mlx-code 0.0.1a2__tar.gz → 0.0.1a3__tar.gz

mlx_code-0.0.1a3/PKG-INFO

@@ -0,0 +1,204 @@
+Metadata-Version: 2.4
+Name: mlx-code
+Version: 0.0.1a3
+Summary: Local Claude Code for Mac
+Home-page: https://github.com/JosefAlbers/mlx-code
+Author: J Joe
+Author-email: albersj66@gmail.com
+License: Apache-2.0
+Requires-Python: >=3.11
+Description-Content-Type: text/markdown
+Requires-Dist: mlx-lm>=0.19.0
+Requires-Dist: PyYAML
+Dynamic: author
+Dynamic: author-email
+Dynamic: description
+Dynamic: description-content-type
+Dynamic: home-page
+Dynamic: license
+Dynamic: requires-dist
+Dynamic: requires-python
+Dynamic: summary
+
+# mlx-code: Local Claude Code for Mac
+
+<img width="1024" height="1536" alt="mlxcodelogo" src="https://github.com/user-attachments/assets/df605c1d-f8e7-433c-9aaf-176c73e3d628" />
+![mlxcode](https://github.com/user-attachments/assets/86004bd2-549f-44b8-9a05-a9b6d9f1c4a2)
+
+### Quick Start
+
+```bash
+brew install --cask claude-code
+pip install mlx-code
+```
+
+### Commands
+| Command | What it does | Example |
+|--------|--------------|--------|
+| `mlx-code` | Start interactive mode | `mlx-code` |
+| `mlx-code "task"` | Run a one-time task | `mlx-code "fix the build error"` |
+| `mlx-code -p "query"` | Run one-off query, then exit | `mlx-code -p "explain this function"` |
+| `mlx-code -c` | Continue most recent conversation in current directory | `mlx-code -c` |
+| `mlx-code -r` | Resume a previous conversation | `mlx-code -r` |
+| `mlx-code commit` | Create a Git commit | `mlx-code commit` |
+| `/clear` | Clear conversation history | `/clear` |
+| `/help` | Show available commands | `/help` |
+| `exit` or `Ctrl+C` | Exit Claude Code | `exit` |
+
+<details><summary>Click to expand</summary><pre>
+╭─── Claude Code v2.1.74 ──────────────────────────────────────────────────────────────────────────────────╮
+│                                                    │ Tips for getting started                            │
+│                    Welcome back!                   │ Run /init to create a CLAUDE.md file with instruct… │
+│                                                    │ ─────────────────────────────────────────────────── │
+│                       ▐▛███▜▌                      │ Recent activity                                     │
+│                      ▝▜█████▛▘                     │ No recent activity                                  │
+│                        ▘▘ ▝▝                       │                                                     │
+│                                                    │                                                     │
+│ mlx-community/Qwen3.5-4B-Opti… · API Usage Billing │                                                     │
+│                  ~/Downloads/tmp                   │                                                     │
+╰──────────────────────────────────────────────────────────────────────────────────────────────────────────╯
+
+❯ list the files in this directory and explain what each one does
+
+⏺ The user wants to understand the files in the current directory, so I'll use the Glob tool to list all files first.
+
+⏺ Searched for 1 pattern (ctrl+o to expand)
+
+⏺ I need to fix the Glob tool call - the parameter should be pattern, not key.
+
+⏺ Searched for 1 pattern (ctrl+o to expand)
+
+⏺ The user wants me to explain what each file does. I should organize these by project directory and explain the purpose of each file. Let me read a few key files to provide accurate explanations.
+
+  This directory contains several project folders. Let me explain each one:
+
+  Main Projects
+
+  cyberpunk/
+
+  - main.zig - Main entry point for a Zig-based cyberpunk project
+  - build.zig - Build configuration file for Zig
+  - map.c - C source file, likely for map rendering or game logic
+  - help.py - Python helper script, possibly for documentation or utilities
+
+  rts/
+
+  - main.zig - Main entry point for a real-time strategy (RTS) game in Zig
+  - build.zig - Build configuration for Zig
+  - battle_sim.c - C file for battle simulation logic
+
+  splat_zig_raylib/
+
+  - main.zig (two versions - one in splat_zig/, one in root) - Main entry point for a Raylib-based game
+  - build.zig - Build configuration
+  - gsplat_transform.py - Python script for splat transform calculations
+  - index.html - Web entry point for the game
+  - build_web.sh - Shell script for building web version
+  - README.md - Project documentation
+
+  tarot/
+
+  - tarot.py - Python script for tarot-related functionality
+  - index.html - Web entry point
+
+  yorha/
+
+  - index.html - Web entry point (likely a simple web page)
+
+  jina.py
+
+  - jina.py - Python script, possibly related to Jina AI or a utility
+
+  mlx-code-v1.py & mlx-code.py
+
+  - These appear to be different versions of the mlx-code agent wrapper (the local Claude Code-style agent for Apple Silicon)
+
+  tmp.md
+
+  - tmp.md - Temporary markdown file (contents unknown)
+
+  mlx_trace.log
+
+  - mlx_trace.log - Trace log file from the mlx model runtime
+
+  .DS_Store files
+
+  - macOS metadata files (can be ignored)
+
+  ---
+  Summary: This is a temporary download folder (/Users/jjms/Downloads/tmp) containing multiple small projects:
+  - Game projects in Zig (cyberpunk, RTS, Raylib-based game)
+  - Python scripts for various utilities
+  - Web entry points (HTML files)
+  - Build configurations (Zig build files)
+
+  The Zig projects appear to be the main focus, with game-related code and build systems.
+
+✻ Brewed for 42s
+
+❯ explain how output is parsed in mlx-code.py
+
+⏺ The user wants to understand how output is parsed in mlx-code.py, so I need to read that file first to understand its structure and find the output parsing logic.
+
+⏺ Read 1 file (ctrl+o to expand)
+
+⏺ I need to fix the Read tool call - it should use file_path not key.
+
+⏺ Read 1 file (ctrl+o to expand)
+
+⏺ Now I have the full file. Let me explain the output parsing logic, which is the key part the user asked about.
+
+
+  The output parsing logic is in the parse_output(raw: str) function (lines 233-315). Here's how it works:
+
+  Output Parsing Flow
+
+  1. Extract Thinking Blocks (lines 237-244)
+
+  First, it extracts any `
+
+⏺ The output parsing logic is in the parse_output(raw: str) function (lines 233-315). Here's how it works:
+
+  Output Parsing Flow
+
+  1. Extract Thinking Blocks (lines 237-244)
+
+  First, it extracts any `
+
+⏺ ` blocks by finding the earliest occurrences of both markers and splitting them out:
+  while "
+
+⏺ " in remaining:
+      s = remaining.index("
+
+⏺ ")
+      # Extract thinking block
+      remaining = remaining[e + 8:].strip()
+
+  ### 2. **Extract Tool Use Blocks** (lines 247-289)
+  It searches for `<tool_call>...</tool_call>` markers and tries to parse the content:
+  - **First attempt**: Try parsing as JSON with `name` or `tool` field
+  - **Fallback**: Parse as XML-like format with `<function=...>` and `<parameter=...>` tags
+  - Creates tool blocks with `type: "tool_use"`, unique ID, name, and input
+
+  ### 3. **Handle JSON Tool Response** (lines 291-311)
+  If no `<tool_call>...</tool_call>` blocks were found, it checks if the remaining output is a JSON object with `name` field (indicating a tool response).
+
+  ### 4. **Return Result** (line 315)
+  Returns a tuple of `(blocks, stop_reason)`:
+  - `blocks`: List of parsed blocks (`{"type": "text"}`, `{"type": "thinking"}`, `{"type": "tool_use"}`)
+  - `stop_reason`: `"tool_use"` if tools were detected, `"end_turn"` otherwise
+
+  ## Key Patterns Used
+
+  | Pattern | Purpose |
+  |---------|---------|
+  | `
+
+⏺ | Thinking block markers | |...| Tool invocation markers | | JSON withname/tool| Tool response format | |<function=...>/<parameter=...>` | Alternative tool format |
+
+  The parser is designed to handle both the tool invocation format (<tool_call>...</tool_call>) and tool response format (JSON with name field) used by the system.
+
+✻ Worked for 1m 9s
+</pre></details><br>
+

