[blog.cloudflare.com/cloudflare-computer](https://blog.cloudflare.com/cloudflare-computer/) + [changelog 2026-08-03](https://developers.cloudflare.com/changelog/post/2026-08-03-cloudflare-computer/). Virtual FS living *inside a Durable Object*, SQLite-backed, with three pluggable exec backends (FUSE-mounted container, just-bash isolate, JS isolate in Dynamic Workers), capnweb RPC, MCP examples. +2,856 stars this week (8.2k total), #2 trending TypeScript. Marked PREVIEW ONLY.
+4,423 stars this week (21.8k total), v2.0.0 shipped, v2.0.1 in progress. Four-tier model: L0 raw conversation → L1 atomic facts/preferences → L2 project scenarios → L3 persona. Ships Chat Memory, Skills, Wiki, and CodeGraph as reusable assets; served via an OpenAI-compatible proxy (`/v3/tools/list`, `/v3/tools/call`) — **not** MCP.
+1,549 stars this week (3.6k total), Rust, #4 trending overall. No control plane, no consensus: each object is its own SQLite DB, replicated to S3/GCS, nodes coordinate through the bucket alone. Deploys existing Wrangler bundles. Only 5 commits on main — very new, unproven.
+5,135 stars this week, Python, #2 trending overall. "Open Source Palantir for AI Agents." Neo4j / FalkorDB / Apache AGE / Neptune backends plus vector stores, entity resolution, temporal tracking, Datalog/SPARQL/Rete reasoning without an LLM in the loop. v0.6.5 was a security release fixing 5 externally-reported vulns in the Explorer API.
+1,718 stars this week. Parses 13+ languages (TS/TSX/JS, Python, Rust, Go, C#…) with Tree-sitter into **Memgraph**, converts NL → Cypher, adds ast-grep structural matching and runtime call tracing.
[TS v0.117.0](https://github.com/anthropics/anthropic-sdk-typescript/releases) and [Python v0.122.0](https://github.com/anthropics/anthropic-sdk-python/releases), both Aug 13, 2026: "add `output_behavior` to dream creation (create new memory store or update input store in place)." Dreaming was previously strictly copy-on-write — a dream consumed a store + up to 100 sessions and emitted a *separate* store. Also in 0.117.0: streaming `message_delta` accumulation now applies all fields, and tool-runner forwards container IDs across turns.
[workers-sdk releases](https://github.com/cloudflare/workers-sdk/releases), Aug 13: c3 2.72.0 — "the Next.js framework scaffold now defaults to Vinext, with OpenNext available as an alternative variant." Vinext itself shipped [1.0.0-beta.6 on Aug 15](https://github.com/cloudflare/vinext/releases) (callable cached functions via RSC plugin API, Worker prerender param fixes) plus `@vinext/cloudflare` beta.6.
[workers-sdk releases](https://github.com/cloudflare/workers-sdk/releases), Aug 13: mock Access identities configurable in `wrangler.json` for local dev. Paired with an Aug 14 [Cloudflare changelog](https://developers.cloudflare.com/changelog/) entry enabling Access on a single Worker or all Workers at once (private-by-default with public exemptions). Wrangler 4.122.0 (Aug 12) also auto-detects `nodejs_compat` from compatibility date, fixing redundant-flag startup failures.
[Ollama releases](https://github.com/ollama/ollama/releases), all Aug 14: 0.32.11 adds web search to the OpenAI-compatible Responses API and `ollama launch dsh` / `ollama launch muse` agent harnesses; 0.32.12 adds Qwen 3.8 27B with an MLX variant for Apple Silicon ("substantial gains across coding and long-horizon agentic tasks"); 0.32.13 adds developer-instruction support for it.
https://www.mixedbread.com/blog/toast-1 (Aug 13, 2026; HN 206 pts / 62 comments). Mixedbread ships a dedicated search agent — decomposes queries, gathers evidence, curates context — claiming parity with Opus 5 / GPT-5.6 Sol on BrowseComp Plus and OfficeQA Pro at $0.30/M in, $0.72/M out, ~8s median per query, backend-agnostic. Harvey LAB result: 3.5× fewer tokens, same accuracy.
https://mun-logadan.github.io/why-does-opus-5-feel-worse/ (Aug 14; HN 876 pts / 795 comments — biggest LLM-practice thread of the week). Claim: Opus 4.7/4.8 stopped to ask when intent was unclear; Opus 5 reinterprets plans unprompted. Argues benchmarks are self-contained and therefore reward confident assumptions over clarifying questions.
https://claude.com/blog/maximizing-the-value-of-your-claude-code-sessions (Aug 14; HN 230 / 127). Concrete: `/context` exposes preloaded MCP tool definitions so you can disable unused ones; changing `/model` or `/effort` mid-session invalidates the prompt cache and forces full re-prefill; `/compact` before a break is cheaper than after the 1-hour cache TTL expires.
https://huggingface.co/Qwen/Qwen3.8-27B-FP8 (Aug 14; HN 1150 pts / 698 comments). Hybrid Gated DeltaNet + Gated Attention, 64 layers, 262K native / 1M via YaRN, `reasoning_effort` control, thinking preserved across turns. 73.0 Terminal Bench, 61.7 SWE-bench Pro, 84.3 OSWorld-Verified. FP8 block-128 for consumer hardware.
https://softwaredoug.com/blog/2026/08/10/hypothetical-classifications (Aug 10; HN 234 / 96). Doug Turnbull: skip structured-output constraint over a taxonomy. Let a small model emit a plausible-but-fake label, embed it, dot-product against embedded canonical vocabulary. Removes schema-size limits and works with weaker models.