Skip to content

Release candidate - #152

Draft
jwrosewell wants to merge 106 commits into
mainfrom
rc
Draft

jwrosewell wants to merge 106 commits into
mainfrom
rc

Conversation

@jwrosewell

@jwrosewell jwrosewell commented Nov 15, 2026 •

Copy link
Copy Markdown
Contributor

Releases what has collected on rc since the last release, #83, which merged into main on 13 November 2026 at 10:49 UTC. Its Publish run 43412143722 failed in one example test before it published anything, so npm still has 4.5.102 of every package, and this release takes #83's changes to npm too. Everything #83 listed is listed here again, because no one has had it yet.

This is the first release in which each package has its own version (#37). A package is released when its own code changed since its last release, or when a package it depends on is released, and only the patch number moves. Since 4.5.102 every package has changed or depends on one that has, so all 16 are released, each at 4.5.103.

This release comes four days after #83 merged, sooner than the week common-ci's CONTRIBUTING.md asks between releases, and its date is Tuesday 17 November. npm last had a release of these packages on 4 November, so the Publish workflow's weekly check, which counts from each package's last release, does not refuse it.

This release moves Node.js toward one behavior across every 51Degrees language. Each element it changes now follows the published 51Degrees specifications, which in turn follow the RFCs they cite, so that the same input gives the same answer in .NET, Java, Node.js, Python, PHP, Go, Rust and C. The native modules take the C library at 178371ac, the head of the C library's release draft 51degrees-cxx#191, which is not yet released, so this note gives it no version number.

Conformance with the specification

This release brings Node.js into line with the 51Degrees specifications at 706de1f0 in the places below. Each area quotes the sentence it turns on, and each pull request quotes the sentences it meets in full.

Evidence and caching

"An Element MUST accept only the Evidence that can change its results, because the Pipeline's accepted Evidence decides the cache keys, the ETag and the Vary header." (advertize-accepted-evidence.md lines 25 to 28)

"Keys MUST be case-insensitive and will usually be split into <prefix>.<field>." (evidence.md line 10)

"All Flow Elements and Pipelines MUST be capable of programmatically describing the Evidence keys that they can make use of." (advertize-accepted-evidence.md lines 5 and 6)

"Cache keys MUST always be generated deterministically." (caching.md lines 66 and 67)

  • Evidence keys and Property names are compared ignoring letter case, so header.User-Agent and header.user-agent are one key, and flowData.device.get('IsMobile') works in any letter case (#71, #68).
  • An Engine's cache key is its accepted Evidence sorted by key, with each key in lower case and each value as given, made once for each request before processing, so the same Evidence added in another order or with keys in other letters is served from the cache (#125).
  • Element Data is found by its key in any letter case, so flowData.get('Device') and flowData.Device work as flowData.device does (#122).
  • The Sequence Element counts a given query.sequence only when it is a BigInt or text made only of digits, from 0 up to 2,147,483,646. Any other value gives a sequence of 1 (#134).
  • A Pipeline describes the Evidence it accepts through a new evidenceKeyFilter, which accepts a key when any of its elements does, and FlowData.getEvidenceDataKey gives the Evidence the Pipeline accepts, where it always threw (#117).
  • Until discovery has answered, the Cloud Request Engine accepts no Evidence and keeps every key it is given, and the two on-premise engines do the same until their data file arrives, so an engine that cannot yet say which Evidence changes its results no longer accepts every key. What is kept still reaches the cloud service and the engine's cache key, so two early requests with different User-Agents do not share one cache entry (#182).

Web integration

"The Sequence Element MUST be present. If not, it MUST be added as the first element in the Pipeline." (web-integration.md lines 33 and 34)

"The ETag MUST be sent as a weak entity tag, W/"...", because it is not a strong validator." (web-integration.md lines 273 and 274)

"If this feature is enabled, then all Flow Elements that maintain any such state information MUST ensure that they have the means to keep track of isolated instances of this state for each Pipeline to which they are added and use the correct state information when processing." (multi-pipeline-elements.md lines 16 to 19)

  • Helpers.writeClientSideResponse writes the response of a client-side JavaScript or JSON endpoint with Cache-Control, Vary and an ETag. Node.js has no web integration package, so call it from your own endpoints, as the web examples do (#44).
  • The ETag is a weak entity tag, W/"...", and If-None-Match is compared weakly, as RFC 9110 section 8.8.3 and section 13.1.2 ask. A match gets 304 for GET and HEAD, and 412 for any other method. A response to a request whose processing failed has Cache-Control: private, no-cache and no ETag. The same Evidence gives the same ETag in every process (#44).
  • Evidence.addFromRequest adds the web request as a read-only view under server.http-request (#45).
  • Helpers.writeClientSideNotModified answers a request whose If-None-Match matches before the Pipeline processes it, with 304 for GET and HEAD and 412 for any other method, as the web integration specification asks. A 304 carries a good response's Cache-Control, Vary and ETag, during an outage too, and writeClientSideResponse gives a matching request the same answer (#120).
  • A Set Header Property's identifier may hold digits and other characters, and the header name starts at the next upper-case letter, so SetHeaderBrowser2Accept-CH sets Accept-CH (#120).
  • PipelineBuilder adds the Sequence Element, the JSON Builder, the JavaScript Builder and the Set Headers Element only where the Pipeline has none with that key, compared ignoring case, so one the caller added is the one used, and building twice adds none of them twice (#105).
  • One Set Headers element added to several Pipelines gives each Pipeline the headers its own elements set, where it gave every Pipeline the first one's (#177).

JavaScript Builder and JSON Builder

"Each Element MUST: ... Publicize details of the Properties it will populate." (conceptual-overview.md lines 73 to 78)

"Where there is no Sequence Element, the sequence MUST be rendered as 1 and the session id as an empty string, so the script always parses." (javascript-builder.md lines 60 to 62)

"The keys MUST be written in lower case, as the other keys of the JSON Builder Element are." (weighted-values.md lines 56 and 57)

  • The JavaScript Builder checks the callback URL before it writes it into the client script. The protocol must be http or https, and the host must be ASCII labels or an IPv6 address in brackets, with no user information, query, fragment or zone identifier, following RFC 3986. A URL from the request that fails is left out of the script and logged (#31, #51).
  • The script always carries the page request's query parameters, whether or not the builder has a callback URL, as step 7 of the JavaScript Builder specification asks (#97).
  • The script takes its session id and sequence from the Sequence Element's data, with sequence 1 and no session id where there is no Sequence Element, and the JSON writes each item of a list of Element Data by the same rules as the Element Data's own Properties (#118).
  • The JSON sent to the browser writes each item of a weighted value with rawweighting and value, keeping weight and rawWeight beside them for this release (#139).
  • The JSON Builder publicizes the Property it populates, json, and the JavaScript Builder publicizes javascript, each a string with the description the specification's own table gives. Neither published any Property metadata, so an application reading the Pipeline's metadata was not told either element produces anything (#187).

Cloud Request Engine and cloud engines

"The Engine MUST attempt both discovery requests when it is built (in its builder or constructor), and MAY make them in parallel, so that in the normal case a built Engine is fully initialised and immediately ready to process Flow Data." (cloud-request-engine.md lines 241 to 244)

"Unlike a Resource Key it names no Properties of its own, so a caller that authenticates with a License Key alone MUST also supply the list of Properties it wants." (cloud-request-engine.md lines 67 to 69)

"Similarly to the errors array, any entries in the warnings array in the response MUST be logged as warnings." (cloud-request-engine.md lines 460 and 461)

"A license key MUST NOT appear in full in a log entry or in an error message, at any log level." (logging.md lines 20 and 21)

"Where a WKT value is written to JSON, for example by the JSON Builder Element or in a response from the remote server, it MUST be written as an object with one member, value, holding the WKT text, for example {"value": "POINT(2 4)"}. ... A reader MUST accept this form." (value-types.md lines 123 to 131)

  • The Cloud Request Engine can authenticate with a License Key, alone or with a Resource Key, through licenseKey. With a License Key alone the Properties wanted must be named with requestedProperties. Both keys now travel in the request body (#64).
  • A License Key appears in log entries and error messages only as its last 4 characters, with the rest as *. The packages mask a key where one can enter a message, so the mask no longer runs on each detection, and Pipeline.log passes the message it is given to its listeners unchanged (#58, #85).
  • A cloud response that holds a 51Did is never stored in or served from the cache that cacheSize sets (#57).
  • The Cloud Request Engine fetches its accessible Properties and accepted Evidence keys before its first request, whatever else is in the Pipeline, so a Pipeline with a cache and no Cloud Engine serves a repeated request from the cache (#73).
  • Each entry of a cloud response's warnings array, from discovery or processing, is logged as a warning with any License Key masked, and a product the Resource Key does not include is logged once for each engine at warning level (#123).
  • The IP Intelligence cloud engine gives each weighted value as value, weight and rawWeight, as the on-premise engine does, keeping the service's rawweighting beside them for this release (#124).
  • An empty answer, or one whose errors array has entries, fails the request with the specification's messages, which name the URL (#140).
  • The Cloud Request Engine makes both discovery requests when it is built, in parallel, rather than when the Pipeline first processes, so a built engine is ready to process. One attempt is shared by the build and by any process that starts while it is under way. A transient failure at build, being no answer or a 5xx, logs one warning, completes the build and is tried again on first use (#180).
  • A Cloud Request Engine can be built from a saved copy of a discovery answer, with cloudDiscoveryState, so a process that already has one makes no discovery request, and a saved copy that names no products still makes them (#183).
  • The IP Intelligence cloud engine reads a WKT value the service sends as that object into its WKT text, for a value on its own and for each weighted item, so it gives Areas and Location as the on-premise engine does and either engine can take the other's place (#199).

Usage Sharing

"The Usage Sharing Element MUST NOT accept any Evidence." (usage-sharing-element.md line 8)

"Repeated failed sends MUST pause sending, so that a service that is down or overloaded is not sent a message for every batch, and so that usage data waits in the queue rather than being sent and lost. Only sending pauses. Processing MUST NOT wait for, or be stopped by, the back-off." (usage-sharing-element.md lines 334 to 337)

  • Usage Sharing accepts no Evidence, so a Pipeline that holds it no longer accepts every header, and a header no other Flow Element reads takes no part in a cache key, the ETag or the Vary header. The Flow Data still keeps what Usage Sharing shares. evidence.add keeps a value when a Flow Element accepts its key or when Usage Sharing would share it. An application that adds its own Evidence therefore has nothing to change, and the values it adds are shared under the same rules as the values of a request. evidence.addFromRequest also adds the whole request as server.http-request, which is Constants.evidenceHttpRequest in fiftyone.pipeline.core, a frozen object that holds each value of the request under the key it has as Evidence. Read it with evidence.get('server.http-request'). evidence.getAll() leaves it out, and it is never shared whole (#45).
  • After 10 failed sends in 100 seconds, sending pauses, and the pause doubles up to 300 seconds while sends keep failing. A new trackSession option tells sessions apart (#36, #35).
  • The record names the version of the Pipeline that wrote it, read from fiftyone.pipeline.core, where every release wrote 4.4.7, so the spread of a release can be measured. The session id and sequence are recorded as SessionId and Sequence whatever the letters of their keys (#109).
  • The record marks a value with replaced characters escaped="true", where it was replaced="true", and cuts a value at 10,000 code points, where it was 512 characters (#54).
  • A new shareAllQueryStringParameters option, true unless set to false, shares every query string and form value. Set to false, only the names that start 51D_ and those in queryWhitelist are shared. No cookie but 51D_ ones is shared, shareAll included (#135).
  • With shareAllQueryStringParameters set to false, a queryWhitelist of null is an empty list and shares only the names that start 51D_, where it shared every query string and form value, as the filter of the specification shares no parameter that is not on the list once that option is off (usage-sharing-element.md lines 71 to 83, #204).

