[PFR] Add CODA (Coding Agents on Databricks Apps) as a supported workshop coding assistant

[varunrao]

Product Feature Request: CODA Support

Summary

The workshop currently supports Cursor, VS Code + GitHub Copilot, and Claude Code as AI coding assistant targets. Issue #2 proposes adding Genie Code as a Databricks-native option. This PFR requests adding CODA (Coding Agents on Databricks Apps) as another supported coding assistant -- giving participants a browser-based, multi-agent environment that runs entirely on Databricks Apps with zero local setup.

CODA is uniquely suited for the workshop because it provides multiple coding agents (Claude Code, Codex, Gemini CLI, OpenCode) pre-wired to the Databricks AI Gateway, with built-in MLflow tracing, Unity Catalog integration, and 39 pre-installed Databricks skills -- all running in a browser terminal on Databricks Apps.

---

1. Current State

The workshop assumes participants use an external, locally-installed IDE:

Assistant	Setup Required	How Prompts Are Used
Cursor	Install Cursor, clone repo, configure Claude model	Copy prompt from workshop, paste into Cursor Agent panel
VS Code + Copilot	Install VS Code + Copilot extension, clone repo	Copy prompt, paste into Copilot chat
Claude Code	Install Claude Code CLI, clone repo	Copy prompt, paste into terminal

All three require: local installation, Git clone, local file system access, and manual configuration. The setup step (SetUpProjectStep.tsx) explicitly instructs users to "Open Cursor or VS Code from your Applications folder" and configure models.

---

2. What CODA Provides

CODA is an open-source template that runs coding agents directly on Databricks Apps:

Multi-Agent Support

Agent	Description
Claude Code	Anthropic's coding agent with 39 Databricks skills + 2 MCP servers
Codex	OpenAI's coding agent, pre-configured for Databricks
Gemini CLI	Google's coding agent with shared skills
OpenCode	Open-source agent with multi-provider support

All agents start pre-wired to the Databricks AI Gateway -- models, auth tokens, and base URLs are configured at boot. No API keys to manage.

Enterprise Features

Feature	Benefit
Unity Catalog integration	All data access governed by UC permissions
AI Gateway routing	All LLM calls go through a single control plane -- swap models, set rate limits
MLflow tracing	Every session automatically traced -- review prompts, tool calls, and outputs
39 Databricks skills	Pre-installed skills covering AI/agents, analytics, data engineering, storage
Workspace sync	Every git commit auto-syncs to /Workspace/Users/{you}/projects/
Browser-based terminal	xterm.js with split panes, themes, WebSocket I/O, image paste, voice input

39 Pre-Installed Databricks Skills

Organized across categories directly relevant to workshop steps:

• AI & Agents: agent-bricks, genie, mlflow-eval, model-serving
• Analytics: aibi-dashboards, unity-catalog, metric-views
• Data Engineering: declarative-pipelines, jobs, structured-streaming
• Development: asset-bundles, app-apx, app-python, python-sdk
• Storage: lakebase-autoscale, lakebase-provisioned, vector-search

These skills align directly with the workshop's chapters: Databricks App (asset-bundles, app-python), Lakebase (lakebase-autoscale/provisioned), Lakehouse (declarative-pipelines, structured-streaming), Data Intelligence (genie, aibi-dashboards, agent-bricks, metric-views).

---

3. Why CODA Is a Strong Fit for the Workshop

Zero Local Setup

• No IDE installation, no Git clone to local disk, no local Python/Node.js
• Participants open a browser URL and start coding immediately
• Workshop facilitators don't need to troubleshoot local environment issues

Multi-Agent Flexibility

• Participants can choose their preferred agent within a single CODA instance (Claude Code, Codex, Gemini, OpenCode)
• Switch between agents mid-workshop without reconfiguration
• Compare agent behavior on the same workshop step

Pre-Wired Databricks Integration

• CODA instances are already authenticated to the participant's workspace
• Unity Catalog permissions enforced automatically
• AI Gateway handles model routing -- no manual endpoint configuration
• The Databricks CLI is pre-configured and ready to use

Built-In Observability

• MLflow tracing captures every session -- facilitators can review participant work
• Token usage, cost, and latency tracked per user via AI Gateway
• Workshop completion can be verified from trace data

