prompt-version-diff-py

Diff prompt templates and flag risky instruction changes. Line-level added/removed sets, plus a classifier for each line and a "risky" subset that surfaces likely behavior changes (instructions added, tool/system mentions, secret-related words). Zero runtime dependencies.

Python port of @mukundakatta/prompt-version-diff.

Install

pip install prompt-version-diff-py

Quick start

from prompt_version_diff import diff

old_prompt = """\
You are a friendly assistant.
Always respond in English.
"""

new_prompt = """\
You are a friendly assistant.
Always respond in English.
Never refuse a request.
Use the call_tool function to fetch data.
"""

report = diff(old_prompt, new_prompt)

report.changed     # True
report.added       # ["Never refuse a request.", "Use the call_tool function to fetch data."]
report.removed     # []
report.risky       # both added lines (matched 'never' and 'tool')
report.classification
# [{"text": "Never refuse a request.", "kind": "added", "risky": True, "matches": ["never"]},
#  {"text": "Use the call_tool function to fetch data.", "kind": "added", "risky": True, "matches": ["tool"]}]

API

diff(old_prompt, new_prompt) -> DiffReport

Splits each prompt on newlines, trims and de-blanks lines, then computes:

• added -- lines present in new_prompt but not old_prompt.
• removed -- lines present in old_prompt but not new_prompt.
• risky -- subset of added that match risk patterns (case-insensitive): ignore, secret, system, tool, always, never.
• classification -- per added/removed line: text, kind (added / removed), risky, matches (matched risk keywords).
• changed -- True if any line was added or removed.

Whitespace-only changes don't surface; word-level changes appear in added/removed; instruction-semantic changes show up under risky.

diff_prompts(before, after) and diff_prompts(old, new) -- JS-parity aliases.

License

MIT