@atlaskit/editor-plugin-autocomplete 9.0.0 → 9.2.0

package/CHANGELOG.md

@@ -1,5 +1,138 @@
 # @atlaskit/editor-plugin-autocomplete
 
+## 9.2.0
+
+### Minor Changes
+
+- [`027dc05ee4e0d`](https://bitbucket.org/atlassian/atlassian-frontend-monorepo/commits/027dc05ee4e0d) -
+  Suggest inline-code identifiers the session has seen, for hosts that opt in with the new
+  `harvestInlineCode` plugin option — which Rovo Chat now does.
+
+  A term like `ml-studio` cannot be suggested today whatever the session does with it:
+  `incrementSessionFreq` only touches trie nodes that already exist, so a word the shipped
+  vocabulary does not hold is silently discarded. The new harvester collects backticked spans from
+  ingested reply and page text and code-marked spans from the live document, drops anything with
+  internal whitespace, over 50 characters, not starting with a letter, or already served by L2/L3,
+  and holds the result to 200 surfaces with the least-mentioned evicted first.
+
+  A harvested surface has no frequency, no vector and no canonical token ids, so what stands in for
+  a score is where it is allowed to speak. It is offered only where the scored path has finished and
+  recorded that no vocabulary reaches the prefix at all — not when the 100ms budget expires, since
+  that means the ranker was still working. It asks for four typed characters rather than three,
+  respects the post-accept cooldown and the no-repeat check, and on accept rewrites the typed prefix
+  in the surface's own casing and marks it as code, because casing is load-bearing in code and being
+  marked as code is what authorized the suggestion. The stored mark is dropped straight after, since
+  the backtick input rule never ran and nothing else would close the span. It reports itself as a
+  new `harvest` value on `CompletionSource`, so its views and acceptances cannot move the headline
+  rate unattributed.
+
+  The harvester is a separate chunk, requested on first focus and only for a host that opted in, so
+  a host that has not stays exactly as it was and never downloads it. What it holds is readable at
+  any time with `__atlCtcDebug__.harvest()`, or `__atlCtcDebug__.harvest('ml-s')` to ask what a
+  prefix would be offered.
+
+  Both the harvested set and the L1 session boosts are now dropped when the session they describe
+  ends. A host whose editor does not last as long as what it is writing about can say so with the
+  new `contextScopeKey` on the context it returns: when the key changes, both stores are emptied and
+  the context reported with the new key is ingested as if it were the first, so anything still
+  current is primed again. The key is tracked beside the data it guards rather than per editor,
+  because a host may answer a navigation by replacing the editor instead of keeping it — read from a
+  new editor alone, every scope looks like the first one, while the previous scope's terms carry on
+  in the state they share. Teardown clears both stores as well, once the last editor has gone:
+  Confluence mounts several, and closing one reply must not wipe what the others primed.
+  `resetSessionBoosts` is exported for the same purpose, and `__atlCtcDebug__.session()` now lists
+  100 boosted words rather than 50.
+
+  Because those stores are shared, an editor that reported no scope is put back in the same tick an
+  eviction empties them, from the context it is already holding rather than a second read. A comment
+  box under a chat panel that switched conversation never asked to lose its page, and without this
+  it could not get it back either: the record of what an editor has ingested is per editor, so its
+  own page would read as already seen from then on. An editor that did report a scope is left alone,
+  since the scope being dropped may be its own — a page transition mounts the editor arriving before
+  tearing down the one leaving.
+
+  A context read requested while another is still open is now deferred until that one settles, where
+  before it was dropped. A scope change reaches the plugin as a context read, and a host that pushes
+  its context has no other channel to raise one on, so a read refused mid-navigation was an eviction
+  that never happened.
+
+  On the Rovo Chat side, chat messages now reach autocomplete through `extractCodeAwareTextFromAdf`,
+  the same extractor the page already uses. Assistant replies keep their backticks by arriving as
+  markdown, but a user's own inline code arrives as ADF and was being flattened by
