Repository navigation
[0.4.5] — 2026-11-21
Added
- A job now records a clause-by-clause audit of the instruction contract it ran under: the merge gate evaluates the contract against the pull request it holds, stores
pass,violationorunverifiableper clause, and a violation blocks the merge request itself. The job page lists every clause with its verdict, labelled so an unchecked clause is never read as a pass. - Operators can set
GATEWAY_TENANT_METERING=onto publishfledge_gateway_requests_total{tenant}for each authenticated model-gateway request — once per user and once per repository for a repository token — and the gateway's two 429 answers gainretryAt, the epoch-millisecond time that address's bucket refills. The switch is off by default; with it off the limiter key, the window, the 429 body and the requests admitted are exactly what they were. - The admin metrics page draws the four series of its daily-trend table — jobs, succeeded, failed and merged — as compact inline sparklines above the table, each with its total as text and an
aria-labelfor screen readers. The graphic is server-rendered SVG with no chart library and no external asset, it keeps the page free of horizontal scroll at 360, and a window with no daily data keeps the existing empty state rather than drawing a flat zero line. - Backup retention is now a policy the repository owns:
BACKUP_KEEP_PLAINTEXT(default 1) andBACKUP_KEEP_ENCRYPTED(default 30) are read by the server and by the newnode scripts/backup-prune.mjsdry run, which lists the copies beyond each class's count together with the manifest and-wal/-shmsidecars that belong to them, and deletes nothing. The/adminbackup card shows the same count as "Expired copies: N (dry run)", and reports a directory it cannot read asscan failed: <errno>beside the configured counts instead of failing the page. No--applymode ships here, so the first real deletion stays the owner's. - Backup and restore-drill health now has a machine-readable surface:
/metricspublishesfledge_backup_age_hours,fledge_backup_copies,fledge_backup_encrypted_copies,fledge_restore_drill_age_hoursandfledge_restore_drill_ok, and two new Prometheus rules page when the newest copy is older than the expected interval or the last drill failed or aged out. A stale copy or a failed drill also raises one in-app ops alert, deduplicated by the existingOPS_ALERT_COOLDOWN_MINUTES, so a night that stopped backing up no longer waits for somebody to open the admin page.
Changed
- A job's instruction contract now records the hash of the playbook text the team is told rather than the goal's fingerprint, so two
v8playbooks that differ in the protocol, a role text or the team label are different records, and the repository page prints the first twelve characters of that hash under the goal form. The text is the one the instance applies — built from the repository goal, never from a job's own prompt, which for review, discover, triage, scope and verify and for any carry-over follow-up differs in the section the text calls authoritative. The sealed context bundle carries the same value; the rules commit hash stays on the job and its event. - Error pages now say what the reply was and offer a way forward that fits the reader. The page prints the status the route already replied with (
404,403,500, in the tab title too), and the action is chosen from the viewer: a signed-in person returns to the console, an anonymous visitor is sent to GitHub sign-in, and a route that has its own way out (the OAuth 403, a 5xx report form) keeps it. Server errors still never show a stack trace. - The job page now prints the job cost on its header summary line, in the same amount form the result card uses, so the number is visible without scrolling to the bottom. The Timeline card is folded behind a disclosure that names how many events it holds; every event row still renders when it is opened, and
#timelineand#review-mergestill resolve. - The dashboard no longer answers "nothing needs you" for a repository that cannot run. A disabled repository, one with automatic work off, a paused schedule and a workspace that is not ready each name the lock and the action that lifts it, and an account whose every repository is in that state leads with the lock instead of the all-clear.
- The Admin page's one-click release now shows the daily release planner's merged-pull-request and feature counts and its verdict instead of counting commit messages itself, so it no longer refuses a release the nightly run would make and the numbers a maintainer reads are the rule's.
Fixed
- The schedule console now says, in both languages, that a repository whose role was never chosen has its scheduled runs paused: each due instant is skipped until a role is chosen, while Run now keeps working. The repository page, the role card and the dashboard already said it; the schedule console was the one surface where the owner saw a working Run now and nothing about the scheduled half.
- The console's disclosure markers are drawn from border geometry instead of a typed
▸/▾character in the stylesheet, so a screen reader no longer hears a triangle inside the summary's name and the mark keeps one weight and baseline on every platform. - Error pages a thrown refusal produces now speak the console's language: the 22 messages the app throws (missing installation, job, work item, release or notification, and the permission refusals) have English forms, so an English console no longer shows a Chinese sentence under an English heading.
