v1.8.0

Breaking changes

• Rework prompting for annotate operation: use separate prompts for different file-types when asking LLM to create a file summary. Now annotate operation is usable even with the small OSS models like Yi-Coder (9B) or DeepSeek Coder V2 Lite (16B) and similar.

  • For using local models with ollama see comments, tips and tricks here
  • The file annotations provided by LLM are now more specific and consistent, potentially improving overall results. The annotations are now larger, but because they are more specific, this can reduce the amount of data LLM will request in the final stages of code implementation.

• The implement operation now excludes unit test source files from processing by default, reducing LLM context pressure and your costs. If you need to work with unit tests, use the new -u flag to disable the unit test source file filter and include them in processing.

NOTE: you will need to reinitialize your project by running Perpetual init -l <lang>

New Features

• Added support to the implement, doc, report operations to provide additional custom filters to exclude certain files from processing, exclude unit test source files from processing by default.

Improvements

• Added more file type mappings for Markdown code block markup: bat/cmd, perl
• Added an additional safety check when generating a source file annotation: it must not contain any code blocks.
• Improve python projects support: include shell scripts and bat files into the file-list by default, LLM prompts updated

Bug Fixes

• Fixed checking for reaching the maximum number of tokens for Ollama