Skip to content

interleaved-thinking

Generated from plugins/interleaved-thinking/README.md.

Provider extension for the OpenAI-compatible provider that omits thinking blocks from previous assistant turns when building a request.

What it does

When enabled, every outgoing request strips prior-turn thinking in both forms it appears in:

  • <think>...</think> segments inside assistant message content (models that interleave thinking directly in content, for example through Ollama).
  • reasoning / reasoning_content fields on assistant messages (the separate-field form used by Z.ai GLM and other OpenAI-compatible providers).

The session history itself is never modified — only the request payload. The current turn's thinking is unaffected.

Why it exists

"Preserve previous-turn thinking" is otherwise provider-dependent: Ollama has no request-side setting for it, and providers differ in whether prior thinking even round-trips. Including or excluding this plugin makes the switch uniform:

Plugin Previous-turn thinking in requests
included omitted, on every OpenAI-compatible provider
excluded the provider's default behavior

The task-guidance evaluation suite uses it as the preserve-thinking switch for its producer variants.

Usage

Add the plugin to an agent configuration that uses the reference_openai_compatible provider:

{
  "plugins": [
    "path:${env:BUILTIN_PLUGINS}/interleaved-thinking"
  ]
}

Tests

python -m pip install -e "plugins/interleaved-thinking[dev]"
pytest plugins/interleaved-thinking/tests -q