Full manual

This is the operator’s manual: how to install the application, how to use it from its window, and how to use it from the command line.

Read Known limits before you rely on it.

The command line is for people running from source

Arguments do not reach the packaged builds (the Microsoft Store version and the .dmg version) — they arrive as argv=['']. This is a limit of the way Flet packages the application, and it cannot be fixed on our side. If you installed the application, read only “Using the window”.

The command line works when you clone the repository and run it with uv.

Installing and the first launch

Nothing else has to be installed. Siegfried (format identification) and ClamAV (virus scanning) are bundled.

The first launch takes a few minutes and the window may appear to do nothing. The Python runtime inside the application (about 100 MB) is being unpacked. Later launches are quick.

macOS may say the application cannot be opened because the developer cannot be verified. It is signed with a Developer ID and notarised by Apple, so this normally does not happen; if it does, open it once from the right-click menu (“Open”) rather than by double-clicking.

The first thing to do, if you will use the virus scan, is to download the definition database (several hundred megabytes). Use “Download / update definitions” under “Virus definition database”. See below for details.

Switching the interface to English

The window opens in the language your operating system is set to. Japanese for a Japanese system, English for everything else — German, Korean and so on get English rather than Japanese, because the source language of the code is not a reason to show someone a screen they cannot read.

The language selector is at the top right of the window, next to the version number. It is placed in the header on purpose: someone who reads only English should not have to find it inside a Japanese screen. Once you choose a language it is remembered, and it always beats the system setting — so an English interface on a Japanese machine is a choice you can make and keep.

The interface is translated. The packages are not. Column headings in the CSV files, the text of report.txt, and the PREMIS event records are written in Japanese whatever the interface language is. This is deliberate: an information package is meant to last decades, and core/sip_reader.py reads those headings back. If the headings changed with the interface language, a SIP made in English could not be read by an application set to Japanese, and the other way round.

So a non-Japanese reader gets an English interface over packages whose spreadsheets and reports are in Japanese. Translating the package contents is open work — see issue #11.

Using the window

Using the window goes through every part of the window with screenshots. This section is the short version.

The window as it opens (macOS, v0.1.6). Windows looks the same apart from the window frame.

1. Choose what to create

Three choices at the top:

Normally you accession with “Create a SIP”, check what it found, and only then go on to the AIP. They are separate steps because there are things a person has to look at while the material is still a SIP: personal information candidates, files whose extension disagrees with their content, and files whose format could not be identified.

2. Choose the input and the destination

• Input — the folder holding the material to accession. A ZIP can be given instead.
• Destination — where the package is written. Keep it outside the input folder (if you choose the input folder or a folder inside it, a red message says so and “Run” stays disabled).

The input folder is only read. Nothing is ever written into it.

3. Choose the options

“Options” is collapsed when the window opens, and the line under it says which options are on (“Defaults” when none are).

The PII scan recognises five categories, all defined around Japanese conventions — Japanese individual numbers (マイナンバー) and Japanese postal codes in particular. Email addresses and card numbers are not Japan-specific; phone numbers assume Japanese digit counts. Measured precision and recall are in docs/pii-accuracy.md.

4. Enter the descriptive metadata

• Identifier — the identifier of the transfer (e.g. 2026-transfer-general-affairs). It becomes the folder name.
• Title — required. The name of the body of material.

“Dates”, “Scope and content” and “Archivist” are optional. The archivist’s name is recorded in the AIP’s PREMIS events as the agent that performed the work.

What you enter here is written into the package as the accession record.

5. Run it, and read what it says

“Run” shows progress, then the location of the finished package and a list headed “Points to check by eye”. Four kinds appear there:

• ウイルス検出: — a file the virus scanner flagged, with the signature name
• PII候補: — a file holding something that looks like personal information, and how many
• 未識別: — a file whose format could not be identified
• 拡張子不一致: — a file whose extension disagrees with its actual format

These lines keep their Japanese prefixes, for the reason given above: they come from core/, which does not go through the translation table.

None of these are acted on automatically. What to do about them is a person’s decision.

6. Look inside

“Look inside” shows the package as tables. What you can see only here is the content of the METS and the PREMIS records — the XML is not readable as it stands.

1. Overview — what is in it, how many, when it was made
2. Preservation events — when, what, with which tool, and with what outcome
3. Workflow — the same events grouped by stage (ingestion → virus scan → format identification → normalization → validation → checksum → fixity check). Select a stage to see the tools used, the counts and any files with problems. A stage with no records is shown as such, not hidden
4. Files — format, PRONOM identifier, size, SHA-256, virus scan result

The originals themselves (a PDF, a Word file) do not open in the application. Open the folder and use whatever you normally use.

