[github.com/tinyhumansai/openhuman](https://github.com/tinyhumansai/openhuman) — +1,321 stars this week, 39.4k total, "#1 on GitHub nine days running" after its beta. Rust core + Tauri shell, Ollama for fully-local inference, Privacy Mode enforcing local-only. Memory is a "Memory Tree" of scored Markdown in SQLite, mirrored as an editable Obsidian vault — explicitly rejecting "vector-soup black box."
[github.com/abhigyanpatwari/GitNexus](https://github.com/abhigyanpatwari/GitNexus) — +1,174 stars this week (46.9k total). Tree-sitter → LadybugDB (ex-KuzuDB, embedded graph *with* vector support), BM25 + embeddings via reciprocal rank fusion, 17 MCP tools. Key design bet: precompute execution flows and community clusters at index time so one query returns full context instead of 4+ traversals.
[github.com/rohitg00/agentmemory](https://github.com/rohitg00/agentmemory) — +476 stars this week, 28k total. Claims 95.2% R@5 retrieval. SQLite + in-memory vector index, four tiers (working/episodic/semantic/procedural), and three ingest paths: 54-tool MCP server, 12 lifecycle auto-capture hooks, and a REST API on :3111.
[github.com/anthropics/claude-plugins-community](https://github.com/anthropics/claude-plugins-community) — +1,212 stars this week. Read-only nightly mirror of the reviewed marketplace; submissions go through clau.de/plugin-directory-submission with automated security scan + manual approval. Install via `claude plugin marketplace add anthropics/claude-plugins-community`. Notably it does *not* cover MCP servers (those route to the separate MCP Registry).
[github.com/apache/maka](https://github.com/apache/maka) — +991 stars this week (4.6k total), incubating at the ASF, no official release yet. Model messages, tool calls, results, permission decisions and termination events all persist to an append-only log that feeds UI projections and context management; `packages/mcp/` is a provider-neutral MCP client.
`claude-fable-5-1` / `claude-mythos-5-1`: 1M context default, 128k max output, always-on adaptive thinking, $10/$50 per MTok — but **cache reads dropped to $0.25/MTok (0.025× base input)**. Breaking: `tool_choice: "any"` and `"tool"` now return 400 on these models (use strict tool use / structured outputs); thinking blocks only replay from the producing model or newer. ([release notes](https://platform.claude.com/docs/en/release-notes/api), [announcement](https://www.anthropic.com/news))
TS `0.122.0` / Python `1.2.0` — `client.beta.files` and `client.beta.skills` drop beta headers and return GA shapes. Breaking: `beta.skills.delete()` now deletes a Skill *and all its versions*; `BetaSkill` → `BetaContainerSkill`; `webhooks.unwrap()` now **requires** headers. Follow-up `0.123.0` / `1.3.0` (Sep 1) adds memory-store and toolset schema updates. ([TS 0.122.0](https://github.com/anthropics/anthropic-sdk-typescript/releases/tag/sdk-v0.122.0))
New `observability.redact_query_string` config strips query strings from Worker logs; Workflows list gains `--date-start` / `--date-end` (ISO 8601); deploy output now reports exact raw + gzip bundle sizes in structured form. Install size cut 6.16 MiB (31.4%) for Wrangler, 1.06 MiB for Miniflare. ([release](https://github.com/cloudflare/workers-sdk/releases/tag/wrangler%404.128.0))
As of Sep 1, D1 queries on the Workers Free plan **error out** once daily row read/write limits are hit (email alert at threshold, reset midnight UTC) — previously not enforced. Separately, a single Durable Object can now host up to 10 concurrent distinct Dynamic Workers, up from 4.
0.33.2 fixes the Claude Desktop proxy interrupting in-flight requests during model-catalog updates, plus macOS multi-instance and appearance bugs. 0.33.3-rc2 adds **cached prompt token reporting** and honors GGUF-defined default params. Builds on 0.33.0's caching work (TTFT ~995ms → ~524ms).
[embracethered.com](https://embracethered.com/blog/posts/2026/breaking-claude-code-opus-5-and-automode/) — HN Aug 31, 397 pts / 121 comments. Attacker forces a fallback from WebFetch to `curl` (HTTP 415), ships a ZIP whose `struct.py` shadows stdlib when Claude writes *its own* decoder and imports `base64`. 60–80% success across variants; Auto Mode's classifier only ever saw the innocuous decoder, and then *blocked cleanup*. Anthropic rated it "Informative" — Auto Mode is convenience, not a security boundary.
[github.com/carloslfu/slotstream](https://github.com/carloslfu/slotstream) — HN Sep 1, 227 pts / 110 comments. Keeps the 3.8GB dense trunk resident, `pread`-streams the 68GB of routed experts through an auto-sized slot pool (~33GB on a 48GB machine) instead of mmap'ing 100GB. Byte-identical output across cache sizes. MIT, Apple Silicon only.
[calv.info](https://calv.info/small-models-have-arrived) — Calvin French-Owen (Segment co-founder, ex-OpenAI). HN Aug 27, 799 pts / 348 comments. Splits work into novel problem-solving (frontier) vs. high-volume "token spewer" routine work, now ~$0.10/task at ~100 tok/s vs ~$1 a generation ago. Claims ~95% of one engineer's workload is bucket two.
[anthropic.com](https://www.anthropic.com/claude-fable-and-mythos-5-1) — Sep 1. HN 1,391 pts / 1,355 comments. $10/$50 per MTok, cache reads $0.25/MTok (75% cut); ~25% cheaper typical, ~45% on agentic tasks. Explicitly cites improved prompt-injection safeguards and long-running unattended task persistence.
[blog.cloudflare.com](https://blog.cloudflare.com/botbase-for-operators/) — Aug 28. Operators self-declare action, content-use level, and operator-vs-intermediary status; Cloudflare auto-verifies IP lists, rDNS, and Web Bot Auth signatures. Low HN traction, so signal is thin — but it lands on Sri's host.