dsh-proof: Evidence-Driven Completion Proof and Regression Attribution Plugin
`dsh-proof` is an open-source plugin designed to convert self-reports from agents in DeepSeek Harness (DSH) into verifiable evidence chains. Through content addressing, signed checkpoints, and offline anchors, it builds a tamper-resistant chain of trust, providing evidence-driven proofs of completion, regression attribution, and real-time verification. The plugin adopts a layered architecture with a core layer decoupled from the host, and includes a DSH plugin, an MCP server, and various adapters. It is applicable to safety-critical development, stringent regression testing, and third-party auditing, shifting agent development from "blind execution" to "verifiable execution."
Read Moredsh-proof: Add read-only validation layer for DeepSeek Harness
dsh-proof is an independent, read-only verification layer plugin for DeepSeek Harness, maintained by EvilIrving (version 0.1.0) under the MIT license. Its core functionality involves intervening via `agent/turn-stopping` when a top-level dialogue turn is about to close, launching a read-only verifier to perform an independent verification of the results. If verification fails, the plugin sends the specific gap and subsequent steering back to the driver agent to correct the process. The plugin relies on the `subagents` service and restricts verifier permissions via `toolFilter.deny`, defaulting to blocking the use of `write`, `edit`, `bash`, and other modification tools.
Read More