Module llm_prompt 

LLM prompt template renderer.

Consumes a Diagnosis, never raw case data. Output is a fully-formed prompt: a system role, a structured Evidence/Hypotheses/Unknowns block, and explicit constraints that forbid the LLM from inventing facts.

The binary makes no network calls. This subcommand only renders the prompt; sending it to a model is the caller’s choice.

Sanitization at the boundary

Free-text values inside Evidence (DNS error messages, validation error strings, etc.) originate in log lines and HTTP response bodies. In a production setting those are attacker-controllable. Before each rendered evidence line is concatenated into the prompt, it is passed through sanitize_for_prompt which replaces newlines with literal \n, escapes backticks, strips other control characters, and caps the line length. This neutralizes the structural injection vectors; semantic attacks (e.g. a base64-encoded directive inside an evidence string) remain a residual risk that the human review step in the “Suggested usage” flow is responsible for catching. See docs/llm_assisted_workflow.md.

Functions

render_prompt

Render the prose form of the LLM prompt for a given diagnosis.

render_prompt_json

JSON envelope variant of render_prompt.

sanitize_for_prompt

Sanitize a rendered evidence line for inclusion in an LLM prompt.