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.
Qwen3.8-Max is now callable from Model Studio, with open weights a week out
Alibaba made the 2.4-trillion-parameter Qwen3.8-Max widely accessible to global users through Alibaba Cloud's Model Studio APIs, alongside a public beta of its QwenWork agent platform on web and desktop. The model takes up to a million tokens of context and handles multimodal input, including recreating software from screenshots and turning 2D floor plans into 3D. Weights for Qwen3.8-Max are scheduled to publish next week and the smaller Qwen3.8-27B is also going open source; no licence and no benchmark figures have been named.
An open harness scores coding agents on real schema, auth and RLS tasks
Supabase open-sourced the benchmark it runs internally, scoring Claude Code, Codex and OpenCode on scenarios like building a schema, debugging a failed Edge Function and fixing a broken row-level-security policy. Loading skills moved Sonnet 5 from 78% to 100% on the build stage and GPT-5.6 Sol from 89% to 100%, while Opus 5 and Kimi K3 reached 100% with no skills at all. Supabase also reports that its Claude Code agents check documentation in under 40% of scenarios; the suite is at github.com/supabase/evals and the published results at supabase.com/evals.
A model can reach one team while staying off for the rest of the organisation
GitHub Enterprise customers on Copilot Business or Copilot Enterprise can set each model to Enabled, Disabled or Optional across the enterprise, then grant the Optional ones to named enterprise teams by role or function. Model access was previously decided at the organisation level, so a model one team wanted had to be switched on for everybody or nobody. Administrators can create the teams and stage their assignments before turning Enterprise teams mode on; the preview opened on 31 July and reaches most enterprises by 3 August.
A slash command switches Copilot CLI's approval mode without leaving the session
Copilot CLI 1.0.78 adds a /permissions command for switching approval modes mid-session, and a sandbox setting, allowDevToolCaches, that grants build access to toolchain resources by default. Long session transcripts render progressively and session resumption reads history in parallel, which the notes call dramatically faster; ACP mode gained a closeSession request. A follow-up build on 1 August makes piped stdin prompts fire sessionEnd hooks consistently with the -p flag, and stops extension slash commands running once per loaded extension.
An agent walks a COBOL program's branches to test its Java replacement off-mainframe
The Locksmith Loop instruments a COBOL source and its generated Java target with mocks, runs both on commodity hardware, then loops an agent through witness search over input mocks to reach unexercised branches and applies parity-preserving mutations. Across three programs of 430 to 4,114 lines, two open source and one production-like, the generated Java matched the COBOL reference under deterministic parity checks in every accepted test case, reaching 91.90% branch coverage on the production-like program. When the search stalls at a routing boundary, the analyser names the blocking Locked Paragraph instead of reporting a pass.
Repeated strings in an agent's JSON logs collapse to references and back
condense-json takes a JSON document and a dictionary of replacements, then rewrites every matching substring as a reference of the form {"$r": ...}, and an uncondense call restores the original exactly. Simon Willison uses it to shrink the SQLite logs his LLM tool writes, where the same long strings recur across related records. He tagged 1.0 after eighteen months of iteration on the idea.
