Repository navigation
feat: dashboard home, Chat, Fleet and a Storybook catalog - #73
Merged
Merged
Conversation
The app's pages showed what the TUI acts on but sent every action to the CLI or Chat. They now carry the TUI's actions as buttons, each confirmed in the app's dialog where the TUI asks: - Inbox: Ack per item, and Ack all. - Reclaim · Processes: Kill and Force kill for verified idle trees on this machine; young, unverified and remote trees say why not. - Reclaim · Disk: Remove for ready candidates; others name the command. - A project's page: Summarize. One endpoint, /app/action, replaces /chat/api/flow-action. It checks each request against the page's snapshot (local rows only, verified idle and old enough for kills, a confirmation for every change) and runs the same actions Chat's tools run, which check again before they act. The menus gain what a Mac app and the browser offered: File > Close Window (Cmd+W), View > Back and Forward (Cmd+[ and Cmd+]), View > Inbox (Cmd+4), Reclaim Processes and Reclaim Disk, and Settings and About in the tray menu. Claude-Session: https://claude.ai/code/session_01T2vtcW6ihn1ywy9vxUq9a4
Driga's own panel format runs over the same snapshot as the app's views: ten sources (sessions, inbox, projects, orphans, cleanup, activity, memory, costs, machines, summary), filters with $variables, grouping and aggregation. The default Home dashboard is ordered by what people supervising agents need most: the inbox, finished work, runaway processes, stuck sessions, sessions to resume, then disk. Dashboards are versioned files in the data directory. Edits are typed operations checked against the version read; restoring saves a new version; an agent's replacement keeps dragged positions. driga mcp gains describe_panels, show_panels, open_dashboard, list_dashboards, list_dashboard_versions and the dashboard writes, which return fragments rendered by the ui://driga/panels.html MCP App. Chat may use them; dashboard writes need no approval card because they change only the user's dashboards and can be restored. Claude-Session: https://claude.ai/code/session_01T2vtcW6ihn1ywy9vxUq9a4
The app opens on a dashboard the user owns: Home by default, ordered by what needs them. Edit mode drags and resizes panels on a 12-column grid, and every move saves a version. Chat sits beside it and opens with the inbox and a one-line status, as if an assistant had watched while they were away. When Chat calls a panel tool, the card in the conversation runs the spec again for live data and can be added to the dashboard. The panels render with ECharts through Driga's own spec, never options from the model, and are escaped before any tooltip. The same renderer ships as one self-contained file for MCP App hosts (ui://driga/panels.html), with no eval and no network. ui/ is a Bun, React and Vite workspace; just ui builds into the committed embeds, so a Go build needs no JavaScript toolchain. The palette comes from the console theme. Claude-Session: https://claude.ai/code/session_01T2vtcW6ihn1ywy9vxUq9a4
Every view is now a dashboard. Flow, Projects and Machines ship as editable built-ins beside Home, listed as tabs; the app's menus open them. New sources (flow, waits) and panel kinds (cards, timeline) cover what those pages showed: each session drawn from start to end with its waits on you marked, and projects and machines as tiles. A sum with no readings is shown as a dash, never zero. Claude-Session: https://claude.ai/code/session_01T2vtcW6ihn1ywy9vxUq9a4
Rows keep hidden identity columns so the app can act on them. Tables, lists and cards offer the TUI's actions for their source: stop or open a session, kill a verified idle process tree, remove a ready cleanup candidate, summarize a project. Anything that changes the machine asks the app's dialog first. The inbox lists pending permission prompts with Allow and Deny buttons that arm after a moment. Claude-Session: https://claude.ai/code/session_01T2vtcW6ihn1ywy9vxUq9a4
The app is Chat with a row of metrics above it. The header is a stored dashboard of stat panels, Need you, To review, Possibly stuck, Leftover processes and Reclaimable by default, that the user or Chat can change. Every other dashboard is one the user or Chat made; opening one docks Chat beside it. Chat's brief reads the whole inbox whatever is open. The Flow, Projects, Machines and Home presets are gone, as are the htmx pages, their templates, scripts, styles and routes, the update long poll and flow.json. The View menu loses its page items and Help loses Keyboard Shortcuts. Dashboards gain create and delete in the app and a delete_dashboard MCP tool that Chat runs only after approval. Claude-Session: https://claude.ai/code/session_01T2vtcW6ihn1ywy9vxUq9a4
The app opens on Home again, a shipped dashboard under the metrics header with Chat docked on the side. A viewed dashboard leaves no blank space: list-like panels take the rows their content needs, every panel moves up, the tallest give up rows and scroll inside when content is taller than the window, and each panel then widens and grows into the free space beside and below it. Chat's brief leaves the inbox to the dashboard when it shows one, and chat cards can join the open dashboard or start a new one. Claude-Session: https://claude.ai/code/session_01T2vtcW6ihn1ywy9vxUq9a4