mlx_code-0.0.1a3/README.md

@@ -0,0 +1,182 @@
+# mlx-code: Local Claude Code for Mac
+
+<img width="1024" height="1536" alt="mlxcodelogo" src="https://github.com/user-attachments/assets/df605c1d-f8e7-433c-9aaf-176c73e3d628" />
+![mlxcode](https://github.com/user-attachments/assets/86004bd2-549f-44b8-9a05-a9b6d9f1c4a2)
+
+### Quick Start
+
+```bash
+brew install --cask claude-code
+pip install mlx-code
+```
+
+### Commands
+| Command | What it does | Example |
+|--------|--------------|--------|
+| `mlx-code` | Start interactive mode | `mlx-code` |
+| `mlx-code "task"` | Run a one-time task | `mlx-code "fix the build error"` |
+| `mlx-code -p "query"` | Run one-off query, then exit | `mlx-code -p "explain this function"` |
+| `mlx-code -c` | Continue most recent conversation in current directory | `mlx-code -c` |
+| `mlx-code -r` | Resume a previous conversation | `mlx-code -r` |
+| `mlx-code commit` | Create a Git commit | `mlx-code commit` |
+| `/clear` | Clear conversation history | `/clear` |
+| `/help` | Show available commands | `/help` |
+| `exit` or `Ctrl+C` | Exit Claude Code | `exit` |
+
+<details><summary>Click to expand</summary><pre>
+╭─── Claude Code v2.1.74 ──────────────────────────────────────────────────────────────────────────────────╮
+│                                                    │ Tips for getting started                            │
+│                    Welcome back!                   │ Run /init to create a CLAUDE.md file with instruct… │
+│                                                    │ ─────────────────────────────────────────────────── │
+│                       ▐▛███▜▌                      │ Recent activity                                     │
+│                      ▝▜█████▛▘                     │ No recent activity                                  │
+│                        ▘▘ ▝▝                       │                                                     │
+│                                                    │                                                     │
+│ mlx-community/Qwen3.5-4B-Opti… · API Usage Billing │                                                     │
+│                  ~/Downloads/tmp                   │                                                     │
+╰──────────────────────────────────────────────────────────────────────────────────────────────────────────╯
+
+❯ list the files in this directory and explain what each one does
+
+⏺ The user wants to understand the files in the current directory, so I'll use the Glob tool to list all files first.
+
+⏺ Searched for 1 pattern (ctrl+o to expand)
+
+⏺ I need to fix the Glob tool call - the parameter should be pattern, not key.
+
+⏺ Searched for 1 pattern (ctrl+o to expand)
+
+⏺ The user wants me to explain what each file does. I should organize these by project directory and explain the purpose of each file. Let me read a few key files to provide accurate explanations.
+
+  This directory contains several project folders. Let me explain each one:
+
+  Main Projects
+
+  cyberpunk/
+
+  - main.zig - Main entry point for a Zig-based cyberpunk project
+  - build.zig - Build configuration file for Zig
+  - map.c - C source file, likely for map rendering or game logic
+  - help.py - Python helper script, possibly for documentation or utilities
+
+  rts/
+
+  - main.zig - Main entry point for a real-time strategy (RTS) game in Zig
+  - build.zig - Build configuration for Zig
+  - battle_sim.c - C file for battle simulation logic
+
+  splat_zig_raylib/
+
+  - main.zig (two versions - one in splat_zig/, one in root) - Main entry point for a Raylib-based game
+  - build.zig - Build configuration
+  - gsplat_transform.py - Python script for splat transform calculations
+  - index.html - Web entry point for the game
+  - build_web.sh - Shell script for building web version
+  - README.md - Project documentation
+
+  tarot/
+
+  - tarot.py - Python script for tarot-related functionality
+  - index.html - Web entry point
+
+  yorha/
+
+  - index.html - Web entry point (likely a simple web page)
+
+  jina.py
+
+  - jina.py - Python script, possibly related to Jina AI or a utility
+
+  mlx-code-v1.py & mlx-code.py
+
+  - These appear to be different versions of the mlx-code agent wrapper (the local Claude Code-style agent for Apple Silicon)
+
+  tmp.md
+
+  - tmp.md - Temporary markdown file (contents unknown)
+
+  mlx_trace.log
+
+  - mlx_trace.log - Trace log file from the mlx model runtime
+
+  .DS_Store files
+
+  - macOS metadata files (can be ignored)
+
+  ---
+  Summary: This is a temporary download folder (/Users/jjms/Downloads/tmp) containing multiple small projects:
+  - Game projects in Zig (cyberpunk, RTS, Raylib-based game)
+  - Python scripts for various utilities
+  - Web entry points (HTML files)
+  - Build configurations (Zig build files)
+
+  The Zig projects appear to be the main focus, with game-related code and build systems.
+
+✻ Brewed for 42s
+
+❯ explain how output is parsed in mlx-code.py
+
+⏺ The user wants to understand how output is parsed in mlx-code.py, so I need to read that file first to understand its structure and find the output parsing logic.
+
+⏺ Read 1 file (ctrl+o to expand)
+
+⏺ I need to fix the Read tool call - it should use file_path not key.
+
+⏺ Read 1 file (ctrl+o to expand)
+
+⏺ Now I have the full file. Let me explain the output parsing logic, which is the key part the user asked about.
+
+
+  The output parsing logic is in the parse_output(raw: str) function (lines 233-315). Here's how it works:
+
+  Output Parsing Flow
+
+  1. Extract Thinking Blocks (lines 237-244)
+
+  First, it extracts any `
+
+⏺ The output parsing logic is in the parse_output(raw: str) function (lines 233-315). Here's how it works:
+
+  Output Parsing Flow
+
+  1. Extract Thinking Blocks (lines 237-244)
+
+  First, it extracts any `
+
+⏺ ` blocks by finding the earliest occurrences of both markers and splitting them out:
+  while "
+
+⏺ " in remaining:
+      s = remaining.index("
+
+⏺ ")
+      # Extract thinking block
+      remaining = remaining[e + 8:].strip()
+
+  ### 2. **Extract Tool Use Blocks** (lines 247-289)
+  It searches for `<tool_call>...</tool_call>` markers and tries to parse the content:
+  - **First attempt**: Try parsing as JSON with `name` or `tool` field
+  - **Fallback**: Parse as XML-like format with `<function=...>` and `<parameter=...>` tags
+  - Creates tool blocks with `type: "tool_use"`, unique ID, name, and input
+
+  ### 3. **Handle JSON Tool Response** (lines 291-311)
+  If no `<tool_call>...</tool_call>` blocks were found, it checks if the remaining output is a JSON object with `name` field (indicating a tool response).
+
+  ### 4. **Return Result** (line 315)
+  Returns a tuple of `(blocks, stop_reason)`:
+  - `blocks`: List of parsed blocks (`{"type": "text"}`, `{"type": "thinking"}`, `{"type": "tool_use"}`)
+  - `stop_reason`: `"tool_use"` if tools were detected, `"end_turn"` otherwise
+
+  ## Key Patterns Used
+
+  | Pattern | Purpose |
+  |---------|---------|
+  | `
+
+⏺ | Thinking block markers | |...| Tool invocation markers | | JSON withname/tool| Tool response format | |<function=...>/<parameter=...>` | Alternative tool format |
+
+  The parser is designed to handle both the tool invocation format (<tool_call>...</tool_call>) and tool response format (JSON with name field) used by the system.
+
+✻ Worked for 1m 9s
+</pre></details><br>
+

{mlx_code-0.0.1a2 → mlx_code-0.0.1a3}/main.py

@@ -1,5 +1,6 @@
 import argparse
 import json
+import logging
 import os
 import re
 import subprocess
@@ -19,27 +20,62 @@ DEFAULT_MODEL      = "mlx-community/Qwen3.5-4B-OptiQ-4bit"
 DEFAULT_SKILL_DIRS = ["./skills", os.path.expanduser("~/.claude/skills")]
 LOG_FILE           = "mlx_trace.log"
 
