Telemetry and Stats

Dwight tracks every AI invocation locally — no data leaves your machine. Use the dashboard to understand costs, success rates, and which features consume the most tokens.

How It Works

1. Every invocation is logged — timestamp, model, tokens, cost, status, feature, and mode.
2. View quick stats — :DwightUsage shows a compact floating window.
3. Full dashboard — :DwightStats opens a detailed split buffer.
4. Export — :DwightStats export saves data as JSON or CSV.

Quick Usage

:DwightUsage

A floating window with lifetime and session stats: invocation count, tokens sent/received, estimated cost, and per-model breakdown.

Full Dashboard

:DwightStats

Opens a full-width telemetry dashboard in a split buffer with:

• Overview cards — total invocations, cost, success rate, tokens
• Session summary — current session duration, calls, cost
• Daily trend — 14-day sparkline showing cost, invocations, successes, and failures
• Per-feature breakdown — which features consume the most tokens (sorted by cost)
• Per-model breakdown — token usage and cost split by model
• Mode distribution — invocations across operation modes
• Job log summary — recent success/failure counts

:DwightStats features      " Feature-only view

Export

:DwightStats export        " JSON (default)
:DwightStats csv           " CSV format
:DwightStats json          " JSON format

Exports are saved to .dwight/stats-export.{json|csv} and include lifetime data, daily breakdown, per-feature stats, and per-model stats.

Lessons

Related to telemetry, lessons track what Dwight learns from agent sessions:

:DwightLessons                 " View all lessons with metadata
:DwightLessons consolidate     " Merge similar lessons
:DwightLessons stats           " Summary: total, stale, usage counts
:DwightLessons clear           " Remove all lessons

Lessons older than 30 days are flagged as stale in :checkhealth dwight. Run :DwightLessons consolidate to merge overlapping lessons and remove outdated ones.

Data Storage

All telemetry is stored in .dwight/usage.json (included in .dwight/.gitignore by default). Data tracked per invocation: timestamp, duration, model, tokens sent/received, estimated cost, success/failure, feature name, and operation mode.

Session data resets when Neovim restarts. Lifetime data persists across sessions.

Tips

• Check the daily trend for cost spikes. A sudden increase usually means a task ran more retries than expected or used an expensive model.
• Monitor per-feature costs. Features that consistently consume high tokens may need splitting (see Feature Management).
• Export before clearing. If you want to start fresh, export first — the data is useful for tracking improvement over time.
• Consolidate lessons monthly. Stale lessons add noise to agent prompts. Regular consolidation keeps the set focused.

Commands

See Also

• Core Concepts -- how lessons fit into the learning system
• Agent Mode -- sessions that generate telemetry and lessons
• Auto Mode -- multi-step sessions with per-task tracking
• Session Replay -- step through individual sessions