moonlight-lupin/agent-skills/receipt-compiler

Receipt Compiler

Overview

Receipt Compiler turns a folder of phone-camera receipt photos into one A4 PDF that looks like a photocopied expense-claim pack: each receipt is straightened (perspective warp, deskew), converted to a B&W "scanned" look, tiled on A4 with uniform visual scale, numbered, and summarised on an optional cover page.

Four subcommands with an enforced confirmation gate:

1. scan — straighten + scanify + OCR each photo → manifest.json
2. review — print the extracted expense table
3. confirm — stamp user approval + bind the reviewed data (digest)
4. pack — build the A4 PDF (refuses to run until confirm has run)

When to Use

• "Compile my receipts into a PDF for expense claim"
• "Make a claim pack from these receipt photos"
• "Straighten and B&W these receipts and line them up on A4"
• Not for: scanning via flatbed scanner, or PDFs of invoices (use pdf skill).

Prerequisites

pip install opencv-python-headless pillow pillow-heif pytesseract reportlab numpy
# PyMuPDF is needed only to run the self-test (tests/), not the script itself

Plus system tesseract-ocr on PATH. Check with tesseract --version.

Workflow (agent must follow the gate)

Step 1 — scan

python3 ~/.hermes/skills/productivity/receipt-compiler/scripts/receipt_compiler.py \
    scan <photos_dir> -o <workdir> [--mode bw|photo]

Default --mode bw gives the photocopy look the user asked for. Use --mode photo for grayscale. Completion criterion: manifest exists with one entry per photo; ocr_confidence and needs_review flags set.

Step 2 — review + confirm (MANDATORY, user in the loop)

python3 .../receipt_compiler.py review <workdir>
# ... apply any corrections to manifest.json, THEN re-run review so the user
# ... sees the corrected table ...
python3 .../receipt_compiler.py review <workdir>
# ... after the user approves the corrected table ...
python3 .../receipt_compiler.py confirm <workdir>

Present the printed table to the user in chat. Ask:

1. Are the extracted dates/merchants/amounts right? Fix manifest.json fields.
2. Optional requirements — ask every time, do not assume:

• cover page (with claimant/period/notes)
• per-receipt numbering badges
• per-receipt captions (date · merchant · amount)
• currency symbol on totals

1. Exclusions — any receipt to drop ("include": false)?
2. Purpose/notes text for the cover.

Only after the user approves: run confirm <workdir>. This sets confirmed: true AND binds a review_digest of the exact reviewed fields at approval time. Any edit after confirm (dates, amounts, currency, inclusion) makes pack refuse until you re-run confirm — which means showing the user the changed table again. There is no bypass flag. Apply user corrections to manifest.json BEFORE running confirm.

Step 3 — pack

python3 .../receipt_compiler.py pack <workdir> -o claim.pdf \
    [--cover --number --captions] \
    --title "Expense Claim — <purpose>" --claimant "MH" --period "..." --notes "..."

Completion criterion: PDF exists, page count reported in JSON output. Then send the PDF to the user as a document/file on the active channel.

Layout rules (how receipts line up)

• A4 portrait, 15 mm margins, 5 mm gaps.
• Very tall receipts (aspect ≥ 2.2) are packed 3-per-row full-height;

normal receipts 2-per-row, capped at 52% of content height (≈47% of the page). Scale is per-cell (each receipt fills its cell); rows with fewer receipts get wider cells.

• Every image is scaled to preserve aspect ratio — no distortion, no cropping.
• Receipts flow across pages; 30+ receipts pack fine.
• Tall receipts print before normal ones; re-order via the manifest if needed.

Common Pitfalls

1. Skipping the confirmation gate — pack hard-fails on confirmed: false,

non-boolean confirmation values, and post-confirmation edits (digest check). There is no bypass flag; the user must see the table first.

1. Dark/shadowed photos — the 2/98 percentile stretch handles most; if OCR

confidence is < 55, or no amount was extracted, the entry is flagged needs_review; surface these to the user.

1. Perspective warp failure — glossy/wrinkled receipts may not produce a clean

quad; the script falls back to small-angle deskew of the full frame. Check method in the manifest; if deskew on a receipt that clearly needs cropping, fix the crop manually with PIL or accept the framing.

1. HEIC photos — iPhone photos arrive as .heic; pillow-heif handles them if

installed. If missing, ask the user for JPGs or install pillow-heif.

1. Multiple currencies — totals are grouped per currency (review output and

cover page print one line per currency, no combined total). For a cleaner claim, keep one currency per pack; never zero or edit valid amounts to force a combined total.

1. Long descriptions in cover table — descriptions truncate at 70 chars; keep

cover notes short or put detail in notes.

Verification Checklist

• [ ] scan produced manifest.json with an entry per photo
• [ ] review table shown to user; corrections applied to manifest
• [ ] Optional requirements asked each time (cover/numbering/captions)
• [ ] confirm run only after user approval (never hand-edit confirmed)
• [ ] pack output PDF exists, page count matches expectation
• [ ] PDF sent to the user with the correct output path