agentcode-cli 1.1.0__py3-none-any.whl → 1.2.2__py3-none-any.whl

agent.py

@@ -18,6 +18,8 @@ import litellm
 from rich.console import Console
 from rich.markdown import Markdown
 
+from xml_tool_parser import looks_like_xml_tool_call, parse_xml_tool_calls, strip_think
+
 from tools import TOOL_DEFINITIONS, execute_tool
 from router import ModelRouter, display_routing_decision
 from mcp_client import MCPManager
@@ -346,6 +348,18 @@ def run_agent_loop(
                 cost,
             )
 
+        # Open-weight fine-tunes (e.g. Vigp17/agentcode-27b) may emit tool
+        # calls as inline XML in the text response. Detect and convert before
+        # treating this as a pure text turn.
+        if not tool_calls_accum and looks_like_xml_tool_call(full_text):
+            cleaned, xml_calls = parse_xml_tool_calls(full_text)
+            if xml_calls:
+                full_text = cleaned
+                for i, tc in enumerate(xml_calls):
+                    tool_calls_accum[i] = tc
+        elif not tool_calls_accum:
+            full_text = strip_think(full_text)
+
         # Pure text response — done
         if not tool_calls_accum:
             conversation.messages.append({"role": "assistant", "content": full_text})

{agentcode_cli-1.1.0.dist-info → agentcode_cli-1.2.2.dist-info}/METADATA

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: agentcode-cli
-Version: 1.1.0
+Version: 1.2.2
 Summary: An open, multi-model agentic coding CLI — inspired by Claude Code
 Author: Vignesh Pai
 License: MIT License

agentcode_cli-1.2.2.dist-info/RECORD

@@ -0,0 +1,14 @@
+agent.py,sha256=lBZ8vsduMGt30E7VsPKbByA96Hum8B5pEIE_B7-QlrY,17784
+cli.py,sha256=FhaY-UCKQdmAcKeLw6Cvd2kDowh6yRlvmaKGfGECk80,25074
+mcp_client.py,sha256=2PviTqJtXM4UC_fsYLbAOAfWJvayWy7Q8VOQIIsDiqQ,6710
+router.py,sha256=IjomIOaLVmGHQG9rNgc3xNpGhr82rCnn5h8X-QvZhRU,11573
+server.py,sha256=KZhq4rYsKlRBth1g2fkqYHdCelJR4DD4-ZugC4jAgtw,10334
+settings.py,sha256=Qjc3tiVbT1cqIrnQW6m2UG8Xsvqsxl9qXPMTueqwn50,6903
+tools.py,sha256=MBYy0OeSIZjqyGOezaMvZ6AMW3WRDESU-2u48sifd0Q,24741
+xml_tool_parser.py,sha256=T62DigRj8VCH30dq5AnxV8hsJXf0f2vyubg5F4nzDDk,3138
+agentcode_cli-1.2.2.dist-info/licenses/LICENSE,sha256=BqTzyKKaSaVQoumXzhYCj1UgOSPCgvn-sxV6BIuT558,1068
+agentcode_cli-1.2.2.dist-info/METADATA,sha256=oRCW6r2goUhjL-nrryIi0i0Su-2A45r_128KbameN3Y,12709
+agentcode_cli-1.2.2.dist-info/WHEEL,sha256=aeYiig01lYGDzBgS8HxWXOg3uV61G9ijOsup-k9o1sk,91
+agentcode_cli-1.2.2.dist-info/entry_points.txt,sha256=xP_zeySufuVhL5v10_EqCooKRSxGYB8QuRVNUj4_m1E,39
+agentcode_cli-1.2.2.dist-info/top_level.txt,sha256=684sAEdroiDabB0I7Zud0vYWXcDbCl2WK0LjTQzxiUk,66
+agentcode_cli-1.2.2.dist-info/RECORD,,

{agentcode_cli-1.1.0.dist-info → agentcode_cli-1.2.2.dist-info}/top_level.txt

@@ -5,3 +5,4 @@ router
 server
 settings
 tools
+xml_tool_parser

cli.py

@@ -151,6 +151,8 @@ def handle_slash_command(
             config.model = arg
             if config.router:
                 config.router.enabled = False
+                config.router.default_model = arg
+                config.router.provider = config.router.detect_provider(arg)
             console.print(f"[success]✓ Switched to model: {arg} (routing disabled)[/success]")
         else:
             console.print(f"[info]Current model: {config.model}[/info]")

router.py

@@ -47,11 +47,20 @@ GEMINI_TIERS = [
     ModelTier("gemini/gemini-2.5-pro", "heavy", "Gemini 2.5 Pro", 1.25, 10.00),
 ]
 
