Frame composition — the full screen as a value

The TUI knows how to paint a screen, but historically it did so in two places that could drift:

• the room body went to terminal scrollback via tea.Println out of the transcript flush, never returned to a caller; and
• the bottom chrome (live routing line, action-required banner, divider, mode-prefixed prompt, per-room footer, framework status row) was assembled inline inside RootModel.View().

The only artifact a headless caller could get — the trace's view_rendered — was a width-80, ANSI-stripped projection of the body only: neither the right width, the right styling, nor the whole screen.

The composer

internal/tui/frame.go introduces one seam that returns "the full screen a human sees" as a value:

frame := tui.ComposeFrame(&m, width, height) // body + chrome, at this width

Frame carries both projections plus a typed sidecar:

field	meaning
ANSI	styled, screenshot-ready screen
Text	ansi.Strip(ANSI) — agent-readable; the twin holds
Width/Height	geometry the frame is paint-equivalent to
Metadata	FrameMeta{State, Mode, AllowedIntents, WorldDigest}

Metadata is read straight from the model and the machine — the AllowedIntents are exactly what Orchestrator.AllowedIntents(state, world) reports (the same source inspect/turn use), so a consumer trusts the sidecar without re-parsing Text.

One paint, two consumers

ComposeFrame is width-parameterised: it resizes a copy of the model to the requested geometry through the live resize() seam, so a screenshot at 100 cols is byte-identical to a real 100-col terminal. The body region is the transcript's last flushed entry (transcriptModel.LastBody()); the chrome region is built by composeChromeParts + joinChromeParts.

RootModel.View() now routes its own bottom region through those same two helpers — so the live screen and every headless capture are the same bytes. View() emits chrome only (the body lives in scrollback), so its output is byte-identical to the pre-composer assembly.

Frame (what ComposeFrame returns)         Source renderer (unchanged)
┌─────────────────────────────────┐
│ <last room body, reflowed @W>   │  ◀── transcript.LastBody()
│ <live routing/thinking line>    │  ◀── transcript.LiveLine()
│ <action-required banner>        │  ◀── inbox.ActionRequiredBanner()
│ ─────────────────────────────── │  ◀── blocks.Renderer.Divider()
│ > <prompt + mode prefix>        │  ◀── composePromptAndBanner()
│ <per-room footer>               │  ◀── footerStoryLine()
│ room · state · mode · queue     │  ◀── blocks.Renderer.StatusRow()
└─────────────────────────────────┘

Why this matters

ComposeFrame is the seam the rest of the mcp-studio epic consumes — the headless driver, the screenshot tool, and the MCP render.* tools all read one Frame instead of re-deriving the screen. Because the live TUI is the first consumer of its own frame, a headless capture and a real terminal can never disagree.

See internal/tui/frame_test.go for the contract: frame-equals-live-paint (body + every chrome element, in order), width fidelity (no chrome overflows the requested width; a wide frame reflows past the 80-col fossil), metadata-matches-the-machine, and ansi.Strip(ANSI) == Text.