The app and the panels MCP App follow LabStack's Canon: Mantine components themed through ui/src/theme, Phosphor icons, TanStack Table for sortable tables, react-markdown for text panels, Mantine notifications, and one ECharts wrapper that reads chart colors from the same tokens. The tokens map Canon's surface, text and tone roles onto the console palette, which internal/view now writes as ui/src/theme/palette.json, with a contrast test for both schemes. The header is a row of stat tiles with the real brand mark and wordmark; a stat over readings in time returns its trend, drawn as a sparkline with the change over the hour. The app page carries a per-response style nonce so Mantine's theme variables load under the strict policy. Status shows as a badge with an icon and text. The leftover flow-board browser script is removed. Claude-Session: https://claude.ai/code/session_01T2vtcW6ihn1ywy9vxUq9a4
…ents One process per machine, the Driga service, collects, stores, hosts Chat and posts notifications from login, whether or not the app or the terminal UI is open. It is drigad, the same program started under that name, and `driga service run` runs it in the foreground; `driga collect` is gone. It installs as run.driga.service on macOS, the driga systemd unit on Linux and the "Driga service" task on Windows. A driga-dev build installs its own label and log, so it never takes over the release's. Installing removes the run.driga.collector agent, unit or task earlier releases installed. The app no longer runs a collector child. On launch it installs or starts the service with the login shell's environment, restarts it when it runs another version, and shows its state in the tray; quitting leaves it running. The child supervisor, the notification and lifeline pipes, the app owner kind and the service replacement are removed, and notification clicks from drigad jump through the driga CLI beside it. Releases ship drigad: macOS copies it into Driga.app and signs it before the bundle, and Linux and Windows build it beside driga. Generated third-party notices for the bundled UI replace the vendored scripts' license file. Doctor checks are renamed service, service_rss, other_services and service_definition, and messages and docs say Driga service throughout. Claude-Session: https://claude.ai/code/session_01T2vtcW6ihn1ywy9vxUq9a4
…arklines The header's metrics drop their borders and backgrounds and are set apart by space, as Visitors' and Fingerprint's stat strips are: a Phosphor icon and label, the value in mono with its unit, a badge with the change over the hour, and an ECharts sparkline whose area fades to the baseline and marks the latest reading. A count that needs the person takes its tone with a dot; a hover tint shows the tile asks Chat. Claude-Session: https://claude.ai/code/session_01T2vtcW6ihn1ywy9vxUq9a4
Chat runs in the Driga service, and it started its MCP server, jumps and session actions with its own executable: drigad mcp --chat ran another service instead, so Chat answered that the driga MCP server was unavailable. internal/self names the driga beside drigad, and every place Driga starts itself again uses it, hook paths included. On macOS a drigad started through a symlinked directory restarted itself with "service run" already added, and added it again; the restart now passes the original arguments. TestChatThroughTheService builds driga and drigad, runs the service in an isolated home with a stand-in claude that completes the MCP handshake and calls show_panels, and asks a question from the CLI. It fails without either fix. Claude-Session: https://claude.ai/code/session_01T2vtcW6ihn1ywy9vxUq9a4
Every row repeated its buttons, Open pane and Stop on each line. A row now opens its session's pane when clicked or on Enter, and one ⋯ button, shown on hover, keyboard focus or touch, lists its actions with icons: the main one first, then Stop, Kill, Force kill, Remove or Summarize in red, each still asking the app's confirmation. Tables, lists, cards and the inbox share it. Claude-Session: https://claude.ai/code/session_01T2vtcW6ihn1ywy9vxUq9a4
A failed turn's error reached the message only while the service ran, and only within the capture policy, so a failure was gone after a restart. The service log now records it, scrubbed. Claude-Session: https://claude.ai/code/session_01T2vtcW6ihn1ywy9vxUq9a4
driga dev serve serves the app's page and API over HTTP on loopback, from the data directory the environment names, with the same handlers as the app: dashboards, Chat through the Driga service, and local actions. just dev runs it under watchexec, restarting on Go changes, behind Vite's dev server, which hot-reloads the UI and proxies the API. Without the app's native dialog, confirmation falls back to the browser's. Claude-Session: https://claude.ai/code/session_01T2vtcW6ihn1ywy9vxUq9a4
Home now answers "what needs me" more honestly and with fewer clicks.
