adit

ADIT Core

AI Development Intent Tracker — The Transparent Time Machine

A local-first flight recorder for AI-assisted development that captures the complete context of every interaction, reasoning step, and code modification.

Design Philosophy

ADIT is built on three pillars:

1. Atomic Traceability — Every code change links to the prompt and Chain of Thought that produced it
2. Git-Native, Non-Polluting — Checkpoints stored as refs/adit/checkpoints/<id>, never cluttering commit history
3. Observable Execution — Tool calls, subagent calls, and skill invocations are all recorded

Features

• Integrated Prompt & CoT Capture — Records the full dialogue including Chain of Thought
• Shadow Git Checkpoints — Instant revert to any working state via adit snapshot revert <id>
• Interactive Revert Picker — adit snapshot revert (no ID) shows a numbered list of checkpoints with SHA, timestamp, file count, and summary
• Session Resume — adit snapshot resume [branch] restores working tree from the latest checkpoint on a branch, prints session context, and outputs platform-specific continue commands
• Squash-Merge Support — Resume and revert work after squash merges; automatically finds checkpoints from deleted branches with SHA reachability validation and graceful fallbacks
• Content-Aware Secret Redaction — Shannon entropy scoring with 35+ regex patterns (AWS, GitHub, Stripe, etc.), configurable thresholds and skip rules, custom pattern support
• Hook Chaining — Installs alongside other tools' hooks (Entire CLI, linters, formatters) without overwriting; reinstalls cleanly replace only ADIT entries
• Multi-Actor Timeline — Distinguishes assistant, user, tool, and system actions
• Environment Snapshotting — Captures git state, dependency versions, runtime context, Docker detection, shell info, CPU/memory, and package managers
• Environment Drift Detection — Automatically detects and records environment changes between sessions
• Cloud Sync — Device-code and token-based authentication, cursor-based incremental push with per-project cursors, auto-sync after stable hook events, transcript upload, and rolling-window circuit breaker
• Auto-Sync — Fire-and-forget cloud sync triggered by record count threshold or time elapsed, with fail-open error handling
• Auto Project-Link — Background sync of git metadata and project documents on session-start, with staleness-based caching
• Platform Session Tracking — Correlates events to platform-native session IDs for accurate multi-session handling
• SpecFlow Planning — Structured Intent → Roadmap → Tasks workflow
• Interactive TUI — Full terminal UI built with React/Ink with live detail updates, search, diffs, and environment snapshots
• Platform Adapter System — Pluggable architecture for AI platform integration (Claude Code and OpenCode fully supported, extensible to Cursor, Copilot, Codex, and others)
• Hook Management — Install, validate, and manage platform hooks via adit hook with --all auto-detect and --clean data removal
• Advanced Search — Full-text search with actor, type, date range, branch, and checkpoint filters
• Flexible Export — Export events and sessions as JSON, JSONL, or Markdown reports with optional gzip compression
• Fail-Open Hooks — Recording errors never block the AI agent

Architecture

TypeScript pnpm monorepo with six packages:

Package	Description
@adit/core	Types, SQLite database, config, sync primitives (ULID, vector clocks), content-aware secret redaction
@adit/engine	Git operations, temp-index snapshots, change detection, timeline management, environment capture
@adit/hooks	Hook handlers with platform adapter registry (Claude Code adapter implemented)
@adit/cli	Commander.js CLI with interactive TUI (React/Ink)
@adit/cloud	Cloud sync client — device-code auth, HTTP client with token refresh, incremental sync engine, transcript upload
@adit/plans	SpecFlow-inspired plan artifact generator

Platform Adapters

ADIT uses a pluggable adapter pattern for AI platform integration. Each adapter maps platform-specific hook events to normalized ADIT handlers: