PR #8339 - superset-sh/superset

Links

Summary

Tool runs start collapsed (SUPER-2622). A run of tool calls is one summary line ("Ran 4 commands") while the turn runs, as in Codex. A run opened for an approval stays open after the approval.
Agent icon in the chat pane header (SUPER-2627) and reply actions (SUPER-2616): tooltips, a split icon for Branch, and the date on a reply from an earlier day. No hidden action row under a running turn's latest message, which used to leave a 22px gap above the live run.
Commands, files and pages as chips (SUPER-2664). A picked /command or skill, an @file, and a new @page read as a dashed-underlined chip, in the composer and in sent messages. Hover shows the full /name with the agent's description, the file path, or the page URL.
Background work across several chats. A job started by a subagent no longer stays in the activity menu forever, and each row names its chat when several chats have work.

Collapse: transcriptRows defaults every run to collapsed unless it waits on an approval. Transcript records an approval-opened run as expanded so approving does not collapse it. lastReplyKeys only gives a settled turn a last reply; the working row carries running from the turn status, since completedAtMs is optional on settled turns.
Chips: the composer records each chip's place in the sent text as protocol elements (UTF-8 byte range, kind, and a new optional label). UserMessageRow splits the text back into chips. Older messages without elements fall back to a leading /command the agent advertises, so a path stays plain text. The look lives in a shared @superset/chat-ui/Chip. Picking a command only inserts it; Enter no longer sends.
Slash menu, after Zed: groups by ACP _meta.command_category ("Commands", "MCP Server Commands"), shows the argument hint, and refuses an unknown leading /command with a toast while keeping the text (onSubmit may return false to keep the draft). Claude does not set the category yet, so its commands share one heading, as in Zed.
Background work: the Claude ACP adapter routes a subagent's task updates to the subagent's session and drops them once it finishes (native-subagents.js, routeOfToolCall), so the "finished" update never came. AcpAdapter now ends a subagent's jobs when the subagent finishes.

Done in the dev app over CDP with real Claude sessions:

A 9-step turn (about 35 tool calls): every run collapsed at every 15-second sample, the live one included, and after a reload
No gap above the live run; the final reply gets Copy, Branch and the time once the turn ends
Chat pane header shows the Claude logo; Copy and Branch tooltips; split icon
/goal then Enter inserts the chip and sends nothing (12 user messages before and after)
Command chip hover shows /goal and its description; file chip shows notes/release-plan.md; page chip shows the page URL
File and page chips stay chips after a full reload; the agent read the file by its path
/notacommand hello is refused with a toast and the text stays; /Users/me/file.ts style text still sends
Two Claude chats: the activity menu counted 5 (3 processes, 2 subagents), each row named its chat, a click opened the owning chat, Stop stopped only that chat's job
A subagent's background sleep left the list when the subagent ended (it stayed for 273 s before)
Not seen in the app: a run waiting on an approval (the session auto-approved; covered by unit tests), the dated time for an older reply (formatter output checked)

bun test: desktop v2-workspace suites 500 pass; chat 54, chat-ui 10, chat-runtime ACP adapter 59, all pass. New tests: approval-opened runs, running turns have no last reply, a subagent's job ends with it, command category passthrough, command chip labels, element byte ranges with non-ASCII text.
typecheck: chat, chat-ui and chat-runtime are clean. Desktop's only errors are in packages/auth and packages/trpc better-auth types in this worktree, none in changed files.
Biome clean. bun run check:i18n: two new strings ("MCP Server Commands", "Unknown command /{name}"), translated in all 16 locales.

A process a subagent leaves running drops off the activity menu when the subagent ends; the adapter sends nothing more about it. The real fix is upstream: route those updates to the main chat instead of dropping them.
The agent receives a page as its URL. A plain HTTP fetch of a page needs a signed-in viewer, so an agent without the Superset CLI cannot read it.
File and page chips come from elements stored with the message. They survive a window reload, but after a host-service restart the chat is rebuilt from the agent's own replay, which has plain text, so those messages show the path or URL as text. A leading /command still shows as a chip.
Mid-turn, finished messages have no Copy or Branch until the turn ends (the trade-off for removing the gap).
"Goal set" still reads as an agent reply. claude-agent-acp sends structured goal data only to JetBrains AIR clients; a goal banner is follow-up work.

New Features
- Chat supports mentions for cloud pages and workspace files, plus slash-command chips with descriptions and hints.
- Available commands are grouped by type, and workspace activity can identify the pane associated with background work.
- Copy and branch controls display tooltips.
Improvements
- Message timestamps show the time for today and the date for older messages, with the year when needed; hovering reveals the full timestamp.
- Tool runs are collapsed by default unless they need approval.
- Replies from still-running turns are excluded from the latest completed reply.
- Terminal pane icons can reflect the associated agent.
- Unknown slash commands are flagged, and rejected submissions keep the draft in the editor.
- Failed mention searches clear stale results and stop showing as pending.