- The inbox collapses a session's finished turns into one row, and
acknowledging it clears every turn.
- Sessions show the state without the "~" marker, outline inferred
states, and explain them with the evidence in a tooltip. Unknown
memory reads "–" rather than 0 B.
- The stuck panel and the header count match "~stalled?" as well as
"stalled?".
- The memory tile shows the total ("26 / 36 GB"), and its chart axis
steps in powers of 1024.
- Clicking a header tile opens a menu: show the panel on Home, or ask
Chat about it.
- Empty panels collapse into an all-clear strip above the grid.
- Chat cards are kept whatever the capture policy, Chat notes when
earlier messages were dropped, and a failed reply points at the
service logs.
Claude-Session: https://claude.ai/code/session_01T2vtcW6ihn1ywy9vxUq9a4
Finished turns now say what there is to review, past sessions can be found and resumed, and the page says when its numbers are stale. - Each finished inbox item carries the files its session edited and its checkout's uncommitted and unpushed counts and latest commit. The inbox shows them as "To review" with a Mark reviewed button on the row. - ⌘K, or the header's Search button, opens a Spotlight over live sessions and those that ended in the last two weeks. A live session opens its pane; an ended one copies its resume command, or asks Chat when its agent has none. The board snapshot keeps that history. - "Need you" counts Chat's permission prompts as well as waits. - When a refresh fails the page says since when it is stale, and the prompts last seen stay listed. - Chat stays mounted while hidden, so a draft survives, and whether it is open is remembered. - Working is no longer green: success is kept for verified outcomes, and each state has its own icon. - Chart cards state their reading in words and can show their data as a table. Machine memory draws one line per machine, and project memory says it covers this machine only. - `just check` now runs the UI tests and fails when the committed UI build is behind ui/; CI installs a pinned Bun to run it. The build committed with the last change was stale, which this would have caught. Claude-Session: https://claude.ai/code/session_01T2vtcW6ihn1ywy9vxUq9a4
Every Chat turn failed with 'unexpected tool "mcp__driga__delete_dashboard"; Chat requires driga read tools only'. Chat refuses to start when its MCP server offers a tool it does not expect, and the server offered every dashboard tool: delete to Chat with actions, and all the writes to read-only Chat. Chat's server now registers dashboard writes only for Chat with actions, and delete_dashboard never; outside MCP clients keep all of them. The fake claude in the end-to-end test now lists the server's tools in its init event as claude does, hiding app-only tools, so the test fails without this fix. Claude-Session: https://claude.ai/code/session_01T2vtcW6ihn1ywy9vxUq9a4
- "Need you" counts machine warnings still in effect, such as memory pressure and low disk, as well as waits. The inbox lists warnings in their own Machines section with a Dismiss button; they no longer pass for finished work. - Every review row says what is left: its files, uncommitted and unpushed counts, "clean" when the checkout has nothing left over, or "no checkout". It shows the turn's summary, or else the last commit, and a visible Reviewed button. - When the board is taller than the window, the last panels give up rows first, charts down to a readable floor, so the inbox keeps its rows. A panel that grows to its content now pushes the panels below it down instead of overlapping them. - Sessions list waiting first, then stalled or dead, working, idle and finished, newest first within each. An untitled session off the main line is named with its branch. - The tray's session items open that session's pane, and the Find menu opens search. - A failed Chat reply offers Retry, and a send that never reached the service gives its text back. - Inferred states keep the one light badge style with a dashed edge; badges keep their text, long titles and branches truncate, and every bar is labelled. Claude-Session: https://claude.ai/code/session_01T2vtcW6ihn1ywy9vxUq9a4
Using Chat end to end showed four bugs:
- Chat's own claude session was collected as one of the person's agents
("You are driga Chat…" in project "chat"). Chat's prompts open with
model.ChatPreamble, and the collector now skips sessions that do.
