Changelog
What changed in each release of searchcode enterprise, newest first. Read every version between the one you run and the one you are moving to before you upgrade, and read its Upgrade notes first: they are the part that can need something from you before the restart. The upgrade procedure itself is in Upgrades.
searchcode version prints the version you are running. This file ships in the
tarball as CHANGELOG.md, and each release's section is attached to the release
as its notes.
Unreleased
The first numbered release.
Upgrade notes
- Every repository is analysed again after this upgrade, because imports are now read from Java, C#, Rust, C, C++, Swift, Ruby and PHP, and because the line counter changed how it counts Scala, Kotlin, Zig, Perl, Nim, COBOL, Swift, F#, Twig, shell heredocs, the strings of C#, VB, Pkl, Dart, PowerShell, Coq and Nix, character literals in Elixir, OCaml and Nim, VB comments and keywords, and a few complexity checks, and now recognises IBM i languages, ProGuard, G#, Faber, Topaz and Lispex. Binary documents with a long text header (Illustrator, some PDFs) are also now refused as binary, and anything the analysis had made of them is cleared by the same re-analysis. Search answers throughout; a repository not yet re-analysed reports a gap. On a 450-repository estate with 4 CPUs this has taken between 14 and 50 minutes.
- Builds from before this release report a commit and no version. Upgrade as
Upgrades describes: back up
subjects.db, replace the binary or image, and restart. - If your proxy injects a group to make someone an administrator, because your
identity provider sends no groups, you can name the person instead with
authz.adminUsers(andauthz.auditorUsers, andusers:on a binding). The injected group keeps working; naming the person is what the proxy recipes now do. The lists apply inauth.mode: proxyonly. - API keys issued by an earlier build have the holder's email filled in from their sign-in record on the first start. Each key's email is then fixed at the value it had when issued.
search.concurrencyneeds a restart to change. A reload that changes it is refused and names the field. On earlier builds the reload was accepted and the new value was shown, but searches kept running at the old concurrency.- Running without a licence now means 3 users and 5 searchable repositories (it was 25 and 50, and the repository number only warned). Nobody already using the instance is turned away: a fourth person it has never seen is refused. On an estate of more than 5 repositories, the first 5 by name stay searchable and the rest are cloned and not indexed, and every search says so. If you run without a licence and need more, ask for one before upgrading.
- A licence that states no repository or user limit is now unlimited, as
licensegenalways said. Until this release such a licence got the community limits instead.
Security
- searchcode no longer opens Multipath TCP (MPTCP) listeners. Go 1.24 and
later open them by default, and CVE-2026-98432 lets a remote attacker with
no credentials take over an unpatched Linux kernel through the IPv4 TCP
handshake on an MPTCP listener. The binary and the image now listen on plain
TCP whatever the environment sets.
ss -Mlnpshould list none of its ports. Any other Go program in front of searchcode, such as Caddy or oauth2-proxy, needsGODEBUG=multipathtcp=0in its own environment until your kernel is patched. - Changing a repository's
groups:now takes effect when the configuration reloads. Before, it waited for that repository's next fetch: up to about 90 minutes at the default refresh interval, or until a restart. A repository moved OUT of a group stayed visible to that group for that time, while the reload was reported as applied. This applies torepositories[].groupsand to a connection'sgroups:. The change costs no fetch. If you changed a group recently and want to be sure it is in force, look at the repository'sgroupsonGET /api/v1/repos.
Added
- Bitbucket Cloud connections. A connection of
type: bitbucket-cloudwithurl: https://bitbucket.orgenumerates whole workspaces (include.orgs) or single projects (include.projects: [workspace/KEY]), andsearchcode discover --provider bitbucket-cloud --workspace NAMEdoes the same from a terminal. It is also on the connect screen and insearchcode setup. Use a workspace access token withrepository(read), available on the Premium plan, or an API token withread:repository:bitbucketon any plan; app passwords no longer work, since Atlassian removed them on 2026-07-28. Bitbucket Cloud is a public host, soconnections.egress: internal-onlyrefuses it. A self-hosted Bitbucket (Data Center) is not yet a connection type. See connections. - File dependencies for Java, C#, Rust, C, C++ and Swift.
file_dependenciesand the graph view now cover all ten languages the analysis parses. A Java class resolves by its package and file name, a C# using to the namespace's directories, a Rustmodoruseto its module file, an include to its header, and a Swift import to its module. Files of one Java package, C# namespace or Swift module are linked to each other, and tests namedFooTest.java,FooTests.csorFooTests.swiftare matched to the code they test. - Ruby: outlines, code navigation, file dependencies and the graph. Ruby
files now have an outline and exact definitions, and
file_dependenciesreadsrequire_relative,require,loadandautoload. Because most Ruby code requires nothing it uses (a Rails application loads by naming convention), a Ruby file is also linked to the files defining the constants it names, found the way Ruby looks a constant up. These rows have the kindconstantand are alwaysprobable. A constant that no file in the repository defines, such asStringor a gem's class, is not reported as a missing import. RSpec*_spec.rband minitest*_test.rbfiles are runnable tests, matched to the file they are named for. - PHP: outlines, code navigation, file dependencies and the graph, for
Composer projects and for older code such as Drupal 7 and WordPress. A
useresolves to the file declaring the class, or throughcomposer.json's PSR-4 prefixes.requireandincludeare read, and so is Drupal'smodule_load_include: a path built on__DIR__is exact, and one built on a constant such asABSPATHorDRUPAL_ROOTis matched by its tail. Drupal 7's.module,.inc,.install,.theme,.profile,.engineand.testfiles are read as PHP when they start with<?php, so the analysis now counts and parses them too. Code outside a namespace is linked to the files declaring the classes it names (kindconstant) and the global functions it calls (kindfunction, new, PHP only). Both are alwaysprobable, and a function declared in two files links to neither. On Drupal 7 core this takes the share of files with no link from 54% to 14%, and on WordPress from 35% to 5%. Files of one namespace are linked to each other, and PHPUnit's*Test.phpfiles are runnable tests. - Code search across every repository you connect, with the file view, blame, history, and code navigation that labels each answer exact or probable. See Searching the estate.
- One HTTP API under
/api/v1, with its reference served by the instance at/api/v1/docs, and an MCP server whose tools are the same calls. See API and MCP tool reference. - Sign-in through your own SSO proxy, with built-in accounts, or open; access by repository group and by named person, explained on request. See Signing in and Who sees what.
- Per-person, per-client API keys, sign-in for MCP clients, and service accounts declared in configuration. See Which credential do I use?.
- An audit stream on every tier, to a file, stdout, stderr or syslog. See The audit log.
- GitHub and GitLab connections that keep the repository list current, and
searchcode discoverfor a list you review yourself. See Code-host connections. - Analysis of every repository: security findings, code quality, dependencies and the licences manifests state, hotspots, change coupling, and trends over time. See searchcode enterprise — what it does.
- Headless mode, with
server.uiListen: off, for the API and MCP without the web UI. - The licences of the third-party software and data in the binary, on a new
admin screen at
/admin/licenses, shown to administrators only. They are compiled into every copy of the binary. - How often searches run out of
search.timeout: on the admin page, on/api/v1/status, assearchcode_search_timeouts_total, and in the shippedSearchcodeSearchTimeoutsalert. A steady count means raising the budget or adding CPU. See Monitoring. - Prometheus metrics on a listener of their own, and
searchcode doctor, which diagnoses a stopped instance and writes a redacted support bundle. See Monitoring and searchcode doctor. - A Helm chart, a tarball with systemd units, and a container image. See Install with Helm and Install on a VM with systemd.
- Project direction, at
/directionandGET /api/v1/project-direction: is each project getting better, worse, or just larger? The page opens on a list of every repository you can see, worst first, each with one word and one sentence; choosing one shows a word and a sentence for each of size, ease of change, findings, vulnerable dependencies, how many people the work is spread across, and effort, plus what the page cannot tell you. "Direction" replaces "Trends" in the header; the trend pages are reached from a project's ease of change. See kb/project-direction.md. Findings and vulnerable dependencies have history from the day you upgrade; size is rebuilt from git history as well, so it is measured from the first day. A repository first committed inside the window reads "New in this window". Each sentence states the numbers that moved ("worse: busiest files 24% more complex per line"), and a project with no commits in the window says so rather than being marked worse on who does the work; nor is a drop from three people to two. The readings are kept ininstance.dataDir/snapshots/, which belongs in your backup (see Backup and restore); older trees are checked out one at a time intoinstance.dataDir/snapshot-backfill/to be counted, so leave room there for your largest repository. - The complexity trend page opens with one or two plain sentences saying what the number is. Every note it showed before is still there, behind "How this was measured", and the dashed control line now says what it is for.
