tercel/gemini-forge/git-researcher

Git Researcher

User Guide (How to use)

You can trigger this skill by asking me to research GitHub projects. Here are some examples:

• Basic Search: "Research Python LLM agent projects with more than 100 stars."
• Complex Search: "Search for topic:ai topic:rag stars:50..500 language:python. Analyze the top 2 and give me a Chinese report (--lang zh)."
• Specific Project: "Analyze the repository 'langchain-ai/langchain' from technical, user, and investor perspectives."
• Batch Analysis: "From my last search results, analyze project #3 and #5."

Parameters you can specify:

• Keywords/Topics: topic:ai, llm, agent, etc.
• Metrics: stars:>100, forks:10..50.
• Language: --lang zh (Chinese report), --lang en (English report, default).
• Limit: "Show me the top 10 results."

Architectural Principles

• Orchestration: The main agent acts as an orchestrator, delegating heavy lifting to sub-agents.
• Persistence: Search results are cached locally in search_results.json to avoid redundant API calls.
• Isolation: Each repository analysis is handled by a fresh sub-agent (e.g., generalist or codebase_investigator) within its specific directory.

Workflow

1. Search & Cache Phase

• Use gh search repos with user criteria.
• Action: Save the raw JSON output to search_results.json in the current directory.
• Example: gh search repos "..." --json ... > search_results.json
• Logic: If the user asks for more projects from a previous query, check search_results.json first before re-searching.

2. Selection & Delegation

• Read search_results.json and present a numbered list to the user.
• For each selected project, delegate the process to a sub-agent.

3. Execution Phase (Sub-Agent Delegation)

For each project, spawn a sub-agent (using the generalist tool) with the following specific mission:

1. Clone: Clone the repository into a unique directory.
2. Investigate: Use codebase_investigator inside that directory to understand the Technical, User, and Investor perspectives.
3. Draft: Create the git-research.md report based on references/report_guide.md.
4. Report Back: Return only a concise summary of the analysis to the Orchestrator.

4. Reporting & Synthesis

• The Orchestrator summarizes the status of all analyzed projects.
• Ensures the workspace remains clean by isolating project-specific artifacts.

Guidelines

• Token Efficiency: Never read the entire codebase into the main context. Always use codebase_investigator via a sub-agent.
• Anti-Redundancy: Verify search_results.json and existing directories before every major action.
• Multi-lingual: If --lang zh is used, ensure the sub-agent generates the report in Chinese.