[Feature] Add multilingual keyword matching for Codex App skill auto-invocation

[Mineru98]

Problem

Ouroboros skills are currently described with English-only SKILL.md metadata. For example, core packaged skills use descriptions such as:

---
name: interview
description: "Socratic interview to crystallize vague requirements"
---

Codex documents that skills are available in the Codex CLI, IDE extension, and Codex App. It also documents two activation paths: explicit invocation and implicit invocation. The implicit path depends on matching the user's task against the skill description.

That creates a Codex App compatibility gap for Korean and other non-English users: a natural-language Korean prompt such as “요구사항을 먼저 정리해줘” may not reliably auto-select the interview skill because the discoverable trigger text is English-only.

This issue is about multilingual keyword/description matching for automatic skill selection. It is not about translating command names.

Why now

Codex App users often interact through natural-language prompts and visual skill/plugin surfaces rather than terminal-style ooo <skill> commands. If Ouroboros skills only expose English trigger text, non-English Codex App users must know and invoke the canonical English skill name manually.

Related context:

• RFC: Runtime-backed skill execution capability guides #1008: Codex-specific skill compatibility guidance matters.
• [Bug] Codex에서 official auto skill이 요구하는 ouroboros_auto MCP tool이 노출되지 않음 #1045: Codex skill/MCP compatibility problems are visible to users.
• docs: add Korean translation of README #156: Korean translation already exists as an accessibility precedent.
• [Bug] /resume slash command invokes ouroboros:resume skill instead of Claude Code's built-in session picker #490 / Skill name resume collides with Claude Code's built-in /resume command #499: skill naming and invocation behavior can affect host-app UX, so routing changes must stay conservative.

User / persona

• Korean-speaking Codex App users who describe goals in Korean.
• Non-English users who expect installed skills to be selected from localized natural-language intent.
• Maintainers who need better auto-invocation without changing canonical commands.

Current behavior

• Core Ouroboros skill description fields are English-only.
• Codex implicit skill invocation depends on the skill description.
• Ouroboros does not provide localized trigger keywords or localized descriptions for matching.
• Users can still explicitly type canonical commands such as ooo interview, but that assumes they know the English skill name.

Desired behavior

Ouroboros should expose multilingual matching metadata so Codex App can auto-select the correct skill from non-English user prompts.

Examples:

• Korean prompt “요구사항을 먼저 명확히 해줘” should be able to match interview.
• Korean prompt “시드를 생성해줘” should be able to match seed.
• Korean prompt “실행 결과를 평가해줘” should be able to match evaluate or qa depending on intent.

Canonical routing must remain unchanged:

• ooo interview still invokes interview.
• /ouroboros:interview still invokes interview.
• MCP tool names and dispatch identifiers stay English/canonical.

Proposed solution

Add localized matching metadata for packaged skills while keeping canonical skill ids stable.

Implementation options to evaluate:

1. Add localized trigger metadata beside the documented Codex App skill metadata surface, agents/openai.yaml.
2. Generate runtime-specific Codex skill guidance that includes locale-specific trigger phrases.
3. Extend Ouroboros' skill registry/packaging layer to emit localized descriptions or aliases for matching, while keeping command routing canonical.

Example concept:

# agents/openai.yaml or generated Codex-facing metadata
matching:
  localized_triggers:
    ko:
      - 요구사항을 정리해줘
      - 먼저 인터뷰해줘
      - 애매한 요구사항을 명확히 해줘

The exact metadata shape should be verified against current Codex App behavior before implementation. If Codex App does not directly consume custom localized matching fields, Ouroboros should document and generate the closest supported fallback.

Constraints

• Do not translate canonical skill names.
• Do not rename skill directories.
• Do not change MCP tool names.
• Do not add localized slash commands unless Codex officially supports that contract.
• Do not weaken collision-sensitive routing behavior from [Bug] /resume slash command invokes ouroboros:resume skill instead of Claude Code's built-in session picker #490 / Skill name resume collides with Claude Code's built-in /resume command #499.
• Keep multilingual matching as discovery/auto-invocation metadata, not a new command namespace.

Non-goals

• Translating full SKILL.md instruction bodies.
• Translating ooo commands or slash prefixes.
• Changing MCP dispatch contracts.
• Replacing README/documentation translation workflows.
• Supporting every language in the first implementation.

