CEREBRO

Researched and written with 310,659 tokens (in 10 · out 32,212 · cache-read 66,075 · cache-create 212,362) across 5 Claude calls

Coding agents: extensibility and mods

You can now mod Claude Code: Claude Code mods are a few lines of TypeScript, or built by Claude itself, that change behaviour, customize the UI, or swap in features, and they ship inside plugins via /plugin.
Introducing the Mods Hub for Command Code: open-source registry to discover, install, and publish tools, commands, hooks, and rules for Command Code, with 30+ mods already live.
Claude code new project feature - beta: Reddit thread on a beta "project" feature in Claude Code; no detail beyond the title.
JetBrains/thinkrail: IDE built around the pi coding agent, with isolated workspaces, durable project knowledge via living specs and reusable skills, and extension support coming.
Sync session data between different CLI harnesses: a way to carry session history across agent CLIs, which matters if you switch between Claude Code, Codex, and others.

The State of Context Engineering: Breunig's AGNTCon talk argues the job is knowing when to be deterministic and when probabilistic. Commenters agree that both too little and too much context hurt, but reject a fixed 200-300k reset threshold because newer models need that much loaded to perform well.
TokenGate: token-gated attention makes long-context prefill up to 11.3x faster: Stanford and UC Berkeley attention variant that cuts long-context prefill time, which is directly relevant to long-context agent latency and cost.
How I stopped Qwen2.5-Coder-14B from hallucinating UI & workflows: pairs a small local model with a deterministic C# ValidationResolver on a single 24GB A10, an example of the deterministic-guardrail pattern from the Breunig piece.

Opus 5.5 orchestrating sonnet agents: Reddit report on the Opus-orchestrates-Sonnet-workers pattern; no detail beyond the title.
Opus 5.5 writes slop code, but succeeds anyways: SlopCodeBench-inspired research suggesting messy code quality and task success are decoupled for current frontier models.
What Claude Code actually runs for my one-person company after six weeks: field report on real production use over six weeks, including five things it still can't do.
multica-ai/andrej-karpathy-skills: a single CLAUDE.md distilled from Karpathy's observations on LLM coding pitfalls (wrong assumptions made on your behalf, then run with), cheap to drop into a project.
mattpocockuk on /grill-me: a prompt that has the agent ask one question at a time when the problem is badly underspecified, avoiding one big Q&A round that sends the work in the wrong direction.
mattpocockuk on /wayfinder: an escape-valve command for tasks that obviously need more than one session to plan.
ADRs for AI: versioned-in-Git deliberation records for human-AI work, so the reasoning from sessions doesn't evaporate. HN commenters call it standard ADRs with marketing gloss.

Proc-CLI: agent-friendly process and port manager where proc kill :3000 ends whatever listens on that port, with confirmation and --dry-run, across macOS, Linux, and Windows.

AgriciDaniel/claude-seo: Claude Code SEO plugin with 26 sub-skills and 19 sub-agents covering technical SEO, schema, GEO/AEO, and agent readiness (WebMCP, llms.txt), running a full audit in 10-15 minutes.