Design: the Charm v2 stack, liberally and tastefully — what we use, what we measured and removed, and the menu of what's left on the table

Owner direction: cobra plus the Charm v2 family should give this CLI everything it needs for a genuinely beautiful terminal experience, used liberally — and the house taste rules are what keep liberal from becoming loud. This issue is the design record: the current inventory, the decisions already made (including one removal with numbers attached), and the adoption menu with the bar each candidate has to clear.

In the tree today

• lipgloss v2 — all styling, exclusively through cmd/flow/internal/ui/theme.go's role palette (charmtone pairs per role, light/dark per stream). No raw ANSI anywhere else; adherence is currently clean.
• bubbletea v2 — flow watch/flow run's live view (watchmodel.go), drawing to stderr so stdout stays pipeable.
• colorprofile — the entire NO_COLOR/CLICOLOR_FORCE/TERM/tmux decision, deliberately not reimplemented (ui.go:64-67).
• x/term, x/exp/charmtone — TTY detection per stream; the named palette.

Decided, with reasons that still hold

• fang: removed, measured. Its mustColorscheme queries the terminal background before any option can prevent it, which cost 4.05s on a pty answering nothing — on the help and error paths, the two surfaces a person reaches when something already went wrong (cmd/flow/help.go:16-34). Re-adopting fang requires that upstream behavior to change first; until then the custom help/error renderers are the design, not a stopgap. Any future Charm adoption inherits the same test: no library gets to ask the terminal a question this repo's ui package hasn't already answered.
• No spinners/progress-percent in watch — by design (watch.go:42-47); a durable run's progress isn't a fraction, and pretending otherwise is the kind of flash the charter rules out.

The menu — each with its bar

1. vhs for transcripts. The mirror tests already require that output quoted in README/docs regenerate when output changes (#327's constraint). vhs makes that mechanical and visual: .tape files checked in, GIFs/text rendered in CI, drift caught the same way flow docs generate catches prose drift. This is the highest-leverage item because it turns "beautiful in form" into something CI can hold — every styled surface gets a recorded, reviewed appearance. Bar: deterministic output (fixed width, FLOWSTATE_BACKGROUND pinned, fake clock where elapsed time shows).
2. bubbles components where the watch view grows real interaction (a run picker for bare flow watch, filterable flow list on a TTY). Bar: every interactive surface remains optional — the flag/plain path stays first-class, TTY-only enhancement never gates function (#327's degrade-cleanly rule; the MCP/agent path never sees a prompt it can't answer).
3. glamour for exactly one job: rendering the generated markdown under docs/reference/ in-terminal (flow docs view <topic> or similar), where the source really is markdown. Not for help prose — that's one backtick construct, handled in-house (#375), and importing a markdown engine to style help would invert that decision.
4. ultraviolet — bubbletea v2's renderer substrate; tracked, not acted on. When the watch view needs finer redraw control it's the sanctioned path; nothing today does.
5. harmonica — physics animation. The honest answer is probably never, and writing that down is the point of a taste record: motion in a CLI for durable workflows communicates nothing the status pills don't. If a use ever appears (smooth scroll in a long watch view), it rides bubbletea, not a new surface.

The taste rules, restated once

Already in force and worth naming as the charter for all of the above: nothing padded to a width a pipe cannot see; every heading survives losing its color; color is an accelerator, never the carrier; symbols not emoji, ASCII fallback always (symbols.go:3-16); stdout is the answer, stderr is the account; and one style token per concept, defined in the ui package, never per command.

Related: #327 (applying the language to informational output), #375 (inline code spans), #370/#372 (the sad paths adopting the same next-command element), docs/CLI_DESIGN.md (the charter this extends).

Read docs/CLI_DESIGN.md and the current UI entry points in cmd/flow/internal/ui/theme.go, ui.go, watchmodel.go, watch.go, and symbols.go. Review the measured fang removal and the adoption menu, then identify a concrete candidate whose stated bar can be tested. Done requires an agreed design decision and evidence that it preserves the charter and degrade-cleanly behavior.