-model     = None
-tokenizer = None
-model_id  = None
-skills    = {}
-_call_counter = 0
-
-def trace(prompt: str, raw: str, elapsed: float):
-    global _call_counter
-    _call_counter += 1
-    sep = "=" * 80
-    entry = (
-        f"\n{sep}\n"
-        f"CALL {_call_counter}  {time.strftime('%Y-%m-%d %H:%M:%S')}  ({elapsed:.1f}s)\n"
-        f"{sep}\n"
-        f"--- PROMPT ---\n{prompt}\n"
-        f"--- OUTPUT ---\n{raw}\n"
-    )
-    with open(LOG_FILE, "a") as f:
-        f.write(entry)
-
-def parse_frontmatter(text):
+logging.basicConfig(
+    filename=LOG_FILE,
+    level=logging.DEBUG,
+    format="%(message)s",
+)
+logger = logging.getLogger(__name__)
+
+class AppState:
+    def __init__(self, model, tokenizer, model_id: str, skills: dict):
+        self.model     = model
+        self.tokenizer = tokenizer
+        self.model_id  = model_id
+        self.skills    = skills         
+        self._counter  = 0
+        self._lock     = threading.Lock()
+
+    def trace(self, prompt: str, raw: str, elapsed: float):
+        with self._lock:
+            self._counter += 1
+            count = self._counter
+        sep = "=" * 80
+        logger.debug(
+            "\n%s\nCALL %d  %s  (%.1fs)\n%s\n--- PROMPT ---\n%s\n--- OUTPUT ---\n%s\n",
+            sep, count, time.strftime("%Y-%m-%d %H:%M:%S"), elapsed, sep, prompt, raw,
+        )
+
+    def skill_body(self, name: str) -> str:
+        if name not in self.skills:
+            available = ", ".join(self.skills) or "none"
+            return f"Unknown skill '{name}'. Available: {available}"
+        try:
+            return Path(self.skills[name]["path"]).read_text()
+        except Exception as e:
+            return f"Error reading skill: {e}"
+
+    def generate(self, prompt: str, max_tokens: int, temp: float, top_p: float) -> str:
+        from mlx_lm import generate as mlx_gen
+        t0  = time.time()
+        raw = mlx_gen(
+            self.model, self.tokenizer,
+            prompt=prompt,
+            max_tokens=max_tokens,
+            verbose=False,
+        )
+        self.trace(prompt, raw, time.time() - t0)
+        return raw
+
+    def encode(self, text: str) -> list:
+        return self.tokenizer.encode(text)
+
+    def apply_chat_template(self, messages: list) -> str:
+        return self.tokenizer.apply_chat_template(
+            messages, tokenize=False, add_generation_prompt=True
+        )
+
+def parse_frontmatter(text: str):
     m = re.match(r"^---\n(.*?)\n---\n", text, re.DOTALL)
     if not m:
         return {}, text
@@ -57,7 +93,8 @@ def parse_frontmatter(text):
             fm[k.strip()] = v.strip()
     return fm, body
 
-def scan_skills(dirs):
+
+def scan_skills(dirs: list) -> dict:
     found = {}
     for d in dirs:
         p = Path(d)
@@ -75,37 +112,17 @@ def scan_skills(dirs):
                 print(f"  warn: {f}: {e}", flush=True)
     return found
 
-def skill_body(name):
-    if name not in skills:
-        return f"Unknown skill '{name}'. Available: {', '.join(skills) or 'none'}"
-    try:
-        return Path(skills[name]["path"]).read_text()
-    except Exception as e:
-        return f"Error reading skill: {e}"
+ANTI_LOOP_INSTRUCTION = (
+    "After you have written your response, stop immediately. "
+    "Do not re-read the question, do not re-draft the answer, "
+    "do not explain what you just did. Output your answer once and stop."
+)
 
-def load_model(path):
-    global model, tokenizer, model_id
-    from mlx_lm import load
-    print(f"Loading {path} …", flush=True)
-    model, tokenizer = load(path)
-    model_id = path
-    print("Ready.\n", flush=True)
 
-def generate(prompt, max_tokens, temp, top_p):
-    from mlx_lm import generate as mlx_gen
-    t0  = time.time()
-    raw = mlx_gen(
-        model, tokenizer,
-        prompt=prompt,
-        max_tokens=max_tokens,
-        verbose=False,
-    )
-    trace(prompt, raw, time.time() - t0)
-    return raw
-
-def skills_system_addon():
+def skills_system_addon(skills: dict) -> str:
+    base = ANTI_LOOP_INSTRUCTION
     if not skills:
-        return ""
+        return base
     entries = "\n".join(
         f"<skill><n>{n}</n><description>{s['description']}</description></skill>"
         for n, s in skills.items()
@@ -113,10 +130,12 @@ def skills_system_addon():
     return (
         "\n\n<available_skills>\n" + entries +
         "\nUse the read_skill tool to get full instructions before attempting "
-        "any task that matches a skill.\n</available_skills>"
+        "any task that matches a skill.\n</available_skills>\n\n"
+        + base
     )
 
-def tools_to_text(tools):
+
+def tools_to_text(tools: list) -> str:
     header = (
         "You have access to these tools. "
         "To call a tool output ONLY a <tool_call> block:\n"
@@ -131,7 +150,8 @@ def tools_to_text(tools):
         lines.append(f"- {name}({params}): {desc}")
     return "\n".join(lines)
 
-def build_messages(body, extra=None):
+
+def build_messages(body: dict, skills: dict, extra: list = None) -> list:
     msgs = []
 
     sys_parts = []
@@ -139,15 +159,16 @@ def build_messages(body, extra=None):
     if isinstance(raw, str) and raw:
         sys_parts.append(raw)
     elif isinstance(raw, list):
-        t = "\n".join(b.get("text","") for b in raw if b.get("type")=="text")
-        if t: sys_parts.append(t)
-    sys_parts.append(skills_system_addon())
+        t = "\n".join(b.get("text", "") for b in raw if b.get("type") == "text")
+        if t:
+            sys_parts.append(t)
+    sys_parts.append(skills_system_addon(skills))
 
     if skills:
         sys_parts.append(tools_to_text([{
             "name": "read_skill",
             "description": "Read the full SKILL.md instructions for a skill before using it.",
-            "input_schema": {"type":"object","properties":{"name":{"type":"string"}},"required":["name"]},
+            "input_schema": {"type": "object", "properties": {"name": {"type": "string"}}, "required": ["name"]},
         }]))
 
     system = "\n".join(p for p in sys_parts if p).strip()
@@ -166,7 +187,7 @@ def build_messages(body, extra=None):
             if t == "text":
                 parts.append(block["text"])
             elif t == "thinking":
-                parts.append(f"<think>\n{block.get('thinking','')}\n</think>")
+                parts.append(f"<think>\n{block.get('thinking', '')}\n</think>")
             elif t == "tool_use":
                 args = "".join(
                     f"<parameter={k}>\n{v}\n</parameter>"
@@ -176,7 +197,7 @@ def build_messages(body, extra=None):
             elif t == "tool_result":
                 rc = block.get("content", "")
                 if isinstance(rc, list):
-                    rc = "\n".join(c.get("text","") for c in rc if c.get("type")=="text")
+                    rc = "\n".join(c.get("text", "") for c in rc if c.get("type") == "text")
                 parts.append(f"<tool_response>\n{rc}\n</tool_response>")
         msgs.append({"role": role, "content": "\n".join(parts)})
 
