Proposal: label-gated OpenCode Go PR reviewer

[simlal]

I've setup agentic PR review using OpenCode for work and I like it. I’d like to propose a similar approach for OpenCode-based PR review agent.

Goal

Leverage the existing AGENTS.md to make an initial review for PRs

Initial Scope

• Automatically review opened and updated PRs from branches in this repository.
• Use a basic prompt in the workflow, fine tune as you go
• Use a dedicated OpenCode Go key stored as an Actions secret (I think you should do org secret so you can enable to other ublue repos if you like!).

External-fork PRs are excluded initially. This avoids exposing the provider key to untrusted code and does not require pull_request_target (hello supply chain attack).

Authentication

OpenCode Go uses an API key. The OpenCode GitHub App is very convinient, otherwise a minimally scoped GITHUB_TOKEN can post review comments instead.

Cost for OSS
I know the Go Sub is not free, but using something like DeepSeekV4-pro you could easily cover hundreds of PRs and stay within the 5hours/weekly/monthly usage windows for 10$/month. Maybe you could even use the free-tier Zen's API key + the Big Pickle model as a starting point?

a opencode-review.yml workflow could look like:

```yaml
name: opencode-review

on:
  pull_request:
    types: [opened, reopened, synchronize, ready_for_review]
    branches: [main]

jobs:
  review:
    if: |
      github.event.pull_request.draft == false &&
      github.event.pull_request.head.repo.full_name == github.repository
    runs-on: ubuntu-latest
    timeout-minutes: 30
    permissions:
      contents: read
      pull-requests: write
      issues: write
    steps:
      - uses: actions/checkout@latest # or commit sha if you want
        with:
          fetch-depth: 1
          persist-credentials: false

      - name: Preflight # Optional but I like if someone messes with secrets/vars
        env:
          OPENCODE_MODEL: ${{ vars.OPENCODE_MODEL }}
          OPENCODE_API_KEY: ${{ secrets.OPENCODE_GO_API_KEY }}
        run: |
          failed=0

          if [ -z "$OPENCODE_MODEL" ]; then
            echo "::error::Required org variable 'OPENCODE_MODEL' is not set."
            failed=1
          fi

          if [ -z "$OPENCODE_API_KEY" ]; then
            echo "::error::Required secret 'OPENCODE_GO_API_KEY' is not set."
            failed=1
          fi

          exit "$failed"

      - uses: anomalyco/opencode/github@latest # or commit sha if you want
        env:
          OPENCODE_API_KEY: ${{ secrets.OPENCODE_GO_API_KEY }}
          GITHUB_TOKEN: ${{ github.token }}
          OPENCODE_PERMISSION: '{"bash":{"*":"deny","gh api *":"allow"}}' # adjust to your liking, sometimes agent will use /tmp so I like to provide '{"external_directory":"allow"}' 
        with:
          model: ${{ vars.OPENCODE_MODEL }}
          use_github_token: true
          share: false
          prompt: | # Adjust to your liking, this is just a basic prompt
            Review pull request #${{ github.event.pull_request.number }} on $GITHUB_REPOSITORY.

            Focus on:
            - Security regressions, secret exposure, and unsafe CI changes.
            - Shell-script correctness, quoting, error handling, and idempotency.
            - Containerfile and image-build regressions.
            - Package-source and repository-security concerns.
            - GitHub Actions permissions, action pinning, and supply-chain risks.
            - Missing or incorrect validation for modified build logic.

            Only report actionable findings introduced by this PR. Do not report style-only
            preferences unless they can cause a reliability, maintainability, or security issue.

            Post a formal GitHub review using gh api:
            - CRITICAL or WARNING findings: REQUEST_CHANGES
            - Suggestions only or zero findings: COMMENT
            - Never APPROVE

            Include:
            ## Review Summary
            <1-2 sentences>

            ## Findings
            <findings or "No actionable findings.">

            ## Residual Risks
            <risks or "None identified.">

            Add line comments only when a precise affected line exists.