ingest

name	ingest
description	Turn material the user already has — journals, notes exports, message archives, old therapy notes, a CV, letters — into context entries with provenance. Use when the user points at files or a folder and says "read this", "here's my journal", "use these notes".

personal-context: ingest

The faster route in. Most people have years of relevant material lying around; reading it beats interviewing them about the same period.

Before starting

Resolve paths (sh "${CLAUDE_PLUGIN_ROOT}/scripts/paths.sh" --json), read the config, read INDEX.md (so you can update rather than duplicate), and ask two questions:

1. What is this material? A journal is not a message export is not a CV; the reading strategy differs, and knowing the genre prevents category errors.
2. What date range does it cover?

Then say what you're going to do: read, propose entries, and write nothing until they've seen each one.

What to extract

Look for the durable and skip the situational. The test from the contract: would this still be true and useful in a workspace about something else?

Extract	Leave
People, and how the relationship works	What someone said on a Tuesday
Chapters with dates	Individual days
Events the person themselves marks as significant	Every dated occurrence
Recurring patterns they name about themselves	Moods
Care history, treatments, what helped	Session-by-session detail
Stated preferences about how to be treated	Passing irritations

A hundred-page journal usually yields ten to twenty entries. If you have sixty candidates, you are transcribing rather than extracting.

Provenance

Every entry produced here carries:

source: ingested
confidence: stated        # if the person wrote it as fact about themselves
                          # inferred  if you concluded it from what they wrote

and a history line naming the file and, where the format allows, the location:

## History
- 2026-08-14 — ingested from `journals/2021.md`, entries of 2021-03 to 2021-06

Provenance is the whole reason ingestion is trustworthy. An entry that cannot say where it came from should not be written.

Procedure

1. Inventory first. List the files, sizes, formats and date coverage. Show it. Large exports: agree an order and work in passes rather than reading everything into context at once.
2. Read a pass. One source, or one date range.
3. Draft candidates. Group by type. For each: the entry as it would be written, and the line or passage it came from.
4. Review together. They will correct a third of them, and the corrections are usually more informative than the entries.
5. Write the confirmed ones via the write protocol. Update INDEX.md.
6. Report coverage — which files were read, which weren't, and what date ranges are still dark. Feeds gap-scan.

Merging with existing entries

If a candidate matches an existing entry (same person, same period, same pattern), update rather than duplicate: revise the body, append a history line naming the new source, bump updated. Match on id and aliases first, title second. When unsure whether two things are the same person, ask — a wrongly merged person entry is hard to unpick later.

Rules

• Never write the raw material into the store. The store holds entries; the journal stays where it is. Reference it by path.
• Never infer sensitive attributes — health conditions, sexuality, religion, politics, immigration status — from material where they were not stated. Not at any confidence level.
• Quote sparingly and mark it. A sentence of theirs is often the best body text; a page is a copy of the journal.
• Old material is dated, not current. Something true in 2019 goes in as a period or event, not as a situation. This is the most common ingestion error.
• Heavy material: if what you're reading is distressing, say so plainly before summarising it back at them, and ask whether they want it in the store at all.
• Ask before reading anything they did not explicitly point at, even if it sits in the same folder.