nanyuu

Collect and locally extract ride-sharing receipt data from Gmail, store it in JSON/SQLite, and generate anonymized ride-behavior insights using a local OpenC...

Security Scan

The skill's code mostly matches the described purpose (fetch Gmail receipts, extract via a local Gateway, store/anonymize locally), but there are multiple incoherences you should understand before installing.

Purpose & Capability

The required binaries (gog, python3) and the declared gateway env vars align with the stated purpose of fetching Gmail receipts and using a local OpenClaw Gateway. However the registry-level name/slug/owner/version differ from the SKILL.md and included _meta.json (SKILL lists 'ride-insights' / DataHive, registry lists 'nanyuu'/'nansui'), which is an inconsistency in provenance that should be explained by the publisher.

Instruction Scope

SKILL.md instructs the agent to prompt for user consent before sending raw email JSON/HTML to the Gateway and to always use a local loopback Gateway only; the implementation sends the full per-email JSON to the Gateway /v1/responses endpoint. The code does not itself prompt for user confirmation (prompting is an agent-level behavior), so the runtime depends on the agent following the SKILL.md. Also the SKILL.md and code explicitly read ~/.openclaw/openclaw.json as a token fallback — this file access is not declared in the skill's required config-paths metadata.

Install Mechanism

No install spec (instruction-only); files are bundled in the skill and there are no external downloads or extract steps. This is the lower-risk install pattern.

Credentials

Declared env vars (OPENCLAW_GATEWAY_TOKEN, OPENCLAW_GATEWAY_URL, OPENCLAW_GATEWAY_MODEL) are appropriate for talking to a Gateway. However the code also reads a local config file (~/.openclaw/openclaw.json) as a fallback for the token — that config path was not declared in the skill metadata. The skill also requires the gog CLI to be authenticated to the user's Gmail account, which is expected for this functionality but is a sensitive capability the user should explicitly consent to.

Persistence & Privilege

The skill writes persistent artifacts to data/ride-insights (emails.json, rides.json, rides.sqlite) as documented. It does not request 'always: true' or modify other skills. The only elevated access beyond declared env vars is reading ~/.openclaw/openclaw.json (local config).

What to consider before installing

This skill largely does what it claims (fetch Gmail receipts via the gog CLI, send raw email JSON to a local OpenClaw Gateway for extraction, store results locally, and export anonymized CSV). Before installing or running it, consider the following: - Provenance mismatch: the registry metadata (name/slug/owner/version) does not match the SKILL.md and _meta.json inside the package. Ask the publisher to clarify why, and prefer only skills whose package metadata matches the registry. - Local config access: the code will read ~/.openclaw/openclaw.json as a fallback for the Gateway token but the skill did not declare this config path. If that file contains other secrets, verify what it holds before proceeding. - Gmail access: the skill requires the gog CLI to be authenticated to your Gmail account. That gives the skill access to emails; run the gog commands yourself to confirm account(s) and inspect the gog auth configuration before granting access. - User confirmation: the SKILL.md instructs the agent to ask for explicit user consent before sending raw receipt JSON/HTML to the Gateway. The bundled scripts do not enforce interactive confirmation — ensure the agent will prompt you and do not run the extraction scripts unattended unless you have reviewed the data flow. - Local-only guarantee: the code enforces that the Gateway URL resolves to localhost/127.0.0.1/::1. Verify your OPENCLAW_GATEWAY_URL and OPENCLAW_GATEWAY_TOKEN are set correctly and that the Gateway is indeed local. If you do not trust the publisher or cannot confirm the provenance discrepancy, run the provided Python scripts manually (inspect them first), and avoid giving the agent direct autonomous permission to run them. If you proceed, back up/remove any sensitive files (emails.json) when finished.

ride-insights

A privacy-first mobility intelligence skill developed by DataHive AI.

Use this skill to collect ride-sharing receipts from Gmail, extract structured trip data locally, and generate rich personal insights on spending, habits, repeated routes, likely anchor locations, time-of-day patterns, and more.

You can also share an anonymized data report with DataHive AI to participate in missions and earn rewards (missions blog).

This skill uses a unique local-processing approach that showcases the real power of OpenClaw: agent intelligence processes sensitive source data locally and ensures that no raw data is shared externally.

Local-only by design: this skill requires a loopback OpenClaw Gateway and will fail if the Gateway URL points anywhere else.

Generate an anonymized/shareable CSV version of the ride history when the user wants to upload it to DataHive and earn points without exposing raw receipt emails or obvious personal identifiers.

