diff --git a/lib/hooks.sh b/lib/hooks.sh
index 0a1d875..0ecf4c0 100644
--- a/lib/hooks.sh
+++ b/lib/hooks.sh
@@ -1,19 +1,50 @@
 #!/bin/bash
-# Stop hook management for Claude Code loop continuation.
+# Stop hook management for the agent loop.
 #
-# NOTE: Hooks are currently no-ops. The loop uses `claude --print` (non-interactive),
-# which runs to completion and exits naturally — no Stop hook is needed to signal
-# iteration boundaries. The install/remove interface is preserved so that a future
-# interactive mode can be added without changing loop.sh's call sites.
+# The Stop hook sends SIGINT to the loop's parent process when a claude session
+# finishes. This is how claude signals "I'm done" back to the bash loop —
+# the Ralph pattern (ghuntley.com/ralph).
 #
-# If interactive mode is added, the hook mechanism will need redesign: `kill -INT $PPID`
-# targets the hook runner's parent (Claude Code), not loop.sh. A sentinel-file or
-# named-pipe approach would be more reliable.
+# Without this hook, claude would exit to an interactive prompt instead of
+# returning control to the loop script.
+
+SETTINGS_FILE="${PROJECT_ROOT}/.claude/settings.local.json"
 
 install_hooks() {
-    :  # no-op — see note above
+    if [ ! -f "$SETTINGS_FILE" ]; then
+        mkdir -p "$(dirname "$SETTINGS_FILE")"
+        echo '{}' > "$SETTINGS_FILE"
+    fi
+
+    if command -v jq &>/dev/null; then
+        jq '.hooks.Stop = [{"matcher": "", "hooks": [{"type": "command", "command": "kill -INT $PPID || true"}]}]' \
+            "$SETTINGS_FILE" > "${SETTINGS_FILE}.tmp" && mv "${SETTINGS_FILE}.tmp" "$SETTINGS_FILE"
+    else
+        python3 -c "
+import json, os
+p = '$SETTINGS_FILE'
+s = json.load(open(p)) if os.path.exists(p) else {}
+s.setdefault('hooks', {})['Stop'] = [{'matcher': '', 'hooks': [{'type': 'command', 'command': 'kill -INT \$PPID || true'}]}]
+json.dump(s, open(p, 'w'), indent=2)
+"
+    fi
+    log "Stop hook installed"
 }
 
 remove_hooks() {
-    :  # no-op — see note above
+    if [ -f "$SETTINGS_FILE" ]; then
+        if command -v jq &>/dev/null; then
+            jq 'del(.hooks.Stop)' "$SETTINGS_FILE" > "${SETTINGS_FILE}.tmp" && mv "${SETTINGS_FILE}.tmp" "$SETTINGS_FILE"
+            jq 'if .hooks == {} then del(.hooks) else . end' "$SETTINGS_FILE" > "${SETTINGS_FILE}.tmp" && mv "${SETTINGS_FILE}.tmp" "$SETTINGS_FILE"
+        else
+            python3 -c "
+import json
+s = json.load(open('$SETTINGS_FILE'))
+s.get('hooks', {}).pop('Stop', None)
+if not s.get('hooks'): s.pop('hooks', None)
+json.dump(s, open('$SETTINGS_FILE', 'w'), indent=2)
+"
+        fi
+        log "Stop hook removed"
+    fi
 }
diff --git a/loop.sh b/loop.sh
index ebefa46..07e73aa 100755
--- a/loop.sh
+++ b/loop.sh
@@ -188,40 +188,33 @@ fi
 
 # --- Agent runner ---
 # Runs a prompt through the selected AI tool.
-# Two modes:
-#   Interactive (default when TTY available): runs claude in full interactive mode.
-#     The user sees the complete CC session (tool calls, file edits, etc.) in the terminal.
-#     No output capture — state is tracked via prd.json and .verdict file.
-#   Headless (no TTY or LOOP_HEADLESS=true): uses claude --print for fully autonomous operation.
-#     Output is captured to a temp file for verdict parsing.
 #
-# The function prints captured output to stdout (headless) or nothing (interactive).
+# Interactive (default): Pipes prompt to claude WITHOUT --print.
+#   This gives the full interactive CC UI — tool calls, file edits, etc.
+#   A Stop hook (installed at startup) sends SIGINT to the loop when claude
+#   finishes, which returns control to the while loop for the next iteration.
+#   State is tracked via files (prd.json, .verdict), not stdout.
+#
+# Headless (LOOP_HEADLESS=true): Uses claude --print for CI/background.
+#   Output captured to file for verdict parsing.
 run_agent() {
     local prompt="$1"
-    local role="${2:-}"  # "generator" or "evaluator" — used for verdict file
+    local role="${2:-}"
 
-    # Clean up any previous verdict file
     rm -f "$LOOP_DIR/.verdict"
 
-    # Determine whether we can run interactively
-    local has_tty=false
-    if [ "${LOOP_HEADLESS:-false}" != "true" ] && { true > /dev/tty; } 2>/dev/null; then
-        has_tty=true
-    fi
-
-    # Run in subshell so a non-zero exit from the AI tool doesn't kill the loop.
     local agent_exit=0
-    if [ "$has_tty" = true ]; then
-        # --- Interactive mode ---
-        # Run claude directly in the terminal — full interactive UI visible.
-        # No output capture. State tracked via files (prd.json, .verdict).
+    if [ "${LOOP_HEADLESS:-false}" != "true" ]; then
+        # --- Interactive mode (Ralph pattern) ---
+        # Pipe prompt to claude without --print — full interactive UI.
+        # The Stop hook handles session exit.
         (
             case "$TOOL" in
                 claude)
-                    claude --dangerously-skip-permissions "$prompt"
+                    printf '%s\n' "$prompt" | claude --dangerously-skip-permissions
                     ;;
                 amp)
-                    amp --dangerously-allow-all "$prompt"
+                    printf '%s\n' "$prompt" | amp --dangerously-allow-all
                     ;;
                 *)
                     log "ERROR: Unknown tool '$TOOL'"
@@ -230,7 +223,9 @@ run_agent() {
             esac
         ) || agent_exit=$?
 
-        # In interactive mode, read verdict from file if evaluator wrote one
+        sleep 2  # Brief pause between sessions
+
+        # Read verdict from file if evaluator wrote one
         if [ "$role" = "evaluator" ] && [ -f "$LOOP_DIR/.verdict" ]; then
             cat "$LOOP_DIR/.verdict"
         fi