Copilot Usage Dashboard

A self-contained VS Code extension that gives you full visibility into your GitHub Copilot token usage.

Features

• Daily AI Credit Limit Guard (v1.9.x) — set a hard daily spend in credits or dollars; status-bar countdown turns amber → orange → red as you approach; animated mascot shield webview when limit is hit; three enforcement modes (soft / pause / strict); snooze / resume / end-override controls with day rollover
• GitHub Copilot Agent Hooks (v1.9.0+) — installs PreToolUse + UserPromptSubmit lifecycle hooks at ~/.copilot/hooks/ that deny tool calls in Copilot CLI, local custom agents (.agents/*.md), and the Copilot cloud agent when the daily limit is hit. Fail-OPEN by design.
• Unified usage across 3 sources — VS Code Copilot Chat, Oh My Pi (OMP) agent sessions, and Pi coding-agent sessions, all rolled into a single AIC budget view
• AI Credits (AIC) tracking — per-model credit costs with configurable rates (June 2026 billing model)
• Budget monitoring — uncapped percentage (shows true overage e.g. 494%), "days of runway at current pace" indicator, projected end-of-cycle usage, overage cost estimates
• Redesigned dashboard (v1.7.6) — hero KPI cards, tabbed Breakdown (Model / Project / Tool / Subagent), collapsible expanders, side-by-side Trend charts with auto-generated insight captions
• Token counts (prompt, output, cached) per session, model, project, and day
• Per-source breakdown table: Sessions, Turns/LLM Calls, Tokens, and AIC Credits split across VS Code, OMP, and Pi
• Session browser with title, preview, duration, tools, and subagent usage
• Clickable links to session log and transcript JSONL files
• Model breakdown across Claude, GPT, Gemini families with per-model credit costs
• Daily usage trends (stacked bar chart) + monthly credits calendar heatmap
• Tool and subagent call tracking
• Live OpenTelemetry receiver (OTLP HTTP on port 14318)
• Auto-refresh (30s / 1m / 2m / 5m / Off) and manual refresh
• Status bar with session count and token totals
• Multi-root workspace support

Install

From source:

cd copilot-usage-extension
npm install
npm run compile
npx @vscode/vsce package --allow-missing-repository
code --install-extension copilot-usage-dashboard-*.vsix

Open: Command Palette > Copilot Usage: Open Dashboard

Data Sources

The scanner auto-detects the VS Code workspaceStorage root across platforms (override with copilotUsage.workspaceStoragePath):

Inside that root the extension reads:

1. chatSessions JSONL at {root}/{hash}/chatSessions/*.jsonl
   • Parsed by src/scanner.ts parseSessionContent. Handles kind=0 (current — session metadata + embedded v.requests[]), kind=1 (legacy — [customTitle] / [requests, N, result]), and kind=2 (latest prompt snapshot). All three must keep working.
   • Tokens here are snapshots only — authoritative per-call data comes from debug-logs.
2. debug-logs at {root}/{hash}/GitHub.copilot-chat/debug-logs/{session}/
   • main.jsonl — per-event stream (session_start, turn_start, llm_request, child_session_ref). Each llm_request carries attrs.model, attrs.inputTokens, attrs.outputTokens, attrs.cachedTokens, and attrs.copilotUsageNanoAiu (exact API billing × 1e9). Per-event model attribution is required so auxiliary calls (e.g. gpt-4o-mini for title generation, claude-haiku-4.5 for subagents) appear in per-model rows instead of being collapsed into the parent turn's model.
   • runSubagent-*.jsonl — subagent rounds. Credits are merged into the parent turn that spawned them (via child_session_ref).
   • title-*.jsonl — title generation. Fires before any turn_start, so the parser attributes it to turn 0 to avoid losing its small-model credits.
3. transcripts at {root}/{hash}/GitHub.copilot-chat/transcripts/
4. Oh My Pi (OMP) agent sessions at ~/.omp/agent/sessions/**/*.jsonl — scanned concurrently with mtime caching; contributes LLM calls, tokens, and AIC credits to the shared budget
5. Pi coding-agent sessions at ~/.pi/agent/sessions/**/*.jsonl — same scanning model as OMP
6. Live OTel (optional) — built-in VSCode OTLP HTTP receiver on port 14318

The scanner handles both legacy (kind=1) and current (kind=0) JSONL formats. All file I/O is fully async with concurrent reads (16-worker pool) and mtime caching. OMP/Pi token counts are reported as all-time historical; AIC credits for those sources are scoped to the current billing cycle (Jun 1+).

Contributors: .agents/agents.md documents the JSONL parsing contract and invariants that must not regress. tests/scan-june-workspace.ts is an independent audit script that re-implements llm_request parsing and must continue to print Extension credit display matches API ground truth (0.0% drift) after any parser change.

Configuration

The extension auto-configures OTel settings on first activation:

A VS Code reload is needed after first install for Copilot to start exporting telemetry.

AI Credits (AIC) Configuration

Since June 1, 2026, GitHub Copilot uses usage-based billing with AI Credits.

Configure your plan in Settings → search copilotUsage.aic:

Plan Defaults (per user/month, pooled at billing entity)

1 AI credit = $0.01 USD. Credits are pooled — an org with 10 Business users gets 19,000 credits shared.

Custom Model Costs

Override or add model pricing via copilotUsage.aic.customModelCosts:

"copilotUsage.aic.customModelCosts": [
  {
    "model": "claude-opus-4.6",
    "inputCreditsPerMillion": 500,
    "outputCreditsPerMillion": 2500,
    "cachedInputCreditsPerMillion": 50,
    "cacheWriteCreditsPerMillion": 625,
    "tier": "premium"
  }
]

Credits are calculated as: (net_input_tokens / 1M) × inputRate + (output_tokens / 1M) × outputRate + (cached_read_tokens / 1M) × cachedRate

Anthropic models also incur cache write costs: (cache_write_tokens / 1M) × cacheWriteRate

Daily AI Credit Limit Guard

Cap your daily spend independently of the monthly AIC budget. Configure in Settings → search copilotUsage.dailyLimit (lives in its own settings group: Copilot Usage — Daily Limit Guard):

How the agent hook works

When installAgentHooks: true, the extension writes:

• ~/.copilot/hooks/copilot-usage-limit.json — registers PreToolUse (decision = deny) + UserPromptSubmit (notification) hooks
• ~/.copilot-usage/check-limit.{ps1,sh} — reads live state, emits {"permissionDecision":"deny","permissionDecisionReason":"..."} when the daily limit is hit
• ~/.copilot-usage/limit-state.json — rewritten on every snapshot with { blocked, percent, dollarsSpent, dollarsLimit, resetIn }

Coverage:

The check script is fail-OPEN — if it errors, Copilot continues working. Only a successful blocked == true read produces a deny.

You can install/uninstall hooks manually via Command Palette:

• Copilot Usage: Install Agent Hooks
• Copilot Usage: Uninstall Agent Hooks

Commands

Requirements

• VS Code 1.85+
• GitHub Copilot Chat extension

Privacy

All data stays local. The extension reads from your VS Code workspaceStorage directory and listens on 127.0.0.1 only.

License

MIT