open-swe/evals/reviewer

Reviewer Eval

Offline LangSmith eval for the Open SWE Reviewer graph against the 50 PRs from withmartian/code-review-benchmark.

Layout

evals/reviewer/
├── golden_comments/      # 50 PRs × golden comments (copied from martian benchmark)
├── build_dataset.py      # martian JSON → LangSmith dataset (resolves SHAs via gh)
├── config.toml           # default benchmark run config
├── judge.py              # claude-opus-4-5 pairwise match evaluator + aggregate
├── target.py             # invokes the reviewer graph over langgraph_sdk
├── store_reporter.py     # publishes live progress to the dashboard store record
└── run_eval.py           # client.aevaluate entrypoint

Prerequisites

• LANGSMITH_API_KEY set in your env.
• gh authenticated (gh auth status) — needed for build_dataset.py.
• ANTHROPIC_API_KEY set — judge runs claude-opus-4-5.
• A running reviewer graph (local langgraph dev or deployed assistant id) with REVIEWER_ASSISTANT_ID env var pointing at it. Defaults to assistant reviewer on http://localhost:2024.

1. Build the dataset (once)

# Dry run — writes evals/reviewer/dataset_dryrun.json without uploading
uv run python -m evals.reviewer.build_dataset --dry-run

# Upload for real
uv run python -m evals.reviewer.build_dataset --dataset-name openswe-reviewer-v1

Each example carries: repo, pr_number, pr_url, base_sha, head_sha, base_ref, head_ref, pr_title. The dataset is frozen at upload time — upstream PR drift can't invalidate it.

2. Run the eval

The reviewer graph must be running and accept a pr input matching the example schema, and must emit a submit_review tool call (or set state["review"]["comments"]) with [{file, line, severity, body}, ...].

uv run python -m evals.reviewer.run_eval

Smoke-test with 3 PRs first:

uv run python -m evals.reviewer.run_eval --limit 3

From the GitHub Action (recommended for full runs)

Trigger the Reviewer eval workflow (.github/workflows/reviewer-eval.yml) from the Actions UI or gh workflow run reviewer-eval.yml --ref prod -f limit=3. Run it on the prod branch so the harness/judge match the deployed reviewer it scores. Running it on a durable runner (instead of inside the serving deployment) means a deploy or container recycle can't kill a long run.

The Action sets REVIEWER_EVAL_REPORT_STORE=1, so run_eval publishes live status/progress/logs to the LangGraph store record the dashboard reads — watch it at Admin → Reviewer eval (/admin/evals), which is now a read-only progress view (status, completed / total, log tail, LangSmith experiment link, and a link back to the GitHub run). If the Action is cancelled/killed, the heartbeat goes stale and the dashboard flips the run to failed within ~60s.

Required repository config:

• secrets: LANGSMITH_API_KEY, ANTHROPIC_API_KEY (the judge runs in-process; reviewer-model keys are not needed — the reviewer runs in the deployment).
• secret or var: LANGGRAPH_URL — the deployment URL the eval drives and reports to.

Tracing project

Eval traces are routed to the open-swe-evals LangSmith project (set via langsmith_project in config.toml, default open-swe-evals) so they stay out of the deployment's production tracing project. The admin-triggered run forces the same project via the LANGSMITH_PROJECT env var; override the default with EVAL_LANGSMITH_PROJECT.

The runner reads benchmark settings from evals/reviewer/config.toml. Set the deployment URL there (or leave it blank to use LANGGRAPH_URL / local dev). The target sets reviewer_eval for every run, so publish_review does not post to GitHub.

Per-repo review style prompts

At runtime the reviewer loads a custom style guide from LangGraph Store when configurable.repo is set (owner + name → store key owner/name). This applies to eval runs too, as long as a completed style profile exists for that repo.

The Martian benchmark uses these upstream repos (10 PRs each):

• getsentry/sentry
• keycloak/keycloak
• grafana/grafana
• discourse/discourse
• calcom/cal.com

Before scoring with repo-specific styles, run Review styles analysis in the dashboard for each repo (or copy prompts into store). Re-run make dev so the reviewer graph sees the same store.

By default the judge scores final add_finding calls. Set score_mode = "surfaced_findings" in the config to score only findings that would pass the production threshold/cap.

model_id and reasoning_effort in the config are passed to the reviewer run, so isolated benchmark deployments can test a specific model/effort without changing deployment-wide defaults.

Notes

• No GitHub forks needed — both upstream repos and martian's benchmark forks (ai-code-review-evaluation/*) are public.
• judge_match charges judge LLM tokens proportional to n_candidates × n_goldens per example. For 50 PRs with ~3 goldens each and agents emitting ~10 candidates, expect ~1500 judge calls per experiment.