[FEATURE] No native way to surface reasoning from type: "thinking" transcript blocks

Preflight Checklist

• [x] I have searched existing requests and this feature hasn't been requested yet
• [x] This is a single feature request (not multiple features)

Problem Statement

Claude Code session transcripts already store full model reasoning in type: "thinking" blocks, separate from the visible reply. There's no built-in way to search, reference, or reason over that content later -- it's only accessible by digging into the raw JSONL yourself.

I ran into this building a personal nightly memory-review tool that reads back through recent sessions to catch things worth remembering. It originally only looked at visible user/assistant turns. Once I extended it to also read thinking turns, it started surfacing real decisions and context that were reasoned through but never made it into the final reply -- genuinely useful signal that's otherwise invisible unless you hand-parse the transcript files yourself.

Proposed Solution

Some way to include thinking-block content in built-in transcript search/reference tooling, not just visible turns -- even something as simple as a documented, stable way to query thinking content, so people building memory/reflection tooling on top of Claude Code don't have to reverse-engineer the JSONL format themselves.

Clearer documentation that thinking blocks live in the session JSONL, and how to access them reliably, would help a lot even without new tooling.

Alternative Solutions

Hand-rolled a script that reads the raw session JSONL directly, filters for type: "thinking" records, and extracts the thinking field's text alongside the normal visible turns. Works, but it's undocumented/unofficial behavior I had to reverse-engineer by inspecting a real transcript file, and it could break if the internal format changes.

Priority

Medium - Would be very helpful

Feature Category

Developer tools/SDK

Use Case Example

1. I built a nightly tool that reads back through my Claude Code session transcripts to find things worth saving to long-term memory.
2. Initially it only looked at visible user/assistant messages.
3. I added logic to also read type: "thinking" blocks from the same JSONL files.
4. This surfaced real decisions/reasoning that never appeared in the visible reply -- information I would have otherwise lost.
5. With native support, anyone building similar memory/reflection tooling on Claude Code wouldn't need to hand-parse the JSONL format themselves.

Additional Context

Not a bug -- a product gap noticed while building something for personal use, that seems bigger than just my own setup.