PeakHatch: A Contract-Compiled Runtime for Reliable Interactive Agents

Interactive workflows combine reusable task requirements with facts that can only be established in the current interface. Assigning both to an online model loop makes the model responsible for task control as well as environment interpretation. We present PeakHatch, a contract-compiled runtime that compiles obligations for future execution, selects cognition according to current evidence needs, and retains action admission and workflow advancement in the system runtime. Natural-language authoring, linked action-effect-assertion contracts, and an OS-grounded execution lifecycle implement this division. The prototype uses off-the-shelf Qwen3.5-4B without additional task-specific weight fine-tuning. A source-level example exposes the compilation and consumption interfaces; a historical social-post preparation execution in WeChat Moments demonstrates six bounded cognitive requests, deterministic text verification without LLM/VLM dispatch or replay, and a final publication action reserved for the human. Across 300 serial runs of the same workflow, the collection ledger records a unique matched archive for every start; 296 runs reached the predefined human-confirm Runtime endpoint, an observed proportion of 98.67%. Four early terminations caused by the same class of runtime privacy-scope defect remain in the denominator. Two incomplete development executions provide supplementary observations of local cross-application verification and recovery, without contributing to that proportion. The contribution is a concrete system interface linking task obligations, cognitive needs, and execution consequences, together with bounded execution evidence. Endpoint reachability is not equated with content quality, final publication, or overall cross-application reliability.