Plan with Claude Code and let a local 9B model do the engineering

The poster runs Ornith 1.5-9B at 8-bit on their own hardware as a coding agent on real engineering work. Claude Code acts as the planner: it writes the spec and a plan broken into small task cards, and verifies the result before merge. The local model does the long mechanical middle, which is where most tokens go: running commands and builds, polling jobs and doing git work. Each card names exact files and line ranges, absolute paths, one to three line edits, exact function signatures, every rule, a command with its expected output at the end of each step, and explicit stop rules. The small model runs one task per fresh session with the plan file as its memory. The poster published these rules as a briefing file to hand to the planning model.

The 9B carried out a 75-commit merge across two repositories and every regression test passed. Its own notes on what it had done were wrong more than once, and the checks written into the plan plus a final verification caught each case. It was much slower: most of a day, against about an hour for Claude alone on the same phase. The poster says it cut their Claude token spend.

Trust the repo, not the small model's report: verify every claim against git and logs. The poster found the 9B does not explore on its own, does not know when it is done and slows at long context, so the plan has to carry all of that. The rules were learned with one model, and the local harness was not named.