Releases

FlowLakes ships as versioned releases. Each release entry on this page records what changed for admins and builders, any configuration you should review after the upgrade, and compatibility notes. Upgrades are applied by the FlowLakes team; your data, applications, and configuration carry forward automatically.

Release 1.57.0

One address for AI assistants, ready connection snippets, a desk toolset, and fixes to incident routing, create forms, agile forms and privacy (2026-11-04). No columns are added or removed. Incident Management offers an update (1.43.10) in Applications after the upgrade.

What’s new

One address per tenant for AI assistants. Your MCP endpoint is https://<your-tenant>.flowlakes.com/mcp. Connected assistants shows that address in a ready command for Claude Code, a Cursor snippet, and a VS Code snippet that asks for the token once. The older /mcp/mcp and /sse addresses keep working.
A desk token lists twelve tools. A token on the analyst, desk assistant or desk lead profile lists the twelve desk tools by default: create and read tickets, your queue, comments, knowledge search, user lookup and the documentation. The rest stay callable and can be listed on request. See Connect an AI assistant.
Incidents with no category reach the Service Desk. An incident created without a category is routed to the Service Desk before the table’s default team. When two active groups are both named Service Desk, the incident still saves, without a team, and its routing trail says why. On grouped reports and lists, a blank value reads “Uncategorised”.
A cloned incident keeps its caller. A reference field prefilled from another record shows the person that record names, not the person signed in. A package update no longer changes which fields you made required.

Fixes

The Agile create dialog shows a refusal under the field it names, story points take their range from the field, and a release can be created as Released.
Resolving an incident sends the caller one resolve email: the Incident Management rule now declares that it replaces the platform’s Task Resolved notification (package 1.43.10).
A literal such as [p:high] in a read request no longer refuses the request; the privacy check for unresolved placeholders now runs on writes and matches only placeholders FlowLakes itself mints.
The anonymous provisioning status poll answers the status only. Unknown subdomains answer at once.
llms.txt and the site footer name the real MCP address.

Review after the upgrade

Open Applications and update Incident Management to 1.43.10.
On Connected assistants, copy the new /mcp address into your clients when convenient. Existing connections keep working.
A revoked token can keep answering for up to two minutes while the server’s cache expires. Revoke early when a secret may have leaked.
Self-hosted and dedicated stacks: database pool sizes are now settings (TENANT_DB_POOL_SIZE, TENANT_DB_MAX_OVERFLOW and their pooled and core counterparts). The defaults are unchanged; a shape that does not fit the instance’s connection ceiling is refused at start.

Release 1.54.1

Fixes the Incident Management update that stopped on 1.54.0 (2026-11-02). No columns are added or removed.

What’s new

Incident Management can finish its update. The update stopped because category rows still named sla_hours, a column that was removed. Version 1.43.5 no longer sends that column. The deploy applies the update.
A cancel reason stored in plain words is saved as its choice. A value such as “raised in error” is stored as the matching choice when those are the choice’s own words. A value that is not one of the choices is left as it is, and that field stays a text field until those rows fit.

Review after the upgrade

Open Applications. Incident Management should show as installed. If the card still offers Retry after the deploy has finished, use Retry once.

Release 1.54.0

Fixes after 1.53.0 (2026-11-02). One data migration updates two connected-assistant tools so they accept either of the argument names they already documented. No columns are added or removed. Agile, Incident Management, and Request Management offer an update in Applications after the upgrade.

What’s new

A language you enable stays enabled. Adding a language on Settings, such as German, now stays on your site. You still choose it, or set it as the default, before the site renders in that language. A site left on English only stays on English until you add a language.
Settings are read from your own workspace. Sign-in, session timeouts, languages, AI spend limits, and branding follow the values saved for your workspace.
A connected assistant gets a time limit and a plain refusal. A call that cannot finish within 30 seconds stops. Connecting with no credential is refused before it waits. Tool descriptions no longer include internal notes. The connection instructions state the limits: 60 calls a minute per token, and 600 a minute for the workspace.
A record you cannot change looks read-only. Setup stays hidden from people who cannot open the admin console. Approvals in the staff view are shown to people who approve. Access denied is shown as access denied.
A missing field is asked for in your language. Changing a record’s state names the fields that change requires. Cancelling work previews the tasks it would cancel. A refusal names the field, in your language, and does not show an internal table name or a database error.
Categories route to a team. Shipped incident and request categories get a default team. Assigning a person fills their team when the record has none.
Dates, story points, and sprints. A project or sprint whose end is before its start is named on the later date. Story points accept 0 to 100, including halves. A new sprint keeps its number after an upgrade.
A message that cannot be sent is named. A rule with no public address is switched off. A message that matched nobody is recorded. An email a script tried to send before the workspace has an address is held until that address exists.
Lists and forms. My requests shows a loader while it is still loading. The menu for adding a sub-task stays above the Related drawer, and the add row shows priority. The application grid counts the applications it shows. Searching incidents uses the same filter as the list.
A rediscovered device keeps the same configuration item. A failed link does not leave a second one. A proposed plan that nobody decided for 14 days is withdrawn.

