Thinking content no longer present in session JSONL files

As of ~today, (Claude Code v2.1.112), natural language thinking block content is no longer written to session JSONL files for CLI Claude Code sessions. For the last few weeks, a showThinkingSummaries: true workaround has been required to show thinking inline in the CLI, and now even that doesn't work. Only the VSCode extension remains as the last bastion of visible "thinking" at the moment.

The model's thinking output, even heavily summarized (and even with that summarizer leaking its system prompt ~10% of the time, which I'm sure many fellow users have noticed), is extremely valuable. It can save millions of tokens in conversations that don't need to be had.

To provide a % estimate on the value of "type: thinking" vs. regular "type: text" messages in Claude code, I'd say for cases where the user (speaking only for myself here) is actually sitting there interacting with the terminal (so not autonomous workflows), it ranges from 15%-40%. That is to say, it's a huge chunk of the overall value of the product.

The feature was removed silently, with no explanation. Users who follow the space have likely inferred it's a measure to defend against "distillation attacks", but it's still a relatively aberrant deviation from Anthropic's ethical best practices.

It also leaves a hotkey-bind in the CLI (ctrl+o) completely functionless (at least on Windows 11, with my Claude Code version).