Skip to content

v0.49.4-beta

Pre-release
Pre-release

Choose a tag to compare

@github-actions github-actions released this 03 Nov 00:44
· 66 commits to main since this release
Immutable release. Only release title and notes can be modified.

Classifarr v0.49.4-beta

Understand destination evidence, review original decisions, and let automatic summaries guide the next fix.

This beta distinguishes declared policy purpose from inferred or unknown support,
adds an explicit assessment in History, and collects bounded prospective evidence
automatically. Existing Evaluation progress gains quiet, expandable summaries;
routine use does not require a CLI command or a new dashboard. Background database
work, capture ownership and logging also gain tighter failure boundaries.

Classifarr has reached 21,997 Docker Hub image pulls as of November 2, 2026
(live repository statistics).
These are image pulls, not people or unique installations, and exclude GHCR.

Before you upgrade

  • Back up the database, persistent appdata and encryption keys. Keep mounts,
    ownership and credentials intact. An older image is not a database downgrade
    plan; restore a matching backup when schema rollback is necessary.
  • Eight application migrations add frozen outcome/policy evidence, review
    references, automatic aggregate checkpoints and private capture claims. Normal
    startup applies them. Stop older Classifarr writers sharing the database;
    the new claim checks do not make mixed-version writers safe.
  • Node 24.22.1, npm/npx 12.2.1 and the digest-pinned Alpine base remain unchanged.
    Strict installer approvals remain enforced. Ordinary Compose and Unraid use
    need no new template setting. This remains a beta release.
  • Existing History and media placements are preserved. New prospective evidence
    accumulates from supported live evaluations and successful explicit reviews;
    missing historical proof remains a gap rather than a guessed label.

Purpose evidence that says what it can prove

Pending and library-comparison cards distinguish declared, inferred, mixed and
unknown purpose. Declared specialized support requires a genuine declared-rule
match bound to the item and candidate. Inferred template/profile rules retain
their ranking roles without being presented as verified declarations.

This does not rewrite saved policies, numeric scores or thresholds. The shared
authority correction addresses one proof boundary; it is not a measured accuracy
improvement across every library, documentary or seasonal classification.

Assess the original decision in History

For an eligible completed automated classification, an administrator can review
the preserved original destination and record an explicit assessment. Current
outcome, original decision and shortlisted choices remain distinct. References
are short-lived and actor-bound; stale state is rejected, and uncertain responses
recover the same receipt instead of submitting another vote.

The assessment uses the existing feedback/statistics transaction. It does not move
media, reclassify the item, create policy rules or add learning authority. Missing
original proof, imported/manual sources and existing receipts remain conservative
eligibility boundaries. History entry controls retain native keyboard access.

Automatic summaries without another operator workflow

Independent provider-free jobs audit retained evidence and compare supported
explicit choices with frozen original and policy references. They save compact
aggregates and refresh through existing ownership, resource admission, cancellation
and backoff. Status polling only reads saved results.

Empty, unavailable, deferred, failed and stale states stay distinct; stale or
invalid counts are withheld even when presentation is paused. Confirmation and
destination-change cohorts remain separate. Agreement is not accuracy, and a
disagreement does not establish cause. Some older or outcome-only paths remain
unassessed. Optional private developer reports are not required for ordinary use.

Frozen policy payloads retain the complete compatible pool only within the
64-candidate / 64-KiB boundary. Overflow and binding failures are explicit gaps.
Thirty-day payload expiry preserves tombstones; configuration exports omit these
diagnostics, and restores reset aggregate/prepared-review state. History deletion
cascades applicable child evidence. No historical backfill is performed.

Storage and background work keep their failure boundaries

Database-proven disk exhaustion now produces fixed incident diagnostics and a
one-hour pause before heavy description/profile refresh work retries. Completed
database cache entries are retained. The pause is process-local and is not proof
that storage recovered; no files are deleted and no service is restarted.

Three-second readiness/workload checks cover pool wait and result acceptance.
Fixed admission slots remain occupied through bounded cleanup and late acquisition
settlement. Stop fences reject late results while preserving genuine database
diagnoses. These cooperative limits do not promise immediate remote SQL/provider
termination or bounded shutdown of an entire worker.

