Overview - Ghost

7 days30 days

Latest CI Pipeline Executions

Fuzzy

Succeeded
30451
e7e04da9 Changed the field type catalog to separate structure from validation ref https://linear.app/ghost/issue/BER-3794/members-manage-their-own-custom-fields-in-their-account The catalog answers two questions that are usually asked by different people. What parts a field type has — an address has six, and the last of them holds a country code — is what anything drawing a form needs. What counts as a valid value is what the server needs when someone submits one. Both were declared together, built out of zod, so asking the first question meant importing the second. Measured through Portal's own build, with the import actually used: taking the whole catalog costs seventeen and a half kilobytes gzipped and puts zod in the bundle. Taking the structure alone costs a hundred and seventy-seven bytes and puts none there. Ghost's server enforces the rules and can afford them; Portal only draws inputs, and it loads on every page view of every themed site. Asking it to carry a validator it will never run, to find out that an address has six parts, is the wrong trade. So the structural half now stands alone and imports nothing. A renderer that wants it gets the types with it, rather than hardcoding part names as strings and hoping they match. Anything needing both halves still imports one place; the split is about what a bundle pays for, not about where the catalog lives. The two cannot drift, because the rules are now built from the structure rather than declared beside it: an address's parts are stated once, and the schema for each is looked up from that statement. A part added without a rule, or a rule without a part, fails to compile. Splitting them also emptied something. A field type used to carry a map of its parts pairing each name with both its declared type and its rule, because the two were needed in different places and there was nowhere else to state the type. The type is stated in the structure now, and nothing reads that map any more, so it goes and takes the last of the coupling with it: a field type is a kind and a rule, and what it is made of is described elsewhere.
by Rob Lester
R
Succeeded
main
d41f8a27 Fixed the member custom field test helper missing an access setting A member field's access setting became a required part of the field shape, but the helper that builds fields for these tests still omitted it, and the partial overrides it accepts cannot supply it, so the helper's return value no longer satisfied the type. It defaults to closed, which is what the server reads a field with no access set as.
by Rob Lester
R
Succeeded
main
d41f8a27 Fixed the member custom field test helper missing an access setting A member field's access setting became a required part of the field shape, but the helper that builds fields for these tests still omitted it, and the partial overrides it accepts cannot supply it, so the helper's return value no longer satisfied the type. It defaults to closed, which is what the server reads a field with no access set as.
by Rob Lester
R
Failed
30637
ba22c851 Converted limit-service to TypeScript The package now builds, lints and tests the way every other internal package does, and ships types of its own rather than leaving Admin to hand-write a declaration for it. Nothing about what it does has changed. The count queries still live here and still take the database they are given, the checks still compare the way they did, and the tests are the ones that came with the package rather than replacements for them. Where the compiler found behaviour that looks like a mistake, such as a limit whose name is spelled another way silently arriving without its settings, it is recorded and left alone. Deciding what this package should do differently now that it lives here is worth doing with fresh eyes, and doing it inside a conversion would hide it.
by Rob Lester
R
Succeeded
30631
14850e57 Fixed the member custom field test helper missing an access setting A member field's access setting became a required part of the field shape, but the helper that builds fields for these tests still omitted it, and the partial overrides it accepts cannot supply it, so the helper's return value no longer satisfied the type. It defaults to closed, which is what the server reads a field with no access set as.
by Rob Lester
R
Failed
30631
403d897b Fixed the CI typecheck job never running pnpm 12 rejects the -s flag, so the command that lists projects with a test:types target produced no output at all. Piping it into tr masked the non-zero exit, leaving the projects list empty, which made the typecheck job's if condition false and silently skipped it on every commit.
by Rob Lester
R
Succeeded
30451
ae9b4a69 Changed the field type catalog to separate structure from validation ref https://linear.app/ghost/issue/BER-3794/members-manage-their-own-custom-fields-in-their-account The catalog answers two questions that are usually asked by different people. What parts a field type has — an address has six, and the last of them holds a country code — is what anything drawing a form needs. What counts as a valid value is what the server needs when someone submits one. Both were declared together, built out of zod, so asking the first question meant importing the second. Measured through Portal's own build, with the import actually used: taking the whole catalog costs seventeen and a half kilobytes gzipped and puts zod in the bundle. Taking the structure alone costs a hundred and seventy-seven bytes and puts none there. Ghost's server enforces the rules and can afford them; Portal only draws inputs, and it loads on every page view of every themed site. Asking it to carry a validator it will never run, to find out that an address has six parts, is the wrong trade. So the structural half now stands alone and imports nothing. A renderer that wants it gets the types with it, rather than hardcoding part names as strings and hoping they match. Anything needing both halves still imports one place; the split is about what a bundle pays for, not about where the catalog lives. The two cannot drift, because the rules are now built from the structure rather than declared beside it: an address's parts are stated once, and the schema for each is looked up from that statement. A part added without a rule, or a rule without a part, fails to compile. Splitting them also emptied something. A field type used to carry a map of its parts pairing each name with both its declared type and its rule, because the two were needed in different places and there was nowhere else to state the type. The type is stated in the structure now, and nothing reads that map any more, so it goes and takes the last of the coupling with it: a field type is a kind and a rule, and what it is made of is described elsewhere.
by Rob Lester
R
Succeeded
30511
b3651369 Fixed Admin failing to count emails when the cap is not periodic The same fault the server had, and it predates the move: a host capping emails outright rather than per period leaves the counter with no period, and formatting the missing date throws. The publish flow reports whatever it catches, so a publisher was told their sending was disabled because of an invalid time value. Admin now counts the whole history when there is no period to count within, which is what an outright cap covers.
by Rob Lester
R
Failed
30511
1f54ae5e Added where limit-service came from to its README The history came across with the package, but nothing in the repository says so and `git log --follow` cannot reach it, because it does not traverse the merge that brought it in. Someone following a file sees it appear from nowhere at the move into src. The commit that history hangs off is recorded here with the command that reads it, so the provenance the import preserved is something a reader can actually find.
by Rob Lester
R
Succeeded
main
ad9e9c94 Changed the browser tests to one timeout everywhere ref https://linear.app/ghost/issue/BER-3864/differentiate-public-and-private-member-projections A test had half as long to finish on a laptop as it had on CI, so a correct test could fail in front of the person writing it and pass once pushed. That is the worst way for a suite to be wrong: it costs an investigation every time, and it teaches everyone to disbelieve a local failure, which is the one signal that arrives early enough to be cheap. Half a minute is also not the difference between a fast iteration loop and a slow one, so the shorter budget was buying nothing to offset that. It was not a decision anybody made. The two were deliberately brought together at thirty seconds once; a later commit raised CI to sixty because CI needed longer, and splitting them again was a side effect of that rather than its intent. The identical branches left behind on the assertion timeout are the fossil of the same tidy-up, and go here too.
by Rob Lester
R
Succeeded
30535
120ea99d Changed the browser tests to one timeout everywhere ref https://linear.app/ghost/issue/BER-3864/differentiate-public-and-private-member-projections A test had half as long to finish on a laptop as it had on CI, so a correct test could fail in front of the person writing it and pass once pushed. That is the worst way for a suite to be wrong: it costs an investigation every time, and it teaches everyone to disbelieve a local failure, which is the one signal that arrives early enough to be cheap. Half a minute is also not the difference between a fast iteration loop and a slow one, so the shorter budget was buying nothing to offset that. It was not a decision anybody made. The two were deliberately brought together at thirty seconds once; a later commit raised CI to sixty because CI needed longer, and splitting them again was a side effect of that rather than its intent. The identical branches left behind on the assertion timeout are the fossil of the same tidy-up, and go here too.
by Rob Lester
R
Succeeded
30512
ee6b31b3 Added a host limit that can switch custom member fields off per site Custom member fields are intended for one Ghost(Pro) plan and above, but nothing could express that. A labs flag says a feature does not exist yet, which is the wrong thing to tell someone whose plan simply does not include it, and it is the same answer for every site. This adds limitCustomFields alongside the other flag limits the host already sets, so the decision about which plan carries the feature is configuration on the hosting side rather than a change to Ghost. It stays open unless a host switches it off, so nothing changes for anyone until that configuration exists. The routes that exist only to change field definitions sit behind one guard, and Admin asks the same question before offering any of it, so a publisher whose plan excludes the feature is told so rather than shown a door that refuses them.
by Rob Lester
R
Failed
30511
814cb8e8 Added where limit-service came from to its README The history came across with the package, but nothing in the repository says so and `git log --follow` cannot reach it, because it does not traverse the merge that brought it in. Someone following a file sees it appear from nowhere at the move into src. The commit that history hangs off is recorded here with the command that reads it, so the provenance the import preserved is something a reader can actually find.
by Rob Lester
R
Succeeded
30535
4bad7481 Changed the browser tests to one timeout everywhere ref https://linear.app/ghost/issue/BER-3864/differentiate-public-and-private-member-projections A test had half as long to finish on a laptop as it had on CI, so a correct test could fail in front of the person writing it and pass once pushed. That is the worst way for a suite to be wrong: it costs an investigation every time, and it teaches everyone to disbelieve a local failure, which is the one signal that arrives early enough to be cheap. Half a minute is also not the difference between a fast iteration loop and a slow one, so the shorter budget was buying nothing to offset that. It was not a decision anybody made. The two were deliberately brought together at thirty seconds once; a later commit raised CI to sixty because CI needed longer, and splitting them again was a side effect of that rather than its intent. The identical branches left behind on the assertion timeout are the fossil of the same tidy-up, and go here too.
by Rob Lester
R
Failed
30535
983bf48a Changed the browser tests to one timeout everywhere ref https://linear.app/ghost/issue/BER-3864/differentiate-public-and-private-member-projections A test had half as long to finish on a laptop as it had on CI, so a correct test could fail in front of the person writing it and pass once pushed. That is the worst way for a suite to be wrong: it costs an investigation every time, and it teaches everyone to disbelieve a local failure, which is the one signal that arrives early enough to be cheap. Half a minute is also not the difference between a fast iteration loop and a slow one, so the shorter budget was buying nothing to offset that. It was not a decision anybody made. The two were deliberately brought together at thirty seconds once; a later commit raised CI to sixty because CI needed longer, and splitting them again was a side effect of that rather than its intent. The identical branches left behind on the assertion timeout are the fossil of the same tidy-up, and go here too.
by Rob Lester
R
Succeeded
30535
c8ccda60 Changed the browser tests to one timeout everywhere ref https://linear.app/ghost/issue/BER-3864/differentiate-public-and-private-member-projections A test had half as long to finish on a laptop as it had on CI, so a correct test could fail in front of the person writing it and pass once pushed. That is the worst way for a suite to be wrong: it costs an investigation every time, and it teaches everyone to disbelieve a local failure, which is the one signal that arrives early enough to be cheap. Half a minute is also not the difference between a fast iteration loop and a slow one, so the shorter budget was buying nothing to offset that. It was not a decision anybody made. The two were deliberately brought together at thirty seconds once; a later commit raised CI to sixty because CI needed longer, and splitting them again was a side effect of that rather than its intent. The identical branches left behind on the assertion timeout are the fossil of the same tidy-up, and go here too.
by Rob Lester
R
Succeeded
30535
5501379d Changed the browser tests to one timeout everywhere ref https://linear.app/ghost/issue/BER-3864/differentiate-public-and-private-member-projections A test had half as long to finish on a laptop as it had on CI, so a correct test could fail in front of the person writing it and pass once pushed. That is the worst way for a suite to be wrong: it costs an investigation every time, and it teaches everyone to disbelieve a local failure, which is the one signal that arrives early enough to be cheap. Half a minute is also not the difference between a fast iteration loop and a slow one, so the shorter budget was buying nothing to offset that. It was not a decision anybody made. The two were deliberately brought together at thirty seconds once; a later commit raised CI to sixty because CI needed longer, and splitting them again was a side effect of that rather than its intent. The identical branches left behind on the assertion timeout are the fossil of the same tidy-up, and go here too.
