Dashboard Data Pipelines

- how each panel gets its numbers -

Enter PIN to continue

Every panel follows the same shape: a source is sampled by a PowerShell script, transformed into a JSON snapshot, and rendered by an HTML page on Vercel.

Manual trigger Scheduled trigger Session-only trigger Pipeline stages

Every pipeline, by panel

Panel	Source	Refresh script	Snapshot
claude.html	`ccusage` reading `~/.claude/projects/*.jsonl` (daily totals + `blocks --json` for active block & P90 caps)	`refresh-claude-dashboard.ps1`	`data/claude-usage.json`
agents.html	Frontmatter + body of every `*.md` in `ai-shared-brain/agents/` + git log for last-modified dates	`refresh-agents-dashboard.ps1`	`data/agents.json`
agents.html (activity)	Agent invocation logs collected over time	`sync-agent-logs.ps1`	`data/agent-stats.json`
brain.html	`git log` + file counts from `ai-shared-brain` repo	`refresh-brain-dashboard.ps1`	`data/brain.json`
mcps.html	`claude mcp list` + `claude mcp get <name>` for servers; plugin / skill directories on disk for slash commands	`refresh-mcps-dashboard.ps1`	`data/mcps.json`
sites.html	Vercel REST API for 5 active projects (deployments, domains, env vars) + curl to each live URL for HTTP 200	`refresh-sites-dashboard.ps1`	`data/sites.json`
routines.html (cloud)	Anthropic `RemoteTrigger list` API — token only available inside a Claude Code session	`refresh-routines-dashboard.ps1`	`data/routines.json`
routines.html (local)	Opted-in Windows scheduled tasks via `Get-ScheduledTask`	`refresh-local-tasks-dashboard.ps1`	`data/local-tasks.json`
tasks.html	`*.md` in `ai-shared-brain/tasks/{backlog,in-progress,done}/` — status derived from folder, not frontmatter	`refresh-tasks-dashboard.ps1`	`data/tasks.json`

Why this shape

Pure static pages + offline-computed snapshots. The dashboard is just HTML/CSS/JS on Vercel — no backend, no database, no live API calls at view-time. Every "live" number on the page is actually a snapshot computed by a PowerShell script and committed to git. That keeps the page fast, free, and trivially debuggable: the value you see on screen is byte-for-byte what's in data/X.json at the head of main.

Why PowerShell scripts not serverless functions. Most sources are local-machine-only — ~/.claude/projects/ JSONL, the brain repo, Windows scheduled tasks, the claude CLI. A serverless function on Vercel can't reach any of those. Running locally and pushing the result is the simpler shape; the script IS the function.

The routines panel is the awkward case. The Anthropic RemoteTrigger API needs the OAuth token that lives inside a running Claude Code process. Mark refreshes that one by hand from inside CC; the unattended scheduler can't reach it. Three possible fixes (build a save-snapshot endpoint that a cloud routine POSTs to, manual-token CLI wrapper, or just live with the manual refresh) — currently option 3.

Burn-rate stack (added 2026-05-15). Live in two places: ccusage statusline in ~/.claude/settings.json for live readout inside Claude Code's CLI; and a "Live session" panel on claude.html that shows P90-calibrated cost/token/message caps from the same JSONL data. Both read the same underlying logs — different presentation. Companion TUI claude-monitor (alias ccm) installed via uv tool install for a richer real-time view in a side terminal.