open-swe/.github/workflows/reviewer-eval.yml

# Runs the reviewer benchmark (evals/reviewer/run_eval.py) on a durable runner instead of as a
# subprocess inside the serving deployment. Trigger it from the Actions UI / `gh workflow run`,
# selecting the `prod` branch so the harness + judge match the deployed reviewer it scores.
#
# Progress is streamed to the LangGraph store record the dashboard reads, so the run shows up live
# at /admin/evals. Required repository config:
#   secrets: LANGSMITH_API_KEY, ANTHROPIC_API_KEY   (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 + reports to)
name: Reviewer eval

on:
  workflow_dispatch:
    inputs:
      model_id:
        description: Reviewer model id
        type: string
        default: google_genai:gemini-3.5-flash
      reasoning_effort:
        description: Reasoning effort
        type: string
        default: medium
      dataset_name:
        description: LangSmith dataset
        type: string
        default: openswe-reviewer-v1
      experiment_prefix:
        description: Run name (LangSmith experiment prefix)
        type: string
        default: openswe-review-confidence
      max_concurrency:
        description: Max concurrent PRs
        type: string
        default: "5"
      score_mode:
        description: all_findings | surfaced_findings
        type: choice
        default: surfaced_findings
        options:
          - all_findings
          - surfaced_findings
      severity_threshold:
        description: Severity threshold (surfaced_findings only)
        type: choice
        default: low
        options:
          - low
          - medium
          - high
          - critical
      cap:
        description: Max surfaced findings per PR (surfaced_findings only)
        type: string
        default: "6"
      limit:
        description: Run only the first N examples (blank = full dataset)
        type: string
        default: ""
      langsmith_project:
        description: LangSmith tracing project for eval traces
        type: string
        default: open-swe-evals
      assistant_id:
        description: Reviewer assistant id
        type: string
        default: reviewer

concurrency:
  group: reviewer-eval
  cancel-in-progress: false

permissions:
  contents: read

jobs:
  reviewer-eval:
    name: Reviewer eval
    runs-on: ubuntu-latest
    timeout-minutes: 360
    steps:
      - uses: actions/checkout@v7
      - uses: astral-sh/setup-uv@c771a70e6277c0a99b617c7a806ffedaca235ff9 # v9.0.0
      - name: Install dependencies
        run: uv sync --locked
      - name: Run reviewer eval
        # Inputs are passed via env and referenced as quoted "$VARS" — never
        # interpolated into the script — so dispatcher-supplied text is treated
        # as data, not shell syntax.
        env:
          LANGSMITH_API_KEY: ${{ secrets.LANGSMITH_API_KEY }}
          LANGCHAIN_API_KEY: ${{ secrets.LANGSMITH_API_KEY }}
          ANTHROPIC_API_KEY: ${{ secrets.ANTHROPIC_API_KEY }}
          LANGGRAPH_URL: ${{ secrets.LANGGRAPH_URL || vars.LANGGRAPH_URL }}
          REVIEWER_EVAL_REPORT_STORE: "1"
          INPUT_MODEL_ID: ${{ inputs.model_id }}
          INPUT_REASONING_EFFORT: ${{ inputs.reasoning_effort }}
          INPUT_DATASET_NAME: ${{ inputs.dataset_name }}
          INPUT_EXPERIMENT_PREFIX: ${{ inputs.experiment_prefix }}
          INPUT_MAX_CONCURRENCY: ${{ inputs.max_concurrency }}
          INPUT_SCORE_MODE: ${{ inputs.score_mode }}
          INPUT_SEVERITY_THRESHOLD: ${{ inputs.severity_threshold }}
          INPUT_CAP: ${{ inputs.cap }}
          INPUT_LIMIT: ${{ inputs.limit }}
          INPUT_LANGSMITH_PROJECT: ${{ inputs.langsmith_project }}
          INPUT_ASSISTANT_ID: ${{ inputs.assistant_id }}
        run: |
          set -euo pipefail
          limit_args=()
          if [ -n "${INPUT_LIMIT}" ]; then
            if ! [[ "${INPUT_LIMIT}" =~ ^[0-9]+$ ]]; then
              echo "limit must be a positive integer, got: ${INPUT_LIMIT}" >&2
              exit 1
            fi
            limit_args=(--limit "${INPUT_LIMIT}")
          fi
          uv run python -m evals.reviewer.run_eval \
            --model-id "${INPUT_MODEL_ID}" \
            --reasoning-effort "${INPUT_REASONING_EFFORT}" \
            --dataset-name "${INPUT_DATASET_NAME}" \
            --experiment-prefix "${INPUT_EXPERIMENT_PREFIX}" \
            --max-concurrency "${INPUT_MAX_CONCURRENCY}" \
            --score-mode "${INPUT_SCORE_MODE}" \
            --severity-threshold "${INPUT_SEVERITY_THRESHOLD}" \
            --cap "${INPUT_CAP}" \
            --langsmith-project "${INPUT_LANGSMITH_PROJECT}" \
            --assistant-id "${INPUT_ASSISTANT_ID}" \
            "${limit_args[@]}"