Repository navigation
Description
Problem
App connectors keep becoming unavailable across desktop chats, including after an app update and a full quit/reopen. Slack, Gmail, GitHub and Plugin Management calls can all fail together. I need a supported fix so I do not have to keep asking chats to retry or reconnect accounts.
Verified environment
- ChatGPT/Codex desktop for macOS: 26.1106.51428, build 26605.
- Bundled CLI: 0.167.0-alpha.4.1.
- macOS 15.3.1, Apple Silicon / arm64.
- Observed November 12–13, 2026; timestamps below are UTC.
Reproduction and result
- Open an existing desktop Codex chat with enabled connected apps.
- Ask it to perform a small read-only Slack operation, such as listing workspaces or finding a channel.
- Tool metadata may be present, but invocation fails with:
MCP tool `codex_apps/slack.slack_list_workspaces` is not available to the model
MCP tool `codex_apps/gmail.get_profile` is not available to the model
MCP tool `codex_apps/github.search_issues` is not available to the model
Plugin Management calls have failed the same way. On November 12, Slack workspace lookup, Slack channel search and Gmail profile lookup succeeded in one chat while another chat's Slack calls failed. On November 13, Slack/Gmail failed again after a full desktop restart, and a later Slack retry also failed.
Sanitized diagnostic evidence
Fresh desktop logs after reopening contain repeated failures of the shared app MCP service:
2026-11-13T04:57:54.933Z
MCP client for `codex_apps` failed to start: MCP startup failed: timed out awaiting tools/list after 30s
2026-11-13T04:58:31.126Z
MCP client for `codex_apps` failed to start: MCP startup failed: timed out awaiting tools/list after 30s
Previous-day logs also contain failed shared-service initialization/discovery requests with HTTP 403 and an HTML Unable to load site / Please try again later response, as well as HTTP transport failures. Example failure timestamps: 2026-11-12T09:49:24.245Z and 2026-11-12T10:11:12.515Z. IP addresses, request identifiers and HTML have been omitted. It is not established that those 403 responses and the later timeouts have the same root cause.
Checked for stale component versions
A prior, distinct browser incident involved a confirmed bundled-skill version mismatch. That was specifically rechecked for this incident:
| Component | App bundle | Staged marketplace | Installed cache |
|---|---|---|---|
| browser | 26.1106.51428 | 26.1106.51428 | 26.1106.51428 |
| chrome | 26.1106.51428 | 26.1106.51428 | 26.1106.51428 |
| unified-computer-use | 26.1106.51428 | 26.1106.51428 | 26.1106.51428 |
| codex-app-tools | 26.1106.51428 | 26.1106.51428 | 26.1106.51428 |
No Bundled CUA skill version mismatch or Plugin skill catalog reload failed events were found in the inspected November 12–13 desktop logs. Slack, Gmail and GitHub plugins are enabled. No current version mismatch or expired provider credentials have been demonstrated.
Recovery attempted and limits
- Updated the desktop app, then fully quit/reopened it; the failure persisted.
- Rediscovered currently exposed tools and retried bounded read-only calls; those failed again.
- Temporarily tested the documented
mcp_optional_startup_grace_ms = 0setting. It persisted across restart but did not fix the shared service's 30-second tools/list failure. Only that experimental setting was reverted. - Accounts were not disconnected or reauthorized without evidence of credential expiry.
- No security checks, permission gates or TLS validation were disabled or bypassed.
- The separately authenticated GitHub CLI can fetch repository information while the GitHub app connector is unavailable. This does not prove the app connector's transport or authenticated requests are healthy. One CLI search also encountered a TLS handshake timeout, so network reliability remains a possible contributor, not an established cause.
Expected behavior / request
Please investigate shared codex_apps startup and recovery on this desktop build. When an upstream failure is transient, reconnect and restore tools to affected existing chats without requiring repeated user prompts. Preserve and expose actionable diagnostics distinguishing HTTP rejection, transport failure, authentication, tools/list timeout and stale catalog state. Please provide a supported repair or fixed build; increasing a generic user-MCP timeout is not established as a valid fix for the host-owned app service.
Related symptom reports: #20167 and #29396. This report adds current macOS desktop evidence and a completed component-version comparison; a common root cause is not claimed.
Privacy
The user explicitly requested this report. It contains sanitized diagnostics only: no credentials, cookies, personal paths, email addresses, workspace/channel names or IDs, private URLs, raw transcripts, or business content.
Activity
github-actions commentedon Nov 13, 2026
Potential duplicates detected. Please review them and close your issue if it is a duplicate.
Powered by Codex Action
TimBonal commentedon Nov 18, 2026
November 18 update: this is still happening on macOS after the desktop update and a full restart. It's more frequent today and is interrupting daily work across several connectors.
codex_appsstartup timing out attools/listafter 30 seconds. A separately configured MCP server returned an authenticated response.MCP startup failed: tools/list pagination timed out after 30s. The timeout continued on subsequent reads.app/listrequest withforceRefetch: truewas still pending after 45 seconds.Can you trace the 10:16:02 UTC attempt and tell me where the live tool discovery request stalled and what caused the timeout? I need a fix for the shared connector service that restores all affected connectors and lets existing chats recover. Is there a service fix or desktop build that addresses this?
If a request or session ID would help you trace it, let me know where I can share it privately. Thanks.
TsangKwunHei commentedon Nov 20, 2026
The cached catalog count could make a recovery check look green while calls still fail. I’d use the same test Doc title in two fresh local chats on the same build, one with the built-in Drive connector selected and one using only the separately registered MCP server. Record which server actually answered and whether it returned the original Doc link. That gives support a timestamped live read to compare with
tools/list, and it avoids sharing the document itself.I work on Zortex. If you need a temporary way to do Drive lookups while the shared service is investigated, I can help try that second path using one sample Doc and a Drive-only connection on macOS. It would leave your built-in connectors as they are.