Grok Build might just be the best local AI harness

Grok Build, once modified for local models and other LLMs, is OpenGrok, the coding-agent client I actually use. I wanted one place to set which model gets the turn, when context gets compacted, and how hard local subagents are allowed to hit one Mac.

This is post 0 of 4 on the Home Base series. Next: Mac hub, Flash Next MTP vs 27B+DFlash, speculative decode / parent+1.

What it is

OpenGrok runs on gHost64 (and on ghost128). For local coding it points at the Home Base brain over Tailscale (homebase-brain), not straight at a raw model server. The brain owns routing, memory, and compaction; the client owns the session UI and the knobs below.

This series stays on what the stack actually uses. Soft note: workshop / build-in-public material for destroyrebuild.xyz.

Features this series actually uses

  • Model picker / Home Base route: coding traffic goes through the brain; explore/plan can pin elsewhere (I pin those to gpt-5.6-luna).
  • Auto-compact at 40%: auto_compact_threshold_percent = 40 hits the brain POST /v1/compact (Jev keep/drop). Per-turn chat_compact stays off on purpose so prefix cache stays warm.
  • Subagent settings: max_concurrent, sampling_limit, max_depth, limit_behavior, plus which models explore/plan use.
  • cancel_subagents_on_turn_cancel: cancel the turn, cancel the kids.

Those concurrency knobs are why post 3 exists. Without them, four fat local streams looked hung.

Why a client I own

Stock clients were fine until I needed one place to aim every coding turn at a local hub, compact through Jev instead of every message, and fail closed when fan-out would thrash a single GPU. OpenGrok is that client. Home Base is the Mac hub behind it.

Series

All ship logs