fix(claude-code): filter injected harness turns from retention

hindsight-integrations/claude-code/scripts/lib/content.py

@@ -48,6 +48,28 @@ def strip_memory_tags(content: str) -> str:
     return content
 
 
+def strip_harness_blocks(content: str) -> str:
+    """Remove <system-reminder> and <task-notification> blocks.
+
+    Claude Code injects these into user turns. They are harness output rather
+    than conversation, so retaining them teaches Hindsight about hook and tool
+    plumbing instead of the user's work.
+    """
+    content = re.sub(r"<system-reminder>[\s\S]*?</system-reminder>", "", content)
+    content = re.sub(r"<task-notification>[\s\S]*?</task-notification>", "", content)
+    return content
+
+
+def is_synthetic_user_message(content: str) -> bool:
+    """Return True for skill bodies Claude Code persists as user chat.
+
+    Claude Code records an invoked skill's full SKILL.md as a normal user
+    message. Retaining it stores documentation as if the user had written it,
+    and can create very large retain payloads.
+    """
+    return content.lstrip().startswith("Base directory for this skill: ")
+
+
 # ---------------------------------------------------------------------------
 # Recall: query composition and truncation
 # ---------------------------------------------------------------------------
@@ -244,7 +266,7 @@ def _extract_message_blocks(content, role: str = "") -> list:
       - Channel message tool_use blocks get their text extracted inline.
     """
     if isinstance(content, str):
-        cleaned = strip_channel_envelope(strip_memory_tags(content)).strip()
+        cleaned = strip_harness_blocks(strip_channel_envelope(strip_memory_tags(content))).strip()
         return [{"type": "text", "text": cleaned}] if cleaned else []
 
     if not isinstance(content, list):
@@ -257,7 +279,7 @@ def _extract_message_blocks(content, role: str = "") -> list:
         block_type = block.get("type", "")
 
         if block_type == "text":
-            text = strip_channel_envelope(strip_memory_tags(block.get("text", ""))).strip()
+            text = strip_harness_blocks(strip_channel_envelope(strip_memory_tags(block.get("text", "")))).strip()
             if text:
                 blocks.append({"type": "text", "text": text})
 
@@ -342,6 +364,15 @@ def prepare_retention_transcript(
 
     allowed_roles = set(retain_roles or ["user", "assistant"])
 
+    target_messages = [
+        msg
+        for msg in target_messages
+        if not (
+            msg.get("role") == "user"
+            and is_synthetic_user_message(_extract_text_content(msg.get("content", ""), role="user"))
+        )
+    ]
+
     if include_tool_calls:
         return _prepare_json_transcript(target_messages, allowed_roles)
     return _prepare_text_transcript(target_messages, allowed_roles)
@@ -384,7 +415,7 @@ def _prepare_text_transcript(messages: list, allowed_roles: set) -> tuple:
 
         content = _extract_text_content(msg.get("content", ""), role=role)
         content = strip_channel_envelope(content)
-        content = strip_memory_tags(content).strip()
+        content = strip_harness_blocks(strip_memory_tags(content)).strip()
 
         if not content:
             continue

hindsight-integrations/claude-code/tests/test_content.py

@@ -10,9 +10,11 @@
     compose_recall_query,
     format_current_time,
     format_memories,
+    is_synthetic_user_message,
     prepare_retention_transcript,
     slice_last_turns_by_user_boundary,
     strip_channel_envelope,
+    strip_harness_blocks,
     strip_memory_tags,
     truncate_recall_query,
 )
@@ -67,6 +69,54 @@ def test_strips_multiline_block(self):
         assert strip_memory_tags(raw).strip() == ""
 
 
+# ---------------------------------------------------------------------------
+# strip_harness_blocks
+# ---------------------------------------------------------------------------
+
+
+class TestStripHarnessBlocks:
+    def test_strips_system_reminder_block(self):
+        raw = "before\n<system-reminder>plugin chatter</system-reminder>\nafter"
+        result = strip_harness_blocks(raw)
+        assert "plugin chatter" not in result
+        assert "before" in result
+        assert "after" in result
+
+    def test_strips_task_notification_block(self):
+        raw = "text <task-notification>background job done</task-notification> text"
+        result = strip_harness_blocks(raw)
+        assert "background job done" not in result
+
+    def test_strips_multiline_block(self):
+        raw = "<system-reminder>\nline1\nline2\n</system-reminder>"
+        assert strip_harness_blocks(raw).strip() == ""
+
+    def test_passthrough_clean_text(self):
+        raw = "no harness blocks here"
+        assert strip_harness_blocks(raw) == raw
+
+    def test_keeps_unpaired_mention(self):
+        raw = "the <system-reminder> tag is injected by the harness"
+        assert strip_harness_blocks(raw) == raw
+
+
+# ---------------------------------------------------------------------------
+# is_synthetic_user_message
+# ---------------------------------------------------------------------------
+
+
+class TestIsSyntheticUserMessage:
+    def test_detects_skill_body(self):
+        raw = "Base directory for this skill: /tmp/skills/foo\n\n# Foo\n\nDocs."
+        assert is_synthetic_user_message(raw) is True
+
+    def test_detects_leading_whitespace(self):
+        assert is_synthetic_user_message("\n  Base directory for this skill: /tmp/x") is True
+
+    def test_rejects_ordinary_message(self):
+        assert is_synthetic_user_message("what is the base directory for this skill?") is False
+
+
 # ---------------------------------------------------------------------------
 # slice_last_turns_by_user_boundary
 # ---------------------------------------------------------------------------
@@ -323,6 +373,36 @@ def test_strips_memory_tags(self):
         assert "leaked" not in transcript
         assert "actual question" in transcript
 
+    def test_strips_harness_blocks(self):
+        msgs = _msgs(("user", "<system-reminder>plugin chatter</system-reminder> actual question"))
+        transcript, _ = prepare_retention_transcript(msgs, retain_full_window=True)
+        assert "plugin chatter" not in transcript
+        assert "actual question" in transcript
+
+    def test_skips_skill_body_user_message(self):
+        msgs = _msgs(
+            ("user", "Base directory for this skill: /tmp/skills/foo\n\n# Foo\n\nDocs."),
+            ("user", "real question"),
+            ("assistant", "real reply"),
+        )
+        transcript, count = prepare_retention_transcript(msgs, retain_full_window=True)
+        assert "Docs." not in transcript
+        assert "real question" in transcript
+        assert count == 2
+
+    def test_skips_skill_body_in_json_mode(self):
+        msgs = _msgs(
+            ("user", "Base directory for this skill: /tmp/skills/foo\n\n# Foo\n\nDocs."),
+            ("user", "real question"),
+        )
+        transcript, count = prepare_retention_transcript(msgs, retain_full_window=True, include_tool_calls=True)
+        assert "Docs." not in transcript
+        assert count == 1
+
+    def test_returns_none_when_only_skill_body(self):
+        msgs = _msgs(("user", "Base directory for this skill: /tmp/skills/foo\n\n# Foo\n\nDocs."))
+        assert prepare_retention_transcript(msgs, retain_full_window=True) == (None, 0)
+
     def test_filters_by_retain_roles(self):
         msgs = _msgs(("user", "user msg"), ("assistant", "assistant msg"))
         transcript, _ = prepare_retention_transcript(msgs, retain_roles=["user"], retain_full_window=True)