CloudCodeTree · Journal
AI News
Daily field notes on AI-assisted engineering.
Per page
Topics
Jul 13, 2026 · 20:12 UTC · 3 min read
Keep CLAUDE.md Under 200 Lines: Use @import to Modularize Project ContextEvery Claude Code session injects your full CLAUDE.md before any tool call. Use @import to split it into focused files and keep the main file under 200 lines — reducing per-session token overhead w…
Jul 13, 2026 · 20:11 UTC · 4 min read
When Retrieval Goes Wrong, Fix It: Build a Self-Correcting RAG Pipeline With LangGraphCRAG adds an LLM grader after each vectorstore retrieval. If scores are low it rewrites the query and falls back to web search — your RAG pipeline catches its own bad retrieval before hallucinating…
Jul 13, 2026 · 12:13 UTC · 3 min read
From Cron to Cloud Agent: Automate Any Dev Workflow With Claude Code RoutinesClaude Code Routines run your agent on a schedule, a webhook POST, or a GitHub event — all on Anthropic-managed cloud infrastructure, no machine required. Three trigger types, one setup form.
Jul 13, 2026 · 12:12 UTC · 2 min read
Any Hugging Face Model at Native vLLM Speed: One Flag Changes EverythingThe vLLM transformers backend now dynamically fuses layers at runtime via torch.fx, matching native vLLM throughput for any Hugging Face model. Enable it with --model-impl transformers.
Jul 13, 2026 · 12:11 UTC · 3 min read
Retrieve Small, Return Big: Better RAG With LangChain's ParentDocumentRetrieverParentDocumentRetriever solves the RAG precision-context tradeoff: embed small child chunks for accurate retrieval, return large parent chunks to the LLM for full context. Under 30 lines of Python.
Jul 13, 2026 · 04:05 UTC · 4 min read
Five Attacks That Break MCP Servers — and the Fixes the Official Spec RequiresThe official MCP Security Best Practices spec documents five exploitable attack classes — confused deputy, token passthrough, SSRF, session hijacking, and OAuth URL injection — each with concrete c…
Jul 13, 2026 · 04:04 UTC · 2 min read
PyTorch 2.13 Adds 12x FlexAttention Speedup on Apple Silicon; Mistral Opens Early Access to Frontier-Class Open-Weight MoEPyTorch 2.13 brings FlexAttention to Apple Silicon MPS with native Metal kernels (up to 12x speedup); Mistral opens July early access to a new "fat but sparse" open-weight MoE family.
Jul 13, 2026 · 04:03 UTC · 4 min read
Your Embedding Model Sees Chunks, Not Documents: Three Splitting Strategies Every RAG Engineer NeedsFixed-size, recursive, and semantic chunking — learn when each wins and wire LangChain's RecursiveCharacterTextSplitter and SemanticChunker in under 20 lines of Python.
Jul 12, 2026 · 20:05 UTC · 2 min read
Safe AI Code Execution in One API Call: Google Cloud Run Sandboxes Hit Public PreviewGoogle Cloud Run Sandboxes (public preview, July 9) give you millisecond-startup isolated execution environments for AI-generated code within existing Cloud Run services — zero outbound network by…
Jul 12, 2026 · 20:04 UTC · 3 min read
Merge Two Fine-Tuned LLMs Into One — No GPU Needed: Model Merging with mergekitCombine two fine-tuned LLMs into a single model that inherits both skills — no GPU, no retraining. mergekit SLERP/TIES/DARE + Maxime Labonne's free LazyMergekit Colab notebook.
Jul 12, 2026 · 12:05 UTC · 2 min read
Claude Desktop Is Now on Linux: Install the Full Dev Environment on Ubuntu or Debian in 3 CommandsAnthropic's official apt repo brings the complete Claude Desktop — parallel sessions, visual diff review, live app preview, Claude Code + Cowork + Chat — to Ubuntu 22.04+ and Debian 12+.
Jul 12, 2026 · 12:04 UTC · 1 min read
Claude Code 2.1.207 Breaks Hooks That Read User Config: Switch to exec Form Before Your Next Updatev2.1.207 rejects shell-form hook/MCP commands that interpolate ${user_config.*} variables. If your .claude/settings.json hooks use this pattern, they will silently stop running after the update.
Jul 12, 2026 · 12:03 UTC · 2 min read
Managed Vector Search, Zero Infrastructure: Pinecone Serverless From API Key to Production in 10 MinutesPinecone serverless manages your vector index in the cloud with no Docker or infra needed. This walkthrough covers the BYOV path with sentence-transformers: create index, upsert with metadata, quer…
Jul 12, 2026 · 04:14 UTC · 3 min read
Claude Code Now Has a Browser Built In: Read Docs, Click Designs, and Test Live Sites Without Leaving the AppClaude Code's desktop app now ships with a sandboxed in-app browser: Cmd+Shift+B opens a Chromium panel with an isolated profile, so Claude can read docs, click designs, and interact with live prev…
Jul 12, 2026 · 04:13 UTC · 4 min read
Payload Filtering Is Qdrant's Superpower: Go From Docker to Production Vector Search in 15 MinutesQdrant combines semantic similarity search with exact metadata filters using a filterable HNSW index. Build a filtered vector search engine from Docker pull to first query in 15 minutes with Python…
Jul 11, 2026 · 20:05 UTC · 2 min read
Tell Claude What Done Looks Like: Add a Grader Agent to Your Pipeline With Managed Agents OutcomesClaude Managed Agents Outcomes wires a rubric-graded iterate-grade-revise loop into your pipeline — define success once, send one event, and the harness iterates until the output passes.
Jul 11, 2026 · 20:04 UTC · 1 min read
Rotate API Keys Before They Burn You: Anthropic Console Now Sets Key Expiration DatesAnthropic Console now lets you set an expiration on any API key — presets from 3 hours to 30 days, custom durations, or Never, with email reminders and an expires_at field in the Admin API.
Jul 11, 2026 · 20:03 UTC · 3 min read
Your Agent Is Failing Silently: Design Rubric-Based Evals Like Anthropic DoesHow to build rubric-based agent evaluations with an LLM judge, trajectory checks, and CI integration — evaluation patterns from Anthropic's engineering blog.
Jul 11, 2026 · 12:07 UTC · 2 min read
Claude Code Auto Mode Is Now On By Default on Bedrock, Vertex, and Foundry — Here's What to ConfigureClaude Code auto mode is now enabled by default on AWS Bedrock, Google Vertex AI, and Azure Foundry — no env var needed. Bedrock also now defaults to Claude Opus 4.8. Here is how to configure it.
Jul 11, 2026 · 12:06 UTC · 2 min read
GitHub Copilot Desktop Goes Free for All and Lands in JetBrains With Codex AgentGitHub opened its agent-native Copilot desktop app to all tiers including Free on July 7, and shipped Codex as an agent provider in JetBrains IDEs — each session runs in an isolated git worktree.