[github.com/volcengine/OpenViking](https://github.com/volcengine/OpenViking) — 32.8k ★, +3,799 in 7d, #1 Python trending. Unifies agent memory + knowledge RAG + skills as a virtual filesystem under a `viking://` protocol; agents browse context with `ls`/`tree`/`find` instead of querying a black-box vector store. Tiered loading (L0 abstract / L1 overview / L2 details), MCP client support, AGPLv3 core. Backed by the VikingMem paper (VLDB 2026).
[github.com/semantica-agi/semantica](https://github.com/semantica-agi/semantica) — 10.6k ★, +2,317 in 7d. Context graphs where every decision is a first-class, causally-linked object with W3C PROV-O provenance on each fact. Backends include **Neo4j and pgvector**, plus a full MCP server (`python -m semantica.mcp_server`) exposing entity extraction, decision recording, and graph queries.
[github.com/anthropics/claude-plugins-community](https://github.com/anthropics/claude-plugins-community) — 1.2k ★, +406 in 7d. Read-only nightly mirror of the approved community marketplace for **Claude Cowork and Claude Code**. Submission is via `clau.de/plugin-directory-submission` with automated security scanning + manual review; direct PRs auto-close. Notably it does *not* cover MCP servers — those live in the separate GitHub MCP Registry.
[github.com/RyanCodrai/turbovec](https://github.com/RyanCodrai/turbovec) — 16.3k ★, +1,483 in 7d, MIT. TurboQuant (ICLR 2026) is data-oblivious: **no training pass**. Claims 10M docs in 4GB vs 31GB float32, 3.4× faster than FAISS IndexPQFastScan at 4-bit, single-vector insert in 6–20µs, O(1) delete by ID.
[Aug 19, 2026 release notes](https://platform.claude.com/docs/en/release-notes/api) — `files-api-2025-04-14` and `skills-2025-10-02` beta headers no longer required; new `browser_toolset_20260801` (accessibility-tree-driven, element refs, form input, tab management) plus `computer_toolset_20260801` out of beta with batch actions. Separately: Managed Agents sessions on self-hosted sandboxes can now attach memory stores, which the TS/Python/Go SDK workers sync to a `mount_path`.
[v1.0.0, Aug 20](https://github.com/anthropics/anthropic-sdk-python/releases). HTTP layer moves to `httpx2` (Pydantic's maintained fork); `temperature`/`top_p`/`top_k` removed from Messages methods (now `extra_body=`); `output_format=` → `output_config={"format":...}`; async `.with_raw_response` needs `await response.parse()`; Bedrock requires explicit region. [Full migration guide](https://github.com/anthropics/anthropic-sdk-python/blob/main/MIGRATION.md).
[workers-sdk releases, Aug 20](https://github.com/cloudflare/workers-sdk/releases). `wrangler dev` gains a `connect` trigger (`{"connect":[{"protocol":"tcp","port":5432}]}`) delivering raw sockets straight to a `connect(socket, env, ctx)` handler, bypassing the HTTP entrypoint (experimental flag required). Also `previews.containers`, batch Workflow instance deletion, and a `kv bulk put` binary-corruption fix. Separately [Aug 19](https://developers.cloudflare.com/changelog/): `@cloudflare/vitest-pool-workers` → `@cloudflare/vitest-plugin` v1.0.0 with a `vitest:pool-workers-to-vitest-plugin` codemod.
[Aug 20](https://bun.com/blog/bun-v1.4). Adds `Bun.cron()`, `Bun.WebView` (headless browser), `Bun.Image`, `Bun.markdown`, `Bun.Terminal` as built-ins; HTTP server memory down up to 48%, idle CPU 5×; +1,517 Node test-suite tests passing — Next.js 16, Playwright and vitest now run on Bun; `bun test --parallel/--shard`.
[Aug 20–23 releases](https://github.com/ollama/ollama/releases). 0.32.15 adds model metadata caching that cuts time-to-first-token roughly in half, plus Qwen 3.8 system-message normalization. 0.33.0-rc0→rc2 add Claude model management in the desktop app (toggle individual models, select from available Ollama models) and fix prefill restore points causing agent client hangs.
[openrouter.ai/blog](https://openrouter.ai/blog/announcements/openrouter-is-joining-stripe/) (Aug 19) — 957 pts / 497 comments on HN. OpenRouter (10T+ tokens/day) acquired at a reported ~$7B; promises "same product, same roadmap," routing "doesn't bend to any parent company." HN thread is skeptical: neutrality-under-Stripe, Stripe seeing all corporate inference traffic, and whether routing is commoditizing. Named alternatives surfaced: Cortecs.ai, EUrouter.ai, Requesty.ai. OpenAI's DevEx lead commented in-thread.
[Level1Techs forum](https://forum.level1techs.com/t/why-your-local-llm-feels-dumber-than-it-is/253917) (Aug 22, 417 pts / 171 comments). Same weights, different serving stack → measurable divergence: attention backend choice (FlashAttention 2 vs Flash Inference vs Triton) produced 0–8% top-1 token disagreement rising with context depth; KV-cache quantization to INT4 caused unrecoverable **tool-calling failures**; NVFP4 hit ~50% token flips at 88k context, while INT8 W8A16 beat official FP8 numbers.
[Wiz research](https://www.wiz.io/blog/red-agent-snowflake-copilot-cicd-bug) (Aug 17, 424 pts / 157 comments). GitHub Actions script injection via crafted issue title (`${{ github.event.issue.title }}` interpolated into a shell `echo`); a quote broke out and exfiltrated Jira creds. Introduced by PR #1218; Copilot reviewed it and marked it all-clear, and GitHub Advanced Security's scan also missed it.
[raphaelbauer.com](https://www.raphaelbauer.com/posts/postgresql-everything/) (Aug 19, 438 pts / 267 comments). Argues Postgres subsumes Elasticsearch, Redis, Kafka (`SELECT ... FOR UPDATE SKIP LOCKED`), ClickHouse — and **Neo4j via Apache AGE** — with pgvector/pgai covering embedding + retrieval in-database.