5.7. Generating and editing scripts

Code requests carry your original wording, recent conversation, resolved workflow and project context, retrieved evidence, and output requirements to the code generator. Name the stations, quantities, units, filters, filenames, and formats you need. Unusual requirements remain in the original request; they are not discarded just because the workflow parser has no matching field.

For example:

Write a Python script to load data/3edis, run QC, and save the table
to qc_results.csv in an output folder.

Supported workflow templates remain available in Offline mode. A template returned for a detailed request is labelled for review: Offline mode has not applied arbitrary request-specific constraints. Select a local or cloud model for custom compositions. An unspecified request such as “write me a script” asks what the script should do, rather than defaulting to QC.

If a figure request specifies three panels but omits their quantities, the assistant asks which quantities to show. Your answer is combined with the pending request. Missing data directories are explicitly labelled as placeholders. API evidence includes bounded static inspection of science source signatures and documentation; this does not import or execute modules.

5.7.1. Follow-up edits

The conversation stores generated code and its artifact metadata. A recent script can therefore be edited with requests such as:

Change only the output directory to results/qc_review.
Keep the same figure but save it at 300 dpi instead.
Remove the correction step but keep the factor-estimation table.

The first two forms have a deterministic path for recognized literal output paths and explicit savefig(..., dpi=...) calls. This changes only those source locations. More complex expressions or edits go to the model with the complete prior script. An explanation request does not regenerate the code. Edited artifacts receive a new filename when the destination already exists, preserving the earlier script.

Recent context is limited to six turns and 8000 characters of whole messages. For a tight local context budget, optional templates are omitted first, then retrieved evidence and API documentation are compacted with a recorded note. The current request and prior script remain intact. Exact directory and DPI edits bypass retrieval and model generation. Older omitted turns are reported. The current request and previous script are not silently shortened to fit a local model’s context; an oversized prompt can require a larger configured context or a smaller task.

5.7.2. Validation limits

Generation does not execute the script. Static source/signature checks and supported literal scientific checks are reported separately from skipped runtime checks; see Generated-code validation and repair. They do not establish general scientific validity or fulfillment of every requirement. Model edits still need review. Small local models may ignore instructions or invent APIs despite receiving accurate evidence. Narrow checks flag absent requested filenames and recognized prohibited static-shift correction calls. A draft that fails these checks is shown for review without saving a script. Passing them is not a general correctness test.

5.7.3. Python interface

The shared request can also be supplied directly:

from pycsamt.agents import CodeGenerationAgent, GenerationInput

request = GenerationInput(
    original_request="Run QC and save the table as qc_results.csv.",
    workflow_config={"workflow": "qc", "data_path": "data/3edis"},
    output_requirements=["Export the QC table to qc_results.csv"],
)
result = CodeGenerationAgent(llm_provider="ollama").execute({
    "generation_input": request,
    "output_dir": "results/qc",
})

The existing workflow_config-only interface remains available for template exports. The result records the generation mode, assumptions, script path, and whether the edit used the narrow deterministic path. It never marks a model-generated script as having all requirements verified.