TUI observe-first

A TUI is judged rendered in a real terminal, never from its source. Render it, screenshot it, compare it against strong CLIs at the same terminal grid, and cite a screenshot for every finding and fix. Terminal sibling of ui-fix-observe-first.

The comparative loop

1. Reference set: spawn the CLIs you must beat (e.g. claude, codex, droid) and your TUI in the same rendering stack and viewport. Reuse the same browser/PTY tab for every capture — a resized tab makes an unfair comparison.
2. Capture the same CUJ moments: welcome, empty composer, slash palette, streaming/working, tool receipts, permission dialog, error, interrupt, quit.
3. Teardown: every finding cites a screenshot. No screenshot = not a finding. Rank by impact; fix one issue-class at a time.
4. Re-render after every rebuild, then dogfood real turns (tools, interrupts). Many worst bugs only appear live.

What a strong agent TUI usually has

Element	Shape
Welcome header	Logo + name/version, model · permission-mode, cwd · session — few lines
Bottom bar	Mode + key hints left; model · session · cost · ctx% right
Working line	Spinner + verb + elapsed + interrupt hint + queued count
Turn / tool receipts	Compact status; collapsed summaries, not raw dumps
Text	Markdown-lite: bold, inline code, tinted fences
Palette	Ranked search; clear selected row; dim aliases
Dialogs	Clear selection + quick keys + hint row
Input	Distinct prompt; status lives outside the input line
Flow	Type-to-queue while running; Esc interrupts; Ctrl+C arms quit

Render-loop laws

• Never await a shared engine/session lock in the render loop. Clone handles out and poll with try_lock, or the spinner, timer, and keyboard freeze.
• Every new-turn path goes through one begin_turn() that stamps start time. Setting Running without a stamp freezes a 0s spinner.
• Drain completion channels before finalizing a turn, or leftover deltas duplicate response blocks.
• Clear the full frame before drawing — otherwise stale cells bleed through.
• Set the terminal cursor explicitly from the edit column.
• Repair histories at session load: dangling tool_use without tool_result breaks resume (API 400). Insert synthetic error tool_results at load.
• TestBackend tests gate layouts — they are not proof; screenshots are.

Capture traps

• Compare only at the same font/grid/renderer as your references.
• Browser attach pages can swallow Escape / chords — reproduce keyboard bugs against the raw PTY before changing code.
• Scrollback / ready-regex must match the current UI string after redesigns.
• cargo test | tail can eat exit codes — capture status yourself.

Common mistakes

Mistake	Reality
"It compiles / tests pass, ship it"	Many TUI bugs have green tests. Render + dogfood.
Diagnosing look/flow from source	Freeze, bleed-through, grid width, timing do not live in code.
Different tabs/sizes for comparisons	PTY grid follows viewport — same tab, same size.
Restyling without a reference set	Capture references first, then rank gaps.

Cross-references

• ui-fix-observe-first — web sibling (render → measure → cite → fix)