Step summary is cluttered with `thinking_tokens` System Message blocks

[shibayu36]

Describe the bug

The step summary (Claude Code Report) shows many "⚙️ System Message" JSON blocks with subtype: "thinking_tokens". They make it hard to read Claude's actual responses and tool executions.

To Reproduce

1. Run a workflow with anthropics/claude-code-action@ba0aafd (Claude Code 2.1.204) on AWS Bedrock with --model global.anthropic.claude-opus-4-6-v1
2. Open the step summary after the run
3. Many thinking_tokens System Message blocks are displayed

Note: the same workflow and model with anthropics/claude-code-action@2fee155 (Claude Code 2.1.185) did not show these blocks. Also, the same action version (ba0aafd) with global.anthropic.claude-opus-4-7 did not show them, so the behavior appears to be model-dependent.

Expected behavior

thinking_tokens looks like an internal event for UI progress display. It should not appear in the step summary, or it should be aggregated into a single entry.

Screenshots

API Provider

• Anthropic First-Party API (default)
• AWS Bedrock
• GCP Vertex

Additional context

N/A

[anishesg]

Put together a fix in #1479.