Skip to content

Latest commit

 

History

History
134 lines (101 loc) · 6.19 KB

HISTORY.md

File metadata and controls

134 lines (101 loc) · 6.19 KB
Raw

Release history

v0.13.0

  • Only git commit dirty files that GPT tries to edit
  • Send chat history as prompt/context for Whisper voice transcription
  • Added --voice-language switch to constrain /voice to transcribe to a specific language
  • Late-bind importing sounddevice, as it was slowing down aider startup
  • Improved --foo/--no-foo switch handling for command line and yml config settings

v0.12.0

  • Voice-to-code support, which allows you to code with your voice.
  • Fixed bug where /diff was causing crash.
  • Improved prompting for gpt-4, refactor of editblock coder.
  • Benchmarked at 63.2% for gpt-4/diff, no regression.

v0.11.1

  • Added a progress bar when initially creating a repo map.
  • Fixed bad commit message when adding new file to empty repo.
  • Fixed corner case of pending chat history summarization when dirty committing.
  • Fixed corner case of undefined text when using --no-pretty.
  • Fixed /commit bug from repo refactor, added test coverage.
  • Benchmarked at 53.4% for gpt-3.5/whole (no regression).

v0.11.0

  • Automatically summarize chat history to avoid exhausting context window.
  • More detail on dollar costs when running with --no-stream
  • Stronger GPT-3.5 prompt against skipping/eliding code in replies (51.9% benchmark, no regression)
  • Defend against GPT-3.5 or non-OpenAI models suggesting filenames surrounded by asterisks.
  • Refactored GitRepo code out of the Coder class.

v0.10.1

  • /add and /drop always use paths relative to the git root
  • Encourage GPT to use language like "add files to the chat" to ask users for permission to edit them.

v0.10.0

  • Added /git command to run git from inside aider chats.
  • Use Meta-ENTER (Esc+ENTER in some environments) to enter multiline chat messages.
  • Create a .gitignore with .aider* to prevent users from accidentaly adding aider files to git.
  • Check pypi for newer versions and notify user.
  • Updated keyboard interrupt logic so that 2 ^C in 2 seconds always forces aider to exit.
  • Provide GPT with detailed error if it makes a bad edit block, ask for a retry.
  • Force --no-pretty if aider detects it is running inside a VSCode terminal.
  • Benchmarked at 64.7% for gpt-4/diff (no regression)

v0.9.0

  • Support for the OpenAI models in Azure
  • Added --show-repo-map
  • Improved output when retrying connections to the OpenAI API
  • Redacted api key from --verbose output
  • Bugfix: recognize and add files in subdirectories mentioned by user or GPT
  • Benchmarked at 53.8% for gpt-3.5-turbo/whole (no regression)

v0.8.3

  • Added --dark-mode and --light-mode to select colors optimized for terminal background
  • Install docs link to NeoVim plugin by @joshuavial
  • Reorganized the --help output
  • Bugfix/improvement to whole edit format, may improve coding editing for GPT-3.5
  • Bugfix and tests around git filenames with unicode characters
  • Bugfix so that aider throws an exception when OpenAI returns InvalidRequest
  • Bugfix/improvement to /add and /drop to recurse selected directories
  • Bugfix for live diff output when using "whole" edit format

v0.8.2

  • Disabled general availability of gpt-4 (it's rolling out, not 100% available yet)

v0.8.1

  • Ask to create a git repo if none found, to better track GPT's code changes
  • Glob wildcards are now supported in /add and /drop commands
  • Pass --encoding into ctags, require it to return utf-8
  • More robust handling of filepaths, to avoid 8.3 windows filenames
  • Added FAQ
  • Marked GPT-4 as generally available
  • Bugfix for live diffs of whole coder with missing filenames
  • Bugfix for chats with multiple files
  • Bugfix in editblock coder prompt

v0.8.0

  • Benchmark comparing code editing in GPT-3.5 and GPT-4
  • Improved Windows support:
    • Fixed bugs related to path separators in Windows
    • Added a CI step to run all tests on Windows
  • Improved handling of Unicode encoding/decoding
    • Explicitly read/write text files with utf-8 encoding by default (mainly benefits Windows)
    • Added --encoding switch to specify another encoding
    • Gracefully handle decoding errors
  • Added --code-theme switch to control the pygments styling of code blocks (by @kwmiebach)
  • Better status messages explaining the reason when ctags is disabled

v0.7.2:

  • Fixed a bug to allow aider to edit files that contain triple backtick fences.

v0.7.1:

  • Fixed a bug in the display of streaming diffs in GPT-3.5 chats

v0.7.0:

  • Graceful handling of context window exhaustion, including helpful tips.
  • Added --message to give GPT that one instruction and then exit after it replies and any edits are performed.
  • Added --no-stream to disable streaming GPT responses.
    • Non-streaming responses include token usage info.
    • Enables display of cost info based on OpenAI advertised pricing.
  • Coding competence benchmarking tool against suite of programming tasks based on Execism's python repo.
  • Major refactor in preparation for supporting new function calls api.
  • Initial implementation of a function based code editing backend for 3.5.
    • Initial experiments show that using functions makes 3.5 less competent at coding.
  • Limit automatic retries when GPT returns a malformed edit response.

v0.6.2

  • Support for gpt-3.5-turbo-16k, and all OpenAI chat models
  • Improved ability to correct when gpt-4 omits leading whitespace in code edits
  • Added --openai-api-base to support API proxies, etc.

v0.5.0

  • Added support for gpt-3.5-turbo and gpt-4-32k.
  • Added --map-tokens to set a token budget for the repo map, along with a PageRank based algorithm for prioritizing which files and identifiers to include in the map.
  • Added in-chat command /tokens to report on context window token usage.
  • Added in-chat command /clear to clear the conversation history.