GitHub Trending: Impeccable Gives Coding Agents a Design Linter, and antirez's DwarfStar Runs DeepSeek V4 Locally

GitHub's daily trending page on October 5, 2026 again leads with agent tooling. Several entries -- ponytail, Agent-Reach, claude-mem -- have been covered here before. Two new ones are worth a look because they address problems developers hit after an agent is already in use: output that all looks the same, and the cost of running models remotely.

pbakaus/impeccable -- a design linter for agents

Impeccable (JavaScript, Apache 2.0, about 76k stars, +1,171 on the day) calls itself "the design language that makes your AI harness better at design." Its README names the problem directly: models "trained on the same SaaS templates" produce "the same handful of tells on every project: Inter for everything, purple-to-blue gradients, cards nested in cards."

• Workflow: one /impeccable command with subcommands such as init, shape for UX planning, audit, critique, polish and live for iterating in a browser. A PRODUCT.md file records audience, purpose, constraints and voice so the agent stays consistent across screens.
• Systems: it installs into 17-plus harnesses, including Claude Code, Cursor, GitHub Copilot, Codex, Gemini CLI and OpenCode, via npx impeccable install, which detects harness folders and adds hooks where supported.
• The interesting design choice: 61 deterministic detector rules -- overused fonts, gray text on colored backgrounds, bounce easing, nested cards -- that run without an API key or model calls, and can report before or after an agent's edit.

antirez/ds4 -- local inference, narrowly

DwarfStar (antirez/ds4, C, MIT, about 23k stars, +211) is a native inference engine from the creator of Redis. Instead of being a general model runner, it builds and tests a small set of models end to end -- loading, prompt rendering, tool calls, KV cache, an HTTP server and coding-agent use. Supported models include DeepSeek V4 Flash, V4.1 Flash and V4 PRO, GLM 5.2 and 5.3, and Qwen 3.8 Flash Next, in Q2 and Q4 quantizations. Backends are Metal (Apple Silicon with 96 GB or more), CUDA (aimed at DGX Spark and Ada-generation cards) and ROCm (Strix Halo). The README says Q2 DeepSeek Flash fits on 128 GB machines, reports about 126 tokens per second aggregate across 16 sessions on eight L40S GPUs, credits llama.cpp and GGML without linking to them, says it was developed with heavy AI coding-agent assistance under human direction, and labels itself beta.

Analysis: Impeccable's deterministic rules are the more transferable idea. A model asked to "make it look good" grades its own work generously; a rule that flags nested cards does not. Teams adopting coding agents can apply the same pattern to any house standard -- accessibility, copy tone, API naming -- as cheap checks that run in the harness's hooks. DwarfStar's adoption constraint is hardware: 96-128 GB of unified or GPU memory is still a workstation, not a laptop. But a maintained engine tuned to a handful of strong open models makes running a capable coding agent fully on-premises, with no code leaving the building, a realistic option for teams that have that hardware.