用户用Claude Code规划并由本地Ornith执行
一名Reddit用户公开了一套方法:先用Claude Code编写规格说明和小型任务卡,再交给本地运行的8位量化Ornith 1.5-9B执行。在该工作流中,本地模型负责运行命令、构建、轮询任务和Git操作,较强模型在合并前核验结果。作者称最终在两个代码仓库完成了75次提交且回归测试全部通过,但也承认模型对自身完成情况的记录曾多次出错。
这则经验展示了一种明确分工:由Claude Code生成小模型可执行的计划,让Ornith 1.5-9B在本地硬件上完成令牌消耗较多的实施工作。
I've been running Ornith 1.5-9B (8-bit) as a local coding agent on real engineering work. It can do the work, but only when the plan is written for a small model. I wrote down what I learned into a public file meant to be handed to a stronger planning model (I use Claude Code):
The split
- The stronger model writes the spec and a plan of small task cards, then verifies the result before merge.
- The local model does the long, mechanical middle: running commands, builds, polling jobs, git work. That's where most of the tokens go, and it runs on my own hardware.
What the file covers
- What the small model is good at, and what to expect (work more reliable than its reports; doesn't explore on its own; doesn't know when it's done; slows down at long context).
- How to brief it: exact files and line ranges, absolute paths, 1–3 line edits, exact function signatures, every rule written down, every step ending in a command with its exact expected output, explicit stop rules.
- How to run it: one task per fresh session, small context, the plan file as memory.
- Before anything lands: verify claims against git and logs, not the model's summary.
Results so far: a 75-commit merge across two repos with every regression test passing, done by the 9B. Its notes about what it did were wrong more than once; the plan's checks and a final verification caught every one.
Limitations: slower (most of a day vs about an hour for Claude alone on that phase), and you need hardware that runs a decent local model. The rules are harness-agnostic but were learned with one model; I'd like to hear how they hold up with others.