[TechNomadCode]

Slice 20: the visitor sees when the model is thinking and for how long, never what it thinks.

• While an answer is on its way and no step runs and no reply arrives, the moving line reads "Thinking" (opencode's row, 13px medium on a 16px line) instead of "Working…".
• Each reply follows a line "Thought 8s" ("1m 5s" from a minute on), the word in white, the time in grey. It doesn't open: the reasoning's words still never reach the browser.
• The browser times each wait itself: the chat's transport pipes each answer's stream through timeThoughts, from the request or from the moment the last step running finished, to the next reply's first text or step. It adds a data-thought part to the answer at that place, so it's kept with the conversation and outlasts a reload. Answers kept before this have none and show none, so STORED_FORMAT stays 2.
• Searches in a row fold together with the thought lines around them, as opencode's groupContent does. The count still counts only the searches.
• An answer stopped before it finished anything is still left out of what goes back, even with a thought's time in it.

src/server: no files changed. src/lib: no files changed. The times go back to the server inside the answers. safeValidateUIMessages accepts them, and convertToModelMessages leaves them out (ai 7.0.143). A unit test runs the fake model on a conversation with and without the times and shows the model's input is identical.

Checks (local, fakes): verify passed (488 unit tests). test:e2e passed all 67 in the second full run; the first failed 6 waiting on the fake model's answer (KI-7, noted). Page checked at 1440px and 375px on a dev server with the fakes.

Unverified: the real model's waits and times on staging; Safari and Firefox.