Claude Code Projects remembers dev tasks

Hi there, this is your daily ☕️ Devshot.

In today's Devshot:

🤖 Claude Code Projects remembers dev tasks

🍎 Xcode's new .xcproj files are JSON

🔀 htmx 4.0 ships fetch-based rewrite

🧭 Safari 27 adds MCP support for devs

🦆 DuckDB-Wasm now persists to browser disk

Plus: 🎁 7 other news you might like, 🧰 6 tools, and 📚 5 papers.

🤖 Claude Code Projects remembers dev tasks LINK

Anthropic today launched Claude Code Projects in beta, adding a persistent coordination layer that keeps a shared memory of an ongoing development effort across multiple Claude Code sessions so developers no longer manually reconstruct context each time.

A coordinator conversation sits above worker threads, each running as its own cloud session on a separate branch, routing new natural-language requests into existing or fresh threads while tracking decisions, dependencies and merge order across repositories.

The beta reaches select Claude Pro and Max subscribers using cloud sessions, expanding to more users over the coming week; threads can't yet touch local files or internal networks, though Anthropic says local support arrives "very soon."

🍎 Xcode's new .xcproj files are JSON LINK

Xcode now offers a JSON-based project format called .xcproj that replaces the notoriously merge-hostile project.pbxproj build graph inside the .xcodeproj bundle, and you enable it from the file inspector.

The format swaps the old flat ID table for a tree mirroring the UI, with files, targets, and build-settings structures, so a file declares its own target-membership and adding one no longer means editing several cross-referencing spots.

Apple also shipped xcode-project-format, a Swift library modeling the format as typed value types under an XCSchema namespace, giving generators, linters, and validators a shared model instead of each reverse-engineering the JSON.

🔀 htmx 4.0 ships fetch-based rewrite LINK

htmx 4.0 has landed, its first major release since 2.0 in 2024, rewriting the library's transport layer onto the fetch API while keeping the script around 14KB and skipping version 3 entirely.

The rewrite unlocks native streaming, idiomorph-powered morphing swaps that preserve DOM state like input focus, and a new tag that updates several targets from one response, a cleaner replacement for out-of-band swaps.

Attribute inheritance is now explicit and can silently break CSRF tokens with a 403, so htmx ships under npm's next tag, keeps 2.x supported indefinitely, and adds a CLI checker plus LLM skills files for upgrades.

🧭 Safari 27 adds MCP support for devs LINK

Safari 27.0 ships a built-in MCP server that lets coding agents like Claude Code or Codex drive the browser window, so they can see how a developer's code actually renders without manual screenshot handoffs.

The server exposes the DOM, network requests, screenshots, and console output, letting agents compare computed styles across browsers and flag accessibility problems like missing labels, bad ARIA attributes, and poor contrast.

Enabled via Safari > Settings > Developer with "Allow remote automation and external agents," it runs locally, makes no network calls, and sends captured data only to your agent, not to Apple.

🦆 DuckDB-Wasm now persists to browser disk LINK

DuckDB-Wasm can now persist a full database to browser disk via the Origin Private File System, so a file opened at an opfs:// path survives page reloads and browser restarts without manual Parquet serialization or IndexedDB juggling.

Opening opfs://analytics.duckdb in READ_WRITE mode produces a standard .duckdb file with a write-ahead log and checkpoints; the resulting file works unchanged in the CLI or Python client if you pull the bytes back out.

Because browser tabs rarely close cleanly, call CHECKPOINT after each write batch to flush the WAL to OPFS, and pin version 1.32.0 or 1.33.1-dev64.0, since the current npm latest creates files but never actually persists writes.

Other news you might like

• Intel squeezed a 1.58-bit LLM down to 1.485 bits without changing a single weightLINK

• HCP Vagrant deprecation: important dates and migration guidanceLINK

• AWS's OpenSearch pitch ahead of OpenSearchCon: retrieval, serverless costs and agent memoryLINK

• AlmaLinux 9.9 Beta is hereLINK

• Auditing in the age of (good enough) AILINK

• Chess in the terminalLINK

• Be alert: targeted attacks on prominent RustaceansLINK

🧰 Trending tools

SKI: adds voice control to AI coding agents like Claude Code and Codex, letting you dictate commands and hear responses hands-free while codingLINK

Dograh: open-source voice agent platform with a visual flow builder, telephony, human transfer, and self-hosted QA monitoring across 30-plus model integrationsLINK

tiun: an AI-native backend combining authentication, payments, customer database, and analytics in one platform, so teams ship and scale faster.LINK

Humalike x Hermes: gives AI agents turn-taking, timing, and memory APIs to decide when to speak, wait, or interrupt, ideal for game characters and roleplay.LINK

Kilo Code for JetBrains: an open-source AI coding agent for JetBrains IDEs, letting you build and iterate with flexible model choices across projects.LINK

CartAI: automates checkout across merchant sites via one API, handling payments, shipping, and order tracking without building per-site integrations.LINK

📚 Trending papers & reports

Telemetry-to-insight automation turns raw, noisy app event logs into business-ready metrics without hand-built rules, lifting semantic quality from 50 to 80+ on a 100-point scale while cutting maintenance work by 80%.LINK

Coding agent design breaks down the wrapper software around AI coding tools into swappable parts, showing that trimming a model's working memory prevents crashes and that planning saves money for stronger models rather than boosting accuracy.LINK

Financial filing assistants can now fix their own recurring mistakes after launch without breaking answers that already worked, cutting the error rate from 20% to about 13% while only approving 6 of 33 proposed fixes.LINK

Spreadsheet question-answering gets more accurate when a system labels each cell's role so the AI understands the grid's layout, though the authors argue truly reliable answers require flattening spreadsheets into plain text.LINK

Compliance auditing turns regulations into checklist-style code an AI can run against a case, giving founders and legal teams accurate privacy assessments that show which rule was applied and what evidence backs each decision.LINK

See you tomorrow for a new dose of ☕️ Devshot!