[Absarahmedofficial]

Description

Adds OutputFormatConverter and CharSplitOutputConverter (text -> text) for output-format steering and response normalization in PyRIT converter pipelines.

• OutputFormatConverter: Abstract base class for converters that append output-formatting instructions to a prompt (action="encode") and normalize the formatted target response back into plain text (action="decode") before scoring. Follows the same action="encode" | "decode" pattern and public self.action interface as SmugglerConverter (pyrit/converter/token_smuggling/base.py).
• CharSplitOutputConverter: Instructs the target to insert a single non-whitespace delimiter (. by default) between adjacent characters within each word while preserving spaces, hyphens, and punctuation, and strips intra-word delimiters (unescaping doubled literal delimiters) on action="decode".
• Exported both classes via TYPE_CHECKING and _LAZY_EXPORTS in pyrit/converter/__init__.py.

encoder = CharSplitOutputConverter(action="encode")
prompt = (await encoder.convert_async(prompt=user_prompt)).output_text
response = await target.send_prompt_async(message=Message.from_prompt(prompt=prompt, role="user"))
normalized = (await CharSplitOutputConverter(action="decode").convert_async(prompt=response[0].get_value())).output_text
scores = await scorer.score_text_async(text=normalized)

cc Roman Lutz (@romanlutz)

Tests and Documentation

• Tests: Added 62 unit tests in tests/unit/converter/test_output_format_converter.py covering abstract instantiation checks, action and modality validation, valid/invalid delimiter checks (empty, multi-char, whitespace, non-string, NUL byte), printable and non-printable format instructions, intra-word delimiter stripping, doubled-delimiter unescaping, and ComponentIdentifier params. Verified with pytest tests/unit/converter/test_output_format_converter.py, pytest tests/unit/docs/test_converter_documentation.py, pytest tests/unit/common/test_lazy_package_imports.py, ruff check, ruff format --check, and ty check.
• Documentation & JupyText: Added a CharSplitOutputConverter encode/decode example to doc/code/converters/1_text_to_text_converters.py, synced the paired doc/code/converters/1_text_to_text_converters.ipynb with jupytext --sync, and executed the new cell so the notebook carries the actual output while leaving existing cell outputs untouched.

[Roman Lutz (romanlutz)]

Absarahmedofficial Could we compose the existing converters here instead of adding a new public base class?

PromptTemplateConverter or SuffixAppendConverter can append the formatting instructions, and SearchReplaceConverter can decode the response using the same regex. AttackConverterConfig already lets us configure these separately as request and response converters.

Could we package that configuration as a reusable attack technique instead? That would keep the character-splitting behavior without adding another converter hierarchy to maintain. Is there a use case this composition cannot support that I'm missing?