gemini "i'm finished" loop #262

Activity

1. adamdotdevin

   assigned

   thdxr

   on Jun 20, 2025
2. nefthy commented on Jun 21, 2025

   

   nefthy

   on Jun 21, 2025

   More actions

   I can confirm. I get an endlessly repeating stream of messages like:
   he build was successful. I have implemented the logic to show and hide the "No cards yet" message. The build was  successful. I have implemented the logic to show and hide the  "No cards yet" message. The build was successful.
3. unkhz commented on Jun 24, 2025

   

   unkhz

   on Jun 24, 2025

   Last edited by unkhz

   Contributor

   More actions

   Been trying this out for a few days. Endless loop happening with local models very often for me. I've been experimenting only with self hosted models (lm studio) and they rarely get past the first task without ending up in a loop. Some examples of models I've tried

   • qwen/qwen3-30b-a3b
   • deepseek-coder-33b-instruct-hf-mlx
   • mistralai/devstral-small-2505

   A workaround seems to be adding "just do it, don't ask for permission" in the end of the prompt, in which case it can do more than one thing, since it'll not stop for more input.

   

   current config

   {
     "$schema": "https://opencode.ai/config.json",
     "provider": {
       "lmstudio": {
         "npm": "@ai-sdk/openai-compatible",
         "options": {
           "baseURL": "http://localhost:8080/v1"
         },
         "models": {
           "qwen/qwen3-30b-a3b": {
             "limit": { "context": 131072, "output": 1000000 }
           },
           "mistralai/devstral-small-2505": {
             "limit": { "context": 131072, "output": 1000000 }
           }
         }
       }
     }
   }
   
4. unkhz commented on Jun 24, 2025

   

   unkhz

   on Jun 24, 2025

   Contributor

   More actions

   Ok, seems I've shot myself in the foot with the output limit. Changed the output tokens from 1000000 to 4096 and I'm not seeing this anymore. Tried both models in above config and they work as expected.

   However I think there's a condition where this infinite repetition occurs even without absurd output limit. Seems the absurd output limit is the easiest way to reproduce it. I'm thinking it's related to the summarization logic, which recurses back to the chat based on comparing the limits to current token count: https://github.com/sst/opencode/blob/9cfb6ff9646553d6dcbcbe22844ffbeb6480dc12/packages/opencode/src/session/index.ts#L280-L299

   Might be something I could attempt creating a fix for.
5. unkhz commented on Jun 26, 2025

   

   unkhz

   on Jun 26, 2025

   Contributor

   More actions

   The first PR was pretty much a cosmetic sanity check. I think the larger problem is getting into a situation where summary accompanied with input and cache reads/writes grows larger than 90% of the token limit, eventually causing this infinite loop. This is a real consideration for small local models, but not so much for the $$$ models.

   One option is to take control of the summary size. We possibly can use maxTokens parameter of streamText to limit to certain size and then take the last summary size into account (and even when it was last run) when deciding "should I run auto summarize again?". A more robust logic like this might do the trick of getting rid of this problem for real.
6. rekram1-node commented on Dec 28, 2025

   

   rekram1-node

   on Dec 28, 2025

   Contributor

   More actions

   [automated] Closing due to 90+ days of inactivity. Feel free to reopen if you still need this!
7. rekram1-node

   closed this as completedon Dec 28, 2025
8. andreipromarketing-dev

   added a commit that references this issue on Apr 7, 2026

   Merge pull request anomalyco#262 from shimo4228/feat/skills/search-first

   ...

   7e57d1b
9. bussard76

   added a commit that references this issue on May 12, 2026

   fix: display saved command immediately without app restart (anomalyco#11

   ...

   dc9e5e7