Review after the upgrade

If you enabled a language before this release and it did not appear, open Settings and add it again. An English-only list is left as you saved it.
Open Applications and apply the offered updates for Agile, Incident Management, and Request Management. The platform upgrade does not change an application’s installed content by itself.
A large Microsoft Graph change that could not find your workspace’s database stays waiting and is picked up by the regular resume. Nothing is marked finished that did not run.

Release 1.53.0

Launch readiness, round 5: required settings are checked up front, access fails closed, and a field’s rules come from one place (2026-11-01). Two migrations run during the upgrade; one removes a stored password column.

What’s new

A send blocked by missing configuration is held, not failed. If your tenant’s address, or a connection a message depends on, is not set up yet, the message now waits with a named reason instead of being marked failed and discarded; retry it from the Outbox once the setting is fixed and it sends as originally written.
Setup names facts it could not check before. A missing address, an unconfigured embedder, a connection switched on with no working credential, and an application’s categories with no team to route to now show up as findings on Setup, instead of being discovered the first time something needed them.
Turning on a connection that needs a credential checks for one. Enabling Microsoft Entra Graph Query with no credential bound, or one that does not resolve, is now refused on the spot, naming why.
A table with no access rule of its own is no longer open to every reader by default. Seeing every record of a kind now always requires holding that table’s administer grant, the same requirement a table with a declared rule already has.
An approval ending now closes every pending approver, not only the one who acted. A decision, a timeout, a withdrawal, or switching off the policy that raised it all record who closed each pending approver and when, and write the record’s own lifecycle field correctly.
Approvals stay visible only to the people they concern. Listing or reading someone else’s approval by id now checks the same participant, record-read, or oversight rule a person’s own approvals page already applied.
A blueprint component keeps its identity across a rename. A package can now declare a component’s stable key, or name what it was called before, so a later release can reword a rule, a notification, a form, or a dashboard without leaving the old one behind as a duplicate. Removing a component from a package now retires every copy of it, wherever it was installed.
A connected AI assistant gets clearer refusals. Writing a field a table does not have now names the closest real field name it probably meant; a Service Hub submission naming a catalog item that does not exist, or one you cannot see, answers the same generic not-found message either way.
A new /mcp endpoint for AI clients, alongside the existing one, whose session survives a backend restart without dropping your place; a session now also refuses a request from any token other than the one that opened it.
A busy MCP server says so. A call that cannot be served fast enough now answers a plain “busy, try again” with how long to wait, instead of hanging or dropping the connection.
Every MCP access token now has an expiry. A token created before this was required is given one during the upgrade, so a connected assistant keeps working.
The person-name scan now tells configuration apart from work records. A short name-shaped field on a role, a rule, or an application is left alone; every text field of an incident, a request, or a task, short fields included, is still scanned the same as before.
One rule decides who may write a field. Each field now says whether it can be written always, only when the record is created, never, or only by the platform. A field the platform manages, such as an incident’s escalation level, is refused from any form or connected assistant, whatever value is sent.
Text fields have length limits. The short description of an incident, a request and their tasks holds at most 255 characters, and a description at most 20,000. A longer value is refused with the limit named. Records that already hold a longer value keep it and still save.
The upgrade applies a field’s rules to tables you already have. A length limit, a number range or a “must be unique” rule a package adds to an existing field now reaches a tenant installed earlier, not only a new installation. Existing data is never shortened or removed; a duplicate that blocks a new uniqueness rule is named in the upgrade notes and left as it is.
A cancelled request needs a reason. Cancelling a request now asks for a cancel reason, the same as an incident.
Agile work items follow their states. A work item reaches Done only from In Progress, and only with someone assigned. Moving a card on the board follows the same rules as editing the work item.
Shared task fields read the same everywhere. title and summary are accepted in place of the short description on every task table, and priority, impact and urgency on request and change tasks are choices, not free text.
Sprint numbers are filled in. A new sprint gets its number, such as SPR0001, like any other numbered record.
Blueprint packages are checked more strictly. A cancelled state with no cancel reason, a text field whose description lists its options instead of being a choice, and an unknown top-level key are refused when a package is validated.

