Skip to main content

SharpOMatic 10.0.7

SharpOMatic 10.0.7 is now available.

The headline addition is fallback models for Model Call nodes. A node can now define an ordered list containing its primary model and one or more fallbacks. When a provider is unavailable, rate limited, times out, or encounters another eligible transient failure, SharpOMatic can continue the logical call with the next enabled model. Fallbacks can use a different model, connector, or provider, and the editor highlights capability or configuration differences before the workflow is run.

This release also adds OpenTelemetry tracing for workflow runs, node executions, and model provider calls; repeat counts for evaluation rows; compilation caching for C# scripts and expressions; and support for passing a broader range of file URLs to compatible models. The editor gains clearer navigation, a dedicated string and JSON value viewer, live run duration, and expanded fallback-aware metrics.

Changes​

  • Added ordered primary and fallback models to Model Call nodes, with support for multiple fallback entries.
  • Added matching model cards in the editor, including capability warnings and controls to add, remove, disable, and re-enable models.
  • Added compatibility checks for text, image and file input, tool calling, structured output, and active model parameter values.
  • Reused compatible primary Tool Calling and Structured Output settings when executing a fallback while retaining fallback-specific values where settings are incompatible.
  • Added conservative default fallback handling for rate limits, provider HTTP 5xx responses, network failures, and timeouts.
  • Prevented fallback after visible response output or tool execution has started, avoiding duplicate output and repeated tool side effects.
  • Added IEngineNotification.ModelFallbackOverride so hosts can customize whether an eligible failure should use the next model.
  • Added provider-specific failure translation for OpenAI, Azure OpenAI, Anthropic, Azure AI Foundry Anthropic, and Google model callers.
  • Added automatic in-memory upgrade of version 1 Model Call nodes to the version 2 ordered model representation. Upgraded workflows are persisted on their next normal save.
  • Added one model-call metric per provider attempt, with logical call IDs, attempt numbers, normalized failure categories, and provider HTTP status codes.
  • Expanded the Metrics page with logical-call totals, fallback usage, recovery rates, final failure rates, fallback-aware charts and breakdowns, and expandable attempt details.
  • Added OpenTelemetry spans for workflow runs, node executions, model provider calls, and tool execution using standard activity and GenAI conventions.
  • Added workflow, node, model, connector, token, cost, status, conversation, fan-out/fan-in, batch, switch, and Gosub attributes to traces.
  • Recorded original exceptions on failed node spans, including exception type, message, and stack trace.
  • Added telemetry options for disabling engine tracing or explicitly enabling sensitive prompt, response, and tool content.
  • Added a fixed Repeat value to evaluation rows. A row can run from 0 to 10,000 times, with zero skipping that row.
  • Added evaluation row repeat support to the editor, execution engine, CSV import/export, transfers, and repository providers.
  • Cached compiled C# scripts and expressions for Code nodes, Switch branches, context expressions, and evaluation inputs, avoiding repeated Roslyn compilation across executions.
  • Used non-streaming model calls when all incremental AG-UI response output is suppressed, while preserving the configured stored and forwarded events.
  • Added URL input support for PDFs, Office documents, spreadsheets, presentations, CSV, JSON, JSONL, HTML, Markdown, plain text, XML, and YAML in addition to images.
  • Updated Model Call editor labels from image input to file input where the configured model accepts broader content types.
  • Fixed Fan In completion so a retired pre-merge thread cannot supply the workflow's final context.
  • Added back buttons to workflow, connector, model, evaluation, and evaluation-run pages.
  • Added a full-screen viewer for string context values, including automatic formatted JSON display and clipboard copying.
  • Added current run status and duration to the workflow trace panel, with duration updating while the run is active.
  • Added Claude Fable 5 metadata for Anthropic and Azure AI Foundry Anthropic.
  • Added GPT-5.6 Luna, Sol, and Terra metadata for OpenAI and Azure OpenAI.