stage.gitlab.md

Merge request: {{inputs.mr}} Repository: {{repository}} This run's checkout: {{cwd}} This run's directory: {{run.dir}}

A note you hand on — to a later step, another agent or a human — goes in {{run.dir}}/handoffs/, never /tmp: the Run's record lists it, and it is kept as long as the Run is.

This Run puts the merge request's head on stage for the human to test, unmerged, and does nothing more. This checkout is the merge request's source branch, at the head Collie read: never commit or push, never post on the merge request or change its state, and never switch this checkout's branch.

A package publishes from its tag pipeline and is deployed by nobody; an application has deploy jobs or GitLab environments (stage, prod) of its own.

assess

Nothing shared is touched yet: no Helle claim is held, so read only.

Decide first whether this repository is a package or an application: read its .gitlab-ci.yml and glab api projects/<id>/environments. Deploy jobs or a staging or production environment make it an application; a publish job on tags and no deploy makes it a package. Say which in your Output, with the evidence.

For an application, also name from its CI configuration, in stage_jobs, the jobs that deploy to stage: those a merge request's pipeline runs (merge_request), and those a tag's pipeline runs (tag). Either list may be empty.

Name in applications every application the merge request's pipeline deploys to stage, by the name its deploy jobs give it: spilnu for stage:spilnu. A repository that deploys several, such as a monorepo, names each one it deploys. Collie claims each one's Helle project before anything goes on stage, and asks the human where one has none.

Then write your Output. is_package is the judgement everything after this branches on, so say it plainly and say what the evidence was.

stage

This Run holds the repository in Helle, or the repository has no Helle project: either way stage is yours until the Run ends. Deploy {{inputs.mr}} to stage from its own pipeline, at its head {{head}}, for the human to test there. Nothing is merged: merge, tag and push nothing here.

• The merge request's pipeline at {{head}} deploys to stage with {{stage_jobs}}. Play, retry and watch those jobs to success — a manual one is played with glab api -X POST projects/<id>/jobs/<job id>/play — and then the jobs the repository itself checks stage with: an e2e-stage or smoke job of the same pipeline.
• Where a stage deploy or a stage check fails, Roll stage back to the newest release production carries, so stage is never left broken, and only in the stage environments those jobs deployed to: the ref of the newest successful production deployment — glab api projects/<id>/deployments?environment=<production>&status=success&order_by=created_at&sort=desc&per_page=1 — and its pipeline's stage deploy job, played or retried. Watch the rollback to success, say in your log what stage is running now, and write your Output: Collie reads stage back and ends the Run with what failed.
• Never play, retry or ask about a production job, nor a job that deploys an application not named above.

Then write your Output, naming in stage_jobs the stage deploy and stage check jobs of the merge request's pipeline you played or checked. Collie reads each of them, and stage, back from GitLab, not from your Output.

unstage

Stage carries {{inputs.mr}} at {{head}}, unmerged, and it comes off: {{failure}}.

Roll stage back to the newest release production carries, unless it is already there, and only in the stage environments this Run deployed {{stage_jobs}} to: the ref of the newest successful production deployment — glab api projects/<id>/deployments?environment=<production>&status=success&order_by=created_at&sort=desc&per_page=1 — and its pipeline's stage deploy job, played or retried. Watch it to success. Never play a production job, and merge, tag or push nothing.

Then write your Output, naming in stage the release stage runs now. Collie reads stage back from GitLab, not from your Output.