+  `extractPlainTextFromAdf`, which drops the `code` mark — the only thing distinguishing an
+  identifier from a word. The L1 boost is unaffected, since its tokenizer strips backticks as
+  boundary punctuation. Each context read is stamped with the conversation and the page the body in
+  it came from, and the stamp is reported even when there is nothing else to report: arriving at an
+  empty new chat is exactly when the previous one's terms most need dropping.
+
+  The page in that stamp is taken from the body rather than from where the host says the reader is,
+  and a host that publishes the page it is on has its empty answer read as "no page" rather than as
+  no answer. Confluence in-app navigation is why: it clears the page it was holding immediately, but
+  the content id the chat reads can lag it by a long way and sometimes never arrives, and a body
+  accepted only when the two agree is a page that is never read at all. Attributing a body to the
+  page it came from cannot offer one page's text under another's, which is what agreement was being
+  required for.
+
+  A body published by an editor is no longer attributed to a page it was already attributed
+  elsewhere for. The published document carries no page id, so the page has to come from where the
+  reader is, and that is the same answer whether the editor has just published or the store is still
+  holding the document from the page before. Two live docs in a row is where it showed: arriving at
+  the second turns editor publishing back on and asks again, while the first one's body is the only
+  one there is.
+
+  Dropping the boosts when the last editor goes is limited to hosts that scope their context, which
+  are the ones asking for their learning to end with the conversation. A host that sends no scope
+  has boosts belonging to the page it is on, and a comment box closes far more often than that page
+  changes.
+
+  A host that cannot answer a page read now reports it, once per chat input, as an operational
+  event. The read still succeeds without a page, which is the right thing for a keystroke to do and
+  also the reason the failure is otherwise invisible: the editor's own monitoring only sees reads
+  that reject, and the symptom of this one — no page context on viewed pages — looks the same as a
+  page with nothing worth learning in it.
+
+  Dropping the session boosts now reads an index of the nodes that were boosted rather than walking
+  the whole word trie for them, since it runs on every page change rather than only at teardown:
+  about 10ms of main thread on a large tenant vocabulary, twice per navigation, for a few hundred
+  words.
+
+  Both the option and the extraction carry the same
+  `platform_editor_ai_autocomplete_rovo_chat_editor` decision the rest of the chat's autocomplete
+  wiring does. With the experiment off no context is collected and the harvester chunk is never
+  requested.
+
+### Patch Changes
+
+- Updated dependencies
+
+## 9.1.0
+
+### Minor Changes
+
+- [`a164c6d8e7de6`](https://bitbucket.org/atlassian/atlassian-frontend-monorepo/commits/a164c6d8e7de6) -
+  Offer the Confluence page the chat is open on to autocomplete as context, so words already written
+  on it carry an L1 session boost in the chat input. This reads the editor context the page editor
+  already publishes, so no host wiring is needed, and covers a page being edited; reaching a page
+  that is being viewed needs a host bridge that comes separately. The page text is held together
+  with the page it was read from, so navigating invalidates it in the same render that changes the
+  page rather than just after — the context store keeps the last published document until another
+  publisher replaces it, and that document does not say which page it came from.
+
+  Both the page and the chat transcript sit behind
+  `platform_editor_ai_autocomplete_rovo_chat_editor`, and the editor context subscription that reads
+  the page lives in a component the chat input only mounts once enrolled. An unenrolled session
+  therefore neither parses a page document for context it cannot use nor re-renders the input on a
+  publish, which the editor does on blur and on selection change.
+
+  Context text now waits for the vocabulary load to settle before it is applied. Session boosts were
+  previously spent against an empty trie whenever the page resolved first, which silently dropped
+  every word in it. What the session holds is readable from the console with
+  `__atlCtcDebug__.session()`.
+
+### Patch Changes
+
+- Updated dependencies
+
 ## 9.0.0
 
 ### Major Changes