pi-codex-apply-patch

A Pi coding-agent extension that adds an apply_patch tool and a V4A patch harness so GPT-5.2 / GPT-5.2-Codex can propose structured diffs which Pi applies.

This follows the OpenAI Apply Patch / V4A diff flow:

• The model calls apply_patch with a list of operations
• Each operation is one of: create_file, update_file, delete_file
• The diff format is V4A (context/add/delete lines)

What you get

• A Pi tool named apply_patch
• Path validation (prevents directory traversal)
• Atomic writes
• Streaming progress updates in the Pi TUI

Install

Option A: Copy into Pi extensions directory

mkdir -p ~/.pi/agent/extensions
cp ./codex-apply-patch.ts ~/.pi/agent/extensions/codex-apply-patch.ts

Then start Pi normally.

Option B: Load the extension explicitly

pi -e /absolute/path/to/codex-apply-patch.ts

Usage

When the active model is gpt-5.2 or gpt-5.2-codex, the extension:

• enables the apply_patch tool
• disables Pi's built-in edit/write tools (so the model must use structured patches)

Tool schema

apply_patch expects:

{
  "operations": [
    { "type": "create_file", "path": "...", "diff": "..." },
    { "type": "update_file", "path": "...", "diff": "..." },
    { "type": "delete_file", "path": "..." }
  ]
}

Notes:

• create_file.diff: full file content in V4A create mode (each line starts with +)
• update_file.diff: V4A diff using @@ sections and +/-/ lines
• delete_file: no diff

Benchmarks & quality notes (informal)

These are notes from local A/B runs to understand how “forced apply_patch” behaves vs Pi’s normal edit/write flow.

• baseline = Pi built-in edit/write
• apply_patch = this extension (forces structured diffs)
• Each number below is the mean of 10 runs on the same repo snapshot and prompt.
• Cost is taken from Pi’s JSONL usage.cost.total.

Small change (single-file fix)

In a tiny one-file fix, forcing apply_patch was usually slower and more expensive.

• gpt-5.2: baseline $0.0467 / 44.6s vs apply_patch $0.0587 / 50.1s
• gpt-5.2-codex: baseline $0.0291 / 22.6s vs apply_patch $0.0405 / 33.1s

Multi-file change (feature across a few files)

In a multi-file fix (3–5 files), results depended on the model:

• gpt-5.2: baseline $0.1557 / 136.4s vs apply_patch $0.1470 / 129.7s
• gpt-5.2-codex: baseline $0.1002 / 82.6s vs apply_patch $0.1180 / 95.0s

Tool reliability note (10-run totals): on gpt-5.2-codex, forcing apply_patch reduced tool errors (34 → 17), but it cost more and took longer.

Code-quality observations from diffs

• apply_patch gives a nice audit trail (explicit file ops + diffs). That’s the main “quality” win.
• It doesn’t automatically guarantee cleaner code. In a few runs, models used shortcuts to get tests green (e.g. adding a hidden field like __timesLeft or escaping types with any).
• Some runs also touched unrelated files (import specifier changes, etc.). That’s not always wrong, but it adds churn.

Practical takeaway: if your priority is reviewability/reproducibility, forcing apply_patch is a good default. If you optimize for tiny one-line fixes, baseline tooling is often cheaper.

Development

This is a single-file extension. You typically don't need to build anything; Pi loads the TypeScript directly.

References

• OpenAI Platform docs (Apply Patch tool): https://platform.openai.com/docs/guides/apply_patch
• OpenAI Agents SDK (TypeScript) applyDiff reference implementation: https://github.com/openai/openai-agents-js/blob/main/packages/agents-core/src/utils/applyDiff.ts
• OpenAI Agents SDK (Python) apply_diff reference implementation: https://github.com/openai/openai-agents-python/blob/main/src/agents/apply_diff.py
• OpenAI Agents SDK examples:
  • TypeScript apply patch tool example: https://github.com/openai/openai-agents-js/blob/main/examples/tools/applyPatch.ts
  • Python apply patch tool example: https://github.com/openai/openai-agents-python/blob/main/examples/tools/apply_patch.py
• Pi coding-agent extensions docs: https://github.com/badlogic/pi-mono/blob/main/packages/coding-agent/docs/extensions.md

License

MIT