fix(parsing): schema-aware tool-arg coercion for XML-style parsers (#28)

Author: hallerite

renderers/base.py

@@ -297,8 +297,23 @@ def render_ids(
         """Render messages to token IDs (without attribution metadata)."""
         ...
 
-    def parse_response(self, token_ids: list[int]) -> ParsedResponse:
-        """Parse completion tokens back into a structured message."""
+    def parse_response(
+        self,
+        token_ids: list[int],
+        *,
+        tools: list[ToolSpec] | None = None,
+    ) -> ParsedResponse:
+        """Parse completion tokens back into a structured message.
+
+        ``tools`` is the same list passed to ``render`` for this turn.
+        XML-style formats (Qwen3.5, GLM, MiniMax, Laguna) render argument
+        values verbatim inside ``<arg_value>`` tags with no quoting, so
+        a value like ``true`` is ambiguous between bool and the string
+        ``"true"``. When ``tools`` is supplied, the parser consults each
+        parameter's declared JSON-schema type to preserve string args
+        verbatim. Without ``tools``, parsers fall back to the historical
+        ``json.loads``-with-text-fallback behavior.
+        """
         ...
 
     def get_stop_token_ids(self) -> list[int]:

renderers/client.py

@@ -150,7 +150,7 @@ def _prepare():
     completion_ids = choice.get("token_ids") or []
 
     parsed = await _maybe_offload(
-        renderer, lambda: renderer.parse_response(completion_ids)
+        renderer, lambda: renderer.parse_response(completion_ids, tools=tools)
     )
 
     # ChatCompletionLogProbs flatten: {"content": [{"logprob": ...}, ...]}

renderers/deepseek_v3.py

@@ -221,7 +221,12 @@ def render_ids(
             add_generation_prompt=add_generation_prompt,
         ).token_ids
 
-    def parse_response(self, token_ids: list[int]) -> ParsedResponse:
+    def parse_response(
+        self,
+        token_ids: list[int],
+        *,
+        tools: list[ToolSpec] | None = None,  # noqa: ARG002 — args land in a ```json fence, schema not needed
+    ) -> ParsedResponse:
         return parse_deepseek_v3(
             self._tokenizer,
             token_ids,

renderers/default.py

@@ -167,7 +167,12 @@ def render_ids(
             messages, tools=tools, add_generation_prompt=add_generation_prompt
         )
 
-    def parse_response(self, token_ids: list[int]) -> ParsedResponse:
+    def parse_response(
+        self,
+        token_ids: list[int],
+        *,
+        tools: list[ToolSpec] | None = None,  # noqa: ARG002 — DefaultRenderer relies on configured tool_parser, schema not consulted here
+    ) -> ParsedResponse:
         # 1. Extract tool calls while we still have token ids (most formats
         #    use special-token delimiters, so id-level matching is reliable).
         if self._tool_parser is not None:

renderers/glm45.py

@@ -214,7 +214,12 @@ def render_ids(
             add_generation_prompt=add_generation_prompt,
         ).token_ids
 
-    def parse_response(self, token_ids: list[int]) -> ParsedResponse:
+    def parse_response(
+        self,
+        token_ids: list[int],
+        *,
+        tools: list[ToolSpec] | None = None,
+    ) -> ParsedResponse:
         return parse_glm(
             self._tokenizer,
             token_ids,
@@ -227,6 +232,7 @@ def parse_response(self, token_ids: list[int]) -> ParsedResponse:
             arg_key_end_id=self._arg_key_end,
             arg_value_id=self._arg_value,
             arg_value_end_id=self._arg_value_end,
+            tools=tools,
         )
 
     def get_stop_token_ids(self) -> list[int]:

renderers/glm5.py

@@ -226,7 +226,12 @@ def render_ids(
             add_generation_prompt=add_generation_prompt,
         ).token_ids
 
-    def parse_response(self, token_ids: list[int]) -> ParsedResponse:
+    def parse_response(
+        self,
+        token_ids: list[int],
+        *,
+        tools: list[ToolSpec] | None = None,
+    ) -> ParsedResponse:
         return parse_glm(
             self._tokenizer,
             token_ids,
@@ -239,6 +244,7 @@ def parse_response(self, token_ids: list[int]) -> ParsedResponse:
             arg_key_end_id=self._arg_key_end,
             arg_value_id=self._arg_value,
             arg_value_end_id=self._arg_value_end,
+            tools=tools,
         )
 
     def get_stop_token_ids(self) -> list[int]:

renderers/gpt_oss.py

@@ -307,7 +307,12 @@ def render_ids(
             add_generation_prompt=add_generation_prompt,
         ).token_ids
 
-    def parse_response(self, token_ids: list[int]) -> ParsedResponse:
+    def parse_response(
+        self,
+        token_ids: list[int],
+        *,
+        tools: list[ToolSpec] | None = None,  # noqa: ARG002 — harmony args land in a JSON object, schema not needed
+    ) -> ParsedResponse:
         return parse_gpt_oss(
             self._tokenizer,
             token_ids,

renderers/kimi_k2.py

@@ -279,7 +279,12 @@ def render_ids(
             add_generation_prompt=add_generation_prompt,
         ).token_ids
 
-    def parse_response(self, token_ids: list[int]) -> ParsedResponse:
+    def parse_response(
+        self,
+        token_ids: list[int],
+        *,
+        tools: list[ToolSpec] | None = None,  # noqa: ARG002 — section-JSON wire format quotes strings, schema not needed
+    ) -> ParsedResponse:
         return parse_kimi_k2(
             self._tokenizer,
             token_ids,

renderers/kimi_k25.py

@@ -904,7 +904,12 @@ def render_ids(
             add_generation_prompt=add_generation_prompt,
         ).token_ids
 
-    def parse_response(self, token_ids: list[int]) -> ParsedResponse:
+    def parse_response(
+        self,
+        token_ids: list[int],
+        *,
+        tools: list[ToolSpec] | None = None,  # noqa: ARG002 — section-JSON wire format quotes strings, schema not needed
+    ) -> ParsedResponse:
         stop_ids: set[int] = {self._im_end}
         if self._endoftext is not None:
             stop_ids.add(self._endoftext)

renderers/laguna_xs2.py

@@ -238,7 +238,12 @@ def render_ids(
             add_generation_prompt=add_generation_prompt,
         ).token_ids
 
-    def parse_response(self, token_ids: list[int]) -> ParsedResponse:
+    def parse_response(
+        self,
+        token_ids: list[int],
+        *,
+        tools: list[ToolSpec] | None = None,
+    ) -> ParsedResponse:
         return parse_laguna_xs2(
             self._tokenizer,
             token_ids,
@@ -247,6 +252,7 @@ def parse_response(self, token_ids: list[int]) -> ParsedResponse:
             think_end_id=self._think_end,
             tool_call_id=self._tool_call,
             tool_call_end_id=self._tool_call_end,
+            tools=tools,
         )
 
     def get_stop_token_ids(self) -> list[int]: