ai-research-arm: Skill for Codex

Gen Research Ko Backfill

Publish a Korean translation of an existing ARA generative-research article by dispatching the repo's translate-generative-research.yml ("Korean Generative Research Backfill"). That workflow is the publication path: it resolves the canonical source, translates in an isolated adapter, checks source/translation parity, writes through the canonical writer, verifies the exact commit, and lands it as a review PR (or auto-merges on request). main is PR-only, so do not translate and commit locally instead. The dashboard is static and manifest-driven: do not add runtime translation.

First-Principles Check

Before dispatching, answer:

• Which article is the target, and does research/generative/index.json on the default branch contain its slug?
• Is there already a Korean artifact under translations.ko? If yes, did the user ask to replace it?
• Does the user want the review PR (the default) or an auto-merge?

If the target slug does not exist, stop and ask for the correct article.

Workflow

1. Resolve the repo and the target article from an up-to-date default branch.

cd "$(git rev-parse --show-toplevel)"
git fetch origin
git show origin/main:research/generative/index.json \
  | jq -r '.[] | select(.slug=="<slug>") | [.slug, .file, (.translations.ko.file // "")] | @tsv'

1. Dispatch the backfill.

gh workflow run translate-generative-research.yml -f slug="<slug>"

Add -f overwrite=true only when a Korean version exists and the user asked to replace it. Add -f auto_merge=true only when the user asked to skip the review PR.

1. Follow the run to completion.

gh run list --workflow translate-generative-research.yml --limit 3
gh run watch <run-id> --exit-status

A failed run is a failed backfill: report the failing step and its log excerpt (gh run view <run-id> --log-failed), and do not retry by publishing another way. The run summary names the artifact path and, in review mode, the PR (branch prefix automation/ko-backfill).