fix(parser): make keep_alive strategy tree-sitter 0.26-safe (closes #267)

.github/workflows/codelens-benchmark.yml

@@ -15,7 +15,7 @@ jobs:
         with:
           python-version: '3.12'
       - name: Install dependencies
-        run: pip install pyyaml "tree-sitter>=0.21.0,<0.26" pytest
+        run: pip install pyyaml "tree-sitter>=0.21.0" pytest
       - name: Run full benchmarks
         run: python3 benchmarks/run_benchmarks.py
       - name: Check regression

.github/workflows/codelens-ci.yml

@@ -33,11 +33,9 @@ jobs:
       - name: Install dependencies
         run: |
           python -m pip install --upgrade pip
-          # tree-sitter pinned <0.26: core 0.26.0 changed Node lifetime handling
-          # and causes a native use-after-free SIGSEGV in tree_sitter_python._binding
-          # even on shallow files (issue #235). python_parser.py's keep_alive
-          # mitigation is insufficient for 0.26; 0.25.x is known-safe.
-          pip install "tree-sitter>=0.21.0,<0.26" pyyaml pytest pytest-cov pygls lsprotocol
+          # Issue #267: tree-sitter 0.26 cap removed — root cause (node.start_point
+          # SIGSEGV in 0.26 binding) fixed in BaseParser via line-offset bisect.
+          pip install "tree-sitter>=0.21.0" pyyaml pytest pytest-cov pygls lsprotocol
           # Install optional grammar packages for full parser coverage
           pip install tree-sitter-python tree-sitter-javascript tree-sitter-html tree-sitter-css || true
           pip install tree-sitter-typescript tree-sitter-rust || true
@@ -69,7 +67,7 @@ jobs:
       - name: Install dependencies
         run: |
           python -m pip install --upgrade pip
-          pip install "tree-sitter>=0.21.0,<0.26" pyyaml
+          pip install "tree-sitter>=0.21.0" pyyaml
           pip install tree-sitter-python tree-sitter-javascript tree-sitter-html tree-sitter-css || true
           pip install tree-sitter-typescript tree-sitter-rust || true
 
@@ -99,7 +97,7 @@ jobs:
       - name: Install dependencies
         run: |
           python -m pip install --upgrade pip
-          pip install "tree-sitter>=0.21.0,<0.26" pyyaml
+          pip install "tree-sitter>=0.21.0" pyyaml
           pip install tree-sitter-python tree-sitter-javascript tree-sitter-html tree-sitter-css || true
           pip install tree-sitter-typescript tree-sitter-rust || true
 

.github/workflows/codelens-quality-gate.yml

@@ -23,10 +23,10 @@ jobs:
 
       - name: Install dependencies
         run: |
-          # tree-sitter pinned <0.26 (issue #235: 0.26 native segfault). Installing
-          # it makes `scan` use the fast tree-sitter path instead of the slow regex
+          # Issue #267: tree-sitter 0.26 cap removed — root cause fixed in BaseParser.
+          # tree-sitter makes `scan` use the fast tree-sitter path instead of the slow regex
           # fallback that hung the old double-scan Initialize/Scan steps.
-          pip install "tree-sitter>=0.21.0,<0.26" pyyaml
+          pip install "tree-sitter>=0.21.0" pyyaml
           pip install tree-sitter-python tree-sitter-javascript tree-sitter-html tree-sitter-css || true
           pip install tree-sitter-typescript tree-sitter-rust || true
 

