feat(ssh): AGit PR creation in ssh-only mode - push refs/for/<base>, deterministic topic, closes #N linkage, metadata degradation #327

New issue

Closed

opened 2026-07-21 13:57:46 +00:00 by hexajon · 0 comments

hexajon commented

2026-07-21 13:57:46 +00:00

(Migrated from codeberg.org)

Follow-up from the #298 spike (decision: keep, scoped). AGit becomes the PR-creation path for ssh-only mode (#292's madtea.git-protocol = ssh-only): git push origin HEAD:refs/for/<base> with push options creates or updates the PR entirely over SSH, no token. Scope is bare create/update - merge and all richer PR metadata stay API-side, so in ssh-only they degrade per #295's typed-degradation model, not silently.

Design, from the probe evidence on #298:

finish's ssh-only push leg targets refs/for/<base> with -o topic=<derived deterministically from the local branch name> so every re-push of the same branch updates the SAME PR; -o title=/-o description= are passed on the FIRST push only (the server ignores them on updates - do not pretend otherwise in output).
Non-fast-forward pushes need -o force-push=true; detect the rejection (the server names the option in its error) and either retry with the option or surface a teaching error - decide during build, but never force silently on the first attempt.
The PR number/URL comes from the push's stderr remote: lines (pulls/(\d+)); parse it tolerantly ("Visit the existing pull request" appears on both create and update; * [new reference] vs <old>..<new> distinguishes them). This is text scraping - keep the parser small, pinned by fixtures of real push output, and degrade to "PR created, number unknown - check the forge" rather than failing the finish if the format shifts.
No head branch exists server-side (head is refs/pull/N/head, empty label) and none locally either - so branch-name issue auto-close cannot work on this path. Put closes #N in the PR body from the branch-detected issue instead, honoring the same dependency safety check as the normal path.
Watch out documented in the probe: UNKNOWN push options are silently ignored by the server - a typo'd option gives rc=0 and no feedback. Our own option strings are constants, so this mainly matters for tests (pin the exact options we emit).
After merge there is no topic branch to delete, locally or remotely - the cleanup legs skip with a note, not an error.

Tests (fixtures from the probe transcript): create, ff update same topic, non-ff rejected then force-push retry, nonexistent base rejection, no-new-commits rejection, stderr URL parse on create and update wording, closes-linkage in body.

Out of scope: any change to token-mode finish; merge over SSH (stays API/web); rich metadata over AGit (the -o vocabulary is topic/title/description/force-push only).

Refs #187, #298. Depends on #294 (the ssh push leg) landing first.

Follow-up from the #298 spike (decision: keep, scoped). AGit becomes the PR-creation path for ssh-only mode (#292's `madtea.git-protocol = ssh-only`): `git push origin HEAD:refs/for/<base>` with push options creates or updates the PR entirely over SSH, no token. Scope is bare create/update - merge and all richer PR metadata stay API-side, so in ssh-only they degrade per #295's typed-degradation model, not silently. Design, from the probe evidence on #298: - finish's ssh-only push leg targets `refs/for/<base>` with `-o topic=<derived deterministically from the local branch name>` so every re-push of the same branch updates the SAME PR; `-o title=`/`-o description=` are passed on the FIRST push only (the server ignores them on updates - do not pretend otherwise in output). - Non-fast-forward pushes need `-o force-push=true`; detect the rejection (the server names the option in its error) and either retry with the option or surface a teaching error - decide during build, but never force silently on the first attempt. - The PR number/URL comes from the push's stderr `remote:` lines (`pulls/(\d+)`); parse it tolerantly ("Visit the existing pull request" appears on both create and update; `* [new reference]` vs `<old>..<new>` distinguishes them). This is text scraping - keep the parser small, pinned by fixtures of real push output, and degrade to "PR created, number unknown - check the forge" rather than failing the finish if the format shifts. - No head branch exists server-side (head is `refs/pull/N/head`, empty label) and none locally either - so branch-name issue auto-close cannot work on this path. Put `closes #N` in the PR body from the branch-detected issue instead, honoring the same dependency safety check as the normal path. - Watch out documented in the probe: UNKNOWN push options are silently ignored by the server - a typo'd option gives rc=0 and no feedback. Our own option strings are constants, so this mainly matters for tests (pin the exact options we emit). - After merge there is no topic branch to delete, locally or remotely - the cleanup legs skip with a note, not an error. Tests (fixtures from the probe transcript): create, ff update same topic, non-ff rejected then force-push retry, nonexistent base rejection, no-new-commits rejection, stderr URL parse on create and update wording, closes-linkage in body. Out of scope: any change to token-mode finish; merge over SSH (stays API/web); rich metadata over AGit (the -o vocabulary is topic/title/description/force-push only). Refs #187, #298. Depends on #294 (the ssh push leg) landing first.

No milestone

No project

No assignees

1 participant

Notifications

Due date

The due date is invalid or out of range. Please use the format "yyyy-mm-dd".

No due date set.

Dependencies

No dependencies set.

Reference

sixfold-space/madtea#327

No description provided.

Rows
Columns