v3.3.0: Native Claude Code hooks — self-evolving sessions without markdown

Add dhee/hooks/claude_code/ — 6 lifecycle hooks that make every Claude Code
session learn from its own execution. No CLAUDE.md bloat, no SKILL.md files,
no static routing tables. Vector memory with decay + token-budgeted XML
context injection (~630 tokens for rich context, constant regardless of
memory volume).

Hooks: SessionStart (context injection), UserPromptSubmit (per-turn memories),
PostToolUse (outcome capture with privacy filter), PreCompact (state survival),
Stop/SessionEnd (checkpoint with learnings).

CLI: `dhee task "..."` starts Claude Code with hooks pre-configured.
     `dhee install` / `dhee uninstall-hooks` for manual hook management.

42 tests covering renderer, privacy filter, installer, dispatch handlers.
878 existing tests unaffected.

Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>

Author: Ashish-dwi99

README.md

@@ -78,12 +78,42 @@ d.context("fixing auth bug")
 d.checkpoint("Fixed it", what_worked="git blame first")
 ```
 
+### Claude Code — Native Hooks (v3.3.0)
+
+One command. Every Claude Code session becomes self-evolving.
+
+```bash
+dhee install
+```
+
+That's it. Dhee hooks into Claude Code's lifecycle — no CLAUDE.md bloat, no SKILL.md files, no markdown accumulation. Structured XML context injection, budgeted to ~630 tokens regardless of how much memory you have.
+
+**What happens automatically:**
+
+| Hook | When | What Dhee does |
+|:-----|:-----|:---------------|
+| `SessionStart` | Session opens | Injects last session, insights, performance trends, relevant memories |
+| `UserPromptSubmit` | Every prompt | Surfaces memories relevant to what you just asked |
+| `PostToolUse` | After Edit/Write/Bash | Captures what Claude did (secrets auto-stripped) |
+| `PreCompact` | Before context compaction | Checkpoints state so nothing is lost |
+| `Stop` | Session ends | Records outcomes — what worked, what failed, learnings |
+
+Or start Claude Code directly with context:
+
+```bash
+dhee task "fix the flaky auth test"
+```
+
+**Why not CLAUDE.md?** Markdown files are static. After 6 months of accumulated knowledge, they rot — stale patterns sit at equal weight to current ones, no retrieval ranking, no forgetting. Dhee uses vector memory with strength-based decay. Relevant memories surface. Irrelevant ones fade. The context budget stays constant at ~630 tokens whether you have 50 memories or 50,000.
+
 ### CLI
 
 ```bash
 dhee remember "User prefers Python"
 dhee recall "programming language"
 dhee checkpoint "Fixed auth bug" --what-worked "checked logs"
+dhee install          # install Claude Code hooks
+dhee uninstall-hooks  # remove them
 ```
 
 ### Docker
@@ -219,8 +249,18 @@ These are surfaced through `context()` and `checkpoint()` automatically when ena
 ## Architecture
 
 ```
-Agent (Claude, GPT, Cursor, custom)
+Claude Code (or any agent)
   │
+  ├── SessionStart hook ──→ dhee.context() ──→ XML renderer ──→ system prompt injection
+  ├── UserPromptSubmit ───→ dhee.recall()  ──→ ranked memories ──→ per-turn context
+  ├── PostToolUse ────────→ dhee.remember() ─→ privacy filter ──→ stored (0 LLM)
+  ├── PreCompact ─────────→ dhee.checkpoint() + re-inject context
+  └── Stop ───────────────→ dhee.checkpoint(what_worked, what_failed, outcome_score)
+```
+
+The 4-operation API under the hooks:
+
+```
   ├── remember(content)     → Engram: embed + store (0 LLM)
   ├── recall(query)         → Engram: embed + vector search (0 LLM)
   ├── context(task)         → Buddhi: performance + insights + intentions + memories

dhee/__init__.py

@@ -32,7 +32,7 @@
 # Default: CoreMemory (lightest, zero-config)
 Memory = CoreMemory
 
-__version__ = "3.2.0"
+__version__ = "3.3.0"
 __all__ = [
     # Memory classes
     "Engram",

dhee/cli.py

@@ -318,6 +318,58 @@ def cmd_uninstall(args: argparse.Namespace) -> None:
         print("Cancelled.")
 
