Aider-AI
aider will get into a loop trying to answer a question
Issue
A couple of times aider would response by start to loop and print out the answer very slowly as if it is calling the LLM for every word and it consumes tokens a high rate (take note of simple answer required but token sent went from 22k to 313k) and never quite finish the reply, I would CTL-C it since it takes too long just to print out a sentence like:
Tokens: 22k sent, 21k cache write, 2.4k received. Cost: $0.12 message, $0.21 session. Applied edit to main.py Commit 6014826 feat: Improve timestamp extraction from call log
pip install fuzzywuzzy python-Levenshtein Run shell command? (Y)es/(N)o/(D)on't ask again [Yes]: n You can use /undo to undo and discard each aider commit. ──────────────────────────────────────────────────────────────────────────────────────────────────────────────────────── Qprompt.txt main.py prompt.txt
/ask did you make the changes in the script?
Yes Yes, Yes, Yes, Yes, Yes, Yes, Yes, Yes, Yes, Yes, Yes, Yes,
^C again to exit
Tokens: 313k sent, 13 received. Cost: $0.22 message, $0.43 session.
another example:
Tokens: 21k sent, 19k cache hit, 964 received. Cost: $0.03 message, $0.19 session. ──────────────────────────────────────────────────────────────────────────────────────────────────────────────────────── Qprompt.txt main.py prompt.txt
/ask what happens if the same sentence is spoken several times in the diaglogue text, will this method of matching timestamp will lead to errors?
You You've You've identifie You've identifie a You've identifie a vali You've identifie a vali concern You've identifie a vali concern. You've identifie a vali concern. You've identifie a vali concern. You've identifie a vali concern. You've identifie a vali concern.
^C again to exit
Tokens: 257k sent, 11 received. Cost: $0.21 message, $0.40 session.
Version and model info
Aider v0.58.0
Main model: claude-3-5-sonnet-20240620 with diff edit format, prompt cache, infinite output
Weak model: claude-3-haiku-20240307
Git repo: .git with 19 files
Repo-map: using 1024 tokens, files refresh
Use /help
aider invoked with --cache-prompts --no-stream --cache-keepalive-pings 5 --model claude-3-5-sonnet-20240620
I'm experiencing this after upgrading to version 0.5.8 as well. I never experienced this with older versions.
Thanks for trying aider and filing this issue.
Can you share the output where it happened to you, along with the announce lines and any active params/switches?
I am not currently able to reproduce this issue. If you find a way to reliably reproduce the problem, please let me know. I will continue to investigate.
It happens every two or three prompts for me, until I /exit and start again. So I have no idea why it is doing this.
A few things to note: I install this version using pipx:
pipx install aider-chat --force --pip-args "--extra-index-url https://download.pytorch.org/whl/cu118"
Also I invoke this for sonnet as:
aider --anthropic-api-key sk-xxxxxxxxxxxxxx --cache-prompts --no-stream --cache-keepalive-pings 5 --model claude-3-5-sonnet-20240620
See if this will trigger that issue.
Looks like this is similar to what I was seeing but I made the mistake of walking away and chewed through millions of tokens:
https://github.com/paul-gauthier/aider/issues/1842
I'll add my switches to my post. Since mine was a clean start on a new example project, it might be easier to reproduce for Paul.
Am I correct that you were all using the /architect chat command, rather than invoking aider with --architect ?
Nope, never used the /architect chat command. I was just try to fix up using the usual /add main.py /ask etc. This error is very consistent for me, it comes up in about 2 to 3 prompts.
Sounds like folks are seeing this without using architect mode. And without using /ask or /architect style single-message mode switch command?
Have you been seeing this with streaming turned on or off, or both?
Looks like maybe our common denominator is: --cache-prompts --no-stream --cache-keepalive-pings 5
I used /ask a lot just to check what aider will do, before I gave it the go ahead, saves me a lot of /undo
Have any of you checked to see if the token charges were real? Do they show up in your Anthropic dashboard? Just trying to see if it's an internal aider token counting issue or really a loop sending to the LLM.
