Repository navigation
Release - #284
Open
pirate-bot wants to merge 20 commits into
Open
Release#284pirate-bot wants to merge 20 commits into
pirate-bot wants to merge 20 commits into
Conversation
…ash on it (#278) * fix(kb): load the tokenizer vocabulary through the HTTP API, never crash on it The tiktoken library fetches its vocabulary with fopen() on a remote URI and needs a writable temp directory. On a host with allow_url_fopen off or a locked temp directory it threw a RuntimeException that nothing caught, so every self-hosted add crashed the REST request (#241). A new Vocab_Loader streams the download with wp_remote_get() into the temp directory, or uploads when that is not writable, under the file name the library uses so an existing cache is reused. Streaming is the fast path: buffering the 1.7 MB body through the HTTP API took over a minute where streaming took seconds. With nowhere to write it buffers in memory as a last resort. Downloads are checksummed, written beside the final name and renamed, so a reader never sees a partial file. ingest_document() now reports a tokenizer failure as an error on the row instead of letting the exception escape. The encoder is built once per request. Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com> * fix(tokenizer): trim page context by characters when the vocabulary is unavailable Co-Authored-By: Claude Sonnet 5.5 <noreply@anthropic.com> * docs(page-context): describe the character-trim fallback Co-Authored-By: Claude Sonnet 5.5 <noreply@anthropic.com> * fix(kb): keep a tokenizer failure on the row and retry it from the cron QA on the PR build found the "could not be tokenized" reason only in the snackbar: the row showed nothing once it faded, and nothing after a reload. Development has since started recording every ingest error against the post (#276), so after the rebase the reason stays on the row and counts toward Needs attention. The update cron now records it as a retry rather than as something the admin must fix, since the post stays queued. The #275 encoder guard in pinned_chunks() dropped the chunk that needed a trim; with trim_to_limit() falling back to a character cut, the chunk is kept trimmed and the guard goes. Its test now blocks the download the way the tokenizer tests do, because an unusable temp directory alone no longer stops the vocabulary from loading. Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com> * fix(tokenizer): remember a failed vocabulary load for the rest of the request With ingest no longer crashing on it, update_posts() went on to the next post and tried the download again, up to five times a run; a slow CDN could block the cron for one timeout per post. The first failure is now kept and rethrown for the request. Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com> * style(tokenizer): satisfy PHPCS on the remembered-failure change Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com> --------- Co-authored-by: Claude Fable 5.1 <noreply@anthropic.com>
Contributor
Author
🌍 i18n String Review Report📊 Summary
➕ Added Strings (35) - Click to expand
➖ Removed Strings (14) - Click to expand
|
Bumps [codeinwp/themeisle-sdk](https://github.com/Codeinwp/themeisle-sdk) from 3.3.67 to 3.3.68. - [Release notes](https://github.com/Codeinwp/themeisle-sdk/releases) - [Commits](Codeinwp/themeisle-sdk@v3.3.67...v3.3.68) --- updated-dependencies: - dependency-name: codeinwp/themeisle-sdk dependency-version: 3.3.68 dependency-type: direct:production update-type: version-update:semver-patch ... Signed-off-by: dependabot[bot] <support@github.com> Co-authored-by: dependabot[bot] <49699333+dependabot[bot]@users.noreply.github.com>
) * fix(kb): keep tel: and mailto: link targets in knowledge base text A contact page often holds its phone number and email only as link targets: `<a href="tel:+441274998877">Call us</a>`. `html_to_text()` strips every tag, so the number never reached the knowledge base and the bot could not answer "what is your phone number?" from a page that shows one. Both modes are affected, since self-hosted and Connect ingest both go through the same extractor. The target of a `tel:` or `mailto:` link is now folded into the anchor text before the strip, as "Call us (+441274998877)", unless the text already shows it. Phone numbers are compared on digits from the end so national and international forms match, a `+`-prefixed different number is not taken for the target, `;ext=` becomes " ext. N", and an icon-only link stores the bare value. No other scheme is folded, so navigation URLs stay out of chunks. An unclosed contact link is bounded by the next anchor and cannot take the rest of the page with it. Connect reconcile hashes the extracted text, so already-synced pages with a contact link re-upload once after this ships. Fixes Codeinwp/hyve#350 Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com> * fix(chat): extract page context like the knowledge base Page awareness turned a scraped or rendered page into text with a plain tag strip, predating the structured extractor the knowledge base uses. A contact page not in the index lost its tel:/mailto: targets there, and tables and headings fused as they once did on ingest. Both page-context paths now go through Tokenizer::html_to_text(). Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com> * test(connect): let the vocabulary download through the sync test's HTTP mock The contact-link sync test answered every request with the canned SSE body, so on a runner with no cached vocabulary the tokenizer's download got that body, failed its checksum, and the failure was remembered for the rest of the run. The mock now answers only the Connect workflow. Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com> * fix(kb): address contact-link review findings - A quoted tel:/mailto: target may hold spaces ("tel:01274 998877", as some widgets print a phone field raw); only an unquoted one ends at whitespace. - The shown-number check tries each contiguous span of digit groups in a number-like run, so "01274 998877 (24 hours)" and "Mon-Fri 9-5 01274 998877" count as showing the number. A run with its own "+" country code is read from its start and must carry the whole linked number, so "+44 1274 998877" shows tel:01274998877 while "+1 274 998877" does not show a UK target. - An icon-only link's bare value is padded with spaces so it does not fuse with the text or link after it. - Connect reconcile rehashes an unedited post that was synced by an older extractor: the synced marker now records the extractor version (Tokenizer::EXTRACTOR_VERSION), the manifest reuses a stored hash only when that matches, and a post whose text is unchanged adopts the current version so later reconciles reuse its hash again. Without this, a contact page synced before the fold kept its old text on the platform until edited. - Tests: page-context coverage for the loop-page path, and the linear-time test now ends with a closed link that must be folded, so a regex that gives up fails it. Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com> * fix(kb): a coded label must be the same number as a coded tel: target "+49 441274998877" ends with the digits of tel:+441274998877, so the suffix compare took it for the target and dropped the fold. When both the label and the target carry a "+" country code they now have to be the same number. A label with a code may still stand for a national target that has none. The span loop is capped at 32 digit groups per run so a table of numbers inside one anchor cannot stall extraction. Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com> * fix(kb): read a 00 dialling prefix as a country code Only "+" marked a number as carrying its country code, so tel:00441274998877 with the label "+49 441274998877" (or the reverse pairing) was taken for the same number by the suffix compare and the target was dropped. A leading "00" now counts the same as "+" on both sides, so two coded numbers must match exactly. Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com> * docs(kb): say a 00 prefix counts as a country code in same_number() Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com> --------- Co-authored-by: Claude Fable 5.1 <noreply@anthropic.com>
…eval (#283) * feat(chat): fuse a keyword search with the vector search for KB retrieval A question with a term the embedding does not carry (a product code, a name, a rare word) scored below the floor and the chunk never reached the model. Retrieval now runs two passes: the cosine scan or Qdrant search as before, and a keyword pass over the chunk text through a FULLTEXT index on the local table (prefix-matched terms, substring matching for scripts without word boundaries). Each list keeps the matches within its own relative cutoff; the two are merged by reciprocal rank with a title bonus and then fitted to the token budget. The keyword pass runs on the local table in Qdrant mode too, and alone when the embeddings call fails, which no longer fails the turn. The trace records both scores per chunk. Weights and the candidate cap are filterable through `hyve_hybrid_retrieval`. The table version moves to 1.3.0 so dbDelta adds the index on upgrade. Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com> * fix(chat): keep the question first in the keyword query, run it in the ability, show keyword hits Review findings. The keyword pass reads ten terms, so a long thread could crowd the current question's words out; the message now leads the keyword query. The test-retrieval ability runs the keyword pass too, carries on keyword-only when embeddings fail and says so with `embedding_failed`, and its chunks carry `lexical`. The Messages debug panel labels a chunk only the keyword search found. Equal scores order by row id. The retrieval tests commit their cleanup so seeds and options no longer leak into other classes, and new tests cover the upgrade index, the parser edges, the fusion filter and the source order. Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com> * fix(messages): show the keyword score next to the vector score, keep the upgrade test's version committed A chunk both passes found showed only its vector score, so the debug panel could not tell it from a vector-only hit; it now reads "0.50 · keyword 1.20". The upgrade test's ALTERs commit implicitly, which left the downgraded 1.2.0 version behind after rollback; the upgrade is redone and committed in `finally`. Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com> * test(e2e): cover the dual and keyword-only source scores in the debug ledger Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com> * fix(chat): keep combining marks in keyword query terms, match marked words as substrings Bengali, Hindi and other abugida questions such as "মূল্য কত" produced no keyword terms: the term splitter dropped combining marks (vowel signs, viramas), breaking every word into fragments shorter than the FULLTEXT minimum. Keep marks in the words, and also match marked words as substrings, since InnoDB FULLTEXT can tokenize at marks inside a word. Covers the QA findings on Codeinwp/hyve#323. Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com> * fix(chat): reach precomposed content from decomposed accented queries Keeping combining marks in keyword terms broke decomposed input: a Vietnamese keyboard sends "giá" as "gia" + U+0301, and that term matches neither NFC content in FULLTEXT nor as a substring. A Latin word carrying marks now also adds its mark-free form, after the whole words so it never displaces one under the ten-term cap. Other scripts keep their marks. The long-input test now checks that a later marked word survives a long unmarked one, without a timing assertion. Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com> * test(chat): cover the three-character guard on the mark-free keyword term Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com> * fix(chat): match short Arabic, Hebrew and Hangul words when nothing reaches the index "سن من" has two two-letter words, below InnoDB's three-character FULLTEXT minimum, and Arabic script was not substring-matched, so the keyword pass found nothing for a chunk holding the phrase verbatim. Two-letter words in those scripts are content words; they are now matched as substrings, but only when no word in the question reaches the index: index scores scale with the corpus while a substring hit does not, so a preposition such as "در" would otherwise outrank and evict a real indexed match. Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com> * test(chat): cover the five-word cap and deduplication of short keyword words Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com> * fix(chat): match single-syllable Korean words when nothing reaches the index "내 집" is two one-syllable words, so the two-letter rule left the keyword pass empty. A Hangul syllable is a whole morpheme; accept one or two of them, still only when no word reaches the index. Single Arabic and Hebrew letters stay out: they are particles. Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com> * fix(chat): exclude lone jamo from short Hangul queries * fix(chat): retain exact keyword hits through rank fusion * fix(chat): link keyword-only sources only at the best query coverage With the keyword list no longer cut by the score ratio, every included keyword-only post became a source link, so a question matching one common word in many chunks linked them all. A keyword-only post now links only when its chunk matches as many question words as the best keyword candidate: indexed words at a word start, segments as substrings, accents folded. The score-ratio and similarity-threshold docblocks now say they apply to vector matches only. Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com> * fix(chat): ignore index stopwords when counting source-link coverage The FULLTEXT index never matches its default stopwords, so a chunk full of "what", "is" and "the" must not out-cover the answer for the source link. Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com> --------- Co-authored-by: Claude Fable 5.1 <noreply@anthropic.com>
…ns or fails (#285) * fix(qdrant): keep the knowledge base answering while the migration runs or fails Qdrant was marked active before the asynchronous migration ran, and chat then searched Qdrant only. A migration that failed, returned an unconfirmed upload, or never ran because WP-Cron was blocked left every unmigrated chunk unreachable, so a previously indexed knowledge base looked empty. The local table keeps every vector even after a row is marked Qdrant-backed, so retrieval now uses the local vectors while any processed chunk is still unmigrated, and falls back to them when the Qdrant search itself fails. The migration treats an unconfirmed upload as a failure, retries a transient failure after five minutes, ends after a terminal error that deactivated Qdrant, and only migrates processed rows. A batch holds