Release 0.13.2 · LIUTod/scream-code

What's new in 0.13.2

Subagent tree lines never break: long descriptions/activity no longer
cut the │ connectors when multiple agents run in parallel (WolfPack and
plain multi-agent calls alike).
No more height bouncing: streaming activity is rendered as a single
stable row, so panels stop growing/shrinking while agents work.
Single-agent preview rows stabilized: thinking/output previews are
truncated to one row and no longer push the layout around.
Approval panel keeps its buttons: very long commands (e.g. big
heredocs) are folded to 10 lines with a "N more lines hidden" hint, so
the Allow Once / Deny options stay on screen.

First-run configuration card: fresh installs get a guided card that
jumps straight into model configuration (Enter to configure, Esc to skip).

reportedModel in wire logs: each LLM call now records the model the
provider actually served (gateway routing included), alongside the
requested model — visible in the session wire.jsonl (step.end).

Faster startup on Windows: host-environment detection (which walks
PATH and probes dozens of paths) is now cached on disk, cutting startup
by seconds on subsequent launches.