Windows desktop: recurring missing integration entries, marketplace path failures and intermittent prompt submission failure (26.908)

Version and platform

Windows desktop AppX package OpenAI.Codex 26.908.4834.0, installed package Status=Ok. Bundled plugin/runtime application version is 26.908.40834. Evidence collected September 13, 2026; timestamps below are UTC.

Problem

The user repeatedly loses integration entries in Settings and intermittently cannot submit prompts without several desktop restarts. The supplied screenshot shows the Integrations section containing only Plugins. The exact missing entry names are not yet independently verified. Repeated troubleshooting consumes subscription usage and disrupts work.

Observed evidence

• September 12 23:35:25: runtime marketplace written with 8 plugins; subsequent bundled_plugin_status_unknown errors report Windows os error 3 for codex-app-tools, browser, unified-computer-use and chrome. A sites install reports that the plugin was not found in openai-bundled.
• September 12 05:48:04: plugin/installed request timeout at 30,000 ms, pendingCount=8. This startup also records node_repl/cua_repl path-not-found failures.
• September 13 00:00:06, 00:00:17 and 00:00:50: bundled_plugins_reconcile_skipped_features_unavailable on window focus. The current startup has the same event at 00:01:02. These indicate feature availability was absent at those moments, not a proven service outage or entitlement loss.
• Across September 12-13 main-process t0 logs: 47 unknown-conversation messages, 13 Conversation state not found, 4 Item not found in turn state. These counts are occurrences, not distinct failed user actions.
• Current startup reports node_repl, codex_apps and notebooklm-mcp ready. Config parses, current MCP executable paths exist, and connector tools are exposed to the active session. Missing settings entries therefore must not automatically be described as all integrations being uninstalled.
• Native Computer Use inspection currently fails with native pipe unavailable / os error 2. No computer-use pipe was enumerated. The available unified CUA surface is browser-only, preventing live native UI validation in this investigation.
• Remote-control websocket reports HTTP 409, Remote app server already online. Multiple app servers belong to Desktop, a usage tray and a Chrome extension host; none was proved orphaned. Local connection state remains connected. Causality with these symptoms is unproven.

Important diagnostic limitation

The current log also contains latestTurnStatus=completed with markedStreaming=true, similar to issue #44102. Read-only inspection of the installed main bundle shows this logging site writes markedStreaming as the literal constant true. That field alone does NOT prove a stuck composer. Please correlate actual composer state and turn/start dispatch with lifecycle events.

Reproduction and expected behavior

User-reported sequence: open Settings and notice missing integration entries; in repeated separate occurrences, type a prompt and attempt Send, which does not submit until repeated app restarts. Exact deterministic reproduction and affected menu labels remain unverified.

Expected: installed/available integrations remain discoverable or show an explicit loading/error/retry state. After turn completion, submitting another prompt should work or show an actionable error without repeated restart cycles.

Requested investigation

Please investigate feature-availability initialization/recovery, atomic publication and reconciliation of the bundled marketplace across app/runtime versions, and conversation hydration/event ordering around composer readiness. Please clarify the semantics of markedStreaming logging. Treat these as hypotheses requiring engineering validation, not an asserted single root cause.

No cache deletion, profile reset, auth changes, feature-gate override or binary patch was performed. Current executable existence and MCP readiness are not a UI acceptance test. Related reports: https://github.com/openai/codex/issues/44102 and https://github.com/openai/codex/issues/43366.

Prior repair / recurrence clarification

The user confirms a prior repair on September 12. Its local report records five stale runtime path replacements, matching bundled executable hashes, and successful MCP initialize/tools-list. Browser use subsequently worked; native Computer Use pipe still failed in that prior investigation. Current runtime executable paths exist and node_repl is ready, so repeating the same path replacement is not presently justified. Please distinguish a prior browser-runtime repair from an unresolved native UI issue and the recurring settings/composer symptoms.

Follow-up: Computer Use specifically remains missing after restarts

The user clarified that Computer Use is the missing integration. Both computer-use and unified-computer-use remain enabled in config, while the generated unified MCP manifest contains CUA_REPL_ENABLED_SURFACES=browser only. Native Computer Use pipe enumeration is empty and sky.list_apps fails with os error 2. This is a current failure, not merely historical startup evidence.

Read-only bundle inspection shows native service startup is conditional on computerUse and computerUseNodeRepl, and contains a Windows-specific process environment switch. A bounded fresh-process launch experiment with CODEX_ELECTRON_ENABLE_WINDOWS_COMPUTER_USE=1 did not restore the pipe or computer surface. Propagation through packaged app activation was not independently verified, so this is not conclusive evidence of override behavior. A normal launch was restored afterwards; no persistent environment setting, binary modification or entitlement change was made. The app still logs skipped_features_unavailable. Please investigate feature availability delivery and native service lifecycle rather than recommend repeated blind cache resets.

Start with the Windows desktop feature-availability initialization and recovery, bundled marketplace reconciliation, native Computer Use service lifecycle, and conversation/composer turn-start events described in the report; no source files or tests are named. Use the supplied logs and a reproducible restart/prompt sequence to identify the failure, then verify that Computer Use remains discoverable or reports an actionable state and prompt submission works without repeated restarts.