Async REPL conversion with prompt

Async REPL conversion with prompt_async | Generated by AI

Home 2026.06

Here’s the full breakdown of commit f148313:

Title: refactor: convert REPL loop to async with prompt_async

What changed and why:

The iclaw REPL was using session.prompt() (synchronous) from prompt_toolkit. This blocks the thread while waiting for user input, which is fine for simple CLIs but limits future extensibility — e.g., running background tasks, handling async tool callbacks, or integrating with async-native libraries.

iclaw/main.py (2 changes):

Added import asyncio at the top.
Split main() into two functions:
- main() — thin wrapper that calls asyncio.run(_main())
- async _main() — the actual REPL logic, now async
Changed session.prompt("> ") → await session.prompt_async("> ")

The rest of the REPL logic (tool-call loops, token refresh, command handling) remains synchronous inside the async function — await is only on the prompt call for now. This is a stepping stone: the async shell is in place, and individual operations can be migrated to await incrementally.

tests/test_main.py (2 changes):

_mock_session() — updated to mock prompt_async instead of prompt. Added an async def _async_side_effect() that yields inputs in order (raising exceptions for error inputs). Keeps .prompt() as a no-op fallback for safety.
test_token_refresh — the tricky one. asyncio.run() internally calls time.monotonic() for its shutdown/generator cleanup, which consumed entries from the side_effect=[0, 99999, ...] list unpredictably. Replaced the static side_effect list with a callable _monotonic() that tracks call count: returns 0 for the first 2 calls (asyncio internals + startup), then 99999 to trigger the token refresh check. This makes the test deterministic regardless of how many internal monotonic calls asyncio makes.

Pre-commit:

First commit attempt failed — ruff fixed an unused import asyncio in the test file (it was added but asyncio wasn’t directly called in the module scope; the async def uses it implicitly).
Second attempt passed both ruff and ruff-format.

Net effect: +41 lines, -3 lines across 2 files. No behavior change for users — the REPL works exactly the same, just runs under an async event loop now.

Back Donate