[glayyiyi]

Summary

With Claude Opus 5.5 on Amazon Bedrock, some Flash turns come back with no content and these stop details:

stop_reason: refusal
stop_details: {type: refusal, category: reasoning_extraction}

The turn has 0 output tokens. The run then has nothing to act on.

Two pieces of the Flash prompt ask the model to write out its thinking:

• the cognitive-protocol line in flash_runner.md: "think through observation, planning and action … your reply carries one short paragraph of connected prose reflecting those three phases";
• REASONING_REMINDER: "Your previous turn carried no visible reasoning text (internal thinking is not shown) …".

This PR rewords both as a progress note: what the screen shows, which milestone it is on, and the next action. The step log receives the same information, and nothing else changes.

Evidence

I captured the refused requests through a logging proxy and replayed them unchanged against the same endpoint.

request	change	Opus 5.5	Sonnet 5.5	Haiku 5.5
turn 1 (system prompt only)	none	refusal 3/3	tool call 2/2	tool call 2/2
turn 1	reword the protocol line	tool call 4/4
turn 5 (with the reminder)	reword the protocol line only	refusal 2/2
turn 5	reword the reminder only	refusal 2/2
turn 5	reword both (this PR)	tool call 3/3

The refusal depends on the request content. On some screens the original wording does go through, so it does not happen on every run.

With this change, plus #217, the two cases I use for on-device acceptance pass 4/4 rounds on Opus 5.5, with no refusals. One case is applying a theme in a custom-drawn store app; the other is Settings → Dark theme.

Tests

The existing Flash runner and reminder tests pass. They compare against the REASONING_REMINDER constant, not its text. make test, make lint and make typecheck pass.