[0.53.0] - 2026-08-25
Bug Fixes
- Ship SLF4J backend + close CRLF/credential-leak gaps in kuml-cli/desktop/mcp (logging)- AI panel token-usage counter permanently stuck at 0/0/$0.0000 (desktop)- Bump project version to 0.53.0 (build)
Chore
- Consolidate kuml-logo-simple.png into kuml-logo.png (docs)- V0.53.0 (release)
Features
- AI provider settings dialog + fix Main-dispatcher crash (kuml-desktop)- Real tool-calling in desktop AI panel + fix context-loss/race bugs (ai)- App-Icon, Ansichtsmodus, echte Ollama-Modellliste + Tool-Call-Crash-Fix (desktop)- V3.7.4 — macOS Keychain credential misrouting fix + desktop UX polish (desktop)
Other
- Add MiniMax-M2.7 GCR benchmark results via GonkaRouter (Gonka network)
New run-gonka.py / run-gonka-examples.py scripts for benchmarking models
served over GonkaRouter, an independent third-party broker for the
decentralized Gonka compute network (distinct from kUML's built-in Gonka
provider, which is hardcoded to the official broker).
MiniMax-M2.7's baseline run had a 73.3% "stuck" rate: the model's
reasoning trace routinely exhausted the 32,768-token budget before
producing any code, making every FCR/SF figure for this model the lowest
recorded in this benchmark - conditioned on the minority of cells that
finished in time, not a representative sample. kuml.examples MCP tool
access more than doubled completion and Structural Fidelity, at the cost
of a near-tripled Hallucination Rate.
Also fixes two pre-existing handbook inaccuracies caught during doc
review: a deepseek-r1:70b FCR figure (70.0% -> 62.0%, mixed up with its
PlantUML number) and a stale "six configurations" count in the MCP
takeaways section that predated the Grok 4/MiniMax-M2.7 additions.
Co-Authored-By: Claude Sonnet 5 noreply@anthropic.com