fix(config): don't warn when SEARCH_RERANK_URL points at the gateway - #1358

Claude finished @cbcoutinho's task in 2m 1s —— View job

Claude's Review

• Gather context (read diff, changed files, CLAUDE.md)
• Understand the request (code review)
• Review nextcloud_mcp_server/config.py changes
• Review tests/unit/test_config.py changes
• Post review feedback

Summary

This PR narrows the "gateway-namespaced model on a direct rerank endpoint" warning (added in #1356) so it stays quiet when SEARCH_RERANK_URL is explicitly pinned to the gateway's own /v1/rerank — a legitimate configuration. The fix and accompanying test are well-targeted and the doc comment clearly explains the why.

One correctness issue in the new logic, plus a test-coverage gap:

🟡 str.startswith is not a safe way to check "is this URL the gateway"

nextcloud_mcp_server/config.py:1741-1746:

_direct_rerank = bool(self.search_rerank_url) and not (
    self.embedding_gateway_url
    and self.search_rerank_url.startswith(
        self.embedding_gateway_url.rstrip("/")
    )
)

startswith does a raw character-prefix match, not a host/authority comparison, so a lookalike or subdomain host is indistinguishable from a genuine sub-path of the gateway. E.g. with EMBEDDING_GATEWAY_URL=https://gw.example:

• SEARCH_RERANK_URL=https://gw.example.attacker.com/v1/rerank → startswith("https://gw.example") is True, so this is (incorrectly) treated as "pointing at the gateway" and the warning is suppressed even though it's a completely different host.
• Less severe, but also wrong the other way: EMBEDDING_GATEWAY_URL=https://GW.example vs SEARCH_RERANK_URL=https://gw.example/v1/rerank (case differs) would not match even though hostnames are case-insensitive, so the warning would fire on a working setup — the exact false positive this PR is trying to eliminate.

A more robust check parses both URLs and compares scheme + netloc (optionally also verifying the path is a proper sub-path), e.g.:

from urllib.parse import urlsplit

def _same_origin(url: str, base: str) -> bool:
    u, b = urlsplit(url), urlsplit(base)
    return (u.scheme, u.netloc.lower()) == (b.scheme, b.netloc.lower())

Given the module already imports comment-heavy validation logic like this, this is a small, self-contained fix.

Fix this →

🟢 Test coverage gap

The new test (test_rerank_url_pointing_at_the_gateway_is_quiet) only covers the "quiet when pointed at the gateway" case. There's no test for the case this feature must still catch: EMBEDDING_GATEWAY_URL and a different, non-gateway SEARCH_RERANK_URL both set with a gateway-namespaced model — i.e. confirming the warning still fires when both env vars are present but genuinely point at different hosts. Right now that combination is only implicitly exercised by test_direct_url_warns_about_the_gateway_namespaced_default_model, which doesn't set EMBEDDING_GATEWAY_URL at all, so it wouldn't catch a regression in the and/startswith branch specifically.

Other notes

• Logic and control flow are otherwise sound: bool(self.search_rerank_url) and not (...) short-circuits correctly, so there's no risk of calling .startswith() on None.
• Comment style matches the file's existing convention of explaining why non-obvious validation decisions were made — consistent with CLAUDE.md.
• The third_party/astrolabe submodule bump appears incidental to this stacked branch and unrelated to the fix itself.
  · Branch

Quality Gate passed

Issues
0 New issues
0 Accepted issues

Measures
0 Security Hotspots
0.0% Coverage on New Code
0.0% Duplication on New Code

See analysis details on SonarQube Cloud