From 67011befe3ae31397644a4796cef493b84f94c31 Mon Sep 17 00:00:00 2001
From: ciregenz <ciregenz@berkeley.edu>
Date: Tue, 4 Aug 2026 16:43:07 -0700
Subject: [PATCH] [eric] mcp: discovery refuses a credentialless env_vars
 server, and the failure translator reads the whole tail not the npx dump

---
 backend/apps/tools_lib/mcp_discovery.py       |  3 +-
 backend/apps/tools_lib/tools_lib.py           |  4 +++
 backend/tests/test_mcp_failure_reason.py      | 34 +++++++++++++++++++
 frontend/src/app/pages/Tools/integrations.tsx |  2 +-
 4 files changed, 41 insertions(+), 2 deletions(-)

diff --git a/backend/apps/tools_lib/mcp_discovery.py b/backend/apps/tools_lib/mcp_discovery.py
index be8bd501..01b9bc0c 100644
--- a/backend/apps/tools_lib/mcp_discovery.py
+++ b/backend/apps/tools_lib/mcp_discovery.py
@@ -182,7 +182,8 @@ async def discover_mcp_tools_stdio(command: str, args: list[str] | None = None,
                     await asyncio.wait_for(asyncio.shield(stderr_task), timeout=1.0)
                 except (asyncio.TimeoutError, asyncio.CancelledError, Exception):
                     pass
-                tail = "".join(stderr_tail[-10:]).strip()
+                # Full window, not the last few lines: an npx wrapper prints ~20 lines of Node crash dump AFTER the server's one useful fatal line, so a short tail hands the translator pure noise.
+                tail = "".join(stderr_tail).strip()
                 # A Go server's dying breath is a JSON line with a goroutine dump. Handing that to
                 # the UI hides the one fact the user can act on, which is usually "sign in again".
                 raise HTTPException(status_code=502, detail=readable_mcp_failure(tail))
diff --git a/backend/apps/tools_lib/tools_lib.py b/backend/apps/tools_lib/tools_lib.py
index 3a3fedde..039c4ff1 100644
--- a/backend/apps/tools_lib/tools_lib.py
+++ b/backend/apps/tools_lib/tools_lib.py
@@ -362,6 +362,10 @@ async def delete_tool(tool_id: str):
 async def discover_tools(tool_id: str):
     tool = load(tool_id)
 
+    # A credential-driven server with no credentials dies at boot with a crash dump; refuse to spawn and say the useful thing instead.
+    if tool.auth_type == "env_vars" and not tool.credentials:
+        raise HTTPException(status_code=409, detail=f"{tool.name} isn't connected yet. Connect it first, then discover its tools.")
+
     if tool.auth_type == "oauth2" and tool.auth_status == "connected":
         if tool.oauth_tokens.get("refresh_token"):
             if tool.name.lower() == "airtable":
diff --git a/backend/tests/test_mcp_failure_reason.py b/backend/tests/test_mcp_failure_reason.py
index 36cd2b21..524e037e 100644
--- a/backend/tests/test_mcp_failure_reason.py
+++ b/backend/tests/test_mcp_failure_reason.py
@@ -25,6 +25,32 @@ SLACK_NO_TOKENS = (
     '"app":"slack-mcp-server","stacktrace":"provider.New\\n\\tapi.go:682"}'
 )
 
+# Byte-for-byte the FULL stderr of a tokenless `npx -y slack-mcp-server` run (2026-08-04): the Go
+# fatal comes FIRST, then the npm wrapper's execFileSync crash dump buries it under ~20 Node lines.
+# This is the exact toast Haik and Eric saw; a translator fed only the last few lines can never win.
+SLACK_NO_TOKENS_WITH_NPX_DUMP = SLACK_NO_TOKENS + """
+node:child_process:963
+    throw err;
+    ^
+
+Error: Command failed: /Users/x/.npm/_npx/2f12aed4e6049c73/node_modules/slack-mcp-server-darwin-arm64/bin/slack-mcp-server-darwin-arm64 --transport stdio
+    at genericNodeError (node:internal/errors:983:15)
+    at wrappedFn (node:internal/errors:537:14)
+    at checkExecSyncError (node:child_process:924:11)
+    at Object.execFileSync (node:child_process:960:15)
+    at Object.<anonymous> (/Users/x/.npm/_npx/2f12aed4e6049c73/node_modules/slack-mcp-server/bin/index.js:64:14)
+    at Module._compile (node:internal/modules/cjs/loader:1692:14)
+    at TracingChannel.traceSync (node:diagnostics_channel:322:14) {
+  status: 1,
+  signal: null,
+  output: [ null, null, null ],
+  pid: 96234,
+  stdout: null,
+  stderr: null
+}
+
+Node.js v24.4.0"""
+
 
 def test_the_real_slack_failure_becomes_reconnect_advice():
     out = readable_mcp_failure(SLACK_REAL)
@@ -46,6 +72,14 @@ def test_missing_tokens_reads_differently_from_expired_ones():
     assert never != expired
 
 
+def test_the_npx_crash_dump_never_buries_the_real_reason():
+    # The regression that survived 1.7.2: the useful fatal line is FIRST and the Node noise last.
+    out = readable_mcp_failure(SLACK_NO_TOKENS_WITH_NPX_DUMP)
+    assert "signed in" in out
+    for leak in ("TracingChannel", "execFileSync", "node:child_process", "status: 1", "Node.js v"):
+        assert leak not in out, f"{leak!r} leaked into what the user reads"
+
+
 def test_revoked_access_says_so():
     assert "revoked" in readable_mcp_failure('{"error":"token_revoked","message":"bad"}').lower()
 
diff --git a/frontend/src/app/pages/Tools/integrations.tsx b/frontend/src/app/pages/Tools/integrations.tsx
index 68958a4b..a0969a32 100644
--- a/frontend/src/app/pages/Tools/integrations.tsx
+++ b/frontend/src/app/pages/Tools/integrations.tsx
@@ -185,7 +185,7 @@ export const INTEGRATIONS: Integration[] = [
     id: 'slack',
     name: 'Slack',
     description: 'Search messages, send messages, read channels, DMs, and threads in Slack workspaces.',
-    mcp_config: { type: 'stdio', command: 'npx', args: ['-y', 'slack-mcp-server@latest', '--transport', 'stdio'], env: { SLACK_MCP_ADD_MESSAGE_TOOL: 'true' } },
+    mcp_config: { type: 'stdio', command: 'npx', args: ['-y', 'slack-mcp-server@1.3.0', '--transport', 'stdio'], env: { SLACK_MCP_ADD_MESSAGE_TOOL: 'true' } },
     color: '#4A154B',
     website: 'https://github.com/korotovsky/slack-mcp-server',
     icon: (