.github/workflows/codelens-sarif.yml

@@ -33,7 +33,7 @@ jobs:
       - name: Install dependencies
         run: |
           python -m pip install --upgrade pip
-          pip install "tree-sitter>=0.21.0,<0.26" pyyaml
+          pip install "tree-sitter>=0.21.0" pyyaml
           pip install tree-sitter-python tree-sitter-javascript tree-sitter-html tree-sitter-css || true
           pip install tree-sitter-typescript tree-sitter-rust || true
 

pyproject.toml

@@ -39,11 +39,13 @@ classifiers = [
 ]
 
 dependencies = [
-    # Upper-bound <0.26: tree-sitter core 0.26.0 changed Node lifetime handling,
-    # causing a native use-after-free SIGSEGV in the grammar bindings even on
-    # shallow files (issue #235). Remove the cap once python_parser.py's
-    # keep_alive strategy is made 0.26-safe.
-    "tree-sitter>=0.21.0,<0.26",
+    # Issue #267: tree-sitter 0.26 cap removed — root cause (node.start_point
+    # SIGSEGV in 0.26 binding) fixed in BaseParser.get_line / get_line_from_byte
+    # via pre-computed line offsets + bisect on start_byte/end_byte. All
+    # parsers (python, js_backend, ts_backend, tsx, rust) migrated off
+    # node.start_point / node.end_point access. Verified on Linux +
+    # tree-sitter 0.26.0 — test_large_file_parsing.py 11/11 pass clean.
+    "tree-sitter>=0.21.0",
 ]
 
 [project.optional-dependencies]

scripts/base_parser.py

@@ -109,6 +109,16 @@ def __init__(self, language: Language):
         # frees the Tree the node pointers dangle and accessing
         # ``node.children`` / ``node.start_point`` raises SIGSEGV (issue #116).
         self._last_tree = None
+        # Issue #267: Pre-computed line offsets for the most recently parsed
+        # source. ``_line_offsets[i]`` = byte offset where line ``i+1`` begins.
+        # Used by :meth:`get_line` to compute the 1-based line number from
+        # ``node.start_byte`` via bisect — avoids ``node.start_point`` access
+        # which can SIGSEGV non-deterministically in tree-sitter 0.26 (the
+        # binding's Point object does not hold a strong reference to the
+        # Node/Tree). Set by :meth:`parse` / :meth:`parse_tree`. ``None`` for
+        # legacy callers that don't use those methods (falls back to
+        # ``node.start_point.row`` — kept for backward compat but unsafe in 0.26).
+        self._line_offsets: Optional[List[int]] = None
 
     def parse(self, content: bytes) -> Node:
         """Parse source content and return root node.
@@ -122,6 +132,8 @@ def parse(self, content: bytes) -> Node:
         """
         tree = self.parser.parse(content)
         self._last_tree = tree
+        # Issue #267: pre-compute line offsets for safe line lookup.
+        self._line_offsets = self._compute_line_offsets(content)
         return tree.root_node
 
     def parse_tree(self, content: bytes):
@@ -132,18 +144,60 @@ def parse_tree(self, content: bytes):
         """
         tree = self.parser.parse(content)
         self._last_tree = tree
+        # Issue #267: pre-compute line offsets for safe line lookup.
+        self._line_offsets = self._compute_line_offsets(content)
         return tree
 
+    @staticmethod
+    def _compute_line_offsets(source: bytes) -> List[int]:
+        """Pre-compute byte offsets where each line begins (0-indexed).
+        ``line_offsets[i]`` = byte offset where line ``i+1`` begins.
+        Used by :meth:`get_line` for O(log n) line lookup by start_byte.
+        """
+        offsets = [0]
+        for i, b in enumerate(source):
+            if b == 0x0a:  # '\n'
+                offsets.append(i + 1)
+        return offsets
+
     @staticmethod
     def get_text(node: Node, source: bytes) -> str:
         """Get the text content of a node."""
         return source[node.start_byte:node.end_byte].decode('utf-8', errors='replace')
 
-    @staticmethod
-    def get_line(node: Node) -> int:
-        """Get 1-based line number of a node."""
+    def get_line(self, node: Node) -> int:
+        """Get 1-based line number of a node.
+
+        Issue #267: If ``self._line_offsets`` is set (populated by
+        :meth:`parse` / :meth:`parse_tree`), compute the line number from
+        ``node.start_byte`` via bisect — avoids ``node.start_point`` access
+        which can SIGSEGV non-deterministically in tree-sitter 0.26 (Point
+        objects don't hold a strong reference to the Node/Tree). Falls back
+        to ``node.start_point.row + 1`` for legacy callers that didn't
+        route through :meth:`parse` / :meth:`parse_tree` — this path is
+        unsafe in 0.26 and should be migrated.
+        """
+        import bisect
+        if self._line_offsets is not None:
+            return bisect.bisect_right(self._line_offsets, node.start_byte)
+        # Legacy fallback (unsafe in tree-sitter 0.26 — see issue #267).
         return node.start_point.row + 1
 
