todo.knowledge.md

[x] Should indicate if diff is still generating somehow... - send to client the file changes still planned. Wait for them before allowing user input?

• Improve speed by prompting it to rely more on diffs instead of reproducing the whole file [x] Prompt engineer it to automatically make knowledge changes when you correct it [x] it's doing a poor job loading all the relevant files into context. Need to add more explicity steps for this. [x] Indenting fails when it makes multiple changes with different indent levels. [x] More chain of thought for main prompt. blocks, <code_review>, , etc. [x] Delete files [x] Shift + Enter, and pasting in stuff with new lines. Update: not really possible

[x] Failed to replace in the new contents (but did expand it for the old content): // ... (keep the existing setup code) [x] Continuation 2 led to infinite responses [x] cancelling before the prompt starts being genereated leads to error of two user messages in a row

• Show commands available in a bar under the input (not sure if this is possible) [x] Escalate when old diff does not match. Got case where a helper function in the middle of two others was moved up, but diff didn't match because it ignored the helper function. [x] Stuck in loop: Prompting claude num messages: 21, continuing to generate [x] Backspace after typing more than one line is buggy

[x] Dawson's case of wanting it to scrape documentation from a url and answer questions with it.

• x% knowledge written in menu based on number of commits to knowledge files [x] How it could run bash:

1. First, let's build the common package:

cd common
yarn build
cd ..

Important. Can use commandline to search for code. Could move files.

[x] Try asking first if the complete file is listed, and then automatically generate diffs.

[] create some structure in a knowledge file for users to initially fill out.

• Project overview
• Goals
• Project structure
• Coding do's and don'ts

[] reverting doesn't work with created files [x] File paths using too much context? [x] Error on server didn't print the error (or maybe it did and the terminal overwrote it...) [x] Change ip address to be from api.manicode.ai

Notes from Janna & Stephen test [x] backspace on windows doesn't clear the char, just moves cursor [x] Tried to do left arrow and it didn't work [x] Failed to apply any edits. They all failed even after a second attempt. This was on Windows [x] CTRL-C doesn't work

[] Kill if it continues too far without user prompt. [] Prompt it not to generate the whole file when just making a local edit. Consider just reproducting the function edited. Or a block a code. - Before editing a file, get it to say what changes it will make and then edit just those sections. [x] Consider confirming with the user whether to go ahead and make a change if not that confident or want more input from the user [] Force updates: run the command to update app. [] Store previous user input's and always include that history in system prompt. - Can also allow you to press up arrow to go through whole history [x] Changes prompt is printing object for most previous messages in message history [x] It keeps requesting files that are already in its context. Need to highlight those paths again somewhere? - Requests a file before editing that it just got. [] Knowledge files should be treated more like regular files, but with system prompts to frequently include them

-- Instead, of below, apply edits immediately. Track all changes for a message via an id. Fix conccurent request by subscribing/unsubscribing to response with id. [x] Give it a new marker token that will await all file changes, so then it can run tsc or tests, etc. - It is using grep to see if changes went through, but they haven't so gets in a loop. - Realized we could just apply the edits before the tool call. We were already passing them through. [x] Was able to start a concurrent request after cancelling the last one... [] Changes behavior like removing if guard if statements at the top, removing cases of logs/errors, adds

. It should only do the minimal change and assume existing code is there for a reason. [x] Terminal command that doesn't finish bricks manicode [x] Easy to forget to run it in root directory. [x] Allow pasting new lines based on timing strategy [] It's not saving useful info like that this is a Windows machine and it shouldn't be using grep into a knowledge file. [x] Budget of tokens when reading files, and skip files that put it over the limit. [x] Still does too many extra things [x] Doesn't consistently request new files when it needs to [x] Scrape urls by default, like we request files by default [x] The user may have edited files since your last change. Please try to notice and perserve those changes. Don't overwrite these please! [x] Show a message when manicode is out of date. Or figure out how to automatically update. [] The "// add this blah" comments are really annoying. Strip them out in the diff step [x] The comprehensive files list generates a huge amount of thinking that is probably slow [x] Try a lower temperature. Might prevent it from doing random stuff. - apparently it's been 0 this whole time, huh (for claude) - Also openai defaults to 0 [] Add current file diff from main into context (git status/git diff?)

[x] It thought that update_file_context would create a file? (a knowledge file?) [] Claude doesn't leave a comment that there's more code in between when it should. Then lots gets deleted [] Try using the native stop marker feature [] Use tree sitter to include exported tokens ttps://github.com/tree-sitter/node-tree-sitter See how github implemented code search: ttps://docs.github.com/en/repositories/working-with-files/using-files/navigating-code-on-github#precise-and-search-based-navigation [] Multiple file blocks for the same file could fail to patch [] Still hallucinating without getting the file: Totally destroyed my package .json by removing random scripts & dependencies, changing a bunch of stuff around [] Create tests for the full user input. Set up fake project data for this. [] Use gpt-4o mini to grade tests

[] Updated a function in a file, but didn't include comments at the top or bottom for existing code [] Looks in wrong directory for a file, e.g. web/components/comments instead of web/components/buttons web/components/profile/profile-comments.tsx instead of web/components/comments/profile-comments.tsx [] Ari: editing is failing: deleting sections of code it shouldn't. - [J] Removes commented out code, other comments [] Doesn't give up control after running terminal commands. Just keeps running more [] Says reconnected, but actually doesn't go after [] Run in a containerls - Maybe distribute as brew package which would have more permissions? - Check files being edited to be within project dir (no '..') [x] Send a command, then just run it. if it goes through, print. - thefuck util that could change - should look more like shell - could have two panel one for you, other shows claude's response. Claude's commands go on your side [] Got file path wrong: backend instead of npm-app for project-files [] Still is overwritting a user's change every time