rcedit was passed "%APP_VERSION%"", which stamped ProductVersion as
e.g. 9.17" on Windows installer builds.
Co-authored-by: Cursor <cursoragent@cursor.com>
paramiko 5.0 removed DSSKey entirely, whilst sshtunnel 0.4.0 still refers to
paramiko.DSSKey in SSHTunnelForwarder.get_keys(), which _consolidate_auth()
calls from the constructor. A major bump therefore does not merely drop DSA
key support, it raises AttributeError before any SSH tunnelled connection can
be established, and nothing in CI covers SSH tunnels so it looks green.
sshtunnel has had no release since 0.4.0 in 2019, so there is nothing newer to
move to on that side.
Ignore major paramiko updates until sshtunnel is fixed or replaced. The
exclusion is repeated under the /web/regression entry because
web/regression/requirements.txt starts with "-r ../../requirements.txt", so
that entry sees the root pins too.
Flask-Security-Too 5.8.2, released on 12 August 2026, fixed a long-standing
inversion in its login forms: `LoginForm.validate()` previously read a `True`
return from `UserMixin.is_locked()` as "not locked, carry on", and the base
implementation unconditionally returned `True`. Our `User.is_locked()` was
written against that inverted convention, so as soon as CI began resolving
5.8.2 through the loose `Flask-Security-Too==5.8.*` pin, an unlocked user
returned `True`, form validation failed, the login POST redirected with a 302
and every subsequent request arrived as `AnonymousUser`. The server-mode data
isolation tests caught it, though the breakage is not limited to tests: on
5.8.2 nobody could log in at all.
`User.is_locked()` now returns `True` when the account is locked, matching the
corrected upstream contract, and the dependency is floored at 5.8.2 so that we
cannot silently resolve a release which reads the value backwards. The two
conventions are mutually exclusive, hence a floor rather than a version check
in the model. The regression tests are updated to assert the fixed contract.
Python 3.14 is now supported, so add the trove classifier for it to the
pip packaging metadata and move the desktop builds onto it: the macOS
bundle now defaults to 3.14.7, and the Windows build looks for an
interpreter in C:\Python314 by default, with both build READMEs updated
to match. The minimum supported version is unchanged at 3.9.
Whilst here, the SonarQube scanner's Python compatibility list had drifted
somewhat, still naming 3.7 and 3.8 and stopping at 3.11, so it has been
brought into line with the versions we actually support.
pkg/mac/build-functions.sh reverted to bd841b882's state - keeps the
notarization diagnostic fix (print REQUEST_STATUS + fetch notarytool
log on failure), drops everything from 0296de48e through 6891b5e12.
Five iterations trying to fix _fixup_imports's otool handling of
"pgAdmin 4 Helper (GPU)" (newline-based path list, otool-classic
fallback via bare name, via resolved absolute path, absolute-path
retries for otool/install_name_tool) all failed to get the appbundle
build past this file. Confirmed along the way: the path itself is
byte-for-byte correct and the file verifiably exists at the moment
otool runs (od -c / ls -la evidence from a live buildfarm run); both
otool and otool-classic, called directly with correct quoting and a
resolved absolute path, still fail identically on it. That's a real
bug/limitation in Apple's tooling on this build node, not fixable by
further changes to how this script constructs or passes the path.
Reverting rather than continuing to guess. The stale nspkg.pth /
--system-site-packages fix in pkg/linux/build-functions.sh (verified
working, pgadmin4-rpm-build #180 succeeded) is untouched by this.
pgadmin4-appbundle-build #116 (macos-arm64) failed even with
otool-classic invoked directly via its real resolved path
(669c2f244's fix) - identical error, byte-for-byte, to calling plain
`otool -L`:
can't open file: .../pgAdmin 4 Helper (No such file or directory)
That rules out PATH resolution and otool's internal re-exec dispatch
as the cause too. Both otool -L and otool-classic -L, invoked directly
with correct argv quoting and a verified-correct, verified-present
path, fail identically on this one binary. Whatever's actually broken
is inside these tools' own handling of "(" in a path they're given.
