Browse (TUI)

arc-skill-eval browse opens an interactive terminal UI built with Ink. It reads grading, timing, tool-use, and comparison data from evals-runs/.

# the current directory (a skill dir or a repo root)
arc-skill-eval browse

# one skill
arc-skill-eval browse ./skills/arc-conventional-commits

# a repo root (discovers every SKILL.md + evals/evals.json pair)
arc-skill-eval browse .

If <input> contains evals/evals.json, the browser loads that skill’s newest run. If <input> is a repository root, it lists discovered skills. When no run artifacts exist, it prompts you to run the suite first.

Layout

Four panels on the left control the detail pane on the right.

Panel	Contents
[1] Skills	Discovered skills with pass fraction and with-skill delta. Skills loaded only through --extra-skill appear dimmed with a distractor badge.
[2] Cases	Eval cases for the selected skill, with pass/fail status
[3] Assertions	Deterministic and judged assertions for the selected case
[4] Runs	Run / iteration history with the compare marker and exit code

For a case, the detail pane shows the tabs below. For an assertion, it shows the claim, evidence, and source from evals.json. For a run, it shows the reconstructed command, exit code, and output paths.

Case detail modes

When a case is selected, the detail pane is tabbed. Cycle the tabs with [ / ] (the strip shows in the pane header); v jumps straight to Raw.

Mode	Contents
Overview	Prompt, expected output, per-assertion grading with evidence, metrics (model, duration, tokens, cost, context-usage bar), and the with/without summary
Response	The final assistant response (assistant.md)
Diff (compare runs only)	A line diff from the without_skill response to the with_skill response
Trace	Tool-call digest from tool-summary.json (calls by name, bash/file/skill activity, external calls) plus the produced outputs/ listing
Context	The context-manifest.json — attached skills colored by role (target/extra/ambient), active/available tools, MCP tools, and ambient flags
Raw	The raw grading.json

Keybindings

The Keybindings reference is generated from src/tui/keymap.ts by scripts/gen-keymap-docs.mjs. The ? overlay uses the same source. The sections below explain multi-step actions.

Filtering and sorting

Press / to filter Skills and Cases by name. Press ↵ to keep the filter or Esc to clear it. F toggles a failures-only view. s sorts Skills by name, pass rate, with-skill delta, or cost. Active filters appear in the status bar.

Writing feedback (f)

With a case selected, press f to enter a note. Press ↵ to write or merge the note into feedback.json in the newest run directory. The browser sets the case status to reviewed and stores the text in notes. arc-skill-eval improve --from-feedback reads the same file.

Detail-pane cursor

The detail pane scrolls freely by default. Press →, l, or ↵ to place a cursor in it. Use j and k to move. In the Skills and Cases views, ↵ opens the selected case or assertion in its side panel. In the Assertions and Runs views, the cursor moves through section headers. Press ←, h, or Esc to return to free scrolling.

Re-running (r, R, o)

Press r to run the selected skill or case in the TUI. A console overlays the browser with a spinner, elapsed time, and per-case pass bars. When the run finishes, press ↵ to reload the artifacts or Esc to abort. In the Skills panel, r runs the whole skill. In the Cases panel, it runs only that case. R adds --compare.

Press o to enter custom run flags such as --model, --iteration, --extra-skill, or --context-mode ambient. The prompt starts with --model in Cases and --iteration in Runs. Press ↵ to run the command and reload, or Esc to cancel. This action starts a subprocess, so arc-skill-eval must be on PATH. Set ARC_SKILL_EVAL_BIN to use another binary path.

New eval case (n)

In the Skills or Cases panel, press n to enter an ID, prompt, and expected result. Saving appends a valid case skeleton with a placeholder file-exists assertion to evals/evals.json. Edit the file to replace the placeholder.

Cross-iteration comparison (c)

In the Runs panel, press c to pin the selected run as a baseline. The Runs view then shows pass-rate and cost changes relative to that run.

Flags

--no-baseline

Hide the without_skill comparison rows in the detail pane (handy when you only care about the with-skill result):

arc-skill-eval browse ./skills/arc-conventional-commits --no-baseline

Terminal compatibility

Colors and glyphs adapt to the terminal:

• Truecolor terminals use the Tokyo Night palette. Lower-color terminals use 256-color or named ANSI values. Set NO_COLOR or FORCE_COLOR=0 to disable color.
• UTF-8 locales use block bars, status marks, and arrows. Other locales use ASCII. Set ARC_TUI_ASCII=1 to force ASCII.
• Set ARC_TUI_THEME=tokyonight|gruvbox|nord to choose a palette. The default is tokyonight.

# force a plain, ASCII, no-color render
NO_COLOR=1 ARC_TUI_ASCII=1 arc-skill-eval browse .

# Gruvbox palette
ARC_TUI_THEME=gruvbox arc-skill-eval browse .

See also

• CLI reference for non-interactive commands
• Artifacts for the files under evals-runs/
• Dogfooding and authoring loop for interpreting the with/without delta