Ingest

ingest

When to use

Anything new is arriving: files in inbox/, a link or DOI in the chat, a citation the human read out.

Method

1. Files already in inbox/, or a path you were given: magi ingest auto — it routes by type and finalizes. Stop here unless it refuses. (A run that skipped cleanup: magi ingest finalize <the file> --project-dir ..)
2. A link, DOI, arXiv id or citation: identify it first — that is the step that needs you, not the pipeline. An id you typed from memory goes in as magi ingest url "<id>" --expect "<title fragment>"; then magi ingest batch-run.
3. magi ingest review lists what is waiting. Show its findings to the human before committing — surface identity-mismatch, figure-count-mismatch and image-path-not-portable every time: they mean the file is not what its name says.
4. magi ingest review --item <id> --decision approve per row, then magi ingest review --commit.
5. After a batch: magi math check --json. Ingestion's signature failure is a $$ that lost its pair and swallowed the paragraph after it — valid LaTeX, so nothing catches it. Work that list with the tidy skill.

Rules

• Never transcribe pages through vision because a converter failed: it costs one sub-agent call per page. Not until a person has said yes; 10 at once.
• Never commit a batch the human has not seen the findings for.
• A converter that exits non-zero: report its stderr and stop. An unreadable file: name it and carry on. Report partial work as partial — "9 of 12".
• Which project a source belongs to is never your guess: when it is ambiguous, stop and ask. Collect the questions the work you delegated could not ask (Invariant 4) and put them all at once.