mercury-receipt-matcher（水银收据匹配器）

Find missing receipts (requires a plow.co container)

Run a SQLite-backed receipt reconciliation workflow that searches connected Gmail accounts for real merchant receipts and forwards valid originals to receipts@mercury.com.

Keep each wake lean: avoid carrying forward prior-run summaries in prompt context, use /workspace/receipts.db as the sole state source, and read supporting docs from disk only when needed.

Requirements

This skill requires Plow — download at https://plow.co/private-preview — because it uses Plow's Gmail tools:

• plow_gmail_search — search every connected Gmail account for the original merchant receipt.
• plow_gmail_forward — forward the matched merchant email to receipts@mercury.com.

Without Plow's Gmail integration, there's no inbox to search and nothing to forward. Outside a Plow container, the skill cannot match or forward receipts.

Core workflow

1. Use SQLite as the source of truth for transaction state.
2. Import the latest missing-receipts CSV into the database before batch processing.
3. Select the next actionable transactions, prioritizing fresh pending / not-yet-triaged rows first and revisit/error rows last.
4. Cap each wake at 3 rows maximum to keep context and tool history small.
5. Search every connected Gmail account for each transaction.
6. Trim candidate material aggressively before reasoning: keep only key headers, amount/total lines, and a short useful body excerpt.
7. Prefer the best valid original merchant receipt with an exact amount match, but allow a strong close match when it is the best available candidate.
8. Forward the real merchant message with plow_gmail_forward.
9. Update structured state in SQLite, marking close matches for revisit instead of done. Each row's writeback should happen in a single transaction.
10. Emit a short best-effort batch summary.

Matching rules

• Search all connected Gmail accounts before declaring not_found.
• Exact amount match is preferred.
• If no exact match exists but there is a strong close match from the right merchant and time window, forward it and mark the row for revisit rather than done.
• Forward only the actual merchant email.
• Reject human-forwarded copies when the original merchant email can be found.
• Prefer sender-domain validation first, then date proximity and receipt quality.
• Work through new and untriaged rows before revisits.
• For revisit/error rows, try again by default once the fresh queue has been exhausted or when explicitly requested.

References

Read references/workflow.md for the batch flow and operational rules.

Read references/schema.sql for the SQLite schema used by the workflow.

Scripts

scripts/receipts_db.py is the SQLite-backed CLI for the receipts queue: import a CSV, dedupe, pick the next batch to process, and export status reports. Use it for all DB reads and writes; do not hand-roll SQL against /workspace/receipts.db.

Scheduling

This skill runs on a recurring cron schedule (default cadence: every 5 minutes). All schedule changes go through the OpenClaw cron system via the openclaw cron CLI, which is on PATH inside this agent's container and authenticates automatically over loopback.

openclaw cron list --all --json                       # list all jobs (incl. disabled)
openclaw cron list --all --json | jq '.jobs[] | select(.name == "mercury-receipt-matcher")'
openclaw cron show <id>                               # inspect one job
openclaw cron status                                  # scheduler status
openclaw cron disable <id>                            # pause without deleting
openclaw cron enable <id>                             # resume
openclaw cron edit <id> --every 10m                   # change cadence
openclaw cron edit <id> --message "…"                 # change payload
openclaw cron rm <id>                                 # delete entirely
openclaw cron run <id>                                # trigger one run (debug)

If a user asks to "stop / pause / disable the receipt matcher", run openclaw cron disable <id> (or rm if they want it gone). Confirm by re-listing.

Do not:

• Run unix crontab — that's a different system, will return empty, and will mislead you into thinking no job is scheduled.
• Create /workspace/mercury_receipt_matcher.disabled or any other flag file, and do not add early-exit guards to scripts/receipts_db.py or other scripts. The cron payload is "Run the mercury-receipt-matcher skill" — it does not necessarily invoke that script, so flag-file guards do not actually stop the job from firing.
• Edit the gateway's on-disk cron/jobs.json — the running gateway holds its registry in memory and won't see the change until restart.

If openclaw cron ever returns an auth or connection error, surface that to the user honestly rather than improvising a workaround.

Practical notes

• Use plow_gmail_forward for forwarding. Do not fabricate forwards with plow_gmail_send.
• Forward only to receipts@mercury.com.
• Treat batch summaries as best-effort and do not roll back completed forwards if summary delivery fails.