[calpaterson.com/memoryfields.html](https://calpaterson.com/memoryfields.html) — 190 pts on HN, Aug 31. Proposes markdown pages + YAML frontmatter + a SQLite vector index, zipped. Explicitly names "systems requiring pgvector, Neo4j, and dedicated LLMs" as over-engineering that "confuses models and doesn't scale," and argues graph traversal is slow and strips context.
[github.com/workweave/router](https://github.com/workweave/router) — 3,578 stars, **2,510 gained this week** (Go trending #2). Per-action routing in <50ms via in-process ONNX cluster scorer (Avengers-Pro), not heuristics. Self-hostable with Docker/Postgres, keys stay local. Claims Robinhood, PostHog, Reducto as users.
[github.com/abhigyanpatwari/GitNexus](https://github.com/abhigyanpatwari/GitNexus) — +1,149 stars this week. LadybugDB (ex-KuzuDB) compiled to WASM + Tree-sitter, builds the graph fully in-browser. Precomputes community detection, call-chain flows and confidence scores at index time rather than handing raw edges to the LLM; exposed over MCP.
[github.com/apache/maka](https://github.com/apache/maka) — +1,285 stars this week, 4,218 commits, nightly releases. Records model messages, tool calls, results, permission decisions and termination as an **append-only event log**; SQLite operational state; provider-neutral MCP client. Key design point: old tool output can be dropped from the prompt without discarding the saved evidence.
[github.com/anthropics/claude-plugins-community](https://github.com/anthropics/claude-plugins-community) — +1,634 stars this week (Python trending #1); the official `claude-plugins` repo added +2,014. Curated mirror of the Claude Code / Cowork marketplace; submissions via clau.de/plugin-directory-submission, gated on automated security scanning.
[TS releases](https://github.com/anthropics/anthropic-sdk-typescript/releases) / [Python releases](https://github.com/anthropics/anthropic-sdk-python/releases). Three minors in a week: **Standard Schema support for structured outputs and tools** (v0.121.0), **beta files/skills namespaces moved to GA shapes + dated beta header pins dropped** (v0.122.0), and **memory-store and toolset schema updates** (v0.123.0). Python v1.2.0 also fixed the memory tool to "preserve exact file bytes… no newline translation."
[Advisories](https://github.com/honojs/hono/security/advisories): parseBody() unbounded dot-notation nesting → heap exhaustion (GHSA-g6gw-c38x-mqfc), toSSG() path traversal (incomplete fix for CVE-2026-39408), and a query-parser cache-key differential. All patched in 4.13.5.
[workers-sdk releases](https://github.com/cloudflare/workers-sdk/releases). Wrangler dropped **6.16 MiB (31.4%)** via workspace dependency optimization, added `--date-start`/`--date-end` filtering to Workflow instance listings, query-string redaction in observability config, and now prints raw + gzip bundle metrics on deploy.
[Ollama releases](https://github.com/ollama/ollama/releases). v0.33.1 added **MLX Qwen3.8 Flash Next support** and **structured output support in mlxrunner**, plus a fix for Metal GPU timeouts loading models from slow storage. v0.33.3 (pre-release) honors GGUF-defined default parameters.
[calpaterson.com/memoryfields.html](https://calpaterson.com/memoryfields.html), HN 190pts/93 comments Aug 31 ([thread](https://news.ycombinator.com/item?id=49508317)). Proposes memory as a ZIP of markdown pages + YAML frontmatter + optional SQLite/nomic-embed index; agents write prose directly, no chunking/re-ranking. Explicitly attacks needing "pgvector, Neo4j graph database and an LLM" to decide what's worth remembering. Comments split: "just markdown + RAG"; docheinestages — "The Achilles' heel of this approach is the RAG"; memory-poisoning/"heresies" thread on pruning > retrieval.
[vercel.com/blog/the-best-workflow-engine-is-a-programming-language](https://vercel.com/blog/the-best-workflow-engine-is-a-programming-language), Aug 27. `"use workflow"` / `"use step"` directives; compiler splits into workflow/step bundles. Runs as a library over a "World" interface — swap Postgres/Redis/Kafka without touching workflow code. Apache-licensed client, not Next.js-specific. Paired with [Run SDK](https://vercel.com/blog/introducing-run) (sandboxed agent eval, pause for human approval) Aug 25.
[calv.info/small-models-have-arrived](https://calv.info/small-models-have-arrived) (Calvin French-Owen, Aug 26) — HN 799pts/348 comments. Splits work into "IQ 180" vs "token spewer," claims ~95% is the latter; a task that cost ~$1 now ~$0.10. Names GPT-5.6-Luna (~100tps), GLM 5.3, Fable 5, Sonnet-class. Adjacent: GLM-5.3-Flash (1,130pts) and GLM-5.3 open weights (805pts) both landed this week.
[acceptmarkdown.com](https://acceptmarkdown.com/), HN 176pts/108 comments Aug 26. Content negotiation returns markdown to agents instead of HTML — up to 97% bandwidth cut. Adoption thin (Anthropic's Claude Code docs, `.md` URL suffixes). Criticism is sharp: cloaking/prompt-injection surface, "can never be trusted except for a whitelisted set."
[embracethered.com](https://embracethered.com/blog/posts/2026/breaking-claude-code-opus-5-and-automode/) (wunderwuzzi, Aug 26), HN 397pts/119 comments. Auto Mode's safety classifier bypassed via a chain: HTTP 415 pushes Claude from WebFetch to curl → ZIP with poisoned `struct.py` → module shadowing on import → RCE. 60–80% success. Not MCP-specific; Bash + Python stdlib shadowing.