@@ -184,22 +205,23 @@ def build_messages(body, extra=None):
         msgs.extend(extra)
     return msgs
 
-def build_prompt(body, extra=None):
-    msgs = build_messages(body, extra=extra)
-    return tokenizer.apply_chat_template(msgs, tokenize=False, add_generation_prompt=True)
 
-def parse_output(raw):
+def build_prompt(body: dict, state: AppState, extra: list = None) -> str:
+    msgs = build_messages(body, state.skills, extra=extra)
+    return state.apply_chat_template(msgs)
+
+def parse_output(raw: str):
     blocks    = []
     remaining = raw
 
     while "<think>" in remaining and "</think>" in remaining:
-        s = remaining.index("<think>")
-        e = remaining.index("</think>")
+        s      = remaining.index("<think>")
+        e      = remaining.index("</think>")
         before = remaining[:s].strip()
         if before:
             blocks.append({"type": "text", "text": before})
-        blocks.append({"type": "thinking", "thinking": remaining[s+7:e].strip()})
-        remaining = remaining[e+8:].strip()
+        blocks.append({"type": "thinking", "thinking": remaining[s + 7:e].strip()})
+        remaining = remaining[e + 8:].strip()
     remaining = remaining.replace("</think>", "").strip()
 
     tool_blocks = []
@@ -213,8 +235,10 @@ def parse_output(raw):
             name = obj.get("name") or obj.get("tool")
             args = obj.get("arguments") or obj.get("input") or {}
             if isinstance(args, str):
-                try:    args = json.loads(args)
-                except: args = {"raw": args}
+                try:
+                    args = json.loads(args)
+                except Exception:
+                    args = {"raw": args}
             if name:
                 parsed = {"name": name, "input": args}
         except Exception:
@@ -252,8 +276,10 @@ def parse_output(raw):
             args = obj.get("arguments") or obj.get("input") or {}
             if name:
                 if isinstance(args, str):
-                    try:    args = json.loads(args)
-                    except: args = {"raw": args}
+                    try:
+                        args = json.loads(args)
+                    except Exception:
+                        args = {"raw": args}
                 blocks.append({
                     "type": "tool_use",
                     "id":   f"toolu_{uuid.uuid4().hex[:8]}",
@@ -268,35 +294,44 @@ def parse_output(raw):
         blocks.append({"type": "text", "text": remaining})
     return blocks or [{"type": "text", "text": raw}], "end_turn"
 
-def resolve_read_skill(blocks, body, max_tokens, temp, top_p):
-    extra = []
+
+def resolve_read_skill(
+    blocks: list, body: dict, state: AppState,
+    max_tokens: int, temp: float, top_p: float,
+) -> tuple:
+    extra       = []
+    stop_reason = "tool_use"
+
     for _ in range(5):
-        skill_calls = [b for b in blocks if b.get("type")=="tool_use" and b["name"]=="read_skill"]
+        skill_calls = [b for b in blocks if b.get("type") == "tool_use" and b["name"] == "read_skill"]
         if not skill_calls:
             break
         for c in skill_calls:
-            name    = c["input"].get("name","")
-            content = skill_body(name)
+            name    = c["input"].get("name", "")
+            content = state.skill_body(name)
             args    = f"<parameter=name>\n{name}\n</parameter>"
             extra.append({"role": "assistant", "content": f"<tool_call>\n<function=read_skill>\n{args}</function>\n</tool_call>"})
-            extra.append({"role": "user", "content": f"<tool_response>\n{content}\n</tool_response>"})
-        prompt = build_prompt(body, extra=extra)
-        raw    = generate(prompt, max_tokens, temp, top_p)
+            extra.append({"role": "user",      "content": f"<tool_response>\n{content}\n</tool_response>"})
+        prompt = build_prompt(body, state, extra=extra)
+        raw    = state.generate(prompt, max_tokens, temp, top_p)
         blocks, stop_reason = parse_output(raw)
-    stop_reason = "tool_use" if any(b.get("type")=="tool_use" for b in blocks) else "end_turn"
+        if stop_reason == "end_turn":
+            break
+
     return blocks, stop_reason
 
-def sse(event, data):
+def sse(event: str, data: dict) -> bytes:
     return f"event: {event}\ndata: {json.dumps(data)}\n\n".encode()
 
-def blocks_to_sse(blocks, msg_id, stop_reason, in_tokens, out_tokens):
+
+def blocks_to_sse(blocks: list, msg_id: str, stop_reason: str, in_tokens: int, out_tokens: int) -> bytes:
     out = bytearray()
 
     out += sse("message_start", {
         "type": "message_start",
         "message": {
             "id": msg_id, "type": "message", "role": "assistant",
-            "model": model_id, "content": [], "stop_reason": None,
+            "model": "local", "content": [], "stop_reason": None,
             "stop_sequence": None,
             "usage": {"input_tokens": in_tokens, "output_tokens": 0},
         },
@@ -349,77 +384,89 @@ def blocks_to_sse(blocks, msg_id, stop_reason, in_tokens, out_tokens):
 
     return bytes(out)
 
-class Handler(BaseHTTPRequestHandler):
+def make_handler(state: AppState):
 
