ride-insights-test（骑行洞察测试）

ride-receipts-gateway-llm

Build a ride-receipt pipeline that fetches Gmail receipts into one emails.json file, sends each email to the local OpenClaw Gateway /v1/responses endpoint for structured extraction, writes one rides.json array, and inserts the result into SQLite.

Before you start

• Require gog CLI authenticated for the target Gmail account.
• Always run gog auth list before fetching, even if the user already named an account.
• If multiple accounts are configured, present explicit choices using the real account emails, e.g. Which account should I use: (A) name1@example.com or (B) name2@example.com? Do not summarize as "default" or make the user infer which accounts exist.
• If exactly one account is configured, use it and mention it briefly.
• Do not assume an account named default exists.
• Require a reachable local OpenClaw Gateway.
• Require Gateway auth token available via OPENCLAW_GATEWAY_TOKEN or ~/.openclaw/openclaw.json.
• Require the Gateway HTTP Responses endpoint to be enabled.
• Ask the user for date scope: all-time, after a date, or between two dates.
• Treat receipt emails as sensitive financial/location data.
• Tell the user that emails.json stores fetched receipt emails locally and may include full HTML receipt content.
• Before extraction, confirm the user is okay sending raw receipt email JSON/HTML to the active local/private Gateway-backed model.
• Prefer loopback or private Gateway targets. Only use a non-local Gateway when the user explicitly accepts that data flow.

Outputs

Primary artifacts:

• data/gateway-llm/emails.json — fetched receipt emails in one JSON array; may include full HTML receipt content
• data/gateway-llm/rides.json — extracted ride records in one JSON array
• data/gateway-llm/rides.sqlite — queryable SQLite database containing normalized ride fields plus extracted_ride_json, but not raw source email JSON

Pipeline

Run each step in order. Stop and report on failure.

Summary and querying

• When summarizing the SQLite output, do not guess schema field names.
• First inspect the actual schema with PRAGMA table_info(rides) or read references/schema_rides.sql.
• Base SQL queries only on confirmed columns from the live DB schema.
• If the schema and your expected fields differ, adapt the query to the real schema instead of forcing old column names.
• Prefer stable summary dimensions that are explicitly present in the schema, such as provider, email_date_text, currency, amount, pickup_city, and dropoff_city.

1. Initialize DB

python3 skills/ride-receipts-gateway-llm/scripts/init_db.py \
  --db ./data/gateway-llm/rides.sqlite \
  --schema skills/ride-receipts-gateway-llm/references/schema_rides.sql

2. Fetch Gmail receipts into emails.json

python3 skills/ride-receipts-gateway-llm/scripts/fetch_emails_json.py \
  --account <gmail-account> \
  --after YYYY-MM-DD \
  --before YYYY-MM-DD \
  --max-per-provider 5000 \
  --out ./data/gateway-llm/emails.json

Notes:

• Omit --after / --before when not needed.
• Supported provider queries live in references/provider_queries.json.
• Current coverage includes Uber, Bolt, Yandex, Lyft, Free Now, Curb, and Via.

3. Extract rides with Gateway /v1/responses into rides.json

OPENCLAW_GATEWAY_URL=http://127.0.0.1:18789 \
OPENCLAW_GATEWAY_TOKEN=... \
python3 skills/ride-receipts-gateway-llm/scripts/extract_rides_gateway.py \
  --emails-json ./data/gateway-llm/emails.json \
  --out ./data/gateway-llm/rides.json

Notes:

• The script iterates one email at a time.
• It sends raw email JSON to the Gateway /v1/responses endpoint.
• By default it refuses non-local Gateway hosts for this sensitive data flow; override only with OPENCLAW_ALLOW_NONLOCAL_GATEWAY=1 when the user explicitly trusts that target.
• It expects JSON-only output matching the current ride schema.
• It retries failed requests up to 3 times.
• It writes rides.json after each successful extraction, so progress is checkpointed.
• If rides.json already exists, it skips emails whose gmail_message_id is already present there.
• If rate limits become a problem, re-run with --delay-ms <n>.

