thinkingContext function
The extra_context that sets thinking for a LiteRT-LM conversation, the
one source for the native and the web path.
enable_thinking is what the templates read — Qwen3, Gemma 4, SmolLM3, and
upstream's own web chat app sends the same key. Always an explicit bool: a
Qwen3 template (enable_thinking|default(true)) thinks when the key is
absent, so sending nothing for "off" left those bundles reasoning —
Qwen3-0.6B_dynamic spent a 768-token budget on reasoning and never
answered. A template that ignores the key (the legacy Qwen3
prompt_templates) is unaffected, and Gemma 4 renders the same prompt for
false as for no key.
Implementation
Map<String, Object> thinkingContext(bool enableThinking) => {
'enable_thinking': enableThinking,
};