[Bug] Important user-facing results are emitted only in Thinking and become invisible after session resume

[user-lzy]

Description

DeepCode CLI sometimes emits important user-facing information — including the final generated prompt, cost estimation, confirmation context, and task execution status — only inside Thinking messages.

These Thinking messages are visible during the live session, but after exiting and resuming the session they are permanently collapsed/not rendered to the user.

The underlying information is not actually lost from the model context: after resuming, I explicitly asked the agent to reproduce the previous final prompt and generation parameters without recalculating them, and it successfully returned the complete prompt and all parameters correctly.

Therefore, this appears to be a UI/message-routing issue rather than conversation-state loss.

Steps to reproduce

1. Start DeepCode CLI and ask the agent to perform a task that produces substantial intermediate/final information before AskUserQuestion.
2. Observe that the agent places important user-facing content inside Thinking. In my case this included:
   • complete final video-generation prompt
   • cost estimate
   • generation parameters
   • confirmation context
3. Answer the AskUserQuestion and continue the task.
4. Exit DeepCode CLI.
5. Resume the same session.
6. Observe that the previous Thinking content is no longer accessible/rendered to the user.
7. Ask the resumed agent to reproduce the previous information without recalculating it.
8. The agent can reproduce it correctly, demonstrating that the content remains in the model/session context but is no longer visible in the CLI.

Actual behavior

Important user-facing information may exist only in Thinking. After session resume, this information becomes invisible to the user even though it remains available to the agent.

Expected behavior

Information required for user decisions, confirmations, or final task results should not depend exclusively on transient Thinking output.

Such information should be emitted as durable user-visible assistant/tool content and remain accessible after session resume.

Alternatively, if Thinking messages are intentionally part of the user-visible transcript, the CLI should provide a way to inspect historical Thinking after resume.

Additional evidence

The attached screenshots show:

1. The complete final prompt and cost estimate being emitted inside Thinking.
2. AskUserQuestion containing only a summarized confirmation context.
3. After confirmation, additional task execution/status information continuing to be emitted through Thinking.

After resuming the session, I asked the agent to reproduce the previous prompt and parameters without recalculating them. It successfully returned the complete prompt and all parameters, confirming that the information was retained in the model context but was no longer rendered to the user.