Module harmony 

OpenAI Harmony format parsing for GPT-OSS models.

The Harmony format uses channels to separate different types of content:

• analysis: Chain-of-thought reasoning (internal, not for end users)
• commentary: Tool call preambles and explanations
• final: User-facing response content

Tool calls in Harmony are indicated by the recipient field being set to:

• functions.tool_name for user-defined tools
• browser.search, browser.open, browser.find for browser tool
• python for python tool

This module provides incremental parsing of Harmony-formatted token streams.

Structs

HarmonyAccumulated

Accumulated content for each channel

HarmonyContext

Context for tracking Harmony parsing state within a sequence.

HarmonyDelta

Incremental delta from Harmony parsing

HarmonyToolCall

A tool call parsed from Harmony format

Enums

HarmonyChannel

Channel types in Harmony format

Functions

is_harmony_encoding_ready

Check if the Harmony encoding has been initialized.

is_harmony_template

Check if a chat template uses Harmony format by looking for Harmony markers.

prewarm_harmony_encoding

Pre-initialize the Harmony encoding. This MUST be called from a non-async context (e.g., during pipeline loading) before any async code runs. The openai-harmony crate uses reqwest::blocking which creates its own tokio runtime, so it cannot be called from within an existing async context.