smooth-ai-review

AI Review Report

TL;DR

Automated, AI-driven pull-request code review. A GitHub Actions gate diffs each PR, splits the changes into context-aware chunks, and runs them through the OpenCode CLI — the provider-agnostic model transport — which calls the configured LLM at whatever endpoint the selected provider points to: a LiteLLM proxy, a provider's native API (Google Gemini, OpenAI, GitHub Copilot), or OpenCode's own gateway (OpenCode Go). The gate then posts one consolidated review back to the PR — an executive summary plus collapsible per-chunk detail, with findings categorized by priority (Critical / High / Medium / Low). Runs automatically on PRs and on demand via /ai-review.

Two skills back it:

• ai-review-report — generates the review (the CI gate; also runnable locally).
• ai-review — consumes a posted review and applies fix/skip decisions (/ai-review).

Implementation details and decisions live in .agents/skills/ai-review-report/SKILL.md.

Four ways to consume this repo

Channel	What you get	Best for
Reusable workflow	The CI gate via a ~80-line caller workflow; scripts fetched at run time, version-pinned	Repos that want the gate with minimal footprint and easy upgrades (@v1)
Claude Code plugin	The three skills (ai-review-report, ai-review, git-commit-review-push) inside Claude Code — not the CI gate	Developers who want /ai-review and the local review tooling without touching the repo
opencode plugin (npm)	The same three skills for opencode users — linked into .agents/skills/ at startup, nothing vendored (GitHub Packages registry: needs a one-time read:packages PAT per developer)	Repos/developers driving the skills from opencode instead of Claude Code
Copy-install	Workflow + skills copied into the repo; everything editable in place	Repos that customize the gate or vendor everything

The channels coexist: a repo can use the reusable workflow for CI while developers install the plugin for /ai-review. To set up a repo end-to-end (gate + skills + credentials), follow Install into another repo (AI-agent driven) — its default path is the reusable workflow + the plugin at project scope.

Use as a reusable workflow

Instead of copy-installing the 1,400-line gate, call it as a reusable workflow. Copy .docs/examples/code-review-caller.yml into your repo as .github/workflows/pipeline-code-review-report.yml — it carries the triggers, permissions, concurrency, and the model_preset dispatch dropdown, and delegates the job:

jobs:
  review:
    uses: generic-automation-and-it/smooth-ai-report-review/.github/workflows/pipeline-code-review-report.yml@v1
    with:
      pr_number: ${{ inputs.pr_number || '' }}
      model: ${{ inputs.model || '' }}
      model_preset: ${{ inputs.model_preset || '(repository default)' }}
    secrets: inherit

How it works:

• Scripts are fetched, not installed. The called workflow detects that your repo has no ai-review-report skill and checks out this repo into a .smooth-ai-review-tools/ side path, locked to the same ref the workflow was called at (override with the tools_ref input). If your repo does have the skill installed (copy-install), the local copy wins — no fetch.
• Secrets: pass secrets: inherit. The gate reads the canonical OPENCODE_*_API_KEY names and only uses the selected provider's key.
• Variables: vars.OPENCODE_REVIEW_REPORT_* resolve against your repo/org automatically — configure them exactly as in GitHub configuration; Steps 3–4 of the installer section apply unchanged.
• Inputs: runner (default ubuntu-latest; set self-hosted for private-network gateways), tools_ref, mandatory_context_files / agents_md_exempt_paths (override the context lists without editing any workflow), plus the dispatch passthroughs pr_number / model / model_preset.
• Versioning: pin @v1 (floating major) or an exact tag/SHA. The model_preset dropdown options in your caller must match the preset mapping in the called workflow — when a release adds presets, update your caller to expose them.

Install as a Claude Code plugin

The three skills — ai-review-report (review generator + local driver), ai-review (/ai-review fix/skip executor), and git-commit-review-push — are packaged as the Claude Code plugin smooth-ai-review, with this repo doubling as its marketplace:

/plugin marketplace add generic-automation-and-it/smooth-ai-report-review
/plugin install smooth-ai-review@smooth-ai-report-review