Repository navigation
Changed
np workflow publishrequires--expected-version, and the MCP toolspublish_workflowand
update_workflow_definitionrequireexpectedVersion. A version that is out of date is rejected
with409 workflow_version_conflictinstead of overwriting newer changes. Scripts and agent
prompts must pass it (#708).- Sign in once more per profile with
np auth loginafter the update. Every profile now has its
own session file, so profiles such asprod.aandprod_ano longer share one login (#620). - A
matchescondition with an invalid pattern, a regex timeout or an input longer than 1,048,576
characters fails the run and names the edge, instead of counting as "no match" (#680). - Two active nodes can no longer share an output variable, regardless of case: saving, importing
or an MCP change with such a pair is rejected. A disabled node is ignored, so a disabled copy can
no longer take over the alias of the active one. A workflow saved earlier with such a pair keeps
running and triggering as before, with the last node holding the alias; the designer flags the
pair, and it must be resolved before the next save. An older export with such a pair is
rejected on import and names the alias (#703, #789). GET /api/executions/{id}/paused-steps,np exec paused-stepsand the MCP tool
list_paused_stepsreturn one snapshot per paused step (step id, label, pause time, reason and
redacted variables) instead of a list of step ids. The MCP result field is nowpausedSteps
(#606).- Folder and workflow lists reuse one set of permission data per user and request, avoiding an
additional database query for each folder while preserving access rules (#576).
Fixed
- Designer
- Read-only workflows keep note and group editing controls disabled. Collapsing or expanding a
group changes only the local view, without autosave or an unsaved-changes warning. Save and
publish also stop queuing writes when the edit lock is lost (#584). - Renaming an output variable updates every reference to it: templates in any field, a dynamic
target machine or credential, breakpoint conditions, decision cases and edge conditions,
including templates in their fixed comparison values (#689, #789). - Copying, pasting and duplicating give each copy its own output variable (
name_2) and rewrite
the references inside the copy, so it no longer takes over the original's bindings (#703). - Inserting an activity on a conditional edge keeps the condition in front of the new activity,
so the branch still runs only when the condition holds (#693). - Deleting an edge from the command palette marks the workflow as changed and can be undone
(#694). - Undo right after typing in a property field restores the value from before the edit (#595).
- A save, lock change, publish or run that finishes after you switched to another workflow no
longer acts on the workflow you opened in the meantime (#692). - Undo and a restored draft also update open property panels: an edge switches back to its
expression, an agent's host list shows the restored values (#695). - Text-file delete can switch to line-range and pattern mode again (#601).
waitForConditioninportOpenmode saves the default port 443 it shows, and a probe without
host or valid port is flagged. A port given as a template or as a number in quotes is accepted,
as the engine accepts it (#702, #790).scheduledTaskregistration accepts only built-in or managed service accounts and no longer
suggests ordinary domain users with delegation. A task registered for a managed service account
(DOMAIN\gmsa$) now runs under it; before, it was registered but never started (#705, #790).- The step details of a past run label the workflow's current configuration as such and show the
target machine the run actually used (#603). - A
forEachparallelism of 0 is saved as 0 (parallel up to the engine limit) instead of 1 (#602). - After a reconnect or a reload, the designer shows a paused debug step with its variables again
and catches up on steps that changed during the disconnect (#606). - Variable edits at a breakpoint belong to that one pause and are dropped when another step
pauses. Global variables and engine keys are read-only in the inspector, and the API rejects
such overrides with 400 while the step stays paused, instead of failing the run after the
resume (#696, #697).
- Read-only workflows keep note and group editing controls disabled. Collapsing or expanding a
- Management pages and settings
- Saving credentials, machines, users, global variables, maintenance windows, custom activities,
alerting rules and policies or a settings section locks the form until the server answers. A
late answer no longer closes or resets a record opened in the meantime (#690).
- Saving credentials, machines, users, global variables, maintenance windows, custom activities,
- AI
- The AI chats keep working when browser storage is full or blocked. A note says what survives a
reload; once the storage budget is reached, only the newest history is kept (#701). - Switching between conversations or regenerating an answer in the designer's AI chat no longer
carries a proposal's discarded, applied or deselected state over to a different proposal
(#815).
- The AI chats keep working when browser storage is full or blocked. A note says what survives a
- Running workflows
- A breakpoint condition can use the run's inputs (
{{manual.*}}) (#685). - Input values that look like
{{…}}stay plain text and no longer fail the step as unresolved
references (#686). - Editing a text file keeps its permissions (#683).
wmiQuerymethod calls pass their arguments in the types the CIM method declares, so a UInt32
above 2147483647 such ashDefKey2147483649 (HKCU) works. A value outside the declared range
fails the step with a message that names the argument (#616).- A
fileWatcherTriggeron a busy server no longer reports its directory as unreachable when the
check simply waited for a free worker thread. The time limit now covers only the directory
access itself.
- A breakpoint condition can use the run's inputs (
- Notifications
- CLI and MCP
np exec --watchshows the final state even when the live connection drops the last event
(#725).- The MCP server rejects broken nodes and edges before it merges a workflow change (#726).
