Koios v1.0.2
Patch release fixing model-history gaps for in-memory inputs, hardening model upload validation, and tightening up a handful of UI rough edges.
Fixed
Model-history gaps for in-memory inputs
AI models with inputs sourced from in-memory tags (e.g. component outputs, expression results) were not writing entries to the per-model history bucket. Long-term history was unaffected — the gap only existed in the model-bucket history that powers per-model prediction views.
The expression evaluator now writes per-model ai_history entries for in-memory tag inputs the same way it always did for device-sourced inputs. A periodic safety net also catches any AI-binding-changed events that the pub/sub channel missed (e.g. during evaluator restart), so configuration changes always settle within one refresh cycle.
Predict engine bucket creation no longer fails after a restore
After restoring a backup, the predict engine could fail to start models whose InfluxDB buckets already existed, logging bucket with name model_<id>_predictions already exists and refusing to run. Bucket creation is now idempotent — if the bucket is already present, the predict engine adopts it instead of erroring.
Model upload rejects incompatible ONNX files at upload time
ONNX files whose IR version exceeds what the platform's ONNX runtime can load (e.g. files produced by onnx>=1.21 defaulting to IR 13 when the runtime caps at IR 11) are now rejected at the upload step with a clear message. Previously these files uploaded successfully but failed later in the predict engine with Failed to load interpreter.
UI polish
- Component library tray — long component names no longer cause horizontal overflow in the canvas side drawer.
- Binding rows — bindings sourced from in-memory tags now correctly display "In Memory" instead of being mislabeled as "Expression".
- Bulk export — exporting all tags from the tag table no longer fails when no filters are applied.
Changed
Model upload now uses the official koios-model-utils parser
The webapp's ONNX metadata extraction now consumes the canonical koios-model-utils library (the same library data scientists use to embed metadata into their models) instead of a hand-rolled parser. This eliminates a class of subtle wire-format drift bugs and lets the library's strict validators catch malformed metadata at upload instead of letting it produce blank rows or silent misconfiguration later.
No user-facing behavior change for correctly-formed models.