July 19, 2026

Set runs-on to xcode-27 or xcode-27-xlarge to use the preview on a GitHub-hosted macOS runner. The new support model names images around a major Xcode version and supports one major Xcode line per image, making the selected toolchain more explicit than an operating-system-centered label.

The image is arm64-only and its bundled tools differ from earlier images. Add it first as a non-blocking matrix job and compare warnings, tests, signing, simulator behavior, package resolution, and duration with the pinned current image. Keep a known-good release path until the preview's installed software and availability meet the team's requirements.

VulnHunter goes beyond listing suspicious locations by exploring paths from inputs to sensitive operations, documenting the conditions and code evidence, and proposing a remediation. The repository provides a quickstart, architecture documentation, and annotated workflows. Apache 2.0 permits teams to inspect, modify, and redistribute the implementation.

The initial implementation assumes Claude Opus 4.8 and Claude Code; portability to other harnesses and models is described as potential, not a verified capability. Evaluate it on a fixed mix of known vulnerabilities and harmless controls, recording recall, false positives, evidence accuracy, post-fix tests, cost, and execution permissions. Do not close a release gate using only the same model's self-validation.

The four-person software team reports that median Pebble 2 Duo battery life increased from 17 days last summer to more than 30 days, with Pebble Time 2 around 21 days. SDK work exposes Round 2 touch, speaker, RGB backlight, and quick-launch capabilities. These are medians, so application mix and usage patterns still need device-level measurement.

On iOS, the legacy arrangement in which the app exposes the PPoGATT service blocks AccessorySetupKit. Reversing that relationship requires recovery-firmware upgrades across installed watches before ASK and eventual EU notification replies can be enabled. Unresolved step and sleep accuracy, accelerometer stops, and touch behavior make recovery, data integrity, and long-duration power testing as important as a successful OTA update.

Comic Chat rendered IRC as a comic rather than a text stream. It selected characters, orientation, expressions, gestures, balloon shape, and panel changes from conversational cues. Development began in 1995 using Visual C++ 4.0 and MFC, and the client shipped with Internet Explorer 3 in 1996. The archive spans pre-1.0 source through the 2.5 beta line.

The modern folders demonstrate Visual Studio 2022 builds, high-DPI work, current IRC behavior, and optional TLS, but they are worked examples rather than a polished rerelease. The repository is archived, and Microsoft frames the release around preservation and study. Pin source and toolchains for research; for reuse, separately audit legacy MFC dependencies, networking, encodings, and the rights surrounding bundled art assets.

The presentation starts from distributing Claude Code across the organization in June 2025 and Codex in September 2025 without mandating a usage pattern. The observed structure was that requirements, testing, and operations did not accelerate at the same rate as implementation. More delegation therefore increases the importance of designing judgment and context flow.

The proposed response is to express acceptance as deterministic contract tests and contain change through observability, staged rollout, and rollback. Architecture should localize the context and blast radius required for one change. Measure requirement queues, review queues, integration failures, rework, and operating intervention alongside coding time, then judge the intervention by end-to-end lead time rather than the fastest local stage.

The deck begins with a capacity problem: more AI-generated implementation also creates more review work downstream. Specification-driven development can reduce rework but does not guarantee that the specification is correct or cover independent security and operating concerns. The proposed structure adds evaluator agents separate from producers and combines multiple review perspectives, tests, and output scoring.

The target division is a producing agent, an evaluating agent, and a human final check. Humans define review policy, success conditions, thresholds, and permissions, then improve the system from evaluation results. In practice, give producer and evaluator distinct context, use read-only judgment, deterministic tests, preserved evidence, loop limits, and human approval for material changes. Measure misses and rework, not the number of automated reviews.