T001: Filter boilerplate from SESSION_STATE.md

new_session.py

@@ -206,6 +206,27 @@ def _tool_summary(name, inp):
         return f"{name}()"
 
 
+# Patterns that indicate a user message is pure boilerplate (hook/system noise)
+# and should be dropped entirely from SESSION_STATE.md
+_BOILERPLATE_PATTERNS = [
+    re.compile(r'^Stop hook feedback:\s*\nDO NOT STOP', re.MULTILINE),
+    re.compile(r'^You are a self-analysis agent\.'),
+    re.compile(r'SESSION START INSTRUCTIONS:'),
+    re.compile(r'^Context was reset\. Do not ask what to do\.'),
+]
+
+
+def _is_boilerplate_user_msg(text):
+    """Return True if the user message text is pure hook/system boilerplate."""
+    stripped = text.strip()
+    if not stripped:
+        return True
+    for pat in _BOILERPLATE_PATTERNS:
+        if pat.search(stripped):
+            return True
+    return False
+
+
 def _parse_and_render_tail(jsonl_lines, max_chars=32000):
     """Parse raw JSONL lines into readable conversation text.
 
@@ -269,6 +290,15 @@ def _parse_and_render_tail(jsonl_lines, max_chars=32000):
         if isinstance(content, str):
             content = [{"type": "text", "text": content}]
 
+        # Skip user messages that are pure boilerplate (hooks, system prompts)
+        if role == 'user':
+            raw_text = ' '.join(
+                b.get('text', '') for b in content
+                if isinstance(b, dict) and b.get('type') == 'text'
+            )
+            if _is_boilerplate_user_msg(raw_text):
+                continue
+
         # Format timestamp
         ts_short = ''
         if ts:

scripts/test.py

@@ -154,6 +154,38 @@ def test(name, condition):
 ctx_boundary = context_reset._parse_and_render_tail(boundary_entries)
 test("boundary shown", "compacted" in ctx_boundary and "150,000" in ctx_boundary)
 
+# --- _is_boilerplate_user_msg ---
+print("\n=== _is_boilerplate_user_msg ===")
+test("empty is boilerplate", context_reset._is_boilerplate_user_msg(""))
+test("whitespace is boilerplate", context_reset._is_boilerplate_user_msg("   \n  "))
+test("stop hook is boilerplate", context_reset._is_boilerplate_user_msg(
+    "Stop hook feedback:\nDO NOT STOP. DO NOT SUMMARIZE. DO NOT LIST OPTIONS. Follow this order:\n1) Check TODO.md..."
+))
+test("self-analysis is boilerplate", context_reset._is_boilerplate_user_msg(
+    "You are a self-analysis agent. A user interrupted Claude mid-response."
+))
+test("session start is boilerplate", context_reset._is_boilerplate_user_msg(
+    "SESSION START INSTRUCTIONS: Check TODO.md in $CLAUDE_PROJECT_DIR for pending tasks."
+))
+test("context reset prompt is boilerplate", context_reset._is_boilerplate_user_msg(
+    "Context was reset. Do not ask what to do. Pick up where the last session left off."
+))
+test("real user message is NOT boilerplate", not context_reset._is_boilerplate_user_msg(
+    "please fix the bug in auth.py"
+))
+test("looks good is NOT boilerplate", not context_reset._is_boilerplate_user_msg("looks good"))
+
+# Test boilerplate filtering in _parse_and_render_tail
+boilerplate_entries = [
+    _json.dumps({"type": "user", "message": {"role": "user", "content": [{"type": "text", "text": "Stop hook feedback:\nDO NOT STOP. DO NOT SUMMARIZE. keep going"}]}}),
+    _json.dumps({"type": "assistant", "message": {"role": "assistant", "content": [{"type": "text", "text": "Continuing work"}]}}),
+    _json.dumps({"type": "user", "message": {"role": "user", "content": [{"type": "text", "text": "real feedback here"}]}}),
+]
+ctx_bp = context_reset._parse_and_render_tail(boilerplate_entries)
+test("boilerplate user msg filtered out", "DO NOT STOP" not in ctx_bp)
+test("real user msg kept", "real feedback here" in ctx_bp)
+test("assistant after boilerplate kept", "Continuing work" in ctx_bp)
+
 # --- extract_session_context (integration) ---
 print("\n=== extract_session_context ===")
 with tempfile.TemporaryDirectory() as d: