Real-time user clarifications/modifications/pending comments during active agent tasks

[PratikCreates]

Feature Description

The Challenge Currently, Kiro's agentic flow is a one-way street once execution begins. If a user spots a misunderstanding or needs to pivot (e.g., changing a host limit or a duration range), they are forced to wait for completion or restart the job entirely.

The Request Enable a "Pending Comments" queue that allows users to provide real-time modifications while the agent is active. Instead of being a "black box," the agent should check for user overrides immediately after every "Thinking" or "Tool" cycle.

This allows the agent to:
Self-Correct: Acknowledge mid-stream changes without killing the process.
Reconcile Inputs: Aggregate multiple pending notes to ensure the final output reflects the most recent user intent.

Use Case

The Correction: Kiro starts working on a 10-20 minute duration range. The user types "min duration can be 1 minute too" while it’s still editing. Kiro picks this up at the next step and adjusts the logic.

The Refinement: During a logo generation task, the user adds "no black" and "no green" as the agent is thinking. Kiro consumes these constraints before the final execution, avoiding a useless generation.

Additional Context

:Image 1: Perplexity Inspiration
Shows a "Working" state with a 3-minute countdown. It features a persistent "Add details or clarifications" input field with a specific user correction: "No I meant 5". This illustrates the need for a non-blocking UI where corrections can be queued during long-running tasks.

Image 2: Kiro Mid-Task Friction
Demonstrates Kiro actively editing files like data_models.py and voice_library.py. While the agent is "Working," the user is forced to send separate chat messages like "Max is 20 mins" and "min duration can be 1 minute too". Without a pending queue, these corrections may be missed or require a full restart of the multi-step implementation.

Image 3: Antigravity "Pending Comments" Model
Displays the "Antigravity" approach where an agent is "Generating with Gemini". Below the active thought/generation block, there is a dedicated "Pending comments" section holding two distinct user inputs: "no black" and "no green". Each comment is timestamped and queued to be consumed by the agent once its current "Thinking" or "Tool" cycle completes, ensuring the final output reconciles all recent changes.

Also Gusto had expressed his pain to solve the problem in discord.
My probable solution:
1)SLM-Driven Reconciliation To ensure the main agent doesn’t receive conflicting or messy instructions mid-task, a Small Language Model (SLM) could act as a local gatekeeper:

2)Conflict Detection: The SLM compares "Pending Comments" against the active system prompt and previous thought blocks to identify contradictions.

3)Prompt Merging: If multiple comments are added (e.g., "no black" then "no green"), the SLM merges them into a single coherent instruction.

4)State Management: The SLM determines if the update is a "minor tweak" (inject at next cycle) or a "critical pivot" that requires pausing the agent for user confirmation.

5)Also since it is just to find conflict exists so maybe we just compare the last 5 prompts and see if active prompt has conflict with scheduled prompt