Data updates

"The file system watcher runs on its own setting, whether autoUpdate is enabled or disabled, because watching a file downloads nothing." (data-updates.md lines 104 and 105)

"These messages MUST also be consistent across languages to make life easier for 51Degrees support." (data-updates.md lines 481 and 482)

  • Data update log entries and errors use the text and levels of the specification's error table, which every language now shares. Every check reports a status, and an updateComplete listener that throws is logged (#69).
  • With updateOnStart and a License Key or an update URL, an on-premise Engine downloads a missing Data File, where the build threw. DataFileUpdateService.updateFromMemory is new (#52).
  • The file system watcher runs on its own fileSystemWatcher setting, whether automatic updates are on or not, and needs no License Key. This reverses a change on rc that made the watcher wait for automatic updates, and so for a License Key (#52), which was a mistake. That change never reached npm, so the watcher behaves as in 4.5.102 (#90).
  • A data file has one polling timer. An update started by hand with updateDataFile replaces the waiting timer rather than adding another, so the file is checked once in each polling interval however many updates are started (#63).
  • A file system watcher event that arrives while the data file is updating is reported as AUTO_UPDATE_IN_PROGRESS, where it was dropped with no report (#119).
  • A refresh from a data file is tried 10 times, 200 ms apart, and the message says "after 10 tries", and the file system watcher waits up to 10 minutes for a changed file to open before it refreshes, reporting the new AUTO_UPDATE_TIMEOUT if it never does, as the .NET service does (#157).
  • A downloaded data file that cannot be read is reported as AUTO_UPDATE_UNKNOWN_ERROR, the status of the last row of the specification's table of data update errors, with that row's message, where it was AUTO_UPDATE_MASTER_FILE_CANT_RENAME, or AUTO_UPDATE_ERR_READING_STREAM for a compressed download. Both old names stay, marked deprecated (#205).

Properties and missing Properties

"When the user attempts to access a Property that is defined for an Engine, but does not exist in the result set for some reason, the system MUST throw an error/exception with a message explaining why the Property is not available." (properties.md lines 116 to 118)

  • Reading a Property an Engine has but that its results lack throws the message the specification's table gives for the reason, which is decided in the specification's order, with names compared ignoring case. The messages it replaces are kept, deprecated, and no longer thrown (#163).

Device Detection and IP Intelligence

"After loading the data source into the native code, call the getKeys function to return a list of the accepted Evidence keys. These values will then need to be stored to prevent repeated calls to the native code. This MUST be done at start-up and any time the data is refreshed." (device-detection-on-premise.md lines 81 to 86)

"A value converted from Azimuth or Declination MUST be given to the caller rounded to 3 decimal places, whether it is given as a number, as a string or as a point in a WKT shape." (value-types.md lines 99 to 101)

"The Engine MUST be able to give the ids of the profiles that hold a given value of a given Property, such as a TAC or a native model name, through the native code's EngineHash::getProfileIdsForPropertyValue()." (device-detection-on-premise.md lines 479 to 482)

  • Both on-premise engines read their accepted Evidence keys again after each data refresh, as the specification asks, so the keys a new data file accepts are used from then on (#130).
  • Each match metric Property of the on-premise Device Detection engine has the category "Device Metrics", the data files "Lite", "Premium", "Enterprise" and "TAC" and a default value, and they are listed in the order of the specification's table, after the data file's Properties (#137).
  • The on-premise Device Detection engine gives the ids of the profiles holding a value of a Property through the native lookup, matching the name and the value exactly, and takes the Properties to index (#174).
  • Each match metric Property of the on-premise Device Detection engine lists its possible values through getValues and getNumberOfValues, as a data file Property does, Method giving NONE, PERFORMANCE, COMBINED and PREDICTIVE from the specification's table (#176).
  • Each profile the cloud hardware lookup gives is device data read through get, with the letter case handling the on-premise device data has, and keeps every Property as a member, so a caller that reads a member is unaffected (#178).

51Did

"A reader MUST refuse a Payload whose version it does not know." (identifier-layout.md line 260)

  • A 51Did payload too short for its header whose flags byte names version 1, 2 or 3 is refused as an unsupported version naming that version, as one of normal length is (#115).

Derived properties

"Every reference implementation MUST provide one example for the Derived Property Element." (required-examples.md lines 30 and 31)

  • A new package in this repository, pipeline/fiftyone.pipeline.derivedproperty, adds a DerivedPropertyElement that runs every script of derived-properties, so a Node.js pipeline can offer HumanConfidence and IsGdpr, as .NET, Rust and PHP do. It is not published to npm in this release, as when to publish a new package is a separate choice (#98, #154).
  • fiftyone.pipeline.derivedproperty has the derived property example the pipeline specification asks for. It needs no key, data file or network, and shows a script held as text, a script read from a file, the same Pipeline built from a configuration file, and the message a caller sees when a source Property is missing (#155).

Translation

"SourceElementDataKey : The Element data that the Translation engine retrieves, that contains values to be translated from." (translation-engine.md line 6)

  • TranslationEngine accepts the specification's option names, sourceElementDataKey and missingTranslationBehavior. sourceDataKey and behavior still work and are deprecated (#49).

Configuration

"Configuration files MUST be human-readable and consistent between languages." (pipeline-configuration.md line 50)

  • A list setting may be given as one comma separated string, as the pipeline configuration specification says a configuration file gives a list. The engines' restrictedProperties and usage sharing's queryWhitelist and headerBlacklist read it as its items, through the new listSetting of fiftyone.pipeline.core, and an array works as before (#129).

Errors

"In all cases, standard language features SHOULD be used to let IDEs and users know what exceptions/errors a particular function can throw." (exception-handling.md lines 17 to 19)

  • Each public function of fiftyone.pipeline.core that throws says what it throws, and when, in a typed JSDoc @throws tag, which the TypeScript types carry. Nothing throws anything different (#166).

Examples

"Those tests MUST include a run of each example with no arguments or settings, as someone would run it from a fresh clone, so that the run uses the free data files." (required-examples.md lines 20 to 22)

  • Every on-premise example runs with no settings on the free Data Files, which it downloads the first time. A Property that only the Enterprise Data File holds is shown as needing that file, with a link to ask for a copy (#32).
  • Every JavaScript example, console and web, ends with a Find out more section giving the repositories and tagged 51degrees.com pages to follow up, including where it stops early for want of a key or a data file (#103).
  • Each example's 51d.json names its engines by package, so the same file works in this repository and with the installed packages, where eleven named them by a path into the source (#110).
  • The IP Intelligence mixed API example runs with no settings on the free Device Detection and IP Intelligence Lite files, where it searched the disk for paid files first. It names each Property only the Enterprise data files hold as needing them, with a link to ask for a copy (#111).
  • The data update examples check for a License Key before they look for a data file, the IP Intelligence ones use the small ASN file, every automatic update example ends with status 0 when it succeeds, and both update data file examples read _51DEGREES_LICENSE_KEY first. The cloud metadata example says a resource key is needed when none is set, and the match metrics example prints a documentation link a person can follow (#101).
  • fiftyone.pipeline.cloudrequestengine has the star sign Cloud Engine example the pipeline specification asks for, using the real Cloud Request Engine against a star sign service it starts on the same machine, so it needs no key or network. The simple star sign example takes a date of birth as its argument (#121).
  • The IP Intelligence on-premise web example shares usage when its configuration names the usage sharing element, as its own 51d.json does and as the Device Detection one does, with a comment saying production should leave usage sharing on. The console examples still turn it off (#131).
  • Every cloud example run with no Resource Key ends without an error and with status 0, names _51DEGREES_RESOURCE_KEY, gives a tagged link to where a key is created and ends with its Find out more section, and a test runs each of the 20 with no key (#133).
  • Both IP Intelligence getting started console examples print each value of the Registered Country with its weighting, and a single value with weighting 1 (#171).
  • The IP Intelligence cloud, cloud mixed and on-premise mixed web examples share usage, as the other web examples do, each saying that production should leave usage sharing on (#181).
  • Both update data file examples wait for the status that ends an update check. With the file system watcher on, the example's own download raises a watcher event that is reported as In Progress before the download's own result, as the specification asks, and the examples had taken that first report as a failure and abandoned the run (#197).
  • The IP Intelligence metrics example no longer writes EquivalentRadiusKm, a distance worked out from an area that the data does not give, and three examples label AccuracyRadiusMin in full where they wrote "Accuracy Radius". A test fails any example about where an IP address points that calls the area a circle (#200).
  • The five examples that started a web server on every network interface, four by naming no host and one by naming 0.0.0.0, now listen on 127.0.0.1, so each serves only the machine it runs on. One of them, the IP Intelligence mixed API example, serves the data file in use, so an Enterprise data file set in _51DEGREES_IPI_PATH was offered to every machine that could reach it (#201).
  • The IP Intelligence country translation console example and the Device Detection User-Agent Client Hints web example read _51DEGREES_IPI_PATH or _51DEGREES_DD_PATH before looking in the repository for their free data file, as every other example already did, so a machine that already has the file uses it (#203).
  • The two Device Detection cloud console examples and the usage sharing example read 51d.json from their own folder, so each runs from any folder, where run from another they threw before saying that no key was set (#202).
  • The TypeScript getting started example asks for the three Properties only the Enterprise data file holds, says when the data file in use lacks them with a tagged link to ask for a copy, and ends with the Find out more section, also when it throws, keeping a failure a failure. The Device Detection readme now gives a command that runs it in a clone, npx ts-node examples/onpremise/gettingstarted-console/gettingStarted.ts from device-detection/fiftyone.devicedetection.onpremise, where its old command could not find the file (#198).

What this release does not yet meet

These MUSTs of the specifications at dbaf099c are still not met in Node.js at this head. Each is left for a reason given here, and none is made worse by this release.

