From f02ccd461bb41d392600fd8ab1b3475ae172fbf8 Mon Sep 17 00:00:00 2001
From: mottopanikeiku <176798723+mottopanikeiku@users.noreply.github.com>
Date: Sun, 25 Oct 2026 02:06:53 -0700
Subject: [PATCH] fix(glm45): emit the newline before every tool call after
 assistant text

The GLM-4.5 chat template opens every tool call with "\n<tool_call>", but when an assistant turn had text and two or more calls the renderer only emitted that newline before the first call. I emit it before each later call too and add a parity scenario with text plus two tool calls.

Assisted-by: Claude
Assisted-by: GPT-6 Astra
Signed-off-by: mottopanikeiku <176798723+mottopanikeiku@users.noreply.github.com>
---
 renderers/glm45.py |  6 ++++--
 tests/parity.py    | 29 +++++++++++++++++++++++++++--
 2 files changed, 31 insertions(+), 4 deletions(-)

diff --git a/renderers/glm45.py b/renderers/glm45.py
index cb967552d..e6b111d1e 100644
--- a/renderers/glm45.py
+++ b/renderers/glm45.py
@@ -517,12 +517,14 @@ def _render_assistant(
         elif content.strip():
             emit_text("\n" + content.strip(), msg_idx, is_sampled=True, is_content=True)
 
-        for tc in tool_calls:
+        for tc_idx, tc in enumerate(tool_calls):
             func = tc.get("function") or tc
             name = func.get("name", "")
             arguments = func.get("arguments", {})
 
-            if not content.strip():
+            # The template opens every call with "\n<tool_call>"; only the
+            # first call's "\n" is already carried by the content text above.
+            if tc_idx > 0 or not content.strip():
                 emit_text("\n", msg_idx, is_sampled=True, is_content=True)
             emit_special(self._tool_call_tok, msg_idx, is_sampled=True, is_content=True)
             emit_text(name + "\n", msg_idx, is_sampled=True, is_content=True)
diff --git a/tests/parity.py b/tests/parity.py
index ac34e11cc..ee42375bc 100644
--- a/tests/parity.py
+++ b/tests/parity.py
@@ -202,6 +202,7 @@ def _model(
             "tool-call-content",
             "tool-call-none",
             "multiple-tool-calls",
+            "multiple-tool-calls-content",
             "tool-response",
             "named-tool-response",
             "consecutive-tool-responses",
@@ -217,6 +218,7 @@ def _model(
             "tool-call-content",
             "tool-call-none",
             "multiple-tool-calls",
+            "multiple-tool-calls-content",
             "tool-response",
             "named-tool-response",
             "consecutive-tool-responses",
@@ -246,7 +248,11 @@ def _model(
     _model(
         "meta-llama/Llama-3.2-1B-Instruct",
         bridge=True,
-        excluded={"multiple-tool-calls", "consecutive-tool-responses"},
+        excluded={
+            "multiple-tool-calls",
+            "multiple-tool-calls-content",
+            "consecutive-tool-responses",
+        },
         oracle_defaults={"date_string": "26 Jul 2024"},
         extra_suites={"llama-checkpoints"},
     ),
@@ -254,7 +260,11 @@ def _model(
         "meta-llama/Llama-3.2-3B-Instruct",
         shared=False,
         roundtrip=False,
-        excluded={"multiple-tool-calls", "consecutive-tool-responses"},
+        excluded={
+            "multiple-tool-calls",
+            "multiple-tool-calls-content",
+            "consecutive-tool-responses",
+        },
         oracle_defaults={"date_string": "26 Jul 2024"},
         extra_suites={"llama-checkpoints"},
     ),
@@ -486,6 +496,21 @@ def _call(city: str, *, id: str | None = None) -> dict:
         ],
         tools=TOOLS,
     ),
+    _scenario(
+        "multiple-tool-calls-content",
+        [
+            {"role": "user", "content": "Weather in Paris and London?"},
+            {
+                "role": "assistant",
+                "content": "Let me check both.",
+                "tool_calls": [
+                    _call("Paris", id="call_1"),
+                    _call("London", id="call_2"),
+                ],
+            },
+        ],
+        tools=TOOLS,
+    ),
     _scenario(
         "tool-response",
         [