a lock so a second cron run cannot upload the same rows alongside it, and migration points use a stable id derived from the row so a retry after a lost response overwrites instead of duplicating. The status endpoint reports the remaining count from the table and re-arms a lost cron event; the chat path does the same, so a site with blocked WP-Cron recovers from visitor traffic alone. The Qdrant client throws a bare Error on some connection failures, which the catch (\Exception) blocks never saw; every catch now takes Throwable and reports a readable message. Disconnecting Qdrant also clears its stored error, and the Integrations card shows the last error while migrating. Fixes #265 Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com> * fix(qdrant): reuse the existing connection error string Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com> * fix(qdrant): address migration recovery review findings * fix(qdrant): allow local disconnect and bound HTTP requests * chore: remove PR verification artifacts * fix(qdrant): state the lock's real guarantee and prove an overlap is harmless On a host without cURL, Guzzle's stream handler bounds each read, not the whole request, so an upload with a dripping response can outlive the migration lock TTL and the next run can take the lock while the first is still uploading. That overlap cannot duplicate knowledge: migration points carry a stable id derived from the row, so the repeat upsert overwrites the same points. The lock docblock now says so instead of promising exclusivity, and a test ages the lock past the TTL from inside the upload and asserts both runs sent identical point ids with every row migrated once. Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com> * fix(qdrant): do not count chunks without a vector as migrated A Connect import can insert processed rows with an empty embedding. The migration skipped them but still marked them as Qdrant-backed, so the progress shown in the Integrations panel counted chunks that were never uploaded. Rows with no vector are now excluded from the unmigrated query and count, so they are never pending and never counted as moved, and only rows actually uploaded are marked as Qdrant. A row holding an unreadable vector is normalised to the empty form so it is dropped rather than retried forever. Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com> --------- Co-authored-by: Claude Fable 5.1 <noreply@anthropic.com> Co-authored-by: copilot-swe-agent[bot] <198982749+Copilot@users.noreply.github.com> Co-authored-by: pirate-bot <58979018+pirate-bot@users.noreply.github.com>
) The model had no clock, so it guessed today's date, resolved "tomorrow" to the wrong day and asked visitors for the date. The base prompt also let it write its own apology in English to non-English visitors, describe retrieved excerpts as site navigation, assert that the site lacks something retrieval missed, confirm visitor claims, promise bookings it cannot make, and omit URLs that were in the context. apply_system_prompt() now appends a date line after the base prompt, spelled out as well as ISO because gpt-5.4-nano compares "25 October" against "28 October 2026" reliably but not against "2026-10-28". It goes after the static text so prompt caching keeps a shared prefix. The base prompt gains rules for dates, for what the model cannot know and cannot do, verbatim numbers, links from the context, the visitor's language, and worked examples for the cases small models got wrong. The "I apologize, but I couldn't understand" example is gone: unparseable input returns an empty response with success:false and the site's default_message shows. Connect chat settings send the site timezone so the platform dates its prompt the same way. Partial answers and "I can't do that in this chat, here is the contact" replies carry success:true so the visitor sees them; they drop out of the unanswered-question analytics as a result. Refs Codeinwp/hyve#327. Co-authored-by: Claude Fable 5.1 <noreply@anthropic.com>
Bumps [@wordpress/e2e-test-utils-playwright](https://github.com/WordPress/gutenberg/tree/HEAD/packages/e2e-test-utils-playwright) from 2.1.0 to 3.1.0. - [Release notes](https://github.com/WordPress/gutenberg/releases) - [Changelog](https://github.com/WordPress/gutenberg/blob/trunk/packages/e2e-test-utils-playwright/CHANGELOG.md) - [Commits](https://github.com/WordPress/gutenberg/commits/@wordpress/e2e-test-utils-playwright@3.1.0/packages/e2e-test-utils-playwright) --- updated-dependencies: - dependency-name: "@wordpress/e2e-test-utils-playwright" dependency-version: 3.1.0 dependency-type: direct:development update-type: version-update:semver-major ... Signed-off-by: dependabot[bot] <support@github.com> Co-authored-by: dependabot[bot] <49699333+dependabot[bot]@users.noreply.github.com>
OpenAI shut down gpt-4.1-nano on 23 October 2026, and has scheduled gpt-5 (11 December 2026) and the gpt-4o / gpt-4.1 families (14 January 2027). This is the fallback.
