v3.0.0.rc1

New Features

• Added explain operation - answering arbitrary questions about the project or source code, context-aware by automatic selection of relevant code based on project annotations and the question asked.

NOTE: You will need to reinitialize your project configs by running Perpetual init -l <lang> in order to install the explain operation config file and make it work. Additionally, you may want to manually update your production .env file from the new env.example to add new environment options to fine-tune the LLM for the explain operation.

Breaking Changes

• The report operation now writes its output to stdout instead of stderr if no output file is provided, and all program logs for this operation are redirected to stderr.

v2.1.1

Improvements

• Minor parameter updates and spelling fixes in .env.example.
• Added the inclusion of the Perpetual version number to the generated .env.example config.

v2.1.0

Improvements

• Added base support for C and C++ languages with CMake build. Due to the vast number of different project formats and file types, you will likely need to modify the default configuration provided by Perpetual init -l <c|cpp> to suit your needs before fully utilizing it.
• Added initial support for Arduino C/C++ projects (sketches). Perpetual will not have access to third-party modules' sources if initialized from the sketch directory. It is recommended to run Perpetual init -l arduino from the parent sketch directory and extract the module sources there to allow Perpetual to access them.
• Improved debug string generation for all LLM providers (displayed when running perpetual operations with -v flag).
• Added optional basic-auth support for the Generic provider and the ability to disable authentication completely.
• Added optional basic and bearer authentication support for the Ollama provider (useful for public instances wrapped with an HTTPS proxy).
• Added parameters to set context window sizes for Ollama provider models (per operation).
• Made the temperature parameter optional for all currently implemented providers (none of them enforce setting the temperature).
• Made the max tokens parameter optional for the Generic and OpenAI providers (it is still recommended to set them).
• Added "reasoning effort" advanced parameter support for OpenAI and Generic providers, only works with reasoning models like o1 (full version, not -preview or -mini).
• Added various quirks to better support working with reasoning models for Ollama and Generic providers

Bug Fixes

• Removed unsupported advanced parameters from the Anthropic provider.
• Removed unsupported advanced parameters from the OpenAI provider.
• Fixed advanced parameter support for the Generic provider, allowing any combination of parameters.

NOTE: You will need to reinitialize your project config by running Perpetual init -l <lang>. Also, you may want to manually update your production .env file from new env.example (this is optional, old .env file should continue to work).

v2.0.0

This major release aims to make the implement operation more usable with smaller models (sub-15B models run with Ollama).

Breaking Changes

• User-customizable prompts in the .perpetual/prompts directory have been moved to the base .perpetual directory. Prompts are now grouped together by operation name and stored inside JSON config files. Configs include all needed prompts, text tags, and regex definitions used with each specific operation.
• Split the extra reasoning mode for the implement operation into a dedicated stage, so the implement operation now has 4 stages instead of 3. This allows for separating reasonings and change-detection prompts, producing smaller and simpler instructions. May improve results with Ollama when using smaller models.
• Added support for structured JSON output mode for the implement operation (stages 1 and 3) and for the doc operation (stage 1). This may provide better results with Ollama when using smaller models and potentially reduce costs for OpenAI or Anthropic. For Ollama, the minimum supported version is 0.5.1, and results may vary depending on the model used. For OpenAI, the minimum requirement is gpt-4o and newer. For Anthropic, it should work with Claude 3 models and newer.

NOTE: You will need to reinitialize your project by running Perpetual init -l <lang> to regenerate prompts. You should also update your .env file(s) for implement stages 2, 3, and 4 configurations if not using defaults.

Improvements

• In implement stage 3, when generating a list of files to be changed, user-requested files (with ###IMPLEMENT### comments) are always added to the list by default (can be disabled).
• When running init, the system now warns about obsolete config files in the .perpetual subdirectory that are no longer needed. Use the -c flag to remove them automatically.
• Added the best variant selection strategy for the annotate operation.
• Added support for the o1 series of models for the OpenAI provider. This is only recommended for use with the doc operation as it is slower and less predictable.
• Added support for generic OpenAI-API compatible providers. See .env.example for more info.

Bug Fixes

• Automatically unset env-variables that may affect OpenAI, Anthropic, and Generic providers if using both default and non-default profiles in your .env file, like OPENAI_* and OPENAI1_*.
• On implement stage 3, ensure that new files proposed by LLM satisfy initial project file-selection filters, so it is impossible now to rewrite files that do exist on disk but have been omitted from processing globally.

v1.9.0

Breaking changes

• Improve annotate operation: Added multi-stage annotation generation with support for multiple annotation variants and different selection strategies. The operation now generates multiple variants of source file annotations and selects the optimal one using configurable strategies:
  • short: Selects the shortest variant (default)
  • long: Selects the longest variant
  • combine: Creates an optimal annotation by combining multiple variants through an additional LLM pass
  • Configuration is done via environment variables (see .env.example):
    • *_VARIANT_COUNT_OP_ANNOTATE: Number of variants to generate (default 1 - use old behavior)
    • *_VARIANT_SELECTION_OP_ANNOTATE: Selection strategy to use
    • Separate settings for different LLM providers (ANTHROPIC/OPENAI/OLLAMA)
    • Can be customized per operation using *_OP_ANNOTATE and *_OP_ANNOTATE_POST suffix (for post-processing stage)

NOTE: you will need to reinitialize your project by running Perpetual init -l <lang> to save new annotate-operation prompt-templates

Improvements

• Added multiple profiles support (see .env.example for more info). You can now configure multiple profiles for each LLM provider using a numeric suffix in the provider name (e.g., ANTHROPIC1_, OPENAI2_, OLLAMA3_).

v1.8.2

Improvements

• Improve annotate operation prompts. Add separate prompts for golang unit-tests source files.

• Update example .env file, set more optimal default settings to the current time.

v1.8.1

Bug Fixes

• Various fixes to default regular expressions for selecting and filtering project files.

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

v1.7.4

Bug Fixes

• Check project root dir is not a symbolic link, as this behavior is not supported and will lead to no files to process

v1.7.3

Bug Fixes

• Fixed implement operation bug introduced with 1.7: when using the -p or -pr modes, new files are filtered-out from processing