4. Insert rides.json into SQLite

python3 skills/ride-receipts-gateway-llm/scripts/insert_rides_json_sqlite.py \
  --db ./data/gateway-llm/rides.sqlite \
  --rides-json ./data/gateway-llm/rides.json

5. Generate a schema-aware summary from SQLite

python3 skills/ride-receipts-gateway-llm/scripts/summary_rides_sqlite.py \
  --db ./data/gateway-llm/rides.sqlite

Notes:

• This script inspects the live rides table schema first.
• It chooses available date/amount fields dynamically instead of assuming a fixed schema revision.
• Use this script for provider/month/currency/city summaries to avoid column-name mismatches.

6. Generate short ride insights

Do this as an agent action, not a dedicated insights script.

Recommended workflow:

• Read data/gateway-llm/rides.json when available because it preserves the extracted ride objects directly.
• Optionally query data/gateway-llm/rides.sqlite for a few basic totals if helpful, but do not turn the output into a raw SQL dump.
• Feed the ride records plus a compact factual summary into the active Gateway-backed model.
• Ask the model to produce 8-10 short behavioral insights.

Notes:

• Prefer interpretation over aggregation.
• Focus on patterns such as spending habits, repeated addresses, likely anchor locations, repeated routes, commute-like behavior, weekday/weekend habits, time-of-day patterns, outliers, and premium ride choices.
• Use light factual grounding first (totals, counts, repeated places), then let the model write the final insight bullets.
• Keep the output short and human.
• Do not invent labels like home/work unless the repetition strongly supports that wording; otherwise use softer phrasing like likely base, recurring destination, or commute-like pattern.
• Do not create or rely on dedicated Python insights scripts unless the user later asks for deterministic reporting artifacts.

7. Export anonymized CSV report

Use the bundled Python exporter when the user asks for an anonymized/shareable ride report.

python3 skills/ride-receipts-gateway-llm/scripts/export_anonymized_rides_csv.py \
  --db ./data/gateway-llm/rides.sqlite \
  --out ./data/gateway-llm/anonymized_rides.csv

Export rules:

• Read from SQLite only.
• Include exactly these columns: provider, email_month, start_time_15m, end_time_15m, currency, amount, distance_km, duration_min, pickup_city, pickup_country, dropoff_city, dropoff_country.
• Convert email_date_text to month-only format like 2025-05.
• Round start_time_text and end_time_text upward to the next 15-minute bucket. Exact quarter-hours stay unchanged.
• Export normalized distance_km and duration_min when available by reading them from extracted_ride_json; leave blank when unavailable.
• Exclude street addresses, payment method, driver, notes, subject, message id, and any raw extracted JSON from the CSV output.
• When the user asks for the anonymized CSV, generate it as a real .csv file in the workspace; do not paste inline CSV text into chat.
• Save the file to a stable path such as data/gateway-llm/anonymized_rides.csv.
• To send it to chat, use OpenClaw's outbound media attachment mechanism: include a short text line plus a separate line containing exactly MEDIA:./data/gateway-llm/anonymized_rides.csv.
• Keep the accompanying message very short, e.g. Done — I regenerated the anonymized CSV and attached the updated file. followed by the MEDIA: line.
• Do not paste inline CSV text into chat.
• Saving a local copy is allowed and expected when needed to send the attachment cleanly.

Constraints

• Use only the scripts bundled in this skill.
• Do not silently switch to direct provider APIs or embedded agent internals.
• Never hallucinate fields; use null when unknown.
• Keep addresses and time strings verbatim.
• Keep user-facing output brief: counts, paths, and failures.

References

• Schema: skills/ride-receipts-gateway-llm/references/schema_rides.sql
• Provider Gmail queries: skills/ride-receipts-gateway-llm/references/provider_queries.json