AI Native Notes
Volver a Gemini CLI
Gemini CLIAvanzado3 min de lectura

Ralph Persistence Loops — Self-Healing Execution

How to build infinite verification loops (Ralph Loops) where the agent analyzes errors and fixes code autonomously until tests pass.

ralphpersistenceself-healingtddinfinite-loop

The probability of an AI writing perfect code on the first try is low. True automation lies in creating a "loop that fixes itself when wrong."

1. What is a Ralph Loop?

A Ralph Loop (or self-healing loop) is a pattern where the agent does not just execute a one-off command, but iteratively modifies code and verifies it until a target condition (e.g., all tests passing) is met. Instead of saying "Fix this," you instruct the agent: "Run this test script. If it errors out, analyze the cause, fix it, and run it again. Do not stop until it passes."

2. Writing the Perfect Ralph Prompt

You must establish clear constraints so the agent doesn't fall into an infinite loop or delete unrelated code.

<objective>
Modify the code so that all unit tests in `src/utils/date-parser.ts` pass.
</objective>
 
<execution_loop>
1. Execute the command `npm run test:date-parser`.
2. If the tests pass (Exit Code 0), terminate the task.
3. If the tests fail, analyze the error log to determine the cause.
4. Modify the source code and loop back to step 1.
</execution_loop>
 
<constraints>
- Loop a maximum of 5 times. If you exceed 5 attempts, stop and report back to me.
- NEVER modify the test file (`*.test.ts`) itself. Only modify the implementation file.
- Do not install new external libraries (like moment or date-fns); use only native Date APIs.
</constraints>
 
If you understand these instructions, begin the loop.

3. Synergy with TDD (Test-Driven Development)

The Ralph Loop is most powerful in a TDD environment. When a human developer writes a "failing test (Red)" first, the agent loops to make it "pass (Green)," and finally refines the code (Refactor).

  • Human: Writes rigorous test code including edge cases.
  • Agent: Implements the code, runs the tests, and iterates at high speed to fix failures.

This delegates the judgment of "does this work?" from the human to a clear, impartial judge (the Test Runner).

4. Background Bots via Headless Mode

Don't block your local terminal. Spin up a Ralph bot in Headless mode in another tab. The human can set up the skeleton, and the bot fleshes it out—perfect asynchronous collaboration.

gemini --headless "@tests/auth.test.ts Fix the logic in src/auth/ until this test file passes. Repeat up to 5 times."

5. Handling Loop Breaks

If the agent fails after 5 attempts, it's rarely an intelligence issue; it's usually a lack of context.

  1. The test itself is flawed: Dependency mocking might be missing.
  2. Missing required knowledge: The agent might need the latest specs for an external API. Provide documentation via MCP or give a hint with /memory add.

Next to Read

Guías Conectadas