[FEATURE]: Unleash the full power of the CLI: Callback to the LLM/agent from scripts

[micuintus]

Feature hasn't been suggested before.

• I have verified this feature I'm about to request hasn't been suggested before.

TL;DR

Let the LLM write bash scripts that call back into the running agent — deterministic control flow where code excels, AI judgment where LLMs excel.

---

More detail

Recently I stumbled across this article about — in the end — why Claude Code is so successful (amongst other things): Because it leans into the power of the Unix CLI:

• CLI is a native language of LLMs — they have read millions of Bash examples
• The "everything is text" principle fits very well with LLMs

Second thing I stumbled across: Fabric — AI as a Unix pipe:
yt --transcript https://youtu.be/... | fabric -sp extract_wisdom

This is where I had

The idea

Give the LLM the ability to write scripts that call back into the agent to:

• Create a deterministic execution umbrella like a Ralph Loop — the LLM writes the retry loop, bash runs it reliably, each callback gets fresh context
• Unleash the beauty of something like Fabric in openCode — pipe file contents through AI judgment, map-reduce across a codebase
• Let scripts do what they can do better: more reliably, faster, and save tokens

-> Deterministic work (loops, pipes) goes into in bash where it's reliable. Non-deterministic judgment ("is this buggy?", "what's the fix?") calls back to the LLM where it excels. The LLM writes the script that orchestrates both.

The problem

1. LLMs are good at figuring out things and generating code
2. LLMs are bad at:
   • Iterations — they lose count, skip steps, forget where they were
   • Following through — they lose track of the control flow once grinding too much on one aspect
   • Deterministic control flow — loops, conditionals, batch operations are something LLMs are really bad at