goal: OpenAI-Compatible Responses API (Limited: built-in search & scrape tools)

[locnguyen1986]

🎯 Goal

OpenAI-Compatible Responses API (Limited: built-in search & scrape tools)

📖 Context

We want a drop-in Responses API compatible with OpenAI’s request/response semantics and streaming behavior, but intentionally scoped. Instead of shipping a full assistants/tooling surface (code interpreter, retrieval, etc.), we’ll provide two built-in tools—search and scrape—with strong safety, rate limits, and deterministic outputs. This keeps the surface area small while enabling high-value “grounded” responses.

✅ Scope

• Endpoints & semantics

  • POST /v1/responses (JSON + SSE streaming)
  • GET /v1/responses/{id} to retrieve prior outputs (if persistence enabled)

• Streaming

  • SSE events for: response.created, response.output_text.delta, response.tool_call, response.completed, response.error

• Built-in tools (server-hosted)

  • search
  • scrape

• Tooling policy & safety

  • Allow/deny domain lists, per-org; rate limits & timeouts per tool; concurrency caps
  • Tool call audit trail (who, when, which URL/query); signed logs

• Usage & errors

  • usage counters for tokens; optional tool_units/tool_calls_count
  • Consistent error envelopes with actionable codes (e.g., tool_timeout, disallowed_domain, rate_limited)

❌ Out of Scope

• Non-text outputs (audio/image), function-calling beyond search/scrape
• Code interpreter, retrieval over private corpora, SQL/tools beyond the two built-ins
• Batch/async job APIs and multi-step agents (covered by Conversations/Runs goal)