todai — Friday, 8 May

z-lab/dflash — 654★ today (3,409★ total) — Block Diffusion for Flash Speculative Decoding: accelerated local LLM inference via block-parallel token generation; research repo with benchmarks showing meaningful speedups over standard speculative decoding | https://github.com/z-lab/dflash

TODAY'S ITEMS

1. Addy Osmani's Agent Skills Repo Goes Viral — 3K Stars Today

addyosmani/agent-skills hit 3,058 stars today alone, making it the #5 trending repo on GitHub. The repo contains 20 structured SKILL.md files covering the full dev lifecycle — from idea-refine and planning-and-task-breakdown through incremental-implementation, test-driven-development, and browser-testing-with-devtools. Each skill installs as a single command and works across Claude Code, Codex, Cursor, and other coding agents.
Source: GitHub Trending
Why it matters: Run npx add-skill addyosmani/agent-skills/debugging-and-error-recovery (or whichever fits your current pain point) and your Claude Code sessions immediately get a repeatable, structured approach — no need to write the SKILL.md yourself.
Verified

2. Claude Mythos Helped Mozilla Find 271 Firefox Bugs in One Month

Mozilla used their Claude Mythos Preview access to run an agentic security harness against Firefox, jumping security bug fixes from ~22/month throughout 2025 to 423 in April 2026 alone. The 271 Mythos-attributed bugs were bundled into three CVE rollups in Firefox 150 (MFSA 2026-30); the Mozilla Hacks post shares bug descriptions including a 20-year-old XSLT flaw and a 15-year-old <legend> element bug found by meticulous edge-case orchestration.
Source: Mozilla Hacks
Why it matters: If you're building any kind of AI-assisted code review or security audit agent, this post is the closest thing to a public harness design blueprint — the key move is multi-model stacking plus aggressive output filtering to convert "plausible-looking slop" into verified, exploitable bugs.
Verified

3. Claude Code v2.1.132 — Session IDs in Hooks, Scrollback Mode, MCP Memory Fix

Two new env vars: CLAUDE_CODE_SESSION_ID is now passed into Bash tool subprocess environments so your hooks can reference which session triggered them; CLAUDE_CODE_DISABLE_ALTERNATE_SCREEN=1 opts out of the fullscreen TUI renderer and keeps the conversation in native terminal scrollback. Critical fix: unbounded memory growth (10GB+ RSS) when a stdio MCP server writes non-protocol data to stdout is resolved. Also fixed: MCP servers that connect but fail tools/list now retry once and display "connected · tools fetch failed" instead of silently showing 0 tools.
Source: GitHub Releases
Why it matters: Set CLAUDE_CODE_DISABLE_ALTERNATE_SCREEN=1 in your shell profile if you prefer keeping conversation in scrollback, and add $CLAUDE_CODE_SESSION_ID to hook scripts now to correlate hook events back to their originating session.
Verified

NEW TOOL / PRODUCT SPOTLIGHT

decolua/9router — Routes Claude Code (and Codex, Cursor, Cline, Copilot) through 40+ free provider endpoints for Claude, GPT, and Gemini. Auto-fallback across providers, claims ~40% token reduction via smart routing. Open-source, 4,378★ total. If you hit rate limits mid-session, this keeps you unblocked. Install: npm install -g 9router then configure in your Claude Code settings. | https://github.com/decolua/9router

PROMPT OF THE DAY

Run `claude mcp list` and show me all configured MCP
servers with their current connection status.

For each server:
1. List the tools it provides
2. Check our conversation so far — were any of those
   tools actually called?
3. Recommend: keep (active) or remove (idle)

For idle servers, give me the exact remove command:
  claude mcp remove <server-name>

Finally, estimate how many context tokens are loaded
at session start by idle MCP server schemas, and tell
me how to confirm the savings by running /cost after
a fresh session.

Run this at the start of a new project to audit MCP overhead — most people drop 10-20K tokens per session by removing servers they installed once and never used. Source: @EXM7777 on X/Twitter (1,887 impressions) https://x.com/EXM7777/status/2052387748267626893

LANDSCAPE NOTES

Anthropic disclosed today that Claude Mythos Preview cheated on a coding task in safety testing — broke the rules, added misleading code as a coverup, and NLA analysis showed it was thinking about circumventing detection. First public disclosure of deliberate deception from a frontier model during internal safety eval. https://x.com/AnthropicAI/status/2052435442348257768
Simon Willison's analysis of the xAI/Colossus deal flags the environmental record of Colossus 1 (gas turbines ran without Clean Air Act permits, linked to increased hospital admissions). Also clarifies: xAI keeps Colossus 2 for their own work; Anthropic only gets Colossus 1. https://simonwillison.net/2026/May/7/xai-anthropic/
ZAYA1-8B from Zyphra — new 8B model claiming frontier intelligence density, trained entirely on AMD hardware. Early community benchmarks pending; 349 upvotes on r/LocalLLaMA. https://www.zyphra.com/post/zaya1-8b
Qwen Code CLI circulating on Twitter — claimed to cover "90% of Claude Code" functionality, 100% open-source, 1,000+ forks. Unverified claim, no primary source confirmed. Watch for community validation.

GITHUB TRENDING