Codebase Exploration Context Strategies

Large codebase exploration requires specific strategies to maintain accuracy as context accumulates during extended investigation sessions.

Context Degradation Problem

After exploring many files, the model's responses degrade:

• References become vague ("typical patterns" instead of specific class names)
• Earlier findings are forgotten or contradicted
• Answers become inconsistent across similar questions

Scratchpad Pattern

Have agents maintain scratchpad files that record key findings:

1. After each significant discovery, write findings to a scratchpad file
2. Before answering subsequent questions, re-read the scratchpad
3. This externalizes memory beyond the context window

Subagent Delegation

Spawn subagents for specific investigation tasks:

• "Find all test files" -- subagent explores, returns structured list
• "Trace refund flow dependencies" -- subagent traces, returns dependency graph
• Main agent stays focused on high-level coordination with clean context

Phase-Based Exploration

1. Phase 1: Map structure (subagent explores, returns summary)
2. Inject Phase 1 summary into Phase 2 context
3. Phase 2: Deep-dive specific areas based on Phase 1 findings
4. Use /compact between phases when context is heavy