Verify each finding against the current code and only fix it if needed.

Inline comments:
In @.github/workflows/release.yml:
- Around line 8-9: The scheduled workflow uses GITHUB_REF_NAME (branch) to
derive raw and then validates semver, but scheduled runs set the event to
"schedule" and have a branch like "main", causing semver validation to fail; fix
by detecting scheduled runs and injecting an explicit version branch (e.g., set
raw = "nightly" or "scheduled") before the semver validation step: update the
workflow expression that sets raw (the variable named raw derived from
GITHUB_REF_NAME/GITHUB_REF_NAME) to check github.event_name == 'schedule' (or
compare to the schedule trigger) and assign a stable string like "nightly" for
scheduled runs so the subsequent semver validation logic (the validation step
that currently rejects non-semver raw values) accepts scheduled runs.

In @.vscode/settings.json:
- Line 8: Replace the invalid VS Code setting key "js/ts.tsdk.path" with the
correct key "typescript.tsdk" while keeping the value
"node_modules/typescript/lib"; locate the entry that currently reads
"js/ts.tsdk.path" and rename the key to "typescript.tsdk" so VS Code will use
the workspace TypeScript SDK.

In `@apps/server/src/provider/Layers/CodexAdapter.test.ts`:
- Around line 1030-1033: The test in CodexAdapter.test.ts currently asserts the
native log prefix mistakenly as "NTIVE"; update the assertion to validate the
payload message directly instead of the log prefix: after reading threadLogPath
into contents (variable contents), replace the regex /NTIVE: .*"message":"native
flush test"/ with a regex that only matches the JSON payload message such as
/"message":"native flush test"/ so the test verifies flush behavior rather than
the log prefix.

In `@apps/server/src/provider/Layers/CodexAdapter.ts`:
- Around line 1401-1416: The event consumer fiber (eventFiber) is being forked
into the temporary scope created by startSession (via Effect.scoped) using
Effect.forkChild, so when startSession's scope closes the fiber is interrupted
and runtime.events are dropped; change the fork target to the persistent session
scope (sessionScope) instead of forking as a child of the temporary scope—i.e.,
when creating the fiber from Stream.runForEach(runtime.events,
...)/.pipe(Effect.forkChild) ensure you fork into sessionScope (the session's
persistent scope) so the event pump lives for the full session lifetime rather
than being tied to startSession's ephemeral scope.

In `@apps/server/src/provider/Layers/CodexProvider.ts`:
- Around line 389-394: The probe call in CodexProvider is using process.cwd()
which ignores the configured server workspace; update CodexProviderLive to read
ServerConfig (serverConfig.cwd) and pass that cwd into the probe call (used when
building the options for checkCodexProviderStatus / probe), replacing
process.cwd() with the injected serverConfig.cwd so the probe (function probe
and the checkCodexProviderStatus flow that produces probeResult) runs against
the configured workspace while preserving the existing timeout
(PROVIDER_PROBE_TIMEOUT_MS) and other parameters (binaryPath, homePath,
customModels).
- Around line 218-239: The initialize call is using a hardcoded version
("0.1.0") instead of the helper; replace the inline clientInfo object passed to
client.request("initialize", ...) with the values produced by
buildCodexInitializeParams() (or merge its returned params) so the
clientInfo.version comes from packageJson via buildCodexInitializeParams; locate
the initialization block around the clientContext/CodexClient.layerCommand and
update the initialize variable creation to use buildCodexInitializeParams()
output rather than the hardcoded literal.

In `@apps/server/src/provider/Layers/CodexSessionRuntime.ts`:
- Around line 313-340: The code in buildTurnStartParams is dropping label-only
user-input options (it rejects options when description is missing), causing
valid Codex user-input requests to be suppressed; update the logic that builds
turnInput (inside buildTurnStartParams) to include option objects even when
description is absent—ensure V2TurnStartParams__UserInput entries preserve
label-only options by setting description to undefined or omitting it rather
than filtering the option out, and keep the existing handling for attachments
and text so label-only options are pushed into turnInput the same way other
inputs are.

In `@apps/server/src/provider/Layers/OpenCodeAdapter.test.ts`:
- Around line 79-99: The test double connectToOpenCodeServer currently always
registers a finalizer that mutates runtimeMock.state (closeCalls/closeError),
which makes configured-server sessions (when serverUrl is provided) incorrectly
exercise finalizer failure paths; change connectToOpenCodeServer to only add the
finalizer when serverUrl is not provided (i.e., local/test server path) so
configured external connections return without owning lifecycle, and move any
test that simulates finalizer throws (stopAll close-error coverage) into the
layer or test that calls connectToOpenCodeServer with no serverUrl; reference
connectToOpenCodeServer, runtimeMock.state.closeCalls, and
runtimeMock.state.closeError when making this conditional change.