  • A 4xx answer to the Cloud Request Engine's discovery requests is logged and does not fail the build, where cloud-request-engine.md lines 254 to 256 asks for a critical error (#186). Failing the build would stop an application that builds today, which a patch release does not do.
  • Processing a Flow Data a second time runs every element again rather than failing, where automated-testing.md lines 70 and 71 says a Flow Data can only be processed once. Refusing it would stop an application that does this today, which a patch release does not do.
  • The Pipeline builder does not refuse an element that adds Evidence placed in a parallel group, where evidence.md lines 59 and 60 says it MUST NOT add values. The one such element in these packages, the Sequence Element, is always placed first and outside any group, and refusing a layout an application builds today is not something a patch release does.
  • An on-premise engine built from a configuration file or with new and no licenceKeys option refuses to build, even when it would download nothing (data-updates.md lines 33 to 35). The pipeline builders pass the option, so the documented route works. #114 changes the rest and is not in this release.
  • While an on-premise engine reloads its data file, processing waits, where data-updates.md lines 43 and 44 asks for processing to go on with minimal impact. The native reload runs on the one JavaScript thread, and moving it to another thread is a change to the native modules for its own release.
  • Of the data update errors, a failed programmatic check and a failed update from memory are logged and reported rather than passed to the caller, there is no "No response from data update service" message because the update request has no time limit, and an engine that has its data file carries on with it when the update on start-up fails rather than reporting "Update on startup failed for". Each would change what a caller is given back, which a patch release does not do, and they are #210.
  • The JSON Builder writes a WKT value as its text, where value-types.md lines 124 to 126 asks for an object with one member, value. Node.js has no WKT type to tell the writer which text is WKT, and how a weighted WKT value is written is an open question about the specification.
  • The JSON Builder writes each weighted value with rawWeight beside the lower-case keys, where weighted-values.md line 56 asks for lower case. rawWeight is kept for this release because an application may read it.
  • FodId still has the raw flags byte as _flags, marked deprecated, where package-surface.md lines 131 to 135 says it MUST NOT be public. A patch release removes no public name.
  • A configuration file names each element by its package, with elementName and elementParameters, where the other languages use BuilderName and BuildParameters and pipeline-configuration.md line 40 asks for files as similar as possible. Whether the names, the values or neither should match across languages is an open question.
  • Nothing closes a Pipeline, an element or its data, and the native results are freed by the garbage collector, where resource-cleanup.md lines 5, 51 and 60 ask for a way to clean up. What that means in a language with a garbage collector is an open question.
  • The usage sharing record names no 51Degrees package, where usage-sharing-element.md lines 208 to 210 asks for one element for each. How often the record carries them is being changed in the specification, and #136 waits for that.

Before you upgrade

These are the changes a caller may need to act on, most of them following from the conformance above.

  1. On-premise Device Detection defaults change. performanceProfile now defaults to 'Balanced', where it was 'LowMemory', and updateMatchedUserAgent is now on. Set performanceProfile: 'LowMemory' and updateMatchedUserAgent: false to keep the old behavior (#42).

  2. The JSON Builder needs a Sequence Element. It fails with "The JSON Builder needs a sequence number. Place a Sequence Element before the JSON Builder in the Pipeline." when there is none. PipelineBuilder now adds the Sequence Element first, so only a Pipeline built with addSequenceElement: false, or one that adds JSONBundlerElement by hand, is affected (#53, #74).

  3. The JSON sent to the browser gains a sequence entry, which holds session-id and sequence. A request with a session id and no sequence now gets sequence 1, where it got 2 (#59).

  4. The Cloud Request Engine sends more Evidence. It sends every Evidence value apart from the reserved names resource, license, licence and values, which it writes itself. Before, it sent only the Evidence that a Flow Element in the Pipeline accepted, so do not add Evidence that must not reach the cloud service (#72).

  5. Usage Sharing shares no cookie but 51D_ ones. It shared every cookie. It now shares a cookie only when its name starts 51D_, whatever the settings, shareAll: true included, and never the Cookie header. With trackSession on, the session cookie named by cookie is read for the repeat check and is not sent unless its name starts 51D_. Every query string and form value is still shared by default. To share only those whose names start 51D_ or are in queryWhitelist, set the new option shareAllQueryStringParameters: false (#35, #45, #135).

  6. A Flow Element of your own must name its Evidence. An element with no evidenceKeyFilter accepted every key and now accepts none, and flowData.evidence.get gives undefined for a key that no element accepts. Give the element evidenceKeyFilter: new BasicListEvidenceKeyFilter([...]) with the keys it reads (#45).

  7. Data update log entries change. Each error entry starts Error during check for data update, so a log search or alert on the old text needs to change (#69).

  8. Latitude and Longitude have 3 decimal places, as do the points of WKT values such as Areas, so 18.542741172521133 is now 18.543 (#62).

  9. A client IP value is read only when all of it is one address. Values such as 1.2.3.4/24, 1.2.3.4, 5.6.7.8 and 01.2.3.4 were read and are now skipped. Pass one address in each value (#67).

  10. An Engine built with no properties keeps the first Data File's Properties across every data update. A Property that only a later Data File holds has no value until the Engine is created again (#55).

  11. Settings that were accepted now throw when the element is built. These are a JavaScript Builder protocol other than http or https, an objName such as window or document (#31, #60), and a Translation Engine behavior that names none of Original, EmptyString or FlowError (#49).

  12. A 51Did with white space inside it, or with characters from both base64 alphabets, is refused with the status INVALID_BASE64. FodId.toStandardBase64 returns null for such a text, so a TypeScript caller needs to handle null (#40, #56).

  13. Where one name is under more than one Evidence prefix, the Cloud Request Engine sends it once. It sends the value of the highest prefix in the order query, header, cookie, server, fiftyone, location, and compares prefixes and names ignoring letter case, so header.User-Agent and query.user-agent are sent as one field (#75).

  14. _51DEGREES_CLOUD_ENDPOINT replaces FOD_CLOUD_API_URL, which still works in this release. The Cloud Request Engine and the 51Did client read _51DEGREES_CLOUD_ENDPOINT, and read FOD_CLOUD_API_URL only when the new name is unset or empty. An engine whose endpoint came from the old name logs one warning to its Pipeline, and the first 51Did client that uses it raises one Node DeprecationWarning for the process. The baseURL and endpoint options still come first. Set _51DEGREES_CLOUD_ENDPOINT to the same value as the old name (#77, #88).

  15. Each package now has its own version. All 16 are 4.5.103 in this release, and from the next release on only the packages that changed, and those that depend on them, get a new number, so the numbers will differ between packages. Each package requires the others here with ^, so a caller that pins every fiftyone.* package to one number should pin each to its own (#37).

  16. The shared packages keep their code in lib. fiftyone.devicedetection.shared and fiftyone.ipintelligence.shared moved the modules their index.js exports out of examples and tests. A require of an old path, such as fiftyone.devicedetection.shared/tests/keyUtils, still works, as each old file now re-exports the module in lib. The generated types are only under types/lib (#37).

  17. The JSON sent to the browser marks delayexecution only on a Property with a value, as the JSON Builder specification asks, so a Property with no value no longer carries it (#66).

  18. The client-side endpoints should answer a matching If-None-Match before processing. Call Helpers.writeClientSideNotModified(request, response, flowData, options) once the evidence is added and stop when it returns true, as every web example now does. Without it a revalidation is still answered correctly but processed first. A 304 sent while processing records errors now carries Cache-Control: private, max-age=<max-age>, the Vary header and the ETag of a good response, where it carried private, no-cache and no ETag (#120).

  19. A Resource Key that includes no product a cloud engine reads is logged at warning level, once for each engine, where it was an error, and warnings the cloud service sends are now logged (#123).

  20. FodId's raw flags byte _flags is deprecated. Read type, usage and usageIsIndirect, which name its bits. A later release removes it (#115, #128).

  21. The TypeScript types are written again from the code. HardwareProfileCloudEngine from the Device Detection cloud package now has its real type, where a caller that skips checking type files got any, so that caller's use of it is now checked (#113).

  22. The JSON sent to the browser writes each item of a list of Element Data, such as the cloud hardware lookup's profiles, as plain lower-case values, with null and a nullreason for a Property with no value, where it wrote each Property's _value, hasValue and noValueMessage. With no Sequence Element in the Pipeline, the script carries an empty session id and a sequence of 1, where it carried the evidence's (#118).

  23. A weighted IP Intelligence value from the cloud engine also carries weight and rawWeight, as the on-premise engine's values do. rawweighting is still there and deprecated, so read rawWeight in new code (#124, #138).

  24. A query.sequence added as a JavaScript number now counts as absent. The Sequence Element counts a given sequence only when it is a BigInt or text made only of digits, from 0 up to 2,147,483,646, so an application that adds query.sequence itself as a number now gets a sequence of 1. Add the text of the number, as a query string gives it, or a BigInt. A web integration that hands over the request sees no change (#134).

  25. The match metric Properties' category is now "Device Metrics", where it was "Device metrics", and they are listed in the order of the specification's table. Code that compares the category or relies on their order needs to change (#137).

  26. A built Pipeline cannot be changed. It keeps a frozen copy of the list of elements it was built from, and of each list of parallel elements in it, and freezes pipeline.flowElements. Code that changed a list given to addParallel or new Pipeline, flowElementsChain or pipeline.flowElements after the build needs to build a new Pipeline instead. A method that changes flowElementsChain now throws a TypeError, and so does a change to pipeline.flowElements in strict mode code (#142).

  27. weight and rawWeight on each weighted item in the JSON sent to the browser are deprecated. Each item now also carries rawweighting and value, as the specification writes them, so read those. A later release removes weight and rawWeight (#139, #147).

  28. A cloud answer whose errors array has entries now fails the request, whatever its HTTP status, with an error for each entry, and the error messages name the URL the request went to. A log search or alert on "No data in response from cloud service" or "Cloud service returned status code" needs to change, and a failed answer that is not JSON gives the status code message, not "Error parsing response" (#140).

  29. Usage Sharing's warnings are emitted at 'warn', as every other element's are, and at 'warning' as well, which is deprecated. Listen with pipeline.on('warn', ...). A later release stops emitting 'warning' (#141, #149).

  30. minify: true now minifies the client script. It never did, as the JavaScript Builder threw the minified script away. Minifying takes time on every request for the script, so measure it before keeping the setting (#150, #132).

  31. The message for a missing Property has new words. Reading a Property an Engine has but that its results lack throws the message the specification's table gives for the reason, such as "Property 'hardwarevendor' not found in data for element 'device'. This is because your license or data file does not include this property. It is available with the Premium,Enterprise,TAC license/data for the DeviceDetectionOnPremise." A program or a log search that looks for the old words, such as "can be found in the following datafiles", needs the new ones (#163).

  32. A failed refresh from a data file is reported about 2 seconds later, a watched file that is deleted is waited for, and there is a new status. A refresh from a data file is tried 10 times, 200 ms apart, before it is reported as AUTO_UPDATE_REFRESH_FAILED. The file system watcher waits up to 10 minutes for a changed file to open before it loads it, and an update started meanwhile is reported as AUTO_UPDATE_IN_PROGRESS. A watched file that is deleted is loaded if a file appears at its path within that time, and is otherwise reported as the new AUTO_UPDATE_TIMEOUT, value Timeout, with nothing logged, where it was reported at once as a failed refresh and logged as an error. Code that handles every status needs to handle Timeout (#157). With the file system watcher on, a download writes the file the watcher watches, so the watcher's event for that write is reported as AUTO_UPDATE_IN_PROGRESS before the download's own result. Code that listens with once('updateComplete') and takes the first report as the result of its check needs to wait for the first status that is not AUTO_UPDATE_IN_PROGRESS, as both update data file examples now do (#119, #197).

  33. A Cloud Request Engine sends its discovery requests as soon as it is built. It asked the cloud service for its accessible Properties and accepted Evidence keys when the Pipeline first processed, and now asks when it is constructed, so building an engine with the real request client and a made-up endpoint looks that endpoint up. Give such an engine a stand-in requestClient. A transient failure at build logs one warning and the build still completes (#180).

  34. An engine accepts no Evidence until it can say what changes its results. Until discovery answers, or a data file downloaded at start-up arrives, engine.evidenceKeyFilter.filterEvidenceKey(key) gives false where it gave true, and the Pipeline's accepted Evidence, and so the Vary header, holds none of that engine's keys. Evidence added meanwhile is kept, so it still reaches the cloud service and the engine's cache key (#182).

  35. The IP Intelligence cloud engine gives a WKT value as its text. An application reading Areas or Location from the cloud engine now gets the WKT text where it got { value: '...' }, so anything reading .value of that object reads the text instead, which is what the on-premise engine has always given. Nothing changes for the on-premise engine (#199).

  36. A null list of query parameters to share now shares none. An application that sets shareAllQueryStringParameters: false and queryWhitelist: null now shares only the query string and form values whose names start 51D_, where it shared every one. To share them all, leave shareAllQueryStringParameters at true, its default (#204).

  37. A download that cannot be read reports AUTO_UPDATE_UNKNOWN_ERROR. Code that acts on AUTO_UPDATE_MASTER_FILE_CANT_RENAME or AUTO_UPDATE_ERR_READING_STREAM in an updateComplete report now receives AUTO_UPDATE_UNKNOWN_ERROR when a downloaded data file cannot be read. Nothing else reported either status (#205).

New features

  • DerivedPropertyElement, in the new package fiftyone.pipeline.derivedproperty, runs the derived property scripts, with an example (#98, #155).
  • The Cloud Request Engine authenticates with a License Key, alone or with a Resource Key (#64).
  • DataFileUpdateService.updateFromMemory refreshes a data file held in memory (#52).
  • Pipeline.evidenceKeyFilter describes the Evidence a Pipeline accepts (#117).
  • Helpers.writeClientSideResponse and Helpers.writeClientSideNotModified write a client-side endpoint's response with its cache headers (#44, #120).
  • Usage Sharing's trackSession and shareAllQueryStringParameters options (#36, #135).
  • listSetting in fiftyone.pipeline.core reads a list setting given as an array or as one comma separated string (#129).
  • The data update status AUTO_UPDATE_TIMEOUT (#157).
  • profileIdsForPropertyValue(propertyName, value) on the on-premise Device Detection engine, and its indexedProperties option, a list or one comma separated string, which the pipeline builder passes on (#174).

Performance and test coverage

  • A Flow Data answers then without looking for Element Data. A promise reads then from the Flow Data at each element, and each read ran the lookup for a missing key, which throws, and which the case-ignoring lookup made dearer. Timed against main on Ubuntu ARM, Device Detection gives 28 per cent and IP Intelligence 26 per cent more detections a second (run 43751311143, #175).
  • A test measures what minifying the JavaScript Builder's script adds to each request, with a limit of 160 ms set from measurements on every runner, where minifying added 28 to 53 ms against under 0.15 ms for the rest of a request (run 43765200841, #179).
  • A test holds the three request line values to the bytes the request line carried, over twelve targets covering percent-encoded slashes, a plus, encoded spaces, escapes in the case they were written, doubled slashes with dot segments, a question mark with nothing after it and a full URL, so a later change that decoded or tidied a path or a query would fail the build (#188).
  • A test reads the Property metadata of every element from the built Pipeline alone, finds each element by its Data Key and finds a Property by one of its metadata values, which is what the specification asks to be possible at the Pipeline level (#189).
  • A test holds the rule that only sending pauses when usage sharing backs off, processing 25 requests with the pause held open and the receiver silent, so neither the pause nor a send can be what processing waits for, and with the queue full so that processing is shown to go on once sharing can keep nothing (#190).
  • Node.js reads each on-premise Property by name, because a read by name measured faster than a read by position, at 711 ns against 2,797 ns for one read of IsMobile on Ubuntu ARM with Node.js 22 (run 42995016804, #68). The IP Intelligence engine was timed on its own as well, and a read by name took about a third of the time of a read by position on Ubuntu x64 and ARM (run 43487633019, #81).
  • The pipeline examples' tests run each example in its own process and check what it prints or serves, where they built a function and never called it, and the 51Did offline reader and usage sharing examples have tests of their own (#104).
  • The umbrella getting started tests run each example in its own process and check its answers, and a new test reloads a Device Detection engine from the Lite data file onto the Enterprise data file and checks that it keeps its first Properties, as the C library does (#108).
  • The web integration elements test runs as a unit test, so the Publish tests, which run the integration tests against the packed packages, no longer look for a test file the core package does not pack (#112).
  • The native read timing test compares paired rounds and counts the case work, so it no longer fails at random on macOS (#96).
  • Tests of the cloud Property metadata before and after discovery and of the Evidence a Cloud Aspect Engine accepts, and of a session cookie found only in the web request (#144, #145).
  • The Device Detection no-settings test runs offline processing with no arguments and builds the getting started web example from its own 51d.json in its own folder, so neither run is handed a data file and both show the example finding the free file itself (#156).
  • A test refreshes a real Device Detection and IP Intelligence on-premise engine from a file it cannot use, which throws and leaves the engine answering as before, and part way through a batch of requests, each of which is answered as before (#165).
  • Tests build both on-premise engines through their pipeline builders with the data file path, and the folder for a temporary copy of the data file, given relative to the working folder, and check that each answers a request (#168).
  • A test of each of the four Device Detection and IP Intelligence engines, on-premise and cloud, processes a request and checks that the Element Data is Aspect Data whose Properties are read by name (#169).
  • A test of both on-premise engines checks that processing a request reads no Property value, and that reading one Property reads that Property alone (#170).
  • Tests check that both on-premise engines load the native module built for the system they run on, named for its platform, processor and Node.js major version, and that an engine given no Property names creates the native engine with none (#172).
  • The on-premise Device Detection tests run on the free data file. Two that read files only a licensed CI run has now use the free Lite file and a port the system gives, and the reload test that needs the Enterprise file is skipped, with its reason printed, when that file is not there. A test no longer copies the Lite file under the Enterprise file's name, which had made later tests take one for the other (#206).

Environment variables

Each name below has a new name that starts _51DEGREES_, which every shell can set. The Cloud Request Engine and the 51Did client read the cloud endpoint, and the examples and tests read the rest. Every old name is still read in this release, but only when its new name is unset or empty, and its first use is reported as deprecated, through the Pipeline's logger or a DeprecationWarning in the packages, and in a printed message in the examples and tests. Each constant that held an old name holds it again and is marked @deprecated (#70, #77, #88).

Old name, deprecated New name
FOD_CLOUD_API_URL _51DEGREES_CLOUD_ENDPOINT
51DEGREES_DD_PATH _51DEGREES_DD_PATH
51DEGREES_IPI_PATH _51DEGREES_IPI_PATH
51DEGREES_RESOURCE_KEY, RESOURCE_KEY _51DEGREES_RESOURCE_KEY
LICENSE_KEY _51DEGREES_LICENSE_KEY
LicenseKey, LICENSEKEY (the update data file examples) _51DEGREES_LICENSE_KEY
TEST_SUPER_RESOURCE_KEY, TEST_PLATFORM_RESOURCE_KEY, TEST_HARDWARE_RESOURCE_KEY, TEST_BROWSER_RESOURCE_KEY, TEST_NO_SETHEADER_RESOURCE_KEY _51DEGREES_RESOURCE_KEY_SUPER, _51DEGREES_RESOURCE_KEY_PLATFORM, _51DEGREES_RESOURCE_KEY_HARDWARE, _51DEGREES_RESOURCE_KEY_BROWSER, _51DEGREES_RESOURCE_KEY_NO_SETHEADER

Fixed

  • From the second request on, the JSON sent to the browser lost delayexecution and evidenceproperties for every element but one, because the JSON Builder kept each element's delayed Properties under one shared key (#66).
  • An update started by hand with updateDataFile started another polling timer for the same data file each time, so the file was checked once for each of them in every polling interval (#63).
  • With no callback URL, the client script's renderedParameters returned nothing, and the script then threw a TypeError on a page whose tab held a record of an earlier request and values saved from it (#97).
  • A data update download cut short part way was logged as "No response from data update service", or on Windows as the access error, depending on timing. It is now always the access error, "Error accessing data update service", with its cause, and the part file is removed before the failure is reported (#94).
  • The mixed web example named its IP Intelligence engine by a path into the source folder, so it failed against the published packages, as it did in the 13 November publish. It now names the package. The location examples printed ReferenceError: resourceKey is not defined, and now read the key as the other examples do (#95).
  • The Cloud Request Engine asked the cloud service for its Evidence keys and accessible Properties before every request, so each request made three calls to the cloud service in place of one. Release 4.5.102 and earlier have the fault (#73).
  • In 4.5.102 a Cloud Engine threw TypeError: Cannot convert undefined or null to object when the Resource Key did not include its product. One error is now logged for each Engine and the request carries on (#41).
  • An updateComplete listener that threw ended the Node.js process (#65).
  • The wait before a data update check used only the millisecond part of the Data File's expected update time, so polling did not wait for that time (#52).
  • Usage Sharing put a comma between the <Device> elements of a message, logged a failed send as Usage data sent, and wrote <ClientIP>undefined</ClientIP> for a request with no client IP (#35, #36, #54).
  • The on-premise web examples never showed their warning about an old data file, because each was given the current time as the file's published date and compared days with milliseconds. They now warn from the file's own published date (#102).
  • The client-side evidence and on-premise flow element web examples ended their own server on the first page, because each page read a star sign or hemisphere that the element left out of its data until it was known. Both are now null until known (#104).
  • The IP Intelligence compare, metrics, suspicious and mixed examples stopped with RangeError: Maximum call stack size exceeded while they searched for a Data File (#32).
  • flowData.get('Device') and flowData.Device threw where flowData.device worked (#122).
  • The Device Detection cloud package's types did not compile for a TypeScript caller (TS2307), because its index.js required its own files through .. (#113).
  • A list setting given as one comma separated string stopped the build with a TypeError for restrictedProperties and headerBlacklist, and was split into single characters for queryWhitelist (#129).
  • The simple and client-side star sign examples gave no star sign for 21 June (#121).
  • With no Resource Key set, three cloud web examples stopped with an uncaught error and the 51Did creator context demo exited with status 1 and no link to where a key is created (#133).
  • A list of parallel elements given to addParallel stayed the caller's list, so an element added to it after the build was processed by the built Pipeline, and one taken out was not (#142).
  • The same Evidence added in another order, or with a key in other letters, missed an Engine's cache and was processed again, and the key stored was made after processing (#125).
  • A failed cloud answer with an empty body read as "Error parsing response - [object Object]", and one whose JSON had no errors member threw a TypeError (#140).
  • One JSON Builder added to a second Pipeline failed on any element the first Pipeline lacked, and used the first Pipeline's delayed execution for an element with the same data key. An element whose Properties were replaced after its first request, as a cloud engine's are when discovery succeeds late, kept the delayed execution of its old Properties (#143).
  • Usage Sharing's warnings of a full queue, a pause and each failed send were emitted only as the event warning, which no example or document listens for (#141).
  • Helpers.setResponseHeaders threw when processing stopped before the Set Headers element, which ended the process where it was called in a promise with no catch, as the getting started web examples do. It now sets no headers and logs a warning (#146).
  • With minify: true the JavaScript Builder gave its script unminified, because it used the minified script only when uglify-js reported error as null, where uglify-js leaves it undefined. A failure to minify is now logged as a warning (#150).
  • With usage sharing on and a repeat shared after the interval, the tracker's cache kept two entries for one request, and once it was full the cache threw a TypeError, so a request's process was rejected unless suppressProcessExceptions was set. The cache keeps one entry for each key, and an error in usage sharing is now logged and processing goes on (#116).
  • A query. key spelled in another case, such as Query.Mark, was not passed to the client script as a parameter, and the on-premise engines gave the native engine a key such as Header.User-Agent in the case it was added with, so it was accepted and then not used (#117).
  • A missing Property was said to need another data file whenever the metadata listed it, even when the data file in use has it, and a name in another case found no metadata. The reason is now decided in the specification's order, ignoring case (#163).

Housekeeping and CI

  • The C library moves from 51degrees-cxx 4.5.103 to 178371ac, the head of the C library's release draft 51degrees-cxx#191. That draft says under its own Before you upgrade that looking up a Device Detection profile's value by key can miss some values (51degrees-cxx#142). No JavaScript in this repository makes that lookup, as getValuesForProfile for Device Detection appears only in the generated native wrapper hash_node_wrap.cxx, so the packages' JavaScript API does not reach it. The HighPerformance profile holds in memory only what the named Properties read, for 51D_GetHighEntropyValues and 51D_StructuredUserAgent the query value is used before the cookie value, evidence under a key whose prefix is not in lower case is used, a reload holds the data set in use while it reads it, and the C code builds with GCC 15 (#67, #55, #92, #158).
  • Each package has its own version, tagged <package>/<version> in this repository with one GitHub release for each, and the Publish workflow releases only the packages that changed since their last tag, with every package that depends on one of them. A change only to documentation, tests, examples, or a manifest's description, keywords or links releases nothing (#37).
  • fiftyone.pipeline.core packs only the template its JavaScript Builder reads, JavaScriptResource.mustache, and not the rest of the JavaScript templates repository, its tests, workflows, README and licence (#106).
  • fiftyone.pipeline.did carries the OWID library as a submodule and packs only its v1.js and its licence, where it installed OWID from GitHub as a bundled dependency, so installing the package needs no access to GitHub (#107).
  • The six main Device Detection and IP Intelligence packages describe themselves in the words people search npm for, and name the products they are an alternative to, being UAParser.js, DeviceAtlas and WURFL, and MaxMind and IPinfo (#15).
  • Every package's TypeScript types are written again with npm run tsc, so they match the code, and each package's types compile for a TypeScript caller (#113).
  • Links to the configuration page, in the READMEs, code comments, example output and the messages that say where to create a Resource Key, point at https://51degrees.com/configure, where the page has moved from configure.51degrees.com (#87).
  • The repository has one LICENSE file at its root, the EUPL 1.2 text, and every JavaScript source file carries the standard 51Degrees header. The fiftyone.geolocation package no longer carries a LICENSE file of its own (#79).
  • The Selenium contract tests run for real on Linux. The CI removed .NET to free disk space, so dotnet was missing and the step reported a pass without running a test. The step now fails when dotnet is missing or when no test passed, and a pull request into rc is built on the configurations its changes need, with macOS and the performance tests left to builds of rc and the release (#27).
  • The Publish workflow takes allow-early-release, for a release agreed to go out less than a week after the last, and common-ci-ref, to run common-ci's steps from a branch other than main (#99).
  • The cloud 51Did tests run on pull requests into rc as well as main, so a change to the 51Did package is tested against the live cloud service before it reaches a release (#100).
  • AGENTS.md says that a test failing on timing is fixed rather than run again, and points at common-ci's one list of environment variable names (#93, #89).

Issues this closes

Build

Merges into rc are open until 02:00 UTC on Tuesday 17 November 2026, so this section gives the checks that have run on this release's files and says plainly which one still has to run on the final head.

The checks below ran on earlier commits of this release, and all four run again on the final head once merges into rc close at 02:00 UTC on 17 November. rc is at 3ec5bf34dcfd59112ccf61ff25c3b2ffcd0793ed. The Release Pull Request workflow starts on this pull request each time rc moves and is cancelled when rc moves again before it finishes, so most runs today were cancelled, and the newest one to finish is the one on ae9cedc0 given below. The Publish dry run, the Nightly Pipeline dry run and the timing ran on de78d479e3e892517aefab8a07ed0c4cd73477f4, whose tree 123c8e1ebf4393d9500c742072f271632bd88499 is also the tree of 0be6feeb. Since de78d479, Automation51D's #195 moved the C library pointer back to the commit main has and #196 moved it forward again, #197, #198, #200, #201, #202, #203 and #206 change examples, a readme and tests and no published package's code, #199 changes fiftyone.ipintelligence.cloud, #204 changes fiftyone.pipeline.engines.fiftyone and #205 changes fiftyone.pipeline.engines.

The C library. This release is built on 178371ac, read from the cxx gitlink on rc with git ls-tree origin/rc cxx rather than from any note. main has 7992af70.

  • Release build, with macOS. Run 43946813701, the Release Pull Request workflow on this pull request, is green. It ran from 20:09:30 to 21:01:26 UTC on 16 November 2026 on ae9cedc0, and Configure and all eight Build and Test jobs print Setting 'PullRequestSha' to 'ae9cedc0ae603b8b28442b46be14c9f152c752a9', which is 9 of 9 lines on that commit. All eight Build and Test entries, Ubuntu, Ubuntu ARM, Windows and macOS on Node.js 22 and 24, succeeded, each reporting "2 skipped, 2276 passed, 2278 total" for the unit tests and "4 skipped, 130 passed, 134 total" for the integration tests, and Compare Performance and Complete succeeded. Compare Performance reads DetectionsPerSecond against the average of the last ten results on each runner, not against main.
Detections per second This build Average of the last 10 Standard deviation Acceptable above Change
Ubuntu_Node_24 71,286.41 37,604.79 13,860.65 9,883.49 +89.6 per cent
Ubuntu_ARM_Node_24 39,243.60 34,600.92 1,919.04 30,762.83 +13.4 per cent
Windows_Node_24 35,030.96 28,831.81 6,580.19 15,671.42 +21.5 per cent
macOS_Node_24 56,349.17 40,073.02 10,688.99 18,695.04 +40.6 per cent

ae9cedc0 holds every change to package code in this release but one. After it, #198 and #206 change an example, a readme and tests, and #205 changes fiftyone.pipeline.engines, whose own full build without macOS is green (run 43956276656). An earlier release build on 372c8a4c, run 43876589521, was green too. The release build runs again on the final head after the 02:00 cut-off.

  • Publish releases only what changed. Publish dry run 43807209158 ran on version/owner-node-rel-0457, which is de78d479 with only the two macOS entries taken out of ci/options.json so that the plan prints without waiting for macOS runners. Its Configure plans all 16 packages from 4.5.102 to 4.5.103, fifteen of them as changed since 4.5.102 with the count of changed files for each, and fiftyone.ipintelligence.translation as a package whose dependency is released. Each passes the version check as the patch place, which needs no agreement, against its own <package>/<version> tags and against npm. The run used allow-early-release: false and passed the weekly check.

  • The nightly passes. Nightly Pipeline dry run 43807220965 ran on tmp/node-monday-release-head at de78d479 from 04:58 to 07:11 UTC and is green. Of its 35 jobs 33 succeeded and 2 were skipped, the two being the reports of Compare Performance. Nightly Publish passed Configure, the eight Pre Build jobs, Build, all eight Test jobs on the packed packages with both macOS entries, ComparePerformance and Package. The Submodule Update passed and Nightly Pull Requests built the head on all eight configurations.

    The scheduled nightly on main fails, and this release is what fixes it. Run 43796227383 of 03:07 UTC on 16 November failed in all eight Nightly Publish Test jobs, each reporting "1 failed, 4 skipped, 131 passed, 136 total" with the one failure being "Examples without settings, the mixed web page marks the Enterprise properties". This release changes both the example and that test, and the dry run above passes all eight of those jobs. The dashboard badge turns green once this release is on main and a scheduled nightly has run there.

  • Not slower than main. Run 43807240556 timed main (4d5832f0, C library 7992af70) against this release (de78d479, C library 178371ac) on one ubuntu-22.04-arm runner with Node.js 24, over 21 rounds with the order turned each round. It logs both arms' commits, their C library commits and the four native binaries with their sizes and hashes.

Measure main This release This release better in
Device Detection, detections a second 33,173.13 40,710.23 (+22.7 per cent) 21 of 21 rounds
IP Intelligence, lookups a second 33,445.85 42,198.46 (+26.2 per cent) 20 of 21 rounds
Control, this release's packages run by main's own performance script 41,728.74 (+24.8 per cent) 20 of 21 rounds

This release is faster on both engines, so it is not slower than main under the release rule, which counts a release as slower only when its whole range of rounds sits on the slower side of main's. The control arm, which runs this release's packages through main's own performance script, gains almost as much as the release arm, which says the gain is in the packages and the C library rather than in the script.

