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

Inline comments:
In `@lib/codex-manager/settings-hub.ts`:
- Around line 329-333: The codexCliSessionSupervisor toggle is currently listed
in BACKEND_TOGGLE_OPTIONS but not rendered by BACKEND_CATEGORY_OPTIONS, so
backendSettingsSnapshot() / buildBackendConfigPatch() will include it in backend
resets and silently clear it; fix by removing "codexCliSessionSupervisor" from
BACKEND_TOGGLE_OPTIONS (or alternatively add it to BACKEND_CATEGORY_OPTIONS and
ensure the backend UI renders it) so that the key is only in the shared backend
option set if the backend UI can actually display and preserve it—update the
array where BACKEND_TOGGLE_OPTIONS is defined and adjust any UI rendering logic
if you choose to move it into BACKEND_CATEGORY_OPTIONS.

In `@package.json`:
- Line 54: Update the smoke npm script "test:session-supervisor:smoke" to
include the storage regressions by adding test/storage.test.ts to the list so
the smoke run exercises the changed storage logic (the path-equivalence and
transaction/export safety code in lib/storage.ts around the path-drift and
windows-handling sections). Modify the package.json target that currently lists
tests like test/codex-supervisor.test.ts to also run test/storage.test.ts, then
commit the package.json change so CI's smoke job covers the storage changes.

In `@scripts/codex-supervisor.js`:
- Around line 551-585: Add a short in-source comment above the pendingRefresh
handler explaining that the read-then-write window in pendingRefresh (which
calls readSupervisorLockPayload and then writeSupervisorLockPayload using
lockPath, ownerId, acquiredAt, ttlMs) is not a critical TOCTOU because lock
acquisition uses exclusive create ('wx') so a concurrent supervisor cannot
silently take the same filename, and the code explicitly checks payload.ownerId
and calls stopWithError on mismatch; mention these two protections
(exclusive-create on acquisition and ownerId mismatch check) so future readers
understand the mitigation.
- Around line 34-35: The module-level maps snapshotProbeCache and
sessionRolloutPathById persist across process reloads and need a public reset
entrypoint; add and export a production-safe cache-reset function (e.g., export
function resetSupervisorCaches() or two exports resetSnapshotProbeCache() and
resetSessionRolloutPathCache()) that calls snapshotProbeCache.clear() and
sessionRolloutPathById.clear(), and document/ensure callers (such as the
supervisor runtime reload path) invoke it when reloading; keep the existing
__testOnly.clearAllProbeSnapshotCache and
__testOnly.clearSessionBindingPathCache behavior but provide this new public API
for embedded/long-running use.
- Around line 114-118: The parseNumberEnv helper currently signature
parseNumberEnv(name, fallback, min = 0) ignores an intended max bound; update
the signature to accept a fourth parameter (e.g., max = Infinity) and clamp the
parsed value between min and max (use Math.trunc then Math.max(min,
Math.min(max, value))). Modify parseNumberEnv accordingly and ensure callers
that pass a 4th arg (e.g., places that call parseNumberEnv(..., 0, 100) for
CODEX_AUTH_PREEMPTIVE_QUOTA_5H_REMAINING_PCT) behave correctly without further
changes.

In `@scripts/codex.js`:
- Around line 541-554: The branch after runCodexSupervisorIfEnabled that returns
supervisedExitCode without triggering autoSyncManagerActiveSelectionIfEnabled is
untested for the case supervisedExitCode === 0 and supervisorDidForward ===
false; add a deterministic vitest regression that simulates
runCodexSupervisorIfEnabled returning 0 while supervisorDidForward is false and
rawArgs is an interactive command (use isSupervisorInteractiveCommand to pick
args), then assert autoSyncManagerActiveSelectionIfEnabled was called once
(mock/spied) and the wrapper returns 0; reference runCodexSupervisorIfEnabled,
supervisedExitCode, supervisorDidForward, isSupervisorInteractiveCommand and
autoSyncManagerActiveSelectionIfEnabled in the test and avoid flaky timing by
stubbing dependencies instead of real concurrency or filesystem operations.

In `@test/codex-bin-wrapper.test.ts`:
- Around line 68-125: The fixture always hard-codes codexCliSessionSupervisor:
true in writeSupervisorRuntimeFixture which masks runtime env precedence; change
writeSupervisorRuntimeFixture to accept a boolean (e.g., enableSupervisor =
false) and write codexCliSessionSupervisor accordingly, then update the wrapper
test cases that previously relied on the fixture to call
writeSupervisorRuntimeFixture(fixtureRoot, false) and enable the supervisor only
via CODEX_AUTH_CLI_SESSION_SUPERVISOR=1 in the specific test that needs it so
tests exercise the env-path behavior rather than the hard-coded config.

In `@test/codex-supervisor.test.ts`:
- Around line 1326-1328: The busy-wait loop polling setImmediate to wait for
startedAccounts.size to reach 3 is flaky; replace it with a deterministic
deferred/promise that resolves when startedAccounts.size === 3 (and include a
short timeout fallback). Concretely, create a Promise (deferred) in the test and
have whatever code increments the startedAccounts Set resolve that promise as
soon as it observes startedAccounts.size === 3; then await that promise instead
of the for-loop. Ensure the promise also rejects or resolves after a reasonable
timeout to avoid hanging tests.
- Around line 9-26: The env cleanup list envKeys in
test/codex-supervisor.test.ts is missing the environment variable
CODEX_AUTH_CLI_SESSION_LOCK_TTL_MS which the tests set; add
"CODEX_AUTH_CLI_SESSION_LOCK_TTL_MS" to the envKeys const so the test harness
clears it between tests (update the envKeys array near the existing CODEX_AUTH_*
entries and re-run the test suite to ensure no state leakage).

In `@test/plugin-config.test.ts`:
- Around line 991-1008: The test suite misses the explicit environment opt-out
case: add a deterministic vitest test that sets
process.env.CODEX_AUTH_CLI_SESSION_SUPERVISOR = '0', calls
getCodexCliSessionSupervisor({ codexCliSessionSupervisor: true }) and asserts
the result is false, then cleans up the env var; this verifies the env override
in getCodexCliSessionSupervisor correctly wins over a true config value and
prevents regressions in the opt-out branch.

In `@test/storage.test.ts`:
- Around line 1008-1044: Update the test that simulates Windows platform to also
exercise separator normalization: when running inside
withAccountStorageTransaction (using setStoragePathDirect, exportAccounts,
saveAccounts, testStoragePath, exportPath) set the storage path to a
Windows-style path with backslashes (e.g., transform testStoragePath to use "\"
separators) in addition to uppercasing, so the export path comparison covers
both case-insensitivity and separator-insensitivity; ensure the modified test
still toggles process.platform to "win32" and restores it in finally, and assert
the exported accountId as before.

---

Outside diff comments:
In `@lib/storage.ts`:
- Around line 2146-2155: The persist transaction freezes account file but calls
saveFlaggedAccountsUnlocked which resolves its path at call time; capture the
flagged-path up-front (call getFlaggedAccountsPath() inside persist into e.g.
flaggedPathCaptured) and pass that path into saveFlaggedAccountsUnlocked (and
any helpers that resolve paths) so both accounts and flagged writes target the
same storage root; update cloneAccountStorageForPersistence/state.snapshot usage
unchanged but ensure saveAccountsUnlocked and saveFlaggedAccountsUnlocked are
called with the captured paths, add a regression vitest near the existing
storage tests that simulates handler storage-root rotation mid-transaction to
assert rollback still restores both files, and update any affected queue logic
to explicitly retry/handle EBUSY and 429 per existing queue abstractions as
noted in coding guidelines.