+4,048 stars in 7d (#2 Python trending, 33.1k total); 34 commits on Aug 24–25 2026 alone. Stores memory/resources/skills as a virtual filesystem under `viking://` — agents `ls`/`tree`/`find` their own context instead of querying a vector store. Tiered loading (L0 ~100tok abstract → L1 ~2k overview → L2 full). AGPLv3 core, Apache-2.0 CLI; native Claude Code + MCP client support.
+2,520 stars in 7d (Rust + all-language trending). MIT. Git-versioned markdown wiki as source of truth + SQLite FTS5 for indexing; ships as an MCP server with lifecycle hooks. Pitch: quit Claude Code mid-task, start Codex in the same dir, continue without re-explaining.
+1,534 stars in 7d (Rust trending, 16.4k total), MIT, Python bindings on PyPI. Benchmarks: 3.4× faster than FAISS IndexPQFastScan at 4-bit, ~16× compression on 1536-dim at 2-bit, no training phase. Separately, the [pgvector PR #989](https://github.com/pgvector/pgvector/pull/989) adding `tqflat`/`tqivf`/`tqhnsw` was **closed unmerged** (last activity Jul 29 2026) — ankane wanted benchmarks vs. halfvec/binary quant and pointed at RaBitQ instead.
+615 stars in 7d, Apache-2.0, from NVIDIA-NeMo. Rust proxy *and* embeddable library that preserves native **Anthropic Messages** and OpenAI formats while routing across tiers; includes an escalation router that uses in-conversation signals (tool results, errors) to decide weak vs. strong model. Prometheus metrics built in. Explicitly labeled pre-alpha.
+1,313 stars in 7d (#2 TypeScript trending, 3.1k total). TypeScript/Node/Electron/React on SQLite; model messages, tool calls, tool results, permission decisions and termination events all recorded as an append-only log. BYO model (cloud, local, or gateway) via models.dev catalog. Apache-2.0, incubating at ASF, macOS Apple Silicon only so far.
[v1.0.0, Aug 20](https://github.com/anthropics/anthropic-sdk-python/releases). HTTP layer moves `httpx` → `httpx2`; requires Python 3.10+; removes legacy Text Completions and the `temperature`/`top_p`/`top_k` params on Messages; async `.with_raw_response` now needs `await response.parse()`; `AnthropicBedrock` errors instead of defaulting to `us-east-1`.
Aug 19 release notes. `/v1/files` and `/v1/skills` no longer need `files-api-2025-04-14` / `skills-2025-10-02`. Header-less requests get the *new* response shape: `expires_in_seconds` on upload, `expires_at` on file objects, `page`/`next_page` pagination plus an `ids[]` filter. Computer-use also left beta as `computer_toolset_20260801`.
[[email protected], Aug 20](https://github.com/cloudflare/workers-sdk/releases/tag/wrangler%404.125.0). New `connect` trigger: `{"connect":[{"protocol":"tcp","port":5432}]}` opens a listening socket on 127.0.0.1 forwarding to a `connect(socket, env, ctx)` handler (experimental flag, TCP only). Also `previews.containers` binding container apps to DO classes in previews.
[Bun v1.4, Aug 20](https://bun.com/blog/bun-v1.4). First production release of the Zig→Rust rewrite. Adds `bun audit fix`, `bun dedupe`, `bun prune`; `Bun.cron()`, `Bun.Image`, `Bun.WebView`, `Bun.Terminal`; `bun test --parallel/--shard/--changed`. 5× lower idle CPU, up to 35% less memory, ~7× faster warm CI installs.
[v0.33.0, Aug 21](https://github.com/ollama/ollama/releases) (pre-release). Turn individual Ollama models on/off for use in Claude from the menu bar. Stable 0.32.15 (Aug 19) cut time-to-first-token from ~995ms to ~524ms via model metadata caching.
[level1techs forum, Aug 22](https://forum.level1techs.com/t/why-your-local-llm-feels-dumber-than-it-is/253917) — 505 pts / 205 comments on HN. Reproducible finding: int4 KV-cache quantization causes **tool-calling failures** at long context (int8 recovers); attention backend choice (FA2 vs Triton vs Flash Inference) flips tokens; NVFP4 weights diverge ~50% by 88k tokens; wrong sampler temp → infinite loops.
[earendil.com, Aug 20](https://earendil.com/posts/what-is-a-harness/) — 557 pts / 178 comments. Names four components (system prompt, tools, agentic loop, model translation layer) and argues the harness — not the model — is the layer users can own. Heavy HN debate on lock-in.
[blog.cloudflare.com, Aug 20](https://blog.cloudflare.com/task-based-oauth-consent/) — devs mark scopes `optional`; users deselect at consent; tokens carry only granted scopes. Explicitly motivated by MCP: "an MCP server might request a broad set of permissions… most users would not want an agent to have that much access." Rolling out over weeks.
[openrouter.ai, Aug 19](https://openrouter.ai/blog/announcements/openrouter-is-joining-stripe/) — 962 pts / 497 comments, the week's biggest infra thread. 10T+ tokens/day, 400+ models; "nothing about your integration changes," closing in weeks. HN debate centered on whether a payments company can stay routing-neutral.
[neon.com, Aug 24](https://neon.com/blog/wal-s3-lakebase-storage-for-the-era-of-agents) — WAL becomes source of truth, pages are derived cache. Branch = pointer to an LSN, zero copy; scale-to-zero makes **per-agent databases** viable; time-travel queries without restore.