Jeb

Jeb is a coding agent that runs entirely inside Pythonista on iOS, and is meant to be built by bootstrapping an LLM against itself. There is no separate build step, no maintainer writing the harness by hand up front: you run a single seed file, hand the model one instruction, and let it grow its own tool loop, permission system, and console in place.

miniagent/ is not "the" answer — it's one outcome of running the process below to completion once. Treat it as a reference, not a template to copy verbatim.

Point a coding agent at jeb_bootstrap.py and give it this as its first task:

Split jeb_bootstrap.py into a proper multifile module. Keep every behavior — the system prompt, the permission choices, the keychain scheme, the termination blocklist, the console commands, the first-run config migration — identical. Organize it into focused files with clean import boundaries, and write the documentation a future agent would need to understand and safely extend what you've built.

That single prompt is the whole bootstrap. Everything else — which module names to use, how to split responsibilities, what to document — is left to the agent doing the splitting. miniagent/ is what one run of this produced; miniagent/docs/architecture.md describes the shape it landed on, and miniagent/README.md walks through using the result.

Once the split exists, the agent (and the project) keeps evolving the same way: by editing its own harness in place, in the workspace it's running from. The rules for doing that safely — read before you write, prefer surgical edits, never remove the safety machinery, never touch the files that grant it permissions — are written up for the agent itself, not for a human maintainer, in miniagent/docs/self_editing.md. Read that document before asking an agent to modify the harness further, whether it's the one in miniagent/ or a fresh split of your own.

Related documentation inside miniagent/:

• docs/architecture.md — how the pieces fit together after a split.
• docs/reference.md — generated tool, command and module tables, produced from the code by miniagent/gendocs.py so they cannot drift from it.
• docs/self_editing.md — the guide a coding agent follows to modify its own harness.
• docs/jeb_md.md — how standing project instructions (JEB.md) are discovered and injected.
• INSTALL.md — installing a split package in Pythonista.

The point isn't that jeb_bootstrap.py is hard to split by hand — it's that the harness should be able to build and rebuild itself using the same tool loop it will later use for everything else. A harness that can't safely restructure its own source under agent control hasn't proven the safety model it's supposed to enforce on every other project. Bootstrapping it on itself first is the test.