Patch-22

Patch-22 provides an apply_patch command you can put on your PATH as a safety net when a model tries to run apply_patch as a shell command.

It can also optionally print an LLM-facing warning (or outright refuse to patch) to nudge the model back toward its native editing tool (see Configuration below).

It ensures that running a Codex-style patch block in a shell:

apply_patch <<'PATCH'
*** Begin Patch
*** Update File: some/file.txt
@@
 hello
-world
+there
*** End Patch
PATCH

…actually applies the patch, instead of failing with “command not found”.

Why “Patch-22”?

Because it’s a patch-shaped Catch‑22: if a model is convinced apply_patch is a shell command, it will keep trying it. Adding a real apply_patch makes the failure go away — but also “confirms” the habit.

Alas, sometimes the best solution is a carefully contained hack.

Like a fine wine, Patch-22 goes especially well when paired with the jeffnash/CLIProxyAPI fork (hosted on Railway) so those models can be used inside other coding tools. In that setup, this repo is the safety net: apply the patch anyway, or refuse/warn loudly enough that the model learns to stop trying (see Configuration below).

If you're looking for a coding tool to test this out with, I highly recommend the vastly-underrated Letta Code (and the broader Letta ecosystem).

The jeffnash forks — jeffnash/letta-code and jeffnash/letta — are wired up to route the main model calls through an also-Railway-hosted jeffnash/CLIProxyAPI server, while still using the standard OpenAI API for embeddings (cheap).

The jeffnash/letta fork also includes scripts to make deploying the Letta server to Railway easy.

Links:

• jeffnash/CLIProxyAPI
• jeffnash/letta-code
• jeffnash/letta
• Letta
• Letta Code docs

Install

Option B (recommended): build a standalone Rust binary (Codex reuse)

This repo vendors the codex-rs/apply-patch crate from the Codex repository (see vendor/CODEX_COMMIT) and builds a tiny apply_patch binary that reads from stdin.

cargo install --path . --locked

Ensure ~/.cargo/bin is on your PATH.

Option A: use the Bash/Python fallback script

From this directory:

chmod +x ./apply_patch
mkdir -p ~/.local/bin
ln -sf "$PWD/apply_patch" ~/.local/bin/apply_patch
export PATH="$HOME/.local/bin:$PATH"

Notes

• Reads the patch from stdin.
• Supports *** Add File:, *** Update File: (with optional *** Move to:), and *** Delete File:.
• Option A (script) is a Python implementation intended to match the vendored Codex behavior/output as closely as possible; Option B is still preferred.

Configuration (LLM Guardrails)

The goal is for this tool to be used as little as possible.

It can be configured to either:

• Refuse patching and print an instruction message for the model (refuse mode), or
• Apply the patch and also print an instruction message (warn mode).

In other words: ideally you run this once, it prints a stern note to the LLM about using the model-native patching tool instead, and then it never gets invoked again.

Defaults

• Default mode: apply (applies the patch; prints no LLM banner).
• --apply, --refuse, --warn: aliases that set mode to apply / refuse / warn (persisted).
• Default refuse banner: built-in DEFAULT_REFUSE_MESSAGE (used when mode=refuse and refuse_message is unset).
• Default warn banner: built-in DEFAULT_WARN_MESSAGE (used when mode=warn and warn_message is unset).
• --set-refuse-message <text> / --clear-refuse-message: set custom refuse banner / revert to default.
• --set-warn-message <text> / --clear-warn-message: set custom warn banner / revert to default.

Config Location

• $APPLY_PATCH_CONFIG if set, otherwise $XDG_CONFIG_HOME/.apply_patch/config.json, otherwise ~/.apply_patch/config.json.
• If neither HOME nor XDG_CONFIG_HOME is set and you run a config command (e.g. --show-config), it exits 1 with: Error: could not determine config path (HOME/XDG_CONFIG_HOME not set).

Examples:

apply_patch --show-config
apply_patch --refuse   # persistently refuse patching
apply_patch --warn     # persistently apply + print message
apply_patch --apply    # back to normal behavior (default)

apply_patch --set-refuse-message "..."   # customize message
apply_patch --clear-refuse-message
apply_patch --set-warn-message "..."
apply_patch --clear-warn-message

License & Attribution

• This project is licensed under Apache-2.0 (see LICENSE).
• It vendors codex-rs/apply-patch from OpenAI Codex at vendor/CODEX_COMMIT (see vendor/CODEX_LICENSE and vendor/CODEX_NOTICE).