Prevent include parsing across inline documentation tokens

.pdd/meta/sync_order_python.json

@@ -0,0 +1,16 @@
+{
+  "pdd_version": "0.0.298.dev0",
+  "timestamp": "2026-07-10T06:09:15.384060+00:00",
+  "command": "fix",
+  "prompt_hash": "2aefd06c3a9bdf2c2fc30dcafd174edf7820c33fa56ff40c9826009008c644bb",
+  "code_hash": "22dec5209b81b3ab64575f541212c85b86b534844db64259a71d8d8d18daf750",
+  "example_hash": null,
+  "test_hash": null,
+  "test_files": null,
+  "include_deps": {
+    "context/agentic_common_example.py": "c1538a115a3a33da228e7f6cc048f2a99023cd2c83a0efef382e49c477b24ad3",
+    "context/architecture_sync_example.py": "b6158dbc7ca3fd32665528562225f6739ec521a071d845acab0f7caaea221ae0",
+    "context/auto_include_example.py": "0c278ed716298149777bc7082827eb8aa6d80dac264c66b6cded6871232f2df7",
+    "context/python_preamble.prompt": "0388ed131bf986f8752e1bc4c81e4da0460cfe2908ec8c60b1314edbab768254"
+  }
+}

pdd/prompts/sync_order_python.prompt

@@ -16,7 +16,7 @@ A utility module that parses `<include>` tags from prompt files, builds a depend
      - bare body:        `<include>path</include>`
      - attributed body:  `<include query="...">path</include>` (or any other attribute — discovery doesn't inspect them, just allows them on the opening tag)
      - self-closing:     `<include path="..." />`
-   - Body-form regex: `r'<include(?:\s+([^>]*?))?(?<!/)>(.*?)</include>'` with `re.DOTALL`, capturing the optional attribute string so the parser can honor a `path=` attribute on attributed body includes. The `(?<!/)>` negative lookbehind is required so a self-closing `<include path="..." />` is NOT picked up by the body-form regex (without it, the body-form would absorb everything between the self-closing tag and the next `</include>`, dropping inner includes entirely).
+   - Body-form regex: `r'<include(?:\s+([^>]*?))?(?<!/)>((?:(?!<include\b).)*?)</include>'` with `re.DOTALL`, capturing the optional attribute string so the parser can honor a `path=` attribute on attributed body includes. The `(?<!/)>` negative lookbehind is required so a self-closing `<include path="..." />` is NOT picked up by the body-form regex. The tempered body MUST reject any match that crosses another `<include` opener; otherwise an inline documentation token such as `` `<include>` `` can pair with a later real closing tag and turn hundreds of lines of prompt prose into a bogus path.
    - Self-closing regex: `r'<include\s+([^>]*?)\s*/>'` with `re.DOTALL`, then extract the `path` attribute via `r'path\s*=\s*["\']([^"\']+)["\']'` from the captured attrs.
    - `<include-many>` body regex: `r'<include-many(?:\s+[^>]*?)?>(.*?)</include-many>'` (attributes optional on the opening tag).
    - For `<include-many>` payload, mirror `pdd.preprocess.process_include_many_tags`: split on newlines first, then on commas within each line, strip each piece, drop empties. Both separators must be supported because real prompts write the list either inline (`<include-many>a, b, c</include-many>`) or one-per-line (`<include-many>\na\nb\nc\n</include-many>`).

pdd/sync_order.py

@@ -54,7 +54,13 @@ def extract_includes_from_file(file_path: Path) -> Set[str]:
         #   - self-closing:     ``<include path="..." />``
         # Real prompts use all three; missing any one means a real PDD
         # include form sits outside #739's safety net.
-        # Body-form must exclude self-closing `<include ... />`. Without
+        # Body-form must exclude self-closing `<include ... />` and must not
+        # cross another include opener. Without the tempered body match, an
+        # inline documentation token such as `` `<include>` `` can consume
+        # everything through a later real ``</include>`` and surface hundreds
+        # of lines as a bogus filesystem path.
+        #
+        # Without
         # the `(?<!/)>` lookbehind, the body-form regex would absorb the
         # entire `<include path="..." />\n<include>foo.md` span as one
         # match, losing the inner include and producing garbage.
@@ -68,7 +74,9 @@ def extract_includes_from_file(file_path: Path) -> Set[str]:
         # actually resolves it as ``docs/source.md`` — and the scope
         # guard's allowlist would diverge from the real include graph.
         single_matches = re.findall(
-            r'<include(?:\s+([^>]*?))?(?<!/)>(.*?)</include>', content, re.DOTALL
+            r'<include(?:\s+([^>]*?))?(?<!/)>((?:(?!<include\b).)*?)</include>',
+            content,
+            re.DOTALL,
         )
         for attrs, body in single_matches:
             path_value: Optional[str] = None

tests/test_sync_order.py

@@ -89,6 +89,20 @@ def test_extract_includes_body_form_no_path_attr_uses_body(tmp_path):
     includes = sync_order.extract_includes_from_file(f)
     assert includes == {"docs/source.md"}
 
+def test_extract_includes_does_not_span_inline_documentation_token(tmp_path):
+    """An inline `<include>` mention must not consume a later real include."""
+    f = tmp_path / "test.prompt"
+    f.write_text(
+        "Use `<include>` tags to expand context.\n"
+        "<pdd.context>\n"
+        "  <include select=\"def:example\">context/example.py</include>\n"
+        "</pdd.context>\n",
+        encoding="utf-8",
+    )
+
+    includes = sync_order.extract_includes_from_file(f)
+    assert includes == {"context/example.py"}
+
 # ==============================================================================
 # Unit Tests: extract_module_from_include
 # ==============================================================================