I've been experimenting with agent frameworks and noticed that many workflows reduce to a simple pattern:
- Send input
- Wait for a pattern
- Branch on the match
This is essentially the classic Unix expect model, but applied to LLM conversations.
So I built expectllm — a minimal pattern-matching conversation flow library (365 lines of code).
Example:
from expectllm import Conversation
c = Conversation()
c.send("Review this code for security issues")
c.expect(r"found (\d+) issues")
if int(c.match.group(1)) > 0:
c.send("Fix the top 3")
No chains, no schema definitions, no output parsers.
Features:
- expect_json(), expect_number(), expect_yesno()
- Regex → auto format instructions
- Full conversation history for multi-turn flows
- Auto-detects OpenAI / Anthropic via environment variables
The idea: treat LLM conversations as explicit state machines, where each expect() is a state transition.
Repo: https://github.com/entropyvector/expectllm
PyPI: pip install expectllm
Would love feedback — especially on where this abstraction breaks down.
