Desktop app crash-loops at startup: sqlite state runtime backfill exceeds fixed 30s timeout on large session stores; rebuilds discard pins and custom thread names

Summary

After updating to ChatGPT Desktop 26.825.41651 (bundled codex-cli 0.151.0-alpha.7.1, macOS arm64), the app crash-loops at startup on a machine with a large session store (~17,000 rollout files / ~16 GB under $CODEX_HOME/sessions). Root cause chain:

1. The new sqlite state runtime (state_5.sqlite, includes thread_spawn_edges) performs a full backfill over all rollout files on first initialization.
2. On this store the backfill takes several minutes, but startup waits only 30s:
   Error: failed to initialize sqlite state runtime under $CODEX_HOME: timed out waiting for state db backfill at $CODEX_HOME after 30s (status: running) → exit code 1.
3. Every relaunch kills the in-progress backfill at 30s and restarts it, so the app never comes up (crash loop).

Before the app became fully unable to start, active chats failed with "thread not found"; logs show:
state db list_threads failed: error returned from database: (code: 1) no such table: thread_spawn_edges

Workaround that recovered the app

Running the bundled binary manually (no 30s watchdog) let the backfill complete once (~4 min), after which the app started normally:

"/Applications/ChatGPT.app/Contents/Resources/codex" app-server

Data-loss aspect (worse than the crash)

Each state-db rebuild regenerates thread rows from rollout files and discards user-curated data that is not stored in rollouts:

• is_pinned (pinned chats)
• custom thread names (set via thread/name/set — these land only in the state DB, not in the rollout, so a later rebuild erases them)
• previously generated/curated titles (regenerated from the first user message)

Additionally, any externally detected modification of state_5.sqlite appears to trigger an immediate full rebuild, compounding the loss.

Related observation

codex doctor shows the rollout scan is capped at 9,999 files (rollout DB scan cap reached true); with >10k active files, stale-row and archive-mismatch checks are silently skipped.

Suggestions

• Raise or remove the 30s startup timeout, or make the initial backfill asynchronous/resumable so the UI can start while it runs.
• Persist pins and custom names durably (e.g. in rollout metadata) or preserve these columns across backfills.
• Surface the scan cap instead of skipping checks silently.

Environment

• ChatGPT Desktop 26.825.41651, Codex Framework 151.0.7922.174
• bundled codex-cli 0.151.0-alpha.7.1
• macOS 26.5.2 (arm64)
• ~17k active rollout files, ~16 GB; state DB ~360 MB after backfill

Start with the bundled codex app-server entry point and state_5.sqlite initialization against a large session store, focusing on the 30-second backfill timeout and rebuild behavior. Also inspect the codex doctor rollout scan cap. Done means startup completes reliably without a crash loop or destructive rebuild of pins and custom thread names.