Skip to content
product

Howzer v2.5.3-DE: the quality release — 7 critical fixes, smaller install, honest degradation

A full codebase deep audit turned into a release: 7 critical silent failures fixed (customer history, follow-up feedback linking, empty reports), reports now fully German and print-ready, installation around 6.3 GB smaller (16 → 9 models), roughly 30,000 lines of dead code removed — with zero breaking changes for v1 consumers.

By Howzer Team, Product

What's in v2.5.3-DE

v2.5.3-DE is the quality release that came out of a complete codebase deep audit ahead of handover: every module was reviewed for silent failures, dead weight, and dishonest fallbacks. The result is a pipeline that is smaller, more truthful about its own state, and harder to misconfigure — without changing anything for existing v1 API consumers.

7
Critical silent failures fixed
−6.3 GB
Installation footprint
0
Breaking changes for v1 consumers

Seven critical silent failures fixed

The audit surfaced defects that never threw an error but quietly produced wrong or empty results. All seven are fixed, among them:

  • Customer history: prior feedback was not reliably attached to the customer record, so journey and history views could miss context.
  • Follow-up feedback linking: follow-ups could fail to link to their original case, breaking escalation and SLA continuity.
  • Empty reports: report generation could silently produce empty sections instead of failing loudly.

Reports: fully German, print-ready

Generated reports are now German end to end — headings, labels, and table content — and formatted for printing, so a scheduled PDF can go straight to a stakeholder without manual cleanup.

A much smaller, cleaner install

  • Model inventory reduced from 16 to 9 HuggingFace models (7 in-house ONNX models plus the gbert tokenizer and E5 for policy retrieval) — the installation shrinks by around 6.3 GB.
  • Source package down from 11 MB to 2.8 MB.
  • Roughly 30,000 lines of dead code removed across the codebase.

Honest degradation instead of hidden fallbacks

Sentiment, emotion, and root-cause analysis no longer fall back to a hidden second model stack when the primary ML model is unavailable. They now report an explicit degraded status with a reason — operators see the true state of the system instead of silently lower-quality results.

Hardened installer security and a stricter CI gate

  • Localhost bind is now persisted by the installer — services do not silently re-expose themselves after a restart.
  • Secret protection and least-privilege database roles are applied by default during installation.
  • A blocking persistence and GDPR test gate in CI: releases cannot ship if data persistence or retention behavior regresses.

Upgrade notes

No database migration is required for this release. A small set of v2 endpoints that had no consumers was removed — they are listed in the changelog. v1 API consumers are unaffected: zero breaking changes.