- All MCP requests, including dependency queries, wait until the server connection is ready
(#727). np auth login --password-stdinsends the password as read, including leading and trailing
spaces, so the CLI and the Engine Switcher sign in with such passwords again (#617).np backup manifest,previewandrestorehonour-o jsonand-o yaml(#618).np machine create --sslwithout--portand the MCP toolcreate_machinewithuseSsland no
port use the HTTPS port 5986 (#619).-o yamlkeeps multiline strings, empty lists and objects, and strings such asnull,yes
or a date as they are when the output is parsed again (#621).
- Setup and uninstall
- A server uninstall whose cleanup fails stops before the "Apps & features" entry is removed,
keeps the installation marker and the deploy scripts and exits with a non-zero code, so the
same uninstall can be run again. A failed service deletion stops it before any file is removed.
The desktop uninstaller stops the same way and keeps NodePilot installed (#714). - A desktop update keeps a custom data path and custom service names (#709).
- When a desktop update fails after its services were removed, the rollback registers both
services again with their previous settings and folder permissions, so the old version starts
again (#710). - Deleting the data on uninstall (
-PurgeData,/PURGEDATA=1) also works on German Windows. If
ownership of an item cannot be taken, nothing under the data directory is deleted (#731). - When a reinstall that changes the service account fails, the rollback gives the previous
account its access to the install directory back, so the restored service starts (#735). - The gMSA repair command suggested by the setup check adds this host to the existing principals
instead of replacing them. The deploy README no longer suggests delegation on every target
(#713). - A manual update with
Update-NodePilot.ps1keeps the server setup's uninstaller and its
deploy\scripts, so "Apps & features" can still uninstall an installation updated that way
(#622). - The Create administrator button of the desktop first-time setup has readable text contrast
and a visible keyboard focus ring (#817).
- A server uninstall whose cleanup fails stops before the "Apps & features" entry is removed,
- Monitoring
- The observability summary queries the counter names the exporter actually publishes, so
executions per minute, success rate and failures show values again. The Grafana CPU panels use
dotnet_process_cpu_count(#736). - The Grafana reliability ranking also lists workflows that never succeeded (#737).
- The dashboard, Live-Ops and the execution pages update live again instead of waiting for their
polling interval, and catch up after a reconnect (#706).
- The observability summary queries the counter names the exporter actually publishes, so
- Examples, test suite and tooling
- CLI tests now verify write requests and their payloads. Telemetry
tests verify actual log delivery over HTTP and gRPC, including collector headers and disabled
export settings. - The Windows Update health example applies
freshnessDays(#734). The temp-file cleanup
example writes a new archive per run and keeps the earlier ones (#723). The README showcase
reports a failed probe instead of success (#724). The tech demo uses one name for its child
workflow (#722). The SCCM/AD example of the browser demo is valid PowerShell again, and CI
parses every demo fixture script (#698). - The
powerManagementtest moves to the invasive set and needs a disposable host, because
abortcancels a pending host restart (#720). Acknowledgement files accept only GUID
correlation IDs (#721). Verification checks every run in the window and fails when required
workflows are missing (#718, #719). A first install already enables the two webhook
workflows, and the release check verifies the integration workflows it installs (#776). - Nightly retention deletes only the run folders it created (#716). The failover drill restarts
the original leader also after a failed takeover (#715). - The failover drill recognises a healthy leader again. Its HTTP 200 answer has no
reason, and
reading that field stopped the drill before the takeover (#624). - The SSO lab drills check the requested run and use the configured account prefix throughout
(#717, #645). Their database checks for refused logins, deprovisioning and JIT creation can
pass again, because they match the audit columns the server actually fills (#775). - CI classifies changed files through
scripts/classify-changes.sh, also for long file lists
(#729). - The release lab also runs the failure paths: a reinstall with another service account that
must roll back, an uninstall that must stop and stay repeatable on server and desktop, and a
failed staged desktop update. The account change needs a second lab gMSA in
server.secondGmsaAccount.
- CLI tests now verify write requests and their payloads. Telemetry
Security
- Copying nodes to the clipboard also masks recognised secrets in edge conditions and in lists, not
only in node fields (#574). - Script generation sends the current script only to the AI profile the approval was given for. If
the profile, its endpoint or model changed in between, the request fails with
409 LLM_RECIPIENT_CHANGEDand the dialog asks again. The workflow assistant also treats a model
change as a new recipient for its chat history (#688). - A write query in the database viewer runs only after the confirmation phrase is typed; pressing
Run again while the confirmation is open no longer starts it (#691). - The CLI, the MCP server and the Python admin scripts under
scripts/no longer follow redirects
with the login token, so the token is never sent to another host (#711). - Logging out with the CLI also ends the MCP server's session right away (#712).
