Edge-Echo/dsh-driftwatch
An LLM regression detector for DeepSeek Harness: record a baseline session, change your prompt, model, plugin or settings, then compare two sessions turn by turn and fail the run when a score drops past a threshold. It works as a network-free CLI (npx dsh-driftwatch list/compare) and as a dsh plugin that exposes the same drifts records and sweep tools to the agent, with --markdown output for CI.
dsh-driftwatch is an LLM regression detector for DeepSeek Harness. You record a baseline session, change your prompt, model, plugin or settings, then compare two sessions turn by turn and fail the run when a score drops past a threshold - so a change that quietly made the agent worse is caught instead of discovered in production. The turn-by-turn comparison is deterministic and network-free, and the same records and sweep tools are exposed to the agent when it is installed as a plugin, making it usable both as a CI gate and as an in-session regression check.
Install
dsh plugin --profile web add dsh-driftwatchCopied verbatim from the README's 'As a dsh plugin' install line (npm form). The README also documents a network-free CLI form (npx dsh-driftwatch compare session-A session-B).
Compatibility
Works as a network-free CLI (npx dsh-driftwatch list / compare) and as a DeepSeek Harness plugin that exposes the same drifts records and sweep tools to the agent. The compare command supports --fail-on-drift <threshold> and --markdown <file> for CI. English README with a 中文 translation (README.zh.md).
Details
- Repo: Edge-Echo/dsh-driftwatch
- Category: Coding & Development
- Stars: 3
- Version: GitHub repo Edge-Echo/dsh-driftwatch (3★, MIT); npm package dsh-driftwatch at 0.1.1 (registry check 2026-10-12; registry repository field resolves back to this repo).
Alternatives
030611/dsh-context-provenance · 030611/qiushi-dsh-evidence-audit · 0x3fffff/dsh-desktop-launcher