fix(desktop): latch a resume that paints nothing when the list row is stale

[austinpickett]

What does this PR do?

Fixes one root cause behind a session-resume flow that can leave the primary thread blank with no way out.

resumeSession's cold path computes the final transcript and, when it comes up empty, arms $resumeFailedSessionId so use-route-resume re-attempts the resume (bounded backoff, then an explicit error + manual Retry). That latch armed only when the cached sessions-list row reported message_count > 0:

if (sessionShouldHaveTranscript(stored) && preferredMessages.length === 0) {

stored comes from $sessions / resolveStoredSession — it is a cache of backend truth, and it lags exactly the two flows that report a vanished thread:

• Wake/reconnect ([Bug]: macOS Desktop — after sleep/reopen, previous prompts and right-hand chat timeline disappear #82806): after sleep the local gateway respawns and re-mints runtime ids. The sessions list can still carry the session at message_count 0 while the backend's stored transcript is intact.
• Compression / stale tip ([Bug]: Desktop can place newest user bubble after completed reply following context compression #83154): a compression tip can show 0 rows while the stored transcript is intact.

In both cases a resume that paints nothing falls past the latch, then sets activeSessionId to the resumed runtime and publishes an empty message set. The thread is not "loading" (routedSessionIsLoading returns false once activeSessionId is set and the row has no known history), so the UI shows an empty chat with no error and no retry — history looks lost even though the session is alive.

The resume RPC is the authoritative source and always reports the stored transcript size — tui_gateway/methods_session.py:814 fills message_count from the stored size even when messages_omitted is set for Desktop's cold resume. This PR treats it (and a non-empty REST page) as the other rungs of the same ladder:

const saidToHaveTranscript =
  sessionShouldHaveTranscript(stored) ||
  (resumed.message_count || 0) > 0 ||
  Boolean(prefetchedResult?.messages.length)

if (saidToHaveTranscript && preferredMessages.length === 0) {

A resume that paints nothing now arms the retry latch whenever any source says the session has a transcript, so the window self-heals through use-route-resume's bounded retry and surfaces an explicit error + manual Retry once exhausted, instead of going silently blank. The provisional cached-tail rollback inside the branch is unchanged.

Common root cause, one PR

Both reports are the same flow (use-session-actions/index.ts) where a resume that ends empty leaves the user on a blank thread; the single widened precondition covers both. No second code path was needed.

Parts of each report NOT covered here (deliberately)

• [Bug]: macOS Desktop — after sleep/reopen, previous prompts and right-hand chat timeline disappear #82806 — warm-path empty-REST wipe: already fixed on current main. The guard PR fix(desktop): stop empty REST transcript refresh from wiping a warm resume #82843 proposed —
  if (persisted && persistedMatchesActivatedSession && (persisted.messages.length || !activatedMessages.length))
  — is present at index.ts:1613. fix(desktop): stop empty REST transcript refresh from wiping a warm resume #82843 is closed unmerged, but its change landed another way. This PR does not re-touch it.
• [Bug]: Desktop can place newest user bubble after completed reply following context compression #83154 — "silently startFreshSessionDraft(true)" (:2373): the fresh-draft fallback fires only on goneSessionVerdict → 'draft' (RPC + REST both 404, id verifiably dead). Dropping to a fresh draft without a toast there is the documented contract (apps/desktop/AGENTS.md: "the window drops to a fresh draft without toasting or looping"), and it already carries the stashed composer draft over via announceGoneSessionDraft. Changing it would contradict that contract; the reported defect is the unlatched blank with a live runtime, which this PR fixes.
• [Bug]: Desktop can place newest user bubble after completed reply following context compression #83154 — "paint the prefetched REST transcript before awaiting session.resume": already done at index.ts:1950 ("Paint the persisted transcript as soon as REST returns …", fix(desktop): paint prefetched history during cold resume #90130). The isCurrentResume() bails at 1979/2234/2251/2278/2343 are stale-navigation guards — returning without painting is correct there, since a newer navigation owns the view.