Looks like it only received 1 token? The bottom shows my first task /ask 1606 tokens and then next task "update main.py" 26 tokens, thereafter 1 token (is it keep-alive or just "stuck" repeating?)
1 Oct 2024, 12:36 am | claude-3-5-sonnet-20240620 | 1.01 | Default | 23127 | 1 | HTTP | 1 Oct 2024, 12:36 am | claude-3-5-sonnet-20240620 | 0.99 | Default | 23125 | 1 | HTTP | 1 Oct 2024, 12:36 am | claude-3-5-sonnet-20240620 | 0.81 | Default | 23127 | 1 | HTTP | 1 Oct 2024, 12:35 am | claude-3-5-sonnet-20240620 | 2.36 | Default | 23124 | 1 | HTTP | 1 Oct 2024, 12:32 am | claude-3-5-sonnet-20240620 | 2.29 | Default | 22441 | 1 | HTTP | 1 Oct 2024, 12:27 am | claude-3-haiku-20240307 | 0.63 | Default | 2289 | 26 | HTTP | 1 Oct 2024, 12:27 am | claude-3-5-sonnet-20240620 | 21.12 | Default | 21469 | 1606 | HTTP
Time (GMT+8) | Model | Model latency | Workspace | Input Tokens | Output Tokens | Type | Error
1 Oct 2024, 12:36 am claude-3-5-sonnet-20240620 1.01 Default 23127 1 HTTP 1 Oct 2024, 12:36 am claude-3-5-sonnet-20240620 0.99 Default 23125 1 HTTP 1 Oct 2024, 12:36 am claude-3-5-sonnet-20240620 0.81 Default 23127 1 HTTP 1 Oct 2024, 12:35 am claude-3-5-sonnet-20240620 2.36 Default 23124 1 HTTP 1 Oct 2024, 12:32 am claude-3-5-sonnet-20240620 2.29 Default 22441 1 HTTP 1 Oct 2024, 12:27 am claude-3-haiku-20240307 0.63 Default 2289 26 HTTP 1 Oct 2024, 12:27 am claude-3-5-sonnet-20240620 21.12 Default 21469 1606 HTTPTime (GMT+8) Model Model latency Workspace Input Tokens Output Tokens Type Error 1 Oct 2024, 12:45 am claude-3-5-sonnet-20240620 1.08 Default 23593 1 HTTP
Looks like it only received 1 token?
These look very much like ping keepalive packets, those are capped at 1 token via max_tokens, though their timing look a bit odd.
Do these loops happen with prompt cache/keepalive turned off too?
The fix is available in the main branch. You can get it by installing the latest version from github:
aider --install-main-branch
# or...
python -m pip install --upgrade git+https://github.com/paul-gauthier/aider.git
If you have a chance to try it, let me know if it works better for you.
Since this fix is now released in 0.58.1, I think it is time to close this issue.
Thank you all for discovering and helping fix this problem.
I had an infinite loop scenario today, when starting aider in a deleted folder.
Tokens: 1186k sent, 0 received. Cost: $3.56 message, $11.64 session.
Can you share the announce lines aider prints on launch and some example output of the loop?
I'm seeing similar, here's a copy/paste from my .aider.chat.history.md
Note: This implementation provides the basic structure for authentication middleware. For a complete solution, you would need:
- Actual session management with cookies/tokens
- Database integration for user storage
- Real Google OAuth flow implementation
- Proper error handling and redirect logic
The current code shows how to set up the middleware pattern but doesn't include full authentication details which require backend configuration.
User:
I want to add a new page that requires authentication, so I need to create an auth middleware or similar solution.
Assistant:
To implement authentication middleware for protected routes, we'll:
- Create
src/middleware.ts- Next.js middleware to check authentication- Add
src/lib/auth.ts- Helper functions for session management- Update
src/app/page.tsx- Add logout functionalityThe changes involve:
- Creating a middleware that checks for valid sessions on protected routes
- Adding auth helper functions for login/logout
- Implementing proper redirect behavior when unauthenticated
- Updating the home page to show user info and logout option
Here are the exact code changes:
Then the User and Assistant interaction loops endlessly.
