Agent Frameworks◆ AI-generated · Sourced

Git-temp: A Git scratchpad tool designed for AI agents—zero-Git-status clutter, zero configuration

By Umi IntelligenceUpdated 2026-07-18 16:1510 min read

TL;DR

git-temp is a lightweight CLI tool that enables AI agents to safely generate and manage temporary files (e.g., LLM outputs, debug logs, code drafts) inside Git repositories by isolating them in a `.git-temp/` directory and automatically synchronizing its exclusion into `.gitignore`, without modifying Git config, hooks, or index state.

Core Purpose: Solving the 'write-and-pollute' engineering pain point for AI agents in Git repos

Modern AI agents—including DevOps automation scripts, RAG post-retrieval rerankers, and LLM-based code generators—frequently need to write transient artifacts (e.g., unvalidated code snippets, JSON intermediate states, diff snapshots) directly into local Git working directories. Conventional approaches rely on git update-index --skip-worktree or manual .gitignore edits, which are error-prone, hard to automate, and conflict with agent statelessness. git-temp addresses this by defining a convention-over-configuration protocol for ephemeral files.

Implementation: Isolated directory + intelligent `.gitignore` synchronization

All temporary files are written exclusively under .git-temp/ at the repo root (configurable, but default-conforming);
On first git-temp init, it auto-detects and appends /\.git-temp/ to the project’s .gitignore (creates it if absent), guaranteeing full Git invisibility;
Provides semantic commands: git-temp add <file>, git-temp rm <file>, git-temp list; all operate via filesystem moves and line-level .gitignore validation—no Git internals invoked;
Bypasses Git index entirely: no git status changes, no untracked file warnings, no interference with git commit -a;
Compatible with all Git ≥2.0; requires no hooks, no global config, no .git/config edits.

Technical boundaries and intentional minimalism

What it doesn’t do is precisely why it’s reliable

❌ No git stash or git worktree wrappers: avoids Git internal state coupling;
❌ No filesystem event listeners (inotify/fsevents): eliminates background process contention;
❌ No web UI or daemon: pure CLI, single Go binary (<500KB);
❌ No filesystem abstraction layer: all I/O uses native os package—zero third-party FS libraries.

This minimalism makes git-temp an invisible infrastructure in AI agent workflows—for example, calling git-temp add output.py inside a LangChain Tool, or writing reasoning_trace.json to .git-temp/trace/ during multi-step Ollama + OpenAI function calling—all transparent to both developers and Git.

Use cases and integration patterns

Agent-driven development (ADD): LLM-generated unit tests staged in .git-temp/test_scaffold/, moved to tests/ only after human review;
RAG pipeline debugging: Raw embedding vectors (e.g., numpy.save dumps) stored in .git-temp/embeddings_debug/, keeping data/ clean;
CI/CD LLM validation steps: GitHub Actions job uses sebmellen/git-temp@v0.3.1 action to auto-clean .git-temp/ for build reproducibility;
Local LLM serving (e.g., llama.cpp + REST API): Server saves token-level attention maps to .git-temp/attn_viz/ for downstream visualization—no impact on repo state.

Open-source status and roadmap

Maintained by Seb Mellen, current version v0.3.1 (released 2024-07-12), hosted at https://github.com/sebmellen/git-temp, MIT licensed. Already validated for compatibility with Hugging Face transformers and langchain. Next milestones include: .git-temp/.keep semantic support (for empty-directory preservation), --dry-run mode for CI safety checks, and Python bindings (pip install git-temp) for broader agent runtimes (e.g., AutoGen, CrewAI).

Sources (compliance trail)

https://github.com/sebmellen/git-temp