Implement AI-from-human decoupled workflow using a parallel branch approach (pilot)

Problem

Original thread and idea by @mjsibanez : #764 (comment 2920269825)

To re-iterate, many translation requests have been open for a long time, which means the source English .mmd files those requests are based on have been updated in English since. We cannot create new translation requests while files are locked in active requests, and need a solution, hence the 'decoupling' Otherwise, the Japanese docs site parity with English is going to be (and is now) behind because we are waiting for human review to complete before we can send updated content [I am saying the same thing, but just being verbose just to over-communicate]

The approach we want to test

On 2025-12-02 (agenda here), @yyu-ext proposed treating main-translation like a development environment and keep AI translations flowing continuously into main-translation, then branch off periodically for human linguistic review.

[this is an assumption so need Spartan to confirm -->] We don't need to redesign the Argo request workflow. We will use the same Argo request, thus not. losing the mental and physical connection between AI and human translations. We just change where the AI and HR Translation MRs get created. AI translation MRs would target main-translation and get merged immediately, and Human review MRs would target a separate branch like docs-human-2024-12-09.

[Need to figure the merging strategy with this branching approach. And probably ask more questions. Seeing it as a live pilot would help]

Things to fugue out

Need Spartan's help to figure out what will need to change in the Argo-GitLab integration , so it "knows" to target different branches.

Then there is a question of how we manage these HR branches - create new ones, reuse existing ones, how many there will be, and how are we going to we decide when a subset of humanly reviewed translations is ready to merge upstream, how exactly it's going to be done (who creates MRs, what is the MR template, review workflow, coordination, etc.), how often.

Need to figure out file selection and prioritization (per @emikimura-ext 's notes here)

[is this. even a valid question / scenario? Seeing it as a live pilot would help-->] What happens if we need to merge an HR branch but there\ is already a newer AI translation for the same file in main-translation?

First iteration of a pilot

  • Cancel one of the stuck Phase 2+3 requests, select 10-15 files from it for the pilot, work with @jcole-ext to configure Argo-GitLab integration , create the first HR branch from main-translation and run those files through our first simulated decoupled workflow. Document all questions, what works, what breaks, what surprises occurred. Iterate on!
  • ...
  • ...

*****

cc @mjsibanez @irogova-ext @laurenbarker @rasamhossain @mpehkonen @jcole-ext @yyu-ext @wojciech.froelich @marcel.amirault @mchrastek-ext @emikimura-ext for your feedback

Edited by Oleksandr Pysaryuk