Retry automatically when LLM returns only thinking blocks

[okottorika]

Problem

When Claude models intermittently return a response containing only thinking blocks with no text or tool_use content, the UI shows a blank turn to the user. The model produces a thinking block (indicating it understood the request) but generates zero output tokens beyond the end-of-turn token.

This was reported in the Discord bug channel where message 246 in a conversation received no visible response from the agent. A simple "Continue" retry from the user produced a proper 448-token response, confirming this is intermittent LLM-level behavior rather than a Shelley backend bug.

Solution

Add detection in the conversation loop (loop/loop.go) for responses that contain only thinking/redacted_thinking blocks with no visible content. When detected, automatically retry the LLM request instead of ending the turn. The empty response is not recorded to the database, so the user never sees the blank turn.

Changes

• loop/loop.go: Added respHasNoVisibleContent() helper and a check in processLLMRequest() that retries when the response contains only thinking blocks
• loop/loop_test.go: Added TestRetryOnThinkingOnlyResponse (integration test using a mock LLM service) and TestRespHasNoVisibleContent (unit test covering all content type combinations)

Behavior

Before: LLM returns thinking-only → blank turn shown to user → user must manually retry

After: LLM returns thinking-only → loop detects no visible content → automatic retry → user sees normal response

Fixes #131