Coding
Nov 1, 2026 · 04:18 UTC
The point
Simon Willison released datasette-llm 0.1a9 with a small user experience improvement to the plugin. It now picks up the default model used by LLM. That means Datasette Agent can start without explicitly configuring a default model first.
Why it matters
For builders, one less setting stands between installing the plugin and trying Datasette Agent on a database.
Key details
- New start command shown: uvx --with datasette-agent datasette content.db --internal internal.db
- Before this release, users also needed to specify a model with -s plugins.datasette-agent.default_model gpt-6-astra
Nov 1, 2026 · 02:18 UTC
The point
DecompGym benchmarks coding agents on designing decomposition algorithms. Agents write solvers across 81 instances from nine problem families, and each solver must run in 60 seconds, pass a method audit, and land within 1 percent of the optimal objective.
Why it matters
For builders, output checks alone miss bad methods, so agent feedback needs to name the exact defect.
Key details
- Evaluated 12 GPT, Claude, and Gemini models in coding agent tools
- Reflection turned 66 percent of retries into successes, versus 39 percent for regeneration
Oct 31, 2026 · 12:19 UTC
The point
Cursor says a tampered Claude Code extension 2.1.313 tried to install a persistent SSH backdoor and impacted around 5,000 developer systems. Anthropic says a malicious actor used Cursor marketplace access to publish it.
Why it matters
Builders who used that extension in Cursor should hunt for the backdoor, not just update.
Key details
- Code collected hostname, username, OS, architecture, home directory, Git email, and public IP
- It tried to add an attacker SSH key and open a reverse tunnel
- Version 2.1.314 replaced 2.1.313, but updating does not remove the backdoor from infected systems
Oct 31, 2026 · 06:18 UTC
The point
Anthropic released Claude Code v2.1.316. It adds CLAUDE_CODE_SUBAGENT_EFFORT to set subagent effort, pooled group budgets on the Claude apps gateway, and a switch to turn the ultrathink keyword off.
Why it matters
Builders can tune subagent cost and share one monthly budget across a group.
Key details
- Pooled budgets are set with pooled_budget on the spend limits admin API
- Fast mode for Opus 4.8 will be removed on November 30, 2026
- Fixed shell commands left running on macOS and Linux when Claude Code exits
Oct 31, 2026 · 04:18 UTC
The point
Researchers present ExLabel, a self-evolving LLM assistant that judges task clarity and test coverage in SWE-bench style datasets. It uses an evolving experience bank to guide assessments of new instances.
Why it matters
Builders can filter training and evaluation data before trusting benchmark scores.
Key details
- New SWE-TQ benchmark spans eight languages, cleaned from human annotations in SWE-bench and Multi-SWE-bench
- Tested with three backbone LLMs, ExLabel achieved the best F1 across all settings
- A model fine-tuned on data filtered by ExLabel resolved 66.2% of SWE-bench Verified issues
Oct 31, 2026 · 02:19 UTC
The point
Researchers present DualGauge, the first fully automated framework for jointly evaluating correctness and security of specification-only code generation. DualGauge-Bench has 307 coding tasks, each with functional and security tests from the same specification.
Why it matters
Builders cannot trust functional tests alone to show code is secure and correct.
Key details
- Evaluated 10 representative LLMs across Python, C++, and JavaScript
- Even the strongest model remained below 15% joint security-functionality success in every language
- Codex, OpenHands, and Claude Code were tested, and two failed to outperform a comparable direct model
Oct 31, 2026 · 00:19 UTC
The point
A user whose mobile app reports crashes through Sentry has Meta Muse watch for new errors. When one surfaces, Muse goes to the user's GitHub repo, works out what went wrong, and drafts a fix before the user knows there is a problem.
Why it matters
Triage starts early, and the builder reviews a draft instead of starting cold. Evidence: user-reported, from the Good Bot case list at goodbot.guide, not independently verified.
Key details
- App crash reporting: Sentry
- Code check: the user's GitHub repo
- Output: a drafted fix, not a merged change
Oct 30, 2026 · 22:19 UTC
The point
Phorion found a persistent SSH backdoor in the official Claude Code for VS Code extension 2.1.313 on the Cursor Marketplace. Cursor removed that release. The malware adds an attacker key, opens a reverse tunnel, and installs persistence across macOS, Linux, and Windows.
Why it matters
Builders on 2.1.313 need incident response and full cleanup, not just an update.
Key details
- Planted key comment is svc-telemetry in ~/.ssh/authorized_keys
- Persistence includes macOS LaunchAgent com.apple.sysmond.plist and Linux @reboot cron
- Phorion says uninstalling alone will not remove the backdoor
Oct 30, 2026 · 20:18 UTC
The point
Security researcher Kevin Beaumont reports that Claude Code extension version 2.1.313 contained malware creating a reverse SSH backdoor to desktops. He says that version was removed, 2.1.314 is the fixed version, and auto updates widened impact.
