selfevolve-reviewer

A qwen2.5:7b-instruct variant preloaded with the system prompt and decoding settings from selfevolve — a code-review agent that learns from an engineer’s corrections and stores the lessons locally.

ollama run hoodarunner/selfevolve-reviewer

Paste in a function and it reviews it: a small number of high-signal comments, each with a concrete concern and a specific recommendation. It is told not to restate what the code does, not to flag formatting a linter already handles, and to return nothing when the code is fine.

What this model is not

It does not learn. The learning lives in the application, not in these weights. selfevolve keeps accepted, rejected and edited review comments in a local SQLite file, distils them into scoped rules, and injects the relevant ones into the next review. A model on a public registry cannot contain anyone’s accumulated judgement — and shouldn’t, since those rules are specific to one person’s codebases.

Pointing selfevolve at this model changes nothing. The agent sends its own system message on every request, which takes precedence over the SYSTEM baked into a Modelfile. This exists so the review prompt can be tried in one command, with no Python and no clone.

Why instruct rather than reasoning

Under a JSON-schema constraint a reasoning model’s thinking phase has nowhere to go. On a real review, qwen3:8b timed out at 180 seconds where qwen2.5:7b-instruct finished the same file in 22. Reasoning quality is not the bottleneck for this task: the model supplies general code knowledge, and the learned rules supply the judgement.

The full loop

retrieve lessons ─► review ─► human accepts/rejects/edits ─► reflect ─► persist
       ▲                                                                  │
       └────────────────── one SQLite file ◄──────────────────────────────┘

Runs entirely offline — one runtime dependency, local Ollama for generation and embeddings, and CI that proves the whole loop inside a network namespace with no interfaces.

Source, tests and setup →

MIT.