Reasoning summaries are omitted for default GPT and Claude thinking

[hqhq1025]

Problem

Reasoning-capable GPT and Claude models can return encrypted/signed reasoning items and non-zero reasoning token usage while Maka stores thinking.text as an empty string, so Desktop shows no Thinking disclosure. This affects direct providers, subscriptions, and third-party relays that expose the native Responses or Anthropic Messages protocols.

Root causes

OpenAI Responses

Maka did not explicitly request reasoning.summary when no thinking level was selected. The AI SDK only supplies a default summary when an explicit non-none reasoning effort is present, and compatible relays can omit summaries when effort remains implicit.

Claude / Anthropic

New adaptive-thinking Claude models can default to thinking.display: omitted, which returns thinking blocks and signatures but no visible text. Maka sent effort without requesting display: summarized.

Expected behavior

Routing follows the resolved wire/adapter rather than the configured provider label:

• known GPT models on OpenAI Responses, including custom Responses relays and third-party model overrides, send the documented default medium effort plus summary: auto
• explicit OpenAI thinking-off still sends none and omits summary
• known adaptive Claude models on Anthropic Messages, including custom Anthropic relays, OpenCode, and Copilot Anthropic routes, request display: summarized
• OpenAI Chat-compatible routes continue using their existing reasoning_content transport rather than receiving unsupported Responses fields
• MiniMax, Kimi, and unknown non-Claude Anthropic-compatible models are not assumed to support Claude display options
• returned summary/thinking deltas become Maka thinking events and render through the existing Desktop Thinking disclosure

Acceptance criteria

• custom Responses gpt-5.6-sol request contains reasoning: { effort: "medium", summary: "auto" }
• OpenAI, Codex subscription, and third-party Responses routes use the same summary behavior for recognized GPT models
• direct, subscription, and third-party Anthropic Messages routes contain thinking: { type: "adaptive", display: "summarized" } for recognized Claude models
• explicit off behavior remains unchanged
• real Responses summary SSE deltas survive the SDK/adapter boundary as non-empty Maka thinking text
• focused and full runtime suites pass

[hqhq1025]

Fix is available in draft PR #2645. It explicitly requests GPT reasoning summaries and adds a real Responses SSE → Maka thinking integration test.

[hqhq1025]

Default policy confirmed and implemented in #2645: reasoning-capable GPT/Claude routes now request visible thinking by default; only an explicit off disables it. This includes OpenAI Responses summaries, OpenAI Chat-compatible reasoning_effort, adaptive Claude summarized display, and legacy budget-based Claude with a 1,024-token thinking budget.