[BUG] Desktop: self-managed GitLab MRs bind but CI checks fail with "Unknown GitHub host", while the no-CI-polling rule still applies

Preflight Checklist

• [x] I have searched existing issues and this hasn't been reported yet
• [x] This is a single bug report (please file separate reports for different bugs)
• [x] I am using the latest version of Claude Code

What's Wrong?

In the desktop app's Code tab, a merge request on a self-managed GitLab instance binds to the session: the app log shows [GitHubPrManager] Bound gitlab #<iid> to session …, and get_status returns the MR's number, state and branches. Its CI checks are never read:

"checks": { "available": false, "error": "Unknown GitHub host \"gitlab.example.com\"." }

So the PR monitor never sends a ci-monitor-event, and Auto-fix does nothing for it. Meanwhile every session's system prompt says to "never schedule or poll CI checks yourself (CronCreate, ScheduleWakeup, /loop, Monitor, gh polling)", so Claude declines to watch the pipeline even when asked, and nothing else watches it. All that's left is re-running a one-off status check by hand.

glab is installed and authenticated for the host (glab api --hostname gitlab.example.com projects/<id>/pipelines/<id> works), and the app already reads the MR's number, state and branches through glab. Only the checks lookup appears to go through GitHub-only code, like the gh lookup in #96503.

What Should Happen?

Either of these would fix it:

1. Read CI for GitLab bindings through glab (the MR's pipelines, their jobs, and child pipelines started by trigger jobs), so checks.available is true and Auto-fix / ci-monitor-event work on gitlab.com and self-managed hosts alike.
2. Until then, apply the no-polling rule only when the bound PR's checks are available (the fix #98774 proposes for Azure DevOps), so Claude can poll when the user asks.

Also useful: watching a standalone pipeline that has no MR, e.g. one started with glab ci run. Our team runs scoped E2E suites that way.

Error Messages/Logs

# ~/Library/Logs/Claude/main.log
[info] [GitHubPrManager] Bound gitlab #<iid> to session local_<id>

# mcp__ccd_pr__get_status
"pr": { "number": <iid>, "url": "https://gitlab.example.com/<group>/<project>/-/merge_requests/<iid>", "host": "gitlab.example.com", "state": "open" },
"checks": { "available": false, "error": "Unknown GitHub host \"gitlab.example.com\"." },
"monitor": { "auto_fix": false, "address_comments": false }

Steps to Reproduce

1. Authenticate glab for a self-managed GitLab host: glab auth login --hostname gitlab.example.com.
2. Open a Code-tab session on a repo whose remote is on that host.
3. Push a branch that has an open MR (or call bind_pr with the MR URL). The MR appears in the PR bar.
4. Ask Claude to call get_status: checks.available is false with "Unknown GitHub host".
5. Ask Claude to tell you when the MR pipeline finishes: it declines to poll and quotes the system-prompt rule.

Claude Model

Opus

Is this a regression?

I don't know

Last Working Version

_No response_

Claude Code Version

2.1.289 (bundled with Claude desktop 2.26454.0)

Platform

Anthropic API

Operating System

macOS

Terminal/Shell

Other

Additional Information

• Related: #98774 (the same rule on Azure DevOps, where the monitor can't bind at all) and #96503 (a GitHub-only code path on a gitlab.com binding).
• /loop works for this in the CLI, so the gap is desktop-only.
• Workaround: a function-hooks plugin that polls glab api with $.clock.every and reports through $.ui.status / $.ui.toast. It can tell the person a pipeline finished, but it can't wake the session, so Auto-fix still isn't possible.