A thinner drone card for Qwen3 on Aider — and a real invocation bug

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.

A ≤1-screen card, not a bigger prompt

platforms/aider/work-drone.md is 68 lines: pick and claim a ready plan by running work_once.py/plan_fit.py (never by reasoning about .plans/ lanes by hand), execute the claimed plan's Steps one row at a time, finish to .plans/review-needed/, never merge. It pairs with a new .aider.conf.swarm.yml.example that loads only that card — not the fuller AIDER.md brief or the skills-index roster — and trims the request itself: diff edit format instead of whole-file rewrites, a capped repo-map, a 60s per-call timeout, streaming off. Every key in that config was checked against a live aider --verbose dump on the actual installed binary (Aider 0.86.2), not assumed from memory.

The bug: Aider eats /work before the model ever sees it

Testing the new card against a real Ollama endpoint surfaced something more fundamental than a slow model: Aider treats any message starting with / as one of its own built-in commands. Type a bare /work and Aider answers Invalid command: work itself — the LLM is never called. This affects every Aider platform file shipped so far, for every model size, not just small ones; AIDER.md still tells operators to treat literal /work as the trigger. The two files this plan owns now say what actually works instead: type the word work, with no leading slash. AIDER.md itself needs the same fix as a follow-up — it wasn't in this change's scope to touch.

An honest partial

The card's mechanics — claim, list, thin context loading, correct invocation syntax — are all verified live and working. A full interactive round-trip (the model actually replying) is not: on this WSL2 + Windows-host-Ollama profile, even a trivial message without any drone-card context still failed to complete within a 60-second cap via Aider, while the same Ollama endpoint answered a comparable prompt over raw curl in under 12 seconds — litellm retries a timed-out call automatically, so a per-attempt cap doesn't bound the total wait the way it looks like it should. That gap is documented as an open question rather than papered over. "Interactive 4B stays hopeless on this class of host" was always a named risk for this work, with the fallback already built: work_once.py for unattended execution.