aider-chat

• Full support for Claude 3.5 Haiku
  • Scored 75% on aider's code editing leaderboard.
  • Almost as good as Sonnet at much lower cost.
  • Launch with --haiku to use it.
• Easily apply file edits from ChatGPT, Claude or other web apps
  • Chat with ChatGPT or Claude via their web app.
  • Give it your source files and ask for the changes you want.
  • Use the web app's "copy response" button to copy the entire reply from the LLM.
  • Run aider --apply-clipboard-edits file-to-edit.js.
  • Aider will edit your file with the LLM's changes.
• Bugfix for creating new files.
• Aider wrote 84% of the code in this release. <img width="1003" alt="2024-11-04_12-11-03" src="https://github.com/user-attachments/assets/ceab2c15-848c-419c-bed6-38cd2347539e">

• Load and save aider slash-commands to files:
  • /save <fname> command will make a file of /add and /read-only commands that recreate the current file context in the chat.
  • /load <fname> will replay the commands in the file.
  • You can use /load to run any arbitrary set of slash-commands, not just /add and /read-only.
  • Use --load <fname> to run a list of commands on launch, before the interactive chat begins.
• Anonymous, opt-in analytics with no personal data sharing.
• Aider follows litellm's supports_vision attribute to enable image support for models.
• Bugfix for when diff mode flexibly handles the model using the wrong filename.
• Displays filenames in sorted order for /add and /read-only.
• New --no-fancy-input switch disables prompt toolkit input, now still available with --no-pretty.
• Override browser config with --no-browser or --no-gui.
• Offer to open documentation URLs when errors occur.
• Properly support all o1 models, regardless of provider.
• Improved layout of filenames above input prompt.
• Better handle corrupted repomap tags cache.
• Improved handling of API errors, especially when accessing the weak model.
• Aider wrote 68% of the code in this release.

• Enable image support for Sonnet 10/22.
• Display filenames in sorted order.

• Full support for Sonnet 10/22, the new SOTA model on aider's code editing benchmark.
  • Aider uses Sonnet 10/22 by default.
• Improved formatting of added and read-only files above chat prompt, by @jbellis.
• Improved support for o1 models by more flexibly parsing their nonconforming code edit replies.
• Corrected diff edit format prompt that only the first match is replaced.
• Stronger whole edit format prompt asking for clean file names.
• Now offers to add .env to the .gitignore file.
• Ships with a small model metadata json file to handle models not yet updated in litellm.
• Model settings for o1 models on azure.
• Bugfix to properly include URLs in /help RAG results.
• Aider wrote 49% of the code in this release.

<img width="1021" alt="2024-10-22_11-25-54" src="https://github.com/user-attachments/assets/335e8276-fa4a-4749-bb47-8f30e13351b2">

• Check for obsolete yes: true in YAML config, show helpful error.
• Model settings for openrouter/anthropic/claude-3.5-sonnet:beta

• Improvements to /read-only:
  • Now supports shell-style auto-complete of the full file system.
  • Still auto-completes the full paths of the repo files like /add.
  • Now supports globs like src/**/*.py
• Renamed --yes to --yes-always.
  • Now uses AIDER_YES_ALWAYS env var and yes-always: yaml key.
  • Existing YAML and .env files will need to be updated.
  • Can still abbreviate to --yes on the command line.
• Config file now uses standard YAML list syntax with - list entries, one per line.
• /settings now includes the same announcement lines that would print at launch.
• Sanity checks the --editor-model on launch now, same as main and weak models.
• Added --skip-sanity-check-repo switch to speedup launch in large repos.
• Bugfix so architect mode handles Control-C properly.
• Repo-map is deterministic now, with improved caching logic.
• Improved commit message prompt.
• Aider wrote 77% of the code in this release.

• Fixed bug where cache warming pings caused subsequent user messages to trigger a tight loop of LLM requests.

• Use a pair of Architect/Editor models for improved coding
  • Use a strong reasoning model like o1-preview as your Architect.
  • Use a cheaper, faster model like gpt-4o as your Editor.
• New --o1-preview and --o1-mini shortcuts.
• Support for new Gemini 002 models.
• Better support for Qwen 2.5 models.
• Many confirmation questions can be skipped for the rest of the session with "(D)on't ask again" response.
• Autocomplete for /read-only supports the entire filesystem.
• New settings for completion menu colors.
• New /copy command to copy the last LLM response to the clipboard.
• Renamed /clipboard to /paste.
• Will now follow HTTP redirects when scraping urls.
• New --voice-format switch to send voice audio as wav/mp3/webm, by @mbailey.
• ModelSettings takes extra_params dict to specify any extras to pass to litellm.completion().
• Support for cursor shapes when in vim mode.
• Numerous bug fixes.
• Aider wrote 53% of the code in this release.