by Rob Lester
R
Failed
30512
a4347e66 Added a host limit that can switch custom member fields off per site Custom member fields need to be sellable as part of a plan, which the labs flag cannot express: the flag says whether the feature exists in this build, not whether a given site is entitled to it. A host can now set limitCustomFields and the site stops being able to define fields. Nothing sets it anywhere, so no site behaves differently today; choosing which plan includes custom fields becomes a change to host configuration and nothing else, which is the point. Because limits now live in this repository, that took a name and no release. The refusal covers only defining a field: the routes that create, reorder, edit and delete one. Reading stays open, so a site that drops onto a plan without the feature keeps seeing, filtering and exporting what it already collected. That works because a definition is the only root of the feature, with values and checkout bindings both keyed off one, so a site that cannot define fields has nothing downstream to reach. Those routes are also now a router mounted at the path they share, where each guard is stated once and order carries the rule: a request walks the stack from the top, so the reads are answered before the guards are reached and anything added after them passes through both. One incidental change comes with that: a method the router has no route for is refused by the guards rather than falling through unmatched. Refusing answers 403 rather than the 404 the flag gives, because "your plan does not include this" is a different thing to tell a caller than "this does not exist here", and only the first is something a publisher can act on. It carries whatever wording the host supplies, or the limit service's own, which names the feature. On the Admin side one hook folds the flag and the limit together so no screen can check one and forget the other, and removing the flag at GA touches only that hook. What a publisher on the wrong plan should be shown instead is a design question and is not answered here. ref https://linear.app/ghost/issue/BER-3797
by Rob Lester
Succeeded
30511
6a208b5a Fixed asking whether a site is over any limit on a site with an allowlist The question names no limit, so every limit a site has must be able to answer it. An allowlist limit cannot, because it judges whether one particular value is permitted and the question carries no value, so it raised an incorrect-usage error that escaped the whole check. A site was left unable to answer whether it was over any of its limits purely because it also had a limit on which themes it could use, which every hosted site does.
by Rob Lester
Succeeded
30512
8a29cf40 Added a host limit that can switch custom member fields off per site Custom member fields need to be sellable as part of a plan, which the labs flag cannot express: the flag says whether the feature exists in this build, not whether a given site is entitled to it. A host can now set limitCustomFields and the site stops being able to define fields. Nothing sets it anywhere, so no site behaves differently today; choosing which plan includes custom fields becomes a change to host configuration and nothing else, which is the point. Because limits now live in this repository, that took a name and no release. The refusal covers only defining a field: the routes that create, reorder, edit and delete one. Reading stays open, so a site that drops onto a plan without the feature keeps seeing, filtering and exporting what it already collected. That works because a definition is the only root of the feature, with values and checkout bindings both keyed off one, so a site that cannot define fields has nothing downstream to reach. Those routes are also now a router mounted at the path they share, where each guard is stated once and order carries the rule: a request walks the stack from the top, so the reads are answered before the guards are reached and anything added after them passes through both. One incidental change comes with that: a method the router has no route for is refused by the guards rather than falling through unmatched. Refusing answers 403 rather than the 404 the flag gives, because "your plan does not include this" is a different thing to tell a caller than "this does not exist here", and only the first is something a publisher can act on. It carries whatever wording the host supplies, or the limit service's own, which names the feature. On the Admin side one hook folds the flag and the limit together so no screen can check one and forget the other, and removing the flag at GA touches only that hook. What a publisher on the wrong plan should be shown instead is a design question and is not answered here. ref https://linear.app/ghost/issue/BER-3797
by Rob Lester