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

Inline comments:
In `@src/renderer/hooks/agent/useAgentExecution.ts`:
- Around line 40-41: The hard timeout constant BATCH_HARD_TIMEOUT_MS should be
derived from the configured inactivity timeout (not a fixed 60 minutes) so set
it to the inactivity config value (e.g., use the same value used for
BATCH_WATCHDOG_CHECK_MS/inactivity settings) and update any watchdog logic that
uses BATCH_HARD_TIMEOUT_MS; in the code paths inside useAgentExecution where
kill() is invoked (and in the related watchdog/cleanup logic around lines
referenced), await the kill() promise and do not swallow errors — catch only
expected recoverable errors explicitly and rethrow or log unexpected ones so
they bubble to Sentry, and ensure you only resolve/allow the next task to start
after kill() has completed (or an explicit handled timeout) so the stalled
process is truly terminated before proceeding.

In `@src/renderer/hooks/batch/useBatchProcessor.ts`:
- Line 2020: The current call to Promise.allSettled(...) around
batchProcessIds.map(id => window.maestro.process.kill(id)) swallows failures;
change useBatchProcessor.ts so after awaiting Promise.allSettled on
window.maestro.process.kill for each id you inspect the returned results array,
filter entries with status === 'rejected', and if any exist either log/report
them (or rethrow a combined Error) before clearing/resetting batch state; only
proceed to reset state when all kill promises succeeded, and reference the
batchProcessIds variable and the window.maestro.process.kill calls in your
handling logic.
- Around line 1087-1088: The try/catch inside useBatchProcessor that currently
does "catch { /* Ignore polling errors */ }" silently swallows all errors;
change it to catch the error object (e.g., catch (err)) and explicitly handle
known transient/read-race cases (check err.code for ENOENT, EBUSY, EAGAIN or use
a helper like isTransientReadError), while for any other unexpected errors call
the global error reporter (Sentry.captureException or processLogger.error with
context including function name useBatchProcessor and the file/agent id) or
rethrow so it bubbles up; ensure you reference the same catch block
(polling/read failure handling) and add minimal tests or logging to confirm
non-transient failures are reported.
- Around line 1022-1034: The progress polling cleanup (stopProgressPolling,
progressPollTimeout) is scoped inside the task loop so the hook-level unmount
cleanup can't cancel an outstanding timeout; register the polling teardown with
the hook lifecycle by exposing or adding stopProgressPolling to the hook's
cleanup handlers (e.g., push stopProgressPolling or a wrapper that clears
progressPollTimeout into the hook-level cleanup registry used by
useBatchProcessor) and ensure the hook's useEffect cleanup calls it; also remove
the registered handler from the registry inside both normal and error paths
after you call stopProgressPolling() so no dangling references remain and
onUpdateSession cannot be invoked after unmount.
- Around line 1947-1950: Replace the console.assert in killBatchRun with a
thrown error to enforce the invariant: check that sessionId does not include
'-batch-' and if it does, throw a descriptive Error (e.g., include the sessionId
and the message '[BatchProcessor:killBatchRun] sessionId must not contain
"-batch-"') so the failure propagates to error boundaries/Sentry; update any
tests or callers of killBatchRun to expect an exception instead of a silent
assertion failure.

In `@src/renderer/stores/settingsStore.ts`:
- Around line 1056-1060: The stored setting is read directly into state
elsewhere and bypasses the setter's 1..600 clamp, so ensure any
load/initialization that reads
window.maestro.settings.get('autoRunInactivityTimeoutMin') applies the same
clamp before calling set; reuse the exact clamp logic used in
setAutoRunInactivityTimeoutMin (const clamped = Math.max(1, Math.min(600,
Math.round(value)))) and write that clamped value to state and to
window.maestro.settings where appropriate (i.e., wherever the stored value is
applied during initialization or reload), so stale/manual/external values cannot
set an out-of-range autoRunInactivityTimeoutMin.

In `@src/web/mobile/App.tsx`:
- Around line 2170-2172: The goToAutoRun action currently calls
handleOpenAutoRunPanel (opening the full-screen Auto Run panel) which is
inconsistent with goToFiles/goToHistory that open right-drawer tabs; change
goToAutoRun to call handleOpenRightDrawer with the Auto Run tab key (e.g.
replace handleOpenAutoRunPanel() with handleOpenRightDrawer('autoRun')) so the
shortcut opens the same right-drawer Auto Run tab as the command-palette
navigation.

---

Outside diff comments:
In `@src/main/ipc/handlers/autorun.ts`:
- Around line 116-151: The scanner currently follows symlinked entries (via
resolveLocalEntryType/scanDirectory) which allows .md files outside the Auto Run
base to be included; update scanDirectory to skip symlinks or enforce canonical
containment by calling fs.realpath on each entry and comparing the resolved path
against the Auto Run base realpath (use visitedRealPaths and the same realpath
check wherever entries are processed — e.g., in the resolvedEntries mapping and
the folder recursion block inside scanDirectory) and only include entries whose
realpath startsWith the Auto Run base realpath; ensure you reference
resolveLocalEntryType, scanDirectory, visitedRealPaths, and nodes when
implementing this containment check so symlink targets outside the Auto Run
folder are excluded before adding to nodes or recursing.

In `@src/renderer/hooks/batch/useAutoRunDocumentLoader.ts`:
- Around line 186-252: refreshAutoRunData can throw and those throws become
detached rejections in the remote poll loop (runRemotePoll) and the local
onFileChanged callback; wrap each await refreshAutoRunData() call (inside
runRemotePoll and inside the onFileChanged handler) in a try/catch that captures
and logs the error with context (folderPath, sshRemoteId and selectedFile when
available) and does not rethrow so polling/watching continues; ensure the catch
reports the error to your app/error service (e.g., existing logger/Sentry
wrapper) and then lets the loop/timer or watcher continue as before.

---

Nitpick comments:
In `@src/main/agents/capabilities.ts`:
- Line 263: The renderer tests mock for hasCapabilityCached in
useInlineWizard.test.ts currently lists wizard-capable agents ['claude-code',
'codex', 'opencode'] and thus won’t exercise factory-droid; update that mock to
include 'factory-droid' (i.e.,
['claude-code','codex','opencode','factory-droid']) so the test covers the newly
enabled supportsWizard capability, and run the test to ensure no other
assertions or snapshots need adjustment.

In `@src/main/web-server/web-server-factory.ts`:
- Around line 822-904: The duplicated construction of the WebSettings object in
server.setGetSettingsCallback and inside the success branch of
server.setSetSettingCallback should be extracted into a single helper (e.g.,
buildWebSettings) that reads from settingsStore and returns the WebSettings
shape; replace both inline object literals with const settings =
buildWebSettings() and use settings for broadcasting and returning so future
fields like shortcuts are maintained in one place.

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

Inline comments:
In `@src/web/hooks/useMobileKeyboardHandler.ts`:
- Around line 149-174: The global shortcut handler fires configurable single-key
shortcuts while the user types; add an editable-target guard in
useMobileKeyboardHandler before iterating actions so we skip shortcuts when
focus is inside an editable element. Implement an isEditableElement check (using
document.activeElement or event.target) that returns true for input/textarea or
elements with contentEditable=true but false for xterm (reuse isXtermElement),
and if it returns true && activeSession?.inputMode !== 'terminal' then return
early; keep matchesShortcut, actions and shortcuts logic unchanged otherwise.
- Around line 67-109: The matchesShortcut function can fire for empty or
modifier-only overrides; update matchesShortcut to guard early: if sc.keys is
empty or the derived mainKey is undefined or is one of the modifier names
('meta','ctrl','command','shift','alt'), return false before any key comparisons
or the shiftNumberMap lookup. Specifically, in matchesShortcut, validate sc.keys
and mainKey (the last item in keys) and only proceed to compare meta/shift/alt,
arrow/symbol checks, and the shiftNumberMap lookup when mainKey exists and is
not a pure modifier to avoid matching ordinary typing when shortcuts are
cleared.

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

Nitpick comments:
In `@src/__tests__/web/hooks/useMobileKeyboardHandler.test.ts`:
- Around line 274-293: Update the test case "does not match an empty or
modifier-only shortcut definition" to also cover the modifier-only shortcut:
create a second shortcut entry (e.g., shortcuts.modOnly = { id: 'modOnly',
label: 'Modifier Only', keys: ['Meta'] }) and render useMobileKeyboardHandler
with that shortcut and the same spy action; then dispatch a non-modifier
KeyboardEvent (like key: 'n') and assert that the mod-only action (the spy) is
not called. Ensure the test still includes the existing empty-keys case and
references the same newInstance spy and useMobileKeyboardHandler invocation.

In `@src/web/hooks/useMobileKeyboardHandler.ts`:
- Around line 35-39: The MobileKeyboardSession type is broader than the hook
needs; update the type to only include the field the hook reads (inputMode) and
remove aiTabs and activeTabId from MobileKeyboardSession, then remove the
now-unused AITabData import; locate the MobileKeyboardSession declaration and
the useMobileKeyboardHandler hook to change the exported type to: { inputMode?:
string } and delete any imports or references to AITabData that become unused.