Stream processing flow

This page describes the current end-to-end flow of streaming model output through the implemented Agentsy packages.

Step 1: ingest provider data

Your application receives partial chunks, SSE frames, or provider-specific event payloads.

Use @agentsy/core/sse when you need low-level SSE parsing.
Use @agentsy/providers/normalizers to convert provider-specific chunks into the common Agentsy event model.

Step 2: normalize to a common event vocabulary

@agentsy/providers/normalizers is the compatibility layer between providers and the rest of the ecosystem.

That lets downstream code stop caring whether a delta originally came from OpenAI, Anthropic, Gemini, or another provider-specific format.

Step 3: process incremental events

@agentsy/core/processor is the central orchestrator.

Its current surface includes:

LLMStreamProcessor
createProcessorEventAdapter
transform helpers such as createSmoothStream, createThinkingFilter, and createToolCallFilter

If you want an opinionated provider-coupled SSE pipeline, add createPipeline from @agentsy/providers/pipeline.

Step 4: extract specialized content

Depending on your use case, add targeted helpers:

@agentsy/core/thinking for reasoning tags
@agentsy/core/tool-calls for tool-call extraction and payload building
@agentsy/core/structured for schema-oriented JSON parsing and repair
@agentsy/core/xml-filter for tag filtering and privacy-oriented scrubbing
@agentsy/core/context for context shaping and dedupe
@agentsy/core/formatting for safe display-oriented formatting
@agentsy/core/recovery for snapshot/continuation behavior

Step 5: project into UI or integration surfaces

Once the stream is processed, you can send it into one or more projections:

@agentsy/renderers for plain-text rendering
@agentsy/ui for conversation-state updates
@agentsy/runtime/ag-ui for AG-UI protocol scenarios
@agentsy/vscode for editor-native experiences

Step 6: wrap with an agent loop if needed

@agentsy/orchestrator/agent sits above the processing pipeline when you need multi-step tool use, stop conditions, or loop management.

Why this shape matters

This layered structure lets you change the outer integration surface without rewriting the parsing core. That is exactly the kind of boring architectural win that saves a team from ritualized suffering later.

Stream processing flow ​

Step 1: ingest provider data ​

Step 2: normalize to a common event vocabulary ​

Step 3: process incremental events ​

Step 4: extract specialized content ​

Step 5: project into UI or integration surfaces ​

Step 6: wrap with an agent loop if needed ​

Why this shape matters ​