-    def log_message(self, fmt, *args):
-        pass  
+    class Handler(BaseHTTPRequestHandler):
 
-    def send_json(self, code, obj):
-        body = json.dumps(obj).encode()
-        self.send_response(code)
-        self.send_header("Content-Type", "application/json")
-        self.send_header("Content-Length", str(len(body)))
-        self.end_headers()
-        self.wfile.write(body)
+        def log_message(self, fmt, *args):
+            pass
 
-    def read_json(self):
-        n = int(self.headers.get("Content-Length", 0))
-        return json.loads(self.rfile.read(n))
+        def send_json(self, code: int, obj: dict):
+            body = json.dumps(obj).encode()
+            self.send_response(code)
+            self.send_header("Content-Type", "application/json")
+            self.send_header("Content-Length", str(len(body)))
+            self.end_headers()
+            self.wfile.write(body)
 
-    def path_base(self):
-        return self.path.split("?")[0].rstrip("/")
+        def read_json(self) -> dict:
+            n = int(self.headers.get("Content-Length", 0))
+            return json.loads(self.rfile.read(n))
 
-    def do_GET(self):
-        if self.path_base() == "/v1/models":
-            self.send_json(200, {"data": [
-                {"id": model_id, "object": "model",
-                 "created": int(time.time()), "owned_by": "local"},
-            ]})
-        else:
-            self.send_json(404, {"error": "not found"})
+        def path_base(self) -> str:
+            return self.path.split("?")[0].rstrip("/")
 
-    def do_POST(self):
-        pb = self.path_base()
+        def do_GET(self):
+            if self.path_base() == "/v1/models":
+                self.send_json(200, {"data": [
+                    {"id": state.model_id, "object": "model",
+                     "created": int(time.time()), "owned_by": "local"},
+                ]})
+            else:
+                self.send_json(404, {"error": "not found"})
 
-        if pb == "/v1/messages/count_tokens":
-            self.send_json(200, {"input_tokens": 0})
-            return
+        def do_POST(self):
+            pb = self.path_base()
 
-        if pb != "/v1/messages":
-            self.send_json(404, {"error": f"unknown endpoint {pb}"})
-            return
+            if pb == "/v1/messages/count_tokens":
+                self.send_json(200, {"input_tokens": 0})
+                return
 
-        body = self.read_json()
-        body["model"] = model_id
+            if pb != "/v1/messages":
+                self.send_json(404, {"error": f"unknown endpoint {pb}"})
+                return
 
-        max_tokens = body.get("max_tokens", 8192)
-        temp       = body.get("temperature", 0.7)
-        top_p      = body.get("top_p", 0.9)
-        msg_id     = f"msg_{uuid.uuid4().hex}"
+            body = self.read_json()
+            body["model"] = state.model_id
 
-        prompt = build_prompt(body)
-        raw    = generate(prompt, max_tokens, temp, top_p)
+            max_tokens = body.get("max_tokens", 8192)
+            temp       = body.get("temperature", 0.7)
+            top_p      = body.get("top_p", 0.9)
+            msg_id     = f"msg_{uuid.uuid4().hex}"
 
-        blocks, stop_reason = parse_output(raw)
+            prompt = build_prompt(body, state)
+            raw    = state.generate(prompt, max_tokens, temp, top_p)
 
-        if any(b.get("type")=="tool_use" and b["name"]=="read_skill" for b in blocks):
-            blocks, stop_reason = resolve_read_skill(blocks, body, max_tokens, temp, top_p)
+            blocks, stop_reason = parse_output(raw)
 
-        in_tokens  = len(tokenizer.encode(prompt))
-        out_tokens = len(tokenizer.encode(raw))
+            if any(b.get("type") == "tool_use" and b["name"] == "read_skill" for b in blocks):
+                blocks, stop_reason = resolve_read_skill(blocks, body, state, max_tokens, temp, top_p)
 
-        sse_bytes = blocks_to_sse(blocks, msg_id, stop_reason, in_tokens, out_tokens)
+            in_tokens  = len(state.encode(prompt))
+            out_tokens = len(state.encode(raw))
+
+            sse_bytes = blocks_to_sse(blocks, msg_id, stop_reason, in_tokens, out_tokens)
+
+            self.send_response(200)
+            self.send_header("Content-Type", "text/event-stream")
+            self.send_header("Cache-Control", "no-cache")
+            self.send_header("Content-Length", str(len(sse_bytes)))
+            self.end_headers()
+            try:
+                self.wfile.write(sse_bytes)
+                self.wfile.flush()
+            except BrokenPipeError:
+                pass
+
+    return Handler
+
+def load_model(path: str):
+    from mlx_lm import load
+    print(f"Loading {path} …", flush=True)
+    model, tokenizer = load(path)
+    print("Ready.\n", flush=True)
+    return model, tokenizer
 
-        self.send_response(200)
-        self.send_header("Content-Type", "text/event-stream")
-        self.send_header("Cache-Control", "no-cache")
-        self.send_header("Content-Length", str(len(sse_bytes)))
-        self.end_headers()
-        try:
-            self.wfile.write(sse_bytes)
-            self.wfile.flush()
-        except BrokenPipeError:
-            pass
 
 def main():
     parser = argparse.ArgumentParser()
@@ -433,9 +480,10 @@ def main():
     print("Scanning skills …", flush=True)
     skills = scan_skills(skill_dirs)
 
-    load_model(args.model)
+    model, tokenizer = load_model(args.model)
+    state = AppState(model, tokenizer, args.model, skills)
 
-    server = HTTPServer((args.host, args.port), Handler)
+    server = HTTPServer((args.host, args.port), make_handler(state))
     thread = threading.Thread(target=server.serve_forever, daemon=True)
     thread.start()
 
