Persistent AI Job Execution Using JDR and Runtime State: Semantic Context Compression and Job Resumption Across Chat Sessions

AbstractLong-running interactions with large language models (LLMs) accumulate increasingly large Chat Contexts. Although much of the accumulated history may no longer be necessary for continuing a task, conventional chat-based execution commonly depends on carrying the previous conversational context forward. This paper proposes a persistent AI Job execution model that combines JDR (Job Descriptor), a human-readable declarative description of an AI-assisted Job, with an externalized Runtime State.Runtime State records the semantically valid execution state required to continue a specific Job execution, including the current Stage and Task, completed Tasks, confirmed human decisions, active constraints, unresolved issues, pending human input, generated artifacts, and the next action. The transformation from accumulated Chat Context into this structured state is termed Semantic Context Compression.The proposed architecture separates three kinds of information: JDR Definition, Runtime State, and Execution Log. A Runtime-State-aware self-descriptive JDR header defines startup, suspension, human confirmation, persistence, and resumption policies.A feasibility experiment was conducted using Job_No_003, a daily voice and expression training Job. The Job was executed through Question 7, suspended, represented by a JSON Runtime State, and subsequently resumed in a new Chat session from Question 8 without repeating Questions 1–7. Previously selected execution parameters were also preserved.The experiment demonstrates the feasibility of separating the lifetime of a Chat session from the lifetime of an AI Job. Quantitative measurement of Context reduction, cross-model portability, automatic checkpointing, and larger-scale validation remain subjects for future work.