Untested variable so far: this script pushd's into the bundle dir and
passes relative paths ("./Contents/...") throughout. Try a plain
absolute path with each tool before falling back to the diagnostic
dump - cheap to test, and applies equally to the two
`install_name_tool -change` calls further down that edit this same
file in place, since they could hit the identical issue.
Still not proven - if this also fails, the diagnostic dump (retained
as final fallback) will at least confirm parens themselves are the
blocker regardless of path form, which would mean sidestepping via a
temporary parens-free symlink is the next thing to try.
pgadmin4-appbundle-build #115 (macos-arm64) failed even with the
otool-classic fallback from 2d17e9e15:
otool-classic output: .../build-functions.sh: line 260:
otool-classic: command not found
otool-classic only exists inside the active Xcode toolchain
(/Applications/Xcode.app/.../XcodeDefault.xctoolchain/usr/bin/
otool-classic - the exact path visible in otool's own internal error
message), not as a bare command on $PATH. The previous fallback never
actually tested whether the real otool-classic binary can open the
file - it just failed on a missing command.
Resolve the real path via `xcrun -f otool-classic` (falling back to
deriving it from `xcode-select -p` if xcrun's lookup is unavailable)
and invoke that directly.
Not yet verified against a real buildfarm run.
pgadmin4-appbundle-build #114 (macos-arm64) failed on the same
"pgAdmin 4 Helper (GPU)" binary as before, but this time the
diagnostics from d4364555a actually ran and proved the path itself is
not the problem:
Raw bytes of the path (od -c): plain ASCII throughout, single
spaces, no non-breaking/unicode look-alikes, no trailing garbage.
ls -la of the exact path: file exists, correct size/perms,
timestamped seconds before otool ran - not a race condition either.
otool -L still failed via its internal otool-classic re-exec on that
exact, verified-correct, verified-present file. That's conclusively a
bug in otool's own internal dispatch, not anything this script builds
or passes to it - we can't fix Apple's tool, but we can avoid
triggering its broken code path.
Retry via `otool-classic -L "${TODO_OBJ}"` called directly (our own
argv-based quoting, not otool's internal re-exec, whatever that does
differently) when the initial `otool -L` fails. Every other binary in
the bundle keeps going through plain `otool -L` unchanged - this only
engages for the specific case that's already failing anyway. The
diagnostic dump (od -c / ls -la / otool-classic's own output) is kept
as the final fallback in case even the direct call fails too.
Not yet verified against a real buildfarm run.
pgadmin4-appbundle-build #113 (macos-arm64) failed again on the same
otool -L call, but none of the diagnostic output from 401afbfa3 showed
up. Instead:
The command "OTOOL_OUTPUT=$(otool -L "${TODO_OBJ}" 2>&1)" failed in
"_fixup_imports" with exit code 1.
That's this script's own ERR trap firing on the assignment statement
itself. Under set -e, `VAR=$(cmd)` still propagates cmd's non-zero
exit to the trap immediately - the previous diagnostic's separate
`OTOOL_STATUS=$?` line right after never got a chance to run, so
nothing was ever printed, same as the original bare failure.
Move the assignment into the condition of the `if` itself
(`if ! OTOOL_OUTPUT=$(...); then`) - bash explicitly exempts a
command being tested by if/while/until (or negated with !) from
errexit, so this time the diagnostic block will actually execute.
Still not verified - watching the next buildfarm run.
pgadmin4-appbundle-build #112 (macos-arm64) failed with otool-classic
unable to open a path that our own logging showed as correct
immediately beforehand ("pgAdmin 4 Helper (GPU)"). Rather than work
around it blindly (an otool-classic fallback was tried and reverted -
unverified, and risks masking a real problem instead of fixing it),
capture real evidence the next time this happens:
- raw byte dump (od -c) of ${TODO_OBJ} right before otool runs, and
again on failure - catches non-ASCII or invisible whitespace that
a plain `echo` can't show
- ls -la of the exact path and its containing directory on failure -
confirms whether the file actually exists at that exact moment
(rules out/in a race with something still writing it)
- otool's actual stdout/stderr and exit code, instead of letting its
own error text be the only signal
No fallback behavior - fails loudly (exit 1) with this diagnostic
dump so we can tell what's actually happening from the next real
buildfarm run instead of speculating further.
pgadmin4-appbundle-build #110 (macos-x64) failed with:
error: otool-classic: can't open file: ./Contents/Frameworks/pgAdmin
error: otool-classic: can't open file: ./Contents/Frameworks/Electron
_fixup_imports built its worklist with `awk -F':| '`, a regex
alternation that splits on a literal ':' OR any bare space. `file`'s
actual output format is "path: description" (colon-space as one
token), so any space *inside* the path itself - "pgAdmin 4 Helper
(Plugin).app", "Electron Framework.framework", both routine in macOS
app bundles - also got treated as a split point, truncating $1 to
whatever preceded the first space.
The resulting list was then joined with a single space (ORS=" ") and
iterated via an unquoted `for x in $list`, which word-splits on
spaces again - doubly ambiguous between a path's own spaces and the
list's separator, unrecoverable however it's parsed.
Fix: split only on the literal ": " token file emits (`awk -F': '`,
not `-F ':| '`), and switch the whole worklist to newline-separated
instead of space-separated, iterated with `while IFS= read -r`
instead of `for x in $unquoted_list`, at both the outer (executable)
level and inner (library-copy) level (`TODO="${TODO}"$'\n'"..."`).
Not yet verified against a real buildfarm run - can't reproduce the
macOS codesigning environment locally. Watching the next
pgadmin4-appbundle-build run.
The previous fix (bd841b882) tried to delete the specific broken
sphinxcontrib-jsmath nspkg.pth file from the system site-packages
before creating the venv. It didn't work: pgadmin4-rpm-build #177
failed identically on el-10. The delete silently no-opped - that
directory is root-owned on the build node and the build user doesn't
have write access there, so `find -delete` failed while `-print` (which
runs first) still logged the match, making the fix look like it ran.
Root cause is the --system-site-packages flag itself: it uses Python's
site.addsitedir() internally, which does not just add a directory to
sys.path - it also scans that directory for every *.pth file and
executes any "import ..." lines found in them. That's what runs the
broken nspkg.pth's stale namespace-package bootstrap code and corrupts
sys.path before core stdlib resolves, breaking pip's own subprocess.
We only need --system-site-packages for OS-provided packages that
don't have reliable pip wheels (e.g. dbus-python, a hard runtime
dependency per pkg/debian/build.sh's python3-dbus dep, needs
libdbus-1-dev to build from source). A *plain path line* (no "import")
in a .pth file only appends that directory to sys.path - it does not
trigger a further .pth scan of it. So: create the venv without
--system-site-packages, then write the system site-packages
directories as plain lines into a .pth file inside the venv's own
site-packages (which the build user does own). Same OS-package
availability, without ever asking Python to treat the system directory
as a site directory - the broken nspkg.pth is simply never read.
Not yet verified against a real buildfarm run - can't reproduce the
el-10 environment locally. Watching the next pgadmin4-rpm-build run.
macOS build (build-functions.sh): _notarize_pkg only printed
"Notarization failed." on rejection, giving no indication why. Print
the actual REQUEST_STATUS and fetch the full notary log via
`notarytool log` so future failures are actually diagnosable from the
Jenkins console instead of just "status: Invalid".
Linux build (build-functions.sh): _create_python_virtualenv creates
venvs with --system-site-packages, which pulls in the entire system
site-packages dir - including any stale, improperly-uninstalled
package's namespace-package .pth hook. That legacy pip/setuptools
mechanism runs at interpreter startup, before core stdlib is
guaranteed to resolve; a leftover sphinxcontrib-jsmath nspkg.pth on
the el-10 build node corrupted sys.path early enough to break pip's
own subprocess bootstrap, failing the whole build with a misleading
"No module named 'importlib'"/"'traceback'" error
(pgadmin4-rpm-build #176). Remove that exact known-broken file before
creating the venv - not every *-nspkg.pth, since this touches the
*system* Python install on a build node shared by other jobs, and any
other such file could still be load-bearing for something unrelated.
--system-site-packages itself is left untouched - it's required so
venvs can see OS-provided packages not available as clean pip wheels
on every target platform.
evaluate_and_patch_config() called keyring.get_password() synchronously
at `import config` time to detect a selected-but-unusable OS keyring
backend. On Debian 13 (and similar headless/RDP sessions with no live
D-Bus/GNOME-Keyring session), that call - and even the keyring import
itself - can block forever, freezing the whole desktop app before it
ever starts.
Revert evaluate_config.py to the pre-9.17 synchronous check (backend
name only, no get_password call). Move the usability probe into
pgadmin.utils.keyring_probe, run from create_app() in a background
daemon thread so it never delays startup, isolated in its own process
via subprocess.Popen(sys.executable, '-c', ...) so a hang can actually
be killed - a thread-only timeout can't do this, it would leave
CPython's per-module import lock held forever and wedge any later
`import keyring` in the parent process too.
subprocess.Popen (plain fork+exec), not multiprocessing.Process, is
required here: create_app() runs at the top level of pgAdmin4.py while
it's still being imported, and multiprocessing's spawn start method
refuses to start a child before the current process finishes
bootstrapping its __main__ module. An earlier version of this fix used
multiprocessing and crashed the probe thread with RuntimeError on every
single test run, which corrupted the SQLAlchemy/sqlite session for the
rest of app init and surfaced as an unrelated-looking "attempt to write
a readonly database" error on module_preference inserts.
config.USE_OS_SECRET_STORAGE is only ever read from request handlers
requiring an authenticated session, never at import time or inside
create_app() itself, so the async resolution race is safe in practice.
Tests mock subprocess.Popen for the timeout/kill/config-fallback
orchestration (deterministic, no real 3s wait or backend dependency),
plus one test class that runs the real probe script in a real
subprocess against a fake keyring module injected via PYTHONPATH, so
the script body itself has real coverage.
Verified with the full regression suite (--exclude feature_tests) in
both desktop mode (2134 passed, 0 failed) and server mode (2251
passed, 0 failed); remaining skips are pre-existing pgAgent-dependent
job tests.
Each of the Kerberos/LDAP/OAuth2/webserver login-mocking test classes
sets self.app.PGADMIN_EXTERNAL_AUTH_SOURCE (the Flask app instance
attribute before_request() actually reads) in setUp, but tearDownClass
only restored app_config.PGADMIN_EXTERNAL_AUTH_SOURCE -- a different
object (the config module). Whichever of these classes ran last
(test_webserver_with_mocking, alphabetically) left the live app
attribute stuck at WEBSERVER for the rest of the suite, making
before_request() route every subsequent unauthenticated /login request
through authenticate.login()'s webserver auto-auth path instead of
rendering the login form -- surfacing as an unrelated CSRF-harvest
failure in test_close_requires_auth much later in the run.
Full server-mode regression: 2507 tests, 0 failures/errors (was 2).
get_test_user() created secondary test clients via app.test_client()
but never called setApp() on them, so any test using a second user
crashed with 'NoneType has no attribute config' the first time it
needed to fetch a CSRF token (44 errors in a full server-mode run).
TestSqlEditorCloseRequiresAuth.setUp() logs out the shared
class-level tester before asserting a CSRF token was harvested; when
that assertion failed, unittest skipped tearDown() entirely (by
design, tearDown only runs if setUp succeeds), leaving the tester
logged out for the rest of the suite and cascading into 188 unrelated
failures. Switched to addCleanup(), which always runs.
Full server-mode regression: 2507 tests, was 232 failures/errors,
now 2 (both isolated to this same test's own CSRF-harvest timing in
full-suite order, unrelated to these two fixes).
_is_query_parens_balanced() always treated \' inside a single-quoted
string as an escaped quote, matching psql only when
standard_conforming_strings=off. Under the actual default (scs=on,
every supported PostgreSQL version), psql treats \ as a literal
character, so 'a\' closes the string right there. A crafted query
export payload like SELECT 'a\') TO PROGRAM 'cmd' x' passed the
balance check while the real ) it hid closed the wrapping \copy (...)
context in psql, exposing a live TO PROGRAM clause for RCE.
Since the correct interpretation depends on a server setting we
can't reliably know, reject any backslash inside a single-quoted
string outright instead of guessing.
