[abhijit-s]

TL;DR: panes(), agents(), terminals() and the other snapshot reads hand the script engine one shared copy per revision instead of converting the whole list on every call, and keys() is listed again only when keys.json or the mods folder changes. Mod script time per frame roughly halves in a 30-agent layout.

Why

• Every snapshot read converted the whole list serde_json::Value → HostValue → JavaScript (JS), on every call, even when nothing changed since the last read.
• keys() re-read keys.json and looked up every mod a binding names on every call; a status-bar mod calls it every render.
• fix: avoid repeated pane reads in agent renders #485 and Open worktree hits the render limit in a repo with many worktrees ("This view could not be re-rendered; interrupted") #486 cut repeated reads in the agents and worktree mods; this does the same for the shipped tabs mod (it read panes() twice per tab) and makes each remaining read cheaper for every mod, shipped or not.

What changes

• gpui-kit: HostValue::Json(Arc<serde_json::Value>), host-to-script only. The QuickJS bridge builds the script's copy straight from the shared value (json_into_js). Each call still returns a fresh JS object, so a mod that mutates what it read sees no difference (test a_shared_json_value_reaches_the_script_as_its_own_copy).
• src/mods.rs: reader serves shared_part(key), copied once per revision of that key (the revisions upstream already keeps to redraw only affected mods). keys() serves keys_part(), rebuilt when keys.json or the mods folder's modified time changes.
• defaults/mods/tabs/main.js: reads panes() once per render.
• defaults/AGENTS.md: one line telling mods each list read is a fresh copy, so read once per render.

Numbers

Measured on my fork, whose mod read path matched upstream's apart from this change, with a headless app, a 30-agent layout (5 spaces × 6 tabs) and fake agents switching working/idle; two runs per side, p50 values:

	Before	After
One panes() read, 30 panes	58 µs	45 µs
One terminals() read, 30 terminals	168 µs	122 µs
One keys() read	355 µs	~125 µs
Script time per mod frame, idle	12.9 ms	6.8 ms
Script time per mod frame, status storm	14.6 ms	7.8 ms
Mod frame, idle	21.6 ms	16.6 ms
Mod frame, status storm	25.8 ms	18.9 ms
App instructions, idle / storm (M/s)	166 / 740	137 / 603

• A cached JSON text handed to JSON.parse was tried first and was slower than building the objects (QuickJS's JSON parser).
• Most of the remaining mod frame is GPUI layout, prepaint and materialize of the panels' trees.

Tests

• mods::tests::reads_share_one_copy_of_a_snapshot_part_until_it_changes, mods::tests::keys_are_listed_again_only_when_their_files_change, and the gpui-kit bridge test above.
• just check passes.

[abhijit-s]

TL;DR: the gpui-kit test this PR adds, a_shared_json_value_reaches_the_script_as_its_own_copy, passes. just check doesn't run gpui-kit's tests, so I ran them separately.

• How: cargo test -p gpui-shell --lib in vendor/gpui-kit, with that workspace pointed at the root's [patch] entries and Cargo.lock. Standalone, it resolves a newer rsa that llrt_crypto doesn't compile against, and stock gpui-pre without spin_clock_ms.
• Result on this branch: 717 passed, 5 failed.
• Result on master (e77f112): 716 passed, the same 5 failed, so they come from upstream, not this PR:
  • plugin::tests::runtime_check_returns_the_rendered_description
  • tests::render::a_readonly_script_editor_remains_selectable
  • tests::render::the_bundled_example_application_loads_and_renders
  • tests::render::the_todolist_example_exercises_the_runtime
  • typings::tests::every_element_method_is_accounted_for

[abhijit-s]

TL;DR: merged master, including 029323e (mod code reused across copies).

• Conflict: only src/mods.rs, where this PR's two tests and 029323e's loading_copies_reclaims_closed_command_channels_without_dispatch were added at the same place; all three are kept.
• host_modules.rs: 029323e's rewrite and this PR's HostValue::Json / json_into_js merged without conflict; the PR's diff against master there is still only those additions and their test.
• just check passes locally on macOS. I haven't re-run the gpui-kit suite (cargo test -p gpui-shell --lib) on the merge; the numbers above are from before it.