fix(health): Sign-in CTA for call-time-only OAuth servers (MCP-2084)#634
Merged
Conversation
…P-2084) Some remote MCP endpoints (e.g. Google's sqladmin MCP) allow anonymous `initialize` + `tools/list` but enforce OAuth only at `tools/call`. The proxy connected, listed 15 tools, and the health classifier reported the server as fully healthy/Ready with no Sign-in affordance — every actual tool call failed at runtime with "authorization required", which the operator only discovered by failing a call. Root cause: the health calculator's OAuth/login branch is gated entirely on `OAuthRequired`, which is derived from connect-time OAuth config only. A server that connects anonymously has `OAuthRequired=false`, and a call-time "authorization required" is not a connection error, so the state machine stays Ready and nothing flags the server for sign-in. Fix (backend detection; reuses the existing action=login Sign-in CTA): - managed client: on a tools/call that fails with "authorization required" / 401 from an otherwise-connected server, set a per-server oauthCallRequired flag (cleared by a successful call or a fresh post-sign-in Connect). - runtime: feed the flag into the health calculator input as CallTimeOAuthRequired. - health calculator: a connected server with CallTimeOAuthRequired now returns Level=degraded, Action=login, Summary="Sign-in required" — placed after the connection-state switch so genuine error/disconnected states keep priority. This closes the gap MCP-1819 left (which only handled servers surfacing login-required at connect time). No new config keys, CLI flags, or API endpoints; the documented health contract (action=login) is unchanged. Tests: health calculator branch (+ negative: connection errors keep priority), managed-client call-time signal classification (auth vs 401 vs non-auth vs connection error). Related MCP-1819 (#628/#629/#630).
Deploying mcpproxy-docs with
|
| Latest commit: |
ca4cb9c
|
| Status: | ✅ Deploy successful! |
| Preview URL: | https://98146169.mcpproxy-docs.pages.dev |
| Branch Preview URL: | https://mcp-2084-calltime-oauth-cta.mcpproxy-docs.pages.dev |
|
Codecov Report❌ Patch coverage is
📢 Thoughts on this report? Let us know! |
📦 Build ArtifactsWorkflow Run: View Run Available Artifacts
How to DownloadOption 1: GitHub Web UI (easiest)
Option 2: GitHub CLI gh run download 27398763120 --repo smart-mcp-proxy/mcpproxy-go
|
Dumbris
added a commit
that referenced
this pull request
Jun 12, 2026
#638) The tool-quarantine change-detection hash covers description, input schema, and output schema, but GET /api/v1/servers/{id}/tools/{tool}/diff only returned the description and input schema. When an upstream evolved only its output schema (e.g. Google sqladmin adding a "POSTGRES_20" enum value), the visible description was byte-identical, so the flagged tool looked like a phantom rug-pull false positive. Add previous_output_schema / current_output_schema (already stored on the ToolApprovalRecord) to the diff response so an operator can see exactly what changed before approving. Document the field in docs/api/rest-api.md and add a regression test for an output-schema-only change. Note: change-detection itself is correct — the upstream schema genuinely changed. This deliberately does NOT weaken the rug-pull hash (no auto-approval of additive enum widenings); the fix is transparency, not relaxing security. Related #634 Related MCP-2085
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
2 participants
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
Problem (MCP-2084)
Remote MCP endpoints like Google's
com.googleapis.sqladmin/mcpallow anonymousinitialize+tools/listbut enforce OAuth only attools/call. The proxy connected, listed 15 tools, and the Web UI showed the server Connected / Enabled / Ready with no Sign-in affordance — yet every tool call failed at runtime withtransport error: authorization required. The operator only discovered the OAuth requirement by failing a call.This is not a duplicate of MCP-1819: that epic (#628/#629/#630) handles servers that surface login-required at connect time (
ErrOAuthPending/ 401 during initialize). Here the server connects + lists anonymously and only rejects at call time, so the health classifier never entered a login-required state.Root cause (file:line)
internal/health/calculator.go— the OAuth/login branch is gated entirely oninput.OAuthRequired.internal/runtime/runtime.go:1968—OAuthRequired: oauthConfig != nil, derived from connect-time OAuth config only. Anonymous connect ⇒oauthConfig == nil⇒OAuthRequired = false⇒ login branch skipped ⇒ classified healthy/Ready.internal/upstream/managed/client.goCallTool— a call-timeauthorization requiredis not a connection error, so the state machine stays Ready and nothing flags the server.Fix (backend detection; reuses the existing
action=loginCTA)tools/callthat fails withauthorization required/ 401 from an otherwise-connected server, set a per-serveroauthCallRequiredflag (extracted intorecordCallToolOAuthSignal). Cleared by a successful call or a fresh post-sign-inConnect.IsOAuthCallRequired()and feed it into the health input asCallTimeOAuthRequired.CallTimeOAuthRequirednow returnsLevel=degraded,Action=login,Summary="Sign-in required". Placed after the connection-state switch so genuine error/disconnected/connecting states keep priority.Once
health.action=loginis set, the existing #630 frontend calm Sign-in CTA and #629 macOS CTA render automatically — no new frontend/macOS work.Tests
internal/health: connected +CallTimeOAuthRequired⇒ degraded/Sign-in required/login; negative — connection errors keep priority.internal/upstream/managed: call-time signal classification (auth string vs 401 vs non-auth error vs connection error); flag clears on reconnect.Verification
go test ./internal/health/ ./internal/upstream/managed/ ./internal/runtime/ -race✅go build ./...✅ andgo build -tags server ./...✅./scripts/run-linter.sh✅ (0 issues)Docs
No new config keys, CLI flags, or API endpoints. The documented Unified Health Status contract (
action: login) is unchanged — this only makes it fire in a case it previously missed. No docs diff needed.Related MCP-1819 (#628 / #629 / #630).