- A panel title kept "$project" as written. Results now carry the title
with its variables filled in, or "all projects" when one is empty.
- A filter such as occurred_at gt 2026-10-25T00:00:00Z matched nothing,
since gt and lt compared numbers only; times now compare as times.
That filter made a Chat card empty, and an empty inbox card said
"Nothing needs you right now" over the card's own empty text.
- Text from consecutive model messages ran together ("gate.I didn't");
a new message now starts a new paragraph.
Claude-Session: https://claude.ai/code/session_01T2vtcW6ihn1ywy9vxUq9a4
The Chat column is 400 px wide, and its cards drew whole boards there:
tables scrolled sideways and charts were too small to read.
- A card now says in a few lines what a panel tool showed: each panel's
title and reading (a stat's value, a chart's summary, a list's row
count, or why it failed), with Open, Add to Home and Save.
- Open shows the panels in the main area, marked "Preview from Chat",
with Add to the open dashboard, Save as dashboard and Close.
- When Chat creates, edits, replaces or restores a dashboard, the card
says so: "Saved Home", its version, chips for what changed ("Added
Quota", "Removed Memory"), and Open dashboard.
- Chat is told to answer in lists and put tables in panels; a markdown
table it writes anyway scrolls inside itself.
- Bars grouped by session id are named by the session's title.
Also fixes cards that broke after a restart under capture_text =
"snippets": a reopened conversation cut each card to a snippet, which
is not JSON. Cards are now kept whole when loaded, as when saved.
Claude-Session: https://claude.ai/code/session_01T2vtcW6ihn1ywy9vxUq9a4
Chat cards no longer have Add to Home and Save buttons, and neither does the Chat preview. As in Fanout, a view is kept by asking Chat, and the card for that write is the record of it. A panel card now waits for its tool's result. A dashboard write's card names the dashboard and version it saved, which claude passes as the tool's structured result, so Open dashboard opens what was saved rather than racing the save. A rejected draft says why it was not saved. Claude-Session: https://claude.ai/code/session_01T2vtcW6ihn1ywy9vxUq9a4
As in Fanout, Chat stays where it is and its results open inside the
conversation rather than in the main area.
- A card is collapsed to a few lines, the newest one expanded. Expanded,
its panels stack full width to fit the Chat column, and Full screen
shows them at full size in a dialog. The main-area preview is gone.
- Chat can delete a dashboard. Delete is not pre-allowed, so each one
waits on an approval card; the other writes still run straight away.
- Text panels render GitHub tables (remark-gfm 4.0.1), and cards count
and describe text panels instead of reporting "0 panels".
- A turn's reply shows after its tool rows and cards, as the answer
that ends the turn.
- Chat's own claude session is also recognised by its working
directory, so it stays out of Sessions before its first prompt is read.
- An empty action result, such as delete's "{}", is not shown.
