[VelikovPetar]

Platform alignment for how an agent's steps look. No public API is added or removed. Some theme defaults change.

Tool-call row (AIToolCallView)

• Declined call: the icon now follows the call's status. While the backend has yet to cancel the call, it shows the shimmering hand; after that, the × in the detail colour. The detail line still says "Declined". The duration now stays once the call has finished.
• Waiting shimmer: the icon is drawn in accentColor, and a band in the title's colour sweeps across it. The two colours were swapped before.
• Metrics: the icon defaults to 16 (ToolCallThemeData.iconSize, was 14), with a fixed gap of 8 after it (was max(6, iconSize / 2)).

Parts view (AIMessagePartsView)

• Steps are spaced 10 apart by default (was 8).
• The view fills a bounded width and shrink-wraps an unbounded one.
• AIUnsupportedPartView now reads like a tool-call row: the row's title font in its detail colour, with the same 16 icon and 8 gap. Before, it used bodySmall with an icon 1.25× the text size.

Approval card (AIToolApprovalCard)

• Spacing: 4 from the title to each line, 12 before the buttons and before the failure line, 8 between buttons.
• The sending spinner now sits after the buttons, in the accent colour, instead of inside the pressed button. It still announces "Sending your answer…".
• The Idle / Sending / Sent / Failed states and the "Sent" note are unchanged.

Goldens

CI goldens that change: run update_goldens on this branch.

• test/src/parts/goldens/ci/ai_tool_call_view_{light,dark}.png
• test/src/parts/goldens/ci/ai_tool_approval_card_{light,dark}.png
• test/src/parts/goldens/ci/ai_unsupported_part_view.png
• test/src/conversation/goldens/ci/ai_conversation_view_{light,dark}.png

Testing

• melos run lint:all passes.
• melos run test:all passes: 1287 tests.
• With CI=true, the failures are only the goldens listed above, plus the voice-view and approval-card CI goldens, which already fail on macOS on main.
• Updated widget tests: a declined call's icon follows its status; a declined call keeps its duration; the shimmer's base and highlight colours; icon size 16 and gap 8; default part spacing 10 and full width; the spinner placed after the buttons, with its semantics label; the unsupported placeholder's font, colour, icon size and theme.

Before / after

The demo's refund turn: two approval cards, one approved and one declined. Each image is before on the left, after on the right.

Moment	What changes	Before ← → After
Waiting for approval	Steps 10 apart (was 8), tool icon 16 with a gap of 8, accent-colored hand, unsupported-step placeholder in the row's font
Answer being sent	Spinner after the buttons instead of inside the pressed one
Declined, call still waiting	The hand keeps shimmering until the backend cancels the call (was × straight away)
Declined, call cancelled	× in the detail color, with the call's duration kept (was none)

🤖 Generated with Claude Code

[xsahil03x]

Thanks Petar, letting the icon follow the status and keeping a declined call's duration is the right model, and it is what both other SDKs already do. I compared the paddings, sizes, colors and spacing the description names against stream-chat-android-ai-compose and StreamChatAI on develop, and the regenerated CI goldens; apart from the gap below, they match, though I couldn't render either platform side by side. Analyze is clean, the non-golden tests pass, and 12 of the changed tests fail against main's lib, as they should.

Please do these before merging:

1. Keep the gap between a tool call's row and its approval card at 8. It comes from spacing, so it moved to 10 along with the steps, while Swift and Android keep it at 8.
2. Make the width test fail without the change: with a tight SizedBox it already passes on main.

The rest is optional and inline.

[xsahil03x]

Thanks! The question now sits a fixed 8 under its row whatever spacing is, the width test fails against main, and the spinner matches Android at 12 with a 1.5 stroke. Verified each against the new code.