Summary
We moved continuity off OpenAI servers and now maintain conversation state locally by persisting and replaying encrypted reasoning items. Requests are stateless (store=false) while retaining the performance/caching benefits of the Responses API. This has the side effect of removing complex logic around condensing.

Additional context https://cookbook.openai.com/examples/responses_api/reasoning_items

Why
This aligns with how Roo manages context and significantly simplifies our Responses API implementation while keeping all the benefits of continuity, caching, and latency improvements.

What changed

• OpenAI provider (Responses API everywhere)

  • All OpenAI models now use the Responses API and pass the system prompt via the top-level instructions field, not an input role OpenAiNativeHandler.createMessage(). We build the request with store=false and include=["reasoning.encrypted_content"] OpenAiNativeHandler.buildRequestBody().
  • We extract encrypted reasoning items from the last response for continuity OpenAiNativeHandler.getEncryptedContent().
  • Conversation formatting preserves any reasoning items alongside user/assistant turns OpenAiNativeHandler.formatFullConversation().

• Task runtime and persistence

  • We persist encrypted reasoning items (type: "reasoning", encrypted_content, optional id) into API history when assistant turns arrive Task.addToApiConversationHistory(). Type supports the new stored fields ApiMessage.
  • On subsequent turns, those reasoning items are replayed with the conversation Task.attemptApiRequest(), and the provider forwards them to Responses API input OpenAiNativeHandler.formatFullConversation().
  • API history persistence is atomic via safeWriteJson saveApiMessages().

• Defaults and gating

  • GPT‑5 reasoning summaries default to summary: "auto" when supported; text.verbosity is included only for models that support it OpenAiNativeHandler.buildRequestBody().

Removed (server-side continuity and legacy flags)

• previous_response_id flows and retry logic (tests marked DEPRECATED/skip) openai-native.spec.ts • Task.spec.ts.
• suppressPreviousResponseId and skipPrevResponseIdOnce (no longer needed with stateless continuity) Task.spec.ts.
• persistGpt5Metadata() and previous-response ID plumbing Task.attemptApiRequest().
• GPT‑5 response ID metadata from UI messages; no gpt5.previous_response_id in message schema anymore clineMessageSchema().

Kept (provider-agnostic metadata)

• taskId and mode on requests for downstream integrations (DeepInfra/Roo/Requesty/Unbound) ApiHandlerCreateMessageMetadata.

Result

• Aligned with Roo’s context model: single, local continuity path using encrypted reasoning items.
• Stateless API usage with preserved performance and caching characteristics from the Responses API.
• Simpler, less error-prone code (no previous_response_id branches, retries, or skip flags).
• No behavior regressions expected; Responses API streaming tests validate new behavior openai-native.spec.ts.