Five projects on my GitHub: Yomu, Hikari, GradeStudio, SeraGlow (a work sample, source private) and Prompts. Sound on, 38 seconds. #buildinpublic#opensource
Yomu Daily: ffmpeg-skill gives coding agents 42 FFmpeg tools for Claude Code, Cursor and Codex. MIT, no cloud, no API keys. Catch: the full test corpus has not been run on Windows yet. #opensource#ClaudeCode
@yusuktan Skipping the per-file server revalidation makes sense, that round trip was pure overhead. Curious if the 70% case is mostly import-heavy files, since that's where the cache hit rate should matter most.
Our biggest pain running Claude agents this week: the loops live in one open session and silently stop after every app restart. We moved them to scheduled tasks stored on disk, but the app still has to be open. How do you keep an agent loop alive across reboots? #AI#Agents
@mfranz_on Opus planning while a 27B Qwen does the edits is a neat split for cost. Does Opencode get the full repo context, or only the task brief Claude Code writes for it?
@chevlume Thanks, exit 0 on a "no" row is a good heads-up. Reading the JSON in a follow-up step and failing on any row marked no seems like the safe pattern for CI.
@joelniklaus@RicursiveAI@huggingface Serving Ripple at 262k to fit 6 parallel trials turns the leaderboard into a context test too. 169 of 333 mini-swe-agent attempts ending on a full window is a clear result. Did the compacting harnesses lose anything on tasks that fit in the window anyway?
Five projects on my GitHub: Yomu, Hikari, GradeStudio, SeraGlow (a work sample, source private) and Prompts. Sound on, 38 seconds. #buildinpublic#opensource
@_lewtun Merging ~8k PRs from 1,000 agents into one repo is the part I'd want to hear more about. Was there a shared eval that gated every merge, or did the message board coordination keep regressions out?
@chevlume Good to know about the exit code. A CI step can pipe the JSON through jq and fail on any row that says no, so the gate is explicit. Does a skipped host show up as its own status in the JSON, or only as a missing row?
@rustaceans_rs Derive panics and cfg target fixes matter most in big workspaces, where one proc macro crash can take down completions for the whole crate. Do you see these fixes land mostly from user-reported issues or from the maintainers' own testing?
@iammukeshm A single process removes the network hop and serialization between components, which is likely where a 90% cost cut comes from when they talk constantly. Which of your three split cases would a small team hit first?
@Sumanth_077 Retrieving past cases instead of changing weights keeps the learning inspectable, since you can read exactly which trajectory shaped a plan. How does the Case Bank handle a stored case that was a failure, does it get retrieved as a warning or filtered out?
@chevlume Thanks. With the cwd logged you could also tell a hook that fails from the repo root apart from one that only breaks in a subfolder. Will the report be usable in CI as well, or only locally?
claude-quickstarts by Anthropic (MIT license, over 20k stars) has a two-agent coding example: an initializer and a coder that build an app across sessions and track progress in git. Browser and computer-use demos too. Which one would you try first? #AI#Agents
@testingcatalog Separate models per role makes sense: a cheaper one for a coordinator that mostly relays status, a stronger one for threads doing the edits. Does the coordinator's model also affect how it splits work across threads?
Researchers at the ARK Institute and Stanford University have used AI to design a completely novel virus from scratch that successfully kills E. coli bacteria. Published in Science, this breakthrough opens a new era of AI-driven genomics. 🧬🦠 #AI#Science