Conversation view displays Claude system metadata as assistant replies

Conversation view displays internal Claude Code system metadata inside assistant replies: token counters, environment updates, progress-reminder instructions and skill listings. Local transcript inspection found 43 matching records marked type=api-request-blob with message.role=system. The view appears to be rendering these internal request records as conversation content.

Expected: exclude internal request/system records and retain user-facing messages. Actual: internal metadata overwhelms the assistant's useful output, making long answers difficult to find and read. Observed with Claude Code 2.1.296. No project names, paths, conversation text or screenshot are included.

AgentsRoom diagnostics

• App version: 1.211.0
• OS: win32 10.0.26200 (x64)
• App language: system

Expected behavior

Conversation view excludes internal API request/system records while retaining user-facing messages.

Actual behavior

Token counters and internal system instructions appear inside the assistant reply, cluttering the conversation.

Steps to reproduce

1. Use a Claude Code agent in Conversation view.
2. Run a task that produces multiple messages and internal runtime updates.
3. Inspect the displayed assistant reply: token counters and internal instructions appear alongside normal response text.

Version or build

Claude Code 2.1.296 on Windows; AgentsRoom installation previously reported 1.210.0.

Errors and logs

Read-only transcript inspection found 43 records matching the displayed internal text, with type=api-request-blob and message.role=system. No raw transcript or project-specific content is attached.