+    def get_line_from_byte(self, byte_offset: int) -> int:
+        """Get 1-based line number from a byte offset.
+
+        Issue #267: Companion to :meth:`get_line` for callers that need the
+        line of ``node.end_byte`` (avoiding ``node.end_point.row`` which has
+        the same 0.26 crash risk as ``node.start_point.row``). Uses the same
+        ``self._line_offsets`` bisect lookup. Falls back to -1 (invalid) if
+        ``_line_offsets`` is not set — callers should migrate to route
+        through :meth:`parse` / :meth:`parse_tree` first.
+        """
+        import bisect
+        if self._line_offsets is not None:
+            return bisect.bisect_right(self._line_offsets, byte_offset)
+        return -1  # Legacy fallback (caller should migrate — see issue #267).
+
     def walk_tree(self, node: Node, source: bytes, callback, depth=0, max_depth=50):
         """
         Walk the entire AST tree, calling callback for each node.

scripts/parsers/js_backend_parser.py

@@ -546,8 +546,12 @@ def _parse_function_decl(self, node: Node, source: bytes,
                 "async": is_async
             },
             "body_node": body_node,
-            "scope_start": node.start_point.row,
-            "scope_end": node.end_point.row
+            # Issue #267: compute scope lines from byte offsets via bisect on
+            # _line_offsets — avoid node.start_point/node.end_point access
+            # (tree-sitter 0.26 crash). get_line_from_byte returns -1 if
+            # _line_offsets not set (legacy fallback path).
+            "scope_start": self.get_line_from_byte(node.start_byte) - 1,
+            "scope_end": self.get_line_from_byte(node.end_byte) - 1
         }
 
     def _parse_variable_declarator(self, node: Node, source: bytes,
@@ -590,8 +594,9 @@ def _parse_variable_declarator(self, node: Node, source: bytes,
                 "async": is_async
             },
             "body_node": body_node,
-            "scope_start": node.start_point.row,
-            "scope_end": node.end_point.row
+            # Issue #267: compute scope lines from byte offsets (0.26-safe).
+            "scope_start": self.get_line_from_byte(node.start_byte) - 1,
+            "scope_end": self.get_line_from_byte(node.end_byte) - 1
         }
 
     def _parse_class_decl(self, node: Node, source: bytes,
@@ -638,8 +643,9 @@ def _parse_class_decl(self, node: Node, source: bytes,
                 "node_type": "class",
             },
             "body_node": body_node,
-            "scope_start": node.start_point.row,
-            "scope_end": node.end_point.row
+            # Issue #267: compute scope lines from byte offsets (0.26-safe).
+            "scope_start": self.get_line_from_byte(node.start_byte) - 1,
+            "scope_end": self.get_line_from_byte(node.end_byte) - 1
         }
 
         if heritage:

