autoquality-claw

autoquality-claw is an OpenClaw plugin that intercepts the final reply dispatch, reviews the turn's code edits against a workspace review-rules.md, and can run a deterministic review/fix loop before the user ever sees the final answer.

It is built for teams that want a configurable post-turn quality gate without hardcoding project rules into the plugin itself. The rules live in markdown in the workspace, so changing the bar is a repo edit, not a plugin release.

What It Does

• Reads review-rules.md from the workspace on every handled reply.
• Tracks mutating tool calls across the root run and spawned subagents.
• Uses a gatekeeper LLM to decide whether to approve, review, fix, or stop.
• Dispatches reviewer and fixer subagents when needed.
• Delivers the final user reply only after the loop terminates.
• Appends a compact autoquality-claw summary when review activity occurred or the loop hit a limit.

LLM Provider

autoquality-claw tries to inherit the host's LLM first. If OpenClaw exposes runtime.getGatekeeperLlm() — typical when OpenClaw is already configured for Bedrock, Mantle, OpenAI, or any other provider — the plugin uses that client directly. No plugin-side credentials needed.

Only when the host does not expose a client does the plugin fall back to constructing its own Anthropic SDK client from the env var named by anthropicApiKeyEnv (default ANTHROPIC_API_KEY). The fallback exists so the plugin works standalone; in any real OpenClaw install, provider inheritance is the intended path.

Install

Option A — via OpenClaw plugin CLI (recommended)

From npm:

openclaw plugins install autoquality-claw
openclaw plugins enable autoquality-claw

From a local clone (dev / link mode):

git clone https://github.com/inceptionstack/autoquality-claw-plugin.git
cd autoquality-claw-plugin && npm install && npm run build
openclaw plugins install -l .
openclaw plugins enable autoquality-claw

Option B — plain npm (if you're wiring a plugin loader yourself)

npm install autoquality-claw

The published tarball ships dist/, the manifest, and examples/. No build step on the consumer side.

Option C — from source

git clone https://github.com/inceptionstack/autoquality-claw-plugin.git
cd autoquality-claw-plugin
npm install
npm run build

Enable In OpenClaw

Point OpenClaw at the installed package or the cloned directory using your OpenClaw plugin-loading config. The manifest file is openclaw.plugin.json; the runtime entry is dist/plugin-entry.js.

Example OpenClaw config section:

{
  "plugins": {
    "entries": {
      "autoquality-claw": {
        "enabled": true,
        "config": {
          "enabled": true,
          "rulesPath": "review-rules.md",
          "gatekeeperModel": "claude-haiku-4-5-20251001",
          "defaultReviewerModel": "claude-sonnet-4-6",
          "defaultFixerModel": "claude-sonnet-4-6",
          "reviewerAgentId": "code-reviewer",
          "fixerAgentId": "coder",
          "maxIterations": 4,
          "loopTimeoutSeconds": 600,
          "subagentRunTimeoutSeconds": 180,
          "emitLivenessUpdates": true,
          "mutatingTools": ["edit", "write", "apply_patch"]
        }
      }
    }
  }
}

When the host provides its own LLM, you do not need to set anthropicApiKeyEnv — it is only read on the fallback path.

Configuration

| Key | Purpose | Default | | --- | --- | --- | | enabled | Turns the plugin on/off without uninstalling it. | true | | rulesPath | Workspace-relative path to the markdown rules file. | review-rules.md | | anthropicApiKeyEnv | Fallback env var name for direct Anthropic SDK when the host does not expose an LLM. | ANTHROPIC_API_KEY | | gatekeeperModel | Model used for the decision step. | claude-haiku-4-5-20251001 | | defaultReviewerModel | Default reviewer model when the gatekeeper does not override it. | claude-sonnet-4-6 | | defaultFixerModel | Default fixer model when the gatekeeper does not override it. | claude-sonnet-4-6 | | reviewerAgentId | Default reviewer agent profile. | code-reviewer | | fixerAgentId | Default fixer agent profile. | coder | | maxIterations | Hard cap for loop iterations at the handler level. | 4 | | loopTimeoutSeconds | Total loop timeout. | 600 | | subagentRunTimeoutSeconds | Timeout per reviewer/fixer run. | 180 | | emitLivenessUpdates | Emits block replies such as autoquality-claw: iteration 2. | true | | mutatingTools | Tools whose successful calls count as edits. | ["edit","write","apply_patch"] |

review-rules.md

Ship examples/review-rules.md into the root of the target workspace and adapt it. Minimal example:

---
minIterations: 1
maxIterations: 4
qualityGate: "reviewer reports zero error-severity issues AND every behavior change is covered by a test"
---

# Project quality rules

## When to trigger a review
- Any edit to files under `src/`

## When to skip review
- Docs-only changes

## Reviewer instructions
Check correctness, tests, and boundary safety.

## Fixer instructions
Fix error-severity findings first and add tests for behavior changes.

Development

npm install
npm run build
npm test
npx tsc --noEmit

CI/CD

Every push and PR runs the GitHub Actions workflow in .github/workflows/ci.yml: install → typecheck → test → build. Release tags (v*) additionally publish the tarball as a GitHub Release asset via .github/workflows/release.yml.

License

MIT — see LICENSE.