The figure in Performance and test coverage above for the Flow Data then change comes from its own run and not from this timing, which measures the release as a whole.

Drafts still open into rc are not in this release. #136 names every 51Degrees package in use in the usage sharing record and is not in this release, as the specification is changing how often the record carries them. #114 would need a licence key only for the updates that download. #10 builds and tests on Node.js 26 and is a colleague's. #194 adds a test that an engine uses any cache meeting the simple interface. #191 is superseded by #190, which is in. The draft from the review of this release that is open is #209 (a comment that named what the cloud request engine accepts before discovery). If it merges before 02:00 UTC it is described in the sections above when this body is written again on the final head.

Only the patch number moves.

AGENTS.md gains a section linking common-ci's list of environment variable
names (environment-variables.md, made from environment-variables.json) at
commit 18f9a570, and saying the list decides. README.md points at the same page
where it names the variables the tests or examples read.

Documentation only.
…file before reporting it (#94)

* Test that a download cut short reports the access error

The data update specification's Errors table gives the messages of the
.NET data update service. .NET reads the whole response before its
request returns, so a download cut short fails that request, and .NET
reports it as "Error accessing data update service at '<url>' for engine
'<name>'" with the cause attached.

The test of a download cut short closed the connections 10 milliseconds
after the check started, whether or not the response had arrived, so it
passed or failed by chance. On a slow runner the connection closed before
the response, and the message was the one for a failed connection. The
test now cuts the connection only once the service's request has its
response, from a server that never sends the whole file, and a
connection closed before any response is tested on its own. The test of
a reset part way through the download expects the same error, which a
reset gives whether it comes before or after the response.

* Report a download cut short as the access error, with its cause

A download that ended before the whole file arrived was logged as "No
response from data update service", a message .NET writes only when its
request returns no response at all. .NET reads the whole response before
its request returns, so it reports a download cut short as "Error
accessing data update service at '<url>' for engine '<name>'" with the
cause attached, which is what Node already wrote for a connection that
failed before any response. The service now writes that message for a
download cut short too, with the first error the request or the response
raised after the response arrived as the cause. The status is unchanged.

* Test that a download cut short leaves no part file, with no timer

The test of a reset part way through the download reset the connection 50
milliseconds after the server wrote, and waited 500 milliseconds before it
looked, so its check that the part file had gone held only because of
those two times. It now resets once the service's request has its
response, waits for the update to be reported, and then checks that there
is one report, one scheduled check and no part file. The test of a
download cut short checks the folder too. In both tests the cut comes
before the service has opened its part file.

* Remove a cut short download's part file before reporting the update

When a download was cut short the service destroyed the write stream for
the part file, asked for the file to be removed straight away, and
reported the update. A stream still opening its file has not created it
yet, so the removal could find nothing and the file the stream then
created would stay in the temporary folder. In every order the file could
still be there when the update was reported. The service now removes the
part file once the stream has closed, which is after the file is open,
and reports the update once the file has gone.

* Test that the part file is removed before the update is reported

The folder checks in the two cut short tests passed or failed by chance on
the service before the fix, because fs.unlink removes the part file on
another thread while the test goes on. A spy on fs.unlink now records each
removal as it finishes, and both tests check that the part file's removal
had finished when the update was reported. Before the fix the service
reports the update first, so the check fails every time.
…or (#15)

* Package text: say what each package does in the words people search for

Developers search npm with words such as user agent parser, device
detection, and IP geolocation. The descriptions of the six Device
Detection and IP Intelligence packages now say what each package does in
those words, and the keywords add the matching terms. Existing keywords
that are still true are kept, and keywords naming other companies'
products are removed.

Both package.json and remote_package.json change, because
ci/build-package.ps1 publishes remote_package.json and the two files
should match. Only the description and keywords fields change, and no
version number moves.

* Package text: name the products these packages are an alternative to

People searching npm for WURFL, DeviceAtlas, MaxMind, or ipinfo should
find these packages too. The keywords naming those products are restored
exactly as they were, and each description now ends with the products
the package is an alternative to. Only the description and keywords
lines change, and no version number moves.

* Package text: name UAParser.js in full

The Device Detection descriptions now say "An alternative to UAParser.js,
DeviceAtlas, and WURFL." The 51Degrees fork is called 51Degrees UAParser,
so the bare name could read as our own package. Only the description line
changes, and no version number moves.
…se work (#96)

Reviewed and approved at head 1ef641c. The test times both arms side by side, takes the middle of 200 rounds and holds them to 1.05, set from 293 measurements on every runner (red 43443859591, green 43473620570, full build 43473656997 with macOS).
…ntries (#27)

On Ubuntu the Selenium contract step reports a pass without running a test, so the browser contract has had no Linux coverage. And a pull request into `rc` runs all eight configurations, two of them on macOS, of which only five run at once across the whole organisation.
…ue (#66)

The JSON Builder wrote `delayexecution` for every property whose metadata sets the delay execution flag, before it read the value ([`jsonbundler.js` line 168 at 24592ac](https://github.com/51Degrees/51degrees-node/blob/24592acbee7fd2881a96ac661296535daed6f157/pipeline/fiftyone.pipeline.core/jsonbundler.js#L168)). So a property with no value, and one that could not be read at all, were marked too. The cross-language audit found this, and PHP ([51degrees-php#61](51Degrees/51degrees-php#61)), Go ([51degrees-go#91](51Degrees/51degrees-go#91)) and Python ([51degrees-python#92](51Degrees/51degrees-python#92)) have already merged the same change.
…heck what the location examples print (#95)

The Publish run for release 4.5.103, [run 43412143722](https://github.com/51Degrees/51degrees-node/actions/runs/43412143722) on `main` at 4d5832f (the merge of #83), failed, so nothing was published and npm still has 4.5.102. In each of the eight "Publish / Test" jobs (Ubuntu, Ubuntu ARM, Windows and macOS, on Node.js 22 and 24) exactly one test failed, "Examples without settings › the mixed web page marks the Enterprise properties" in [`examplesWithoutSettings.test.js`](https://github.com/51Degrees/51degrees-node/blob/24592acbee7fd2881a96ac661296535daed6f157/ip-intelligence/fiftyone.ipintelligence.onpremise/examples/onpremise/examplesWithoutSettings.test.js#L234-L244).
[#8](#8). The JavaScript builder sets the template's `_parameters` only when it has a callback URL ([`javascriptbuilder.js` lines 340 to 365 at 24592ac](https://github.com/51Degrees/51degrees-node/blob/24592acbee7fd2881a96ac661296535daed6f157/pipeline/fiftyone.pipeline.core/javascriptbuilder.js#L340-L365)). With no endpoint, the template's line `var renderedParameters = function() { return {{&_parameters}}; };` renders as `return ;`.
… a named branch (#99)

common-ci refuses a release that comes less than a week after the last one unless `allow-early-release` is set, and the refusal tells the person to "run this repository's Publish workflow with allow-early-release set." ([common-ci `.github/workflows/nightly-publish.yml` lines 28 to 31, 202 and 210 at 1c5c0e0b](https://github.com/51Degrees/common-ci/blob/1c5c0e0b09a1f77722e20ba2d10cbd1031caca9e/.github/workflows/nightly-publish.yml#L28-L31)). This repository's Publish workflow has only the `dryrun` input ([`.github/workflows/publish.yml` lines 3 to 8 at 758fea8](https://github.com/51Degrees/51degrees-node/blob/758fea819c13fe0df4275a0a0478cdd03f57d5ed/.github/workflows/publish.yml#L3-L8)), so a release agreed to go out early cannot be started, whatever the message says.
)

The Cloud 51Did workflow runs the live 51Did tests on a pull request only when it targets `main` ([`.github/workflows/cloud-51did.yml` lines 32 and 33 at 758fea8](https://github.com/51Degrees/51degrees-node/blob/758fea819c13fe0df4275a0a0478cdd03f57d5ed/.github/workflows/cloud-51did.yml#L32-L33)). Changes reach `main` through `rc`, so a change to the 51Did package is first tested against the live cloud service when the release pull request from `rc` opens, after it has been reviewed and merged into `rc`.
…es when a value is not yet known (#104)

Five pipeline examples had tests that never ran them, and two examples had no test at all (gaps DX-node-34 and DX-node-35).
…kage (#106)

`fiftyone.pipeline.core` has no `files` list ([`pipeline/fiftyone.pipeline.core/package.json` at 758fea8](https://github.com/51Degrees/51degrees-node/blob/758fea819c13fe0df4275a0a0478cdd03f57d5ed/pipeline/fiftyone.pipeline.core/package.json)), so npm packs the whole `javascript-templates` submodule, its tests, README, LICENSE and GitHub workflows, where the JavaScript Builder reads only `JavaScriptResource.mustache` ([`javascriptbuilder.js` lines 27 to 30](https://github.com/51Degrees/51degrees-node/blob/758fea819c13fe0df4275a0a0478cdd03f57d5ed/pipeline/fiftyone.pipeline.core/javascriptbuilder.js#L27-L30)). Anything the templates repository adds later would be published in this package too, so the submodule could not move on without changing what the package holds. The Device Detection and IP Intelligence packages already have a `files` list, for example [`device-detection/fiftyone.devicedetection.cloud/package.json`](https://github.com/51Degrees/51degrees-node/blob/758fea819c13fe0df4275a0a0478cdd03f57d5ed/device-detection/fiftyone.devicedetection.cloud/package.json). This is gap DX-node-15, item 6 of run book 9.
Step 3.1 of [common-ci#306](51Degrees/common-ci#306). [`ci/build-package.ps1` line 35 at 21fa119](https://github.com/51Degrees/51degrees-node/blob/21fa11996a1e584e855ef548871cc006d23e5873/ci/build-package.ps1#L35) runs `npm version` with the run's one version for every package, so all 16 packages published 4.5.102 on 4 November 2026 ([Publish run 41918901313](https://github.com/51Degrees/51degrees-node/actions/runs/41918901313)) whatever had changed. That spends every package's allowance under the weekly release rule, and the [release of 13 November](#83) would have done the same.
)

`PipelineBuilder.addRequiredElements` adds the Sequence Element, the JSON Builder, the JavaScript Builder and the Set Headers Element whatever the caller already added ([`pipeline/fiftyone.pipeline.core/pipelineBuilder.js` lines 200 to 205 at 758fea8](https://github.com/51Degrees/51degrees-node/blob/758fea819c13fe0df4275a0a0478cdd03f57d5ed/pipeline/fiftyone.pipeline.core/pipelineBuilder.js#L200-L205)). This has three effects.
…he Device Id row (#102)