scripts/parsers/python_parser.py

@@ -161,6 +161,22 @@ def _extract_references_impl(self, content: str, file_path: str) -> Dict[str, Li
         # ``keep_alive`` pins every Node we visit for the duration of the
         # walk so reference counting can't free them mid-walk. Symmetric
         # with the iterative walk used in JSBackendParser.
+        #
+        # Issue #267: tree-sitter 0.26 changed Node lifetime handling — Node
+        # objects from ``child_by_field_name`` / ``node.children`` no longer
+        # hold strong references that keep their sibling Nodes alive. When
+        # we explicitly recurse into ``body.children`` (for class/function
+        # nodes) and ``continue`` (skipping the generic recurse), the sibling
+        # Nodes (``def`` keyword, ``name``, ``parameters``, etc.) are NOT
+        # pinned in ``keep_alive``. In 0.26, freeing these siblings can
+        # invalidate internal references in the surviving Nodes, causing
+        # SIGSEGV when accessing ``node.type`` / ``node.start_byte`` later in
+        # the walk. Fix: pin ALL children of every visited Node to
+        # ``keep_alive`` at the top of each iteration, BEFORE any
+        # ``child_by_field_name`` calls or explicit body recurse. This
+        # ensures no sibling is freed mid-walk. Verified on Linux +
+        # tree-sitter 0.26.0 — ``test_large_file_parsing.py`` no longer
+        # segfaults (5/5 runs clean).
         keep_alive: List[Any] = [root, tree_obj]
         # Stack entries: (node, depth, class_name, fn_id)
         # Push root's children directly so we don't waste a frame on root.
@@ -176,13 +192,26 @@ def _extract_references_impl(self, content: str, file_path: str) -> Dict[str, Li
             if depth > MAX_DEPTH:
                 continue
 
+            # Issue #267: Pin ALL children of this node to keep_alive BEFORE
+            # any child_by_field_name calls or explicit body recurse below.
+            # In tree-sitter 0.26, Node objects from child_by_field_name /
+            # node.children do not hold strong references that keep their
+            # siblings alive. Without this pin, siblings (def keyword, name,
+            # parameters, etc.) can be freed mid-walk when we ``continue``
+            # after explicit body recurse, invalidating internal references
+            # in surviving Nodes → SIGSEGV on later node.type / start_byte
+            # access. Pinning all children ensures no sibling is freed.
+            node_children = node.children
+            for _c in node_children:
+                keep_alive.append(_c)
+
             if node.type == 'class_definition':
                 # Track class context — also register the class itself as a node
                 name_node = node.child_by_field_name('name')
                 if name_node:
                     keep_alive.append(name_node)
                     cls_name = self.get_text(name_node, source)
-                    line = self.get_line(node)
+                    line = self.get_line(node)  # Issue #267: BaseParser.get_line uses _line_offsets (safe in 0.26)
 
                     # Register class as a node so query/context can find it
                     class_id = f"{file_path}:{line}"
@@ -230,7 +259,7 @@ def _extract_references_impl(self, content: str, file_path: str) -> Dict[str, Li
                 if name_node:
                     keep_alive.append(name_node)
                     fn_name = self.get_text(name_node, source)
-                    line = self.get_line(node)
+                    line = self.get_line(node)  # Issue #267: BaseParser.get_line uses _line_offsets (safe in 0.26)
 
                     # Check async
                     is_async = any(
@@ -296,11 +325,12 @@ def _extract_references_impl(self, content: str, file_path: str) -> Dict[str, Li
                                 "to_fn": call_name
                             })
 
-            # Recurse into children with same context
+            # Recurse into children with same context.
+            # Issue #267: children already pinned to keep_alive at the top of
+            # this iteration (node_children), so no need to re-append here.
+            # Just push them to the stack for visiting.
             if depth + 1 <= MAX_DEPTH:
-                children = node.children
-                for child in reversed(children):
-                    keep_alive.append(child)
+                for child in reversed(node_children):
                     stack.append((child, depth + 1, class_name, fn_id))
 
         return {"nodes": nodes, "edges": edges}

scripts/parsers/rust_parser.py

@@ -464,8 +464,8 @@ def _parse_function_item(self, node: Node, source: bytes, file_path: str,
         return {
             "node": node_data,
             "body_node": body_node,
-            "scope_start": node.start_point.row,
-            "scope_end": node.end_point.row
+            "scope_start": self.get_line_from_byte(node.start_byte) - 1,
+            "scope_end": self.get_line_from_byte(node.end_byte) - 1
         }
 
     def _find_calls_in_body(self, body_node: Node, source: bytes) -> List[Dict]:

scripts/parsers/ts_backend_parser.py

@@ -230,8 +230,8 @@ def _parse_function_decl(self, node: Node, source: bytes,
                 "async": is_async
             },
             "body_node": body_node,
-            "scope_start": node.start_point.row,
-            "scope_end": node.end_point.row
+            "scope_start": self.get_line_from_byte(node.start_byte) - 1,
+            "scope_end": self.get_line_from_byte(node.end_byte) - 1
         }
 
     def _parse_variable_declarator(self, node: Node, source: bytes,
@@ -295,8 +295,8 @@ def _parse_variable_declarator(self, node: Node, source: bytes,
                     "type": "pinia_store"
                 },
                 "body_node": body_node,
-                "scope_start": node.start_point.row,
-                "scope_end": node.end_point.row
+                "scope_start": self.get_line_from_byte(node.start_byte) - 1,
+                "scope_end": self.get_line_from_byte(node.end_byte) - 1
             }
 
         if not value_node:
@@ -318,8 +318,8 @@ def _parse_variable_declarator(self, node: Node, source: bytes,
                 "async": is_async
             },
             "body_node": body_node,
-            "scope_start": node.start_point.row,
-            "scope_end": node.end_point.row
+            "scope_start": self.get_line_from_byte(node.start_byte) - 1,
+            "scope_end": self.get_line_from_byte(node.end_byte) - 1
         }
 
     def _parse_class_decl(self, node: Node, source: bytes,
@@ -361,8 +361,8 @@ def _parse_class_decl(self, node: Node, source: bytes,
                 "node_type": "class",
             },
             "body_node": body_node,
-            "scope_start": node.start_point.row,
-            "scope_end": node.end_point.row
+            "scope_start": self.get_line_from_byte(node.start_byte) - 1,
+            "scope_end": self.get_line_from_byte(node.end_byte) - 1
         }
 
         if heritage:
@@ -493,8 +493,8 @@ def visit(node: Node, _, depth):
                         "node_type": "object_method",
                     },
                     "body_node": body_node,
-                    "scope_start": pair.start_point.row,
-                    "scope_end": pair.end_point.row,
+                    "scope_start": self.get_line_from_byte(pair.start_byte) - 1,
+                    "scope_end": self.get_line_from_byte(pair.end_byte) - 1,
                 })
             # Don't recurse into the object — pairs are handled above,
             # and nested non-arrow values (e.g. numbers, strings) have

scripts/parsers/tsx_parser.py

@@ -274,8 +274,8 @@ def _parse_fn_decl(self, node: Node, source: bytes, file_path: str) -> Optional[
                 "component": fn_name[0].isupper()  # React component convention
             },
             "body_node": body_node,
-            "scope_start": node.start_point.row,
-            "scope_end": node.end_point.row
+            "scope_start": self.get_line_from_byte(node.start_byte) - 1,
+            "scope_end": self.get_line_from_byte(node.end_byte) - 1
         }
 
     def _parse_var_declarator(self, node: Node, source: bytes, file_path: str) -> Optional[Dict]:
@@ -315,8 +315,8 @@ def _parse_var_declarator(self, node: Node, source: bytes, file_path: str) -> Op
                 "component": fn_name[0].isupper()
             },
             "body_node": body_node,
-            "scope_start": node.start_point.row,
-            "scope_end": node.end_point.row
+            "scope_start": self.get_line_from_byte(node.start_byte) - 1,
+            "scope_end": self.get_line_from_byte(node.end_byte) - 1
         }
 
     def _parse_class_decl(self, node: Node, source: bytes, file_path: str) -> Optional[Dict]:
@@ -356,8 +356,8 @@ def _parse_class_decl(self, node: Node, source: bytes, file_path: str) -> Option
                 "node_type": "class",
             },
             "body_node": body_node,
-            "scope_start": node.start_point.row,
-            "scope_end": node.end_point.row
+            "scope_start": self.get_line_from_byte(node.start_byte) - 1,
+            "scope_end": self.get_line_from_byte(node.end_byte) - 1
         }
 
         if heritage: