Dashboard project activity
Dashboard opens a mosaic of activity cards. a opens activity and e opens
Attention; neither needs to load a complete event or session history. The same
React components render in the TUI, browser and desktop app.
Each verified GitHub organization gets one card for its connected repositories. Repositories owned by personal GitHub accounts get separate cards. GitHub's public owner profile API supplies the actual owner type; names and group slugs do not establish it. Unknown owners and other Git providers stay separate. Cards never combine different devctrl users' data.
Cards are ordered by the selected metric’s latest recorded activity, newest first. They show the exact
UTC date/time and elapsed time, agent CLIs used, native token totals, non-merge
commits, and added/deleted lines. Periods are today, the last 7 UTC calendar
days, and the last 30 UTC calendar days, including today. They are not
rolling 24-hour windows. A 91-day contribution graph can display tokens (t) or
commits (c). Last tokens shows the latest recorded native token activity;
Last commit shows the latest reachable Git commit. Switching metrics changes both
the timestamp and sorting. Session starts, status changes and polling do not count
as token consumption or commits.
Token totals use k for thousands and m for millions (1.2m, rather than
1200.0k). Unfilled graph cells consistently use · for no recorded activity;
green squares show increasing activity. An unfilled cell does not establish zero
when coverage is incomplete. The card explicitly labels complete, incomplete or
unavailable history. ≥ identifies a recorded lower bound; — means unavailable.
Complete Git counts stay complete while refreshing, and a failed token scan does
not make an otherwise complete Git scan partial.
Projects without a Git checkout still collect native Claude/Codex token history. The card labels them Workspace · no Git repository and leaves commits and line deltas unavailable. Git and native usage collection are independent: a failure in one cannot discard the other. Each graph also shows its recorded 91-day total, so older work remains visible even when it falls outside Today, 7d and 30d. A dash in those shorter periods means no reliable total is available for that period; it does not erase recorded older usage or establish zero. Native histories remain scoped to exact owned accounts, workspace state and provider session IDs.
Dashboard visibility
Select a card and press h (Hide) to remove its projects/repositories from your
Dashboard. Press v or choose Hidden to see hidden cards, then h (Restore)
to show them again. An organization card hides/restores all repositories currently
included in that card. Press Enter or choose Repositories to open that
organization’s individual repositories and hide/restore only the selected one.
Escape returns to the organization overview. Hidden repositories are excluded
from visible organization totals and counted only in Hidden; they still collect
activity normally. This never stops a session, removes a project, or changes
Projects visibility.
Preferences persist on the controller per signed-in user and work across TUI,
browser and desktop. Operator preferences do not change another user’s Dashboard.
PUT /v1/dashboard/activity/visibility accepts {projectIds, hidden} only after
checking every project’s ownership; mixed authorized/unauthorized batches change
nothing. Operator MCP can discover and call the same operation. Session MCP
get_project_activity remains a read-only view of its own repository, independent
of Dashboard visibility.
With content focus, arrows select cards and Page Up/Down moves through the visible
mosaic. n and p load older/newer pages; r checks cached summaries again.
Escape returns to Attention from the overview. Tab moves focus to the navigation
without changing the activity view or selected metric. The legend and action bar
stay at the bottom of the content area while cards scroll independently. Web cards scroll; small terminals explicitly request
more height rather than silently hiding a card.
Collection and freshness
GET /v1/dashboard/activity?limit=12&cursor=0&metric=tokens&view=visible returns authenticated, user-scoped
summaries, at most 24 cards per page. metric=tokens|commits controls Last and ordering;
view=visible|hidden filters saved visibility. group=<organization-card-id> opens
its individual repository cards. Counts report visible and hidden repositories
within the selected scope. It immediately returns saved data and queues
at most two read-only scans. Each repository is scanned at most once every five
minutes while Dashboard is open. The host executes accepted scans even if the
interface disconnects. No provider session or Project container is restarted.
Execution hosts poll a narrow, authenticated activity-command queue every two seconds, independently of Docker inventory and storage reconciliation. The poller uses the existing authority fencing and command dispatcher; a poll, WebSocket message and general queue cannot run the same command twice. Before reading a checkout, the host rechecks its exact current command and workspace. The controller also rejects late snapshot writes after cancellation, deletion, relocation, ownership changes or a changed checkout path. At most two activity scans are admitted globally, with one scan executing at a time on each host. Other command types, including urgent stops, keep their independent dispatch. No terminal or browser needs to stay connected.
