Claude engine: pass --strict-mcp-config to avoid a TCP_DENIED for api.anthropic.com on every run

[benissimo]

Summary

With gh-aw v0.89.21 (AWF v0.28.23), every Claude-engine run logs one TCP_DENIED for api.anthropic.com:443 from the agent container, within a second of Claude Code starting. The run itself is fine. But the denial lands in the firewall summary on every run, and users read it as a regression after upgrading.

Cause

Claude Code fetches Anthropic's MCP registry (GET /mcp-registry/v0/servers on api.anthropic.com, unauthenticated) at startup. The result only marks which configured MCP servers count as "official" for Claude Code's analytics events; nothing functional reads it. The fetch ignores ANTHROPIC_BASE_URL, so it doesn't go through the api-proxy. Since api.anthropic.com is no longer on the agent allowlist, squid denies it, and claude-debug.log records:

[ERROR] Failed to fetch MCP registry: self signed certificate

Claude Code skips this fetch when it runs with --strict-mcp-config.

Proposal

Have the Claude engine pass --strict-mcp-config next to the --mcp-config argument it already adds (pkg/workflow/claude_engine.go). gh-aw is the sole MCP source for the agent, so no gh-aw-configured server is lost. Strict mode also stops Claude from loading MCP config committed in the target repo (for example .mcp.json), which looks consistent with gh-aw owning the agent's MCP surface.

Re-allowlisting api.anthropic.com for the agent would also silence it. But that reopens the direct agent-to-Anthropic path the api-proxy closes, for a call with no functional value.

Evidence

Same workflow (agent-review-pr), Claude Code 2.1.285, Sonnet 5.5, compiled with gh-aw v0.89.21:

• Without the flag: sandbox/firewall/logs/access.log has 172.30.0.20 api.anthropic.com:443 ... CONNECT 200 TCP_DENIED:HIER_NONE, and the debug log has the error above.
• With engine.args: ["--strict-mcp-config"]: 0 TCP_DENIED and no api.anthropic.com connection from the agent. There's no registry line in the debug log, and the GitHub MCP server via the MCP gateway still connects (MCP server "github": Successfully connected (transport: http)).

A local check on Claude Code 2.1.287 matches: the debug log shows [mcp-registry] Loaded 294 official MCP URLs by default and no registry fetch with --strict-mcp-config. Setting CLAUDE_CODE_DISABLE_NONESSENTIAL_TRAFFIC=1 also suppresses it, but it turns off feature-flag fetching too, so it seems a poorer default.

As a workaround, users can add engine.args: ["--strict-mcp-config"] to the workflow's engine: block.