How engine routing picks a model for each task
ScodyX Advanced classifies each task in a run, then routes it to GPT, Claude or Gemini. Here is what the classifier looks at, and what happens when a provider stalls.
A single prompt rarely produces a single kind of work. Planning a data model, rewriting a component tree and summarising a long file are different problems, and the engines differ in how they handle them. ScodyX Advanced treats routing as a per-task decision rather than a per-run one.
Task classes
- Reasoning and planning: decomposing the outcome into a task graph.
- Code synthesis and refactoring: writing or restructuring files.
- Long-context reading: summarising an existing tree before editing it.
- Verification: reviewing a diff against the stated plan.
The classifier reads the task type produced by the planner, not the raw prompt, so the decision is made after the plan exists. Each class carries a preferred engine and an ordered set of substitutes.
Fallback
If a provider returns a transient error or stalls, the run retries with backoff and then substitutes the managed Core engine so the run still completes. The substitution is written into the timeline entry, so you can always see which engine produced a given phase.
plan → reasoning engine
build → synthesis engine (per file group)
verify → reviewing engine (independent of build)
repair → synthesis engine, same contextNote: Routing is a heuristic. It is not a guarantee that a specific engine handles a specific prompt, and provider behaviour can change without notice.