Bring the branch onto current main. Import overwrite/skip already
landed there as onDuplicate (fail/skip/overwrite), so drop the 1.9.x
boolean overwrite/skip API, controller, and composer branch pins.
Co-authored-by: Jake Barnby <abnegate@users.noreply.github.com>
The membership privacy policy is documented as controlling what members
can see about *other* members, but it was applied to every membership in
the list, including the requester's own. New projects are created with
`membershipsUserId` disabled, so a standard user listing team
memberships got an empty `userId` on every row - including their own -
and could no longer tell which membership was theirs. That also left
them unable to leave a team, since removing a membership needs its ID.
Treat the requester's own membership like a privileged read in both
`GET /v1/teams/:teamId/memberships` and
`GET /v1/teams/:teamId/memberships/:membershipId`, matching how
`Memberships/Delete.php` and `Memberships/Update.php` already identify
the caller's own membership. None of these fields are secret from the
member anyway - `GET /v1/account` returns all of them.
Fixes#13767
Text defaults to a minimum length of one, so an empty replyToName was
rejected. Use Text(128, 0) as the provider create actions do, letting an
empty string clear the custom reply-to name.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
Assert an empty string clears the reply-to, and check both the address
and display name on the delivered mail.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
Accept an empty string for replyToEmail so a draft can drop a custom
reply-to and fall back to the provider or sender, matching the worker's
empty-value handling.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
Add e2e coverage for reply-to fields: persistence and validation on
create/update, and the delivered reply-to header over SMTP.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
buildEmailMessage now reads replyToEmail/replyToName from the message
data, falling back to the provider options when unset.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
Add optional reply-to address and name params to the create and update
email endpoints, persisted alongside the rest of the message data.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
The explicit family=c3d labels were only there to measure the family change on
this branch, since the orchestrator reads .github/gcrunner.yml from the default
branch on a public repository. The measurement is done and the family moves into
the presets in #13782, so the labels go back to runner=<name>.
n2d-standard-4 is EPYC Rome/Milan where RunsOn was handing out Zen 4, and these
lanes are short and largely single-thread bound, so the core generation shows up
directly: 42% slower than RunsOn on n2d against 21% on c3d, measured on #13779.
Only the family changes, so every existing runner=<name> keeps resolving.
Per-step timing shows the gap is image pulls, not compute: Run tests is 3s on
RunsOn against 5s here, while Pull Docker Image is 12s against 54s and Pull
Docker Compose images 50s against 87s.
The app image already comes from GCRUNNER_REGISTRY_PULL, but jobs spread over
eight regions, so each regional read-through cache is cold for a new SHA and the
image is fetched cross-region once per region. europe-west3 is where the source
registry lives, so pinning there should make every job after the first a local
hit.
The preset run came out 42% slower than RunsOn across 38 of 39 jobs, because
n2d-standard-4 is EPYC Rome/Milan against the Zen 4 machines RunsOn was handing
out for runner=4cpu-linux-x64/family=c7. These lanes are short and largely
single-thread bound, so the core generation shows up directly.
Explicit labels rather than presets, because the orchestrator reads
.github/gcrunner.yml from the default branch on a public repository, so a
preset edit on this branch would not affect this branch is own CI. Measure
first, then move the family into the presets and restore runner=<name>.
Each label names the shape it needs and the shape is defined once, in
.github/gcrunner.yml. No preset names an image, so the default resolves the
image family rather than a build name that the image workflow can prune.
Depends on the presets being on main first: the orchestrator reads that file
from the default branch on a public repository, so these labels resolve only
after #13778 lands.
Nothing references them yet. The orchestrator reads this file from the default
branch on a public repository, never from the pull request head, so a change
that adds the file and switches the labels together references presets that do
not exist on main and every job is held queued. The definitions have to land
first.
The pinned image was pruned by the gcrunner image workflow, which keeps only the
newest three per family, so every job here queues forever behind
`The referenced image resource cannot be found`. The family always names the
newest build and cannot be collected out from under CI.
`dart run jaspr_cli:jaspr build` only resolves jaspr_cli from the project's
own dependency graph, and `jaspr create` does not add it to pubspec.yaml --
Jaspr treats the CLI as part of the environment. A stock Jaspr project
therefore failed to build with:
Could not find package `jaspr_cli` or file `jaspr_cli:jaspr`
The Flutter runtime now preinstalls the CLI and puts it on PATH
(open-runtimes/open-runtimes#572), so the plain binary is the correct call.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
Sites map the `processing` status onto the `building` image, but the
extension was still derived from the raw status, so a processing
deployment asked for `status-building-light.png`, which does not exist.
Only the `.gif` variants are shipped for `building`.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
The proxy create and re-verify endpoints run the multi-region DNS validator,
write the rule as verifying and enqueue generation. The worker then ran the
identical validator again seconds later: doubled DNS traffic when both pass,
and when the second run failed on a transient it wrote unverified over a
status the endpoint had just written. The certificate message now carries
skipDomainValidation, set by every enqueuer that verified DNS itself, and
the worker honours it. The renew check stays in place, unlike a forced job.
Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
The provider whitelist no longer holds every OAuth provider, so github is
refused by the parameter before the action runs and never reaches the
PROJECT_PROVIDER_UNSUPPORTED guard.
The previous test's subject was the fake itself -- it drove
receiveBatch() directly and asserted the broker's return. Review called
that implementation-coupled and it is: a test whose subject is a fixture
proves the fixture, not the worker.
What matters about batching here is the boundary between queues. Batching
changes how many messages one receive claims, not how many run at once,
so the way it could go wrong is by occupying slots another queue is
serialised by, or by losing or duplicating what it claimed. The combined
-mode case now runs its parallel sibling batched and asserts what came
out: every message once, databases still serial, the cap still the cap.
Worth knowing what this cannot see: a batch size that is ignored outright
behaves identically at this boundary, because the whole point of a batch
is fewer round trips and nothing else. Claiming is not observable from
out here.
Review flagged rightPopMany() for ignoring $timeout. It does, as every
pop in this fake does -- nothing here blocks, an empty pop yields, and
that is what keeps a busy receive loop from starving the handlers it
shares a thread with. Making one method block would be a new behaviour,
not a fix.
The real defect underneath was quieter: each message after the first went
back through pop(), so a batch shorter than asked ended on a yield the
real connection never takes. LMPOP waits for the first message only and
takes the company it finds. The rest now comes straight off the list.
A batch the fake cannot serve looks exactly like an idle queue, so the
failure it would cause is silence. The test claims a three-message
backlog in one receive and asserts the order and the drained queue --
seen red against a rightPopMany() that returns nothing.