Released Nov 2 with about 48B active parameters and a 1M-token context, it's trained on long agent runs and real agent failures, and Meituan reports 88.4 on Terminal Bench v2.1 and 38.5 on v4.0 (its own numbers). It's API and online chat only for now at a limited-time $0.30/$1.20 per million tokens (pricing), with GitHub and Hugging Face weights promised later.
Shown on HN Oct 28 under Apache 2.0, the single Go binary runs a local ergo IRC server where long running sessions (say a planner and a builder) tell each other when work lands and ask you when they need a call, while you read every line in any IRC client and only your registered nick carries authority. A breaker stops waking the sessions after a burst of posts or a long stretch without you, and it's only been used live on macOS; Claude Code launches need dangerously load development channels server:parley because its channel isn't on Anthropic's allowlist…
OpenAI's updated help article says Codex, Work and other agentic ChatGPT use will draw from one combined allowance measured in tokens, so heavier models and longer tasks burn more of it, while Instant chat stays free of the meter. There's no exact December date or per plan amount yet; combined remaining usage and reset times will show in Settings Usage, with credits on sale when you run out…
Released Oct 29, /permissions policies approve for me in an interactive V3 session has a Claude Sonnet 4.6 classifier approve actions it judges safe, block unsafe ones and ask when unsure, but only for actions your existing deny, ask and allow rules don't already settle. It's saved per workspace and doesn't work in cloud sessions, and approval prompts can now save rules at a subcommand tier like git stash (docs)…
Shown on HN Nov 1, Winch tracks your keywords in 175 countries, drafts and scores listing rewrites, localizes them, sorts reviews into themes and draws store art, using your own Claude Code, Codex, OpenCode, Ollama or OpenRouter setup instead of selling AI credits, with all 84 tools also exposed over MCP. It's macOS 13+ only, keeps data on your Mac, asks before anything is written to a store, and costs $69 once for the first 100 licenses (then $89) after a one day free trial…
Shown on HN Nov 1 under MIT, npx amidumb reads /.claude/projects and /.codex/sessions and computes a fixed, local score across specification, steering, verification, discipline and leverage, then has your own claude or codex write a roast and three habits to paste into CLAUDE.md or AGENTS.md . The roast sends about 150 redacted, trimmed prompts to your CLI ( no roast sends nothing), and the author says the rubric is better at the extremes than in the middle…
Spotted something we missed? Start a thread.