feat(hooks): add BeforeToolCallback to intercept tool calls before execution

[ousamabenyounes]

Summary

Adds a before_tool_callback hook to AgentLoop that lets callers intercept every tool invocation before execution and optionally rewrite the arguments.

Motivation: External tools like RTK (Rust Token Killer) need to transparently rewrite shell commands before they run (e.g. git status → rtk git status). Claude Code and Gemini CLI both provide equivalent hooks (PreToolUse and BeforeTool). This PR brings Vibe to parity. Tracked in rtk-ai/rtk#800.

API

async def before_bash_rewrite(
    tool_name: str, args: dict[str, Any]
) -> dict[str, Any] | None:
    if tool_name == "bash":
        cmd = args.get("command", "")
        return {**args, "command": rewrite_to_rtk(cmd)}
    return None  # passthrough for other tools

agent_loop.set_before_tool_callback(before_bash_rewrite)

The callback receives (tool_name, args_dict) and may:

• Return None → args unchanged (passthrough)
• Return a modified dict → the tool receives the rewritten args

Changes

File	What changed
vibe/core/types.py	New BeforeToolCallback type alias
vibe/core/tools/base.py	before_tool_callback field in InvokeContext
vibe/core/agent_loop.py	Callback invocation before invoke(), set_before_tool_callback() public method
tests/test_before_tool_callback.py	5 new tests

Design notes

• Follows the existing set_approval_callback / set_user_input_callback pattern exactly
• Zero impact on callers that don't set the callback (None check guards every path)
• Applied after permission checks (tool is already approved), before args are passed to run()
• Async to allow I/O in the callback (e.g. subprocess rewrite logic)

Test plan

• test_callback_receives_tool_name_and_args — correct inputs to callback
• test_callback_none_leaves_args_unchanged — passthrough behaviour
• test_callback_can_rewrite_args — rewrite reaches the tool
• test_no_callback_tool_executes_normally — no regression without callback
• test_rewritten_args_produce_successful_tool_result — no errors with rewritten args
• Full existing test suite: 345 passed, 0 regressions

🤖 Generated with Claude Code