sift

Sift

Sift through usage data across all AI coding tools (Claude Code, Copilot CLI, VS Code Copilot, Cursor, Gemini CLI, Codex CLI). Produces metrics, reports, a dashboard, and a machine-readable JSON export.

Running Sift

The binary is bundled as a zipapp. All commands use ${CLAUDE_PLUGIN_ROOT}:

# Full analysis (all time, all sources)
python3 ${CLAUDE_PLUGIN_ROOT}/skills/usage-analysis/analyzer.pyz

# Scoped analysis
python3 ${CLAUDE_PLUGIN_ROOT}/skills/usage-analysis/analyzer.pyz --days 7
python3 ${CLAUDE_PLUGIN_ROOT}/skills/usage-analysis/analyzer.pyz --days 30 --source claude-code
python3 ${CLAUDE_PLUGIN_ROOT}/skills/usage-analysis/analyzer.pyz --project my-repo

# Discover available sources and projects
python3 ${CLAUDE_PLUGIN_ROOT}/skills/usage-analysis/analyzer.pyz --list

Always start with --list if the user hasn't specified a scope, to understand what data is available.

Output Artifacts

All outputs go to ~/.sift/:

File	Purpose
export.json	Primary artifact — machine-readable, all metrics + per-session data
report.md	Human-readable markdown report
dashboard.html	Interactive HTML dashboard (open in browser)
prompts/	User prompts grouped by project

Working with export.json

After running the analyzer, read ~/.sift/export.json for programmatic analysis. Structure:

{
  _schema: "sift-export-v1"
  summary: { total_sessions, total_cost_usd, total_tokens, total_tool_calls, projects }
  metrics: {
    <key>: { title, order, data: { ... } }
  }
  sessions: [
    { session_id, source, project, date, model, cost_usd, tokens: {...}, activity: {...}, extras: {...} }
  ]
}

Key Metrics to Examine

Metric Key	What It Tells You
cost_efficiency	Total spend, cost per session/action/output token
session_outcome	Heuristic success vs failure rate, wasted cost
retry_ratio	Token waste from debug loops and retries
model_routing	Savings available from routing light tasks to cheaper models
cache_efficiency	Cache hit rate, amortization, savings from caching
session_health	Median/P90/max session sizes, bloat index
duration_trend	Are sessions getting longer over time?
prompt_length	Under/over-specification patterns in user prompts
daily_burn	Daily cost trend with 7-day rolling average
platform_comparison	Per-source cost and session breakdown

Interpreting Results

When presenting findings to the user, focus on:

1. Cost drivers — What's expensive and why? Check platform_comparison, model_mix, top_sessions.
2. Waste — Sessions that failed (session_outcome.failure), retry-heavy sessions (retry_ratio), bloated sessions (session_health.sessions_over_50m).
3. Optimization opportunities — Model routing savings (model_routing.potential_savings), cache improvement (cache_efficiency), compaction events (child_metrics.compaction_count).
4. Trends — Is spend growing (daily_burn)? Are sessions getting longer (duration_trend)? Is adoption increasing (project_adoption)?
5. Prompt quality — Short prompts cause exploration waste, long prompts waste context (prompt_length).

Offer the dashboard (open ~/.sift/dashboard.html) for interactive exploration.

Prerequisites

• Python 3.10+
• pyyaml (for Copilot CLI source parsing)