---

name: local-codebase-research description: Fact-driven retrieval for local codebases. Use codebase-wide search first, English long-form queries, then read complete files (optionally in parallel) and use CLI grep/ripgrep for keyword pinpointing when helpful.

Local Codebase Research

Core Principles

1. Fact-driven: If evidence is missing, retrieve more. Quote file paths + snippets.
2. Codebase-first: Start with mcp_augmentcode_codebase-retrieval, then narrow to files.
3. Parallel reads: Read multiple candidates in parallel when comparing implementations.

Tools

• Primary: mcp_augmentcode_codebase-retrieval
• File reading: read_file (default: startline=1, endline=1000)
• Optional: CLI tools (rg/grep) for exact symbol/key pinpointing

Retrieval Strategy

1. Ask with English long-form queries

   • ✅ "Where is reward computation implemented for RL tasks and how are rewards aggregated?"
   • ❌ "reward manager"

2. Codebase-wide retrieval first

   • Iterate multiple times (broaden if nothing found, narrow once you see candidates)

3. Read complete files when needed

   • Parallel reads for multiple candidates
   • Default 1000 lines; continue from where you left off if longer
   • Skip if retrieval provides enough evidence

4. CLI pinpoint (optional)

   • Use rg/grep for exact symbols, config keys, or registry patterns
   • Confirm behavior by reading authoritative file(s)

Output Requirements

• Provide file paths + symbol names as evidence
• Quote only what is supported by retrieved text
• State what is missing and what to retrieve next if uncertain

Anti-Patterns

❌ Jumping to read_file without retrieval
❌ Short keyword queries (except CLI pinpoint)
❌ Reading tiny chunks
❌ Assuming behavior without confirming in code