Update NEWS.

NEWS

@@ -2,6 +2,45 @@ Note: There is no guarantee that version mismatched client and server will
 be able to talk with each other. Network protocol breakages won't be listed
 here.
 
+vx.xx.x (2026-xx-xx)
+--------------------
+
+- External frames are now omitted in the single-line call stack list visible
+  in messages list, or in memory allocation info window.
+- System tracing on Windows has been refactored to be more robust.
+- Tracing on Arm macOS will now have more precise timer readings.
+- Extended CUDA support to track some previously missing memory operations.
+- Added support for setting message's source and severity, through the
+  TracyLogString macro.
+- Various LLM integration improvements.
+  - The protocol has been updated to use model templates. As a result, tools
+    are now specified in a common way and the reasoning is performed in a
+    separate content stream.
+  - Smaller models are now viable to use. Models as small a 4B parameters do
+    now work really well. You can run such models on virtually all hardware.
+  - Added horizontal scroll bars to code segments.
+  - LLM thinking regions are now hidden by default.
+  - Markdown tables are now properly rendered.
+  - Sampled execution costs are now included in assembly attachments.
+  - Source code retrieval now has an optional line context parameter.
+  - Added ability to search the code for keywords.
+  - Calls in assembly attachments are now annotated with function names.
+- Full user name is now stored in trace info.
+- You can now attach complete symbol assembly to LLM.
+- Tweaked high-resolution scroll handling on Wayland.
+- Entry call stacks can be now attached to LLM (previously it was only
+  regular call stacks).
+- External functions can be filtered out in the sampling statistics view.
+- Tweaked external paths heuristics.
+  - Check for both 64-bit and 32-bit versions of Program Files directory.
+  - Hidden unix files and directories are now also considered external. For
+    example: $(HOME)/.cache/cpm/somelib/file.h.
+- LLM configuration now also expects a "fast model" to be set. It is intended
+  for quick queries that do not require tool use or reasoning.
+- Call stack window can now provide LLM summaries. It uses the fast model.
+  - These summaries can be performed automatically. Enable in LLM settings.
+
+
 v0.13.1 (2025-12-11)
 --------------------