Skill Alignment

• CODA's 39 skills map directly to workshop topics
• The genie skill supports Genie Space creation (Step 17)
• The agent-bricks skill supports agent building (Step 18)
• The lakebase-* skills support Lakebase setup (Steps 7-10)
• The aibi-dashboards and metric-views skills support Data Intelligence (Steps 15-17)

---

4. Proposed Implementation

Phase 1: Assistant Selector Update

• Add "CODA (Browser-based)" as an option in the assistant picker (alongside Cursor, VS Code, Claude Code, Genie Code)
• Store selection in session state

Phase 2: CODA-Specific Setup Flow

When CODA is selected, replace the "install IDE + clone repo" setup with:

1. Deploy a CODA instance to the participant's workspace (or use a shared one)
2. Open the CODA URL in a new tab
3. Select an agent (Claude Code, Codex, Gemini, or OpenCode)
4. Verify Databricks CLI is configured (databricks auth status)
5. Clone the workshop template repo from within CODA

The setup step in SetUpProjectStep.tsx should render CODA-specific instructions instead of "Open Cursor or VS Code from your Applications folder."

Phase 3: Adapted Prompt Delivery

For CODA participants:

• Prompts formatted for terminal-based agents (Claude Code CLI style, not IDE panel style)
• "How to Apply" instructions reference the CODA browser terminal, not Cursor/VS Code menus
• One-click copy still works -- participants paste directly into the CODA terminal

Phase 4: Observability Integration

• Link workshop session to CODA's MLflow experiment
• Show step completion status derived from MLflow traces
• Facilitator dashboard showing participant progress across CODA sessions

Phase 5: Pre-Provisioned Workshop Mode

• Create a workshop-optimized CODA template with:
  • Workshop template repo pre-cloned
  • Relevant skills pre-activated
  • Workshop-specific MCP servers configured
  • .cursor/rules or CLAUDE.md with workshop context pre-populated

---

5. What Needs to Change

Component	Current	Proposed
SetUpProjectStep.tsx	"Open Cursor or VS Code" instructions	Conditional rendering for CODA: "Open your CODA instance"
LevelSelector.tsx	No assistant selection	Add CODA to assistant picker
workflowSections.ts	cursor_copilot_ui_design section tag	Make assistant-agnostic (also needed for #2 Genie Code)
design_prd.md	"Cursor, GitHub Copilot, or Claude Code"	Add CODA
db/lakebase/README.md	"A coding assistant (Cursor, Copilot, or Claude)"	Add CODA
Prompt templates	IDE-oriented "How to Apply"	Add CODA terminal-based variant
Session state	No assistant metadata	Store selected assistant + CODA instance URL

---

6. Relationship to Other Issues

• Issue [PFR] Add Genie Code support as a workshop coding assistant target #2 (Genie Code): CODA and Genie Code are complementary -- Genie Code is Databricks' built-in AI in notebooks/SQL Editor, while CODA is a browser-based multi-agent terminal. Both eliminate local IDE dependency. The assistant selector should support all options.
• Issue [PFR] Integrate Databricks AppKit SDK to replace custom React + FastAPI + Lakebase stack #1 (AppKit): If the app migrates to AppKit, CODA integration could leverage AppKit's plugin system for tighter coupling.
• Issue [Enhancement] Leverage AppKit Genie Plugin and Agent Plugin for live in-app Agent & Genie integration #4 (AppKit Agent + Genie): CODA's genie and agent-bricks skills could complement the embedded Genie/Agent panels proposed in [Enhancement] Leverage AppKit Genie Plugin and Agent Plugin for live in-app Agent & Genie integration #4.

---

References

• CODA repo: https://github.com/datasciencemonkey/coding-agents-databricks-apps
• CODA docs: https://datasciencemonkey.github.io/coding-agents-databricks-apps/
• Current setup step: src/components/SetUpProjectStep.tsx
• Current workflow sections: src/constants/workflowSections.ts
• Current PRD: docs/design_prd.md
• Related: Issue [PFR] Add Genie Code support as a workshop coding assistant target #2 (Genie Code support), Issue [PFR] Integrate Databricks AppKit SDK to replace custom React + FastAPI + Lakebase stack #1 (AppKit migration)