unreliable/broken tool calling & unescaped string args in chat_template.jinja

[mpetruc]

This might be related to #73.

The stock LFM2.5-8B-A1B chat_template.jinja (https://huggingface.co/LiquidAI/LFM2.5-8B-A1B/blob/main/chat_template.jinja) never tells the model what format to emit a tool call in. The system block it builds is only:

Today's date: ... \n\n List of tools: <JSON>

It never mentions <|tool_call_start|> / <|tool_call_end|> nor shows the pythonic list. On a fresh session (no prior assistant tool-call turn to
imitate), the model sometimes drifts to an OpenAI-style JSON object instead of the pythonic call, so the tool call is silently lost or the turn ends
with no structured output.

Repro: chat request with tools + "explore the repo, then call a tool", temperature ~0.8, empty history. The model sometimes emits something like

...reasoning... {"TaskCreate": {...}, "activeForm": ...}

instead of:

<|tool_call_start|>[TaskCreate(subject='...', description='...')]<|tool_call_end|>

Observed from vLLM and tabbyAPI (both load this same template). Severity is lowest on vLLM, the reference path where the sentinels are registered as special tokens; highest on providers that match the sentinels as plain text.

Second, independent bug: format_arg_value wraps string args in single quotes without escaping, so a value like "It's" renders the malformed

[send_message(text='It's fine')]

into the assistant tool-call history.

Suggested fix (applies cleanly; both changes validated with jinja2):

--- a/chat_template.jinja
+++ b/chat_template.jinja
@@ format_arg_value (escape quotes/backslashes before wrapping):
-        {{- "'" + arg_value + "'" -}}
+        {{- "'" + (arg_value | replace("\\", "\\\\") | replace("'", "\\'")) + "'" -}}
@@ tools system block (append a format directive after the tool list):
-    " + (tools | tojson) -%}
+    " + (tools | tojson)
+    + "\n\nTo call a tool, output EXACTLY one line wrapped in "
+      "<|tool_call_start|> and <|tool_call_end|>: "
+      "[function_name(arg='value', arg2=123)]. Never emit JSON. "
+      "If no tool is needed, answer in prose." -%}

I validated this change with LFM2.5-8B in hermes and pi. With both harnesses the model generated long chains of tool calls, thinking blocks and responses. It's working great!