Description
I have been using Kilo Code with local codebase indexing (implemented by @shssoichiro ) enabled for a great while and I have observed that it barely uses semantic_search at all, even for very open-ended question such as "How is authentication implemented in Kilo Code?". Previously in the legacy version of Kilo Code, it would have used codebase_search (the old name to current semantic_search) for such questions with significantly higher probability.
To provide an example while using Claude Opus 4.6:
I asked Kilo Code to provide me an answer to the prompt "How is authentication implemented for @kilocode-fork/ ?". The first thing it immediately does is to create an Explore Agent that in turn uses glob/grep commands before using a lot of read operations and further glob/grep commands (see the image for full context):
Ideally, it should have started to use the semantic_search tool to get some better suggestions on files it should start digging into using e.g. grep/glob/read, but this is not the case.
I did then confront Kilo Code and asked about why it didn't use semantic_search in this scenario and then I got the following answer:
It seems to me that it treats e.g. Explore Agent and semantic_search tool as opposing to each other and only one of them can be used at a time during the session instead of using these two altogether as separate calls. If my memory isn't betraying me here, surely do the subagents have access to semantic_search that they could call? Strange that the Explore Agent seems to omit the semantic_search tool when not explicitly asked to do use it.
Plugins
VS Code
Kilo version
7.2.52
Steps to reproduce
No response
Screenshot and/or share link
No response
Operating System
Windows 11
Terminal
Git Bash in Windows Terminal
Description
I have been using Kilo Code with local codebase indexing (implemented by @shssoichiro ) enabled for a great while and I have observed that it barely uses
semantic_searchat all, even for very open-ended question such as "How is authentication implemented in Kilo Code?". Previously in the legacy version of Kilo Code, it would have usedcodebase_search(the old name to currentsemantic_search) for such questions with significantly higher probability.To provide an example while using Claude Opus 4.6:
I asked Kilo Code to provide me an answer to the prompt "How is authentication implemented for @kilocode-fork/ ?". The first thing it immediately does is to create an Explore Agent that in turn uses glob/grep commands before using a lot of read operations and further glob/grep commands (see the image for full context):
Ideally, it should have started to use the
semantic_searchtool to get some better suggestions on files it should start digging into using e.g. grep/glob/read, but this is not the case.I did then confront Kilo Code and asked about why it didn't use
semantic_searchin this scenario and then I got the following answer:It seems to me that it treats e.g. Explore Agent and
semantic_searchtool as opposing to each other and only one of them can be used at a time during the session instead of using these two altogether as separate calls. If my memory isn't betraying me here, surely do the subagents have access tosemantic_searchthat they could call? Strange that the Explore Agent seems to omit thesemantic_searchtool when not explicitly asked to do use it.Plugins
VS Code
Kilo version
7.2.52
Steps to reproduce
No response
Screenshot and/or share link
No response
Operating System
Windows 11
Terminal
Git Bash in Windows Terminal