Opus 5 on AI Gateway, Ruff v0.16 ships 413 rules — Dev Signal
Opus 5 on AI Gateway, Ruff v0.16 ships 413 rules
Share:
Tool of the Week
Claude Opus 5 available on AI Gateway now
Opus 5 handles multi-file refactors and end-to-end feature work natively; set reasoning effort via `reasoning: 'medium'` in AI SDK 7, handle safety blocks with model fallbacks.
Long-horizon agentic tasks complete full features instead of stubs, cutting token spend and latency at medium effort levels. Safety classifier blocks require fallback chains—AI Gateway fallbacks let you skip manual retry logic.
Replaces Opus 4 and 4.8 for coding agents; requires AI SDK 7+ and Anthropic provider setup. Reasoning enabled by default (configure via top-level option). Worth trying now if you're already routing through AI Gateway; test `speed: 'fast'` mode first for latency-sensitive paths. Plan for occasional safety blocks on security audit code—fallback to Claude Sonnet 5 or 4.8 as tested strategy.
“handling multi-file features, larger refactors, and end-to-end feature work, and completing full tasks rather than leaving stubs or placeholders”
“Opus 5 is effective at low and medium effort, which produce quality at a fraction of the tokens and latency of higher settings”
“Reasoning is on by default”
“Opus 5 ships with elevated cybersecurity safeguards, so benign security work can occasionally trigger a safety classifier and get blocked”
“use AI Gateway's model fallbacks. List backup models in a `models` array under `providerOptions.gateway`”
Get issues like this in your inbox — free, every weekday.
Quick Signals
Ruff v0.16 enables 413 rules by default
Ruff v0.16 expands default rule set from 59 to 413 rules, replacing need for separate linters like Black, Flake8, and isort while running 10-100x faster.
Developers get stricter linting out-of-the-box without configuration, catching syntax errors and runtime issues automatically. Markdown code block formatting and improved suppression comments reduce boilerplate in mixed-content projects.
Replaces Black, Flake8 (plus plugins), isort, pydocstyle, pyupgrade—install from PyPI and update config for breaking changes (revert with `select = ["E4", "E7", "E9", "F"]` if needed). Ready now; most users update without code changes. Markdown formatting stabilized; ruff: ignore comments add expressiveness over noqa.
“Ruff now enables 413 rules by default, up from 59 in previous versions”
“all while executing tens or hundreds of times faster than any individual tool”
“can be used to replace Black, Flake8 (plus dozens of plugins), isort, pydocstyle, pyupgrade, and more”
“Ruff can now format Python code blocks embedded in Markdown files”
“Ruff v0.16 builds on this to add two additional ruff suppression comments”
ruffpython-lintingtoolingreleaseformatter
Ruff v0.16.0 enables 413 rules by default
Ruff jumped from 59 to 413 default rules, catching syntax errors and runtime issues previously opt-in—unpinned dependencies will break CI until you fix violations or pin the version.
Blanket rule enablement forces immediate remediation across projects; comprehensive test coverage makes upgrades safe but requires active fix or version pinning to prevent CI failures. Caught hundreds of minor issues in production codebases.
Data Point
Benchmark measures LLM data construction and quality
DataPrep-Bench unifies evaluation of data preparation pipelines—construction (raw→training data) and quality scoring (predicts downstream utility)—with two released tools: Data-Construction-Skill agent (+20pts on Finance) and Distributional Alignment Score (DAS) metric correlating r>0.70 across Math/Science/Medical domains.
Developers building data pipelines now have a shared downstream-grounded protocol to measure whether their data prep actually improves model performance, replacing ad-hoc quality heuristics with benchmarked construction methods and correlation-validated scoring functions.
Replaces custom data-prep evaluation with DataPrep-Bench's six-domain test suite; Data-Construction-Skill agent and DAS metric are released. Requires fine-tuning base models on prepared data to score construction outputs. Ready now as a benchmark framework, though integration into existing pipelines requires wiring Pearson correlation validation into candidate dataset selection.
“quality refers to downstream training utility rather than surface-level textual properties”
“Data-Construction-Skill, a skill-guided agent that lifts the Dolly-only baseline by nearly 20 points absolute on Llama-3.1-8B Finance”
“DAS attains the strongest cross-model correlation in four of six domains and is the only metric clearing r > 0.70 simultaneously in Math, Science, and Medical”
Replaces implicit linting configuration discipline with stricter defaults. Requires either pinning ruff version, running `ruff check . --fix --unsafe-fixes` to auto-remediate, or configuring ignored rules in pyproject.toml. Worth upgrading now if your project has strong test coverage; expect 100–1600+ violations depending on codebase size.
“Ruff now enables 413 rules by default, up from 59 in previous versions”
“the number of rules in Ruff has grown from 708 to 968”
“uvx ruff@latest check . --fix --unsafe-fixes”
“Found 1618 errors (1538 fixed, 80 remaining)”
python-toolinglintingci-cdbreaking-changeruff
Claude models eliminate rigid prompt guardrails
Newer Claude models handle context well enough to replace exhaustive rules with tool-design principles, progressive disclosure, and judgment-based directives.
You can reduce system prompt bloat and context waste by shifting from prescriptive rules to interface design; Claude now auto-saves relevant memories and handles dynamic tool references without preloading everything upfront.
Replaces: over-specified system prompts with hardcoded constraints. Requires: restructuring CLAUDE.md around gotchas only, breaking skills into progressive-disclosure trees, using code/HTML as specs instead of markdown descriptions. Worth trying now if you're on Claude 3.5+; use `claude doctor` command to audit existing prompts.
“newer models have better judgement and can handle these decisions well without explicit rules”
“giving examples actually constrains them to a certain exploration space”
“Claude now automatically saves memories that are relevant to the work and to you”
“Claude can reference HTML artifacts created by our new artifacts feature”
“keeping one item in_progress helps define our requested behavior”
Grabette: handheld gripper records robot manipulation data
Open-source handheld gripper (~€490 BOM) + browser-based SLAM processing pipeline converts human demonstrations into LeRobot datasets without needing a robot.
Removes hardware gatekeeping from robot learning data collection—developers can now record training data from any environment and contribute to shared datasets, addressing the actual bottleneck in visuomotor policy training.
Replaces expensive teleoperation rigs and lab-bound data collection. Requires building Grabette from off-shelf components (Raspberry Pi, OAK-D camera, magnetic encoders), uploading recordings to Hugging Face, and running SLAM processing via browser. Production-ready for early adopters; community contribution model is the real test.
“The bottleneck isn't the model. It's the data.”
“Grabette is a handheld gripper instrumented with everything needed to reconstruct a manipulation demonstration.”
“BOM cost ~490€”
“Everything is open source”
“The output is a standard LeRobot dataset on the Hugging Face Hub, so the same data can drive different robots and different learning methods.”
Gemini CLI shuts down June 18, 2026; multi-agent workloads move to Antigravity CLI (Go-based, async agent orchestration) with unified backend shared across desktop and terminal.
Terminal-first developers lose weekly release cadence and must migrate Agent Skills, Hooks, Subagents to new platform. The unified harness means desktop and CLI agents sync automatically—critical if you're splitting work across multiple agents—but feature parity gaps exist at launch.
Antigravity CLI replaces Gemini CLI entirely; requires migration of Skills and plugins before June 18 deadline. Go implementation promises faster execution and async workflows (no terminal lock on large refactors), but deprecation timeline is tight and early feature gaps are acknowledged. Worth piloting now if you're already on multi-agent patterns; risky if you depend on Gemini CLI's full feature set.