Two gaps from the examples audit, both recorded under "Not changed" in [#32](#32).
This draft closes two gaps from the audit of the deferred work and the examples, DX-node-26 and DX-node-27. On `rc` at 758fea8 the examples did not end with a "Find out more" section, which is the part of an example that gives the person running it somewhere to go next.
Eleven of the examples' `51d.json` files name an engine by a path into this repository's source, such as `"../../../deviceDetectionCloud"` in [`device-detection/fiftyone.devicedetection.cloud/examples/cloud/gettingstarted-console/51d.json` at fe81e01](https://github.com/51Degrees/51degrees-node/blob/fe81e01866b5aae0e2c6ec11a25a4f0a5d1ccd31/device-detection/fiftyone.devicedetection.cloud/examples/cloud/gettingstarted-console/51d.json). Such a file only works beside the source. [`ci/install-package.ps1` lines 43 to 55](https://github.com/51Degrees/51degrees-node/blob/fe81e01866b5aae0e2c6ec11a25a4f0a5d1ccd31/ci/install-package.ps1#L43-L55) rewrites two of them so the Publish tests run them against the packed packages, and [`device-detection/run_examples.md` lines 72 to 92](https://github.com/51Degrees/51degrees-node/blob/fe81e01866b5aae0e2c6ec11a25a4f0a5d1ccd31/device-detection/run_examples.md#L72-L92) and [`ip-intelligence/run_examples.md` lines 61 to 81](https://github.com/51Degrees/51degrees-node/blob/fe81e01866b5aae0e2c6ec11a25a4f0a5d1ccd31/ip-intelligence/run_examples.md#L61-L81) tell a person running an example against the published packages to edit the rest by hand. This is gap DX-node-20, left under "What remains" in [#95](#95), which named the mixed web example's engine by package.
This draft was to wait until https://51degrees.com/configure answers. At 17:35 UTC on 13 November it does. `/configure`, `/configure/`, `/configure/gGc3T5pT` and `/configure?utm_source=code` each answer 200 with the page title "Cloud Configuration | 51Degrees", the same title `configure.51degrees.com` serves, and the page at the new address loads the configuration app with `<base href="/configure/">`. On 13 November at 15:01 the same addresses answered 404.
…ata file reload (#108)

Two test gaps, with no package code to change.
…id and sequence in any letters (#109)

Two faults in the usage sharing record, and two cases no test covered, all at [fe81e01](https://github.com/51Degrees/51degrees-node/tree/fe81e01866b5aae0e2c6ec11a25a4f0a5d1ccd31).
… name what needs the Enterprise files (#111)

The IP Intelligence mixed API example, which runs a local copy of the cloud service's API on both on-premise engines, does not run from a fresh clone as the other examples do (gap DX-node-29, left under "Not changed" in [#32](#32) and "What remains" in [#95](#95)). At [7539a0b](https://github.com/51Degrees/51degrees-node/tree/7539a0be284aec3e0aae12cf50cf727ead68fb4c):
…n the packed packages (#112)

The Nightly Pipeline dry run [43578169805](https://github.com/51Degrees/51degrees-node/actions/runs/43578169805), on a throwaway branch holding `rc` at 7539a0b with the drafts then approved merged in, fails one test in all eight Test jobs of its Publish stage, being Ubuntu, Ubuntu ARM, Windows and macOS on Node.js 22 and 24.
fiftyone.pipeline.did took the [OWID library](https://github.com/SWAN-community/owid-js) as the npm dependency `"owid": "github:SWAN-community/owid-js#6c5ff4e7"`, bundled into its tarball, where every other 51Degrees language takes OWID as a git submodule. This takes owid-js as the submodule `pipeline/fiftyone.pipeline.did/owid-js`, pinned at the same commit, and packs only the library's `v1.js` and its `LICENSE` from it.
…le for a TypeScript caller (#113)

Each package publishes the `types/*.d.ts` files committed in this repository, and nothing writes them again before a package is packed. The readmes say that "Calling the generation step is manual for now" ([`device-detection/readme.md` line 236 at a9a3bd2](https://github.com/51Degrees/51degrees-node/blob/a9a3bd2ec192f0006296a39366a710c5ec090f18/device-detection/readme.md?plain=1#L236)). So the published types have fallen behind the code. With TypeScript 5.9.3, `npm run tsc` on `rc` at a9a3bd2 changes 23 of the committed files and writes one that was never committed, `pipeline/fiftyone.pipeline.core/types/callbackUrl.d.ts` ([run 43580575276](https://github.com/51Degrees/51degrees-node/actions/runs/43580575276)).
…and mark the raw flags byte deprecated (#115)

1. **A short payload that names another version is refused as too short, and the version it names is not reported.** `unpack` checks the length against version 0's header before it reads the version ([`pipeline/fiftyone.pipeline.did/internal/payload.js` lines 88 to 113 at 6d5d4fe](https://github.com/51Degrees/51degrees-node/blob/6d5d4fef292afffffae3e7892d3e8c043f1b22c7/pipeline/fiftyone.pipeline.did/internal/payload.js#L88-L113)), so a payload of 1 to 4 bytes whose flags byte names version 1, 2 or 3 gets `PayloadTooShort`, measured against a header that a later version need not have. A payload of normal length is refused naming its version, and tested ([`tests/fodId.test.js` lines 611 to 632](https://github.com/51Degrees/51degrees-node/blob/6d5d4fef292afffffae3e7892d3e8c043f1b22c7/pipeline/fiftyone.pipeline.did/tests/fodId.test.js#L611-L632)).
2. **The raw flags byte is part of the typed surface.** `FodId` declares `_flags: number` ([`types/fodId.d.ts` line 189](https://github.com/51Degrees/51degrees-node/blob/6d5d4fef292afffffae3e7892d3e8c043f1b22c7/pipeline/fiftyone.pipeline.did/types/fodId.d.ts#L189)), set at [`fodId.js` line 132](https://github.com/51Degrees/51degrees-node/blob/6d5d4fef292afffffae3e7892d3e8c043f1b22c7/pipeline/fiftyone.pipeline.did/fodId.js#L132) and [line 480](https://github.com/51Degrees/51degrees-node/blob/6d5d4fef292afffffae3e7892d3e8c043f1b22c7/pipeline/fiftyone.pipeline.did/fodId.js#L480). The published fiftyone.pipeline.did 4.5.102 declares it too ([its `types/fodId.d.ts` line 177](https://unpkg.com/fiftyone.pipeline.did@4.5.102/types/fodId.d.ts)), where the offsets and the Terms index are kept in modules that are not exported.
…e's headers, and read Set Header identifiers that hold digits (#120)

1. **The client-side endpoints process a request before they compare its If-None-Match with the ETag.** `Helpers.writeClientSideResponse` takes "the processed flow data for the request" ([`pipeline/fiftyone.pipeline.core/helpers.js` lines 148 to 168 at 6d5d4fe](https://github.com/51Degrees/51degrees-node/blob/6d5d4fef292afffffae3e7892d3e8c043f1b22c7/pipeline/fiftyone.pipeline.core/helpers.js#L148-L168)), so every example that serves the JSON or JavaScript endpoint runs the Pipeline first, for example the Device Detection on-premise web example ([`gettingStarted.js` lines 233 to 250](https://github.com/51Degrees/51degrees-node/blob/6d5d4fef292afffffae3e7892d3e8c043f1b22c7/device-detection/fiftyone.devicedetection.onpremise/examples/onpremise/gettingstarted-web/gettingStarted.js#L233-L250)). A browser that revalidates a script it already holds still costs a full detection.
2. **A 304 sent while processing records errors loses the ETag and max-age.** The response then carries `Cache-Control: private, no-cache` and no ETag, 304 included ([`helpers.js` lines 185 to 202](https://github.com/51Degrees/51degrees-node/blob/6d5d4fef292afffffae3e7892d3e8c043f1b22c7/pipeline/fiftyone.pipeline.core/helpers.js#L185-L202)), and a test asserts it ([`tests/clientSideCaching.test.js` lines 226 to 230](https://github.com/51Degrees/51degrees-node/blob/6d5d4fef292afffffae3e7892d3e8c043f1b22c7/pipeline/fiftyone.pipeline.core/tests/clientSideCaching.test.js#L226-L230)). So a browser that holds a good response keeps it through an outage but its cache drops the ETag and max-age.
3. **A Set Header identifier that holds a digit ends early.** `getHeaderName` skips only `a` to `z` after the identifier's first character ([`setHeadersElement.js` lines 61 to 72](https://github.com/51Degrees/51degrees-node/blob/6d5d4fef292afffffae3e7892d3e8c043f1b22c7/pipeline/fiftyone.pipeline.core/setHeadersElement.js#L61-L72)), so `SetHeaderBrowser2Accept-CH` gives the header `2Accept-CH`. No property in use today has a digit in its identifier, so no response is wrong today.
4. **No test passes the max-age option**, which `writeClientSideResponse` takes ([line 167](https://github.com/51Degrees/51degrees-node/blob/6d5d4fef292afffffae3e7892d3e8c043f1b22c7/pipeline/fiftyone.pipeline.core/helpers.js#L167)).
Element Data is found only by its key as written. `FlowData.get` looks the key up as given, `const elementData = this.data[flowElementDataKey];` ([`pipeline/fiftyone.pipeline.core/flowData.js` line 191 at 6d5d4fe](https://github.com/51Degrees/51degrees-node/blob/6d5d4fef292afffffae3e7892d3e8c043f1b22c7/pipeline/fiftyone.pipeline.core/flowData.js#L191)), so `flowData.get('Device')` and `flowData.Device` throw where `flowData.device` works. Property names already ignore letter case ([`elementData.js` lines 100 to 123](https://github.com/51Degrees/51degrees-node/blob/6d5d4fef292afffffae3e7892d3e8c043f1b22c7/pipeline/fiftyone.pipeline.core/elementData.js#L100-L123)).
…warning (#123)

1. **The warnings a cloud response carries are never logged.** The cloud service can answer with a `warnings` array beside its results, and no code in `fiftyone.pipeline.cloudrequestengine` reads it. The request engine parses the discovery response for its products ([`cloudRequestEngine.js` lines 404 to 415 at 6d5d4fe](https://github.com/51Degrees/51degrees-node/blob/6d5d4fef292afffffae3e7892d3e8c043f1b22c7/pipeline/fiftyone.pipeline.cloudrequestengine/cloudRequestEngine.js#L404-L415)) and keeps the processing response as it is ([lines 504 to 515](https://github.com/51Degrees/51degrees-node/blob/6d5d4fef292afffffae3e7892d3e8c043f1b22c7/pipeline/fiftyone.pipeline.cloudrequestengine/cloudRequestEngine.js#L504-L515)), so a warning from the service reaches nobody.
2. **A product the key does not include is logged as an error.** The cloud engine logs it once, at error level ([`cloudEngine.js` lines 131 to 141](https://github.com/51Degrees/51degrees-node/blob/6d5d4fef292afffffae3e7892d3e8c043f1b22c7/pipeline/fiftyone.pipeline.cloudrequestengine/cloudEngine.js#L131-L141)), though nothing is wrong with the request and the other products in the answer are used.
jwrosewell and others added 30 commits November 15, 2026 22:34
The empty list filter set in the constructor made the generated types
declare evidenceKeyFilter as the list filter on these three engines, which
would stop TypeScript code assigning any other filter. It is now cast to the
base filter type, so the declarations stay as they were.
The cast through the local name of the filter class gave the constructor's
type, so the declarations still changed. The type is now named through the
core package, as the base Flow Element declares it.
The cloud hardware lookup gave each profile as a plain object holding an `AspectPropertyValue` for each Property ([`device-detection/fiftyone.devicedetection.cloud/hardwareProfileCloudEngine.js` lines 104 to 121 at c6ba938](https://github.com/51Degrees/51degrees-node/blob/c6ba938e6030b344126c3d7500eba254796d050c/device-detection/fiftyone.devicedetection.cloud/hardwareProfileCloudEngine.js#L104-L121)). A profile had no `get` method and no letter case handling, where the device data of the cloud and on-premise engines is Aspect Data read through `get` ([`device-detection/fiftyone.devicedetection.cloud/tests/aspectData.test.js` lines 93 to 101](https://github.com/51Degrees/51degrees-node/blob/c6ba938e6030b344126c3d7500eba254796d050c/device-detection/fiftyone.devicedetection.cloud/tests/aspectData.test.js#L93-L101)). So code that reads `device.get('HardwareVendor')` from the device data could not read a profile the same way. The profile tests read through the examples' `DataExtension.getValueHelper`, which reads a member, so they did not show the difference.
The Set Headers element built its list of headers, and the SetHeader Properties each is made from, the first time it processed, from that Flow Data's Pipeline, and kept it ([`pipeline/fiftyone.pipeline.core/setHeadersElement.js` lines 117 to 119 at c6ba938](https://github.com/51Degrees/51degrees-node/blob/c6ba938e6030b344126c3d7500eba254796d050c/pipeline/fiftyone.pipeline.core/setHeadersElement.js#L117-L119)). One element added to a second Pipeline then read the first Pipeline's SetHeader Properties for every request, so it never gave the headers the second Pipeline's elements set.
…180)

The Cloud Request Engine asked the cloud service for its accessible Properties and accepted Evidence keys only when the Pipeline first processed ([`pipeline/fiftyone.pipeline.cloudrequestengine/cloudRequestEngine.js` line 187 and lines 282 to 294 at c6ba938](https://github.com/51Degrees/51degrees-node/blob/c6ba938e6030b344126c3d7500eba254796d050c/pipeline/fiftyone.pipeline.cloudrequestengine/cloudRequestEngine.js#L282-L294)). Its constructor made no request, so a built engine was never ready to process, and a cloud service that could not be reached at start-up said nothing until the first request.
…#182)

This branch holds the commits of #180, which makes the Cloud Request Engine's discovery requests when it is built, because the change here depends on it. Until #180 merges, the diff here shows both. The current rc (c1a703c) is merged in with a merge commit.
A promise reads `then` from every value it resolves with, to see whether that value is itself a promise. The Pipeline resolves a promise with the Flow Data before its first element and after each one ([`pipeline/fiftyone.pipeline.core/pipeline.js` lines 164, 176, 191, 210 and 221 at c6ba938](https://github.com/51Degrees/51degrees-node/blob/c6ba938e6030b344126c3d7500eba254796d050c/pipeline/fiftyone.pipeline.core/pipeline.js#L164-L221)), and the caller's `await flowData.process()` does it once more.
…ates (#187)

[`pipeline/fiftyone.pipeline.core/jsonbundler.js` line 358 at d8a76c7](https://github.com/51Degrees/51degrees-node/blob/d8a76c7438990c4221fb6c2b70d4ea4c9abb33fb/pipeline/fiftyone.pipeline.core/jsonbundler.js#L358) writes the JSON Builder's Element Data as `contents: { json: output }`, and [`javascriptbuilder.js` line 435](https://github.com/51Degrees/51degrees-node/blob/d8a76c7438990c4221fb6c2b70d4ea4c9abb33fb/pipeline/fiftyone.pipeline.core/javascriptbuilder.js#L435) writes the JavaScript Builder's as `contents: { javascript: output }`. Neither element publicizes a Property, so `FlowElement`'s default of an empty object stands ([`flowElement.js` line 68](https://github.com/51Degrees/51degrees-node/blob/d8a76c7438990c4221fb6c2b70d4ea4c9abb33fb/pipeline/fiftyone.pipeline.core/flowElement.js#L68)), and an application that reads the Pipeline's Property metadata is not told that either element produces anything. The Sequence Element and the Set Headers Element already publicize theirs.
…est (#179)

The JavaScript Builder minifies its script with uglify-js when minify is on ([`pipeline/fiftyone.pipeline.core/javascriptbuilder.js` lines 421 to 430 at c6ba938](https://github.com/51Degrees/51degrees-node/blob/c6ba938e6030b344126c3d7500eba254796d050c/pipeline/fiftyone.pipeline.core/javascriptbuilder.js#L421-L430)). The test of that feature checks that the script is shorter and still parses ([`pipeline/fiftyone.pipeline.core/tests/javascriptMinify.test.js` lines 45 to 52](https://github.com/51Degrees/51degrees-node/blob/c6ba938e6030b344126c3d7500eba254796d050c/pipeline/fiftyone.pipeline.core/tests/javascriptMinify.test.js#L45-L52)), and both performance examples turn the JavaScript Builder off, at [the Device Detection performance console example line 195](https://github.com/51Degrees/51degrees-node/blob/c612563ae6c3bfa57ae0a52b6995b89a921a0a8d/device-detection/fiftyone.devicedetection.onpremise/examples/onpremise/performance-console/performance.js#L195) and [the IP Intelligence one line 192](https://github.com/51Degrees/51degrees-node/blob/c612563ae6c3bfa57ae0a52b6995b89a921a0a8d/ip-intelligence/fiftyone.ipintelligence.onpremise/examples/onpremise/performance-console/performance.js#L192), so nothing measured what minifying costs.
The current `rc`, which holds #180 (the discovery requests made when the engine is built) and #182 (no Evidence accepted before discovery answers), is merged into this branch, so the diff shows only the change described here. Those two drafts are merged, so nothing waits on them.
…yte (#188)

Nothing checked that the three request line values the web integration supplies are the bytes the request line carried. [`pipeline/fiftyone.pipeline.core/evidence.js` lines 37 to 48 at d8a76c7](https://github.com/51Degrees/51degrees-node/blob/d8a76c7438990c4221fb6c2b70d4ea4c9abb33fb/pipeline/fiftyone.pipeline.core/evidence.js#L37-L48) takes the path and the query as substrings of the request target and decodes nothing, and [lines 367 to 370](https://github.com/51Degrees/51degrees-node/blob/d8a76c7438990c4221fb6c2b70d4ea4c9abb33fb/pipeline/fiftyone.pipeline.core/evidence.js#L367-L370) always put all three on the `server.http-request` view. That is right, and no test held it in place, so a later change that tidied a path or decoded a query would have passed every build.
…eline (#189)

Nothing read Property metadata at the Pipeline level. The elements each publish their own through `getProperties()`, and the Pipeline holds them in [`flowElements`](https://github.com/51Degrees/51degrees-node/blob/d8a76c7438990c4221fb6c2b70d4ea4c9abb33fb/pipeline/fiftyone.pipeline.core/pipeline.js#L243) and indexes them by each metadata key and value in [`propertyDatabase`](https://github.com/51Degrees/51degrees-node/blob/d8a76c7438990c4221fb6c2b70d4ea4c9abb33fb/pipeline/fiftyone.pipeline.core/pipeline.js#L299-L350), which is what `FlowData.getWhere` reads. Both routes work and no test used either of them from the Pipeline, so nothing held the capability in place.
Nothing checked that processing goes on while usage sharing is paused. [`pipeline/fiftyone.pipeline.engines.fiftyone/tests/shareUsageBackOff.test.js`](https://github.com/51Degrees/51degrees-node/blob/d8a76c7438990c4221fb6c2b70d4ea4c9abb33fb/pipeline/fiftyone.pipeline.engines.fiftyone/tests/shareUsageBackOff.test.js) covers the pause itself in nineteen cases, including a full queue dropping new evidence. Five of those nineteen process requests during a pause as a means to their own end rather than to check what processing does, and the rest are cases of the back-off object itself or process nothing at all. So the requirement that only sending pauses had no case of its own.
Co-authored-by: Automation51D <51DCI@51Degrees.com>
The nightly submodule update moved the C library to its own main, which is behind the commit this release is built from. Every language releases on the same native code, and PHP needs that commit to build on Ubuntu 26.04.
… examples (#197)

Both update data file examples abandon themselves when the file system
watcher is on, which is how each of them is configured.
The IP Intelligence cloud engine left a WKT value as the object the service
sends, so a caller got a different kind of value from the two engines for
the same Property.
…cyRadiusMin in full (#200)

The IP Intelligence metrics example wrote a figure the data does not give.
Five examples started a web server that listened on every network interface
of the machine, and every one of them tells the person running it to browse
to localhost.
Two examples looked for a free data file with `findFile`, which looks only in the repository's `cxx/data` folder.
…hem all is off (#204)

With `shareAllQueryStringParameters: false`, the usage sharing element is meant to share only the query string and form parameters whose names start with `51D_` and those on its `queryWhitelist`. Given `queryWhitelist: null`, it shared every query parameter instead.
Three examples read `51d.json` by a bare name, so Node.js looked for it in the folder the example was run from rather than the folder the example is in.
…the Enterprise properties (#198)

`device-detection/readme.md` lines 238 to 243 tell people to run
`device-detection/fiftyone.devicedetection.onpremise/examples/onpremise/gettingstarted-console/gettingStarted.ts`
with `npx ts-node gettingStarted.ts`. It printed five Properties the free
Lite data file holds and stopped. It asked for no Property that only the
Enterprise data file holds, gave no link to ask for a copy, and printed no
Find out more section, so someone following the readme saw none of what the
paid data file adds and was given nowhere to go next.
… needs the Enterprise file (#206)

Three on-premise Device Detection tests fail on a machine that has only the free data files, which is a fresh clone after `cxx/data/get-data-files.sh` or a CI run with no Device Detection licence. Two of them do not need anything the free data file lacks, and the third does and fails rather than saying so.
When the data update service cannot read the data file it has just downloaded, it reported the failure with a status the specification does not give for it.
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment