Fetch Readwise Highlights

Show the query list to the user before searching. Let them add, remove, or adjust. Don't silently batch queries.

Step 2 — Run vector searches

For each query, redirect output to a temp file — do not pipe to stdout:

readwise readwise-search-highlights \
  --vector-search-term "<query>" \
  --limit 30 --json \
  > /tmp/rwhl_query_<N>.json

Batch all queries in one bash call.

Step 3 — Merge, dedupe, group by parent document

jq -s '
  [ .[] | .[] ]
  | unique_by(.id)
  | group_by(.attributes.document_title + "|" + .attributes.document_author)
  | map({
      title: .[0].attributes.document_title,
      author: .[0].attributes.document_author,
      category: .[0].attributes.document_category,
      doc_tags: .[0].attributes.document_tags,
      match_count: length,
      top_score: (map(.score) | max),
      highlights: [ .[] | {
        id, score,
        text: .attributes.highlight_plaintext,
        note: .attributes.highlight_note,
        tags: .attributes.highlight_tags
      }]
    })
  | sort_by(-.match_count, -.top_score)
' /tmp/rwhl_query_*.json > /tmp/rwhl_grouped.json

Report to the user: number of unique parent docs, top 10 by match count. Ask them to confirm the inclusion threshold (default: match_count >= 2, or match_count == 1 && top_score > 0.5). Let them prune off-topic results before writing files.

Step 4 — Write per-doc raw files

For each parent doc passing the threshold, write raw/<slug>_highlights.md. The _highlights suffix distinguishes these from full-document raws.

jq -c '.[] | select(.match_count >= 2 or (.match_count == 1 and .top_score > 0.5))' /tmp/rwhl_grouped.json \
| while IFS= read -r doc; do
    slug=$(echo "$doc" | jq -r '(.author // "unknown" | ascii_downcase | gsub("[^a-z0-9]+"; "-") | .[0:20]) + "_" + (.title | ascii_downcase | gsub("[^a-z0-9]+"; "-") | .[0:35]) + "_highlights"')
    {
      echo "$doc" | jq -r '"# Highlights: " + .title + "\n\n**Author:** " + .author + "\n**Category:** " + .category + "\n**Document tags:** " + (.doc_tags | if length == 0 then "none" else join(", ") end) + "\n**Match count:** " + (.match_count|tostring) + "\n**Top score:** " + (.top_score|tostring) + "\n\n> Note: these are the matched highlights only, not every highlight in the doc. Re-run with more queries for broader coverage.\n\n---\n"'
      echo "$doc" | jq -r '.highlights[] | "> " + (.text | gsub("\n"; "\n> ")) + "\n" + (if .note != "" and .note != null then "**Note:** " + .note + "\n" else "" end) + (if (.tags // [] | length) > 0 then "*Tags: " + (.tags | join(", ")) + "*\n" else "" end) + "\n---\n"'
    } > "raw/$slug.md"
  done

Step 5 — Verify without reading the bodies

ls -la raw/*_highlights.md
wc -l raw/*_highlights.md

Do not cat or Read a highlights file unless the user asks.

Step 6 — Report, then chain into ingest

Report: number of parent docs landed, filenames, match counts, and any docs you dropped. Then invoke the ingest skill on all new *_highlights.md raws. Tell the ingest step these are highlight collections — cite individual highlights rather than re-paraphrasing.

JSON shapes (don't re-probe these)

• readwise-search-highlights --json → top-level array. Each item: id, score, attributes.{document_title, document_author, document_category, highlight_plaintext, highlight_note, highlight_tags}.

Rules

• Never run readwise-search-highlights with stdout going to tool output — always redirect to /tmp/.
• Never write JSON files to raw/ — all raw files must be markdown (.md). Temp JSON goes in /tmp/.
• Never Read or cat a raw/ file you just wrote unless the user explicitly asks.
• Always confirm the query set with the user before searching.
• Always chain into ingest unless the user said "just fetch."
• A 30-doc haul is probably too many; 5-15 is a normal batch.