Bumps [codeinwp/themeisle-sdk](https://github.com/Codeinwp/themeisle-sdk) from 3.3.68 to 3.3.69. - [Release notes](https://github.com/Codeinwp/themeisle-sdk/releases) - [Commits](Codeinwp/themeisle-sdk@v3.3.68...v3.3.69) --- updated-dependencies: - dependency-name: codeinwp/themeisle-sdk dependency-version: 3.3.69 dependency-type: direct:production update-type: version-update:semver-patch ... Signed-off-by: dependabot[bot] <support@github.com> Co-authored-by: dependabot[bot] <49699333+dependabot[bot]@users.noreply.github.com>
Bumps [codeinwp/themeisle-sdk](https://github.com/Codeinwp/themeisle-sdk) from 3.3.69 to 3.3.70. - [Release notes](https://github.com/Codeinwp/themeisle-sdk/releases) - [Commits](Codeinwp/themeisle-sdk@v3.3.69...v3.3.70) --- updated-dependencies: - dependency-name: codeinwp/themeisle-sdk dependency-version: 3.3.70 dependency-type: direct:production update-type: version-update:semver-patch ... Signed-off-by: dependabot[bot] <support@github.com> Co-authored-by: dependabot[bot] <49699333+dependabot[bot]@users.noreply.github.com>
Bumps [codeinwp/themeisle-sdk](https://github.com/Codeinwp/themeisle-sdk) from 3.3.70 to 3.3.71. - [Release notes](https://github.com/Codeinwp/themeisle-sdk/releases) - [Commits](Codeinwp/themeisle-sdk@v3.3.70...v3.3.71) --- updated-dependencies: - dependency-name: codeinwp/themeisle-sdk dependency-version: 3.3.71 dependency-type: direct:production update-type: version-update:semver-patch ... Signed-off-by: dependabot[bot] <support@github.com> Co-authored-by: dependabot[bot] <49699333+dependabot[bot]@users.noreply.github.com>
* fix(kb): resume large imports after rate limits * fix(kb): surface rate-limit details and cover the import queue's edges - Show the account's x-ratelimit limit, remaining and reset values when an import gives up, so the admin knows what to raise or wait for. - Record a visible error on a document created by a failed enqueue instead of leaving a hidden post behind. - Clear the stored Qdrant warning once a staged publish or prune succeeds. - Share one polling hook between Indexed content and Needs Attention. - Make the update-cron test fail without the active-job filter, and add tests for a different document during a job, Needs Attention scope, the Qdrant warning, and the progress chip and queued notice in the browser. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> * fix(kb): recover every import, hide unpruned Qdrant points, name sync 429 limits - The watchdog pages through all active jobs instead of the newest 100, so a lost event on an older source is always rescheduled. - When an import gives up pruning the previous Qdrant generation after a day, search answers only from the new generation and the watchdog keeps retrying the prune. Deleting a source's points clears the mark. - Short documents indexed in the request now record the endpoint, model, HTTP status and token limit headers on failure, through the same OpenAI::describe_error() helper the queue uses. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> * fix(kb): keep a pending Qdrant prune from touching newer imports - The watchdog no longer retries a pending prune while an import owns the source; it takes the source lease and re-checks before deleting. Retrying around an older generation would have deleted the newer import's points. - When a new generation is revealed while an older prune is still pending, search switches to the new generation. - Stale generations are excluded inside the Qdrant query, so the result limit only counts visible points. Pending prunes live in one option, so search needs no per-hit lookup. - A later error response no longer ends a cooldown that an earlier 429 started; only a success does. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> * fix(kb): make pending Qdrant prunes safe under concurrency - Store each source's pending prune as its own post meta row instead of one shared option, so requests for different sources cannot overwrite each other. Search still reads them in one query. - Re-read a source's marker under its lease before retrying the prune, so an import that finished in the meantime is never pruned around the generation it replaced. - Move the marker to the new generation before recording the reveal, so a run that stops in between never leaves search naming a removed generation. - Remove the watchdog's page cap; the scan reads until a short page. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> * fix(qdrant): drop pending prune markers when Qdrant is deactivated A marker left behind would hide, then prune, the points a later migration uploads without a generation, leaving the source with rows marked processed but nothing in Qdrant. The job-scan test now fakes 102 pages through posts_pre_query, so it fails with the old 100-page cap, and the test-only page-size parameter is gone. A new test covers a marker cleared while the watchdog prunes another source. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> * fix(qdrant): never write a prune marker once Qdrant is deactivated An import still waiting on a prune when another request deactivates Qdrant could write a marker afterwards. On reconnection it would hide, then prune, the migrated points. set_current_generation() now reads the Qdrant status from the database and refuses to write while inactive, and connecting Qdrant clears any marker left from an earlier connection. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> * fix(qdrant): tie prune markers to the Qdrant connection they belong to A run waiting on a prune while Qdrant was deactivated and connected again could still write its marker for the new connection, and re-saving settings on a live connection cleared valid markers. Each connection now gets an ID when it is made from an inactive state. An import pins that ID when its upload starts, markers store it, and a marker is only written or read for the current connection. Re-running init() on a live connection keeps the ID and the markers. Fresh connections and deactivation still remove leftover marker rows. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> * fix(qdrant): upload again when Qdrant is reconnected mid-publish An import waiting to retry its prune had already uploaded and revealed its generation, but only on the connection it started on. After a reconnection to a new target, the retry went straight to the prune and deleted the source's migrated points, leaving search with nothing. When the pinned connection no longer matches, publish() now restarts the Qdrant upload on the current connection before revealing or pruning. A new test also covers a marker stored for another connection. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> * fix(kb): only blame the API limit when an import waits on it QA found the progress chip saying "waiting for the API limit" after a Qdrant upload error. Any retryable failure schedules a retry, so the chip blamed the rate limit for errors that had nothing to do with it. The job now records why it waits, and the public summary carries it as `wait`: `rate_limit` for a 429 or a learned budget, `retry` for any other retryable error. The chip says "retrying after an error" for the latter. The summary field is additive, so Hyve Pro is unaffected. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> * refactor(kb): build the import progress label from one lookup The six chip labels differed only in their string, so they now live in one table keyed by phase and wait reason, with a single sprintf call. Each string stays a literal __() call with its translators comment, and make-pot still extracts all six with their comments. The valid-sprintf rule cannot follow the lookup, so that one line carries a reasoned suppression. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> --------- Co-authored-by: Claude Opus 5.5 <noreply@anthropic.com>
* feat(chat): rewrite follow-ups into standalone retrieval queries Retrieval used to embed the visitor's last six turns joined with the new message. Follow-ups kept their topic, but a new question was searched mostly as the old ones, so retrieval stuck to earlier topics. A follow-up is now rewritten by one small completion (gpt-5.4-nano, filterable via hyve_retrieval_rewrite_model) into a standalone question, and only that is embedded. A first message is searched as sent. A long message keeps its own wording ahead of the rewrite so a condensing rewrite cannot drop its details, and the keyword pass always starts with the visitor's own words. A failed rewrite searches the message as sent. The debug trace records the searched query and whether it was rewritten, on the self-hosted path and from the platform's job_complete on Connect. Refs Codeinwp/hyve#324 Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> * fix(chat): keep long rewrites whole and cover the rewrite trace A short follow-up can expand past 300 characters with the subject of a long earlier question. The rewrite was cut there and, being short, was searched alone, so the tail of the question was lost. The rewrite cap is now 800 characters, and a separate 300-character threshold decides when a long message keeps its own wording ahead of the rewrite. Behaviour at 300 and 301 characters is unchanged. This mirrors the platform's RetrievalQuery::MAX_CHARS and LONG_CHARS. The Connect trace docblock now names the searched query, the Connect debug test fails without the change, and an e2e case covers the "Query (rewritten)" label. Refs Codeinwp/hyve#324 Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> * fix(chat): reject a rewrite cut off by the token limit OpenAI marks a completion that hit max_completion_tokens with finish_reason "length". Its first line can stop before the subject the rewrite was meant to add, and it was then searched alone. Such a rewrite now fails like any other, so the visitor's message is searched as sent. Refs Codeinwp/hyve#324 Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> * fix(chat): retry the rewrite with the chat model when the key cannot use it A key or project barred from the rewrite model answered every follow-up with model_not_found or permission_denied. The next successful call cleared the saved error, so the site silently searched raw follow-ups with no dashboard notice. Retry once with the site's chat model, which every turn already uses. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> * test(chat): share the self-hosted OpenAI test setup Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> * fix(chat): keep the resolved subject of a long follow-up's rewrite A follow-up over 800 characters with its pronoun at the end comes back as the whole message with the subject resolved at the end. The flat 800-character cap cut that subject off, and the flat 400-token limit cut longer or non-Latin rewrites short. Both now grow with the message. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> --------- Co-authored-by: Claude Opus 5.5 <noreply@anthropic.com>
… batches forever (#292) * fix(connect): send WPBakery raw HTML as text and stop retrying failed batches forever When WPBakery's shortcodes are not registered in the request that extracts a page, [vc_raw_html] blocks reach the knowledge base as a base64 blob that the embeddings provider rejects. The sync treated that as a temporary error and resent the same batch on every cron run. Tokenizer::html_to_text() now decodes [vc_raw_html] blocks and drops [vc_raw_js] before shortcode markers are removed; EXTRACTOR_VERSION is bumped so reconcile rehashes unedited posts once. connect_run_sync() classifies batch errors: per-document 422s fail only the named sources, auth and site-setup faults pause the job, provider outages pause after a bounded retry budget, and other failures mark the batch's sources with _hyve_processing_error so the job moves on. The Connect and Knowledge Base screens show why a sync is paused. Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com> * test(connect): take the batch order from the query in the 422 test Posts created in the same second have no stable order, so the rejected document index pointed at a different post on the CI database. Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com> * fix(connect): keep concurrent edits pending, resume a paused sync by cron and show the pause reason An edit saved by another request during a failing upload only reached the database, so the stale meta cache let the failure write it off. Each source now carries an edit mark (save time plus re-index flag) re-read past the cache. A paused sync schedules a cron event at its resume time instead of waiting for an admin load, the live mode is read through the settings defaults, and both admin screens name the pause and its reason. Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com> * fix(connect): lift the re-index flag for the batch and re-arm an unanswered pause A flag set before the batch was built describes content the batch already carries, yet it shielded the source from a terminal failure, and a re-save in the same second was indistinguishable from it. The batch now lifts the flag on its sources and restores it wherever a source stays pending, so a flag found after the request can only mean an edit made meanwhile. A paused job whose cron check finds the platform still down schedules another wait. Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com> * fix(connect): lift the flag before reading content and settle edited-page failures The batch lifted a source's re-index flag after rendering it, so an edit landing while another source rendered was lost. The lift now precedes the read, in the batch and in the update cron, and a flag found afterwards is a newer edit that keeps the source pending. The update cron stands down while syncing is paused, pauses the site on credential, setup and outage failures instead of re-sending an edited page every minute, and fails a refused page once. A timed pause check only re-arms its cron event, never rewrites the status, and re-arms when the plan is full as well as when the platform is down. Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com> * fix(connect): preserve retry budgets across delayed syncs * refactor(connect): drop the per-save edit marker from the update cron The marker was written on every save to guard a rare overlap of two update-cron runs. The flag claim and the re-check after a failed upload already cover an edit made during an upload; the remaining overlap now re-sends the page once at most. The retry counter stays in post meta, written only while a page keeps failing. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> * fix(connect): retry temporary failures on edited pages and never requeue a removed page A rate limit or maintenance answer on an edited synced page paused the whole site at once; it now gets the batch path's bounded retries first. A removal finishing during the settling probe had its flag written back, so the next cron run re-uploaded the removed page: the flag is now restored only while the page is still in the knowledge base, and in Connect mode the update cron only selects pages that are. The failure handler's docblock lists its outcomes. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> * fix(connect): keep an outage's spent retries across a resume While uploads keep failing but the stats route answers, every resume gave an edited page, and the batch job, a fresh retry budget, so a lasting outage was re-sent five times per pause cycle forever. The spent budget now survives the pause: after a resume one upload probes the outage and the job pauses again if it persists. A stored upload or a new edit clears the count. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> * fix(connect): probe an outage after a resume without spending other retries Carrying the spent retry count across an outage pause made a single 500 after the resume fail an edited page for good, and the resume set the count in a second write a running sync could race. A resume after an outage now seeds a probe marker in the one status write that makes the job runnable. While probing, only another rate-limit or maintenance answer pauses at once; server errors and timeouts keep their retries. The first stored upload, from a batch or an edited page, ends the probe. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> * fix(connect): bound the outage probe and end it on any other answer The probe only ended on a stored upload, so a resume that sent nothing or got another answer left it armed, and one 503 much later paused the whole site without retries. The probe is now a deadline one pause wait ahead, ends on any upload answer other than another rate limit or maintenance answer, and ends when the resumed update run finds nothing to send. Ending it rereads the stored status past the request cache and changes only the probe, so a pause another cron process stored survives. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> * docs(connect): say a transport failure also ends the outage probe The comments said the probe ends because the upload route answered, but a timeout or refused connection ends it too, with nothing answering. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> * fix(connect): show a Sync paused chip on the Connect screen while the sync is paused Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> * fix(connect): show Sync paused over Offline when an outage pause also fails the stats call Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> --------- Co-authored-by: Claude Fable 5.1 <noreply@anthropic.com>
Bumps [codeinwp/themeisle-sdk](https://github.com/Codeinwp/themeisle-sdk) from 3.3.71 to 3.3.72. - [Release notes](https://github.com/Codeinwp/themeisle-sdk/releases) - [Commits](Codeinwp/themeisle-sdk@v3.3.71...v3.3.72) --- updated-dependencies: - dependency-name: codeinwp/themeisle-sdk dependency-version: 3.3.72 dependency-type: direct:production update-type: version-update:semver-patch ... Signed-off-by: dependabot[bot] <support@github.com> Co-authored-by: dependabot[bot] <49699333+dependabot[bot]@users.noreply.github.com>
The "Docs" link (header bar and dashboard help card) and the Qdrant "Learn more" link used old HelpScout URLs (/article/2009-…, /article/2066-…). They redirect to the Hyve docs, but the redirect drops the query string, and the links had no UTM parameters anyway. Both now point at the current docs pages through tsdk_utmify(). Part of Codeinwp/docs-site#1596. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
fix: point the docs links to the current Hyve docs pages with UTMs
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
Linked issues
This release will close the following issues once merged:
Public changelog