The preview file list is mostly the agent's own instructions, not what it produced

[astetic-dev]

What happens

The preview's file dropdown lists every .html/.md file in the working folder (three levels
deep, eighty of them). For an agent that runs a documented work process, most of those files
are the process description — CLAUDE.md, rules.md, identity.md, card-types.md,
response-templates.md, scoring-rubric.md. Those are what the agent reads. What you want
to look at is what it just wrote, and that is somewhere in the same list under a name that
looks no different.

Filtering on .html is not the answer: agent output is as often .md as it is .html.

Three smaller things make it worse:

• The label is the bare file name, so four different acties.md files are four identical rows.
• Opening a file that is not in the list (clicked in the terminal, outside the folder, deeper
  than three levels) shows that file but leaves the dropdown pointing at something else — the
  selector then names a different file than the page you are reading.
• Opening the preview on a specific file first renders the newest file in the folder and only
  then the file you asked for, so you read one page you did not ask for.

Fix

Group by when the file was written, which is the one signal that separates output from input
and does not depend on the extension:

• Since this tab opened — what this run of the agent produced.
• Earlier today / Older — everything else, newest first, as before.

And:

• Label rows with the path relative to the working folder plus the time it was written
  (_index/dashboard.html · 22:47), so acties.md in four folders is four distinct rows.
• Always show what is on screen: a file outside the list is added to the top and selected.
• list_html returns rel (path relative to the working folder) instead of name.
• Refresh (↻) re-reads the file you are looking at instead of jumping to the newest one; it
  still picks the newest when nothing is open yet.

[astetic-dev]

Follow-up: the folder scan is not where the output is

Measured after the first fix: an agent that runs a work process from folder A writes its
report into project folder B (C:/Users/AST/claude/projecten/_index/*.html), and says so in
the terminal ("Open in browser: …"). list_html only scans the session's own folder, so those
files are not in the list at all — the grouping had nothing to lift. Three files touched to
test it stayed invisible.

So the list also takes the paths the agent named: the frontend reads .html/.htm/.md
paths out of the terminal buffer (logical lines, so a path wrapped over several rows is still
one path; both the normal and the alternate buffer, because a full-screen TUI has no scrollback
in the alternate one), and a new stat_files command reports which of them exist and when they
were written. Only those written since the tab opened are added — the agent names the files it
reads just as often, and those are old.