Custom API Endpoints doesn't work

[MH0386]

I tried to use the new Custom API Endpoints via Environment Variables, but it seems to have a bug.

I added this:

engine:
  id: codex
  model: "openrouter/free"
  env:
    OPENAI_BASE_URL: "https://openrouter.ai/api/v1"
    OPENAI_API_KEY: ${{ secrets.OPENROUTER_API_KEY }}

But I found this in the agent job log. It seems to use gpt-5.3-codex model.

2026-03-14T22:21:22.758990Z DEBUG session_init: codex_core::codex: Configuring session: model=gpt-5.3-codex; provider=ModelProviderInfo { name: "OpenAI", base_url: Some("http://172.30.0.30:10000/v1"), env_key: None, env_key_instructions: None, experimental_bearer_token: None, wire_api: Responses, query_params: None, http_headers: Some({"version": "0.114.0"}), env_http_headers: Some({"OpenAI-Organization": "OPENAI_ORGANIZATION", "OpenAI-Project": "OPENAI_PROJECT"}), request_max_retries: None, stream_max_retries: None, stream_idle_timeout_ms: None, requires_openai_auth: true, supports_websockets: true }

---

Knowing that I tried to configure Codex locally with the same base URL (https://openrouter.ai/api/v1) and model (openrouter/free), and it works properly.

~/.codex/config.toml:

model_provider = "openrouter"
model="openrouter/free"

[model_providers.openrouter]
name = "openrouter"
base_url="https://openrouter.ai/api/v1"
env_key="OPENROUTER_API_KEY"

---

The agent summary:

Agentic Conversation

Conversation:

Agent: stream disconnected before completion: stream closed before response.completed

Agent: Reconnect attempts: 5/5

Statistics:

MCP Gateway

• ✓ startup MCPG Gateway version: v0.1.15
• ✓ startup Starting MCPG with config: stdin, listen: 0.0.0.0:80, log-dir: /tmp/gh-aw/mcp-logs/
• ✓ startup Loaded 2 MCP server(s): [github safeoutputs]
• ✓ startup Guards sink server ID logging enrichment disabled
• ✓ backend

  Successfully connected to MCP backend server, command=docker
  

• 🔍 rpc github→tools/list
• 🔍 rpc safeoutputs→tools/list
• 🔍 rpc safeoutputs←resp {"jsonrpc":"2.0","id":2,"result":{"tools":[{"name":"create_pull_request","description":"Create a new GitHub pull request to propose code changes. Use this after making file edits to submit them for review and merging. The PR will be created from the current branch with your committed changes. For code review comments on an existing PR, use create_pull_request_review_comment instead. CONSTRAINTS: Maximum 1 pull request(s) can be created. Title will be prefixed with \"[code-simplifier] \". Labels [\"refactori...
• 🔍 rpc github←resp {"jsonrpc":"2.0","id":1,"result":{"tools":[{"annotations":{"readOnlyHint":true,"title":"Get commit details"},"description":"Get details for a commit from a GitHub repository","inputSchema":{"properties":{"include_diff":{"default":true,"description":"Whether to include file diffs and stats in the response. Default is true.","type":"boolean"},"owner":{"description":"Repository owner","type":"string"},"page":{"description":"Page number for pagination (min 1)","minimum":1,"type":"number"},"perPage":{"descriptio...
• ✓ startup Starting in ROUTED mode on 0.0.0.0:80
• ✓ startup Routes: /mcp/ for servers: [github safeoutputs]

Firewall Activity

▼ 6 requests | 6 allowed | 0 blocked | 1 unique domain

Domain	Allowed	Denied
openrouter.ai	6	0

[pelikhan]

@copilot update the agentic engine to declare the command api base url environment variable like "OPENAI_BASE_URL" (determine the ones from copilot/claude/gemini) and configure the api-target URL passed to awf accordingly. The compiler should detect this environment variable in the yaml and special treat it. Make sure the example above is operational.