Reported by Arpit Jain (arpitjain099).
sqlparse's string-literal lexing can disagree with PostgreSQL's: under
standard_conforming_strings = on (the default), a backslash before a
quote is an ordinary character to PostgreSQL but sqlparse treats it as
escaping the quote, so a payload like
SELECT '\';COMMIT;CREATE TABLE pwn(x int);SELECT 1 --'
passes _validate_readonly_query as a single SELECT while PostgreSQL
executes it as four statements -- the smuggled COMMIT ends the wrapping
BEGIN TRANSACTION READ ONLY and the trailing ROLLBACK is a no-op,
reintroducing the write/RCE bypass the bf47924444 fix was meant to
close (reported by Kai Aizen / SnailSploit).
Run the LLM-supplied query with prepare=True, forcing psycopg3's
extended query protocol. PostgreSQL's own Parse step -- not a
client-side approximation of it -- rejects any text containing more
than one statement, independent of how it's lexed. Threaded through
execute_2darray as an opt-in parameter (default None) so no other
caller's behavior changes. Also set SESSION CHARACTERISTICS AS
TRANSACTION READ ONLY as defense-in-depth against a smuggled COMMIT.
prepare=True alone is not sufficient: psycopg3's PrepareManager.get()
returns Prepare.NO -- checked before it even inspects the prepare
argument -- whenever the connection's prepare_threshold is None, which
is pgAdmin's per-server default ("Prepare threshold" is blank unless an
administrator sets it). On a default-configured server the extended
protocol never actually engaged, so the bypass stayed live. Force
prepare_threshold=0 on the LLM's single-use connection in
_connect_readonly() so the extended protocol -- and PostgreSQL's
single-statement Parse-step guarantee -- is unconditional on this
connection, without touching the server-wide setting or any other
session/caller.
Adds regression coverage at three levels: unit tests pinning that
prepare=True is always passed and that prepare_threshold=0 is forced
on the connection; and an end-to-end test driving the real
/sqleditor/nlq/chat/<trans_id>/stream route (LLM client mocked, real
tool-dispatch path) with the exact smuggled-COMMIT payload, confirming
the guarantee holds from the HTTP entry point down to the driver call.
Follow-up to #10142. Geometry Viewer's custom tile provider was only
configurable per-user (Preferences), with no way for an administrator
to set an organization-wide default (e.g. an internal tile server)
that applies out of the box for every user.
config.DEFAULT_GEOMETRY_VIEWER_PROVIDER is a plain dict; the five
custom_tile_* preference defaults (url, name, crs, attribution,
max_zoom) are sourced from it instead of hardcoded literals.
Preference.get() only reads a per-user DB row if the user has
explicitly saved one, otherwise it falls back to this default --
override semantics, one active provider, no naming-conflict surface.
config_local.py/config_distro.py/PGADMIN_CONFIG_* replace this config
variable wholesale, not merged key-by-key, so a naive
DEFAULT_GEOMETRY_VIEWER_PROVIDER['name'] lookup would KeyError-crash
preference registration (and app startup) on any partial admin
override, or AttributeError on a wrong-typed value.
resolve_geometry_viewer_provider_defaults() guards against this:
falls back per-field on missing/wrong-type values, validates crs
against the 3 supported choices, validates max_zoom is a non-bool int
in [0, 25], degrading to the original hardcoded defaults instead of
crashing.
No new attacker-reachable surface: DEFAULT_GEOMETRY_VIEWER_PROVIDER is
filesystem/deployment-level admin config, same trust boundary as
DEFAULT_BINARY_PATHS/OAUTH2_CONFIG. Existing DOMPurify sanitization on
name/attribution and the http(s):// + {x}/{y}/{z} URL validation in
GeometryViewerUtils.js (from #10142) apply uniformly to config-sourced
and per-user values alike.
Adds test_geometry_viewer_provider_defaults.py (9 scenarios: fully
valid, partial override, non-dict, invalid CRS, out-of-range/non-int/
bool max_zoom, non-string fields, empty dict).
