You signed in with another tab or window. Reload to refresh your session.You signed out in another tab or window. Reload to refresh your session.You switched accounts on another tab or window. Reload to refresh your session.Dismiss alert
Repository map (Aider, continue) allows LLM automatically understand high-level structure of a codebase to be able to query some detailed info about it if needed
RAGs are also allows to optimize LLM work. Tool indexes the whole codebase with embedding LLM, then on user request it appends the most similar files to the context
As far as I can see there are no these features right now. Do you have plans to support them?
@rekram1-node what if my codebase is a large monorepository and I'd like to send only part of it? Also, the files list doesn't provide any info about files content (repository map provides it, for example; RAG too)
@vldF - I was wondering about this as well. I think I could build this functionality outside of core as a plugin/tool, and then we could update the global Agents.md file to make sure it uses the RAG before reaching out to the internet, or using other tools.
@rekram1-node, I am pretty sure this is an abuse of the system prompt esspecially in non-trivial workflows (see #3112) where the context will change as the workflow progresses.
But perhaps there is a better issue to track that discussion?
it is pretty standard in coding agents to include information about the current project, can you outline why it is an abuse? I could see argument for user prompt usage tho
Only in very narrow use cases will the relevant files be unchanging across the session:
... esspecially in non-trivial workflows (see #3112)
When /add lands the system prompt should only include such data if the user configures it, default could be system_prompt_files: true, but in general users will employ /add ... at the start of a CLI session, or in more complex workflows the agent/subgents will /add ...; /drop ... as the workflow progresses.
Activity
rekram1-node commentedon Aug 20, 2025
@vldF we do inject information about environment to llm like your project tree, example (this is excerpt from opencode repo):
^^ this is injected into system prompt
vldF commentedon Aug 20, 2025
@rekram1-node what if my codebase is a large monorepository and I'd like to send only part of it? Also, the files list doesn't provide any info about files content (repository map provides it, for example; RAG too)
rekram1-node commentedon Aug 20, 2025
We only list up to 200 files
correct
justindonnaruma commentedon Aug 20, 2025
@vldF - I was wondering about this as well. I think I could build this functionality outside of core as a plugin/tool, and then we could update the global Agents.md file to make sure it uses the RAG before reaching out to the internet, or using other tools.
Volko61 commentedon Aug 30, 2025
According to https://www.youtube.com/watch?v=m0X_hmLBhwo
It might be very relevent to use AST
taqtiqa-mark commentedon Oct 14, 2025
You might like to add your feedback on the
/addcommand being discussed here:#1990
@vldF is #1990 sufficiently on-point for you to close this issue and track developments there?
taqtiqa-mark commentedon Oct 14, 2025
@rekram1-node, I am pretty sure this is an abuse of the system prompt esspecially in non-trivial workflows (see #3112) where the context will change as the workflow progresses.
But perhaps there is a better issue to track that discussion?
rekram1-node commentedon Oct 14, 2025
it is pretty standard in coding agents to include information about the current project, can you outline why it is an abuse? I could see argument for user prompt usage tho
taqtiqa-mark commentedon Oct 14, 2025
Agreed, the question is which prompt does the information belong.
From the LLM view this is: System-Prompt (one-off), or User-Prompt (multiple)?
Ack. this can be Vendor dependent: X.ai has collections, but I don't believe the node
aipackage supports this yet?In the node
aiabstractions this is: System-Message vs User-Message vs Assistant-Message.Using the abstractions OC introduces this is, IIUC: Mode-Prompt (current behavior?), Agent-Prompt-Text or Context-Text
Generally, "The system prompt to use that specifies the behavior of the model." A list of files does not specify the behavior of the model.
See my earlier doc suggestions
Only in very narrow use cases will the relevant files be unchanging across the session:
When
/addlands the system prompt should only include such data if the user configures it, default could besystem_prompt_files: true, but in general users will employ/add ...at the start of a CLI session, or in more complex workflows the agent/subgents will/add ...; /drop ...as the workflow progresses.vldF commentedon Oct 16, 2025
Yeah, let's move to another discussion