Claude-Session: https://claude.ai/code/session_01T2vtcW6ihn1ywy9vxUq9a4
Chat that could act started claude with every built-in tool, so one Allow could run any shell command Chat wrote, or change any file. It now gets only Read, Glob, Grep and TodoWrite. Everything Chat changes goes through Driga's own tools and their approval cards. Chat refuses to start when claude offers Bash, Edit, Write or any other built-in tool. A conversation that used Bash before may have its model call it again. claude refuses a tool it did not offer and the model goes on, so the turn no longer ends; a refused tool that was offered still ends it. A card for a dashboard deleted since says so instead of "dashboard not found". Claude-Session: https://claude.ai/code/session_01T2vtcW6ihn1ywy9vxUq9a4
Chat was a fixed 400 px column. The board and Chat now share the window as resizable panes (react-resizable-panels 4.14.4): drag the line between them, or focus it and use the arrow keys. Chat's width is remembered, and hiding Chat collapses its pane, which keeps the conversation and any draft, then shows it again at the same width. Claude-Session: https://claude.ai/code/session_01T2vtcW6ihn1ywy9vxUq9a4
Driga's UI now has a local Storybook, following Canon's design review convention and monk's setup: `just design` serves it on 6506 for review, never published, and `just design-check` runs every story, its play, accessibility (failures are errors) and Canon's layout check in Chromium, light at desktop width and dark at laptop width. `just check` and CI run it. - Components/Panel has one story for each kind of panel, with loading, empty, error and narrow states; Home's own panels, the header metrics, the board, Chat, Chat cards, search, row menu, status badge, measure, chart and brand each have theirs. Pages/Home renders the real app; Flows/ walks through answering a prompt, reviewing finished work, asking Chat for a dashboard, and finding and resuming a session. Embedded/Panels app is the MCP App's view; Theme/Tokens the palette. - Stories render the shipping components. Their data is Driga's own panel engine run over a fictional fleet (internal/panel/stories_test.go; DRIGA_WRITE_STORIES=1 regenerates it and the test fails when it is stale), and a fake service answers the app's requests; any other request fails the story. Time and storage are pinned per story. - A unit test fails when a component or a kind of panel has no story, and when app code imports a fixture. - App moves out of main.tsx, and the MCP App's view out of its host connection, so stories can render them. The catalog found real bugs, fixed here: - Truncated text had no way to read it in full: inbox rows, table and card cells, panel titles, Chat card lines and the page description now carry their full value. - At narrow widths inbox rows overflowed and clipped their buttons; the header metrics clipped their trend, and their row could not scroll. Trends and sparklines now show only on wide windows. - A gauge's "of" line was clipped in its default height; a text panel clipped its content instead of scrolling. - Light status colors missed AA on the inset surface Chat sits on (red was 4.16:1). Light text roles now pass on crust, the darkest light surface, in the console palette too, and the theme test checks every surface. - Header tiles and the search button were named differently from what they show; the tile menu's description broke the menu's roles; scrollable panels could not be reached by keyboard; notification close buttons had no name; the search dialog had none and now has a title; two "Sessions" panels made duplicate landmarks, so panels are articles. Claude-Session: https://claude.ai/code/session_01T2vtcW6ihn1ywy9vxUq9a4
Until the open dashboard arrived, the app took it to lack an inbox, so Chat's brief showed its own copy of the inbox and then removed it when Home came in; a click in that moment hit a button that was going away. While the dashboard is unknown the brief now shows no copy. A Pages/Home story holds Home loading and fails without the fix. Claude-Session: https://claude.ai/code/session_01T2vtcW6ihn1ywy9vxUq9a4
A catalog test now loads every story file and every app module, and fails when a story's component is not one the app exports, so a story cannot show a copy or stand-in of a screen. The theme's token sample is the one listed exception. Pointing the Brand story at a stand-in fails it. Claude-Session: https://claude.ai/code/session_01T2vtcW6ihn1ywy9vxUq9a4
Adds @xyflow/react 12.12.0 and elkjs 0.12.0 for a draft Fleet view, the Drafts sidebar group, and a catalog test that keeps app code from importing drafts. Claude-Session: https://claude.ai/code/session_01T2vtcW6ihn1ywy9vxUq9a4
