DavidandClaude Fable 5
6a472f1649
(1) release.yml hatte kein Gate: ci.yml jetzt als workflow_call vor build-tauri (needs: gate). (2) Checkout war target_commitish (Branch) statt Tag -> ref: tag_name, die signierten Bytes entsprechen jetzt dem Tag (AGPL §6). (3) ci.yml rief nur cargo check -> cargo test + clippy (clippy non-gating: scheitert heute schon an secret.rs:108 absurd_extreme_comparisons, Folgefund). (4) enforce-prerelease war No-Op (rel.id === latest.id greift immer) -> prüft jetzt 'von diesem Run gerade veröffentlicht?' statt 'schon Latest?'. (5) discord-announce zählte alte Assets mit -> nur state=uploaded und created_at >= Run-Start. Audit: LUTechnikAudit2.6.6 M6 (release.yml, release-rules.mjs, discord-announce). Verifikation STATISCH (kein Actions-Lauf, nichts öffentlich); release-rules 28 Tests grün, cargo test 539 grün, Full-Suite 6.280. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
93 lines
4.6 KiB
JavaScript
93 lines
4.6 KiB
JavaScript
/**
|
|
* The rules about release state that CI applies, in one testable place.
|
|
*
|
|
* Both callers are top-level-await programs that talk to the GitHub API on
|
|
* import, so the part that can be wrong quietly lives here instead, where a
|
|
* test can reach it without a token.
|
|
*/
|
|
|
|
export const MARKER = '<!-- lu-old-release-banner -->'
|
|
|
|
export const BANNER =
|
|
MARKER +
|
|
'\n> **This is an old release.** Get the current version from the ' +
|
|
'[latest release](https://github.com/PurpleDoubleD/locally-uncensored/releases/latest), ' +
|
|
'or download the Windows installer straight from ' +
|
|
'[lu-labs.ai](https://lu-labs.ai/api/download/windows). ' +
|
|
'Older builds miss fixes and features, and some of them predate the Cloud.\n\n'
|
|
|
|
/** Strips a banner wherever it sits, so re-running never stacks them. */
|
|
export function withoutBanner(body) {
|
|
const at = (body ?? '').indexOf(MARKER)
|
|
if (at < 0) return body ?? ''
|
|
const end = (body ?? '').indexOf('\n\n', at)
|
|
return end < 0 ? '' : (body ?? '').slice(0, at) + (body ?? '').slice(end + 2)
|
|
}
|
|
|
|
const stamp = (rel) => Date.parse(rel?.published_at ?? rel?.created_at ?? '') || 0
|
|
|
|
/**
|
|
* True when this release should carry the banner.
|
|
*
|
|
* `/releases/latest` is the release flagged Latest, which is a deliberate
|
|
* choice (we publish as a prerelease and flip it once verified), not simply the
|
|
* newest. So "not Latest" is NOT the same as "old": a release we published
|
|
* five minutes ago is not Latest yet either, and the old rule stamped it "This
|
|
* is an old release" the moment it appeared. The Discord announcement then
|
|
* quotes that body, so the release everyone was waiting for announced itself as
|
|
* obsolete (review 2026-08-14).
|
|
*
|
|
* A prerelease or draft that is at least as new as the current Latest is the
|
|
* upcoming one and stays clean. Everything older still gets the banner,
|
|
* including a prerelease that was superseded, because that one really is old.
|
|
*/
|
|
export function shouldCarryBanner(rel, latest) {
|
|
if (!latest || rel.id === latest.id) return false
|
|
const upcoming = (rel.prerelease === true || rel.draft === true) && stamp(rel) >= stamp(latest)
|
|
return !upcoming
|
|
}
|
|
|
|
/**
|
|
* True when this release must be forced back to prerelease.
|
|
*
|
|
* release.yml passes `prerelease: true` to tauri-action, and on the workflow's
|
|
* OWN primary trigger that input does nothing. tauri-action only passes
|
|
* draft/prerelease to createRelease (src/create-release.ts): when the release
|
|
* already exists, which it always does on `on: release: [published]`, the
|
|
* action reuses it and never patches the flag. So the guarantee written into
|
|
* that comment ("a build is not a decision") held for workflow_dispatch and
|
|
* for nothing else, and a release published as a full release went straight to
|
|
* Latest with the download routes and every updater following within minutes.
|
|
* That is the 2026-08-10 incident the comment says was fixed.
|
|
*
|
|
* The question this answers is "has a human verified this release?", NOT "is it
|
|
* already Latest". Those look the same and are not: GitHub moves the Latest
|
|
* pointer to any published full release BY ITSELF, immediately, with nobody
|
|
* deciding anything. So the first version of this rule — `rel.id === latest.id
|
|
* -> leave it alone` — was a guaranteed no-op in exactly the case it was
|
|
* written for: a release published as a full release IS already Latest by the
|
|
* time this job runs, the guard fired, and the release nobody had opened kept
|
|
* the Latest flag and the updater traffic that comes with it.
|
|
*
|
|
* Verification happens on a real machine AFTER this build produced installers,
|
|
* so nothing this run publishes can be verified yet: `publishedByThisRun` (the
|
|
* `release: published` event that started this run names this same release)
|
|
* means force it back, whatever the Latest pointer says.
|
|
*
|
|
* The deliberate flip stays safe: on a manual re-run over some older tag no
|
|
* release event named it, `publishedByThisRun` is false, and a release that is
|
|
* the flagged Latest is left alone — that flag is then a human's verdict.
|
|
*
|
|
* @param {object|null|undefined} rel the release this build belongs to
|
|
* @param {object|null|undefined} latest `/releases/latest`, or null if none
|
|
* @param {{publishedByThisRun?: boolean}} [ctx] run context, see above
|
|
*/
|
|
export function shouldForcePrerelease(rel, latest, ctx = {}) {
|
|
if (!rel || rel.prerelease === true || rel.draft === true) return false
|
|
// Published by the event that started this run: not verified, by definition.
|
|
if (ctx.publishedByThisRun === true) return true
|
|
// Otherwise the Latest flag is somebody's deliberate choice — never undo it.
|
|
if (latest && rel.id === latest.id) return false
|
|
return true
|
|
}
|