Claude Code's Claude Mods let plugins rewire deeper agent behavior
Version 2.1.287 also adds a side agent that flags your misses and restores rm safeguards; Codex tightens its reviewer and Ember-1 undercuts Kimi K3 on tokens.

Copy markdown
Claude Mods go deeper than hooks
Claude Code 2.1.287, shipped Oct 1, introduces Claude Mods — plugins that can modify deeper agent behavior rather than just fire hooks around it. If you've wanted to reshape how the agent plans, reviews, or responds instead of bolting on side effects, this is the extension point that was missing.
A built-in watchdog for your misses
The new "You should know" mod runs a side agent that watches your session and flags things you or Claude might overlook. Turn it on with `/plugin enable cc-plugin-you-should-know@builtin` (first-party sessions with telemetry on) for a free second set of eyes on every turn.
rm safeguards are back, plus MCP sign-in prompts
2.1.287 also fixes a regression where dangerous rm commands lost their safeguards with wildcard or tilde redirects — a must-update if you let agents run Bash. It adds support for MCP URL prompts on the 2025-11-25 protocol, so servers can drive sign-in flows inline.
Codex CLI 0.160.0 stops losing the thread
OpenAI's Codex CLI 0.160.0, also Oct 1, makes Guardian review pull in your earlier instructions and context from agent handoffs, and its terminal UI now remembers provider, reasoning-summary, and verbosity settings between runs. Fewer re-explanations, steadier reviews.
Ember-1: Kimi K3 quality for ~40% fewer tokens
Fireworks Research's Ember-1, live on Fireworks Serverless as a research preview, reportedly matches Kimi K3's quality while using about 40% fewer tokens — 71.3% less reasoning-token spend in their testing — scoring 82.0% on Terminal Bench 2.1 and 75.2% on DeepSWE 1.1. If token cost caps your agent loops, it's worth a benchmark.
LTX-2.5 trends: open-weight video with synced audio
Lightricks' LTX-2.5 is trending on Hugging Face — an open-weights world model that generates video with synchronized audio from text, image, or video, with native multishot scenes and low-VRAM quantization for local runs. A weekend build: a local prompt-to-scene pipeline, free under $10M revenue.