fix(codex): Transform Task agent calls to skill references

[terry-li-hm]

Summary

Adds comprehensive content transformation to the Codex converter, bridging syntax differences between Claude Code and Codex.

Problem

Claude Code and Codex have different syntax for invoking commands, agents, and referencing other features:

Feature	Claude Code	Codex
Subagents	Task agent-name(args)	Use $skill to: args
Commands	/command-name	/prompts:command-name
Agent refs	@agent-name	$skill-name skill

The converter was transforming structure (commands → prompts, agents → skills) but not content references within those files.

Solution

Added transformContentForCodex() function that handles all content mappings:

1. Task agent calls → skill references
2. Slash commands → /prompts: namespace
3. Agent references → $skill syntax

Applied to both skill and prompt files during conversion.

Testing

• All 34 tests pass
• Verified with Codex CLI 0.89.0
• /prompts:workflows-plan now correctly invokes /prompts:deepen-plan (instead of failing with "no such file or directory: /deepen-plan")

Before/After

Task calls:

- Task repo-research-analyst(feature_description)
+ Use the $repo-research-analyst skill to: feature_description

Slash commands:

- Run /deepen-plan to enhance
+ Run /prompts:deepen-plan to enhance

Agent references:

- Have @kieran-rails-reviewer check this
+ Have $kieran-rails-reviewer skill check this

---

🤖 Generated with Claude Code

[terry-li-hm]

Tested and Working ✅

Verified locally with Codex CLI 0.89.0:

1. Ran converter: bun run convert plugins/compound-engineering --to codex --codex-home ~/.codex
2. Confirmed both skill and prompt files have correct syntax:

   - Use the $repo-research-analyst skill to: feature_description
   - Use the $learnings-researcher skill to: feature_description
   

3. Ran /prompts:workflows-plan in Codex
4. Codex correctly:
   • Found and read the referenced skill (Read SKILL.md)
   • Executed the skill instructions manually ("I'll run repository research steps manually")

This is the expected behavior - Codex doesn't have subagent spawning like Claude Code, so it reads the skill and executes instructions in the main thread.

Note: During testing, found and fixed an additional bug where the prompt files weren't being transformed (only skill files were). Both are now correctly transformed.

[terry-li-hm]

Full End-to-End Test Passed ✅

Tested the complete workflow in Codex CLI 0.89.0:

1. Ran /prompts:workflows-plan → Created plan successfully
2. Selected option 2 (/deepen-plan) → Skill loaded successfully
3. No more "no such file or directory: /deepen-plan" error

The /command → /prompts:command transformation is working correctly. Codex can now chain commands as intended.

Ready for review.

[kieranklaassen]

TY!! you are the best

[kieranklaassen]

I'll add CI here as well so we can see it pass on PRs

[terry-li-hm]

These are actually my first open-source contributions — made using Claude Code + compound-engineering itself. Thanks for building this!

[kieranklaassen]

Love this!