+# Azure deployment names are user-defined, so tiers use common defaults.
+# Users can override via settings.json model.light/medium/heavy.
+AZURE_TIERS = [
+    ModelTier("azure/gpt-4o-mini", "light", "Azure GPT-4o Mini", 0.15, 0.60),
+    ModelTier("azure/gpt-4o", "medium", "Azure GPT-4o", 2.50, 10.00),
+    ModelTier("azure/gpt-4o", "heavy", "Azure GPT-4o", 2.50, 10.00),
+]
+
 # Provider configs keyed by prefix
 PROVIDER_TIERS = {
     "anthropic": ANTHROPIC_TIERS,
     "openai": OPENAI_TIERS,
     "gemini": GEMINI_TIERS,
+    "azure": AZURE_TIERS,
 }
 
 
@@ -203,7 +212,9 @@ class ModelRouter:
     def detect_provider(self, model_string: str) -> str:
         """Detect provider from a model string."""
         m = model_string.lower()
-        if "claude" in m or "anthropic" in m:
+        if m.startswith("azure/"):
+            return "azure"
+        elif "claude" in m or "anthropic" in m:
             return "anthropic"
         elif "gpt" in m or "openai" in m or "o1" in m or "o3" in m:
             return "openai"

server.py

@@ -12,6 +12,7 @@ from agent import (
     build_system_prompt, load_project_config, load_hooks,
     _run_subagents, _run_hook, _is_denied, _get_permission,
 )
+from xml_tool_parser import looks_like_xml_tool_call, parse_xml_tool_calls, strip_think
 
 
 # ── Wire protocol ─────────────────────────────────────────────────────────────
@@ -106,6 +107,18 @@ def _server_turn(
                         if tc.function.arguments:
                             tool_calls_accum[idx]["arguments"] += tc.function.arguments
 
+        # Some open-weight fine-tunes (e.g. Vigp17/agentcode-27b) emit tool
+        # calls as inline XML in the text response instead of using litellm's
+        # structured tool_calls field. Detect and convert.
+        if not tool_calls_accum and looks_like_xml_tool_call(full_text):
+            cleaned, xml_calls = parse_xml_tool_calls(full_text)
+            if xml_calls:
+                full_text = cleaned
+                for i, tc in enumerate(xml_calls):
+                    tool_calls_accum[i] = tc
+        elif not tool_calls_accum:
+            full_text = strip_think(full_text)
+
         if router and usage:
             try:
                 cost = litellm.completion_cost(

