Coding Agent Guide

This is a Claude Agent SDK project, not an ADK one: app/agent.py declares root_agent (a ClaudeAgentOptions), app/app_utils/claude.py runs it, and app/fast_api_app.py serves it over A2A. Read .agents/skills/agents-cli-claude/SKILL.md before changing any of them.

Experimental. Agent Runtime and Cloud Run are the supported targets. Models are Claude on Gemini platform: the model named in app/agent.py must be enabled in Model Garden for the project. agents-cli playground, eval dataset synthesize and eval optimize need ADK, so this project refuses them, as does publish gemini on Agent Runtime.

The loop

1. Build in app/agent.py, try it with agents-cli run "prompt".
2. Write eval cases by hand in tests/eval/datasets/, then agents-cli eval run until it holds up. agents-cli eval compare shows regressions; agents-cli eval analyze clusters failures.
3. uv run pytest tests/integration before deploying.
4. agents-cli infra single-project --apply and agents-cli deploy only after the user approves them.

Rules

• Change only what was asked. Keep surrounding code, comments, and config values.
• Never change the model unless asked. On "model not available on your vertex deployment", the model is not enabled for the project; ask the user to enable it.
• List the agent's tools explicitly (tools=[...], allowed_tools=[...]); never give a served agent Claude Code's Bash or file-editing tools unless the user asks.
• Run Python through uv: uv run python script.py, after agents-cli install.
• Same error three times means fix the cause, not retry.
• Terraform 409: terraform import, don't recreate.