7. Build the AIP

Once the SIP has been checked, switch to “Create an AIP” and give it the SIP folder as the input. Right after creating a SIP, “Build an AIP from this SIP” under the result does both in one step. If you choose a folder that is not a SIP (the source folder, or the destination folder that holds the SIP), a red message under the input says so.

Building an AIP does three things:

• compares the SIP’s manifest against the files on disk, to confirm nothing has changed since accession
• normalizes to preservation formats, if you asked for it
• writes a METS with PREMIS events embedded, and wraps the whole thing as a BagIt bag

Derivatives produced by normalization are opened again and read back — a TIFF is fully decoded with Pillow, a PDF is re-opened with pypdf and must report at least one page. A derivative that cannot be read back is discarded, and the event is recorded as a failure. This is not format validation: opening a file says nothing about whether it conforms to its specification.

Using the command line

Setting up

git clone https://github.com/nakamura196/archival-packager.git
cd archival-packager
uv sync
uv run archival-packager check

check reports the state of the bundled tools (format identification and virus scanning), the virus definitions and the normalization rules. It is there to find out first what is unavailable, so a missing tool does not make it fail.

The bundled tools are not in the repository. Fetch them if you need them (scripts/fetch-binaries.zsh on macOS, scripts/fetch-binaries.ps1 on Windows). Without them, format identification and virus scanning are skipped, not run, and the report says so — so that “not checked” never turns into “no problems”.

Commands

archival-packager sip --input DIR --output DIR --identifier ID --title TITLE
                      [--scope-note TEXT] [--date-note TEXT]
                      [--bag] [--scan-pii] [--virus-scan]
                      [--sanitize-filenames] [--zip] [--prior-accession CSV]
                      [--json] [--quiet]

archival-packager aip --sip DIR --output DIR
                      [--no-normalize] [--archivist NAME] [--zip]
                      [--json] [--quiet]

archival-packager inspect PACKAGE_DIR [--json]

archival-packager check [--json]

There is no command for the window’s “From source material through to an AIP”. Call sip, then aip (they are kept apart so that a person can check the SIP in between).

sip — build a submission package

uv run archival-packager sip --input ./incoming/2026-03 --output ./packages --identifier 2026-transfer-general-affairs --title "General affairs records" --scan-pii

aip — build a preservation package

uv run archival-packager aip --sip ./packages/2026-transfer-general-affairs --output ./preservation --archivist "Satoru Nakamura"

inspect — show what is inside

Prints, as text, what the window’s “Look inside” shows. With --json it lists every file (the text output stops after the first 20).

uv run archival-packager inspect ./preservation/2026-transfer-general-affairs-AIP
uv run archival-packager inspect ./preservation/2026-transfer-general-affairs-AIP --json | jq '.summary'

What the output promises (the part that matters for automation)

• Progress goes to standard error, results to standard output
• With --json, standard output is a single JSON document. Pipe it straight into jq
• --quiet stops the progress lines. It does not stop the results or the findings
• On failure, --json still returns {"status": "error", "message": "..."}

Exit codes:

A virus detection or a personal information candidate leaves the code at 0. A finding is material for a person to judge; the run itself succeeded. If it counted as a failure, a nightly job could not tell “a failure to stop on” from “a finding to look at”. So that findings are not missed, they are always written to standard error and to findings in the --json output.

A fixity mismatch while building an AIP (the manifest check) also leaves the code at 0. The AIP is built and the mismatch is recorded in PREMIS. To catch it in automation, read .fixity.outcome from the --json output (passed / failed / skipped).

Automation examples

Build a SIP from a watched folder every night (cron + zsh)

Anything longer than a few lines goes in a script rather than being pasted into a terminal (pasting mangles line continuations). An example to save as scripts/nightly-sip.zsh:

#!/usr/bin/env zsh
# If today's material has arrived in the inbox, build a SIP. Called nightly from cron.
# Usage: nightly-sip.zsh
# Requires: the repository cloned and `uv sync` run.
set -euo pipefail

REPO=$HOME/archival-packager
INBOX=/Volumes/transfer/incoming
OUTBOX=/Volumes/transfer/packages
TODAY=$(date +%Y-%m-%d)

if [[ ! -d $INBOX/$TODAY ]]; then print "Nothing received: $TODAY"; exit 0; fi

cd $REPO
uv run archival-packager sip --input $INBOX/$TODAY --output $OUTBOX --identifier $TODAY --title "Daily accession $TODAY" --scan-pii --virus-scan --bag --json --quiet > $OUTBOX/$TODAY.json

print "SIP: $(python3 -c 'import json,sys; print(json.load(open(sys.argv[1]))["sip_path"])' $OUTBOX/$TODAY.json)"

The crontab entry (every night at 2 a.m.):

0 2 * * * /Users/archivist/archival-packager/scripts/nightly-sip.zsh >> /var/log/archival-packager.log 2>&1

2>&1 is there so that the progress goes into the log too. Remove --quiet to get one log line per step.