Two presentational drafts of the 56 px app header under ui/src/drafts/header, for the owner to compare in Storybook with today's header: - ClustersHeader: attention, hygiene, this-machine memory and fleet pulse clusters set apart by hairlines; tiles tinted with a rule and a dot above zero, one pulsing waiting dot, quiet zeros, one all-clear line when nothing needs you, a memory sparkline with a delta chip and a pressure bar, one dot per machine, and the keys 1, 2, 3 and g f. - AttentionHeader: three large counts with their hour faded behind them and a delta with per-metric polarity, resources demoted to a quiet cluster, one chip per machine with a 30-poll tracker strip, on a bar one surface step dimmer than the content. - DeltaChip, the change chip both share; Header.ts, the contract both take; fixtures of a fictional fleet; stories per variant (busy, all clear, one stuck, many machines, laptop width, filtering, keyboard focus) and a Compare page with both drafts and today's header. Claude-Session: https://claude.ai/code/session_01T2vtcW6ihn1ywy9vxUq9a4
…o choose between Two drafts of a Fleet view under ui/src/drafts/fleet, both over one typed contract (machines, projects, sessions with state, agent, memory, parent, files, last tool and waiting reason; onSelect, onOpen and onReply) and one fictional fleet of three machines, five projects and fourteen sessions. A, Sankey: a machine strip, then an ECharts sankey from machines through projects to sessions where link width is memory and node colour is state, with a fixed node order, adjacency on hover, yellow paths for waiting sessions and leftover pseudo-nodes per machine; a Tree mode as an accessible table; an inspector beside it. B, Topology: React Flow with ELK's layered layout, machine boxes holding project boxes holding session cards with a status ring, agent chip, 30-turn strip, memory and Reply or Open; spawn edges and red dashed same-file edges; hover isolates a neighbourhood; colours follow the Mantine scheme through the --xy-* variables; nodes are keyboard-focusable; the canvas loads lazily and the minimap is off by default. Shared parts (header, machine card and strip, inspector, tree, turn strip, empty state), a Compare page that stacks both, fixtures, unit tests for the model, and a story file per component with plays, Docs descriptions and the page stories pinned to the desktop width. Claude-Session: https://claude.ai/code/session_01T2vtcW6ihn1ywy9vxUq9a4
Sixteen widget drafts under ui/src/drafts/widgets, each a component with typed props, fixtures beside it and a story file with a play on every story: KPI tile, bar and ring gauges, chart card with peak, wait and event annotations, status badge, agent chip, inbox item, session card, machine card with its 90-minute strip, tracker bar, category bar, bar list, donut, heatmap, session swimlanes and empty state, plus a Gallery that lays them on one board. Colors come from the theme; the one palette module holds the validated series palette and the agent tints. Lint, the unit tests and the catalog check pass in both schemes. Claude-Session: https://claude.ai/code/session_01T2vtcW6ihn1ywy9vxUq9a4
… drafts The chosen drafts become the app: the clusters header replaces the metrics row inside the shell, the Fleet becomes a page the app opens (fleet pulse, g f, Search, drigaCommand view fleet) built from what the service already sends, and fourteen widgets replace today's inbox rows, status badges, empty states, gauges, bar and timeseries panels and the timeline. One vocabulary across the three sets: working green, waiting yellow, stuck red, idle blue, finished neutral with the check; one status badge, one machine card, one tracker strip, one delta chip in the neutral tint; tint tokens in the theme and a Latte hover that keeps muted text at AA; the chart module registers every chart kind and takes an onEvent; the series slots are re-stepped Catppuccin hues validated for color blindness. The header B, Fleet B (React Flow and ELK), Donut, Gallery and the merged duplicates are deleted. The catalog runs in all four scheme and width cells. Claude-Session: https://claude.ai/code/session_01T2vtcW6ihn1ywy9vxUq9a4
Mantine's ScrollArea paints its bar inside the viewport, so the right-aligned values in a bar list or a table's last column sat under it once the content overflowed, as in Home's "Memory by project" in the desktop app. Every scroll area now offsets its content while a scrollbar is present, set once in the theme. The catalog gains a check, beside Canon's layout check, that fails a story when visible text lies under a vertical scrollbar, measuring wide tables at their scrolled end. A new Panel story overflows a bar list on purpose. Claude-Session: https://claude.ai/code/session_01T2vtcW6ihn1ywy9vxUq9a4
…g panels A card's Full screen and Open dashboard were a row of text buttons under its panels; they are now icon buttons with tooltips at the right of the card's header, as panel actions are elsewhere in the app. The header no longer says how many panels a card holds, since the panels below say what it shows; it says only how many failed and, once saved, the version. A created or replaced dashboard's note reads Created or Replaced. Claude-Session: https://claude.ai/code/session_01T2vtcW6ihn1ywy9vxUq9a4
