feat request: add AskUserQuestion tool

[virattt]

Like Claude Code, we want to add a tool that will let Dexter ask the user a follow-up / clarifying question to get more context when doing research (or before embarking on a research deep dive).

[FnSK4R17s]

@virattt
This is a great addition. Before building the tool itself though, I think there's a deeper issue that would undermine it even if implemented: the context compaction appears to be summarizing user responses rather than preserving them verbatim.
I tested a multi-turn interview flow (asked Dexter to research AAPL, told it to ask me questions one by one) and the agent kept re-asking questions I'd already answered — time horizon, investment style, etc. Looking at the code, it seems like the conversation history is being stored as summaries rather than being fed back as exact messages on the next turn. The specific user choices get lost in compaction.
An AskUserQuestion tool would hit the same problem — the agent asks a clarifying question, the user answers, the answer gets compacted into a vague summary, and the research phase doesn't have the precise context it needs.
Suggested approach:

Fix first: separate compaction into two tiers — keep user interaction messages verbatim, only compact completed tool results
Then build the tool: once the agent can actually remember answers, the AskUserQuestion tool becomes straightforward

Happy to take a crack at either or both if you're open to it. I work on agent orchestration tooling and have run into this exact state management problem before.

[virattt]

@FnSK4R17s are you on the latest version?

We recently updated the context compaction

[FnSK4R17s]

@virattt Yes I ran the code from the main branch:

sk4r@LAPTOP-TR1AI7PB:/apps/dexter$ git log -1
commit 9cc7bf03a3d605dd8eab21abd96b6df53de6f63b (HEAD -> main, origin/main, origin/HEAD)
Merge: a4370c5 c26ccf7
Author: Virat Singh <virat.dot@gmail.com>
Date:   Tue Feb 10 10:05:34 2026 -0800

    Merge pull request #115 from swrzalek/feature/log-api-source-in-errors
    
    feat: add log source

[virattt]

Thanks for digging into this and testing the multi-turn flow. The re-asking behavior is a real issue we want to fix. I looked through the code and the cause is a bit different from what you described; here’s what’s actually going on and what would help.

Compaction: Context compaction in Dexter only applies to tool results (scratchpad); it never touches or summarizes user messages. So the issue isn’t compaction summarizing user responses.

History: Conversation history is fed back as verbatim user queries via getUserMessages() in buildInitialPrompt(). We do store LLM-generated summaries per turn, but they aren’t used when building the prompt (the code that would use them isn’t called).

Re-asking: The more likely cause is that we only pass user messages (“Previous user queries for context”) and don’t pass assistant messages. So the model sees “5 years” but not “What’s your time horizon?” — the Q&A pairing is lost. Fixing that (e.g. include assistant turns or explicit Q&A pairs in the initial prompt) should help before or alongside an AskUserQuestion tool.
That way you correct the “old flow” misconception (compaction + summaries as the cause) and point to the real design gap (missing assistant-side context).

[virattt]

Could definitely use some help on the re-asking problem if you're up to it. Followed by an AskUserQuestion tool 🙏

[FnSK4R17s]

Sounds good to me. I can work on both. Thanks ✌️

[FnSK4R17s]

Hey @virattt

Does this code change make sense to you? Please share your feedback. Thanks

#133

[FnSK4R17s]

While working on the conversation history fix, I noticed the eval suite requires a Financial Datasets subscription. This means external contributors can't run tests before submitting PRs. Want me to set up a parallel test suite scoped to the free tickers (AAPL, NVDA, MSFT, GOOGL, TSLA)? Would lower the barrier for community contributions.