What actually happens
when you run Claude Code?

class ExecutionMode(Enum):
    INTERACTIVE = "interactive"
    ONE_SHOT = "one_shot"
    BRIDGE = "bridge"

async def query(messages, tools):
    while True:
        response = await call_api(messages)
        tool_calls = extract_tools(response)
        if not tool_calls: break

class Tool(ABC):
    @property
    def name(self) -> str: ...
    @property
    def is_read_only(self) -> bool:
        return False  # fail-closed default

class MailboxAgent:
    messages: list  # own history
    budget: float   # own budget
    async def run(self, prompt) -> str:
        # parent only sees the return value

levels = [
    "/Library/.claude/",     # enterprise
    "~/.claude/CLAUDE.md",   # user
    "./CLAUDE.md",           # project
    "src/CLAUDE.md",         # subdirectory
]

class PermissionResult(Enum):
    ALLOW = "allow"  # proceed
    ASK = "ask"      # prompt user
    DENY = "deny"    # block + tell Claude

def create_store(initial):
    state = [initial]
    listeners = []
    def set_state(fn):
        state[0] = fn(state[0])
        for cb in listeners: cb()

---
# skill frontmatter
name: commit
tools: [Bash, Read, Grep]
model: sonnet
---
Create a git commit...

class MCPTool(Tool):
    def __init__(self, defn, conn):
        self._defn = defn
    async def call(self, args):
        return await self._conn.call_tool(...)

class BudgetTracker:
    continuation_count: int = 0
    def check(self) -> str:
        if self.used > threshold * 0.9:
            return "stop"

# Settings merge precedence:
# enterprise > user > project > local
def merge(*levels):
    result = {}
    for level in levels:
        result.update(level)

# Bridge mode: headless agent
# IDE ↔ replBridge ↔ Claude Code
# Web ↔ WebSocket ↔ Claude Code
# MCP ↔ stdio ↔ Claude Code