[Codex] Parallel specialist agents execution pattern not defined

[JeremyDev87]

Summary

The codingbuddy PLAN/EVAL modes recommend parallel specialist agent execution (e.g., security-specialist + accessibility-specialist + performance-specialist). In Claude Code, this is implemented via the native Task tool with run_in_background: true. Codex (GitHub Copilot) does not have a Task tool or equivalent subagent spawning mechanism.

The dispatch_agents and prepare_parallel_agents MCP tools return ready-to-use dispatch parameters, but there is no documented or tested pattern for how Codex should consume these results.

Context

What Claude Code does (working)

parse_mode → dispatchReady.parallelAgents[] → Task tool (run_in_background: true) × N → TaskOutput → consolidate

What Codex receives (untested)

parse_mode returns dispatchReady.parallelAgents[] with dispatchParams for each agent, but Codex has no mechanism to spawn background subagents. The AI in Codex likely:

• Ignores the dispatch params entirely, OR
• Attempts sequential processing without clear structure

Current Adapter Documentation Gap

packages/rules/.ai-rules/adapters/codex.md has no section on parallel specialist agent execution. The MCP tools (dispatch_agents, prepare_parallel_agents) are not mentioned.

What Needs to Be Investigated

1. When parse_mode returns parallelAgentsRecommendation in Codex, what does GitHub Copilot's AI actually do?
2. Can GitHub Copilot Chat spawn multiple simultaneous agent sessions or panels?
3. Is sequential specialist execution a viable fallback? What is the UX impact?

Proposed Solution

Option A: Sequential specialist execution (recommended for now)

Document a Codex-specific pattern in codex.md:

parse_mode returns parallelAgentsRecommendation
↓
Call prepare_parallel_agents to get specialist prompts
↓
For each specialist (sequentially):
  - Announce: "Running [specialist-name] analysis..."
  - Apply specialist context and analyze
  - Record findings
↓
Consolidate all findings into summary

Option B: GitHub Copilot Workspace (if available)

If Copilot Workspace supports spawning parallel sub-tasks, document how to use it with dispatchReady.parallelAgents[].dispatchParams.

Acceptance Criteria

• Actual Codex behavior when receiving parallelAgentsRecommendation is documented
• A working sequential fallback pattern is documented in packages/rules/.ai-rules/adapters/codex.md
• .codex/rules/system-prompt.md updated with explicit instructions for specialist execution in Codex
• OR: a Codex-native parallel approach is found and documented

Files Involved

• packages/rules/.ai-rules/adapters/codex.md — Parallel Specialist Agents section (missing)
• .codex/rules/system-prompt.md — may need sequential specialist instructions
• apps/mcp-server/src/mcp/handlers/agent.handler.ts — dispatch_agents, prepare_parallel_agents
• apps/mcp-server/src/mcp/handlers/mode.handler.ts — buildDispatchReady()