Capture reservations, checkpoints, publication, rotation and finalization require
the acknowledged owner claim. Cleanup stays within admission, valid longer
cooldowns survive, and acknowledged zero-row writes remain distinct from unknown
commit outcomes. Charged attempts are not refunded or replayed after uncertainty.
Real PostgreSQL regressions protect the existing healthy publication boundary;
they do not establish a global fence against older or nonparticipating writers.

Safer logging and reviewed dependency updates

Bounded logging data projection preserves safe context, redaction and log identity
when input is cyclic, hostile or oversized. Database logging failure uses fixed
text with a shared one-minute output limit while preserving normal INSERT
attempts. Native stacks and whole-record limits retain their separate policies.

Runtime parser, redirect-credential and request-counting patches preserve ordinary
HTTP, restore, CSRF, authentication and rate-limit behavior. Axios is updated in
both independent application trees. Build/lint dependencies and SHA-pinned CI
actions are reviewed without weakening checks. The scoped RE2JS patch addresses
its tooling advisory. vue-tsc stays at 3.3.13 after the 3.3.14 candidate failed
five targeted compiler controls; application syntax was not changed to hide them.

Repository maintenance guidance now keeps verified stale development artifacts
bounded while protecting recovery assets, current evidence and other projects.
This is agent-operated development housekeeping, not production cleanup.

Known security and evidence limits

npm 12.2.1 separately bundles vulnerable ip-address 10.7.2, including in the
production image. The application's patched parser and workspace overrides do
not repair that CLI copy. As of the November 2 registry check, the upstream npm
12.2.2 fix is not published. The maintainer explicitly accepts this moderate
residual finding for v0.49.4-beta only, conditional on every other fresh release
gate passing. Verified remediation is required before subsequent releases.
See the release outcome
for the exact decision and fresh evidence before relying on a publication claim.

The paired glob/minimatch tooling patches have dedicated real-package controls.
The server's exact deepmerge advisory has a short-lived OSV triage entry based
on the official CNA rejection, expiring November 9; this is not a patched-package
claim. Full development-audit database lag remains visible. No blanket security
claim, historical ground-truth conversion or guaranteed routing accuracy follows
from this beta. Release publication and live deployment are separate operations.

Full system changes and preserved development history are in the
changelog.


Release evidence

  • Source revision: ef5b50ed1b1d24c9407f669eb11b4c35646bc620
  • Evidence fingerprint: sha256:71590359447617d0a6c8b913947d271894e850720ad677f78d7ae2d23720b10c
  • GHCR image: ghcr.io/cloudbyday90/classifarr@sha256:21d75a2940ffaa49aabc9ec94bad9e44cbb1b8dc7ba8b4a7f1689126146eacc2
  • Docker Hub image: docker.io/cloudbyday90/classifarr@sha256:21d75a2940ffaa49aabc9ec94bad9e44cbb1b8dc7ba8b4a7f1689126146eacc2
  • Consumer smoke image: ghcr.io/cloudbyday90/classifarr@sha256:21d75a2940ffaa49aabc9ec94bad9e44cbb1b8dc7ba8b4a7f1689126146eacc2
  • Consumer smoke completed: 2026-11-03T00:37:52.745Z
  • Provider-fault receipt: sha256:ea1afd875b425d8b5ae800be9939378f16b1bdcfa00d0b529cd026f6d5a7bb35
  • Provider-fault receipt completed: 2026-11-03T00:24:38.094Z
  • Published routing linux/amd64: passed (sha256:0b04fc37352ec4d75aba8f26da75b044c99c104b2574fb67c29715ea6445842a)
  • Published routing linux/arm64: passed (sha256:ce630c865a15fb063c6c74271ab3ac80dd091207e005b50342b92aea6059db23)

The attached release-candidate evidence asset binds this tag to the exact
CI-accepted source revision, passed provider-fault receipt, and digest-only
consumer smoke result, and native AMD64/ARM64 published-image routing checks.