AgentDebugX/AgentDebugX

A local-first debugging framework for agentic AI systems that ingests a live or exported agent trajectory, detects visible failure signals, attributes them to responsible steps or agents, proposes recovery actions and prepares controlled reruns so fixes can be validated instead of guessed. The repository ships dsh-agentdebugx, its DeepSeek Harness plugin, alongside a Python package (pip agentdebugx) and a published CUADebug paper on diagnosing computer-use agent failures.

AgentDebugX turns failed agent runs into structured, auditable debugging artifacts: it ingests a live or exported trajectory, detects visible failure signals, attributes them to the responsible steps or agents, proposes recovery actions and prepares controlled reruns so a fix can be validated instead of guessed. It is aimed at people building multi-agent systems, tool-using agents, computer-use agents, benchmark runners and local agent workflows, and it keeps the whole loop local-first - traces never leave the machine unless you opt in, and every recovery proposal carries its policy and approval metadata across the Rerun boundary.

Coding & Development ★ 64
View on GitHub ↗

Install

dsh plugin --profile web add dsh-agentdebugx

Copied verbatim from the README section on the DeepSeek Harness plugin, released as dsh-agentdebugx v0.1.0 on 2026-08-25.

Compatibility

AgentDebugX is local-first by default - traces stay on your machine, sharing is opt-in, and recovery proposals carry explicit policy and approval metadata into the Rerun boundary. The framework is Python 3.9+ (pip install agentdebugx); the DeepSeek Harness plugin is published separately as dsh-agentdebugx. The project builds on the authors' CUADebug paper (arXiv 2608.02643) for computer-use agent failures.

Details

Alternatives

030611/dsh-context-provenance · 030611/qiushi-dsh-evidence-audit · 0x3fffff/dsh-desktop-launcher

More plugins in Coding & Development

Browse more in Coding & Development

Guides for Coding & Development plugins