A window is the default, in the shell and out of it

[torstei]

graph could only write a file, so seeing a plot meant leaving the shell for timbergraph — whose --window timbersh never exposed. Adds window and ascii clauses, with a window as the default.

graph http_requests from [class=tally] by status;          window at a desk, ASCII over ssh
graph http_requests from [class=tally] window;             force one (refuses if impossible)
graph http_requests from [class=tally] ascii;              force text
graph http_requests from [class=tally] png '/tmp/r.png';   a file, as before

Three conditions, and none implies the others

A window needs a display, and a gnuplot built with an interactive terminal (gnuplot-nox has none), and — for the default only — a terminal to be watching it.

That third one isn't fussiness. timbersh -q < script.tsh > out.txt at a desk would otherwise open a window and write an empty file, turning a working non-interactive run into nothing. Naming window overrides it, since a window isn't stdout. It follows the rule view and into view already use.

Where a window can't be drawn the default is ASCII, so ssh sessions behave exactly as before. A named window that can't be drawn is refused rather than downgraded — the point of saying it is to get one — and the message names both missing pieces.

Two smaller decisions

The probe reads gnuplot's terminal list, not its exit code. set terminal nosuchterm exits 0 and only warns on stderr, so a probe trusting the status would call every terminal available. One gnuplot at ~10 ms, cached, since it's a fact about the machine.

Two destinations is now refused instead of resolved in silence: window png 'f' wrote the file and no window, window ascii was ASCII, and png beside svg quietly preferred the png.

timbergraph's own CLI keeps ASCII as its default — deliberately asymmetric. It's a pipe tool that gets scripted and redirected; timbersh is a console with somebody in front of it. Its --window now names what's missing rather than relaying gnuplot's complaint.

Tests

Six new, and the suite can now say what the machine can do — a fake-gnuplot reporting a chosen terminal list, plus a per-session env override. Without that a suite would inherit the developer's DISPLAY and assert one thing at a desk and the opposite in CI.

Two things I got wrong on the way and fixed, because they're the failure modes worth knowing about:

• My first four tests all ran with no display, so none of them exercised the auto-window default at all — the actual feature. Confirmed by disabling the default and watching them all still pass. The default test now runs on a PTY, since the default is gated on one.
• Two tests asserted only the absence of the window marker, which would also pass if nothing ran. The fake now prints a marker for both paths, so every case asserts what was actually asked for.

a_graph_draws_a_window_by_default_where_one_can_be_drawn fails with the default disabled (verified). 67/67 timbersh tests pass, as do the timbergraph, timberview and timberfs-client suites.

⚠ Not shipped by the timberfs package — timbersh/timbergraph live in timberfs-sh, so this needs a timbersh-v* tag to reach anyone.

🤖 Generated with Claude Code