A single daily edition of announcements from the AI coding tools you use, read from a curated source list, summarised, and linked back to the original. No infinite feed, no account, no advertising.
A fine-tune aimed at long agent loops lifts the cheap open-weight model more than the flagship.
On LiveBench agentic coding, Abacus.AI's fine-tune moved Smaug Flash from 46.8 to 61.1, a gain of 14.3 over its DeepSeek V4 Flash 0731 base, while the same treatment moved the top-end Smaug Agentic only from 62.2 to 64.6. All three models are open-weight and downloadable from Hugging Face, with Flash built on DeepSeek, Mini on Qwen3.8 27B and Agentic on Kimi K3, and all reachable through the company's RouteLLM API. Mini's largest reported gain was 3.5 points, in mathematics, not coding. Parameter counts and licence terms are not stated on the model page, so check the self-hosting terms before committing to one.
Your background agents now land in an inbox instead of a terminal you have to watch.
AWS has open-sourced Pizza Bot under the Apache License 2.0: a server and client you host yourself, where an agent takes a task, runs on its own, and comes back only when it has finished or needs a decision. Approvals queue in an Action list you can approve, edit or reject across several devices, and a task can be started by hand, on a schedule, or from a webhook. Everything a run touches (threads, checkpoints, memories, attachments, settings and logs) stays in one folder you control, as SQLite databases and ordinary files. It runs on LangGraph and DeepAgents and will talk to Bedrock, Anthropic, Gemini, OpenAI, OpenRouter or a local Ollama model, and AWS is explicit that this is a community project rather than a supported service.
Codex runs can now be dialled to max and ultra effort.
The 0.154.0 Python SDK adds max and ultra to the reasoning-effort values a run can ask for, raising the ceiling above the previous top setting. ExternalMessage lets outside content take part in a run or turn call in both sync and async form, and resume and fork gain include_turns for picking which turns come back, alongside turn_service_tier and source metadata. One change will break existing code: HookMetadata handlers now live under .root, and some notifications carry typed payloads instead of the old generic .params. Install it with pip install --upgrade openai-codex==0.154.0, which needs Python 3.10 or later.
The agent cruft in your commit messages never has to reach the public repo.
Simon Willison released commit-rewriter 0.1, an Apache 2.0 local web app that serves a page on 127.0.0.1:8000 where you edit commit messages and then apply them in one go. He wrote it because the first commits on a repository he was opening up were full of coding agent cruft and carried issue IDs from a private repository, which left them unfit to publish. Applying the edits first creates a backup branch of the current state, then rewrites every commit from the earliest edited one forward. Run it with uvx commit-rewriter and a path, or with no path if you are already in the repository.
Undisclosed drift in the models your agent calls now has a public, subscribable log.
TrackLLM, run by researchers at Inria, IRISA and CNRS, has been querying LLM endpoints through the OpenRouter API since June 2025 and publishing every change it can detect. It watches 839 of the 1,743 endpoints in its catalogue, covering 267 models from 53 organisations, and has logged 15.2 million queries and 191 detected changes so far. Detection runs two ways: log-probability tracking where an API exposes logprobs, and black-box border-input tracking where it does not, which account for 77 and 114 of those changes. Per-endpoint RSS feeds mean you can subscribe to just the models your own agent depends on.
