Memory in Grok Build, Zen vs third-party harnesses, a completely different Jev, and Geiger.

Geiger for finding AI agents
https://atomburst.io/geiger
If, like me, you're constantly testing agents, one command npx geiger-scan shows what's actually installed and what it can touch. Zero dependencies, no account and no telemetry, you only need Node 18+. It only reads configs and folders, it doesn't launch anything it finds. It only writes a report file you explicitly request via --json / --html. Finds many known agents and IDEs. Separately, global and project MCP servers including remote ones, editor and browser extensions, AI browsers.

Memory in Grok Build
https://x.ai/news/grok-build-memory
Grok Build now remembers the project between sessions: conventions, decisions, facts. After each step it reviews it in the background and appends only what's durable, while work doesn't stop. These are markdown files, one file per topic. There's a project level and a separate global one for user habits.

The /memory command opens a simple grouped file view with search and preview, /dream organizes fresh stuff by topic and sometimes cleans up in the background on its own. Background Dream cleanup runs single-threaded via dream_lock, old stuff goes to archive.

OpenCode Zen cuts off third-party harnesses
https://github.com/can1357/oh-my-pi/issues/12306
Similar to the Antigravity story from last week (by the way OMP fixed it with a separate plugin), here the gateway cuts out outsiders. Without the x-opencode-client, x-opencode-session, x-opencode-project, x-opencode-request headers and tool profiles, Zen, especially for free models, responds with 429 FreeUsageLimitError or just hangs. Pi extensions now imitate these headers with random ids to look like the CLI. In the threads this is called identity spoofing with a risk of a full account ban.

A completely different model — Jev / System One from TypeSafe
https://console.typesafe.ai
https://typesafe.ai/blog/introducing-system-one-models-and-jev
The first Jev model is named after Jevons. This is not chat, but System One after Kahneman: state in, ready-made decisions for code out.
Send state as text, JSON or array plus a list of typed questions Choice, Score, Noul — and in a single parallel pass you get all answers with calibrated probabilities.

New architecture, parallel sampler, RLCD training instead of RLHF / RLVR. The schema is fixed upfront, so there's simply nowhere to go outside the types or hallucinate text. Access is currently waitlist-only: endpoint POST api.typesafe.ai/v1/systemone, jev-latest is currently jev-1.13.0, plus jev-preview next to it, SDK for Python and JS, skill for Claude Code, key at console.typesafe.ai/settings/keys.

#grokbuild #opencode #piagent #typesafe #mcp