AgentTM

Product Introduction

1. Definition: AgentTM is a local-first debugging tool and time-travel debugging utility specifically designed for AI coding agents. Categorized under developer tooling and AI workflow diagnostics, it records managed Codex sessions, captures execution timelines, and indexes workspace checkpoints into a local SQLite database with content-addressable storage.
2. Core Value Proposition: AgentTM exists to bridge the visibility gap left by traditional version control systems during AI-assisted software development. By offering local-first Codex session history, workspace checkpoints, file diffs, and safe restoration capabilities into isolated Git worktrees, it allows developers to track agent activities, investigate execution paths, and recover historical file states without disrupting their primary workspace.

Main Features

1. Session Timeline and Event Tracking: AgentTM records supported events, tool activities, commands, and failures reported during AgentTM-managed Codex runs. How it works involves capturing operation-boundary observations and maintaining a chronological sequence of agent interactions in a local SQLite index (.agenttm/data/events.db), giving developers a transparent look into how an AI agent arrived at a specific code modification.
2. Workspace Checkpoints and Deduplicated Storage: The tool tracks workspace states by creating snapshot manifests and storing file blobs using SHA-256 content-addressable storage. Identical file bytes are shared across checkpoints inside the .agenttm/ directory to minimize storage overhead, allowing users to browse historical files and compare changes against parent checkpoints alongside the execution timeline.
3. Isolated Git Worktree Restoration: AgentTM allows developers to recover captured files from an eligible checkpoint without overwriting their current working directory. By restoring files into a separate, isolated Git worktree, the tool provides a fresh starting point for code review or experimentation while keeping the original project and existing Git commits completely intact.

Problems Solved

1. Pain Point: AI coding agents frequently modify large codebases rapidly, making it difficult to trace intermediate steps, diagnose regressions, or recover specific pre-modification file versions when the agent's final diff falls short of expectations. Standard Git commits only capture manual saves or explicit commit points, leaving a blind spot regarding intermediate AI tool executions and conversational context.
2. Target Audience: AI application developers, software engineers utilizing AI coding assistants (specifically Codex workflows), full-stack developers, and DevOps professionals who require granular visibility and accountability over automated code generation tools.
3. Use Cases: Debugging unexpected code rewrites produced by AI coding agents, auditing the sequence of tool commands executed during an automated debugging session, recovering an earlier functional iteration of a file without rolling back an entire active git branch, and continuing saved Codex conversations via a native desktop interface.

Unique Advantages

1. Strengths & Limitations (Pros & Cons):
   • Pros: Local-first architecture ensures complete data privacy as history, manifests, and file blobs stay on the user's device; deduplicated SHA-256 storage prevents excessive disk usage; seamless integration with existing Git workflows via isolated worktrees; includes pass-phrase encrypted history backups and integrity diagnostics (agenttm doctor).
   • Cons: Feature availability varies by release (with native desktop previews primarily supporting Windows); workspace scans can be partial rather than atomic snapshots; restoring files does not automatically rewind active Codex conversational states; requires manual setup of Codex authentication and Git 2.42+.
2. Key Alternatives & Differentiation:
   • Git Stash / Reflog: While native Git tracks commits and stashes manually, it lacks awareness of AI agent event timelines, tool execution histories, and conversation context. AgentTM operates alongside Git to add specific AI agent telemetry.
   • Cloud-based AI Logging Platforms: Many enterprise AI tools log prompts and completions on remote cloud servers, posing privacy risks. AgentTM is strictly local-first, keeping SQLite databases and snapshot objects locally within the project directory.
   • IDE Undo/Redo Buffers: IDE memory stacks are ephemeral and session-bound. AgentTM provides persistent, searchable, project-scoped historical indexing across multiple development sessions.

Frequently Asked Questions (FAQ)

1. Question? Does AgentTM replace Git for version control? Answer: No. AgentTM works alongside your Git repository, preserving your standard commits, branches, and code reviews while adding a specialized local record of managed Codex sessions, execution timelines, and workspace checkpoints.
2. Question? Is my AI session data sent to a cloud server? Answer: No. AgentTM is a local-first tool. All session history, SQLite databases (events.db), snapshot manifests, and file blobs are stored locally inside your project's .agenttm directory on your device.
3. Question? Does restoring a checkpoint in AgentTM rewind my active Codex conversation? Answer: No. Restoring files into an isolated Git worktree recovers the workspace state, but it does not rewind or alter the ongoing Codex conversation history.
4. Question? What are the system requirements to run AgentTM? Answer: AgentTM requires Git 2.42+, an initialized Git project, a properly authenticated Codex setup for recording, and WebView2 (for the Windows desktop app). Prebuilt binaries for Go and Node.js are not required.