Problem

When using the localgpt CLI (localgpt chat), the output has formatting inconsistencies:

1. Log lines interleave with streamed model output: Tracing/log output can appear mid-line during streaming, mixing with the model's response text.
2. Missing line breaks before follow-up responses: After tool execution, the follow-up model response concatenates directly with previous content without a proper line break or "LocalGPT:" label.

Root Cause

In crates/cli/src/cli/chat.rs:

Issue 1: Tool follow-up output (line 500)

print!("{}", follow_up);

After tool calls are executed and the follow-up response is printed, there is no leading newline or "LocalGPT:" label. This causes the follow-up text to concatenate directly onto the end of the previous tool call output.

Issue 2: Streaming delta output (line 422)

print!("{}", chunk.delta);

Log lines from tracing/other threads can interleave with the streamed chunks mid-line since stdout and stderr are not synchronized. The tracing subscriber should be configured to write to stderr or a log file, not stdout, to avoid interleaving.

Issue 3: Post-stream newlines (line 520)

println!("\n");

This adds trailing spacing after the full stream, but if tool calls intervene, the newline before the follow-up "LocalGPT:" response is absent.

Required Changes in crates/cli/src/cli/chat.rs

1. Line 500: Change print!("{}", follow_up) to print!("\nLocalGPT: {}", follow_up) so the follow-up response after tool execution starts on a fresh line with the proper label, consistent with the initial response label on line 410.

2. Ensure newlines before tool call notifications (lines 447-450 and 482-485): The tool call markers like [tool_name: detail] should always start on a clean new line. Currently they use println!("\n[...]") which should be sufficient, but verify they produce clean output after streamed content that may not end with a newline.

3. Verify tracing/logging is directed to stderr or a file, not stdout: Check that the tracing subscriber setup in the CLI entry point does not write to stdout. If it does, it should be changed to write to stderr (std::io::stderr()) to prevent log lines from interleaving with the streamed model output. Look at the CLI's main entry point and tracing initialization code (likely in crates/cli/src/main.rs or similar) to ensure the tracing layer uses stderr or a file writer, not stdout.

Expected Behavior

• The "LocalGPT:" label should appear on its own line before every model response, including follow-up responses after tool execution.
• Log/tracing output should not interleave with streamed model output on the same line.
• Each distinct section of output (user prompt, model response, tool calls, follow-up) should be clearly separated by line breaks.