Sample CPU busy share and persist disk totals across the service and Fleet. Serve attention and machine poll history, captured turn outcomes and relative file paths, project tool activity, and targeted desktop notification clicks. Claude-Session: https://claude.ai/code/session_01T2vtcW6ihn1ywy9vxUq9a4
The Busy story ended by unhovering the brand mark, which left the pointer logically over the last button clicked; its tooltip opened 300 ms later, after the check had passed, and failed the layout check whenever a loaded run was slow enough to measure it. rest() now unhovers every control and waits out the tooltip delay before asserting none is open. Claude-Session: https://claude.ai/code/session_01T2vtcW6ihn1ywy9vxUq9a4
Keep retained events and prompt history scoped and bounded, preserve machine and session identity in panels, and enforce dashboard write contracts. Use stuck consistently and share attention counts, CPU, disk and session details between terminal and app views. Claude-Session: https://claude.ai/code/session_01T2vtcW6ihn1ywy9vxUq9a4
Require bound one-use confirmation tokens and trusted request origins, keep notification targets out of process arguments, and restrict navigation. Repair startup dispatch, service installation ordering, Chat connections and snapshot cache cancellation, with regression coverage for forged and stale requests. Claude-Session: https://claude.ai/code/session_01T2vtcW6ihn1ywy9vxUq9a4
Guard stale loads and approvals, preserve unknown readings, and fix chart and dashboard lifecycle, accessible controls and machine-scoped Fleet selection. Update story names, selection payloads, overlap cases and the minimum-width legend to match those contracts. Include rebuilt app and MCP bundles. Claude-Session: https://claude.ai/code/session_01T2vtcW6ihn1ywy9vxUq9a4
Document the separate service lifecycle, editable dashboards, notification routing and stuck vocabulary. Refresh generated references and add checks for the corrected public instructions. Claude-Session: https://claude.ai/code/session_01T2vtcW6ihn1ywy9vxUq9a4
Keep generated JSON checkouts on LF and bound the Chromium dependency installation step. Cover checkout conversion and the installation timeout with regressions. Claude-Session: https://claude.ai/code/session_01T2vtcW6ihn1ywy9vxUq9a4
Give header KPI labels and values room for Linux Plex Mono metrics while keeping the 56 px shell. Keep Fleet inspector headings clear of its scroll viewport, add a tall-font regression story, and rebuild the UI bundles. Claude-Session: https://claude.ai/code/session_01T2vtcW6ihn1ywy9vxUq9a4
macOS carries drigad inside Driga.app, so the driga archive has one binary there and two on Linux; allow that. On Windows drigad.exe goes into the zip as a plain file, like Driga.exe, so Scoop still shims only bin/driga.exe. Claude-Session: https://claude.ai/code/session_01T2vtcW6ihn1ywy9vxUq9a4
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
What changed
The desktop app becomes a dashboard home with Chat docked beside it. Panels come from a dashboard spec and a panel engine shared by the app, the MCP tools and Chat. The header shows the six attention and hygiene counts. A new Fleet page shows a sankey, a session tree and an inspector. Machine cards show CPU, memory and disk. Every component, panel, page and flow has a Storybook story, and the catalog runs those stories as tests. The TUI now says the same things as the app in the same words: one
stuckstate, the same header counts, CPU % and disk, and per-session turn strips and files.A full code review of the branch was applied on top. The fixes include:
Why
The owner asked for a richer dashboard and header than the plain board. Driga needed one renderer for the app, Storybook and MCP Apps so a component cannot drift between them. Drafts were built side by side in Storybook and chosen there. Rejected variants were deleted rather than kept: the topology Fleet with React Flow and ELK, the attention-bar header, and the donut and gallery widgets.
Need you counts sessions blocked on the user (a permission or a question), not Claude sessions sitting at their prompt after a turn. Those are idle, and To review already lists finished turns. The inbox, header, board and pulse now share one live-wait rule, so the counts agree.
Risk and rollout
0004_panel_samplesaddsdisk_totalandcpu_percenttomachine_samples, and creates thepanel_polls,prompt_intervalsandpanel_eventstables.0005_prompt_sessionsaddsprompt_intervals.session_id.~/.claude,~/.codexor other agent settings.capture_text, and no file contents are stored.Contract changes
stalled_after/DRIGA_STALLED_AFTERis nowstuck_after/DRIGA_STUCK_AFTER, with no alias (alpha). Users who set the old key must rename it.stalledsession label is nowstuck.internal/mcp/panels.go) and the panels MCP App (internal/mcp/apps/panels.html).Verified
All runs below were on macOS (host, not sandboxed), at
05d172e:just check, including 1,008 Storybook stories (plays, axe, layout) and 140 UI unit tests.TZ=UTC scripts/isolated.sh go test -count=1 ./....scripts/isolated.sh go test -race -count=1 ./....GOOS=linux go build ./....gitleaksover the branch: no leaks.https://claude.ai/code/session_01T2vtcW6ihn1ywy9vxUq9a4