The host reduces Git records to at most 91 daily rows before sending them to the
controller. SQLite stores compact per-repository snapshots separately from
/v1/state. The interface polls this small endpoint every five seconds, cancels
requests and timers on exit, and updates relative times once per minute. Offline
hosts retain their last snapshot with partial coverage after five minutes;
missing data remains unavailable. An offline queued scan does not keep a spinner
running indefinitely.
The session MCP tool get_project_activity reads only the current repository's
cached card. It does not reveal sibling repositories from its organization or
start a separate scan. There are no caller-supplied project selectors.
Counting rules and limitations
- Git scans use local refs, including unmerged branches, and deduplicate commit SHAs within a repository. They do not fetch GitHub or attribute human commits to an agent. Merge commits are excluded from contribution counts and line deltas; their timestamp can still establish repository activity. Binary changes have no invented line count. Uncommitted working-tree changes are not contributions.
- Scans read at most 5,000 commits and 2 MiB within an eight-second deadline.
Shallow history, truncated output, missing repositories and failures are explicit.
Latest reachable Git activity may predate the 91-day graph.
Commit IDs and dates are collected before calculating file diffs. A diff timeout
or output limit cannot discard already counted commits. Line totals for affected
days/periods are unavailable (
—), rather than invented zeros; commit coverage and optionallineCoverageare reported independently. Large repositories can therefore have complete commit counts while their line statistics are unavailable. - Native token history includes recorded input and output, with cache input counted once. Claude streaming message updates are deduplicated by native message ID. Codex cumulative counters become deltas; inherited parent history in a fork is not counted as new usage. Confirmed child transcripts are included only when their native parent relationship can be verified.
- Token scans inspect only the exact project's selected Claude/Codex native conversations, account roots and verified children: at most 24 parent sessions, 4 MiB per parent conversation together with its children, and eight seconds per scan. Large histories use their recent tail; Codex's first tail counter becomes a baseline, never a fresh lifetime charge. Missing, deleted, compressed, incompatible or byte-limited history is partial or unavailable. Other CLIs may appear in the agent list without token accounting. Copied histories are deduplicated conservatively and marked partial; complete billing reconstruction is not promised.
- Tokens are not money or subscription credits. No API bill or savings estimate is fabricated. A lower bound is useful evidence of work but is not a billing report.
- The selected activity timestamp comes from native usage or Git records. Session creation/termination, lifecycle changes, polling and host heartbeats do not manufacture new activity. Last activity is not a claim that an agent is working.
- Snapshots contain counts, timestamps and CLI identifiers; no prompts, terminal output, commit messages, authors, file paths or secret values are sent to clients.
Owner metadata is checked asynchronously by the optional GitHub provider package. The public account endpoint identifies personal and organization accounts without requiring access to a private repository. Successful lookups are cached for a day; unavailable metadata retries after an hour. A private repository can use only its immutable owner's Git credential, preferring the account explicitly bound to that repository. Public metadata can be read without a credential. Existing user and session access boundaries apply to every API and MCP read.
Browser regression check
After building the TUI/web bundle, run the opt-in
node apps/tui/web/dashboard-activity.browser.integration.mjs with Playwright
installed. DEVCTL_PLAYWRIGHT_MODULE and DEVCTL_BROWSER_EXECUTABLE can point to an
existing installation. It uses an isolated fixture, tests desktop/mobile hide and
restore, metric-specific Last, fixed footer geometry and Attention navigation,
and writes screenshots to a temporary directory. No paid accounts are required.
Navigate the full activity list
A terminal renders only the cards that fit. The range indicator shows the visible window within the loaded page; arrows, Page Up/Page Down and the mouse wheel reach the remaining cards. Newer/Older navigate API pages. Browser/app render all loaded cards in a scrollable grid. Compact terminal cards retain the metric table and seven-day calendar rows. Doctor uses a compact summary so it does not displace the activity view. See Deployment map for placement views.