- A licence can admit service accounts only. With
access: service-accounts, the instance runs with no web UI and only the service accounts inserviceAccountscan use the API and MCP; a person's key, an OAuth token and open mode's anonymous access are refused and audited asauth.access_restricted. Nothing is deleted, and it holds past expiry. See Licensing. - A graph view, in preview.
/repo/<repository>/graphdraws every file in a repository with what it imports and what it changes together with, and the other repositories it imports from as single bubbles you can click through./graphdraws how your repositories depend on each other. Nothing links to either page yet. Both come from data the analysis already holds, and both show only repositories you may see. The data is also on the API asGET /api/v1/repos/{repo}/graphandGET /api/v1/graph. The pages use one JavaScript library, D3, served from the instance itself and on these two pages only; with JavaScript off they show the same graph as a table.
Changed
-
Replacing a secret file now takes effect without a restart or a config edit. A code-host token, a clone credential, a service account's token and the proxy shared secret are all picked up about two seconds after their file is replaced. The
configuration reloadedlog line names the secrets that changed, never their values. Clones that failed under an expired token retry at once with the new one. Write the new value to a temporary file andmvit into place. A service account's old token stops working at that moment, so give the application its new token first. Replacing theconnections.keyfile is refused while running, because tokens saved on the connect screen cannot be opened with a different key. -
A connection refuses
includefields its host does not take, naming the one to use.include.projectsis Bitbucket Cloud's, and on GitHub or GitLab it fails config load. On Bitbucket Cloud,include.usersandinclude.selffail the same way. -
A rate limit that does not say how long to wait now reports a resume time. A connection paused by such a limit used to report that it would resume immediately; it now reports the time of its next attempt.
-
User names are removed from discovered clone URLs. A code host that lists
https://name@host/...no longer puts that name in front of the clone credential you declared. -
Choosing a technology on the results page now narrows, the way choosing a language always has. With a technology selected, the technology list shows only what those files also use, with counts within them. Clicking another technology adds it, so the results narrow to files using both. Clicking a selected one removes just that one. Before this, the list kept every technology the search matched, and clicking one replaced the selection, which could widen the results.
-
Two
tech:filters in one search now mean "uses both". Before this they meant "uses either". The same goes fortech:together with thetechnologyparameter. A comma inside one filter (tech:A,B) still means either, as it does forlang:andrepo:. A search with one technology answers exactly as before. Thetechnologiescount on a search response now honourstech:like every other filter. -
An OAuth access token now works on the REST API as well as on MCP. It used to be refused on
/api/v1/*, so a client that reads both needed a static API key beside its token. No bearer credential of any kind is accepted on the UI listener, where keys are issued, so a token still cannot be used to create a longer-lived key. The full sign-in and refresh sequence for a desktop client is in kb/mcp-oauth.md. -
An expired or revoked API key is now told so. The 401 says
credential_expiredorcredential_revokedinstead of the genericunauthenticated. Both are answered only after the key's secret has been checked, so the caller already holds the key. An unknown key and a wrong secret still get the same generic answer. -
Every API and MCP response says when the credential that made it expires, in an
X-Credential-Expiresheader (RFC 3339), and/api/v1/meand/api/credentials/selfcarry the same instant. A credential with no lifetime, such as a service-account token, gets no header. -
Authorization: Beareris the one header to use for every credential type on both/api/v1and/v1/mcp.X-Api-Keystill works for a key. The 401 message, the OpenAPI document and the account page now all say this. -
searchcode now keeps its memory under the limit it runs in. A new setting,
instance.memoryLimit, defaults toauto: 90% of the container's memory limit, or of the machine's memory when nothing sets one. Near it, the process collects garbage harder rather than growing until it is killed. On a machine with room to spare nothing changes. If you already setGOMEMLIMIT,autouses it. Set an amount such as12GiB, a percentage such as80%, oroffto choose for yourself; the change applies without a restart, and/api/v1/statusshows the limit in force. See Sizing. -
The deep analysis pass uses more of a large machine.
analysis.deepConcurrencynow defaults to0, meaning the larger of 4 and half the CPUs: 4 on a 4-core machine, as before, and 16 on 32 cores, where a fixed 4 left the machine mostly idle and findings, symbols and dependencies took most of a day to complete at 2,000 repositories. These workers run at a lower priority than search. If you setdeepConcurrencyyourself, your value is kept; lower it if memory, not CPU, runs out first. -
Searches for longer words and for several words read far fewer files. The search index now keeps a second filter, of four-letter sequences, beside the three-letter one, so a file that merely contains the letters of a word scattered across other words is no longer opened and checked. Searches for words of three letters or fewer are unchanged. The cost is memory: the search index uses twice what it did,
search.index.bloomSize/4 bytes per file (2 KB at the default), because that setting now sizes both filters. A restart takes somewhat longer to rebuild the indexes; searches answer throughout, from repositories not yet indexed by reading them directly. -
Case-insensitive search allocates far less while checking candidate files. Every near-miss used to build a short-lived list of case variants; on a 2,000-repository estate that was most of the memory a search allocated and most of the garbage collection behind slow searches under load. A search that finds nothing now makes about 1% of the allocations it did. Results are unchanged.
-
Searches over many repositories are much faster. A search used to visit your repositories one at a time, and paid a fixed cost for each one even when the index had already ruled it out, so its time grew with the number of repositories rather than with what matched. A search that found nothing paid that cost twice. Now one search works through all its repositories at once. On a 2,000-repository test estate a search that finds nothing went from about 95 ms to 5 ms, a rare word from 170 ms to 11 ms, and a common word from 390 ms to 30 ms. Results are unchanged.
-
Concurrent searches scale. With eight people searching at once on the same estate, searches per second rose between four and nine times, depending on the query.
search.concurrencystill limits how many searches run at the same time. -
Checking each candidate file is cheaper. Files the search reads and rejects reuse memory instead of allocating it, which cuts CPU per candidate by 11–29% and memory allocated per search by up to 96%. Results are unchanged.
-
When a search runs out of time, the repositories it reports as searched are the ones that finished. The partial answer still says how many of your repositories it covered, and keeps everything they matched. Repositories are still started in order, so the ones covered are nearly the first ones, but no longer exactly the first N.
Fixed
-
Technologies whose names contain a comma can be selected. Clicking one, such as "Sinon (Spies, Stubs, Mocks)", searched for two technologies that do not exist and returned nothing. A quoted
tech:value is now one name. -
Asking for a file or directory that does not exist is no longer audited as
denied. It is nownot_found, the outcome a mistyped repository name already had. Access is decided per repository, so a missing path inside a repository the caller can see refused nobody anything; an agent guessing at a directory layout was filling the field compliance rules alert on. A repository that exists and the caller may not see is stilldenied. -
The Lifetime field on
/accountno longer replays an earlier value. Browsers could fill it from a previous entry, which issued keys that expired after a day without the person noticing. The field now turns autofill off, shows the configured default (api.apiKeys.ttl.default) instead of a fixed "720h", says what an empty field gets, and the new-key panel says when the key expires. Theapikey.issuedaudit record now includes the expiry. -
An OAuth client using a private-use redirect such as
com.example.app:/callbackcan now finish signing in. The consent page's security policy blocked the final redirect to it, so pressing Authorise did nothing. -
project-directionsays what its chart points measure. Each row has a newunitfield, such ascode linesorfindings per 1,000 code lines, and it is empty on rows with no points. Ease of change points are the per cent change since the first reading in the window, not the density itself, and vulnerable dependencies is a count, not a share. Until now nothing said either. -
findingsandfind-usagessay what an answer covered. Both now return ascope(estate or the repositories named, with the count and their names), and findings returns thefiltersit applied,testsincluded. An empty answer used to look the same whichever repository it was about. -
A service account is always counted against the user limit. One whose repository group had the same name as an
authz.adminGroupsentry was treated as an administrator by the user meter and never counted. -
Binary documents with a long text header are no longer searched as text. Illustrator (
.ai) and some PDF files begin with readable PostScript or PDF syntax and only turn binary further in, past the first 10,000 bytes the binary check used to read. They were indexed as text, matched almost every search and were read on each one. The check now reads the first 100,000 bytes. On a corpus of 68,000 files from real repositories this reclassified exactly those files and nothing else. Such files now count as skipped binary files. -
A search result shows the lines that best answer the query, not the first ones in the file. Each result shows up to five matching lines. They used to be the first five, so a search for
func ParseProxyUrlover a file with two hundredfunclines showed five of those and none of the three lines namingParseProxyUrl. Lines are now chosen by how many query terms they hold and how rare each term is in that file, and still shown in file order. For a single-word search, lines where the word stands on its own, matches the case you typed, or appears more than once now come first, so some results show different lines than before. The match count is unchanged. Searches with many matching lines are slightly faster.
These change what some searches return on builds before this release:
- A branded instance shows one logo, the one made for a dark background.