<img width="832" alt="2024-09-29_08-11-37" src="https://github.com/user-attachments/assets/96b05440-e0f5-49fe-9390-adcbf7ad75d1">

• Fixed dependency conflict between aider-chat[help] and [playwright].

• Support for OpenAI o1 models:
  • o1-preview now works well with diff edit format.
  • o1-preview with diff now matches SOTA leaderboard result with whole edit format.
  • aider --model o1-mini
  • aider --model o1-preview
• On Windows, /run correctly uses PowerShell or cmd.exe.
• Support for new 08-2024 Cohere models, by @jalammar.
• Can now recursively add directories with /read-only.
• User input prompts now fall back to simple input() if --no-pretty or a Windows console is not available.
• Improved sanity check of git repo on startup.
• Improvements to prompt cache chunking strategy.
• Removed "No changes made to git tracked files".
• Numerous bug fixes for corner case crashes.
• Updated all dependency versions.
• Aider wrote 70% of the code in this release.

• Enables prompt caching for Sonnet via OpenRouter by @fry69
• Enables 8k output tokens for Sonnet via VertexAI and DeepSeek V2.5.
• New /report command to open your browser with a pre-populated GitHub Issue.
• New --chat-language switch to set the spoken language.
• Now --[no-]suggest-shell-commands controls both prompting for and offering to execute shell commands.
• Check key imports on launch, provide helpful error message if dependencies aren't available.
• Renamed --models to --list-models by @fry69.
• Numerous bug fixes for corner case crashes.
• Aider wrote 56% of the code in this release.

• Offer to submit a GitHub issue pre-filled with uncaught exception info.
• Numerous corner case bug fixes submitted via new pre-filled crash report feature.
• New /settings command to show active settings.
• Startup QOL improvements:
  • Sanity check the git repo and exit gracefully on problems.
  • Pause for confirmation after model sanity check to allow user to review warnings.
• Include important devops files in the repo map.
• Don't ask again in current session about a file the user has said not to add to the chat.
• Only print the pip command when self updating on Windows, without running it.
• Converted many error messages to warning messages.
• Added --tool-warning-color setting.
• Disabled built in linter for typescript.
• Catch /voice transcription exceptions, show the WAV file so the user can recover it.
• Adopted setuptools_scm to provide vX.Y.Z.dev version naming with git hashes.
• Share active test and lint commands with the LLM.
• Added --verbose debug output for shell commands.
• Do not fuzzy match filenames when LLM is creating a new file, by @ozapinq
• Only show cache warming status update if --verbose.
• Bugfix for shell commands on Windows.
• Refuse to make git repo in $HOME, warn user.
• Added --update as an alias for --upgrade.
• Bugfix to completions for /model command.
• Aider wrote 53% of the code in this release.

• Switched to vX.Y.Z.dev version naming.

• Improved printed pip command output on Windows.

• Bugfix to test command in platform info.

• Include important devops files in the repomap.
• Print quoted pip install commands to the user.
• Adopt setuptools_scm to provide dev versions with git hashes.
• Share active test and lint commands with the LLM.
• Catch and handle most errors creating new files, reading existing files.
• Catch and handle most git errors.
• Added --verbose debug output for shell commands.

• Startup QOL improvements:
  • Sanity check the git repo and exit gracefully on problems.
  • Pause for confirmation after model sanity check to allow user to review warnings.
• Bug fix for shell commands on Windows.
• Do not fuzzy match filenames when LLM is creating a new file, by @ozapinq
• Numerous corner case bug fixes submitted via new crash report -> GitHub Issue feature.
• Crash reports now include python version, OS, etc.

• Offer to submit a GitHub issue pre-filled with uncaught exception info.
• Bugfix for infinite output.

• New /settings command to show active settings.
• Only show cache warming status update if --verbose.

• Bugfix for shell commands on Windows.
• Refuse to make git repo in $HOME, warn user.
• Don't ask again in current session about a file the user has said not to add to the chat.
• Added --update as an alias for --upgrade.

• Bugfix to completions for /model command.
• Bugfix: revert home dir special case.