Extends the tool-RBAC fix (permission enforced only on a blueprint's
"front door" route, with the rest relying on @pga_login_required alone)
to four permission types the original fix didn't cover:
* tools_ai ("AI Reports"): none of the 14 Security/Performance/Design
report routes in the llm blueprint checked this permission at all --
only the frontend menu was hidden behind AllPermissionTypes.TOOLS_AI
(llm/static/js/ai_tools.js). A user an admin had denied AI access to
could call the report/report-stream routes directly and have real
server/database/schema content sent to whichever external LLM
provider is configured, bypassing both the access control and the
data-exposure/cost boundary the permission exists to enforce.
* tools_import_export_servers: only 'save' was gated; 'get_servers' and
'load_servers' relied on @pga_login_required alone. Lower impact than
the above: get_servers only returns the calling user's own servers,
and load_servers only previews an already-uploaded file without
persisting (the actual write stays gated at 'save').
* tools_search_objects: only 'search' was gated; 'types' (a static list
of supported object-type names for the dropdown, no DB content) was
not.
* change_password: the self-service password-change route
(browser.change_password) never checked this permission, so revoking
"Change Password" from a role in Role management had no effect --
users under that role could still change their own password. Lower
severity than the others since the route is inherently self-scoped
(can only ever affect the caller's own password).
Extends tools/tests/test_tool_permissions_required.py with 17 new
scenarios for the first three, and adds a dedicated
browser/tests/test_change_password_permission_required.py for the last
(different blueprint, different permission category).
Verified against a live PostgreSQL server in SERVER mode: all 40
scenarios (37 + 3) pass with the fix applied. Negative control:
reverting the four fixes reproduces exactly the reported gaps -- 17 of
the 20 pre-existing scenarios still pass unaffected, the 17 new ones
fail, and change_password returns 200 instead of 403 -- confirming the
tests actually catch the regression rather than passing vacuously.
Supersedes 7 open Dependabot PRs by applying the safe ones and
properly fixing the one with a broken lockfile, in one CI cycle:
- setuptools ==82.* -> ==83.* (#10144, #10145 - duplicate PRs, same
patch). Only touches the `python_version > '3.9'` line; the
`<82; python_version <= '3.9'` gate for Python 3.9 is untouched,
so this doesn't affect Python 3.9 support.
- fast-uri 3.1.2 -> 3.1.4 in both /web and /runtime (#10183, #10195)
- fixes two real CVEs (GHSA-v2hh-gcrm-f6hx, GHSA-4c8g-83qw-93j6).
- tar 7.5.16 -> 7.5.21 (#10182) - patch series, decompression-bomb
and unbounded-recursion hardening only.
- shell-quote 1.8.4 -> 1.10.0 (#10185) - additive opt-in option +
parser fixes, no breaking changes.
- svgo 3.3.3 -> 4.0.2 (#10184) - Dependabot's own PR left yarn.lock
internally inconsistent (dropped the workspace-level `svgo` entry
while merging version-range blocks), so `yarn install --immutable`
failed in CI with "the lockfile would have been modified by this
install". Regenerated properly via `yarn up`/`yarn dedupe` here.
The direct `svgo`/`svgo-loader` deps are not actually wired into
any webpack rule (verified via grep) - the real SVG pipeline is
`@svgr/webpack` -> `@svgr/plugin-svgo` -> svgo 3.3.3, which this
bump does not touch - so the major version jump has no build
impact. `yarn.lock` now correctly keeps that separate 3.3.3
resolution alongside the deduped 4.0.2 one.
Not included (structurally blocked, tracked separately):
- paramiko 3.5.1 -> 5.0.0 (#9927): paramiko 5 removes DSSKey
entirely; sshtunnel 0.4.0 (dormant since 2021) still references
paramiko.DSSKey, so `import sshtunnel` would crash immediately.
- pywinpty 2.0.* -> 3.0.* (#10082, #10084): the existing pin cites
https://github.com/andfoy/pywinpty/issues/545, confirmed still
open ("process read and write not working as expected in 3.x").
Verified: `yarn install --immutable` clean in both /web and
/runtime, `yarn run linter` clean, full `yarn run bundle:dev`
compiles successfully.
Flipper
Martin Manka
byshyandCursor
dependabot[bot]anddependabot[bot] <49699333+dependabot[bot]@users.noreply.github.com>
Dave Page
Ashesh Vashi
Domenico Sgarbossa
Kundan