In `@apps/server/src/provider/opencodeRuntime.ts`:
- Around line 434-440: The code currently masks failures by converting
child.exitCode failures into 0 via Effect.orElseSucceed(() => 0); update the
return so callers can distinguish real success from errors: either remove the
orElseSucceed wrapper and make exitCode an errorful Effect<number,
OpenCodeRuntimeError> (update OpenCodeServerProcess/OpenCodeServerConnection
types accordingly) so failures surface, or, if keeping a successful-effect
signature, replace the mask with a non-colliding sentinel (e.g., -1) and ensure
failures are logged/annotated; locate the return that references child.exitCode
and Effect.orElseSucceed and adjust the behavior and types consistently.

In `@apps/web/src/components/chat/ModelPickerContent.tsx`:
- Around line 330-337: The handler is incorrectly using
targetModelKey.split(":"), which breaks slugs that contain colons; update the
parsing of targetModelKey (used where modelJumpModelKeys and jumpIndex are
referenced and in other handlers around lines noted) to split only on the first
":"—e.g., find the first colon index and take the substring before as provider
and the remainder as slug—then call handleModelSelect(slug, provider);
alternatively centralize encode/decode logic for keys (provider:slug) to a
single helper used by these handlers to avoid truncation bugs (refer to
ProviderKind, targetModelKey, modelJumpModelKeys, jumpIndex, and
handleModelSelect).

In `@apps/web/src/components/chat/ModelPickerSidebar.tsx`:
- Around line 156-207: Remove the hardcoded "coming soon" Gemini and Github
Copilot UI blocks in ModelPickerSidebar.tsx: delete the two
Tooltip/TooltipTrigger blocks that render buttons with
data-model-picker-provider="gemini-coming-soon" and "github-copilot-coming-soon"
(the blocks that include Gemini, GithubCopilotIcon, SOON_BADGE_CLASS and
Clock3Icon and corresponding TooltipPopup using PICKER_TOOLTIP_SIDE and
PICKER_TOOLTIP_CLASS). This eliminates duplicate entries since
AVAILABLE_PROVIDER_OPTIONS now includes copilot and geminiCli and those
providers are rendered elsewhere.

In `@apps/web/src/components/Icons.tsx`:
- Around line 22-27: The CursorIcon component is forwarding a non-SVG prop
`monochrome` to the <svg> because Icon props aren't destructured; update
CursorIcon (and other icons that follow this pattern) to strip `monochrome` from
the props before spreading to the svg—e.g., destructure { className, monochrome,
...props } in the CursorIcon function signature or remove monochrome from the
object prior to {...props} so the svg doesn't receive a non-standard DOM
attribute (follow the same approach used by ClaudeAI).

In `@apps/web/src/environments/runtime/service.ts`:
- Around line 473-480: syncProjectUiFromStore builds logicalKey using
getClientSettings(), so UI keys go stale when grouping settings change; add a
settings change listener that calls syncProjectUiFromStore whenever
sidebarProjectGroupingMode or sidebarProjectGroupingOverrides change (or re-run
the same sync logic) to re-compute logical IDs and call
useUiStateStore.getState().syncProjects again; locate this near the runtime
service initialization where selectProjectsAcrossEnvironments,
getClientSettings, deriveLogicalProjectKeyFromSettings, and
syncProjectUiFromStore are defined and wire the listener to the settings store
or event emitter so changes immediately trigger re-sync.

In `@apps/web/src/hooks/useSettings.ts`:
- Around line 134-140: getClientSettings() currently returns the in-memory
snapshot (getClientSettingsSnapshot()) and can return DEFAULT_CLIENT_SETTINGS if
hydration hasn't started; change the API so there are two clear accessors: keep
getClientSettingsSnapshot() as the synchronous snapshot-only accessor, and make
getClientSettings() an async/hydrating accessor that ensures persisted settings
are loaded before returning (implement an ensureClientSettingsHydrated() or
call/await subscribeClientSettings() startup/hydration flow inside
getClientSettings()), and update callers: runtime init code should await
getClientSettings(), while synchronous callers that truly need a snapshot should
call getClientSettingsSnapshot(); update useSettings() subscribers to continue
starting hydration as before.

---

Outside diff comments:
In `@apps/server/src/provider/Layers/ClaudeProvider.ts`:
- Around line 688-719: The slash-command and SDK probe calls use the raw
claudeSettings.binaryPath which can be blank/whitespace and cause silent probe
failures; normalize the path first (e.g., const normalizedBinary =
claudeSettings.binaryPath.trim() || "claude") and use normalizedBinary when
calling resolveSlashCommands, runClaudeCommand (or any helper that probes the
CLI), and resolveSubscriptionType so all probes receive the same valid binary
path (update references around resolveSlashCommands, slashCommands,
authProbe/runClaudeCommand, and resolveSubscriptionType).

