DOCS
Stream protocol
One POST, one SSE response, a closed event list. Frontend vs backend tool.call targets.
Framing
Each event is an SSE line
data: {json}
followed by a blank line. The stream ends with
data: [DONE].
Client Accept in sveda mode is
application/vnd.sveda.stream+json.
Response
Content-Type
is
text/event-stream.
The runtime sets
x-sveda-protocol-version: 1.0.
POST /sveda/stream
Accept: application/vnd.sveda.stream+json
Content-Type: application/json
x-sveda-embed-token: sveda_embed_…
Content-Type: text/event-stream
x-sveda-protocol-version: 1.0data: {"type":"message.start","chatId":"chat_1"}
data: {"type":"text.delta","delta":"Hello","chatId":"chat_1"}
data: {"type":"message.end","finishReason":"stop","chatId":"chat_1"}
data: [DONE]Event list
Closed set from
StreamEvent.
Do not invent aliases.
message.start Turn opened. Optional chatId, messageId, timestamp.
text.delta Assistant token. Field delta.
reasoning.delta Reasoning token. Field delta.
tool.call toolCallId, toolName, input, target backend | frontend.
tool.result toolCallId, toolName, output; optional renderHint, renderData.
tool.progress Optional phase plus tasks[{ id, label, status, detail }].
context.usage usedTokens, maxTokens, percent.
chat.title Generated title for the chat.
max_steps maxSteps, canContinue. Client may continueAfterMaxSteps().
message.end Optional finishReason and usage. Turn closed.
error code and message.
tool.call target
target
is
backend
or
frontend.
Backend tools run in
sveda-server
(builtins and host MCP). Frontend tools run in the browser via
SvedaToolRegistry
—
MCP and tools.
data: {"type":"tool.call","toolCallId":"call_1","toolName":"search_code","target":"backend","input":{"query":"SvedaClient"}}
data: {"type":"tool.call","toolCallId":"call_2","toolName":"get_selection","target":"frontend","input":{}}