Entries matching “pull-requests”

We've been curious how far local models have actually come for agentic coding tasks, so we ran an experiment. Setup: • Model: Muse Glimmer (30B), packaged as a single llamafile • Agent: Hermes coding agent (connected via llamafile's local server mode, zero API keys needed) • Target: Mozilla AI's Otari gateway The Issue: We pointed Hermes at a real, reported bug in Otari (#183) where the gateway returned a vague 502 error on image requests instead of passing through the actual provider error. What the Agent Did: Hermes read the issue, navigated the repo, isolated the bug, created a branch, ran existing tests, wrote a new regression test, and opened a draft PR (#727). All of it ran locally and offline, with zero code written by hand. It's still draft PR territory rather than a merged fix, but it's a solid signal that ~30B local models are getting genuinely capable for real dev workflows, not just toy demos. Video walkthrough of the run: https://youtu.be/5GAgbT-XgHU?si=vJqEDGm9hssCO5-M Happy to answer questions about the setup, model performance, or how Hermes handled tool calling!

DHH

@dhh

Got muse cli hooked up for both default agent action and lazy-loading. Come give it a try, @finkd 🙏 github.com/omacom/omarchy…

@Xmarlandoj

@Xmarlandoj

Muse is your personal assistant. Zouroboros Workshop is the workshop behind it. Workshop is an MCP service that runs on a private, single-tenant VPS provisioned for each customer. From a normal Muse conversation, a user can dispatch real software work to a coordinated team of seven AI coding executors — Claude Code, Codex, Gemini, OpenCode, Kimi, Hermes, and Pi — routed by health and capability, with automatic fallback when one executor is unavailable. Three things make it different from a single-vendor hosted builder. First, it runs on the subscriptions the user already pays for: there are no Workshop model credits and no second meter. Second, its memory is shared across every executor and survives across sessions, so a decision recorded by one tool is retrieved by another in a later Muse conversation — memory that compounds instead of resetting. Third, work is verified before it is delivered: builds run through a factory lane with deterministic review, and swarm results return with per-task evidence, not just an agent's claim that it is done. The customer-facing promise is simple: turn your existing AI tools into a workshop that delivers tested, reviewed work — a finished pull request, not a wall of code in chat.

Cristian Uibar

@cristianuibar

@theo please add support for Muse Code in T3. I've opened a PR here for it using the official SDK. github.com/pingdotgg/t3co…