Reduce extraneous prompt refreshing

[rolandwalker]

Description

Only refresh prompt escapes when the entered text is empty. The current behavior is to refresh on every keystroke, which for dynamic format strings, such as the current time, is needless and distracting.

With this change, we usually refresh the prompt only after starting a new query.

Changes

• Move the message parameter to prompt() from PromptSession(), so that the value can be loaded with the application itself.
• Recast get_message() as get_prompt_message() for clarity.
• In the get_prompt_message() callback, peek at the content which the user has typed, and return a cached value if there is any current text.

There still can be an extraneous refresh if the user deletes all content from the line, creating an empty line.

Checklist

• I added this contribution to the changelog.md file.
• I added my name to the AUTHORS file (or it's already there).
• To lint and format the code, I ran

  uv run ruff check && uv run ruff format && uv run mypy --install-types .

[github-actions]

Findings

1. High: prompt disappears in secondary prompts
   After moving message off the PromptSession initializer, the calls in handle_editor_command() and the LLM loop no longer pass message. That means those prompts render with an empty/blank prompt (and lose the prompt refresh behavior). This is a regression from previous behavior where PromptSession carried message for all prompts.
   File references: mycli/main.py:791, mycli/main.py:1077.
   Action: pass the same message callable in these prompt() calls (e.g., reuse loaded_message_fn or refactor so both places can call a shared helper), or restore message=get_prompt_message on PromptSession and override only in the main prompt if needed.

Missing tests / edge cases

1. No coverage for prompt display in non-main prompt flows (\e editor flow and LLM flow). If there’s test infrastructure around prompt rendering or prompt arguments, add a small unit test to assert PromptSession.prompt is called with message in those paths. Otherwise, a minimal integration test that exercises \e and LLM command loops should detect a blank prompt regression.

Notes

No security concerns spotted in this change.

[rolandwalker]

I've realized that the call in handle_editor_command() isn't actually interactive,
Edit: this is wrong. It is interactive.

but want to still check the LLM usage.
Edit: checked.

[scottnemes]

Verified prompt only updates after submitting a statement, or after deleting all input as noted.