OpenStack Horizon Agentic Workflows

AI-assisted workflows for OpenStack Horizon development. Currently includes a code review workflow that analyses Horizon Gerrit patches for intent correctness, plugin-API stability, test coverage, and Horizon coding conventions.

Supports Cursor, Claude Code, and the Ambient Code Platform (ACP).

---

Available Workflows

Workflow	Skill command	What it does
horizon-review	/horizon-code-review	Reviews a Horizon Gerrit patch

---

Quickstart — Cursor

Step 1 — Clone this repository

git clone https://github.com/HanzJas/openstack-horizon-agentic-workflows.git
cd openstack-horizon-agentic-workflows

Step 2 — Open the repository in Cursor

1. Open the Cursor app
2. Select File → Open Folder and choose the openstack-horizon-agentic-workflows folder you cloned

The skills and rules are auto-discovered by Cursor from .agents/skills/ and .cursor/rules/ — no additional configuration needed.

Step 3 — Run a review

1. Press Ctrl+L (Windows/Linux) or Cmd+L (Mac) to open the agent chat
2. Make sure the mode selector says Agent — if it shows Ask or Edit, click it and switch to Agent
3. Type the review command, replacing NNNNNN with the Gerrit change number:

/horizon-code-review NNNNNN

You can also pass a full URL:

/horizon-code-review https://review.opendev.org/c/openstack/horizon/+/NNNNNN

The agent fetches the patch from Gerrit, analyses it, and writes the result to artifacts/horizon-review/code-NNNNNN.md.

---

Quickstart — Claude Code

Step 1 — Clone this repository

git clone https://github.com/HanzJas/openstack-horizon-agentic-workflows.git

Step 2 — Start Claude Code inside the workflow directory

cd openstack-horizon-agentic-workflows/workflows/horizon-review
claude

Step 3 — Run a review

Inside the Claude Code session, type:

/horizon-code-review https://review.opendev.org/c/openstack/horizon/+/NNNNNN

---

Quickstart — Ambient Code Platform (ACP)

In the ACP Custom Workflow dialog, enter:

• URL: https://github.com/HanzJas/openstack-horizon-agentic-workflows.git
• Branch: main
• Path: workflows/horizon-review

Then use /horizon-code-review in the workflow chat.

---

Optional Enhancements

The workflow works out of the box with no extra setup. Two optional enhancements are available if you want deeper reviews:

Enhancement	What you gain	How to enable
Gerrit MCP Server	Agent can read prior reviewer comments and list all changes in a topic	Setup instructions below
Local Horizon clone	Agent can read surrounding source code for context on large or complex patches	Setup instructions below

Without either, the agent fetches patch content via the public Gerrit REST API and reviews from the diff alone — which is sufficient for the vast majority of patches.

---

What the Review Covers

The agent does not re-check what CI already enforces (PEP8, import order, JS lint). It focuses on things that require human judgement:

1. Intent — does the change actually do what the commit message says?
2. Plugin-API stability — do changes to horizon/ break downstream plugin projects (Octavia UI, Designate UI, Magnum UI, etc.)?
3. OpenStack API client conventions — correct client instantiation, pagination, error handling
4. Testing adequacy — regression test present for bug fixes, new code covered, JS tests where needed
5. Bug/blueprint reference — is there a linked Launchpad bug or blueprint?
6. Release notes — is a reno note needed?
7. Template/URL consistency — are context variable renames reflected in templates?

What the Review Output Looks Like

artifacts/horizon-review/code-NNNNNN.md

# Code Review: <title>

Verdict: APPROVE / REQUEST_CHANGES / COMMENT

## Summary
...

## Blockers        ← must-fix: plugin breakage, missing regression test
## Suggestions     ← non-blocking improvements
## Nits            ← minor preferences, always prefixed with "Nit:"
## Positive Feedback
## Files Reviewed

---

Optional — Gerrit MCP Server

What you gain: the agent can read prior reviewer comments (useful when a patch is on patchset 2+) and list all sibling changes in a Gerrit topic.

Without it: the agent skips review history and notes it in the output. Everything else works normally.

To enable, set up the MCP server:

1. Install

git clone https://gerrit.googlesource.com/gerrit-mcp-server /opt/gerrit-mcp-server
cd /opt/gerrit-mcp-server
./build-gerrit.sh

2. Configure

cp gerrit_mcp_server/gerrit_config.sample.json gerrit_mcp_server/gerrit_config.json

Edit gerrit_config.json with your OpenDev credentials (get your HTTP password at https://review.opendev.org/settings/#HTTPCredentials):

{
  "default_gerrit_base_url": "https://review.opendev.org/",
  "gerrit_hosts": [
    {
      "name": "OpenDev",
      "external_url": "https://review.opendev.org/",
      "authentication": {
        "type": "http_basic",
        "username": "YOUR_GERRIT_USERNAME",
        "auth_token": "YOUR_HTTP_PASSWORD"
      }
    }
  ]
}

Authentication is optional for read-only access — OpenDev allows anonymous Gerrit reads.

3. Register with Cursor or Claude Code

Cursor — add to .cursor/mcp.json in this repo (project-scoped) or ~/.cursor/mcp.json (global):

{
  "mcpServers": {
    "gerrit": {
      "command": "/opt/gerrit-mcp-server/.venv/bin/python",
      "args": ["/opt/gerrit-mcp-server/gerrit_mcp_server/main.py", "stdio"],
      "env": {
        "PYTHONPATH": "/opt/gerrit-mcp-server/"
      }
    }
  }
}

Claude Code — same JSON, add to ~/.claude.json or .claude/settings.json.

---

Optional — Local Horizon Clone

What you gain: on large or complex patches the agent can read the surrounding source code — not just the changed lines — to better judge architectural fit and local consistency.

Without it: the agent reviews from the diff alone, which is sufficient for most patches.

To enable, clone the Horizon repository alongside this one:

git clone https://opendev.org/openstack/horizon.git

The skill looks for it at /workspace/repos/horizon/ in ACP sessions, or at any path you point it to.

---

Acknowledgements

This repository is based on the architecture and design patterns from sbauza/openstack-agentic-workflows by Sylvain Bauza. The multi-tool discovery layout (AGENTS.md / CLAUDE.md / .agents/skills/ symlinks / .ambient/ambient.json), the design principles (no duplicate CI checks, human decides / agent assists), and the overall workflow structure were all taken from that project. The Horizon-specific content — skill logic, review rules, knowledge base, and agent persona — was built on top of that foundation.

---

Repository Structure

.agents/skills/
└── horizon-code-review/           → symlink for Cursor skill discovery
.cursor/
├── agents/horizon-core.md         → symlink for Cursor persona discovery
└── rules/horizon-rules.mdc        → always-applied behavioral rules
agents/
└── horizon-core.md                Horizon core reviewer persona
knowledge/
└── horizon.md                     Horizon project reference (architecture, conventions)
workflows/
└── horizon-review/
    ├── .ambient/ambient.json      ACP workflow config
    ├── .claude/skills/
    │   └── horizon-code-review/
    │       └── SKILL.md           Canonical skill definition (all tools read from here)
    ├── AGENTS.md                  Model-agnostic project reference
    ├── CLAUDE.md                  Claude Code pointer
    ├── rules.md                   Behavioral rules for the review agent
    └── README.md                  Workflow-level documentation
artifacts/                         Generated review outputs (git-ignored)