todo_write: rename the items param to todos

A top-level arguments key named "items" shadows minijinja's .items() in Together's GLM-5.2 chat template — every replayed turn 400s.
Old key still executes; GUI renders both; regression tests added.

coworker/tools/todo.py

@@ -14,15 +14,20 @@ _STATUSES = {"pending", "in_progress", "done"}
 
 # Explicit schema — the array-of-objects shape can't be auto-generated reliably, and
 # providers reject a bare `list` annotation. Registered via `__coworker_schema__`.
+#
+# The parameter is `todos`, NOT `items`: a top-level argument key named "items" shadows
+# minijinja's `.items()` map method in at least one hosted chat template (Together's
+# GLM-5.2, 2026-07-21 — "object is not callable"), 400-ing every request that replays
+# the call. Any key name that isn't a minijinja map method is safe; never rename back.
 _TODO_SCHEMA = {
     "type": "function",
     "function": {
         "name": "todo_write",
-        "description": "Replace the task list. Provide the full list of items each call.",
+        "description": "Replace the task list. Provide the full list of todos each call.",
         "parameters": {
             "type": "object",
             "properties": {
-                "items": {
+                "todos": {
                     "type": "array",
                     "items": {
                         "type": "object",
@@ -37,7 +42,7 @@ _TODO_SCHEMA = {
                     },
                 }
             },
-            "required": ["items"],
+            "required": ["todos"],
         },
     },
 }
@@ -49,11 +54,12 @@ class TodoList:
 
 
 def todo_tools(todo: TodoList) -> list:
-    def todo_write(items: list) -> dict:
-        """Replace the task list. Each item is an object with `content` and a `status`
+    def todo_write(todos: list = None, items: list = None) -> dict:
+        """Replace the task list. Each todo is an object with `content` and a `status`
         of pending, in_progress, or done."""
+        # `items` stays accepted (models that free-style the old name; queued replays).
         normalized = []
-        for entry in items or []:
+        for entry in (todos if todos is not None else items) or []:
             if isinstance(entry, dict):
                 status = entry.get("status", "pending")
                 if status == "completed":  # common model alias for our "done"
@@ -67,7 +73,7 @@ def todo_tools(todo: TodoList) -> list:
             else:
                 normalized.append({"content": str(entry), "status": "pending"})
         todo.items = normalized
-        return {"count": len(normalized), "items": normalized}
+        return {"count": len(normalized), "todos": normalized}
 
     wrapped = ai.tool(
         todo_write,

surfaces/gui/src/App.tsx

@@ -554,7 +554,8 @@ export function App() {
           setStreaming(""); // finalized into items (or empty tool-only turn)
           break;
         case "tool_proposed":
-          if (d.name === "todo_write" && d.arguments?.items) setTodo(normalizeTodos(d.arguments.items));
+          if (d.name === "todo_write" && (d.arguments?.todos || d.arguments?.items))
+            setTodo(normalizeTodos(d.arguments.todos ?? d.arguments.items));
           setItems((p) => [
             ...p,
             { kind: "tool", id: newId(), name: d.name, args: d.arguments, status: "…" },

surfaces/gui/src/components/Transcript.test.tsx

@@ -195,9 +195,14 @@ describe("humanizeTool", () => {
   });
 
   it("summarizes todo_write by its single item and status", () => {
-    const line = humanizeTool("todo_write", { items: [{ content: "Post the digest", status: "in_progress" }] });
+    const line = humanizeTool("todo_write", { todos: [{ content: "Post the digest", status: "in_progress" }] });
     expect(line.pre).toBe("Updated the plan — ");
     expect(line.obj).toContain("Post the digest");
     expect(line.post).toBe(" → in progress");
   });
+
+  it("still renders pre-rename todo_write histories (legacy `items` key)", () => {
+    const line = humanizeTool("todo_write", { items: [{ content: "Old plan", status: "pending" }] });
+    expect(line.obj).toContain("Old plan");
+  });
 });

surfaces/gui/src/humanize.ts

@@ -56,7 +56,9 @@ export function humanizeTool(name: string, args: any): HumanLine {
     case "git_log":
       return { pre: "Looked through recent git history" };
     case "todo_write": {
-      const items = Array.isArray(a.items) ? a.items : [];
+      // `todos` is current; `items` renders histories from before the rename (the old
+      // key breaks Together's GLM-5.2 chat template — see coworker/tools/todo.py).
+      const items = Array.isArray(a.todos) ? a.todos : Array.isArray(a.items) ? a.items : [];
       if (items.length === 1) {
         const it = items[0] || {};
         const status = String(it.status || "").replace(/_/g, " ");

tests/test_todo_tool.py

@@ -0,0 +1,34 @@
+"""todo_write's wire contract.
+
+The parameter is `todos` — a top-level arguments key named "items" shadows minijinja's
+`.items()` map method in hosted chat templates (Together GLM-5.2, 2026-07-21) and 400s
+every request that replays the call. The old key stays accepted at execution time for
+models that free-style it, but must never reappear in the schema.
+"""
+
+from coworker.tools.todo import _TODO_SCHEMA, TodoList, todo_tools
+
+
+def _write(**kwargs):
+    todo = TodoList()
+    (spec,) = todo_tools(todo)
+    return spec(**kwargs), todo
+
+
+def test_schema_param_is_todos_not_items():
+    props = _TODO_SCHEMA["function"]["parameters"]["properties"]
+    assert "todos" in props
+    assert "items" not in props  # regression guard: see module docstring
+    assert _TODO_SCHEMA["function"]["parameters"]["required"] == ["todos"]
+
+
+def test_todos_key_writes_the_list():
+    result, todo = _write(todos=[{"content": "a", "status": "in_progress"}])
+    assert todo.items == [{"content": "a", "status": "in_progress"}]
+    assert result == {"count": 1, "todos": [{"content": "a", "status": "in_progress"}]}
+
+
+def test_legacy_items_key_still_executes():
+    result, todo = _write(items=[{"content": "b", "status": "done"}])
+    assert todo.items == [{"content": "b", "status": "done"}]
+    assert result["count"] == 1