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) [e-mail] or (B) [e-mail]? 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 at gateway.auth.token (legacy fallback: gateway.token).
• If the extractor is configured through env vars, declare and use only these: OPENCLAW_GATEWAY_TOKEN, OPENCLAW_GATEWAY_URL, and OPENCLAW_GATEWAY_MODEL.
• 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 data/ride-insights/emails.json stores fetched receipt emails locally and may include full HTML receipt content.
• Tell the user that extraction sends the raw per-email JSON payload, including receipt HTML when present, to the local Gateway /v1/responses endpoint.
• Before extraction, confirm the user is okay sending raw receipt email JSON/HTML to the active local Gateway-backed model.
• Always use the local loopback Gateway. If the configured Gateway URL is not local, fail rather than falling back to any remote/private host.
• When describing the local-only guarantee, be explicit that only localhost, 127.0.0.1, and ::1 are accepted Gateway hosts.

Outputs

Primary artifacts:

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

Retention note:

• emails.json persists raw fetched receipt content until the user deletes it.
• rides.json and rides.sqlite persist extracted ride data locally until deleted.
• The anonymized CSV intentionally excludes raw receipt content and direct identifiers, but it is still a derived dataset and should be treated as potentially sensitive.

Pipeline

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

1. Initialize DB

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

2. Fetch Gmail receipts into emails.json

python3 skills/ride-insights/scripts/fetch_emails_json.py \
  --account <gmail-account> \
  --after YYYY-MM-DD \
  --before YYYY-MM-DD \
  --max-per-provider 5000 \
  --out ./data/ride-insights/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.
• Default processing cap is 50 emails total for the selected interval because Gateway extraction is token-heavy.
• Before extraction, count the fetched emails in data/ride-insights/emails.json.
• If the fetched set is 50 emails or fewer, proceed normally.
• If the fetched set is over 50, ask the user whether to process the full count or keep the default cap of 50.
• Unless the user explicitly approves a higher number, process only the first 50 emails.
• Keep the user-facing explanation short and mention that the cap exists to control token usage/cost.

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

python3 skills/ride-insights/scripts/extract_rides_gateway.py \
  --emails-json ./data/ride-insights/emails.json \
  --out ./data/ride-insights/rides.json

Notes:

• Prefer running the extractor without exporting OPENCLAW_GATEWAY_TOKEN when ~/.openclaw/openclaw.json already contains gateway.auth.token.
• The extractor also accepts legacy gateway.token if present, but gateway.auth.token is the expected current config path.
• The extractor may also read OPENCLAW_GATEWAY_URL and OPENCLAW_GATEWAY_MODEL; these should be declared anywhere the skill metadata or packaging contract lists env dependencies.

Notes:

• The script iterates one email at a time.
• It sends raw email JSON to the Gateway /v1/responses endpoint.
• It refuses any non-local Gateway host for this sensitive data flow and does not provide an override.
• It expects JSON-only output matching the current ride schema.
• It retries failed requests up to 3 times.
• It writes data/ride-insights/rides.json after each successful extraction, so progress is checkpointed.
• If data/ride-insights/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>.
• Default extraction cap is 50 emails total unless the user explicitly approves processing more for the selected interval.
• When applying the default cap, use the fetched emails ordered as written in data/ride-insights/emails.json and extract only the first 50.

4. Insert rides.json into SQLite

python3 skills/ride-insights/scripts/insert_rides_json_sqlite.py \
  --db ./data/ride-insights/rides.sqlite \
  --rides-json ./data/ride-insights/rides.json

5. Generate ride insights

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

Recommended workflow:

• Prefer data/ride-insights/rides.json as the primary source because it preserves the extracted ride objects directly.
• Use data/ride-insights/rides.sqlite for lightweight deterministic counts, filters, grouping, and cross-checks.
• Before querying SQLite, inspect the schema with PRAGMA table_info(rides) or read skills/ride-insights/references/schema_rides.sql.
• Base SQL only on confirmed columns from the live DB schema or that schema reference file.
• Feed the ride records plus a compact factual grounding 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 rides.json for rich per-ride context and rides.sqlite for quick factual checks; combine both when useful.
• Keep SQL-derived grounding compact and human-readable; do not turn the output into a raw SQL dump.
• Keep the output compact 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.

6. Export anonymized CSV report

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

python3 skills/ride-insights/scripts/export_anonymized_rides_csv.py \
  --db ./data/ride-insights/rides.sqlite \
  --out ./data/ride-insights/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/ride-insights/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/ride-insights/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-insights/references/schema_rides.sql
• Provider Gmail queries: skills/ride-insights/references/provider_queries.json