[Codex] codex.md adapter audit — MCP integration patterns unverified

[JeremyDev87]

Summary

packages/rules/.ai-rules/adapters/codex.md was written based on design intent, not observed behavior. Several sections describe workflows (MCP tool calls, AUTO mode, skill access) that have not been verified to actually work in Codex (GitHub Copilot Chat / Copilot Workspace). Additionally, the .codex/rules/system-prompt.md references MCP tools without verifying whether Copilot's AI models reliably invoke them.

Scope of Audit

1. MCP Tools — undocumented tools in codex.md

The following MCP tools exist but are not mentioned in codex.md:

Tool	Status in codex.md
parse_mode	❓ Not explicitly documented
dispatch_agents	❌ Missing
prepare_parallel_agents	❌ Missing
analyze_task	❌ Missing
update_context	❌ Missing
generate_checklist	❌ Missing
get_code_conventions	❌ Missing
list_skills / get_skill	❌ Missing (only cat documented)
suggest_config_updates	❌ Missing
search_rules	❌ Missing

2. Mode keyword detection

Current .codex/rules/system-prompt.md may reference PLAN/ACT/EVAL keywords, but:

• Does GitHub Copilot Chat reliably call parse_mode when the message starts with these keywords?
• Is there a mechanism equivalent to Claude Code's UserPromptSubmit hook to enforce parse_mode calls?

3. AUTO mode (codex.md lines 199–255)

Codex documents AUTO mode in detail, but:

• GitHub Copilot has no autonomous loop mechanism
• AUTO keyword handling depends on Copilot calling parse_mode, which is unverified
• The PLAN → ACT → EVAL cycle cannot be enforced without session continuity

4. Skills section (codex.md lines 127–197)

• cat approach only works in monorepo; get_skill MCP tool is not documented
• PR All-in-One Skill section references .claude/pr-config.json (Claude-specific path) — likely wrong for Codex
• /pr-all-in-one slash command syntax is Claude Code-specific and does not apply to Copilot

5. Context document management

• update_context is mandatory per CLAUDE.md but not documented in codex.md
• Without update_context calls, docs/codingbuddy/context.md will not persist PLAN → ACT decisions

Audit Method

For each item above:

1. Open a test project in VS Code with GitHub Copilot and codingbuddy MCP connected
2. Trigger the described behavior
3. Check MCP server logs (MCP_DEBUG=1) to confirm tool calls
4. Document actual vs. expected behavior

Acceptance Criteria

• Each documented pattern marked as: ✅ Verified / ⚠️ Partial / ❌ Not working
• codex.md updated to reflect actual behavior
• Claude Code-specific patterns (slash commands, Skill tool, .claude/ paths) clearly marked as not applicable to Codex
• Missing MCP tools added to a tools table in codex.md
• Known limitations section added for features unavailable in Copilot (Task tool, native Skill tool, session hooks, autonomous loops)
• .codex/rules/system-prompt.md updated with any missing tool instructions

Files Involved

• packages/rules/.ai-rules/adapters/codex.md — full audit target
• .codex/rules/system-prompt.md — system prompt rules
• apps/mcp-server/src/mcp/handlers/ — all handler files (enumerate actual available tools)