An unofficial, community-built tool. Not affiliated with or endorsed by Anthropic.
A live, theme-native dashboard for Claude Code workflow runs. It reads the on-disk run journal and per-agent transcripts that every workflow writes, and renders a continuously-updating view of the loop, every agent, and any review findings — right inside VS Code.
It is workflow-agnostic: the journal/transcript format is identical for all Claude Code workflows, so the dashboard discovers and visualises whatever run is active without any per-workflow setup.
Screenshots
| Dark theme | Light theme |
 |
 |
Agents tab
 |
 |
Findings tab
 |
 |
Timeline tab
 |
 |
Screenshots regenerated with
npm run screenshots(headless Chromium, no VS Code required).
Features
- One-click access — a live status-bar item (bottom-left, shows a pulse
or circuit-board icon depending on whether agents are active) shows the current
phase and agent counts; click it to open the dashboard. There's also an
Activity Bar icon and the keybinding
Ctrl/Cmd+Alt+W. - Tabbed layout — seven tabs (Agents, Findings, Verdicts, Changed, Charts, Timeline, Results) below an always-visible Overview strip. Tab selection and per-tab scroll position are remembered across refreshes.
- Loop & agent overview — phase, pass number, live/done/dead/superseded agent counts, total output tokens, tool-calls, and findings with a severity breakdown.
- Agent sub-windows — one card per agent (role, status, elapsed time, token & tool-call counts, current activity). Click a card to expand its output: a reviewer's findings, a structured result (pretty-printed, e.g. a build result), or the live text / tool-call tail of a working agent.
- Findings — every
findings[]result aggregated, with per-role and per-severity filter chips; click a finding to read its Why and Fix. Paginated at 50 per page. - Timeline — Gantt-style lanes: one row per agent, bars from start to finish (live agents extend to "now"), colored by status. Hover or click to focus an agent card.
- Charts panel — a horizontal token-bar chart (per-agent output tokens) and a cumulative-token trend sparkline. Dependency-free hand-rolled SVG; theme-native colors.
- Per-agent token & tool-call metrics — each agent card shows output tokens, input tokens (when available), cache read/write tokens, tool-call counts, and finding counts in a compact metrics bar.
- Structured results, Verdicts, and Changed files panels.
- Typed result renderers — each known agent type (implementer, verifier, reviewer, completeness-critic) gets a tailored, readable result view — never a raw JSON dump. Unknown shapes fall back to a generic key-value table.
- Superseded agent detection — when the workflow engine retries a stalled agent, the zombie is shown as "superseded" and excluded from the live count.
- Live updates — refreshes instantly on workflow file changes (and on a configurable interval), preserving your scroll position and active tab.
- Run picker — pin any recent workflow run via the Select Workflow Run… command (view-title icon or Command Palette). Default remains "Follow newest"; choosing "Follow newest" at the top of the picker unpins.
- Export Run as Markdown — the Export button (or "Claude Workflow: Export Run as Markdown" command) saves the full run report — findings, verdicts, structured results, and agent metrics — as a Markdown file. A Save dialog or Copy to clipboard option is offered.
- Prompt disclosure — click the Prompt toggle inside any agent card to read the agent's opening system prompt, with a Copy button.
- Collapse all / Expand all — a button in the Agents panel header collapses or expands every agent card at once. Individual cards still fold/unfold independently.
- Keyboard + ARIA — WAI-ARIA tabs with roving tabindex, ArrowLeft/Right/Home/End navigation; state is never communicated by color alone.
Getting started
Open the dashboard any of these ways:
| Method | How |
|---|---|
| Status bar | Click the Workflow Dashboard status-bar item at the bottom-left |
| Activity Bar | Click the Claude Workflow icon, then the Dashboard view |
| Keyboard | Ctrl+Alt+W (macOS: Cmd+Alt+W) |
| Command Palette | "Claude Workflow: Show Dashboard" or "… Open Dashboard in Editor" |
The dashboard auto-discovers the most recent workflow run under ~/.claude/projects
by default — no configuration required. To pin a specific run, use
"Claude Workflow: Select Workflow Run…" from the Command Palette or the
view-title icon in the Dashboard sidebar pane.
Authoring workflows for this dashboard
To get the richest view (structured findings, verify results, clean role labels), write your workflow scripts following WORKFLOW-AUTHORING.md. It's written so a Claude Code session can follow it directly.
Open it without leaving the editor:
- The 📖 Guide button in the dashboard's top bar, or the book icon in the Dashboard view's title bar.
- Command Palette → "Claude Workflow: Open Workflow Authoring Guide".
Settings
| Setting | Default | Description |
|---|---|---|
claudeWorkflow.workflowsGlobBase |
~/.claude/projects |
Base dir searched recursively for the newest wf_* run. |
claudeWorkflow.repoDir |
first workspace folder | Repo whose recently-changed files appear in the Changed files panel. |
claudeWorkflow.refreshMs |
4000 |
Fallback refresh interval (it also refreshes on file changes). |
claudeWorkflow.statusBar |
true |
Show the live status-bar launcher. |
claudeWorkflow.roleRules |
[] |
Optional {re,label,key}[] to label agents per workflow. agentType from agent-*.meta.json is the primary label source; roleRules + classify() is the fallback for agents without a known agentType. |
How it works
For each run the dashboard reads:
journal.jsonl— onestarted/resultrecord peragent()call.resultpayloads are interpreted by shape: an array underfindingsbecomes the Findings panel; any other object becomes a Structured result; a string is the agent's text output.agent-*.jsonl— each agent's transcript, used for token counts, tool-call counts, the current-activity line, and the expandable output tail.agent-*.meta.json— theagentTypefield drives role labelling and typed result renderers.
Agent status is derived as done (a result was recorded), run (its transcript changed within the last 3 minutes), dead (interrupted — no result and no recent activity), or superseded (a newer same-role agent was spawned before this one produced a result).
Nothing is written; the extension only reads these files.
Build & install from source
npm install -g @vscode/vsce # once
vsce package # -> claude-code-workflow-dashboard-<version>.vsix
code --install-extension claude-code-workflow-dashboard-*.vsix --force
Or press F5 in this folder to launch an Extension Development Host.
Regenerating screenshots
npm run screenshots # runs make-sample-run.mjs then screenshot.mjs (headless Chromium)
npm run make-fixture # write the deterministic fixture only (to .review-tmp/fixture-run)
Playwright with Chromium is required (npx playwright install chromium).
Publishing
This extension publishes under the malte-langermann Marketplace publisher via the
GitHub Actions Release workflow (push a vX.Y.Z tag). Nightly pre-releases
are published by the Nightly workflow. CI validates and packages every push.
Roadmap & contributing
The project tracks progress in ROADMAP.md. M0–M4 (TypeScript migration, robustness, metrics + charts, Markdown export, timeline, launch polish, screenshots, community files, Open VSX) are shipped in v1.0.0. Project orientation for contributors lives in CLAUDE.md, and the on-disk run format the dashboard reads is documented in docs/DATA-FORMAT.md.
Disclaimer
An unofficial, community-built tool. Not affiliated with or endorsed by Anthropic.
License
GPL-3.0-or-later — see LICENSE.
Copyright (C) 2026 Malte Langermann.