Abstract: Time-series interferometric synthetic aperture radar （InSAR） processing is characterized by lengthy workflows， strong node dependencies， and strict quality conditions. Accommodating the executability and verifiability required by specialized scien-tific computing is difficult owing to the uncertainty of large language model （LLM） outputs， which manifests as the propa-gation of a language-understanding bias to low-level professional operations， confusion between tool-call success and gen-uine node completion， and distortion of long-flow task states under interaction interruptions and parameter changes. Ac-cordingly， this paper proposes a natural language-driven workflow supervision method for time-series InSAR， which con-fines the LLM to the natural language understanding stage while a deterministic workflow supervision handles the operation decisions and execution verification. A natural language semantic modeling and intent-action mapping method is proposed for addressing the propagation of language-understanding bias to low-level operations， which jointly models the dialogue history， persistent task state， and directed acyclic graph （DAG） execution state； constrains open-ended language understand-ing to a finite set of high-level semantic actions； and maps them deterministically to professional tool calls. Further， a DAG-based two-stage supervised decision-making and evidence feedback model is proposed for addressing the problem that the instantaneous return values of tools only reflect call-level success or failure whereas genuine node completion must be characterized by physical artifacts and quality evidence. This model judges the executability of the action before the exe-cution， based on node dependencies， manual confirmation， and parameter schemas， and verifies the node completion after the execution based on the physical artifacts and quality evidence. Also， a dual-domain memory-based state consistency method is proposed to address the problem that the long-flow task state is prone to distortion under dialogue truncation， parameter changes， and execution interruptions， which ensures that the task state is consistent with the physical evidence under param-eter changes and execution interruptions by coordinating the working memory and persistent memory domains， and through version binding and dependency-invalidation propagation， and supports checkpoint recovery. The three modules are linked in sequence， forming a closed loop comprising semantic modeling， action mapping， supervised decision-making， profes-sional execution， evidence feedback， and state evolution. The practical processing capability of the proposed method was validated on real Sentinel-1 data from Xi’an and Heifangtai， Yongjing County， Gansu Province. The contribution of each su-pervision component was quantified using paired ablation experiments comprising 108 cases， with the complete method achieving an end-to-end accuracy of 86.1%， and outperforming the LLM-only， no-DAG， and without-evidence-supervision controls by 54.6， 25.9， and 26.9 percentage points， respectively. Its node-level accuracy was no lower than 90.7%. The re-sults show that the proposed method decouples the uncertainty of LLMs from deterministic workflow supervision while preserving the natural language interaction capability， thereby reliably maintaining the operation decisions and state progres-sion of long-flow time-series InSAR tasks.