qwen-project-autopilot (local 1→done)
The Qwen-run version of /project-autopilot. No lens diagnosis, no operator psychology: just a tight, gated bead loop. The gates are the safety net.
QWEN LOCAL RULES (full text: ../qwen-forge/references/QWEN-MODE.md)
Requires qwen-forge installed alongside this skill (it holds the shared rules).
- Builder, not architect. Judgment calls belong to the user.
- ONE change at a time; build → tests → lint after each; green → commit; red → revert, one retry, else STOP. Never weaken tests.
- Build ONLY what the Intent needs; extras → SOMEDAY.md.
- Unclear / judgment / outward → DECISIONS.md + STOP.
- Local endpoint required: a local OpenAI-compatible endpoint (vLLM, SGLang, llama.cpp or Ollama) serving a capable Qwen-class coding model. ONE model loaded, work SEQUENTIALLY, leave memory headroom on unified-memory machines.
Before the loop
- Confirm the baseline is green (build + tests). If red → fixing the baseline is the first task; do that under the loop's gates before anything else.
THE LOOP (repeat)
a. `br ready` → take the TOP task. If none → go to "When the beads drain".
b. Make the SMALLEST change that completes ONLY that task. Touch as little as possible.
c. Run build → tests → lint.
green → commit (short msg), `br update <id> --status closed`, report one line.
red → `git restore` the change; try ONE more fix; still red → STOP and ask.
d. If the task needs a design / naming / scope / taste decision → STOP, write it to
.autopilot/DECISIONS.md, move to the next ready task.
e. Go to (a).
When the beads drain (br ready empty)
1. Run qwen-bug-zero (find + fix real bugs).
2. If the Intent cares about speed AND features are settled → run qwen-perf-max.
3. Report: "intent features built, bugs swept" + hand to the user for ship (qwen-ship-it).
Rules that keep it on track
- Never widen scope. New ideas that come up → SOMEDAY.md, not new beads, unless the user says so.
- Small batches only: one bead, one verify, one commit. The green ratchet is the whole safety system here (there is no smart-model review on a local-only run).
Distilled from /project-autopilot. Its references hold the full lens/routing/operator-moves version for a Claude or human driver.