ProofRun Gives AI Coding Agents a Local Verification Receipt

ProofRun, a local verification tool for AI coding agents, landed on Hacker News today with a straightforward pitch: every agent run should produce a cryptographically sound receipt you can inspect offline. For teams running agentic workflows in production, this fills a gap that logging alone does not cover.

Why it matters

AI coding agents now touch real codebases autonomously. The problem is not whether the output is correct, it is whether you can prove, after the fact, what the agent did, in what order, and under what conditions. Logs can be edited. Memory is unreliable. A local verification receipt that is generated at runtime and stored independently changes the accountability model.

This matters even more as multi-model setups become common. A separate signal today shows Moonshot AI's Kimi K3 being tested inside Claude Code, meaning a single coding session may involve multiple models from different providers. Without a receipt layer, attributing a bug or a security issue to a specific model invocation becomes guesswork.

A verification receipt is not a log. It is a commitment: this agent, this context, this output, at this time.

What changes in practice

• Audit trails become portable. A local receipt does not depend on a vendor's dashboard staying up or retaining your data.
• Blame is assignable. In a multi-model pipeline, receipts let you trace which model produced which change, critical when developer tools chain Kimi, Claude, and local models in sequence.
• Compliance gets easier. Teams in regulated industries can point to a receipt as evidence of what an agent was instructed to do versus what it actually did.
• Debugging improves. Receipts give you a reproducible snapshot of agent state, not just the final diff.

How to use it

1. Integrate at the agent boundary. Trigger receipt generation at the start and end of each agent task, not just on errors. You want a record of clean runs too.
2. Store receipts outside the repo. If the agent can write to the repo, it can overwrite a receipt stored there. Use a separate local path or a write-once store.
3. Include model metadata in the receipt. Capture the model name, version, and provider alongside the action log. As multi-model setups grow, this is the field you will wish you had.
4. Pair with a token ledger. A separate project today, RNet, is tackling the fragmented credit problem across agentic IDEs and deployment agents. Combining a token ledger with a verification receipt gives you both cost attribution and action attribution in one place.
5. Review receipts in CI. Add a step that checks receipt integrity before merging agent-generated PRs. Fail the build if the receipt is missing or malformed.

The ProofRun project is early, but the pattern it encodes is one every team running autonomous coding agents should be implementing regardless of the specific tool.

If your agent can act without leaving a verifiable trace, you do not have an agent workflow, you have a liability.