| name | multi-agent-investigator |
| description | Orchestrates a multi-step investigation using specialized agents to find deep codebase problems. Use this when you need a comprehensive quality and security audit on a specific inquiry. |
Multi-Agent Investigator
This skill guides you through a rigorous, multi-step investigation of a codebase. It leverages multiple specialized sub-agents to analyze code quality, security implications, and deeply investigate a specific inquiry.
By separating the investigation into discrete steps and passing the context between specialized agents, this process uncovers problems that traditional, single-pass inquiries often miss.
The Workflow
When a user triggers this skill with an inquiry (e.g., "Investigate the authentication middleware"), follow these exact steps sequentially. Do not combine these steps into a single tool call; execute them one by one to ensure deep reasoning and context building.
Before starting, create a temporary directory to store the investigation artifacts. For example: mkdir -p .gemini/tmp/investigation.
Step 1: Baseline Quality Report
Ask the codebase_investigator agent to generate a general code quality and problems report for the relevant area of the codebase.
- Action: Call the
codebase_investigator tool.
- Prompt: "Generate a comprehensive code quality and potential problems report for [insert relevant files/directories based on the user's inquiry]. Do not focus on the specific inquiry yet; just establish the general baseline quality and identify any immediate code smells."
- Output: Save the full response to
.gemini/tmp/investigation/1.md.
Step 2: Specific Inquiry Investigation
Ask the codebase_investigator agent to deeply analyze the user's specific inquiry.
- Action: Call the
codebase_investigator tool.
- Prompt: "Investigate the following inquiry: [insert user's exact inquiry]. Focus specifically on the architecture, logic, and potential flaws related to this feature."
- Output: Save the full response to
.gemini/tmp/investigation/2.md.
Step 3: Security Engineer Follow-up
Read 1.md and 2.md. Pass this context to the security-engineer agent to review the findings and generate a list of critical follow-up questions.
- Action: Call the
security-engineer tool.
- Prompt: "Review the following baseline code quality report:\n\n[contents of 1.md]\n\nAnd the specific investigation report:\n\n[contents of 2.md]\n\nBased on these findings, what are the most critical follow-up questions, security concerns, logical gaps, or edge cases we need to look into?"
- Output: Save the response to
.gemini/tmp/investigation/3.md.
Step 4: Specialized Engineering Report
Identify the "best engineer for the job" from your available sub-agents based on the domain of the inquiry (e.g., backend-architect for APIs, frontend-developer for UI, ai-engineer for ML models, devops-automator for infrastructure). Ask this specialized agent to answer the questions raised by the security engineer.
- Action: Call the chosen specialized agent tool.
- Prompt: "Review the previous reports and the security follow-up questions:\n\n[contents of 1.md]\n[contents of 2.md]\n[contents of 3.md]\n\nPlease investigate the codebase to answer the follow-up questions from the security engineer. Provide a final, comprehensive report on your findings."
- Output: Save the response to
.gemini/tmp/investigation/4.md.
Step 5: Final Validation
Ask the same specialized engineer (or a dedicated testing agent like testing-reality-checker or evidence-collector) to validate the final findings.
- Action: Call the chosen validation agent tool.
- Prompt: "Review the final engineering report:\n\n[contents of 4.md]\n\nValidate these findings against the current state of the codebase. Are there any discrepancies, false positives, or recommendations that require immediate action? Provide a final verdict."
- Output: Present this final validation summary to the user.
Important Directives
- Patience: Take your time. Each step must be fully completed and written to its respective markdown file before proceeding to the next step.
- Persistence: Always use
write_file to save the intermediate steps (1.md, 2.md, 3.md, 4.md). This builds the necessary context for the downstream agents.
- Transparency: Inform the user briefly about which step you are currently executing. Do not work in silence for the entire workflow.
