ProbeLabs Assistant

This is the AI assistant that ProbeLabs runs internally. It's how we use Probe and Visor ourselves — a real-world, production implementation that covers our entire codebase of 16 repositories spanning Rust, TypeScript, Go, and Node.js.

We open-sourced it so you can see exactly how a Probe-based assistant works at scale: how skills are structured, how code-talk explores multiple repos, how GitHub integration feeds context into conversations, and how the engineer workflow turns questions into pull requests.

Want to build your own? Use probe-quickstart — a minimal template you can clone and customize in minutes. Come back here when you want to see how a full production setup looks.

What We Use It For

• Code exploration — our team asks questions about implementation details across all ProbeLabs projects and gets answers with file paths and line numbers
• GitHub browsing — check PRs, issues, releases, and activity across the probelabs org without leaving Slack
• CI/CD debugging — paste a failed pipeline link and get a root cause analysis with the relevant code context
• Code changes — describe what needs to change and the assistant explores the code, implements the fix, and opens a PR
• Thread summaries — catch up on long Slack discussions

Running It

CLI

git clone https://github.com/probelabs/probelabs-assistant.git
cd probelabs-assistant
cp .env.example .env
# Edit .env with your API keys

npx -y @probelabs/visor@latest run probelabs-assistant.yaml --tui

Slack

npx -y @probelabs/visor@latest run probelabs-assistant.yaml --slack

Requires Slack tokens in .env — see Environment Variables below.

How It Works

A message goes through this pipeline:

1. Intent classification — is this chat or code_help?
2. Skill selection — which skills match the request?
3. Dependency expansion — engineer requires code-explorer, so both activate
4. Knowledge injection — each active skill injects its docs and tools into the AI context
5. Execution — the AI responds, calling code-talk, gh CLI, or Claude Code as needed

Everything is declared in YAML. No application code.

Project Structure

probelabs-assistant.yaml          ← Main config: Slack, scheduler, system prompt
│
├── config/
│   ├── intents.yaml              ← Intent routing (chat, code_help)
│   └── skills.yaml               ← All skills with knowledge + tools
│
├── docs/                         ← Knowledge files injected into skills at runtime
│   ├── probelabs-architecture.md ← How our products fit together
│   ├── github-tools.md           ← gh CLI patterns for all repos
│   ├── github-actions-debugging.md
│   ├── code-exploration-tool.md
│   ├── code-help-mode.md
│   ├── engineer-tool.md
│   └── capabilities.md
│
└── workflows/
    └── engineer.yaml             ← Code changes + PR creation via Claude Code

Skills

Skill	When It Activates	What It Does
github	PRs, issues, releases, branches	Runs gh CLI commands across all ProbeLabs repos
github-actions	CI/CD failures, pipeline debugging	Fetches logs, greps for errors, identifies root cause
code-explorer	Code questions, implementation details	Semantic search across 15 repos via code-talk
code-help	General code Q&A	Injects coding conventions and project context
engineer	Code changes, PR requests	Explores code, implements changes, opens PRs via Claude Code
capabilities	"What can you do?"	Describes available skills
thread-summary	"Summarize this thread"	Condenses Slack conversations

Repositories

The code-explorer skill covers all ProbeLabs projects with smart routing — it knows that "search engine" means probe, "code review" means visor, and "workflow engine" means visor-ee.

Repo	Language	Description
probe	Rust	Semantic code search engine (ripgrep + tree-sitter)
visor	TypeScript	AI code review — GitHub Action + CLI
visor-ee	TypeScript	Workflow engine powering this assistant
goreplay	Go	HTTP traffic capture and replay
logoscope	TypeScript	AI log analysis CLI + MCP server
maid	TypeScript	Mermaid diagram linter
docs-mcp	TypeScript	Turn any GitHub repo into an MCP server
memaris	TypeScript	Persistent memory for Claude Code sessions
afk	Node.js	Remote control for Claude Code via Telegram
big-brain	TypeScript	MCP server for fresh AI insights when stuck
vow	TypeScript	Accountability gates for AI agents
SandboxJS	JavaScript	Safe eval runtime
probelabs.com	Web	Company website
probe-quickstart	YAML	Quickstart template for building assistants
company-data	—	Internal knowledge base — commercial docs, ideas, strategy (private)
ground-control	—	AI chat interface (private)

Example Conversations

"How does Probe's tree-sitter integration work?"
→ code-explorer activates, searches probelabs/probe, returns explanation with file paths

"Show me recent PRs on visor"
→ github activates, runs gh pr list --repo probelabs/visor

"Why is CI failing on probe PR #42?"
→ github-actions activates, fetches logs, identifies the failing test and root cause

"Add a --verbose flag to the probe CLI"
→ engineer + code-explorer activate, explores the code, implements the change, opens a PR

Environment Variables

Variable	Required	Description
GOOGLE_API_KEY	Yes	Google AI API key (Gemini)
SLACK_APP_TOKEN	For Slack	App-level token (xapp-...) for Socket Mode
SLACK_BOT_TOKEN	For Slack	Bot token (xoxb-...) for messages and API
SLACK_SIGNING_SECRET	For Slack	Webhook request verification

Slack App Setup

1. Create app at api.slack.com/apps
2. Enable Socket Mode
3. Add Bot Token Scopes: app_mentions:read, channels:history, groups:history, im:history, mpim:history, chat:write, reactions:read, reactions:write
4. Subscribe to Bot Events: app_mention, message.channels, message.groups, message.im
5. Install to workspace, copy tokens to .env

Tests

npx -y @probelabs/visor@latest test

Building Your Own

This repo is our production config. If you want to build something similar for your team:

1. Start with probe-quickstart — a working assistant in one YAML file
2. Add your repos to the code-explorer skill's projects list
3. Add skills for your tools (Jira, Confluence, etc.) using MCP servers
4. Write knowledge docs in docs/ and reference them with {% readfile %}
5. Use this repo as a reference for how skills, routing, and workflows fit together at scale

Related

• probe-quickstart — Start here to build your own assistant
• probe — The semantic code search engine underneath
• visor — Workflow engine and code review tool
• visor-ee — Enterprise assistant workflows (code-talk, engineer, intent routing)

License

MIT