v0.20.0

An assistant's search shows the chunks that match whole, never cut, and lists
the other documents it found, fifteen in all; browse lists a folder as a
tree within a budget; and every tool gives a document's date, length and
format the same way
(ADR-044,
still proposed). A project measures its search with the platform's own code.

Added

• search takes limit and compact: fewer documents, or titles,
  links, folders and dates alone.
• characters on every document a tool lists, the length of the text
  fetch returns, so an assistant knows what reading one costs; fetch
  gives the whole document's length however much it returned.
• fetch names its format, doc or pdf, and a PDF's pages, which
  the build reads from a new ctcdocs:pdf-pages meta on the document's page.
• browse takes depth, to cap the levels of its tree.
• mcp.browseCharacters, the budget of a browse tree, 24,000 unless a
  project sets another, and mcp.recent.defaultResults and
  mcp.recent.results, twenty and fifty: the numbers browse and
  recent used to fix.
• ctcdocs-eval-search runs a project's own questions,
  evaluation/search-questions.json, through the Worker's searchDocuments
  against its AI Search instance, under variants that change one setting
  each, and writes each run to evaluation/results/<date>-<label>/: rank,
  recall, and whether the passages of the documents a question needs show the
  words its answer holds. It writes no document text.

Changed

• search no longer cuts a passage. It takes the chunks that count in
  the index's order, across documents, and shows each whole while the budget
  holds, at most passagesPerResult of one document; a chunk that does not
  fit is left out and counted in morePassages, and a chunk a shown one of
  its document already holds is shown once. The other documents found are
  listed without text. Measured on one deployment's corpus, the answers
  shown rose from 23 to 25 of 32 and the recall of the list from 0.859 to
  0.916.
• mcp.search.results lists 15 documents unless set, not ten.
  mcp.search.passageCharacters takes 1,000 or more, no longer 100 for each
  result, so every configuration 0.19.0 accepted is still accepted.
• browse returns a nested tree: what is directly in the folder, then its
  folders level by level, those taking the fewest characters first, each
  listed whole or collapsed with its name and count. A document is listed
  with its id, title, day and length, and one links pattern replaces a URL
  on each entry. Without a folder it lists the whole knowledge base, as far as
  the budget holds.
• Dates are YYYY-MM-DD in search, browse and fetch, and to the
  minute in UTC in recent, instead of Drive's time to the millisecond.
• fetch puts metadata before text, and an unknown id says to find
  one with search, browse or recent.
• A note for the assistant is a note field of the answer, not a text
  item beside the JSON: Claude Code gives the model only a result's structured
  content, so it never saw the notes search and browse sent.

Upgrade note

Bump the packages and the workflow pins. Assistants see the new shapes after
their next tool listing; a script that parsed the old ones, search results
always carrying text or browse's flat lists, needs updating.

Every passage is now a whole chunk, so the instance's chunk size decides how
many fit the budget: about seven of AI Search's default 1,024 tokens. Measure
an instance of smaller chunks over the same bucket with
ctcdocs-eval-search --instance <name> before binding it; on the corpus
measured, 800 tokens ranked best in both of its languages. Keep
mcp.search.contextChunks at 0: neighbouring chunks now lengthen every
passage instead of being cut away.