implementation detail · filed under inference & serving
Harmony history stop-token normalization
A generated reply’s trailing return token is replaced with an end token when storing conversation history.
- source
- 1
- model
- 1
- lab adopt it
- 1
- strongest
- used
How sources treat it
One count per evidence span, weakest treatment to strongest.
used 1
Documented in
Evidence
1 span quoted from the sources, strongest treatment first.
When you add the assistant’s generated reply to conversation history for the next turn, replace the trailing <|return|> with <|end|> so that stored messages are fully formed as <|start|>{header}<|message|>{content}<|end|>.
usedinference servingin gpt-ossOpenAI
Filed alongside
Other methods under inference & serving :: agentic scaffolding.
Advisor strategyAutomatic tool choicePython toolFunction callingHarmony formatBrowser toolGeneric tool-call parserMCP tool configurationMulti-agent collaborationRole-based instruction hierarchyThinking with toolsVisual Search ToolXML-based tool-call schema with DSML tokenAgent Team modeAgentENVAgentic searchAgentic tool callingApp-server mode with adapted tool schemasAppArmor and eBPF sandbox policiesAssistant output channelsAsynchronous teammate spawningBash commands for context retrievalBash computer-use agentBrowsing tool with domain filtering