---

Nitpick comments:
In `@apps/server/src/keybindings.test.ts`:
- Around line 195-197: The test currently asserts
defaultsByCommand.get("modelPicker.toggle"), "modelPicker.jump.1", and
"modelPicker.jump.9" only for their key values; add assertions that verify the
corresponding `when` gating as well — e.g. assert that the entry for
"modelPicker.toggle" has when === "!terminalFocus" and that entries
"modelPicker.jump.1" through "modelPicker.jump.9" have when ===
"modelPickerOpen" — so that defaultsByCommand (and any lookups used in the test)
validate the `when` field and prevent regressions/conflicts with similar
bindings like "thread.jump.1".

In `@apps/web/src/components/ChatView.browser.tsx`:
- Around line 5574-5577: The test currently verifies the model picker UI via
document.querySelector(".model-picker-list") and
findComposerProviderModelPicker(), but doesn't assert the composer draft itself;
update the test inside the vi.waitFor callback to also locate the composer text
input/editor (e.g., the composer input element used elsewhere in tests) and
assert its value/textContent no longer contains "/model" and is cleared (or
equals the expected post-replacement string). Reference the existing helpers
used in this file (findComposerProviderModelPicker() and the composer
input/editor selector used by other tests) to locate the composer element and
add the direct assertion that the draft is cleared after selecting "/model".

In `@apps/web/src/components/CommandPalette.logic.ts`:
- Around line 154-173: Replace the Object.assign return with a single object
literal using spread syntax and conditional spreads (e.g., {...(leadingContent
&& { titleLeadingContent: leadingContent })}) to improve readability and
preserve a single inferred object type; keep the "kind: 'action' as const"
assertion, compute searchTerms using normal string literals for defaults (''
instead of ``) and use a normal string for the delimiter in
descriptionParts.join(' · '); ensure the timestamp still calls
formatRelativeTimeLabel(thread.latestUserMessageAt ?? thread.updatedAt ??
thread.createdAt) and the run function still calls await input.runThread(thread)
and include input.icon and other fields unchanged.

In `@apps/web/src/keybindings.test.ts`:
- Around line 410-422: Add an assertion to cover the inverse condition: when
modelPickerOpen is true, verify that thread jump hints remain hidden. In the
same test that calls shouldShowModelPickerJumpHints, also call
shouldShowThreadJumpHints (the thread hint visibility helper) with the same
event(DEFAULT_BINDINGS) and context { platform: "MacIntel", context: {
modelPickerOpen: true } } and assert it returns false so shared shortcuts do not
show thread hints while the model picker is active.

In `@apps/web/src/shortcutModifierState.test.ts`:
- Around line 23-146: Tests mutate module-level shortcut modifier state and lack
isolation; add suite-level hooks to reset state before and/or after each test by
calling clearShortcutModifierState inside a beforeEach (and optionally
afterEach) within the describe("shortcutModifierState") block so each it()
starts with a clean state; locate where tests reference
readShortcutModifierState, setShortcutModifierState,
syncShortcutModifierStateFromKeyboardEvent and insert the
clearShortcutModifierState calls in beforeEach/afterEach to make the suite
order-independent.

Verify each finding against the current code and only fix it if needed.

Nitpick comments:
In `@apps/web/src/components/chat/ProviderModelPicker.browser.tsx`:
- Around line 191-204: The helper buildOpenCodeProvider is only referenced from
skipped tests and is effectively unused; either delete the function to remove
dead code, or move it next to the first test you intend to restore (so it won’t
be flagged as unused) or re-enable at least one of the it.skip consumers so the
helper is actually used; update or remove the function declaration accordingly
(function name: buildOpenCodeProvider) and adjust imports/usages in the nearby
test blocks to match the chosen approach.
- Around line 384-424: The test is implicitly coupled to the production
MODEL_OPTIONS_BY_PROVIDER; instead of hard-coding "Claude Opus 4.7"/"4.5" update
the test to derive expectations from the same source the picker uses (e.g., call
getCustomModelOptionsByProvider(DEFAULT_UNIFIED_SETTINGS, ...) or read
modelOptionsByProvider.claudeAgent.map(m => m.name)) and then assert
ordering/favorite placement against that derived list (use
TEST_PROVIDERS/favorites to mark the favorite first and then compare the
remaining items in production order); locate references to getVisibleModelNames,
mountPicker, TEST_PROVIDERS, getCustomModelOptionsByProvider and change the
assertion block to build expectedNames from the shared model-options helper
rather than literal model strings.
- Around line 299-321: The test hard-codes a 4-element sidebar order including
"copilot", which is brittle; change the assertion to check ordering relations
instead of a fixed prefix: use getSidebarProviderOrder() to assert that
"favorites" is at index 0 and that the active provider "claudeAgent" appears
after "favorites" (e.g., indexOf("favorites") < indexOf("claudeAgent")), and
optionally assert that "claudeAgent" appears before any non-favorite sibling by
comparing its index against other providers returned by
getSidebarProviderOrder(); update the test inside the it("shows favorites
first...") block (where mountPicker and getSidebarProviderOrder are used) and
add a one-line comment pointing to the defaults source that injects "copilot"
(the default client/favorites settings) so future maintainers know why extra
providers may appear.