adk-dev-guide（adk 开发指南）

ADK Development Workflow & Guidelines

Session Continuity

If this is a long session, re-read the relevant skill before each phase — /adk-cheatsheet before writing code, /adk-eval-guide before running evals, /adk-deploy-guide before deploying, /adk-scaffold before scaffolding. Context compaction may have dropped earlier skill content.

DESIGN_SPEC.md — Your Primary Reference

IMPORTANT: If DESIGN_SPEC.md exists in this project, it is your primary source of truth.

Read it FIRST to understand:

• Functional requirements and capabilities
• Success criteria and quality thresholds
• Agent behavior constraints
• Expected tools and integrations

The spec is your contract. All implementation decisions should align with it. When in doubt, refer back to DESIGN_SPEC.md.

Phase 1: Understand the Spec

Before writing any code:

1. Read DESIGN_SPEC.md thoroughly
2. Identify the core capabilities required
3. Note any constraints or things the agent should NOT do
4. Understand success criteria for evaluation

Phase 2: Build and Implement

Implement the agent logic:

1. Write/modify code in the agent directory (check GEMINI.md for directory name)
2. Use make playground (or adk web.) for interactive testing during development
3. Iterate on the implementation based on user feedback

For ADK API patterns and code examples, use /adk-cheatsheet.

Phase 3: Evaluate

This is the most important phase. Evaluation validates agent behavior end-to-end using.

MANDATORY: Activate /adk-eval-guide before running evaluation. It contains the evalset schema, config format, and critical gotchas. Do NOT skip this.

Tests (pytest) are NOT evaluation. They test code correctness but say nothing about whether the agent behaves correctly. Always run adk eval.

1. Start small: Begin with 1-2 sample eval cases, not a full suite
2. Run evaluations: adk eval (or make eval if the project has a Makefile)
3. Discuss results with the user
4. Fix issues and iterate on the core cases first
5. Only after core cases pass, add edge cases and new scenarios
6. Repeat until quality thresholds are met

Expect 5-10+ iterations here.

Phase 4: Deploy

Once evaluation thresholds are met:

1. Deploy when ready — see /adk-deploy-guide for deployment options

IMPORTANT: Never deploy without explicit human approval.

Operational Guidelines for Coding Agents

Principle 1: Code Preservation & Isolation

When executing code modifications, your paramount objective is surgical precision. You must alter only the code segments directly targeted by the user's request, while strictly preserving all surrounding and unrelated code.

Mandatory Pre-Execution Verification:

Before finalizing any code replacement, verify:

1. Target Identification: Clearly define the exact lines or expressions to be changed, based *solely* on the user's explicit instructions.
2. Preservation Check: Ensure all code, configuration values (e.g., model, version, api_key), comments, and formatting *outside* the identified target remain identical.

Example:

• User Request: "Change the agent's instruction to be a recipe suggester."
• Incorrect (VIOLATION): root_agent = Agent(name="recipe_suggester", model="gemini-1.5-flash", # UNINTENDED - model was not requested to change instruction="You are a recipe suggester.")
• Correct (COMPLIANT): root_agent = Agent(name="recipe_suggester", # OK, related to new purpose model="gemini-3-flash-preview", # PRESERVED instruction="You are a recipe suggester." # OK, the direct target)

Principle 2: Execution Best Practices

• Model Selection — CRITICAL:

- NEVER change the model unless explicitly asked. If the code uses gemini-3-flash-preview, keep it as gemini-3-flash-preview. Do NOT "upgrade" or "fix" model names. - When creating NEW agents (not modifying existing), use Gemini 3 series: gemini-3-flash-preview, gemini-3-pro-preview. - Do NOT use older models (gemini-2.0-flash, gemini-1.5-flash, etc.) unless the user explicitly requests them.

• Location Matters More Than Model:

- If a model returns a 404, it's almost always a GOOGLE_CLOUD_LOCATION issue (e.g., needing global instead of us-central1). - Changing the model name to "fix" a 404 is a violation — fix the location instead. - Some models (like gemini-3-flash-preview) require specific locations. Check the error message for hints.

• ADK Built-in Tool Imports (Precision Required): # CORRECT - imports the tool instance from google.adk.tools.load_web_page import load_web_page # WRONG - imports the module, not the tool from google.adk.tools import load_web_page Pass the imported tool directly to tools=[load_web_page], not tools=[load_web_page.load_web_page].
• Running Python Commands:

- Always use uv to execute Python commands (e.g., uv run python script.py) - Run make install (or uv sync) before executing scripts - Consult Makefile and README.md for available commands (if present)

• Breaking Infinite Loops:

- Stop immediately if you see the same error 3+ times in a row - Don't retry failed operations — fix the root cause first - RED FLAGS: Lock IDs incrementing, names appending v5->v6->v7, "I'll try one more time" repeatedly - State conflicts (Error 409: Resource already exists): Import existing resources with terraform import instead of retrying creation - Tool bugs: Fix source code bugs before continuing — don't work around them - When stuck: Run underlying commands directly (e.g., terraform CLI) instead of calling problematic tools

• Troubleshooting:

- Check /adk-cheatsheet first — it covers most common patterns - Search the installed ADK package with Glob/Grep/Read (find it with python -c "import google.adk; print(google.adk.__path__[0])" — use uv run python if using uv) - For ADK docs, see references/docs-index.md in the cheatsheet skill for WebFetch URLs - For framework questions or GCP products, check official documentation - When encountering persistent errors, a targeted Google Search often finds solutions faster

Scaffold as Reference

When you need specific infrastructure files but don't want to scaffold the current project directly, use the /adk-scaffold skill to create a temporary reference project in /tmp/ and copy over what you need.

See /adk-scaffold for full CLI commands and options.

Development Commands

Projects created via the Agent Starter Pack CLI include a Makefile with these commands. To create or enhance a project, see the /adk-scaffold skill for full instructions. For non-scaffolded projects, use the ADK CLI equivalents.

Skills Version

Troubleshooting hint: If you encounter repeated errors, unexpected tool failures, or instructions in these skills that seem incomplete or reference features you can't find, the skills may be outdated. Run: `` uvx --from "google-adk" adk dev skills check ` If it reports outdated skills, tell the user to update: ` uvx --from "google-adk" adk dev skills setup -y `` Do NOT run this check on every session — only when you suspect stale skills are causing problems.