Why it matters
Builders should check installed versions and remediate impacted VS Code and Cursor systems manually.
Key details
- C2 collected hostname, username, OS, architecture, home directory, environment variable names, and Git email
- Reverse SSH let the attacker choose victims for hands on keyboard access
- Beaumont says updating alone does not remove prior backdoors or recon activity
Oct 30, 2026 · 18:19 UTC
The point
OpenAI launched an API credits benefit for eligible ChatGPT subscriptions. If you have a $100, $200, or $500 subscription plan, you now get that same quantity of credits to apply to one of your API organizations.
Why it matters
For builders, subscription spend now also covers API experiments and demos.
Key details
- Anthropic announced a similar benefit earlier this month for Max and Team subscriptions
- Anthropic subscribers at $100, $200, and $500 per month get the equivalent quantity of expiring Anthropic API credits
- OpenAI credits apply to one of your API organizations
Oct 30, 2026 · 16:19 UTC
The point
In a commentary post, Simon Willison highlights Jacob Sansbury describing Spawn, a multiplayer game where players prompt an agent named Savi to modify the world. A swarm of Fable powered agents now develops Spawn autonomously and ships hourly.
Why it matters
Builders see what tight agent loops require: fast shipping, fast rollback, and production feedback.
Key details
- Token spend is over $3 million per month, described as 30x payroll
- About 60% of work starts from production errors, bug reports, logs, or feature requests
- Broken changes can be rolled back in 30 seconds
Oct 30, 2026 · 12:40 UTC
The point
GitHub announced Copilot in VS Code Universe releases for more choice in models, where agents run, and how you stay involved. Cloud sessions now continue across VS Code, the Copilot app, GitHub Mobile, github.com, and Copilot CLI, and keep running after you close your laptop.
Why it matters
Builders can start work in one client and steer the same session from another.
Key details
- Agents window supports parallel model runs with side by side review, in preview
- Computer use lets Copilot interact with desktop apps, with approval per app
Oct 30, 2026 · 06:19 UTC
The point
Researchers present an agentic workflow that ports Fortran kernels to portable Kokkos C++. Specialized LLM agents translate, validate, compile, test, debug and optimize without human intervention.
Why it matters
It offers HPC teams a way to modernize legacy code for GPUs without manual porting.
Key details
- Evaluated over five runs per configuration on an NVIDIA A100, including a production climate model kernel
- GPT-5 and o4-mini-high completed it most reliably, beating Fortran baselines for a few US dollars
- Open weights models handled simpler kernels but struggled with parallel reductions and hierarchical parallelism
Oct 30, 2026 · 02:20 UTC
The point
Anthropic released Claude Code v2.1.314. A new permissions.blockWritesOutsideWorkingDirectories setting makes Write, Edit and NotebookEdit refuse paths outside the working directories in every permission mode.
Why it matters
Builders get a hard stop on writes that stray outside their working folders.
Key details
- Idle compaction in VS Code and Claude Desktop summarizes a 200K+ token conversation left idle for your next message
- Gateway upstreams can take a groups list, so only developers in listed IdP groups reach that upstream
- It now lowers a capped effort level and retries instead of failing turn after turn
Oct 30, 2026 · 00:23 UTC
The point
OpenAI released Codex 0.166.0. The terminal app now supports keyboard driven transcript selection with word, line, and page motions, integrated search, and configurable shortcuts. A new codex queue --resume command resumes unloaded threads after queueing a message.
Why it matters
Builders in the Codex terminal get faster transcript review and fewer lost queued threads.
Key details
- GPT-6.1 Terra was added to the bundled, Vertex AI, and Amazon Bedrock model catalogs
- codex doctor --network-only runs focused configuration, connectivity, and WebSocket diagnostics
- Fixes preserve acknowledged approval and permission settings when resuming or forking threads
Oct 29, 2026 · 01:00 UTC
The point
GitHub Copilot Fusion is generally available. Instead of picking a model for each task, you select Copilot Fusion and an effort level, and it routes and orchestrates models for you using three patterns: Single (one model), Cascade (a cheap model drafts, a check escalates if needed), and Critique (one model drafts, another reviews).
Why it matters
Model-picking per task goes away. Cost control moves to the effort level you choose and the model policies your admins set.
Key details
- Fusion Max scored 60 on the Artificial Analysis Coding Agent Index, within two points of Claude Code with Opus 5.5 Max in GitHub's comparison, at about half the average API cost per task
- Three effort levels: Standard for everyday coding, High for harder tasks, Max for the most demanding reasoning
- Available on all Copilot plans in VS Code, the Copilot app, Copilot CLI, and JetBrains IDEs; you are billed per model used, with a 30 percent discount on model costs for paid plans for a limited time