- The platform-quota tile at the top of the workspace page linked
/settings#model, but the settings page names that sectionid="model-providers"and defines noid="model"anywhere — so the click landed at the top of a long page with nothing marked, while three other entry points to the same section already used the correct fragment. The tile now points at the section it names. A new test renders the workspace, the dashboard and a repository page, follows every in-app#fragmentlink on them and fails when the destination defines no such id, so the next stale fragment is caught before it ships instead of after a reader reports a click that went nowhere. - The
deploy-checksample contract no longer reports a rejection that named the header or the location as one that did not: its eight guards now test the captured text directly instead of piping it intogrep -q, which exited at the first match and killed the producer with SIGPIPE, sopipefailturned a pass into a failure at large outputs. The checks and their messages are unchanged; only the way each one reads its own output is. - The job page's context-bundle card refused a bundle whose stored bytes no longer matched the recorded hash, and then printed, a few lines below the refusal, "The callback echoed the same hash." plus an unexplained field-by-field diff — because the echo sentence read only the callback's own claim about the run and the diff was built from the same bytes that had just failed verification. On the one surface whose job is to be unambiguous, a reader who skimmed saw a refusal followed by what sounded like confirmation. The echo sentence now names whose claim it is and says the stored bytes do not verify, the diff heading says it is read from bytes that did not verify, and a row whose stored bytes cannot be read at all says there is nothing to compare instead of claiming a first or identical bundle. When the bytes and the hash agree, the card keeps its original wording. Both languages carry the new sentences.
- Who may press merge is now the role the person holds on the repository, not the connection's declared role. An invited developer could press "Merge pull request" — or merge up to ten clean pull requests in one press — on a
maintainerconnection, and the owner of adeveloperor never-chosen connection was refused a merge their upstream accepts. Both human entries now take the role the repository guard already resolved and pass it into the merge gate, so a developer is refused before any GitHub call with the sentence the card already carried, and an owner can merge from either connection. Automatic merging keeps reading the connection's role because the scheduler has no actor. The bulk control is drawn from the viewer's own permission and says so when it is withheld. - The job page no longer prints a measured
¥0.00for a job it never priced. A queued or running job reads—; a finished job whose call the price table could not price reads— 未知/— unknown, which is how the store already counts an unknown cost; a finished job with a real zero keeps¥0.00. One helper now feeds both the header line and the details line. - The instruction-contract audit on a job page no longer reports
passfor a requirement the contract never asked for. A clause whose evidence, sections or forbidden-action list is empty now reads— 无须核对 / — not askedinstead of通过 / pass, and the card's summary names the limit ("this contract required no evidence, named no sections and forbade no action") rather than claiming that every checkable clause passed. A contract that does name those lists is unchanged, and no merge outcome moves. - The nightly release opener no longer adopts an open pull request it never re-read. When a run finds the mechanical release pull request already open at the branch and commit from the plan, it now re-reads that pull request's head — the same
maintainer_can_modifycheck it makes on a pull request it has just created — and refuses instead of answeringreusedwhen the head belongs to another repository, when the flag readstrue, or when the readback cannot be made. The refusal names the pull request and the reason, so a red night says why on #80, and it never opens a second pull request on the same branch; a head that still checks out is reused as before. - A release target older than every tag in the repository is now refused as a target that predates the releases, instead of being planned as a first release. A repository that had already shipped
v1.20.0would acceptv0.5.0on an older commit and tag an old tree with a new name, because the first-release path skips both the "nothing to release" and "not higher" checks; the readiness page and the release form now print that refusal before anything is written. A repository with no tags at all still gets its genuine first release.
Security
- Stored encrypted secrets can now name the key that wrote them: a
v2payload carries a key id, and the operator can supply named keys throughTOKEN_ENCRYPTION_KEYS_FILEwhileTOKEN_ENCRYPTION_KEYkeeps decrypting the existingv1values. A payload that names an id the process does not know fails closed instead of retrying with the default key, and a malformed or unreadable key file refuses to boot with the file and the offending id in the message. Nothing re-encrypts or rotates an existing value; the mechanism ships without a rotation action.
Image: ghcr.io/hugoxk/fledge:v0.4.5@sha256:cbb9ef78c2e4a6fca56b7634c3575a81677de63e4add30baf3e37b26145e148f
SBOM: https://github.com/HugoXK/fledge/releases/download/v0.4.5/sbom.spdx.json
Attestation: sha256:cbb9ef78c2e4a6fca56b7634c3575a81677de63e4add30baf3e37b26145e148f