Version 0.4.1

Added

• Better context window handling. Old messages are now redacted based on the context window usage and token buffer. These can be changed with --context-window [TOKENS] and --token-buffer [TOKENS]. For example, GPT-4 can be run with a 4K context window to save money: --model 4 --context-window 4000
• Support for when OpenAI drops -0613 from model names
• Limit autonomous function calls to 10 in order to prevent infinite function loop. Limit can be changed with --loop-limit [NUMBER]

Fixed

• Forcing command to be run synchronously or asynchronously

Changed

• Git commit is now made after every run command as well when using the --git flag

Version 0.4.0

Added

• A system message for creating PowerPoint presentations using Marp
• Ability to refuse commands ChatGPT wants to run with a message to ChatGPT
• Improved cross-platform support
• Creation of a project outline in the beginning of the project. Can be skipped with --no-outline or used automatically with --use-outline
• --git flag to initialize a Git repository in the project automatically and commit every task. Commit message generation can be disabled with --no-commit-msg if you don't want to generate a commit message with GPT
• --simple flag for running simple prompts quickly
• --no-cmd flag to disable the run_cmd function from ChatGPT
• --zip flag to create a zip file of the project instead of writing files directly (--no-cmd is added automatically). When this flag is used, files are created only in memory and saved to a zip file in the end. An optional filename or full path of a zip file can be provided.
• Token saver module that redacts redundant file read and write messages from the message history. The "savings rate" of the token saver can be changed by changing --token-saver-level (lower number saves more, default 3).

Changed

• Changed default model to gpt-3.5-turbo-16k-0613
• Updated system messages
• Improved tasklist creation
• Moved modules to modules directory

Removed

• Disabled append_file and replace_text functions because GPT used them wrong often

Version 0.3.0

Added

• Ability to set default command line flags in config.json
• --step-by-step flag to run tasklist step by step as separate messages to GPT
• Ability to modify original prompt iteratively with GPT
• Confirmation before detecting custom system message with GPT
• Ability to skip a tasklist entirely
• System message for PHP websites

Fixed

• Enforcement of the --questions flag even when GPT responds with more questions
• Exception when no output was generated from a failed terminal command
• System message was detected again for each version when using --versions
• Initial clarifying questions now are added to each version when using --versions
• The tasklist is now added to each version when using --versions

Changed

• By default, GPT-AutoPilot will ask whether you want to go through the tasklist step by step or as a single message. You can change this behavior with the flags --step-by-step and --single-tasklist
• Changed the prompt for the system prompt detector
• Clarification questions and their answers are now added to message history as actual assistant/user messages instead of a combined Q/A message.
• Only the final tasklist after changes is kept in the message history
• Unanswered clarifying questions are omitted from the message history and not sent to GPT
• Default temperature from 0.9 to 1.0

Version 0.2.0

Added

• Ability to run the script from outside the script directory
• More command line flags:
  • --dir [DIRECTORY] to use a custom project folder for reading/writing instead of the code folder
  • --create-dir to create project directory automatically if it doesn't exist
  • --prompt-file [FILENAME] to load the first prompt from a file
  • --allow-cmd [EXACT_COMMAND] to allow specific terminal commands to be run without confirmation
  • --one-task to end script after the first 'task is finished' summary
  • --max-tokens [NUMBER] to limit the maximum number of tokens the project can use
  • --max-price [NUMBER] to limit the maximum amount of money the project can use
  • --continue to continue automatically if ChatGPT responds without a function call
  • --model [MODEL_NAME] to set the ChatGPT API model to be used
• GPT-AutoPilot will now ask 5 questions about the project in the beginning, before starting to code. The behavior can be changed with the following command line flags:
  • --no-questions to not ask any questions
  • --questions [NUMBER] to set number of questions to ask
• GPT-AutoPilot now creates a task list for the project and runs through the list one by one. You can use the following command line flags to change the behavior:
  • --no-tasklist to not generate a task list at all
  • --single-tasklist to send the whole tasklist to ChatGPT at once
• Custom system messages and checklists for specific projects. A fitting system message from the prompts folder is picked automatically and can be changed with these command line flags:
  • --system [SYSTEM_MESSAGE_SLUG] to set which system message from prompts folder is used
  • --use-system to not ask for confirmation before using an automatically detected system message
  • --do-checklist to do all tasks in the checklist automatically after the project is finished
  • --no-checklist to not use any checklist
• Rudamentary tests for sample projects

Fixed

• Issues with GPT-3.5 model
• Fixed and "fixed" pyright errors

Changed

• Default temperature from 1.0 to 0.9

Version 0.1.3

Added

• Ability to run terminal commands asynchronously
• Token count and price calculation

Fixed

• The cd code/<base_dir>; command before running a terminal command was missing the / and therefore didn't work.
• Conversation history filenames in history folder are now incremented properly, even if some conversations are deleted
• Give error to GPT if it tries to replace text in a file that doesn't have the given text
• Create directory did not work because directory was passed to codedir() as a string instead of a variable
• Raise TypeErrors and NameErrors when calling OpenAI API

Changed

• Version folder structure is now versions/<incrementing_number>/v[1..] as opposed to previous versions/code_<timestamp>_v[1..].
• Formatted the output of the script nicer

Version 0.1.2

Fixed

• Issue #6: Iterating over a selected version didn't preserve its message history
• Issue #7: The replace_text function given to GPT didn't work with code

Changed

• The replace_text function no longer uses regex because GPT can not properly format regex that includes code

Version 0.1.1

GPT-AutoPilot can now be downloaded as a portable version that doesn't need any dependencies, on both Windows and Linux.

Added

• Support for Windows
• PyInstaller scripts for creating portable versions

Fixed

• Script sometimes got stuck at "Waiting for ChatGPT..." when the GPT API stalled. Added request_timeout parameter to fix this (30 seconds)

Version 0.1.0

First official release

Added

• Better support for GPT-3.5 models
• More command line arguments
  • --conv to set previous conversation ID to continue from
  • --prompt to set initial prompt
  • --delete to delete existing code if present
  • --versions to create multiple versions of the same project at once
  • --temp to set temperature of GPT API
  • --better to automatically make prompt better with GPT
  • --not-better to not ask to make prompt better with GPT
  • --ask-better to ask for confirmation of using better prompt when automatically making better prompt
  • --better-versions to create a separate better prompt for every version

Fixed

• API loop when invalid model was used in settings (issue #3)

Changed

• Conversation ID is now set with --conv flag, instead of the first command line argument