xml_tool_parser.py

@@ -0,0 +1,87 @@
+"""
+Parse Hermes/Qwen-3-style XML tool calls into OpenAI-compatible tool_calls.
+
+Some fine-tuned open-weight models (e.g. Vigp17/agentcode-27b) emit tool calls
+as inline XML in the assistant's text response instead of using the structured
+`tool_calls` field that litellm/OpenAI clients expect. This module detects that
+shape and converts it.
+
+Input shape we handle:
+
+    <think>...optional reasoning...</think>
+
+    <tool_call>
+    <function=tool_name>
+    <parameter=key1>value1</parameter>
+    <parameter=key2>value2</parameter>
+    </function>
+    </tool_call>
+
+Output: the same payload as litellm's structured tool_calls — a list of
+{"id", "name", "arguments"} dicts plus the cleaned text (think blocks stripped,
+tool_call blocks removed).
+"""
+
+import json
+import re
+import uuid
+
+
+_THINK_RE = re.compile(r"<think>.*?</think>\s*", re.DOTALL)
+# Qwen 3 / similar templates often consume the opening <think> tag during
+# prompt construction but leave the </think>. In that case everything from
+# the start of the response up to and including the first </think> is
+# reasoning that should be stripped.
+_OPEN_THINK_MISSING_RE = re.compile(r"\A.*?</think>\s*", re.DOTALL)
+_TOOL_CALL_RE = re.compile(r"<tool_call>(.*?)</tool_call>", re.DOTALL)
+_FUNCTION_RE = re.compile(r"<function=([^>]+)>(.*?)</function>", re.DOTALL)
+_PARAMETER_RE = re.compile(r"<parameter=([^>]+)>(.*?)</parameter>", re.DOTALL)
+
+
+def looks_like_xml_tool_call(text: str) -> bool:
+    """Cheap check before doing full regex parsing."""
+    return "<tool_call>" in text and "<function=" in text
+
+
+def strip_think(text: str) -> str:
+    """Remove <think>...</think> blocks so users don't see chain-of-thought.
+
+    Also handles the chat-template-ate-the-opening-tag case: if </think>
+    appears but no matching <think>, strip everything up to and including it.
+    """
+    text = _THINK_RE.sub("", text)
+    if "</think>" in text and "<think>" not in text:
+        text = _OPEN_THINK_MISSING_RE.sub("", text)
+    return text
+
+
+def parse_xml_tool_calls(text: str) -> tuple[str, list[dict]]:
+    """
+    Extract XML tool calls from text and return (cleaned_text, tool_calls).
+
+    cleaned_text has the <think> and <tool_call> blocks removed so it can be
+    streamed to the user as a normal assistant reply.
+
+    tool_calls is a list of dicts shaped like litellm's accumulator:
+        {"id": "call_xxx", "name": "tool_name", "arguments": '{"key": "value"}'}
+    """
+    cleaned = strip_think(text)
+    tool_calls: list[dict] = []
+
+    for tc_match in _TOOL_CALL_RE.finditer(cleaned):
+        body = tc_match.group(1)
+        for fn_match in _FUNCTION_RE.finditer(body):
+            name = fn_match.group(1).strip()
+            params_body = fn_match.group(2)
+            args = {
+                p.group(1).strip(): p.group(2).strip()
+                for p in _PARAMETER_RE.finditer(params_body)
+            }
+            tool_calls.append({
+                "id": f"call_{uuid.uuid4().hex[:8]}",
+                "name": name,
+                "arguments": json.dumps(args),
+            })
+
+    cleaned = _TOOL_CALL_RE.sub("", cleaned).strip()
+    return cleaned, tool_calls

agentcode_cli-1.1.0.dist-info/RECORD

@@ -1,13 +0,0 @@
-agent.py,sha256=zcYfTAarYhj1hdLc9DH5sWs6Qqqp1VpmrHGJfHSJZIQ,17118
-cli.py,sha256=lzZzJkfAJ7nzmTZqIkRnVfJg1SsfuhMwge9J7iX9xM4,24948
-mcp_client.py,sha256=2PviTqJtXM4UC_fsYLbAOAfWJvayWy7Q8VOQIIsDiqQ,6710
-router.py,sha256=kjPpF1js6mA7FN2x_NleyN-BLeVzRAy-Q0wbhjelQzM,11109
-server.py,sha256=w7EXPCsXHiSqfIXep3VIcS66t_ZY02ulF2Yn6e7N5TY,9654
-settings.py,sha256=Qjc3tiVbT1cqIrnQW6m2UG8Xsvqsxl9qXPMTueqwn50,6903
-tools.py,sha256=MBYy0OeSIZjqyGOezaMvZ6AMW3WRDESU-2u48sifd0Q,24741
-agentcode_cli-1.1.0.dist-info/licenses/LICENSE,sha256=BqTzyKKaSaVQoumXzhYCj1UgOSPCgvn-sxV6BIuT558,1068
-agentcode_cli-1.1.0.dist-info/METADATA,sha256=u7sKWqB7yfo66ODU2HGTAtEudTIR8dw1QKxkoPaE6AM,12709
-agentcode_cli-1.1.0.dist-info/WHEEL,sha256=aeYiig01lYGDzBgS8HxWXOg3uV61G9ijOsup-k9o1sk,91
-agentcode_cli-1.1.0.dist-info/entry_points.txt,sha256=xP_zeySufuVhL5v10_EqCooKRSxGYB8QuRVNUj4_m1E,39
-agentcode_cli-1.1.0.dist-info/top_level.txt,sha256=PQseaNK25xxImV2WLHRHRIkipStqhLdVWYmrccM5ln0,50
-agentcode_cli-1.1.0.dist-info/RECORD,,