PR #33104 - manaflow-ai/cmux

Leo, 2026-11-03: "the "running" bar above the composer is currently no longer usable bc it doesn't display much". His capture showed the folded bar reading only 1 subagent running · 1 command 4s ›.

What changed

The running tray's folded header now names what runs, as Claude Code's own running-tasks list does: a subagent by its task, a command by its description (or its short command when it has none), a started chat by its title, newest first.

One subagent plus one command: Watch the tests · Explore how the daemon updates itself 4s ›
Five running: the newest two by name, then +3 more. The bar stays one row.
A narrow tray (400px or less, measured) names only the newest. The count and the time never truncate; the names give way first.
Opening it (the ›) works as before. Open, the lines name each item, so the header goes back to its counts.

Regressed by #29885 (merged 2026-11-01), which made every headed tray start folded to its header (for #28875) while that header said only counts. This keeps that fold and its one-row height and puts the names back on it. Takes back #29885: only the count-only folded header text; the default fold stays.

Spec: leo-workspace design/running-tray.md (updated in its own PR). References used: claude/live/panel-context-and-background-tasks and claude/live/session-running-with-side-panels.

Gallery: agent-pane.subagent-tray-names with variants one-plus-one, five and long-narrow, plus an experiment subagent-tray-fold-head with three arms: two (the default: newest two, then +N more), one (newest one, then +N more) and counts (as before). I picked two because it names both items in Leo's case and still fits on one row. The SubagentTray gallery's one-agent-two-commands variant now opens the tray before it reads the counts.

Performance: only the newest two items are labelled, in a useMemo that runs when the items change, never on the one-second clock tick. The 300-agent folded test now allows at most 10 name reads (it was 0), and a rerender adds none. No new timer, observer or per-token work. The count string reuses the existing tools.more ("+{n} more"), so there are no new localized strings.

Test

webviews/test/subagent-tray-names.test.tsx, committed first (62a2a9981ff), checks that the folded bar names the subagent's task and the command's description, that a command with no description gets its short command, and that five running show the newest two plus +3 more.
Command: cmux-ci test --setup webviews -- bash -c 'cd webviews && bun test test/subagent-tray-names.test.tsx'

Base (62a2a9981ff, test only): RED_RESULT
Head: GREEN_RESULT

Observed / Not run

Observed: bun test test/subagent-tray-names.test.tsx test/subagent-tray.test.tsx gave 45 pass, 0 fail at the head. scripts/web-lint.sh (with CMUX_LOCAL_HEAVY_CHECKS=0) and scripts/cmux-next/prepush-gate.py both passed. A fleet run of bun run check plus the tray suites (trailing, columns, composer geometry, expand shift, running strip, bottom-stack edges, motion, dead clicks) gave: FLEET_RESULT
Not run: I have not looked at the gallery captures myself; the advisory gallery run renders them. No DEV build.

Changelog

Fixed: The running bar above the composer names the subagents and commands that are running, not just how many there are.

Agent session: agent session

🤖 Generated with Claude Code

^{Open in [code]smith · Need help on this PR? Tag @codesmith-bot with what you need.}

Summary by cubic

Fixes the running tray header above the composer, which showed only counts when folded (e.g. 1 subagent running · 1 command), so it now names what runs. This restores names to the fold introduced in #29885 while keeping the one-row folded height.

The folded header names the two newest running items, newest first: a subagent by its task, a command by its description (or its short command when it has none), a chat by its title, then +N more.
Narrow trays (≤400px) name only the newest item; the count and elapsed time never truncate, the names give way first.
Opening the tray shows the full list and the header returns to counts.
The behavior sits behind an experiment (subagent-tray-fold-head) with arms two (default), one, and counts for comparison.
Only the newest two items are labeled, computed when the items change and never on the one-second tick.