feat: add BUB_USE_RESPONSES flag

[anvztor]

Feature Description

Expose a BUB_USE_RESPONSES flag so Bub can opt into Republic/any-llm Responses API routing when upstream gateways only implement /v1/responses.

Investigation

Republic/any-llm defaults to chat completions (/v1/chat/completions). Some OpenAI-compatible gateways only support Responses, which leads to 404s when Bub is configured with those endpoints.

Resolution

• Add BUB_USE_RESPONSES to settings and pass through to Republic LLM.
• Document the flag in env.example.

Dependency

• Depends on feat: add opt-in Responses API mode republic#16 (Responses instructions mapping for system/developer prompts).

Verification

• uv --directory vendor/bub run bub --help
• Local mock Responses API test (Bub runtime returns ok when BUB_USE_RESPONSES=true).

Closes #66

[anvztor]

Local mock Responses API test

We validated the Responses routing in Bub with a local mock server:

• Mock: .tasks/BUB-RESPONSES/mock_server.py (simple HTTPServer for /v1/responses and /v1/chat/completions).
• Command (abridged):
  • PORT=18000 LOG_PATH=.tasks/BUB-RESPONSES/mock_requests_local.log python3 .tasks/BUB-RESPONSES/mock_server.py &
  • PYTHONPATH=... BUB_USE_RESPONSES=true BUB_API_BASE=http://127.0.0.1:$PORT/v1 BUB_MODEL=openai:gpt-4o-mini BUB_API_KEY=dummy uv --directory vendor/bub run python - <<'PY' ... build_runtime ... PY
• Result: assistant_output: ok and mock log shows POST to /v1/responses with converted input.

Not added as an automated test because it requires a live local server/port and environment variables (fragile for CI). If you prefer an integration test, I can add a lightweight one behind a marker or skip flag.

mock_server.py