/local-models has always probed your machine and recommended a fit local model. Until now it stopped there — even when Ollama and a fit model were already running, the operator still had to hand-wire endpoints, .env, and Aider config themselves. local_models_autowire.py closes that gap.

/review has always reviewed exactly one plan per invocation — deliberately. A backlog of several review-needed/ plans still meant re-typing /review between each one. --batch [N] removes that friction without touching the one thing that actually matters: every plan still gets its own critic pass and its own required sign-off.

The Aider drone brief (AIDER.md) and its skills index are sized for a mid-class model. A ≤8B local model — Qwen3 4B in particular, Anchor's default swarm-local pick — doesn't have the context budget to reason about either one reliably: a live canary from a 2026-09-08 session shows a plain hello sending a 4B model into a 600-second confusion loop, treating Aider's own edit-format rules as the task. anchor --platform aider now also scaffolds a thinner, script-first card built for exactly that class of model — and along the way, live testing turned up a real bug in how the whole Aider platform is invoked.

AIDER.md — the drone brief anchor --platform aider scaffolds into every Aider project — has told operators since it shipped to type a literal /work to start the drone procedure. That instruction was wrong: Aider intercepts any chat message starting with / as one of its own built-in commands and replies Invalid command: work without ever calling the model. A bare /work never reached the LLM, on any model size.

On WSL2 with Ollama running on the Windows side, a plain aider invocation (model already set in .aider.conf.yml) failed with OllamaError: … Connection refused even though the model server was healthy. scripts/write_aider_env.py fixes it by writing the endpoint where Aider actually looks.

/work has always been able to run in a dedicated per-agent worktree — nothing stopped two agents from working the backlog at the same time. What was missing was picking a batch that's actually safe to run in parallel: nothing checked whether two plans were about to rewrite the same file. /work --batch [N] closes that gap.

A large .plans/ backlog has no notion of which pieces of work ship together. /work picks the highest-priority plan and nothing else, and there's no signal for "this cluster of plans is done — go ship it." Plans can now point at a named release, and /work//review notice when one is fully done.

Yesterday's Aider platform scaffold shipped AIDER.md — a brief that teaches a drone session to run /work and nothing else. It didn't say much about why nothing else: an operator typing /draft or /review at a drone just got a one-line redirect buried in prose. anchor --platform aider now also writes a small skills-index.md that makes that roster explicit.

Anchor's scaffold has covered Claude Code and Grok Build for a while — both harnesses with a real slash-command runtime, so a scaffolded project gets the full skill suite (/work, /draft, /review, and the rest). Aider is different: it's usually the harness running a small local model, and it has no slash-command runtime to hook into. anchor --platform aider now scaffolds it anyway, with a brief scoped to what that actually calls for.

Local Ollama/llama.cpp is the right default for offline work and tiny swarm edits, but it's often the slowest option on a laptop. Plenty of operators already have an NVIDIA Build key sitting in their environment for something else — Anchor now checks for it and prefers it automatically when it's actually live, without touching an operator's existing choice otherwise.