Tool cards

Why the balanced and full cards look the way they do. The user-facing table is in docs/layout.md (Tool cards).

What the others do

Read locally and read-only: pi 0.85.1 (@earendil-works/pi-coding-agent renderers), legacy MendCode (src/mendcode/.../tui/routes/session/index.tsx), .agents/plans/competitor-gaps-2026-10-02.md. Not found locally: Codex CLI sources (only the installed binary, whose TUI cannot be read), OpenClaude (the recovery-research folder is empty), grok-cli. Nothing was cloned or fetched.

Adopt: one row per looking tool with facts after the target (v1, pi); the range in the header (pi); a short tail with an elapsed time (pi, v1); intraline highlights (pi); a status colour instead of a box (pi tints). Avoid: a body for every read (pi shows none until expanded, v1 none); a coloured background per card (it fights the diff backgrounds); one list entry per row; the last lines of a file.

The frame

▤ Read README.md 123 lines 2.4s, then a thin rule │ holding the body.

• Glyph per kind, coloured by status. Target bold, facts muted, duration right-aligned from 1 s. A running call puts its time after the facts: the bottom-right corner belongs to the pet, which hides when a row reaches it.
• Nothing in the body repeats the header: a passing bun test is 241 pass · 0 fail in the header and has no body; a read header says its range and the body starts at the first line.
• Content in the form that suits it: a list is a grid, a search is grouped by file, source is highlighted with a muted gutter, a diff has two number columns and marks what changed inside a line, an unknown tool says the shape of its result.
• Density scales with the view: balanced is the header and at most 3 grid rows or 3 files; full is 6. minimal is untouched.
• Graph cards: the visual is the content. One muted summary row in balanced, a table of at most 6 rows in full, the whole lists only when the card is opened (detail nodes in the render spec).
• A call with a long input (a whole file, a long script) opens its card while the model still writes it, on every engine: the native one (Responses, Anthropic Messages and Chat Completions providers) and Claude Code. After 1,500 characters the card shows the call with what its input names so far and writing 84 lines · 6.2 KB (at most every 2,000 characters), so the working row never reads as stalled; once the input is whole the same card runs the tool, renamed if the whole input names it differently. A stream that stops first closes it as cancelled or failed, and nothing runs. Plan, secrets and task calls wait for their whole input.
• Display only. The model, /find, /copy and /export keep the whole result; finished cards are cached by item id.