To be told on the days something was found, read the JSON:

VIRUS=$(python3 -c 'import json,sys; print(len(json.load(open(sys.argv[1]))["findings"]["virus"]))' $OUTBOX/$TODAY.json)
if [[ $VIRUS -gt 0 ]]; then print "Check: $VIRUS virus detection(s)"; fi

Watch for regressions in CI (GitHub Actions)

Build a SIP from a small sample and check that the result has not changed.

- run: uv sync
- run: uv run archival-packager check
- run: mkdir -p sample && printf 'test\n' > sample/a.txt
- run: uv run archival-packager sip --input sample --output $/out --identifier ci --title CI --json --quiet > $/sip.json
- run: uv run archival-packager inspect $/out/ci --json | jq -e '.overview.file_count == 1'

sample is a small folder that always gives the same result. Keep it in the repository, or create it on the spot as above.

jq -e exits with 1 when the condition is false, so the job turns red on its own. Values such as .summary.unidentified can be watched the same way, but in a CI without the bundled tools, format identification does not run and every file is unidentified. Choose what to watch according to what check reports as available in that environment. Run uv run pytest -q as well (that is what watches the application itself for regressions).

What gets written

SIP

<identifier>/
  objects/                      Copies of the originals (folder structure kept)
  metadata/
    metadata.csv                Archivematica-style description sheet (filled in by a person)
    checksum.sha256             SHA-256, paths relative to objects/
    submissionDocumentation/
      description.csv           AtoM / ISAD(G) description sheet (filled in by a person)
      atom-import.csv           For importing into AtoM (machine column names)
      formats.csv               Technical inventory (format, PRONOM, size, SHA-256, virus scan)
      accession.csv             Accession record (original file names and paths)
      dfxml.xml                 Technical metadata (DFXML)
      report.txt / report.html  Summary for people to read
      checksum.sha256           SHA-256, paths relative to the SIP root
      pii-report.csv            Personal information candidates (only with --scan-pii)
      arrangement-map.csv       Arrangement before and after (only with --prior-accession)

With --bag, all of the above goes under data/, and bagit.txt and manifest-sha256.txt are added.

The two files a person fills in are description.csv and metadata.csv. What is written in metadata.csv goes into the METS as file-level description when the AIP is built.

The column headings of these CSV files and the text of the reports are in Japanese, whatever the interface language (see Switching the interface to English).

AIP

<identifier>-AIP/
  bagit.txt / bag-info.txt / manifest-sha256.txt / tagmanifest-sha256.txt
  data/
    METS.<uuid>.xml             METS with PREMIS embedded (the processing record)
    objects/                    The originals, plus preservation derivatives if normalized
    objects/submissionDocumentation/
                                The SIP's submission documents, and the normalization rules actually applied
    logs/                       Logs of the preservation processing

Read the METS with inspect, or with “Look inside” in the window.

Image properties in dfxml.xml

For image files, an ap:image element is added to the DFXML fileobject, recording pixel dimensions, colour space, bits per sample and DPI. Only values that were actually read are written; nothing is guessed (no DPI in the file, no DPI in the record).

ap: is this application’s own namespace. No DFXML element is added or reordered (the DFXML schema allows elements from other namespaces at the end of fileobject, which is where these go). The version of Pillow used to read the image is recorded in creator.

If an image cannot be opened, ap:image readable="false" is written with the reason, and a warning is shown. It is never silently left out without saying whether it could be opened.

This covers the formats that the normalization rules assign to Pillow, plus TIFF. Adding a PUID to the rules brings it in automatically. Audio and video are not covered (they would need another library).

The virus definition database

The definitions are not bundled. They are several hundred megabytes and change daily, so bundling them would mean shipping something stale from the first day. Download them separately.

• In the window: the “Download / update definitions” button
• archival-packager check prints where they go

If --virus-scan is asked for and no definitions are present, no scan is run, and the report says “skipped (no definition database)”. It never says “nothing found”. Not having checked must not be mistaken for having found nothing.

What it cannot do

Read these before relying on the output:

• Known limits — the range of normalization, the absence of format validation, and the rest
• docs/pii-accuracy.md — the personal information scan covers five categories; measured misses and false positives
• docs/interoperability.md — the comparison with AtoM and Archivematica is against their specifications, not yet against running instances
• docs/rules.md — how to add your own normalization rules
• docs/performance.md — time and memory from 100 to 50,000 files