@@ -453,5 +501,6 @@ def main():
     result = subprocess.run(["claude"] + claude_args, env=env)
     sys.exit(result.returncode)
 
+
 if __name__ == "__main__":
     main()

mlx_code-0.0.1a3/mlx_code.egg-info/PKG-INFO

@@ -0,0 +1,204 @@
+Metadata-Version: 2.4
+Name: mlx-code
+Version: 0.0.1a3
+Summary: Local Claude Code for Mac
+Home-page: https://github.com/JosefAlbers/mlx-code
+Author: J Joe
+Author-email: albersj66@gmail.com
+License: Apache-2.0
+Requires-Python: >=3.11
+Description-Content-Type: text/markdown
+Requires-Dist: mlx-lm>=0.19.0
+Requires-Dist: PyYAML
+Dynamic: author
+Dynamic: author-email
+Dynamic: description
+Dynamic: description-content-type
+Dynamic: home-page
+Dynamic: license
+Dynamic: requires-dist
+Dynamic: requires-python
+Dynamic: summary
+
+# mlx-code: Local Claude Code for Mac
+
+<img width="1024" height="1536" alt="mlxcodelogo" src="https://github.com/user-attachments/assets/df605c1d-f8e7-433c-9aaf-176c73e3d628" />
+![mlxcode](https://github.com/user-attachments/assets/86004bd2-549f-44b8-9a05-a9b6d9f1c4a2)
+
+### Quick Start
+
+```bash
+brew install --cask claude-code
+pip install mlx-code
+```
+
+### Commands
+| Command | What it does | Example |
+|--------|--------------|--------|
+| `mlx-code` | Start interactive mode | `mlx-code` |
+| `mlx-code "task"` | Run a one-time task | `mlx-code "fix the build error"` |
+| `mlx-code -p "query"` | Run one-off query, then exit | `mlx-code -p "explain this function"` |
+| `mlx-code -c` | Continue most recent conversation in current directory | `mlx-code -c` |
+| `mlx-code -r` | Resume a previous conversation | `mlx-code -r` |
+| `mlx-code commit` | Create a Git commit | `mlx-code commit` |
+| `/clear` | Clear conversation history | `/clear` |
+| `/help` | Show available commands | `/help` |
+| `exit` or `Ctrl+C` | Exit Claude Code | `exit` |
+
+<details><summary>Click to expand</summary><pre>
+╭─── Claude Code v2.1.74 ──────────────────────────────────────────────────────────────────────────────────╮
+│                                                    │ Tips for getting started                            │
+│                    Welcome back!                   │ Run /init to create a CLAUDE.md file with instruct… │
+│                                                    │ ─────────────────────────────────────────────────── │
+│                       ▐▛███▜▌                      │ Recent activity                                     │
+│                      ▝▜█████▛▘                     │ No recent activity                                  │
+│                        ▘▘ ▝▝                       │                                                     │
+│                                                    │                                                     │
+│ mlx-community/Qwen3.5-4B-Opti… · API Usage Billing │                                                     │
+│                  ~/Downloads/tmp                   │                                                     │
+╰──────────────────────────────────────────────────────────────────────────────────────────────────────────╯
+
+❯ list the files in this directory and explain what each one does
+
+⏺ The user wants to understand the files in the current directory, so I'll use the Glob tool to list all files first.
+
+⏺ Searched for 1 pattern (ctrl+o to expand)
+
+⏺ I need to fix the Glob tool call - the parameter should be pattern, not key.
+
+⏺ Searched for 1 pattern (ctrl+o to expand)
+
+⏺ The user wants me to explain what each file does. I should organize these by project directory and explain the purpose of each file. Let me read a few key files to provide accurate explanations.
+
+  This directory contains several project folders. Let me explain each one:
+
+  Main Projects
+
+  cyberpunk/
+
+  - main.zig - Main entry point for a Zig-based cyberpunk project
+  - build.zig - Build configuration file for Zig
+  - map.c - C source file, likely for map rendering or game logic
+  - help.py - Python helper script, possibly for documentation or utilities
+
+  rts/
+
