[Feature Request]: Mid-Task Message Queuing, Dynamic Sub-Agent Steering & Auto-Fetch Provider Models Endpoint

Overview & Summary

This request aggregates three high-impact workflow and orchestration enhancements for ZCode submitted by community member Lalit Sharma:

Feature 1: Non-Interrupting Mid-Task Input Queuing (Turn Steering)

• Problem: Currently in ZCode, entering a new message/point while the agent is coding either forces an immediate abort of the ongoing execution turn or delays registration until the entire multi-step task terminates.
• Proposed Enhancement:
  • Implement a turn queue / mid-task ingestion pipeline (similar to Claude Code / Codex).
  • Allow users to submit instructions while the agent is running; the orchestrator consumes queued instructions at the next tool execution checkpoint without killing the current task or requiring manual stop/restart.

Feature 2: Dynamic Mid-Flight Sub-Agent Steering

• Problem: When a sub-agent is actively working, the user cannot steer or redirect it via the main orchestrator without completely stopping the sub-agent task.
• Proposed Enhancement:
  • Enable dynamic mid-execution messaging between the main orchestrator and running sub-agents.
  • Allow users to tell the main model: "Tell the sub-agent to adjust X", and have the main agent pass an updated instruction into the sub-agent’s active tool loop cleanly.

Feature 3: Provider Setup: Auto-"Fetch" Models Button with Capability Tags

• Problem: Custom provider setup currently requires users to manually type or copy-paste each model identifier.
• Proposed Enhancement:
  • Add a "Fetch" button adjacent to the "Add Model" button in the Provider Settings interface.
  • Clicking "Fetch" performs a GET {base_url}/models call, populates a multi-select dropdown with checkboxes, and auto-detects capabilities (e.g. tagging Vision / Tool-calling / Thinking support).

Community Context & Origin

• Reported by: Lalit Sharma on Z.ai Community Discord
• Original Thread: "Please oblige by making these three changes... Adding these three features would be a game-changer..."

Start by locating the orchestrator and sub-agent execution loops, the user message handling path, and the Provider Settings interface. Define completion around queued instructions being consumed without aborting work, steerable active sub-agents, and a Fetch control that retrieves /models and presents selectable models with capability tags.