 
+def cmd_task(args: argparse.Namespace) -> None:
+    """Start Claude Code with Dhee cognition hooks."""
+    from dhee.hooks.claude_code.install import ensure_installed
+
+    result = ensure_installed()
+    if result.already_installed:
+        pass  # hooks already in place
+    elif result.created or result.updated:
+        print(f"  Dhee hooks installed → {result.settings_path}")
+
+    # Find claude executable
+    claude_bin = shutil.which("claude")
+    if not claude_bin:
+        print("Error: 'claude' not found in PATH. Install Claude Code first.", file=sys.stderr)
+        sys.exit(1)
+
+    # Build command
+    cmd = [claude_bin]
+    if args.print_mode:
+        cmd.append("--print")
+    if args.description:
+        cmd.append(args.description)
+
+    # Replace current process with claude
+    os.execvp(claude_bin, cmd)
+
+
+def cmd_install_hooks(args: argparse.Namespace) -> None:
+    """Install Dhee hooks into Claude Code."""
+    from dhee.hooks.claude_code.install import install_hooks
+
+    result = install_hooks(force=args.force)
+    if result.already_installed and not args.force:
+        print("  Dhee hooks already installed.")
+    else:
+        action = "Created" if result.created else "Updated"
+        print(f"  {action} {result.settings_path}")
+        print(f"  Hooks: {', '.join(result.events)}")
+        if result.backed_up:
+            print(f"  Backup: {result.backed_up}")
+
+
+def cmd_uninstall_hooks(args: argparse.Namespace) -> None:
+    """Remove Dhee hooks from Claude Code."""
+    from dhee.hooks.claude_code.install import uninstall_hooks
+
+    if uninstall_hooks():
+        print("  Dhee hooks removed.")
+    else:
+        print("  No Dhee hooks found.")
+
+
 def cmd_benchmark(args: argparse.Namespace) -> None:
     """Run performance benchmarks."""
     import time
@@ -458,6 +510,19 @@ def build_parser() -> argparse.ArgumentParser:
     p_status = sub.add_parser("status", help="Show version, config, and agents")
     p_status.add_argument("--json", action="store_true", help="JSON output")
 
+    # task
+    p_task = sub.add_parser("task", help="Start Claude Code with Dhee cognition")
+    p_task.add_argument("description", nargs="?", default="", help="Task description")
+    p_task.add_argument("--user-id", default="default", help="User ID")
+    p_task.add_argument("--print", dest="print_mode", action="store_true", help="One-shot mode")
+
+    # install (hooks)
+    p_install = sub.add_parser("install", help="Install Dhee hooks into Claude Code")
+    p_install.add_argument("--force", action="store_true", help="Overwrite existing hooks")
+
+    # uninstall-hooks
+    sub.add_parser("uninstall-hooks", help="Remove Dhee hooks from Claude Code")
+
     # benchmark
     sub.add_parser("benchmark", help="Run performance benchmarks")
 
@@ -481,6 +546,9 @@ def build_parser() -> argparse.ArgumentParser:
     "export": cmd_export,
     "import": cmd_import,
     "status": cmd_status,
+    "task": cmd_task,
+    "install": cmd_install_hooks,
+    "uninstall-hooks": cmd_uninstall_hooks,
     "benchmark": cmd_benchmark,
     "uninstall": cmd_uninstall,
 }

dhee/hooks/__init__.py

@@ -0,0 +1,2 @@
+from dhee.hooks.claude_code.install import ensure_installed, install_hooks, uninstall_hooks
+from dhee.hooks.claude_code.renderer import estimate_tokens, render_context