deepseek-harness-reliability-governor

LLM-driven agents often conclude with a 'task complete' status, but the veracity and reliability of this completion are frequently uncertain. A model's self-claimed success can mask factual errors, omissions, or failures caused by environmental changes. This plugin addresses this by introducing an 'evidence contract' mechanism, transforming completion decisions from subjective model assertions into an objective, verifiable process based on deterministic checks. Users can review specific success criteria and validation methods before a task begins, ensuring the agent's execution results align with expectations.

Its core capability lies in building a transparent execution and verification loop. Users set explicit completion contracts via tools like reliability_begin, and the plugin then enforces predefined deterministic checks (e.g., file existence, tool call success, code verification) during execution and at stopping boundaries. If checks fail, the plugin guides the model through a bounded repair process rather than directly declaring success or abandonment. All attempts and terminal outcomes (certified, exhausted, abstained) are recorded with a content receipt in the durable session log, providing an auditable trail.

This plugin is designed for developers, teams, and security researchers building mission-critical agents on the DeepSeek Harness framework. If you prioritize agent execution reliability, require verifiable evidence of success, or aim to mitigate risks stemming from model uncertainty, this plugin offers an effective runtime governance layer. It is particularly suited for automated workflows needing auditable logs, explicit failure-handling mechanisms, and user-in-the-loop decision points.