Add visual delimiters around each agentic-loop iteration (tool call) within a single turn

[leduftw]

Describe the feature or problem you'd like to solve

When I give Copilot CLI a single prompt, the agent runs an agentic loop that performs many iterations within that one turn — each iteration typically being a model step that issues one or more tool calls (file reads, edits, shell commands, MCP calls). Today these iterations stream into one continuous, undifferentiated block of output, with no visual boundary between them. I can't tell where one iteration ends and the next begins, which makes long agentic runs hard to scan, review, and debug.

Proposed solution

Within a single turn, render a horizontal delimiter line above and below each agentic-loop iteration (e.g. each tool call / command execution), so every step is visually bracketed and the boundaries between iterations are obvious.

Key points:

1. These are intra-turn delimiters — they appear between the iterations of the agentic loop that runs in response to one user prompt, not just a single separator at the end of the turn.
2. One delimiter line above and below each step (tool call), so steps are clearly separated from one another and from surrounding assistant text.
3. A plain rule is enough — no metrics, timings, or "worked for N" labels are needed. This is purely about visual separation.
4. Precedent: recent Codex CLI versions already do exactly this — while the agentic loop runs inside one turn, each step (e.g. a tool call) is printed bracketed by delimiter lines. Claude Code (like Copilot CLI today) does not.

Benefits

• Long agentic runs become scannable — easy to see each step and where it begins/ends.
• Faster review and debugging — jump straight to the exact iteration that did something wrong.
• Parity with recent Codex CLI, where this is already a clear usability win.

Example prompts or workflows

1. "Refactor module Y and run the tests" triggers a 10+ step agentic loop; each tool call is bracketed by delimiter lines, so I can scan step-by-step instead of reading one undifferentiated block.
2. A run goes sideways — I scan the per-step delimiters to jump straight to the iteration where the agent took a wrong action, instead of re-reading everything.
3. Reviewing a long transcript later — the per-step delimiters act as natural anchors / fold points, one per iteration.

Additional context

• To be unambiguous: this is not just an end-of-turn separator. The request is for a delimiter around every agentic-loop iteration inside the same turn.
• Precedent is recent Codex CLI behavior (observed first-hand): it prints delimiter lines around each agentic step within a turn. For reference, Codex CLI is open source at https://github.com/openai/codex (the terminal UI lives under codex-rs/tui).
• Copilot CLI version: 1.0.60.