Thinking: ThinkConfig

Thinking is a first-class stream, not mixed into answer text. Local and cloud backends share runa_core::ThinkConfig.

Modes

Mode	Meaning
`off`	No reasoning channel
`on`	Model default thinking
`budget { tokens, grace }`	Hard cap on reasoning tokens, then `grace` extra
`effort` (`low`/`medium`/`high`/`max`)	Mapped to a budget (P3.4 table) or a cloud `reasoning.effort`

--think-budget and --effort override --think on. They cannot be combined with each other or with --think off.

Show vs hide

ThinkConfig.show (CLI --show-reasoning, config [think] show, env RUNA_SHOW_REASONING) prints GenEvent::Reasoning / cloud reasoning deltas on stderr. --no-show-reasoning forces hide. Answer tokens stay on stdout.

Surfaces

CLI: --think, --think-budget, --effort, --show-reasoning
REPL: /think on, /think budget 2048 show, /think effort high, /think hide
Config: [think] in docs/config.md
Cloud: OpenAI reasoning.effort / Responses reasoning; Anthropic adaptive vs enabled + budget (docs/adr/d07-thinkconfig.md)