• Dependency watchdog<5 for docker image.

• When users launch aider in their home dir, help them find/create a repo in a subdir.
• Added missing pexpect dependency.

• Added model settings for gemini/gemini-1.5-pro-exp-0827 and gemini/gemini-1.5-flash-exp-0827.
• Shell and /run commands can now be interactive in environments where a pty is available.
• Optionally share output of suggested shell commands back to the LLM.
• New --[no-]suggest-shell-commands switch to configure shell commands.
• Performance improvements for autocomplete in large/mono repos.
• New --upgrade switch to install latest version of aider from pypi.
• Bugfix to --show-prompt.
• Disabled automatic reply to the LLM on /undo for all models.
• Removed pager from /web output.
• Aider wrote 64% of the code in this release.

• Keep your prompt cache from expiring with --cache-keepalive-pings.
  • Pings the API every 5min to keep the cache warm.
• You can now bulk accept/reject a series of add url and run shell confirmations.
• Improved matching of filenames from S/R blocks with files in chat.
• Stronger prompting for Sonnet to make edits in code chat mode.
• Stronger prompting for the LLM to specify full file paths.
• Improved shell command prompting.
• Weak model now uses extra_headers, to support Anthropic beta features.
• New --install-main-branch to update to the latest dev version of aider.
• Improved error messages on attempt to add not-git subdir to chat.
• Show model metadata info on --verbose.
• Improved warnings when LLMs env variables aren't set.
• Bugfix to windows filenames which contain \_.
• Aider wrote 59% of the code in this release.

• Bugfix for NameError when applying edits.

• Aider now offers to run shell commands:
  • Launch a browser to view updated html/css/js.
  • Install new dependencies.
  • Run DB migrations.
  • Run the program to exercise changes.
  • Run new test cases.
• /read and /drop now expand ~ to the home dir.
• Show the active chat mode at aider prompt.
• New /reset command to /drop files and /clear chat history.
• New --map-multiplier-no-files to control repo map size multiplier when no files are in the chat.
  • Reduced default multiplier to 2.
• Bugfixes and improvements to auto commit sequencing.
• Improved formatting of token reports and confirmation dialogs.
• Default OpenAI model is now gpt-4o-2024-08-06.
• Bumped dependencies to pickup litellm bugfixes.
• Aider wrote 68% of the code in this release.

• Prompt caching for Anthropic models with --cache-prompts.
  • Caches the system prompt, repo map and /read-only files.
• Repo map recomputes less often in large/mono repos or when caching enabled.
  • Use --map-refresh <always|files|manual|auto> to configure.
• Improved cost estimate logic for caching.
• Improved editing performance on Jupyter Notebook .ipynb files.
• Show which config yaml file is loaded with --verbose.
• Bumped dependency versions.
• Bugfix: properly load .aider.models.metadata.json data.
• Bugfix: Using --msg /ask ... caused an exception.
• Bugfix: litellm tokenizer bug for images.
• Aider wrote 56% of the code in this release.

<img width="871" alt="2024-08-20_09-31-40" src="https://github.com/user-attachments/assets/63f05c05-92b6-4f55-8c0f-a7f738480986">

• Bugfix for provider API exceptions.

• Infinite output for DeepSeek Coder, Mistral models in addition to Anthropic's models.
• New --deepseek switch to use DeepSeek Coder.
• DeepSeek Coder uses 8k token output.
• New --chat-mode <mode> switch to launch in ask/help/code modes.
• New /code <message> command request a code edit while in ask mode.
• Web scraper is more robust if page never idles.
• Improved token and cost reporting for infinite output.
• Improvements and bug fixes for /read only files.
• Switched from setup.py to pyproject.toml, by @branchvincent.
• Bug fix to persist files added during /ask.
• Bug fix for chat history size in /tokens.
• Aider wrote 66% of the code in this release.

• Bugfix to /help.

• Add read-only files to the chat context with /read and --read, including from outside the git repo.
• /diff now shows diffs of all changes resulting from your request, including lint and test fixes.
• New /clipboard command to paste images or text from the clipboard, replaces /add-clipboard-image.
• Now shows the markdown scraped when you add a url with /web.
• When scripting aider messages can now contain in-chat / commands.
• Aider in docker image now suggests the correct command to update to latest version.
• Improved retries on API errors (was easy to test during Sonnet outage).
• Added --mini for gpt-4o-mini.
• Bugfix to keep session cost accurate when using /ask and /help.
• Performance improvements for repo map calculation.
• /tokens now shows the active model.
• Enhanced commit message attribution options:
  • New --attribute-commit-message-author to prefix commit messages with 'aider: ' if aider authored the changes, replaces --attribute-commit-message.
  • New --attribute-commit-message-committer to prefix all commit messages with 'aider: '.
• Added openai/gpt-4o-2024-08-06.
• Worked around litellm bug that removes OpenRouter app headers when using extra_headers.
• Improved progress indication during repo map processing.
• Removed obsolete 16k token limit on commit diffs, use per-model limits.
• Aider wrote 61% of the code in this release.

