CEREBRO

Researched and written with 320,343 tokens (in 10 · out 35,878 · cache-read 66,585 · cache-create 217,870) across 5 Claude calls

Claude models & Anthropic platform

Introducing Claude Fable 5.2 and Claude Mythos 5.2: new Fable and Mythos 5.2 models announced, pitched at coding, reasoning, and vision gains.
Claude Code v2.1.321: claude-fable-5-2 becomes default Fable model, 1M context, $8/$40 per Mtok, $0.20/Mtok cache reads; adds --settings <file> to claude remote-control.
Claude Fable 5.2 in GitHub Copilot: GA in Copilot for complex multistep coding across large codebases.
elder-plinius/CL4R1T4S: claimed leak of the ~950k-char Fable 5.2 Claude Code system prompt and tools. Unverified, but useful reference for how Claude Code structures tools and behavior.
Is Opus 5.5 nerfed? Thinking patterns changed on November 3: users report degraded reasoning after Nov 3. Community correction: only the thinking summarizer changed, not the model.

v2.1.309: sandbox fixes on macOS/Linux (shell startup files ran outside sandbox; denyWrite globs missed symlinked paths). Matters if you rely on the sandbox for isolation.
v2.1.310: Monitor tool and plugin monitors now work on Bedrock, Vertex, Foundry and with telemetry off; long foreground sleeps blocked there; gateway fixes for the effort-2025-11-24 beta value.

Agent Skills donated to the Agentic AI Foundation: the standard now sits next to MCP, with 70+ platforms (ChatGPT, Cursor, Copilot, VS Code) integrated since Dec 2025.
Boris Cherny on the Skills donation: Skills began as a Claude Code experiment and moved to neutral governance after industry adoption.
MCP Tasks and Elicitations in Claude: Tasks let plugins run long jobs (e.g. deep research) and report back; a step toward async agent work.
t3dotgg/tstack: daily-driver skill set for Claude Code and Codex, including full-send, which takes a change to a merged PR unattended.
Nutlope/hallmark: anti-AI-slop design skill for Claude Code, Cursor, Codex, OpenCode; picks themes from a 50-theme catalog and lints in a real browser.

Swipe: Agents that work while you sleep: tools to keep agents running with the lid closed, keep API keys out of chat, and hand off work between parallel sessions.
Swipe: Claude in plain English: roundup with a plugin that rewrites Claudish into plain English, an account-swapping menu bar app for rate limits, local agent memory, and a skill against early quitting.
I built a Claude Code plugin with Claude Code: Sonnet builds, Opus audits, Fable only judges: tiered model-routing plugin (MIT); same pattern as build-cheap, audit-expensive pipelines.
Have your agent record video demos of its work with shot-scraper video: storyboard.yml plus Playwright records a video, giving agents a way to prove UI work.
Swipe: review-pr: review contract covering correctness, security, lifecycle, errors, tests. It explicitly excludes speculative warnings and style nits, so findings stay high-signal.
I built a Rust dev tool with Claude Code to check the tests Claude writes: tool to audit agent-written tests. Reddit post has no body, so detail is thin.

Swipe: Graft: tree-sitter symbol and call graph stored as local files, queryable for callers, references, ranked files. No model or API key needed for the structural path.
Swipe: AgentsView: imports sessions from Codex, Claude Code, Cursor, and 20+ agents into local SQLite, with search and cross-agent token/cost reports.
Swipe: mq: jq for Markdown; pull the smallest complete section instead of feeding whole docs to the model, a direct token saver.
ProtoMap: pitched as stopping agents from burning tokens rediscovering dependencies. Same goal as Graft.

Swipe: Deployment made simple: CLI with agent skills for deploying to a VPS from a coding agent, as a Vercel/Cloudflare alternative.
Swipe: an agent where your mouse is: AtAt runs an agent inline in any Mac text field via @@.
Swipe: Search the work your AI agents already did: issue on local search across agent sessions, a codebase map, and a private inbox for agents.