Add plot_metrics builtin: braille line charts of JSONL streams in the conversation

[wilfred-dore]

Motivation

Vibe Work (web) can chart data in Canvas; Vibe Code (CLI) cannot. Terminal-first data/ML users still have to leave their environment (Jupyter, TensorBoard) just to see whether a training run converges — no terminal coding agent offers this natively today.

This PR adds a minimal, dependency-free visualization capability to the CLI.

What it does

New builtin tool plot_metrics: reads a JSONL stream (one flat JSON object per line, e.g. {"epoch": 3, "loss": 0.42, "accuracy": 0.88}), auto-detects numeric fields, and renders per-series braille line charts directly in the conversation. Read-only (permission = ALWAYS), zero new dependencies.

 2.314 ┤⠣
       │ ⠑⠄
       │  ⠈⠢⡀
       │    ⠈⠢⢄
       │      ⠈⠑⠤⣀⡀
       │          ⠈⠑⠒⠢⠤⣀⣀⡀
0.0238 ┤                 ⠈⠉⠉⠉⠑⠒⠒⠒⠒⠒⠒⠒⠤⠤⠤⠤⠤⠤⠤⠤⠤⠤⠤⠤⠤⠤⠤⠤⠤⠤⠤⠤⢄
       └ loss: x ∈ [1, 40]

sequenceDiagram
    participant U as User
    participant V as Vibe agent
    participant T as plot_metrics tool
    U->>V: "train this model, show me the curve"
    V->>V: run training, stream metrics.jsonl
    V->>T: plot_metrics(source="metrics.jsonl")
    T-->>U: braille chart in the conversation
    V-->>U: commentary (trend, convergence)

Loading

The fun part

The rendering reuses render_braille from vibe/cli/textual_ui/widgets/braille_renderer.py — until now used only to animate the mascot and the spinner. This gives that primitive its first product feature. 🐱

Open question for maintainers: this imports cli code from a core tool, which inverts the usual layering. Happy to move render_braille into e.g. vibe/core/utils/ if you'd prefer — kept the diff minimal for review.

Scope & design notes

• Line charts only, intentionally minimal for a first pass. The dot-cell dedup matters: render_braille sums dot indices, so duplicate dots would overflow past the braille block.
• Richer kinds (scatter, histograms, bar charts, ANSI-colored heatmaps, live --follow streaming, a colored Textual result widget) are demonstrated out-of-tree in Vibe Pulse, built at the Vibe Hackathon Paris (July 18, 2026) via tool_paths + a skill. A natural follow-up here would be a PlotextPlot-based result widget using textual-plotext, since the TUI is already Textual.

Testing

• Tool discovered by ToolManager._load_tools_from_file, prompt file loaded.
• Invoked against a real 40-epoch training stream: both series render, output contains only valid braille + box-drawing characters.

[wilfred-dore]

Full showcase of this capability (extended tool with colors/heatmaps/live follow, ml-trainer skill, W&B-style run tracking & comparison, on-device profiling via Qualcomm AI Hub): https://github.com/wilfred-dore/vibe-pulse