Activity Dashboard

A live kanban of every task your agents are working on - Active, Completed, Errors - with drill-down to every tool call.

The Activity Dashboard is a live view of what your agents are actually doing. Every thread your agents touch becomes a task; every reply becomes an activity; every MCP tool invocation becomes a tool call. The dashboard rolls that up into a kanban so you can see in-flight work, recently finished work, and errors at a glance.

Open it from the sidebar under Activity. The board polls every 4 seconds, so fresh work shows up without a reload.

Columns

Column	What’s in it
Active	Tasks with at least one activity still in_progress. Capped at 200 rows - the column is expected to stay small since in-flight activities resolve to success or error.
Recent	Tasks with no in-flight activity and no errored activity. Paginated, newest first.
Errors	Tasks with at least one activity in status='error'. Paginated, newest first.

Each card shows the platform, channel, the initiator, the initial agent, an activity count, and relative time since the last activity. The stats strip above the board labels the Recent bucket as Completed - same tasks, shorter label.

Filters

Narrow the board with the controls above the columns:

• Agent - show only tasks that involved a specific agent.
• Window - 1h, 5h, 24h, 7d, or 30d against each task’s last_activity_at.

Filters are URL-state, so you can link directly to a filtered view - e.g. /activity?agent=<agent-id>&window=24h. The Activity button on the agent detail page uses this.

The API also accepts a user=<initiator-user-id> filter; there’s no UI control for it yet, but you can hit /api/activity?column=active&user=U01234 directly.

Stats strip

Above the board, four cards summarize the current filtered view:

• Active - tasks with in-flight work.
• Completed - finished without any errors.
• Errors - tasks with at least one errored activity.
• Total - sum of the above.

The Active card’s subtext shows N agents working - scoped to the current filter (1 when you’ve filtered to a single agent with in-flight work, zero otherwise).

Usage tab

Activity has two tabs: Tasks (the kanban above) and Usage - the latter is superadmin-only. The switcher only renders for superadmin sessions; admins, editors, and viewers see just the Tasks board, and /api/activity/usage returns 403 for any non-superadmin call. Enforcement is server-side, so direct navigation to /activity/usage by a non-superadmin loads an empty page.

Usage shares the same layout as Tasks - tabs → summary cards → filter → detail - but instead of kanban columns it shows token consumption rolled up for the selected window:

• Totals strip - five KPIs across the selected window: Input, Output, Cache read, Cache written, Turns. All formatted at human scale (1.2M, 340K, or raw digits).
• By agent - horizontal bars ranking agents by input + output tokens in the window. Each row shows the agent name, a proportional bar, and the per-agent breakdown (in · out · turns).
• Power users - top-10 leaderboard ranked by distinct tasks started, with turn count and total tokens as sub-stats. Ties broken by turn count descending.

Token counts come from the Claude Code SDK’s per-turn usage payload (input_tokens, output_tokens, cache_read_input_tokens, cache_creation_input_tokens) and are persisted to the activities table on each reply. Rows that predate this feature show NULL tokens and are treated as 0 in aggregations - expect the first few days after you enable it to under-report.

Check remaining

Usage shows what you’ve spent, not what’s left. The Claude Agent SDK doesn’t expose the subscription-level quota that Claude Code’s interactive /usage command renders, so a Check remaining link in the page header opens claude.ai/settings/usage in a new tab for the authoritative “remaining” number.

Refresh cadence

Usage does not poll - it re-fetches when you change the agent or window filter, or when you click the Refresh button in the header. The kanban on Tasks still polls every 4 seconds for live state; Usage is a reporting view where live polling isn’t worth the requests.

Task detail

Click any card to open /activity/<taskId>:

• Full timeline of activities, newest last.
• Each activity expands to its tool calls, with args and result previews.
• Deep-link button jumps back to the original thread. The label adapts to the platform - Open in Slack, Open in Telegram, etc. Platforms without a public deep-link scheme today (Discord, WhatsApp, Teams) hide the button.

Enabling

Activity capture is gated by a feature flag. If the flag is off, the writer hooks short-circuit - no tasks, activities, or tool calls are recorded, and the dashboard is empty.

# in .env
ACTIVITY_DASHBOARD=1

Restart the runner after toggling:

slackhive stop && slackhive start

Access control

Who sees what on the dashboard depends on role:

Role	What they see
superadmin / admin	Every task across every agent.
editor	Only tasks that touched an agent they created or were granted access to via agent access.
viewer	Blocked entirely - the sidebar link is hidden and all three activity API routes return 403.

Enforcement happens server-side on every request, so URL tampering (e.g. passing ?agent=<someone-elses>) can’t bypass it. Agent access is managed on each agent’s detail page → Access (admin/superadmin only).

Retention

Tasks, activities, and tool calls live in the same SQLite database as everything else (~/.slackhive/data.db). There is no automatic pruning today - back up the DB if you want a long-term record, or truncate the tasks, activities, and tool_calls tables directly when you need to reclaim space.

Related

Logs

Per-agent log stream - the raw runner output behind each activity.

Version History

Snapshots of the agent’s configuration at the time of each run.