Open tools that put our research into practice

A local control layer for code written by AI agents.

DiffWitness separates four jobs that are often blurred together: controlling what an agent does while it works, proving what the evidence establishes about the resulting diff, recording what the change still owes, and explaining all of it without overstating it.

What it does

• Protect optionally blocks a short list of high-confidence dangerous actions through the agent's own hooks, with observe, standard and strict policies.
• Proof runs counterfactual experiments on the real Git patch: contrast, necessity, sufficiency, interaction and stability. It issues a certificate tied to the exact commits and evidence command.
• Guard checks the change an agent made during a session. Gate checks an existing diff, locally or in a GitHub Action.
• Debt Ledger keeps an append-only, hash-chained record of the obligations a change leaves behind. Items close only when a check confirms them.

What it does not claim

It is not a proof of correctness and not a sandbox. It cannot establish requirements that the tests do not encode, and it runs the repository's own test commands, which must be trusted.

Status

Public alpha, version 0.4.0a1

License

MIT

Runs on

Python 3.11 or later and Git, no runtime dependencies

Agents

Claude Code and Codex natively, other agents through dw guard

• Website
• Source on GitHub
• Proof protocol
• Threat model

The human-understanding layer for agentic coding.

IdleProof follows a coding agent through a task and explains the change as it happens. Its aim is simple: a developer should not become a stranger to their own product.

What it does

• Combines the task, the agent's lifecycle events, touched files, symbols, routes, data access, imports and the final Git change.
• Explains in a fixed order: what the agent is doing, what it means, which parts of the project are involved, what risks matter and what to remember.
• Shows a local cockpit on your machine with real file names and a bounded map of related files, services and tests.
• Keeps observed facts and inferences apart, and says so when evidence does not support a claim.

What it does not claim

It is not a sandbox, a security scanner or a formal verifier. Static explanations never claim runtime proof, and some language analysis is heuristic.

License

MIT

Version

0.10.0, installed from source

Runs on

Node.js 20 or later and Git, on Linux, macOS and Windows

Agents

Claude Code and Codex

• Source on GitHub

IdleProof explains, DiffWitness proves, and the Debt Ledger remembers what is still owed. Joining the three through a stable identity for each change is a long-term goal, planned for Paitite's optional commercial Portal.