A Long Assessment Now Reads as One Run

August 16, 2026 · InfraScout Team

A serious troubleshooting conversation is not one question and one answer. The AI plans, calls a tool, reads the result, calls three more, revises, and comes back — and every one of those iterations used to render as its own message bubble. A thirty-nine-call run was a wall of small cards, several of them empty, because an iteration that only fired tools has no prose in it.

You had to scroll past all of it to find the answer. This release folds that whole stretch into one node.

The assessment run

Everything the AI does between two of your messages is now a single assessment run in the timeline: one header, one timestamp range, one dotted rail down the side. The work along that rail is broken into milestones — the narration the AI wrote, with the tool calls belonging to each step attached underneath it. Milestones start folded, so the run reads as a short list of what happened rather than a transcript of how.

The answer is not part of the ledger. The last message in a run renders exactly as an answer always has — its own reasoning block, the full prose, the copy button, the sources footer, and the refusal card when the model declines. A turn that took one iteration is a run of one, and looks precisely like it did before. Nothing about the shape of a short conversation changed.

Cards you interact with stay out of the ledger too. Questions the AI asks you, delegated Task Agent runs, and charts render as full cards next to the answer, because they are content you read and act on rather than steps you skim past.

Inside a milestone, each tool call is a compact row with the tool name as a badge — blue when it succeeded, amber while it runs, red on failure, and grey when it repeats a call already made, so a batch of three identical calls reads as one action. Durations under a second are simply not printed; the ledger used to be full of meaningless 0.0s readings.

Finished work no longer vanishes under your cursor

The first version of this had a real problem, reported plainly: "tools get hidden immediately after they finish. So it's not clear for the user what happened."

Two things were folding with zero dwell. A completed tool call moved into its collapsed group the instant its result landed, and when the iteration finished the whole thing demoted into a folded milestone. The row you were reading disappeared, and the only evidence left was a counter going up.

Both surfaces now hold for eight seconds before folding, and they hand off to each other: a row still holding when its iteration ends keeps reading inside the freshly opened milestone. The rows visibly travel into the ledger instead of blinking out of one place and silently existing in another.

Two rules make that safe rather than merely slower:

• Any deliberate interaction cancels the fold permanently. Click a held row, open a milestone, click inside one, or follow a citation into it, and it stays open. Folding content away from someone who is reading it is the bug being fixed — the fix must not reintroduce it on an eight-second fuse.
• Only work that finishes while you are watching holds. Reopen an old conversation and it renders folded immediately, rather than splaying its entire transcript open for eight seconds.

If you have reduced motion enabled, the transitions drop but the hold stays. The dwell is timing, not decoration.

Room to actually read it

The conversation view was spending about forty percent of its height on chrome, and reserving a third of its width for a user message bubble that rarely needed it.

The audit disclaimer now appears on the start form only. Every conversation begins there, so you still see it every time, but it no longer takes a row from every open chat. The composer collapsed to a single row with the keyboard legend and the send controls at the end of the text line. The header is one row: title, then the context chips for connection, tool groups and cost, then the plan. The task plan moved out of its banner and into a chip in that header that opens the full list of tasks when you click it. And the "Response complete" pill stopped rendering once a response has settled — a finished response is self-evidently finished.

Your own messages now sit in a fixed-width track instead of claiming a third of the panel. On a narrow window nothing changes; on a wide one everything above that width goes to the AI's column, which is where the tables, tool output and charts are.

At a typical laptop size that is roughly a third more message area.

Two smaller fixes

Citations pointing at a question the AI asked you did nothing when clicked. The Sources footer rendered the entry correctly, but the card it pointed at carried no anchor for the reveal to find, so the click landed nowhere. It now scrolls to the card and flashes it like any other citation.

And your own avatar in the timeline read as disabled — a ghosted circle beside the AI's saturated one. It now carries a full-strength accent color, so the two participants in a conversation look like two participants.

See Chat for the full conversation surface.

Questions or feedback? Reach us at info@infrascout.cloud.