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.
Codex keeps its notes when the context window fills, instead of compacting
OpenAI's GPT-6 Astra began rolling out on September 3 to a limited set of organisations, with ChatGPT Plus, Pro, Business and Enterprise tiers, the API and AWS access following within days. In Codex, an experimental mode lets the model write notes that persist across context windows and leaves earlier windows searchable, so requirements and test results from early messages survive a long session rather than being compressed into a summary; it can be switched on now and becomes the default soon. OpenAI calls Astra its best model for software engineering to date, reporting 99.9% on ARC-AGI-3, 98% on FrontierMath Tier 4 and 100% on ExploitBench, and says it is nearly twice as fast at computer use. The staged rollout reflects the model meeting the Critical cyber-capability threshold under OpenAI's Preparedness Framework.
Your agent's default library picks depend on which agent you asked
A study of 16,893 coding sessions published September 3 compared what Claude Code, Codex and Cursor reach for when asked to build the same thing. The three chose identical tools in only 42% of cases: asked for a voice agent, Claude Code picked Twilio, Codex the OpenAI Realtime API and Cursor Vapi. Their research habits diverge too — Codex searched the web in about 94% of sessions and Cursor in 67%, while Claude Code leaned on training data and searched only about 30% of the time. Popularity did not predict selection: LangChain drew 194 mentions and four picks, PayPal 139 mentions and none, and language shifted outcomes sharply, with Resend taking 55 of 89 TypeScript runs and Sendgrid 22 of 24 Python ones.
Handing an LSP to a coding agent can cost tokens and add no accuracy
A harness study published August 12 measured why coding agents fall back to grep instead of semantic tools. On a clean TypeScript codebase an LSP added zero F1 improvement and consumed 16% more tokens; on a noisier one it improved F1 by 0.246 while using 12% fewer tokens. Much of the gap is presentation: returning bare file locations left agents making 15.2 follow-up reads per episode, and adding inline source context cut that to 3.2 — below grep's 4.3 — lifting pass rates from 67% to 83%. The agents were routing by task rather than avoiding precision, picking semantic tools 0-6% of the time for simple location lookups but 45-57% for reference-finding, and forcing a semantic-first path dropped success from 100% to 89%.
Past agent sessions become a local memory 22 tools can query
deja-vu reads the session logs that 22 coding agents already wrote to disk — including months of history from before it was installed — and hands them back through an MCP tool and session hooks, rather than asking anyone to record memories as they go. It uses no LLM and no embeddings, shipping as a single local Go binary that parses JSONL and SQLite stores, redacts credentials and builds an inverted index in a local cache, with an incremental manifest so unchanged sessions are not reprocessed. Recall fires automatically at session start, on each prompt, before a file edit or command runs, and after one fails, so a repeated error surfaces whatever resolved it last time. Claude Code, Codex, Cursor, Copilot CLI, Gemini CLI, Zed and Goose are among the agents supported; the project reports 85.3% hit@1 on LongMemEval-S, a median lookup near 0.4 ms and a 160 MB index over 5 GB of history, under an MIT licence.
Claude Code turned 72,758 lines of 68000 assembly into a playable Godot port
Rabah Shihab used Claude Fable 5 inside Claude Code to port Babylonian Twins, his 1993 Amiga game, across three steps in which he played the nightly builds and made the decisions but did not write the code. His 2010 iOS engine, 34,000 lines of C++, reached a playable character in Godot 4 in 21 minutes and covered all 38 entity types, menus, physics and multiplatform export. The 1993 original, uncommented and undocumented across 26 files, was rebuilt in 1 hour 45 minutes, with level maps, object tables, sprite sheets and copper effects recovered from the source; the model produced byte-identical binaries, built disk images and verified them in the FS-UAE emulator. It added its own debugging flags for automated testing and surfaced real bugs, including a guard collision check running through floors.
