…ped turns

Restore the full tool-call design on main's AgentPanel: a toolCallDisplay map turns each actionType into a friendly title ("Updating color tokens", "Inserting HTML", "Capturing preview"), a compact detail, a category icon (swatch/code/eye/…), and a tone (amber style, green write, red destructive). Names are canonicalised so both historical short names and current site_*/content_* names resolve. Consecutive same-role messages group under one label, and consecutive tool calls stack tightly in one container.

Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>

Conversations were all stuck on the placeholder title "New conversation". The chat handler now derives a short title from the first user prompt (collapse whitespace, cap 60 chars) and applies it via updateConversationForUser — only while the title is still the default, so a user-renamed chat is never clobbered. Adds DEFAULT_CONVERSATION_TITLE + deriveConversationTitle to the conversation store.

Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>

Drop the chat-bubble treatment: both user and assistant messages now render left-aligned at full width with no border or background. The user's prompt reads muted, the assistant's reply bright. Renames the now-inaccurate bubble classes (messageBubble→messageTurn, contentBubble→messageText, markdownBubble→markdownText).

Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>

When the agent runs set_color_tokens, render a row of colour swatches beneath the tool row — one per token, from params.tokens[].lightValue — composited over a neutral base with a hairline ring, matching the Framework home palette. Params are read defensively (they arrive as unknown from the tool-call stream).

Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>

The browser already captures a PNG for render_snapshot but discards it after handing it to the model. Capture it in the toolRequest handler (from the tool output's image channel) and stash it on the matching pending tool-call block, so the panel renders it inline — you see what the agent is looking at. Held in memory only (new screenshotDataUrl field on AgentToolCall); never posted to the server or persisted, so it rehydrates empty after a reload.

Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>

Message text steps up to text-m / weight 500 / muted (both roles), with a larger space-m gap between a turn's blocks. Tool-call rows sit on a stronger --overlay-10 surface with a 12px radius and uniform --space-xs padding.

Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>

…ol rows

Colour swatches now match the Framework home palette: full-width pill cells (input-radius), a little tall, with a 1px gap, and each wrapped in the shared Tooltip showing 'slug · value'. Tool-row title/detail step up to text-m to match the message text.

Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>

Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>

Starting a new chat (or deleting the active conversation) reset the active credential + model to null but never re-resolved the scope default, so the composer dropped to the "choose a model" lock even with a default configured. Both paths now re-run loadScopeDefault after the reset. loadScopeDefault also soft-fails a failed defaults lookup (logs + leaves the picker empty) instead of rejecting, so the fire-and-forget callers can't produce an unhandled rejection.

Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>

Conversations are deleted from the history dropdown, so the header trash button was redundant. Removes it and the now-unused clearAgentMessages binding (the store action stays — startNewAgentConversation still uses it).

Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>

Each turn's role marker now shows a small avatar and a relative timestamp instead of a bare "You"/"Assistant". The user reuses their Gravatar (UserAvatar, same as the toolbar); the agent gets the robot glyph in a matching circle. Relative time comes from the turn's first message; formatRelativeTime is extracted to a shared module so the history list and markers agree.

Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>

Deleting a conversation or loading the conversation list failed silently (console-only). Per the toast-bus rule, these user-triggered operation failures now surface a toast (kept the console.error for operator logs). loadScopeDefault stays a console-only soft fallback — it degrades gracefully and the actionable no-provider error surfaces at send time.

Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>

Colour swatches grow to 24px tall with 6px (space-xs) top/bottom padding around the row (1px gap unchanged). The agent's role avatar swaps the AI-box glyph (which duplicated the panel's header chip) for the sparkles glyph, which reads better as an identity avatar.

Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>

Reuse AiBoxSolidIcon — the shared AI glyph already used in the toolbar toggle and empty state — for the agent's role avatar, for consistency across the AI surfaces.

Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>

AgentPanel.tsx had grown past the 700-line module-size ceiling after the tool-row work. Move ToolCallRow (plus its category-icon map and tone-class helper, and the ~19 icon imports) into ToolCallRow.tsx, dropping AgentPanel.tsx from 737 to 604 lines. Behaviour-preserving.

Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>