PostgreSQL just to try an agent platform? PGlite embedded mode runs Chorus in a single container. Also: the port changed — read this before upgrading.
A 54-page survey uses cognitive science to answer a question the industry keeps rediscovering: why do agents get more reliable when you move things out of the model?
Still reviewing every Claude Code plan line by line? /yolo lets agents review each other so you don't have to.
v0.6.0 ships independent Review Agents, real-time Presence, and IdeaTracker. When 5 agents work in parallel, humans finally stop flying blind.
Anthropic quietly nerfed Claude Code's thinking depth and cut off OpenClaw's subscription access. The real problem isn't what changed — it's that your workflow had zero resilience to it.
Same model, different harness, 17 problems apart on SWE-bench. Why harness engineering matters more than model choice in 2026.
SSE + MCP dual-channel architecture to make AI agents truly live in the workflow — event push for awareness, tool protocol for execution.
Design patterns from the Chorus experience — plugin ecosystem, hooks, skills, and sub-agent context injection for multi-agent collaboration.