Context in large codebase exploration

Key points

1. 1Long exploration sessions degrade: answers start contradicting earlier ones and cite "typical patterns" instead of the specific classes found earlier. The cure is managing context, not a bigger window.
2. 2Spawn subagents for specific questions ("find all test files", "trace refund flow dependencies"). Their verbose output stays in their context; the main agent keeps the high-level map and gets back concise findings.
3. 3Keep a scratchpad file of key findings (class names, file paths, what was ruled out) and consult it before answering later questions. It survives /compact and context boundaries.
4. 4Between phases, summarise what the phase established and inject that summary into the next-phase subagents' prompts. Use /compact when verbose discovery output fills the context.
5. 5For crash recovery, each agent exports its state to a known location; on resume the coordinator loads a manifest and injects it into agent prompts, so only unfinished work is redone.