CEREBRO

Researched and written with 312,720 tokens (in 10 · out 31,337 · cache-read 65,545 · cache-create 215,828) across 5 Claude calls

Prompt caching & context

Codex & Work keep cache forever for logged-in users: Theo calls it the biggest ship of the 28 days; resuming an old conversation no longer costs a cold-cache penalty.
12h API cache, unlimited for logged-in users: Day 16 post spells out the detail: 12h cache in the API, and logged-in Codex & Work users are guaranteed a cache hit after a few hours away.
T3 Code Stable, 1,400+ changes: Includes beta "idle Claude threads compact early so the cache stays warm", a direct caching trick, plus remote hosts and PR diffs split into code, tests and docs.
codex-subscription-router: Spreads new chats across several ChatGPT subscriptions but pins each thread to one account, so follow-up turns keep account-level caching.
Graphify: /graphify skill builds a queryable knowledge graph of a codebase via local deterministic AST parsing with no vector store, pitched as an 80% cut in agent token spend.
Ponytail: "Laziest senior dev" skill; its own benchmark (39 tasks, Opus 5.5) claims 45% fewer tokens and 26% lower cost, with more risky logic covered by tests.

SlopCodeBench 10/2026: Opus 5.5 and Astra results; "xhigh considered harmful", and commenters say medium effort is best for most tasks because higher effort mostly adds unrequested output.
Devin Fusion scores 68 on Coding Agent Index: Opus 5.5 lead plus a GPT-6 Luna sidekick running in parallel beats Claude Code with Opus 5.5 by 2 points at 31% lower cost per task, which supports the lead-plus-cheap-worker pattern.
Human-centered coding agents: Position paper arguing for agents built to collaborate with people rather than run fully autonomously.
Claude Code v2.1.306: Fixes claude -p reporting "Not logged in" in cloud sessions, and a bug where an API error was replaced by "prompt is too long" and triggered needless compaction in sessions with many MCP tools.
Claude Code v2.1.305: Adds sandbox.allowNestedUserNamespaces for Chromium and rootless podman, which also lets sandboxed commands get around the sandbox's file protections, plus CLAUDE_CODE_RETRY_WATCHDOG_USAGE_LIMIT_RECHECK_MS for unattended retry mode.
Praex: Product Hunt launch of a coding agent that runs on its own models; there's no detail beyond the tagline.

Using Claude to reduce the attack surface of code: Anthropic's early research on using frontier models for defence (automated attack-surface reduction) rather than only finding and exploiting vulnerabilities.
cloudflare/security-audit-skill: Multi-phase audit skill (recon, hunting, candidate validation, independent verification, reporting) that seeded Cloudflare's vulnerability discovery harness.
OpenCode Console key theft: An attacker broke into a legacy system on October 16 and stole provider API keys users had saved on OpenCode's servers; some users found out from the bill. Don't store provider keys with a third party.
Nexus-CLI: Go proxy that exposes DeepSeek V3/R1 to Claude Code and Ollama via a free web account instead of an API key; that depends on scraping the web login, so treat it as fragile and a terms-of-service risk.

wshobson/agents: Plugin marketplace of 94 plugins that works across Claude Code, Codex, Cursor, OpenCode, Copilot, Antigravity and Pi.
agentic-awesome-skills: Local catalog of 2,700+ SKILL.md playbooks with a CLI and MCP so the agent searches, selects and previews a plan you can inspect before anything changes.
agency-agents: Roster of persona-style specialist agents, now with a native app that installs them into Claude Code, Cursor, Codex and Gemini in one click.

Reason Network OSS: Free, P2P-encrypted, open-source network that lets your Claude Code or Codex talk to other online agents, including teammates' agents.
call4me: Gives an AI agent a phone number that can place and answer calls.
Morse: VS Code and browser UI for Pi that drives its RPC mode, with multi-project and multi-session support, image paste and interrupt.
Pi Desktop: Standalone cross-platform desktop workbench for Pi that puts chat, file editing, terminal and Git diff review in one window.

Theo: vibe-coded a Cargo alternative: Compiles his Rust apps up to 129x faster; the point is how quickly such capabilities now feel normal.
Charlie Marsh on agentic Rust forks: Agent-built Rust forks and alternative implementations are appearing, aimed at iteration speed. His interest is in the ideas they explore and how those could evolve upstream.