Rework model selector with loadout-backed slots

[senxd]

Summary

• Rebuild the model picker on top of the five-slot loadout: slots carry harness, model, effort, provider model options, and a per-slot activation shortcut
• Per-slot shortcut recording with physical-key conflict detection against the keymap, other loadouts, and reserved composer/app chords; Some("") disables a slot shortcut while None keeps the positional mod-shift-N default
• Slot cards reorder via drag-and-drop or arrow keys, keeping model options and shortcuts attached to the card; provider catalog columns reorder the same way and persist a device-local provider_order
• Slot menus split into frosted root + submenu cards (agent / model / effort / model options / speed / shortcut / remove) with full keyboard navigation, replacing the single glass slab over union bounds
• Catalog requests are scoped to the effective device and workspace cwd, and reload on scope change; pending agent selections never persist an empty model
• Claude harness passes model[1m] idempotently via a shared model_argument helper; picker rows show each row's resolved loadout shortcut badge
• Split the loadout page into loadout/{bindings,cards,catalog,indicator,options}.rs and add extensive unit + gpui tests (shortcut normalization, reorder semantics, menu navigation, macOS shifted-digit aliases)

Screenshots

Redesigned the model selector (Reasoning bar, repositioned layout)	Custom model loadouts

Test plan

• cargo check --workspace --all-targets
• cargo test -p zeron-ui loadout (67 tests)
• cargo fmt --check
• Manual: record/clear/reorder slot shortcuts on macOS and verify Cmd+Shift+1-5 defaults still work
• Manual: drag models from provider columns (incl. OpenRouter partition) into slots

Generated with Devin

[senxd]

Also this was not listed above but the Pi provider in the model selector respects the model filter that is set in the Pi config. It should accurately represent what /models looks like in the pi cli.

[wingleeio]

Security audit

Went through the full diff (39 files, +7910/−500) looking for malicious code, unintended side effects, and accidental exploit surface. Nothing concerning found.

• Dependencies: the only new dep is glob = "0.3" (workspace + zeron-harness). It resolves to the glob 0.3.4 package that was already in Cargo.lock transitively — same crates.io source and checksum, no new package block. Used only for enabledModels pattern matching in acp/mod.rs.
• Process / network / filesystem: no new Command spawns, no new HTTP calls, no new sockets. All UI ↔ engine traffic goes through the existing engine.client().call(methods::LIST_HARNESSES | LIST_MODELS) path. The only new fs reads are read_pi_settings on $PI_CODING_AGENT_DIR|~/.pi/agent/settings.json and <cwd>/.pi/settings.json — read-only, JSON-parsed, failures logged via tracing::warn! and treated as "no patterns". Test code writes only into tempfile::tempdir().
• fake-cursor-shim.sh: test fixture only; the change adds an effort parameter to the fake catalog and a scenario:reasoning branch that echoes canned frames. No commands executed from input.
• ListModelsParams.cwd (new RPC input): client-controlled path forwarded to models_for_cwd. For Pi it becomes the agent's current_dir for the discovery probe and the location of the project .pi/settings.json. This is the same trust level as existing SetChatCwd / history RPCs that already take an unjailed cwd, and the chat itself already runs Pi in that cwd, so no new capability is exposed. Non-Pi harnesses ignore it (default trait impl).
• Shortcut recording: bindings.rs::on_key_down only captures a keystroke while recording_slot.is_some(), stores the normalized combo string in settings, emits RecordingChanged, and stops. No key logging or persistence beyond the slot field.
• Unicode: no bidi/zero-width/homoglyph characters in added lines (only ⌘ ⇧ → – — … in display strings/comments).
• Formatting churn in unrelated files (schema.rs, uploads.rs, e2e.rs, normalize.rs, subagent.rs, browser/*.rs, examples): main currently fails cargo fmt --check on exactly those spans; this PR just brings them in line. Pure whitespace/line-wrapping, verified.

Verification (Linux, PR head 37ef3a45)

Check	Result
cargo fmt --check	clean
cargo check --workspace --all-targets	pass
cargo test -p zeron-ui loadout	67 passed
cargo test -p zeron-ui pickers / shortcut / settings::	28 / 42 / 140 passed
cargo test -p zeron-harness	209 passed (incl. new Pi scope, model_argument, cursor effort tests)
cargo test -p zeron-engine --test e2e	20 passed, 1 ignored
clippy	2 pre-existing deny errors in opencode/tests.rs (untouched by this PR), nothing new

Manual test under Xvfb (X11, real claude/codex/cursor-agent/devin CLIs on PATH, mock engine)

Composer model picker: pinned loadout rows with shortcut badges, harness rail, context-window + reasoning bar	Picker config popover

Settings › Model Loadout after drag-and-dropping three models into slots (catalogs populated from live CLIs)	Slot root menu

Shortcut submenu, recording	Conflict detection (recorded Ctrl+Shift+2 on slot 1)	Slot activated from picker row → composer chip updates

Worked as described: drag-and-drop into slots, checkmarks on catalog rows, slot menus (agent / model / effort / context window / fast / shortcut / remove), shortcut recording + conflict notice, clicking a loadout row in the picker, and Ctrl+N auto-applying slot 1 to a new session. No loadout/picker/toast warnings or panics in the app log.

Findings

1. Default loadout chords Ctrl+Shift+1–5 don't dispatch on Linux/X11 (bug)

Pressing Ctrl+Shift+1…5 in the chat view never fired ActivateLoadout — no model switch on a filled slot and no "slot is empty" toast on an empty one (control: Ctrl+N, Ctrl+, work fine).

Root cause: gpui_linux::keystroke_from_xkb reports Ctrl+Shift+1 as {control, key: "!", shift: false} (shift is intentionally dropped for symbols), and Linux uses DummyKeyboardMapper, so a ctrl-shift-1 binding never matches. The PR already has the fix — loadout_symbol_alias (mod-shift-1 → mod-!) — but it is only bound under #[cfg(target_os = "macos")] in apply_keymap (shell.rs:409). The gpui tests pass because simulate_keystrokes bypasses XKB translation. Suggest dropping the cfg gate (bind the alias on every platform; harmless where the shifted form is what arrives).

The conflict screenshot shows the same mechanism from the recorder side: the notice reads "Ctrl+@ is already assigned to loadout slot 2" — the physical-key comparison correctly resolved it, but the display should fold mod-@ back to Ctrl+Shift+2 (invert SHIFTED_DIGITS in normalize_loadout_shortcut or in display_combo).

2. heal_reserved_shortcuts can silently rewrite a user's custom app shortcut (behavior note)

Because the five default loadout combos are now "owned", any persisted keymap entry on mod-shift-1..5 is reset to its default on load, and if the default is taken it becomes "" (disabled) with no notice. fixed_loadout_shortcuts_heal_persisted_conflicts asserts exactly this. Probably acceptable, but worth a changelog/toast since it changes existing users' bindings on upgrade.

3. Pi discovery error path hides the cause (minor)

In models_for_cwd, Err(_) if patterns.is_some() => Ok(Vec::new()) returns an empty catalog when the probe fails and the user has enabledModels, so the column shows nothing rather than an error. Consider propagating the error (or at least tracing::warn!).

4. Per-cwd Pi probe is uncached

Pi now spawns a fresh discovery process on every LIST_MODELS (the models_cache short-circuit is skipped so project-local enabledModels can apply). The UI caches per scope so it's only on scope change, but worth keeping in mind.

Overall: safe to merge from a security standpoint; #1 is a functional regression on Linux I'd fix before merging.