Alternatives considered

• Translate canonical skill names.
  • Rejected: skill names are routing identifiers and collision-sensitive.
• Only add translated display names.
  • Insufficient: this issue targets automatic keyword/intent matching, not only visible labels.
• Only translate README/docs.
  • Insufficient: Codex App skill auto-selection depends on discoverable skill metadata.

Acceptance criteria

• Ouroboros defines a multilingual matching metadata contract for packaged skills.
• Korean trigger phrases exist for core skills such as interview, seed, run, evaluate, qa, and status.
• English remains the canonical and fallback matching language.
• ooo <skill>, /ouroboros:<skill>, and MCP dispatch behavior is unchanged.
• Tests prove localized matching metadata does not rename or shadow canonical skill ids.
• A Codex App compatibility note documents whether localized metadata is consumed directly or generated as fallback guidance.

References

• Codex Skills: https://developers.openai.com/codex/skills
  • Skills are available in Codex CLI, IDE extension, and Codex App.
  • Codex can activate skills explicitly or implicitly.
  • Implicit invocation depends on the skill description.
• Codex plugins: https://developers.openai.com/codex/plugins
• Related issues: RFC: Runtime-backed skill execution capability guides #1008, [Bug] Codex에서 official auto skill이 요구하는 ouroboros_auto MCP tool이 노출되지 않음 #1045, docs: add Korean translation of README #156, [Bug] /resume slash command invokes ouroboros:resume skill instead of Claude Code's built-in session picker #490, Skill name resume collides with Claude Code's built-in /resume command #499

Checklist

• I searched existing issues and discussions first.
• I explained the problem, not just the solution.
• I included clear scope boundaries and non-goals.
• I listed concrete acceptance criteria.

[ouroboros-agent]

This looks like a multi-PR effort that needs design discussion before implementation. Labeling needs-design for human planning.

Scope assessment:

• Affected surfaces include packaged skill metadata under skills/*/SKILL.md, Codex/plugin artifacts such as .claude-plugin/*, skill artifact resolution in src/ouroboros/skills, Codex runtime dispatch tests, and user-facing compatibility docs.
• The requested contract depends on confirming what Codex App can consume from agents/openai.yaml; if direct localization is not supported, Ouroboros needs a generated fallback artifact and documented behavior.
• Implementation should preserve canonical routing for ooo <skill>, /ouroboros:<skill>, and MCP tools while adding display-only locale metadata and tests proving localized labels cannot affect dispatch, especially for collision-sensitive skills like resume-session.

ouroboros-agent[bot]

[shaun0927]

The stated blocker — "confirm what Codex App can consume from agents/openai.yaml" — rests on a file that does not exist in the repo, so the design is actually unblockable now.

Verified against main:

• No agents/openai.yaml (nor any openai.yaml / openai.yml) is tracked anywhere; there is no agents/ directory at the repo root.
• Codex/plugin skill discovery is driven by .claude-plugin/ (plugin.json, marketplace.json, SKILL_CAPABILITY_GUIDE.md) plus per-skill skills/*/SKILL.md description: frontmatter — implicit invocation matches the English description: text.

So there is no openai.yaml localization channel to "consume." Any multilingual keyword-matching metadata has to be a new generated artifact or an extension to the SKILL.md frontmatter / SKILL_CAPABILITY_GUIDE.md, while preserving canonical routing for ooo <skill> / /ouroboros:<skill>. That removes the needs-design external-fact dependency; the remaining decision is purely the metadata shape, which is internal to this repo.

[Q00]

Triage update: removing the stale needs-design blocker.

The earlier blocker depended on confirming what Codex App can consume from an agents/openai.yaml channel, but current main has no tracked agents/openai.yaml, openai.yaml, openai.yml, or root agents/ directory. Skill discovery is driven by packaged .claude-plugin metadata plus skills/*/SKILL.md frontmatter/descriptions.

So this is no longer blocked on an external Codex artifact question. The remaining work is internal implementation design for multilingual display/trigger metadata while preserving canonical routing (ooo <skill>, /ouroboros:<skill>, MCP tools) and proving localized labels cannot affect dispatch identity. Marking ready-for-dev for a scoped implementation proposal/PR.

[ouroboros-agent]

Created draft PR with automated fix: #1346

Pipeline: interview -> seed -> run -> evaluate (PARTIAL)

Verification note: Python syntax compilation and git diff --check passed, but ruff and pytest were not installed in the execution environment, so full local verification could not run before creating the draft PR.

ouroboros-agent[bot]