pi-apply-patch-plus

English | 简体中文

Bring Codex-style patch-based editing to pi: describe multiple file changes — add, update, delete, move — in one structured apply_patch envelope, with local format verification and a colored diff preview. Smartly recognizes GPT-family models and automatically enables patch editing, taking over edit / write — no hardcoded model list to maintain.

Features

• Patch-based multi-file editing — one patch envelope describes all changes at a glance:

  *** Begin Patch
  *** Add File: src/new.ts
  +export const value = 1;
  
  *** Update File: src/app.ts
  @@
  -const oldValue = 0;
  +const newValue = 1;
  
  *** Delete File: src/obsolete.ts
  *** End Patch
  

• Local verification first — the patch is parsed and validated (envelope, operation headers, context matching, + prefixes) before anything touches the filesystem; invalid patches are rejected instead of silently producing wrong results.

• Smart GPT detection — any gpt-* model gets the tool automatically (gpt-5.6-sol / gpt-5.6-luna / gpt-5.6-terra / gpt-4o, and future GPT models); non-GPT models keep it hidden.

• Codex remote constraint — on gpt-5.2/5.3-codex, patches are sent to the OpenAI Codex API as CFG-constrained freeform tools, constrained at the grammar level.

• Diff UI & live counters — pi's native diff renderer shows colored diffs, with per-file +X -Y counts and progress.

• Interrupted-turn recovery — orphaned custom tool calls are auto-completed with synthetic cancellation output so follow-up prompts never fail on a missing custom_tool_call_output.

• Tool takeover policy — while active, edit / write are disabled in favor of apply_patch, preventing parallel tools from clobbering the same file; they are restored automatically on non-GPT models.

Install

Requires pi. Install the package from npm:

pi install npm:pi-apply-patch-plus

Then run /reload inside pi so the extension is loaded into the current session.

To update later:

pi update npm:pi-apply-patch-plus

Alternatives — install from the source repository:

pi install git:github.com/matsuzaka-yuki/pi-apply-patch-plus

or for local development:

pi install ./

How it works

1. Registers an apply_patch tool in pi;
2. Enforces the per-model tool policy: GPT models get apply_patch and lose edit / write; non-GPT models get the built-in editing tools back;
3. On gpt-5.2/5.3-codex, the tool is sent to the Codex API as type: "custom" + syntax: "lark", and custom_tool_call events are mapped back to pi tool calls;
4. Parses the patch locally → validates → applies file changes → replies with custom_tool_call_output;
5. Renders per-file +X -Y counters and a colored diff throughout (expand the tool output to see the full patch).

Patch format

Envelope structure:

*** Begin Patch
*** Add File: hello.txt
+Hello
*** End Patch

Supported operations:

• *** Add File: <path> — create a file (every line of content prefixed with +)
• *** Delete File: <path> — remove a file
• *** Update File: <path> (optional *** Move to: <path>) — patch in place / rename
• hunks starting with @@, lines prefixed with (context), + (added), - (removed)

Path behavior matches pi's write semantics: relative paths resolve against the current working directory.

Model support

Model	Behavior
Any gpt-* (e.g. gpt-5.6-sol, gpt-5.6-luna, gpt-5.6-terra)	Local support: regular JSON tool, patch in the input field
gpt-5.2-codex* / gpt-5.3-codex*	Full support: additionally sent as CFG/freeform tool over the Codex API
Non-GPT models (claude-*, deepseek-*, glm-*, …)	Tool hidden and blocked

Development

pnpm install
pnpm typecheck
pnpm test

Acknowledgments

This project's feature design and interaction ideas are inspired by the MIT-licensed pi-extension-codex-apply-patch project — thanks to the original author for the open-source contribution. On top of it, this project rewrote the model-detection logic (smart GPT recognition instead of a hardcoded list), documentation, and project packaging to better fit personal use.