vibe: Skill for Claude Code

Analytics

Gets you the data. What to do with it is your call — read the question and answer it, rather than producing a fixed report.

Prerequisites

scripts/export_analytics.py calls load_dotenv(), so a .env at the repo root supplies: BASE_URL, AUTH_SECRET, AUTH_NAME, AUTH_EMAIL, APP_KEY, and optionally APTABASE_REGION (defaults to SH).

Export

Dates are UTC and day-granular: --start-date inclusive, --end-date exclusive. For a window in hours, export the days that overlap it and trim afterwards.

uv run scripts/export_analytics.py \
  --start-date YYYY-MM-DD --end-date YYYY-MM-DD \
  --output scripts/analytics_raw.csv

Also takes --format csv|parquet and --build-mode release|debug (default release — debug rows are developer machines).

Columns

Column
timestamp	UTC
event_name	see below
user_id	anonymous
app_version
os_name / os_version	e.g. Windows, macOS, Linux Mint
string_props	JSON; failures carry error_message with the full string
numeric_props	JSON

The event set

Read it from the code, never from memory — names change:

• desktop/src-tauri/src/analytics.rs
• desktop/src/lib/analytics.ts

The shapes are <action>_started → <action>_succeeded / <action>_failed, and <component>_<failure_type> for infrastructure faults. Work out which of the current names are starts, successes and failures before computing any rate.

Trimming to an exact window

uv run --with pandas python - <<'PY'
from datetime import timedelta
import pandas as pd

df = pd.read_csv("scripts/analytics_raw.csv")
df["timestamp"] = pd.to_datetime(df["timestamp"], utc=True)
df = df[df["timestamp"] >= df["timestamp"].max() - timedelta(hours=48)]
df.to_csv("scripts/analytics_window.csv", index=False)
print(len(df), df["timestamp"].min(), df["timestamp"].max())
PY

From here, inline pandas passes over the CSV answer most questions. A few things worth knowing when you interpret them: a handful of users can repeat the same input mistake and dominate an event-level error rate, so per-user counts often say more than raw ones; and a failure rate is only meaningful against its own start event, when that lifecycle is present in the data.