Bug: HTTP MCP backend tools missing from agent when `tools/list` response is slow

[chrisfregly]

Summary

When a workflow imports an HTTP MCP server whose tools/list response takes longer than
the gateway's tool discovery window, the agent receives an empty tool list for that server
and proceeds without those tools — even though the backend eventually initializes successfully.

Environment

Component	Version
gh-aw compiler	v0.68.1
gh-aw-mcpg	v0.2.17
Engine	Claude Code 2.1.98
MCP server type	http (LiteLLM proxy)

Steps to Reproduce

1. Configure a workflow with an HTTP MCP server whose tools/list response takes >~5s
2. Run the workflow

Expected Behavior

The agent receives the full tool list from all configured MCP backends before execution begins.

Actual Behavior

The agent only sees tools from fast-responding backends. The slow HTTP backend's tools are absent.
The agent reported:

"No OpsLevel MCP tools are available in this environment (only mcp__github__* and
mcp__safeoutputs__* exist). Without catalog validation, no candidates can be confirmed."

The backend itself succeeds — the gateway log confirms successful registration at 21s:

[2026-04-13T16:37:36Z] [DEBUG] [backend] GetOrLaunch called for server: opslevel                 
[2026-04-13T16:37:57Z] [INFO]  [backend] Successfully registered tools from opslevel (took         
21.213743169s)

But by then the agent has already started and its tool list is fixed. No error or warning
is surfaced — the tools are silently absent.

Root Cause (Hypothesis)

The gateway lazily initializes HTTP backends on the first tools/list request. If a
backend's response time exceeds the tool discovery window, the gateway returns the tool
list to the agent before the slow backend completes — silently dropping that server's tools.

Additionally, the compiler always emits "tools": ["*"] in the gateway config regardless
of what is declared in allowed:, which forces a live round-trip to discover tools and
makes the race condition unavoidable for slow backends.

Suggested Fix

• Block tool list delivery to the agent until all configured backends respond (or fail),
  rather than returning a partial list silently
• Surface an error when a backend's tools/list exceeds the discovery window, rather
  than proceeding as if the server has no tools
• Static registration: when allowed: declares specific tool names, have the compiler
  emit those as a static "tools" list in the gateway config — allowing the gateway to
  register them immediately without a round-trip and bypassing the race entirely

[lpcox]

Fixed by PR #3782 — increased default HTTP transport connect timeout from 5s to 30s and made it configurable via connect_timeout per-server field. Also promoted failed backend tool registration from WARN to ERROR with summary logging.