Git tells you what changed, but it remains silent on why. Oqura-ai has released Wisdom, an open-source tool designed to bridge that gap by recording coding-agent sessions locally. The utility links every file modification to the specific request, reasoning, and commands that produced it, effectively creating a ledger of intent that persists beyond the ephemeral context windows of AI agents.

The Problem With Ephemeral Contexts

Current agent workflows suffer from a critical disconnect: transcripts live in proprietary JSONL formats that are unreadable by other tools and vanish when sessions are cleared. Wisdom addresses this by observing file changes directly, including those made via shell commands, and storing them as diffs linked to the originating turn. This ensures that when an agent writes code, the 'why'β€”the prompt, the failed attempts, and the final decisionβ€”remains accessible to subsequent agents or human reviewers.

Inter-Agent Communication and Handoffs

Wisdom normalizes data across major CLI agents, specifically supporting Claude Code (v2.1.319), Codex CLI (v0.166.0), and Pi (v1.2.0). By converting these disparate formats into plain text sized for agent contexts, it enables true interoperability. A standout feature is the wisdom handoff command, which allows a new agent, such as Codex, to start its session by reading the context from a previous Claude Code session. This creates a continuous thread of development that survives tool switching.

Local-First Architecture and Privacy

The system operates entirely on your machine, storing data in a .wisdom/ folder within the project directory. It relies on a small daemon that receives events via agent hooks, ensuring the agent is never blocked even if the daemon is temporarily down. Privacy is maintained by keeping all data local and unredacted in the primary store, while wisdom share and wisdom export functions actively scan for and redact secrets like API keys and tokens before outputting data to teammates or public issues.

Key Takeaways

  • Wisdom links code diffs to specific agent turns, preserving the reasoning behind changes.
  • It supports Claude Code, Codex, and Pi, normalizing their session data for cross-agent queries.
  • The wisdom handoff command enables seamless context transfer between different AI agents.
  • All data is stored locally in a .wisdom/ directory, with built-in secret redaction for sharing.

The Bottom Line

Wisdom solves the 'black box' problem of AI coding by making agent reasoning auditable and portable. For teams mixing multiple LLMs, this ledger is the missing layer of context that turns isolated sessions into a coherent development history.