enhance: simplify the question agent flow

[a7m-1st]

Related Issue

Closes ##1540

Description

Issue 1

• Previously when a question gets sent, it was passing through two stages:

1. Question Agent decides if task is complex / no

is_complex_task = await question_confirm(
    question_agent, question, task_lock
)

2. Question Agent replies if task is simple

simple_resp = question_agent.step(simple_answer_prompt)
if simple_resp and simple_resp.msgs:
    answer_content = simple_resp.msgs[0].content

• The issue with this is speed ofcourse; so instead of making them separate prompts why can't the agent just return this?:

Respond in this exact format:
COMPLEXITY: [SIMPLE|COMPLEX]
ANSWER: [Direct answer only if SIMPLE. Leave blank if COMPLEX]"""

Issue 2

• Also, the question agent had almost no context of the environment. Thus in this PR I gave context to mcp and files attached additionally.

• I have left some TODO's regarding the old implementation. To be covered in upcoming PRs. They include the fact that options is frozen on multi turn, and attached files, mcps and other artifacts remains frozen among turns too. Human reply tool is also affected where no file information is passed along multi turn.

Testing Evidence (REQUIRED)

It doesn't break the SSE.

• Fun fact before you comment; TTFT actually is within 1 - 2 sec in the backend; I guess initialization stretched it longer. We can see that times were 16s, 9s, 8s. It is not very stable, but has overall improvement.

2026-04-06.05-47-04.mp4

• I have included human-verified testing evidence in this PR.
• This PR includes frontend/UI changes, and I attached screenshot(s) or screen recording(s).
• No frontend/UI changes in this PR.

What is the purpose of this pull request?

• Bug fix
• New Feature
• Documentation update
• Other

Contribution Guidelines Acknowledgement

• I have read and agree to the Eigent Contribution Guideline

[a7m-1st]

Update 080426:

1. Parallel calls; inside startTask. Also made history creation async (non awaited).
2. Caching: Fetching user keys was taking 2 seconds, so I have cached it for 5 minutes with TanStack query and invalidated when chat ends SSE and invalidate with default events by default (chatBox/index.tsx load, refocus & model change)

[a7m-1st]

Update

• Model backend hashing implemented for cross project reuse. Lowers the question agent time to:
• 5 seconds for "hi" (from ~14 seconds)
• 2:41 min for "write for me hello word text file then save it." (from ~3:30 min)

[a7m-1st]

This added to bypass error. This event is not used anymore; patching it in #1438 .