v3.3.0: Native Claude Code Hooks

Native Claude Code Hooks — Self-Evolving Sessions Without Markdown

One command makes every Claude Code session learn from its own execution:

dhee install

No CLAUDE.md bloat. No SKILL.md files. No static routing tables. Vector memory with strength-based decay + token-budgeted XML context injection (~630 tokens for rich context, constant regardless of memory volume).

What's new

6 lifecycle hooks wired into Claude Code's native hook system:

• SessionStart — injects full context (last session, insights, performance trends, relevant memories)
• UserPromptSubmit — surfaces per-turn memories relevant to current prompt
• PostToolUse — captures tool outcomes with automatic secret filtering
• PreCompact — checkpoints state to survive context compaction
• Stop / SessionEnd — records outcomes, what worked, what failed

CLI commands:

• dhee task "fix the auth bug" — starts Claude Code with hooks pre-configured
• dhee install / dhee uninstall-hooks — manual hook management

Privacy filter — 10 regex patterns strip API keys, tokens, passwords, and secrets before any tool output enters memory.

XML renderer — priority-ordered sections, budget-enforced. High-priority context (session state, performance) always included. Low-priority sections (episodes, policies) drop gracefully when budget is tight.

Why not CLAUDE.md / RESOLVER.md / SKILL.md?

Markdown files are static. After accumulation they rot — stale patterns at equal weight to current ones, no retrieval ranking, no forgetting. Dhee uses vector memory with Ebbinghaus decay. Relevant memories surface. Irrelevant ones fade. The context budget stays constant whether you have 50 memories or 50,000.

Stats

• 42 new tests (renderer, privacy, installer, dispatch handlers)
• 878 total tests passing, 0 regressions
• ~630 tokens for fully-populated context
• 1500 token budget cap (configurable)