fix: Gemini engine passes unsupported --mcp-config flag to Gemini CLI

[Mossaka]

Bug

The Gemini engine compiler generates a CLI invocation with --mcp-config, which the Gemini CLI does not support:

gemini --mcp-config /tmp/gh-aw/mcp-config/mcp-servers.json --output-format json --prompt "..."

Gemini CLI rejects this immediately:

Unknown arguments: mcp-config, mcpConfig

Reproduction

Triggered Smoke Gemini workflow on gh-aw-firewall after merging github/gh-aw-firewall#974:

• Failed run: https://github.com/github/gh-aw-firewall/actions/runs/22202278807
• Gemini CLI version: @google/gemini-cli@0.29.0
• gh-aw version: v0.47.0

Root Cause

The compiled lock file (line 1373) passes --mcp-config as a CLI flag:

gemini --mcp-config /tmp/gh-aw/mcp-config/mcp-servers.json --output-format json --prompt "$(cat /tmp/gh-aw/aw-prompts/prompt.txt)"

Gemini CLI does not have a --mcp-config flag. Instead, it reads MCP server configuration from settings.json files:

• Global: ~/.gemini/settings.json
• Project: .gemini/settings.json

The mcpServers key in settings.json supports command, url, or httpUrl transports. See: https://geminicli.com/docs/tools/mcp-server/

Impact

• Agent never starts (0 turns, 0 tokens)
• All downstream jobs (detection, safe_outputs) are skipped
• The firewall itself works correctly — only the CLI invocation is broken

Suggested Fix

In the Gemini engine command builder (pkg/workflow/gemini_engine.go or equivalent):

1. Instead of passing --mcp-config <path>, write the MCP config to .gemini/settings.json (project-level) as a pre-step before invoking the Gemini CLI
2. The settings.json format uses mcpServers key with server entries containing command/url/httpUrl
3. Remove --mcp-config from the generated command

Comparison with Other Engines

Engine	MCP Config Method
Claude	--mcp-config <path> CLI flag
Copilot	--mcp-config-file <path> or ~/.copilot/mcp-config.json
Codex	--mcp-config <path> CLI flag
Gemini	~/.gemini/settings.json or .gemini/settings.json file

Audit Data

gh aw audit 22202278807

• 1 error, 0 warnings
• Agent job: failure
• Duration: 3.8m (mostly container build time)