guildlm/go-review

GuildLM · go-review

*A small, sharp Go code-review specialist from the GuildLM Code Guild.*

go-review reads Go and hunts for the bugs a green build hides — correctness errors, race conditions, and un-idiomatic patterns. It is one of three specialists in the GuildLM Code Guild (`go-dev` · `go-test` · go-review) built to work together in a verification-driven agent loop.

Why a reviewer in the loop? A passing test suite proves the cases you thought of. go-review is the adversary that asks what you didn't — the off-by-one inside a branch no test exercises, the data race that only shows under -race, the error that's swallowed. Inside the GuildLM Builder it runs as a non-regressing pass: a fix it suggests is kept only if the project is still green afterward.

Why this isn't "just Qwen with a name"

go-review is a fused, standalone model (no separate adapter) with its own identity — ask who it is and it answers GuildLM go-review. It is an honest Apache-2.0 derivative of Qwen2.5-Coder-7B-Instruct, fine-tuned for one craft: reviewing Go.

What it's for

• —Reviewing diffs and files for real bugs: nil derefs, index errors, unhandled errors, races, deadlocks.
• —Flagging un-idiomatic Go (ignored errors, needless allocations, non-stdlib reflexes).
• —Acting as the review role inside the Builder, catching what a green build and the tests both missed.

Benchmarks

Measured locally with the real Go toolchain.

<!-- BENCH:go-review --> | Benchmark | Metric | go-review | base 7B | |---|---|---|---| | crucible go_review_bench (8 planted bugs) | identify@1 | 6/8 | 7/8 |

Honest note from the [research log](https://guildlm.github.io/research/): for review, the base is already strong (7/8) and per-role fine-tuning is within noise — we don't pretend otherwise. go-review's value is structural, not a benchmark number: it's a distinct second opinion in the loop, prompted only to find faults, separate from the model that wrote the code. That separation — a non-regressing review pass after a green build — catches bugs a single self-reviewing model rationalizes past. Use it as the third member of the guild, not as a solo linter.

Quickstart

Apple Silicon (MLX)

pip install mlx-lm
python -m mlx_lm generate --model guildlm/go-review \
  --prompt "Review this Go for bugs:\n\nfunc Sum(xs []int) int { s := 1; for _, x := range xs { s += x }; return s }" \
  --max-tokens 300

Ollama (GGUF)

ollama run guildlm/go-review "Review this handler for race conditions and unhandled errors: ..."

Inside the agent loop (recommended)

python -m mlx_lm server --model guildlm/go-review --port 8082
guildlm-build --spec specs/myservice.yaml --out ./out \
  --base-url http://localhost:8080/v1 \
  --test-model guildlm/go-test \
  --review-model guildlm/go-review --review-base-url http://localhost:8082/v1

Prompting

Trained with the system prompt:

You are GuildLM go-review, a Go code-review specialist from the GuildLM Code Guild.

Give it Go code and ask what's wrong. It reports concrete, located findings rather than vague praise.

The Guild

• —Agent loop: https://github.com/guildlm/builder
• —Research log: https://guildlm.github.io/research/

License & attribution

Apache-2.0, inherited from Qwen2.5-Coder-7B-Instruct (© Alibaba Cloud). GuildLM fine-tuning, identity, and packaging under the same license. Trained locally on Apple Silicon with MLX — total cloud spend: $0.