OpenQode/Documentation/implementation_plan_multi_agent.md

Real-Time Multi-Agent Visualization Plan

Goal

Show which specific agent (e.g., Security, Planner, Builder) is currently active in real-time during the "Thinking" phase.

User Review Required

Tip

This relies on "Prompt Engineering" to force the AI to self-report its agent usage. It acts like a "verbose mode" for the internal router.

Proposed Changes

1. Prompt Injection (opencode-ink.mjs)

• Check: if (multiAgentEnabled) in prompt construction.
• Append:

  [MULTI-AGENT LOGGING]
  When delegating to a sub-agent or switching context, you MUST start the line with:
  [AGENT: <AgentName>] <Action description>
  Example:
  [AGENT: Security] Scanning for auth vulnerabilities...
  [AGENT: Planner] Breaking down the task...
  

2. Stream Parsing (opencode-ink.mjs)

• Logic:
  • Regex: /\[AGENT:\s*([^\]]+)\]/i
  • If match found: Update thinkingStats.activeAgent.

3. UI Update (ThinkingBlock.mjs)

• Visual:
  • If stats.activeAgent is present, display: 🧠 Thinking (<activeAgent>)... instead of just Thinking....

Verification Plan

1. Setup: Enable /agents mode.
2. Trigger: Ask "Plan a secure login system and check for bugs".
3. Verify: Watch the Thinking Block switch from Thinking... -> Thinking (Planner)... -> Thinking (Security)....