+  - main.zig - Main entry point for a real-time strategy (RTS) game in Zig
+  - build.zig - Build configuration for Zig
+  - battle_sim.c - C file for battle simulation logic
+
+  splat_zig_raylib/
+
+  - main.zig (two versions - one in splat_zig/, one in root) - Main entry point for a Raylib-based game
+  - build.zig - Build configuration
+  - gsplat_transform.py - Python script for splat transform calculations
+  - index.html - Web entry point for the game
+  - build_web.sh - Shell script for building web version
+  - README.md - Project documentation
+
+  tarot/
+
+  - tarot.py - Python script for tarot-related functionality
+  - index.html - Web entry point
+
+  yorha/
+
+  - index.html - Web entry point (likely a simple web page)
+
+  jina.py
+
+  - jina.py - Python script, possibly related to Jina AI or a utility
+
+  mlx-code-v1.py & mlx-code.py
+
+  - These appear to be different versions of the mlx-code agent wrapper (the local Claude Code-style agent for Apple Silicon)
+
+  tmp.md
+
+  - tmp.md - Temporary markdown file (contents unknown)
+
+  mlx_trace.log
+
+  - mlx_trace.log - Trace log file from the mlx model runtime
+
+  .DS_Store files
+
+  - macOS metadata files (can be ignored)
+
+  ---
+  Summary: This is a temporary download folder (/Users/jjms/Downloads/tmp) containing multiple small projects:
+  - Game projects in Zig (cyberpunk, RTS, Raylib-based game)
+  - Python scripts for various utilities
+  - Web entry points (HTML files)
+  - Build configurations (Zig build files)
+
+  The Zig projects appear to be the main focus, with game-related code and build systems.
+
+✻ Brewed for 42s
+
+❯ explain how output is parsed in mlx-code.py
+
+⏺ The user wants to understand how output is parsed in mlx-code.py, so I need to read that file first to understand its structure and find the output parsing logic.
+
+⏺ Read 1 file (ctrl+o to expand)
+
+⏺ I need to fix the Read tool call - it should use file_path not key.
+
+⏺ Read 1 file (ctrl+o to expand)
+
+⏺ Now I have the full file. Let me explain the output parsing logic, which is the key part the user asked about.
+
+
+  The output parsing logic is in the parse_output(raw: str) function (lines 233-315). Here's how it works:
+
+  Output Parsing Flow
+
+  1. Extract Thinking Blocks (lines 237-244)
+
+  First, it extracts any `
+
+⏺ The output parsing logic is in the parse_output(raw: str) function (lines 233-315). Here's how it works:
+
+  Output Parsing Flow
+
+  1. Extract Thinking Blocks (lines 237-244)
+
+  First, it extracts any `
+
+⏺ ` blocks by finding the earliest occurrences of both markers and splitting them out:
+  while "
+
+⏺ " in remaining:
+      s = remaining.index("
+
+⏺ ")
+      # Extract thinking block
+      remaining = remaining[e + 8:].strip()
+
+  ### 2. **Extract Tool Use Blocks** (lines 247-289)
+  It searches for `<tool_call>...</tool_call>` markers and tries to parse the content:
+  - **First attempt**: Try parsing as JSON with `name` or `tool` field
+  - **Fallback**: Parse as XML-like format with `<function=...>` and `<parameter=...>` tags
+  - Creates tool blocks with `type: "tool_use"`, unique ID, name, and input
+
+  ### 3. **Handle JSON Tool Response** (lines 291-311)
+  If no `<tool_call>...</tool_call>` blocks were found, it checks if the remaining output is a JSON object with `name` field (indicating a tool response).
+
+  ### 4. **Return Result** (line 315)
+  Returns a tuple of `(blocks, stop_reason)`:
+  - `blocks`: List of parsed blocks (`{"type": "text"}`, `{"type": "thinking"}`, `{"type": "tool_use"}`)
+  - `stop_reason`: `"tool_use"` if tools were detected, `"end_turn"` otherwise
+
+  ## Key Patterns Used
+
+  | Pattern | Purpose |
+  |---------|---------|
+  | `
+
+⏺ | Thinking block markers | |...| Tool invocation markers | | JSON withname/tool| Tool response format | |<function=...>/<parameter=...>` | Alternative tool format |
+
+  The parser is designed to handle both the tool invocation format (<tool_call>...</tool_call>) and tool response format (JSON with name field) used by the system.
+
+✻ Worked for 1m 9s
+</pre></details><br>
+

{mlx_code-0.0.1a2 → mlx_code-0.0.1a3}/mlx_code.egg-info/requires.txt

@@ -1 +1,2 @@
 mlx-lm>=0.19.0
+PyYAML

{mlx_code-0.0.1a2 → mlx_code-0.0.1a3}/setup.py

@@ -6,13 +6,13 @@ setup(
     author_email="albersj66@gmail.com",
     author="J Joe",
     license="Apache-2.0",
-    version="0.0.1a2",
+    version="0.0.1a3",
     readme="README.md",
     description="Local Claude Code for Mac",
     long_description=open("README.md").read(),
     long_description_content_type="text/markdown",
     python_requires=">=3.11",
-    install_requires=["mlx-lm>=0.19.0"],
+    install_requires=["mlx-lm>=0.19.0", "PyYAML"],
     py_modules=["main"],
     entry_points={"console_scripts": ["mlx-code=main:main"]},
 )

mlx_code-0.0.1a2/PKG-INFO

@@ -1,24 +0,0 @@
-Metadata-Version: 2.4
-Name: mlx-code
-Version: 0.0.1a2
-Summary: Local Claude Code for Mac
-Home-page: https://github.com/JosefAlbers/mlx-code
-Author: J Joe
-Author-email: albersj66@gmail.com
-License: Apache-2.0
-Requires-Python: >=3.11
-Description-Content-Type: text/markdown
-Requires-Dist: mlx-lm>=0.19.0
-Dynamic: author
-Dynamic: author-email
-Dynamic: description
-Dynamic: description-content-type
-Dynamic: home-page
-Dynamic: license
-Dynamic: requires-dist
-Dynamic: requires-python
-Dynamic: summary
-
-# mlx-code: Local Claude Code for Mac
-
-![mlxcode](https://github.com/user-attachments/assets/86004bd2-549f-44b8-9a05-a9b6d9f1c4a2)

mlx_code-0.0.1a2/README.md

@@ -1,3 +0,0 @@
-# mlx-code: Local Claude Code for Mac
-
-![mlxcode](https://github.com/user-attachments/assets/86004bd2-549f-44b8-9a05-a9b6d9f1c4a2)

mlx_code-0.0.1a2/mlx_code.egg-info/PKG-INFO

@@ -1,24 +0,0 @@
-Metadata-Version: 2.4
-Name: mlx-code
-Version: 0.0.1a2
-Summary: Local Claude Code for Mac
-Home-page: https://github.com/JosefAlbers/mlx-code
-Author: J Joe
-Author-email: albersj66@gmail.com
-License: Apache-2.0
-Requires-Python: >=3.11
-Description-Content-Type: text/markdown
-Requires-Dist: mlx-lm>=0.19.0
-Dynamic: author
-Dynamic: author-email
-Dynamic: description
-Dynamic: description-content-type
-Dynamic: home-page
-Dynamic: license
-Dynamic: requires-dist
-Dynamic: requires-python
-Dynamic: summary
-
-# mlx-code: Local Claude Code for Mac
-
-![mlxcode](https://github.com/user-attachments/assets/86004bd2-549f-44b8-9a05-a9b6d9f1c4a2)