With both
branding.logoLightandbranding.logoDarkset, readers whose browser prefers a light colour scheme were shown the light-background logo on the UI's dark header. The UI has only a dark theme, sologoDarkis now always used when set, andlogoLightonly when it is the one logo you configured. Nothing to change unless yourlogoDarkwas not meant for a dark background. - An
.icofavicon is accepted.branding.faviconrefused every ICO file as "not an image" while listing ICO among the accepted types. PNG favicons were unaffected. - A comma list in
ext:now matches every extension in it.ext:ts,tsxsearched.tsfiles only, and dropped any search term written after the list. The API'sinclude_extparameter had the same fault with two or more extensions.ext=ts,tsxmatched every file and now matches the two extensions; the same goes for lists onfileandpath. - In a regular expression,
^and$match at a line, not only at the start and end of the file, and a regular expression is case-insensitive unless the search is case-sensitive, like every other term.(?-i)inside the pattern restores case. - Two notices that misread working queries are gone. One said a
/inpath="apps/auth"started a regular expression; the other asked for brackets around an OR that already had them, or warned about everyNOT. - A
!at the start of a filter value is explained:path="!vendor"looks for the text!vendor, and the notice namespath!=vendor. - The usage page's "Searches that found nothing" says where each search was asked: across every repository, or in which ones. A search narrowed to one repository that finds nothing is often the right answer.
- A regular expression uses the search index when it can. Every regular
expression used to read every file in every repository. One whose matches
must all contain a particular piece of text, such as
/[a-z]+Handler/or/aws_[a-z_]+/, now reads only the files that contain it. A pattern with no such text, or one that matches any of several (/foo|bar/), still reads everything. Within a file, a regular expression that cannot span a line now runs only on the lines that contain that text. - The search index matches case the way search does. A file that spells a
letter in an unusual case form, such as the long s
ſforsor the Greek final sigma, could be missed by an indexed search that a full scan would have found. - A search that only excludes is refused.
NOT teston its own read every file and answered with nothing, which looked as though every file containedtest. It now answers400and asks for a term to look for.handler NOT test, and a browse scoped by a filter such asNOT test lang:go, work as before. - Stopping searchcode with
kill -9, or the out-of-memory killer, stops its git processes too. They kept running and writing into a half-finished clone, and the next start could keep that partial clone instead of discarding it. Under systemd or in a container this was already handled. - The admin page's code-host connections panel shows your connections. It
rendered empty on every instance, because
/api/v1/adminleft theconnectionsblock out of its response. The block is in the response now. - A search that runs out of time keeps what it found, and says so. It used
to answer
400, the search could not be run, as if the query were malformed, and threw away every match it had found. It now answers with the matches from the repositories it finished, and a gap saying how many of them that was and that the totals are lower bounds. A search that finished no repository answers 504 witherror_code: tool_timeout. Both suggest narrowing the search. search.timeoutapplies on reload. On earlier builds a reload showed the new value and searches kept the old one until a restart.- Regular expressions with fixed text in them are faster, case-insensitive
ones most of all: files without the text are skipped before the pattern runs.
/BRMP-[0-9]+/took 200 times longer case-insensitive than case-sensitive and now takes about the same. - Pages no longer show unstyled for a moment on refresh. The stylesheet and scripts were sent with no caching headers, so the browser downloaded them again on every page, and behind an authenticating proxy each download was also a sign-in check. Their links now carry a digest of the file and are cached for a year; an upgrade changes the digest, so the browser never uses an old stylesheet with a new page. Branding assets are cached the same way.
- Search indexes build about four times faster and use about a third of the memory, so searches are fast again sooner after a restart or an upgrade. Building an index allocated memory for every three-letter piece of every word in every file, and copied each file three times to tidy whitespace that the word split ignored anyway. It now does neither. An index built by this release is identical to one built by an earlier release.
- After a restart, the next index starts building as soon as one finishes. Each build slot used to wait up to half a second between builds, so an estate of many small repositories spent about half its warm-up idle.
- The complexity trend no longer counts the same files twice. Files left out for being created inside the window were also reported as files that "produced no measurable point", so one page could say "9 files" in both places for the same nine.
- The web UI counts an estate of more than 1,000 repositories correctly.
It read one page of the repository list and treated it as the whole
estate, so a search over 2,000 repositories said "in 1,000 repositories"
and the repository filter said "Listing 539 of 1,000". Repositories past
the first 1,000 by name can now be filtered on, found through the
repository filter's box, and searched by a repository-group filter, and
their result headings show their names. The repository listing
(
GET /api/v1/repos,list_repositories) gains aphasesobject: the number of matching repositories in each phase, counted before paging.