Behavior

• Before: a cold resume that paints nothing while the sessions-list row is stale leaves an empty thread bound to a live runtime — no retry, no error.
• After: the same resume arms the retry latch; use-route-resume re-attempts with backoff and, if it stays stranded, the chat shows the explicit error + Retry state.

Acceptance coverage

Two regression tests in apps/desktop/src/app/session/hooks/use-session-actions.test.tsx (both fail on pre-fix source, pass with it):

• arms the failure latch when a stale list row hides history the resume RPC still reports — [Bug]: Desktop can place newest user bubble after completed reply following context compression #83154: list row message_count: 0, session.resume returns message_count: 12 with messages_omitted and an empty REST page → latch arms, activeSessionId stays null.
• arms the failure latch when a wake warm-resume bail falls to an empty cold resume — [Bug]: macOS Desktop — after sleep/reopen, previous prompts and right-hand chat timeline disappear #82806: a warm cache is bound, session.activate 404s (respawned backend), the mapping is dropped, the cold resume paints nothing, list row stale → latch arms.

Validation (real output)

$ cd apps/desktop && npm run typecheck
> tsc -p . --noEmit && tsc -p tsconfig.electron.json --noEmit && tsc -p tsconfig.e2e.json --noEmit && ...
(exit 0, no diagnostics)

$ npx vitest run src/app/session/hooks/use-session-actions.test.tsx
 Test Files  1 passed (1)
      Tests  119 passed (119)

$ npx vitest run src/app/session/hooks/use-route-resume.test.tsx \
    src/app/chat/thread-loading.test.ts src/app/session/hooks/use-session-actions/utils.test.ts
 Test Files  3 passed (3)
      Tests  153 passed (153)

$ npx eslint src/app/session/hooks/use-session-actions/index.ts \
    src/app/session/hooks/use-session-actions.test.tsx
(exit 0, clean)

Proving the tests are the fix (source file reverted, tests kept):

$ git stash push -- apps/desktop/src/app/session/hooks/use-session-actions/index.ts
$ npx vitest run src/app/session/hooks/use-session-actions.test.tsx -t "arms the failure latch when"
 ❯ ... > arms the failure latch when a stale list row hides history the resume RPC still reports
   AssertionError: expected null to be 'stored-1'
 ❯ ... > arms the failure latch when a wake warm-resume bail falls to an empty cold resume
   AssertionError: expected null to be 'stored-1'
 Test Files  1 failed (1)
      Tests  2 failed | 1 passed | 116 skipped (119)

No macOS hardware was available to physically exercise sleep/wake; the tests drive the exact code path (warm bind → activate 404 → cold resume painting nothing) and the assertions are on the observable latch/active-session state, not on internal calls.

Reused work

• Based on fix(desktop): stop empty REST transcript refresh from wiping a warm resume #82843 by @chelsealong — its diagnosis of the empty-REST-page race in the warm/activate path (already merged into main as the persisted.messages.length || !activatedMessages.length guard) framed where the residual primary-thread gap is.
• fix(desktop): keep page-omitted prompts before tool activity and replies #119496 (open) explicitly does not claim [Bug]: Desktop can place newest user bubble after completed reply following context compression #83154.

Fixes #83154

hermes-agent #82806 is not closed by this PR. Its original mechanism — the warm-path empty-REST
wipe — is already fixed on main (PR #82843's persisted.messages.length || !activatedMessages.length
guard, credited above); this PR does not re-touch it. What it adds is the failure latch for the
sleep/wake warm-bail flow, i.e. the residual that left the thread blank with no retry. Whether that
fully resolves #82806 is a maintainer call — leaving it open until someone confirms on a build with
this change.

🤖 AI-assisted (Hermes Agent); verification above is real command output.

[alt-glitch]

This was generated by AI during triage.

Duplicate of #83802: its open broader resume repair already uses response message_count to arm the empty-transcript retry latch when sidebar state is stale or absent.