• Added openai/gpt-4o-2024-08-06.
• Worked around litellm bug that removes OpenRouter app headers when using extra_headers.
• Improved progress indication during repo map processing.
• Corrected instructions for upgrading the docker container to latest aider version.
• Removed obsolete 16k token limit on commit diffs, use per-model limits.

• Performance improvements for large/mono repos.
• Added --subtree-only to limit aider to current directory subtree.
  • Should help with large/mono repo performance.
• New /add-clipboard-image to add images to the chat from your clipboard.
• Use --map-tokens 1024 to use repo map with any model.
• Support for Sonnet's 8k output window.
  • Aider already supported infinite output from Sonnet.
• Workaround litellm bug for retrying API server errors.
• Upgraded dependencies, to pick up litellm bug fixes.
• Aider wrote 44% of the code in this release.

• Improvements to conventional commits prompting.

• Commit message improvements:
  • Added Conventional Commits guidelines to commit message prompt.
  • Added --commit-prompt to customize the commit message prompt.
  • Added strong model as a fallback for commit messages (and chat summaries).
• Linting improvements:
  • Ask before fixing lint errors.
  • Improved performance of --lint on all dirty files in repo.
  • Improved lint flow, now doing code edit auto-commit before linting.
  • Bugfix to properly handle subprocess encodings (also for /run).
• Improved docker support:
  • Resolved permission issues when using docker run --user xxx.
  • New paulgauthier/aider-full docker image, which includes all extras.
• Switching to code and ask mode no longer summarizes the chat history.
• Added graph of aider's contribution to each release.
• Generic auto-completions are provided for /commands without a completion override.
• Fixed broken OCaml tags file.
• Bugfix in /run add to chat approval logic.
• Aider wrote 58% of the code in this release.

• Downgraded stray numpy dependency back to 1.26.4.

• New /ask <question> command to ask about your code, without making any edits.
• New /chat-mode <mode> command to switch chat modes:
  • ask: Ask questions about your code without making any changes.
  • code: Ask for changes to your code (using the best edit format).
  • help: Get help about using aider (usage, config, troubleshoot).
• Add file: CONVENTIONS.md to .aider.conf.yml to always load a specific file.
  • Or file: [file1, file2, file3] to always load multiple files.
• Enhanced token usage and cost reporting. Now works when streaming too.
• Filename auto-complete for /add and /drop is now case-insensitive.
• Commit message improvements:
  • Updated commit message prompt to use imperative tense.
  • Fall back to main model if weak model is unable to generate a commit message.
• Stop aider from asking to add the same url to the chat multiple times.
• Updates and fixes to --no-verify-ssl:
  • Fixed regression that broke it in v0.42.0.
  • Disables SSL certificate verification when /web scrapes websites.
• Improved error handling and reporting in /web scraping functionality
• Fixed syntax error in Elm's tree-sitter scm file (by @cjoach).
• Handle UnicodeEncodeError when streaming text to the terminal.
• Updated dependencies to latest versions.
• Aider wrote 45% of the code in this release.

• Use 4o-mini as the weak model wherever 3.5-turbo was used.

• GPT 4o mini scores like the original GPT 3.5 on aider's code editing benchmark, using whole edit format.
• Aider is better at offering to add files to the chat on Windows.
• Bugfix corner cases for /undo with new files or new repos.
• Now shows last 4 characters of API keys in --verbose output.
• Bugfix to precedence of multiple .env files.
• Bugfix to gracefully handle HTTP errors when installing pandoc.
• Aider wrote 42% of the code in this release.

• Default pip install size reduced by 3-12x.
• Added 3 package extras, which aider will offer to install when needed:
  • aider-chat[help]
  • aider-chat[browser]
  • aider-chat[playwright]
• Improved regex for detecting URLs in user chat messages.
• Bugfix to globbing logic when absolute paths are included in /add.
• Simplified output of --models.
• The --check-update switch was renamed to --just-check-updated.
• The --skip-check-update switch was renamed to --[no-]check-update.
• Aider wrote 29% of the code in this release (157/547 lines).

