xtrshow

Code Extraction & Patching Made Easy

xtrshow is a CLI suite that closes the loop between your local codebase and AI coding assistants (ChatGPT, Claude, Gemini, etc.). It removes the two biggest points of friction in AI-assisted development:

1. Getting code INTO the LLM — interactively select files and format them with line numbers so the model can anchor edits to exact locations.
2. Getting code OUT of the LLM — apply the model's changes safely, atomically, and reversibly, without copy-pasting dozens of snippets.

No API keys, no agents, no daemons. It works with any chat interface — including free tiers — because the "protocol" is just text.

Installation

pip install xtrshow

Installs two commands: xtrshow (the extractor) and xtrpatch (the patcher).

Quickstart: Extract → Prompt → Apply

1. Extract context. Select files in the TUI; the formatted output goes to stdout:

xtrshow | pbcopy        # macOS
xtrshow | xclip -sel c  # Linux
xtrshow | clip          # Windows

2. Prompt. Paste the context into your LLM along with the patch-format instructions:

xtrshow -p   # prints a copy-pasteable instruction block

3. Apply. Save the LLM's reply to a file and apply it:

xtrpatch changes.txt

Made a mistake? xtrpatch --revert changes.txt unwinds every file the patch touched.

Tip: Iterating on the same files? xtrshow --update re-exports your last selection without opening the TUI.

👉 Read the Full Getting Started Guide

Why not just use git diffs?

LLMs are unreliable at producing valid unified diffs — line numbers drift, context lines get hallucinated, and one bad hunk poisons the file. xtrpatch uses a Search & Replace block format designed around how models actually fail:

• Fuzzy whitespace matching absorbs the indentation errors models constantly make.
• Wildcard anchors (~~~~) let the model match on stable signatures and skip volatile interior lines.
• Tail context disambiguates near-duplicate blocks with a lookahead check.
• Per-hunk annotations (@ why this change) surface in the apply report, so failures are self-describing and feed straight back into the LLM for self-correction.

Features

xtrshow (The Extractor)

• Interactive TUI: fast, keyboard-driven file selection with directory expand/collapse and live size stats.
• LLM-optimized output: line-numbered content with --- a/path headers the patcher can target.
• Smart filtering: ignores node_modules, .git, build artifacts, and friends by default (--no-ignore to disable).
• Re-export: --update replays your last selection from the saved manifest.
• Multi-file export: --multi writes one file per selection — useful for RAG pipelines.
• Prompt printer: -p emits the LLM instruction block for the patch format.

xtrpatch (The Patcher)

• Safety first: automatic, versioned backups in .xtrpatch/ before every modification.
• Checksum verification: warns when a file was edited outside the patch loop since the last apply.
• Conflict detection: overlapping hunks are caught pre-flight and skipped, never blindly stacked.
• Full lifecycle: modify, insert, create files, delete sections, delete whole files (! DELETE FILE).
• Idempotent: re-applying a patch detects already-applied hunks and skips them.
• Error reports: failed applies generate a .rpterr bundle (original + patch + log) you can paste straight back to the LLM.
• Undo button: --revert restores the most recent backup — per file, or for every file in a patch.

License

Apache 2.0