# CLI Agent Starter Kit
This is a TypeScript CLI agent starter kit with AI SDK and Anthropic integration.
## Important: Private Package
This package is marked as `"private": true` in package.json and is NOT published to npm. Therefore:
- **NO changesets are required** for changes to this package
- This is an example/starter kit, not a published library
- Changes here do not need versioning or changelog entries
## Available Skills
This project has access to Phoenix skills: `phoenix-cli`, `phoenix-tracing`, and `phoenix-evals`.
Use these skills when working with Phoenix tracing, evaluations, or the Phoenix CLI.
## Development
```bash
pnpm dev # Run the agent
pnpm build # Build the TypeScript
pnpm start # Run the compiled output
```
## Environment Variables
- `ANTHROPIC_API_KEY` - Required for Anthropic API calls
## Debugging with Phoenix CLI
### Quick Trace Check
To verify traces are being captured:
```bash
npx @arizeai/phoenix-cli traces \
--endpoint http://localhost:6006 \
--project cli-agent-starter-kit \
--limit 5
```
### View Recent Traces with Key Details
```bash
npx @arizeai/phoenix-cli traces \
--endpoint http://localhost:6006 \
--project cli-agent-starter-kit \
--limit 5 \
--format raw \
--no-progress | jq '.[] | {
traceId: .traceId,
startTime: .spans[0].start_time,
operation: .spans[0].attributes["operation.name"],
model: .spans[0].attributes["ai.model.id"]
}'
```
### Inspect Specific Trace (including tool calls)
```bash
npx @arizeai/phoenix-cli trace <trace-id> \
--endpoint http://localhost:6006 \
--project cli-agent-starter-kit \
--format raw \
--no-progress | jq '{
traceId: .traceId,
spanCount: (.spans | length),
spans: [.spans[] | {
name,
span_kind,
tool: .attributes["ai.toolCall.name"],
model: .attributes["ai.model.id"]
}]
}'
```
### Important Notes
- **Project Name**: Must match `projectName` in `src/instrumentation.ts` (currently `cli-agent-starter-kit`)
- **Endpoint**: Local Phoenix is at `http://localhost:6006`
- **Format**: Use `--format raw --no-progress` for piping to `jq`
- **Telemetry**: ToolLoopAgent requires `experimental_telemetry: { isEnabled: true }` in constructor
### Common Span Attributes
- `operation.name` - Operation type (e.g., `ai.generateText`)
- `ai.model.id` - Model name (e.g., `claude-sonnet-4-20250514`)
- `ai.toolCall.name` - Tool name for TOOL spans
- `ai.usage.completionTokens` / `ai.usage.promptTokens` - Token counts
- `span_kind` - Span type: `AGENT`, `LLM`, `TOOL`, `CHAIN`, etc.
