Add size limits to prevent context overflow in large repos

[JordanCoin]

Summary

• Session-start hook now uses adaptive depth based on repo size (depth 2-4)
• Both hook and MCP get_structure enforce 60KB max output (~15k tokens, <10% of context)
• Truncates cleanly at line boundaries with helpful message

Problem

Large repos (10k+ files like Rails monoliths) were outputting 1.3MB+ of tree structure on session start, consuming 250%+ of Claude Code's context window before any conversation even started.

Root Cause: Hook Output Goes to "Messages"

The critical insight is that hook output gets injected into the "Messages" portion of Claude's context, not into system prompt or tools. This means:

1. Hook output competes directly with conversation history
2. /clear doesn't help because hooks re-run on session start
3. Even with no conversation, "Messages" can show 500k+ tokens
4. Users see "Context limit reached" immediately on fresh sessions

Why This Matters for Hook Architecture

Hooks need to be context-aware. A hook that's helpful for a 500-file project becomes destructive for a 10k-file project. Current hooks assume unlimited output is fine - it's not.

Principles for future hook design:

• Hooks should NEVER output more than ~10% of context window (~20k tokens / ~80KB)
• Output should scale inversely with repo size (bigger repo = less detail)
• Consider making hooks return structured data instead of free-form text
• Add a global hook output budget that hooks share

The Quick Fix (This PR)

• Adaptive depth: >5000 files → depth 2, >2000 files → depth 3, else depth 4
• Hard cap: 60KB max output with clean truncation
• Applies to both CLI hooks and MCP get_structure tool

Future Refactoring Ideas

1. Hook output budget: Global limit shared across all hooks
2. Structured hook responses: Return JSON that Claude Code can format/truncate
3. Lazy loading: Show summary first, let user request details
4. Project-specific config: Allow .codemap/config.json to tune hook behavior
5. Hook priority system: Critical info first, nice-to-have truncated

Test Results

• Tested on gumroad repo (10,144 files, 303MB):
  • Before: 1,375,873 bytes (1.3MB) → 500k+ tokens
  • After: 60,973 bytes (61KB) → ~15k tokens
• Both codemap and codemap-mcp rebuild successfully
• Claude Code sessions now start with normal context usage

Files Changed

• cmd/hooks.go: Adaptive depth + size limit for session-start hook
• mcp/main.go: Size limit for get_structure MCP tool

🤖 Generated with Claude Code