• Added scipy back to main requirements.txt.

• Added build-essentials back to main Dockerfile.

• Moved HuggingFace embeddings deps into [hf-embed] extra.
• Added [dev] extra.

• Use /help <question> to ask for help about using aider, customizing settings, troubleshooting, using LLMs, etc.
• Allow multiple use of /undo.
• All config/env/yml/json files now load from home, git root, cwd and named command line switch.
• New $HOME/.aider/caches dir for app-wide expendable caches.
• Default --model-settings-file is now .aider.model.settings.yml.
• Default --model-metadata-file is now .aider.model.metadata.json.
• Bugfix affecting launch with --no-git.

• Performance release:
  • 5X faster launch!
  • Faster auto-complete in large git repos (users report ~100X speedup)!

• Allow Claude 3.5 Sonnet to stream back >4k tokens!
  • It is the first model capable of writing such large coherent, useful code edits.
  • Do large refactors or generate multiple files of new code in one go.
• Aider now uses claude-3-5-sonnet-20240620 by default if ANTHROPIC_API_KEY is set in the environment.
• Enabled image support for 3.5 Sonnet and for GPT-4o & 3.5 Sonnet via OpenRouter (by @yamitzky).
• Added --attribute-commit-message to prefix aider's commit messages with "aider:".
• Fixed regression in quality of one-line commit messages.
• Automatically retry on Anthropic overloaded_error.
• Bumped dependency versions.

• Fixed /undo so it works regardless of --attribute settings.

• Bump versions to pickup latest litellm to fix streaming issue with Gemini
  • https://github.com/BerriAI/litellm/issues/4408

• Improved context awareness of repomap.
• Restored proper --help functionality.

• Improved prompting to discourage Sonnet from wasting tokens emitting unchanging code (#705).
• Improved error info for token limit errors.
• Options to suppress adding "(aider)" to the git author and committer names.
• Use --model-settings-file to customize per-model settings, like use of repo-map (by @caseymcc).
• Improved invocation of flake8 linter for python code.

• Use --sonnet for Claude 3.5 Sonnet, which is the top model on aider's LLM code editing leaderboard.
• All AIDER_xxx environment variables can now be set in .env (by @jpshack-at-palomar).
• Use --llm-history-file to log raw messages sent to the LLM (by @daniel-vainsencher).
• Commit messages are no longer prefixed with "aider:". Instead the git author and committer names have "(aider)" added.

• Use --vim for vim keybindings in the chat.
• Add LLM metadata via .aider.models.json file (by @caseymcc).
• More detailed error messages on token limit errors.
• Single line commit messages, without the recent chat messages.
• Ensure --commit --dry-run does nothing.
• Have playwright wait for idle network to better scrape js sites.
• Documentation updates, moved into website/ subdir.
• Moved tests/ into aider/tests/.

• Repo map is now optimized based on text of chat history as well as files added to chat.
• Improved prompts when no files have been added to chat to solicit LLM file suggestions.
• Aider will notice if you paste a URL into the chat, and offer to scrape it.
• Performance improvements the repo map, especially in large repos.
• Aider will not offer to add bare filenames like make or run which may just be words.
• Properly override GIT_EDITOR env for commits if it is already set.
• Detect supported audio sample rates for /voice.
• Other small bug fixes.

• Aider can now lint your code and fix any errors.
  • Aider automatically lints and fixes after every LLM edit.
  • You can manually lint-and-fix files with /lint in the chat or --lint on the command line.
  • Aider includes built in basic linters for all supported tree-sitter languages.
  • You can also configure aider to use your preferred linter with --lint-cmd.
• Aider has additional support for running tests and fixing problems.
  • Configure your testing command with --test-cmd.
  • Run tests with /test or from the command line with --test.
  • Aider will automatically attempt to fix any test failures.

• Aider now uses GPT-4o by default.
  • GPT-4o tops the aider LLM code editing leaderboard at 72.9%, versus 68.4% for Opus.
  • GPT-4o takes second on aider's refactoring leaderboard with 62.9%, versus Opus at 72.3%.
• Added --restore-chat-history to restore prior chat history on launch, so you can continue the last conversation.
• Improved reflection feedback to LLMs using the diff edit format.
• Improved retries on httpx errors.