Ski Output Codex

OpenAI Codex Platform Output

Takes a generated GA entity (.agents/) and produces the Codex equivalent (.codex/), applying all platform-specific transformations.

Input / Output

Input:

• entity_type: workflow | agent | command | skill | rule | knowledge-base | resources | script | hook
• entity_name: the entity filename (e.g. age-spe-email-classifier.md)
• source_content: the full content of the generated GA entity
• export_path: base export path (e.g. exports/{name})
• model_ga: the GA model value from frontmatter (e.g. gemini-3.1)

Output:

• Entity file(s) written to the correct .codex/ path
• Hook JSON fragment if entity is a Hook
• Summary of what was generated

Procedure

1. Determine transformation type

Read the directory mapping from ../resources/res-entity-templates-codex.md §3:

2. Transform content based on entity type

For Workflows, Agents, Commands → TOML conversion

Follow the conversion rules in ../resources/res-entity-templates-codex.md §1:

1. Parse frontmatter: Extract name, description, model from YAML frontmatter
2. Map model: Apply model mapping from §2:
   • gemini-3-flash / haiku → gpt-5.4-mini, effort "low"
   • gemini-3.1 / sonnet → gpt-5.4, effort "medium"
   • opus → gpt-5.4, effort "high"
3. Extract body: Everything after the closing --- of frontmatter becomes developer_instructions
4. Escape for TOML: Wrap in ''' (single-quoted literal). If body contains ''', replace with ```
5. Transform agent references: Replace ./age-spe-*.md and ../agents/age-spe-*.md with descriptive subagent invocation text
6. Parse Skills table: Extract skill names and generate [[skills.config]] entries with .codex/skills/ paths
7. Generate nickname: Convert entity name from kebab-case to Title Case, strip prefix
8. Set defaults: sandbox_mode = true
9. Prepend header: # AUTO-GENERATED from .agents/ — Do not edit directly.

Output file: {export_path}/.codex/agents/{entity-name-without-extension}.toml

For Skills

• Direct copy to .codex/skills/ski-[name]/SKILL.md
• Preserve subdirectory structure

For Rules, Knowledge-base, Resources

• Direct copy to corresponding .codex/ directory

For Scripts

• Same as Claude Code: extract executable logic from GA .md → generate .sh or .py
• Write to .codex/scripts/

For Hooks

Generate TWO files:

1. Documentation: Copy .md to .codex/hooks/hok-[name].md
2. JSON fragment: Generate .codex/hooks/hok-[name].json:

{
  "event": "[from frontmatter event field]",
  "matcher": "[from frontmatter matcher field]",
  "hook": {
    "type": "[from frontmatter type field]",
    "command": ".codex/scripts/scp-[related-script].sh",
    "timeout": 600
  }
}

If the hook event has no Codex equivalent (see ../resources/res-codex-conventions.md §2), generate only the .md and skip the .json fragment. Add a note in the .md:

> **Codex note:** This hook event ([event]) has no direct Codex equivalent. The behavioral rule is documented here for reference but cannot be automated via hooks.json.

3. Write the output file(s)

Write to {export_path}/.codex/{destination_path}.

4. Return summary

Codex output: {entity_name} → .codex/{destination_path}
  Type: [TOML agent | direct copy | executable | hook fragment]
  Transforms: [list, e.g. "MD→TOML", "model mapping", "JSON fragment"]

Error Handling

• Destination file already exists: Overwrite silently (GA is source of truth).
• Unknown entity type: Skip and warn — do not block the pipeline.
• Model value not in mapping table: Use gpt-5.4 with "medium" as default, warn.
• Hook event not supported in Codex: Generate .md only, skip .json, add note.
• TOML escaping failure (body contains '''): Replace with backtick code fences and warn.