- The server setup accepts only the configured certificate, by thumbprint, for its health probe and
the first login, does not follow redirects there and no longer turns off certificate validation
for the whole process. Update-NodePilot.ps1does the same for its health probe: it reads the thumbprint from the
installedKestrel:Httpsconfiguration, so nothing new has to be passed, and no longer turns off
certificate validation for the whole process. If that configuration cannot be read, the update
warns and checks the one probe request without the thumbprint.- The desktop package bundles .NET 10.0.14 for the API,
npandnodepilot-mcp; 10.0.12 and
earlier are affected by the Kestrel advisories CVE-2026-88092 and CVE-2026-70188. The build
refuses a stage whose framework declaration or runtime binaries are older. The runtime is part of
the package, so existing desktop installations get it only by installing the new package (#787). - The desktop build extracts Electron with
@electron-internal/extract-zip1.0.8, which carries the
fixes for RUSTSEC-2026-0367, -0368 and -0369 (#773). - The desktop setup hands the first-login token to the account that started setup, also under
Remote Desktop or a switched user, and to the user signed in at the console when software
distribution runs setup as SYSTEM. If the starting account cannot be determined, provisioning stops
instead of handing the token to another profile (#732). - The Switcher checks the SCOrch connection in both switch directions and rejects plain HTTP to a
remote SCOrch, because it signs in there with Windows credentials (#728). - MCP alerting output shows webhook targets only as scheme, host and port, because the rest of the
URL can contain tokens (#730). - GitHub tokens in step output are masked except for their prefix such as
ghp_; before, the whole
token stayed visible (#681). - Log steps and returned data mask private keys before long values are shortened, also when the end
marker is missing (#682). - An OIDC login with an older group list no longer overrides a newer SCIM change, such as a revoked
group. A token issued before the update is rejected; signing in again with a fresh token works. minimatchmoves to 11.0.4 in all three npm trees (GHSA-g3hv-75wg-m249, denial of service through
glob patterns). In the UI and the documentation site it is only used by ESLint; in the desktop project
only desktop packaging uses it, through@electron/asar, which now getsglob14 as well. None of it
ships in a bundle; the CI audit step rejects it all the same.
Downloads and verification
| File | Purpose |
|---|---|
NodePilot-Server-Setup-1.5.4.exe |
GUI installer for the server |
NodePilot-Desktop-Setup-1.5.4.exe |
desktop installer |
NodePilot-1.5.4.zip + .manifest.json + .p7s |
server payload with detached signed manifest |
NodePilot-Deploy-Scripts-1.5.4.zip |
install scripts, verified before the payload is extracted |
NodePilot-Switcher-1.5.4-win-x64.zip |
standalone Engine Switcher |
nodepilot-release-signing.cer |
publisher certificate |
NodePilot-1.5.4.SHA256SUMS.txt |
SHA-256 of every file above |
Release signing certificate thumbprint (SHA-1): 277EAB317A581C88302CE92BE805938C86B4650D
The certificate is self-signed, so Get-AuthenticodeSignature reports UnknownError. Check that the
signer is CN=NodePilot Release Signing and that its thumbprint equals the one above. Windows
SmartScreen warns on first run because the publisher has no reputation yet.
The server setup bundles .NET runtime 10.0.14 and the PostgreSQL 16 client; the desktop setup
bundles PostgreSQL 16.6.
Testing
- Backend: full suite, elevated: Engine 3446, Api 3284, Ai 849, CLI 671, Data 345, MCP 247 and
Switcher 175 passed, none failed (25 tests are skipped in the test code).dotnet list package --vulnerable --include-transitiveis empty. - Frontend: lint, 3848 unit tests, 491 end-to-end tests (21 skipped by the specs); docs site lint,
227 tests and build; desktop 126 tests.npm auditreports no vulnerabilities in the three
projects. CI on the release branch is green. It runs one commit after the build commit: a
security advisory for a test-only dependency (Microsoft.AspNetCore.WebUtilitiesthrough
WireMock.Net, not part of any product project or of the payload) was published after the build, and
the pin to 2.3.14 is the only difference. - Release lab, server setup on the signed 1.5.4 artifacts: 32/32 scenarios passed, including 20
direct updates from 1.5.3, 1.5.2, 1.5.1, 1.5.0 and 1.4.4 (gMSA and LocalSystem against SQL Server
and PostgreSQL), fresh installs, a second installation over an existing one, a remote SQL Server,
a restart, a database outage during start-up, the CRL pre-flight refusal and AllSigned execution
policy. - Release lab, desktop setup: all phases passed (13/13, 103/103 and five direct updates from the same
five releases with 23/23 checks each), including fresh install, restart, over-install with the
shell open, a failed staged update that rolls back, uninstall and reinstall variants and a blocked
uninstall. - Workflow test suite on the long-running lab instance after updating it to 1.5.4: with the
continuousprofile pass=32, fail=0; withcontinuousandintegration37 workflows passed. The
other three integration workflows (email notification, SQL with a named connection, AI agents) are
disabled on this instance because it has no SMTP sink, no PostgreSQL connection and no LLM
endpoint, and the verifier reports them as failed for that reason. They were skipped in earlier
releases for the same reason. Invasive workflows are disabled by profile and triggers are judged
from their cadence. - The lab found one defect before the tag: after a failed desktop update, the rollback restored the
API service without its delayed start until the next reboot. It is fixed, and the backend suite,
both lab matrices and the workflow suite ran again on the final build; the frontend checks ran on
it as well, since the fix touches no frontend code.