Hey, I was thinking about something.
We already have a mechanism to automate issue labels, and it works really well.
The idea is not to make a bot that decides, closes issues, or gives random "try this" answers. Just a small bot that can drop the right docs link when it is obvious.
For example:
- WebView / SABR / playback issue: link the WebView page and troubleshooting.
- Backup / restore / import: link the backup docs.
- Prerelease / update issue: link the prerelease page.
- Downloads, filters, accounts: same thing.
Something short like: "Maybe this can help while we check the issue: ...". If the bot is not sure, it says nothing. No auto-close, no pretending it knows the real cause, no huge wall of text from the bot.
The nice part is that the bot only needs to route the issue. Links and messages can come from a small reviewed map in the repo, so it cannot invent a doc page or some weird technical advice.
For the model side, we can use OpenRouter free models, but I do not think we should pick one from a marketing page. I can try a few of them on real closed PipePipe issues first: can they pick the right docs page, avoid a wrong comment, and follow the tiny output format? Then we keep the winner in a repo variable.
If you think the idea is worth trying, I can make a small POC on my fork. When it is finished, come have a look and tell me what you think, I think it could be nice to try :)
Hey, I was thinking about something.
We already have a mechanism to automate issue labels, and it works really well.
The idea is not to make a bot that decides, closes issues, or gives random "try this" answers. Just a small bot that can drop the right docs link when it is obvious.
For example:
Something short like: "Maybe this can help while we check the issue: ...". If the bot is not sure, it says nothing. No auto-close, no pretending it knows the real cause, no huge wall of text from the bot.
The nice part is that the bot only needs to route the issue. Links and messages can come from a small reviewed map in the repo, so it cannot invent a doc page or some weird technical advice.
For the model side, we can use OpenRouter free models, but I do not think we should pick one from a marketing page. I can try a few of them on real closed PipePipe issues first: can they pick the right docs page, avoid a wrong comment, and follow the tiny output format? Then we keep the winner in a repo variable.
If you think the idea is worth trying, I can make a small POC on my fork. When it is finished, come have a look and tell me what you think, I think it could be nice to try :)