Fixes

A validation refusal on a record write or a bulk update no longer answers a generic internal error.
Retrying a write with the same idempotency key while the first attempt is still finishing no longer answers an incomplete-result error.
A null expiry on an MCP access token is no longer read as “never expires.”
A tool with no declared permission can no longer read every approval, or describe a table’s secret columns, by accident.

Review after the upgrade

Review the Setup findings this release adds. A finding naming a missing credential, address, or unrouted category describes a gap that was there before, just not named until now.
The Marketing package’s provision request no longer stores an admin password, and the stored values are removed by the upgrade. Rotate any password that was entered there.
A role that read every record of a table only because the table had no access rule now needs that table’s administer grant. If someone lost access they need, grant it through their role.

Release 1.36.0

Notification rules, approval policies, and people and roles in plain words, with the security fixes they were built on (2026-10-20). Seven additive data migrations run during the upgrade. The 1.35.2 patch below ships inside this release.

What’s new

Notification rules read in plain words. Each rule is one sentence with one audience. The list and the rule page live under Automate, and “Try it” shows who a real record would reach before you switch the rule on.
Approval policies read in plain words. The Approvals list groups policies by record in the order they are tried. Move a policy up or down, open one to see its When, If, Who decides and Then, and use “Try it” to see who a record would ask. A policy an application installed is read-only here.
People and roles read in plain words. A person’s page shows their level, the areas they configure, one role per desk and their teams, with “Try it” to check what they can do. Every role reads as one sentence and is grouped by desk.
A blueprint names the role its installer receives. Studio and dashboard sharing no longer create roles behind the scenes.

Security

Message templates render in a sandbox.
Every role grant passes one fence; an expired grant no longer counts; deactivating a person ends their session at once.
AI tools that change approval policies respect the same protection as the pages do.
Creating a table or changing its columns requires the schema capability and is audited.

Review after the upgrade

Shipped notification rules and approval conditions that never fired now do. Check Automate for rules you expected to stay silent.

Patch release 1.35.2 (inside 1.36.0)

The general assistant is off by default, a chat or agent run no longer fails on a desk tool, and the search box finds a page by what it does (2026-10-19). No data migrations.

What changed

No Ask button unless the instance turns Flow on. The general assistant competed with the assistants you already use. Connect yours over MCP from Connected assistants instead: it reaches the same tools, and every change it asks for becomes a plan a person approves. The Service Hub intake chat, the script assistant and triage are not affected.
The search box finds a page by its name or by what it does. Typing “mcp” lists Connected assistants.

Fixes

A chat or an agent run that picked one of the desk tools, such as Search Knowledge or My Queue, failed with “Unknown tool”. The same tools worked over MCP. They now work everywhere.

Release 1.35.1

Rules, workflows and scripts in plain words, a rebuilt record page, a quieter shell, an MCP connector and incident triage in one call (2026-10-19). Fourteen additive data migrations run during the upgrade.

What’s new

Rules and workflows read in plain words. The rules list and each rule page say what a rule does. A deleted rule moves to the Archive with its runs. A workflow has its own page, can wait and run on a schedule, and can hold script steps. A script assistant drafts a script for you.
The Agent and Connections pages read in plain words. Add a connection from a picker.
The record page is rebuilt. A card at the top says what to do next, the conversation reads like a chat, the facts sit on the right, and a History card shows what happened. The Audit is a list you can download. The classic view is one click away.
A quieter look. The Home leads with one headline and one action, the side rail shows icons, text and edges are lighter, and the logo is refined.
MCP connector. An outside MCP client can build apps, and every change it asks for becomes a plan that a person approves.
Incident triage in one step. Incident Management 1.32.0 fills category, team, impact and urgency in one call, using the teams, similar resolved incidents and the matching knowledge article.

Review after the upgrade

Update Incident Management to 1.32.0 from the App Marketplace to get one-step triage. The agent is now called Incident Triage.
Business rule scripts now see number fields as numbers. Check any script that compared a number field to a string.

Fixes

The monthly AI budget counts only this month’s spend (also shipped as 1.34.3).
A response SLA breach is recorded on the task, and an approval shows who asked for it.
Installing an app no longer stops on a table that already exists.
Scripts and conditions stay inside their sandbox, and a rule can name only a script of its own instance.

Patch release 1.34.2

Testing a business rule no longer does anything for real (2026-10-15). Before, Test on a rule could start the approvals, workflows, notifications and tasks the rule names. Now a test shows what the rule would change and what it would start, checks the rule’s conditions against the record first, and keeps nothing. No data migrations.

Review after the upgrade

Look for approvals, tasks or notifications created while someone tested a rule before this release. They were real and may need cancelling.

Patch release 1.34.1

Every instance’s App Marketplace now shows the versions a release brings (2026-10-15). Before, an instance kept listing the app versions it was created with, so it never offered Update. No data migrations.

Review after the upgrade

Open Applications or the App Marketplace: Incident Management and Request Management show Update where an older version is installed. Updating stays your choice; nothing installed changes by itself.

Release 1.34.0

Sub-tasks, filters that reach into a referenced record, the catalog item on one page and a quieter Service Hub Admin (2026-10-15). Two additive data migrations run during the upgrade.

What’s new

Sub-tasks on incidents and requests. A Sub-tasks tab lists them and adds one inline. A sub-task shows its parent.
Filters reach one level into a referenced record, for example the caller’s department, and record lists have a quick search.
A catalog item is edited on one page: its questions, the values every request gets, who can see it, who approves and who handles it, with a live portal preview. See Create a catalog item.
Make a category while you edit an item, and see every category with its item count on the Categories tab.
Service Hub Admin lists items under their categories, the way the portal shows them, with search and Shown and Drafts filters.
The Service Portal is one click from the user menu, and Back to work in the portal header takes staff back.

Review after the upgrade

Incident Management and Request Management show Update in Applications. Update them to get the Sub-tasks tab.
The catalog item wizard and the Forms tab are gone. Open an item from Catalog items; rules and calculations are set through Open the form editor on a saved item.
A category that still holds items can no longer be deleted. Move or remove its items first.

Compatibility

No breaking changes. A submission that posts a field the item does not ask for is now refused by name.

Patch releases 1.30.1 to 1.30.4

Three patches for the self-serve trial and one brand correction (2026-10-12). 1.30.4 redraws the mark as the F from the logo pack and sizes it to the lettering. No data migrations; nothing to do on an existing instance except read the review items.

What’s new

A free trial now runs 30 days. Trials that started earlier keep the end date they were given.
Signing up by email ends with choosing a password. The first time you open your instance, a Set a password dialog opens and stays until you choose one. The ready email says so. See Start a free trial.
The sign-up page shows the build in four steps, and no longer flashes the email form after you click the emailed link.
One ready email, with the logo. The second, plain-text ready email is gone, and the logo now shows in mail clients that block remote images.
The new FlowLakes logo across the application, the public site and every email.

Review after upgrade

Control tenant only: the Applications page shows an Update badge for Tenant & Instance Provisioning 2.13.0. Applying it removes the retired ready-email workflow.
No pricing surface yet: the trial banner stays silent until the last week of a trial, then asks you to write to info@flowlakes.com.

Compatibility

Instances created before 1.30.2 have no AI model rows and their assistant stays silent until the FlowLakes team seeds them; a new instance gets them automatically.

Current release: FlowLakes 1.30.0

FlowLakes 1.30.0 is the release this documentation describes. A new instance installs with Incident Management and Request Management; the rest of the catalog arrives app by app as each one passes its own test round, rather than all at once.

What’s new

Session learning. A process can now learn a commonly used line from verified outcomes and offer it on a later record. A person can accept a proposed line early, dismiss it, or retract an active one, and the record keeps the reason either way. The per-record panel also surfaces up to five other open records raised in the last seven days that share the same service or configuration item, with the total count when there are more. Adding a file to a record while it is waiting on someone now wakes it, the same way a comment does.
Learning controls. Admins can set, per table or company wide from House Rules, when a learned line is proposed and when it switches on, and can choose which fields drive the related records list instead of the platform default.
Trust that is earned, not assumed. An approval policy only counts a plan toward auto approval once a person approved it, the plan finished, and any change it made outside FlowLakes was confirmed afterward. Editing the connection or the policy starts the count again. A plan that fails unattended, or one whose outside change cannot be confirmed, now suspends auto approval for that connection until someone reviews and clears it on the Approval Policies page.
A new operations health view. The Specialists page now shows coverage, verified resolution rate, reopen rate, harmful changes, approval wait time, SLA breaches and cost per verified outcome, over a window you choose. Anything the platform cannot measure yet is shown as not measured, never as a zero.
Move a portfolio in. A new tenant data export and a governed import path let you bring a service portfolio in from ServiceNow or Jira, or take your own data out, without losing a record’s number, timeline or inherited fields. See the migration guide.
Verified means checked, not just said. A record can only be marked resolved and verified once the fix has an actual outside check behind it; a person’s sign off is recorded separately and does not stand in for that check.
Fewer wasted approvals. A plan that would fail a record’s own validation rules is now rejected when it is created, instead of being approved and only then failing.
Incidents can name their fix. An incident can record which change will resolve it.
A default agent on every tenant. Every tenant now has a built in resolver agent. A table’s Session tab turns it on, chooses which specialist takes a record, and can hand a case to a person’s judgment instead.
Agents can ask for a connection. If an agent needs an external system you have not connected, it now files a setup request naming what it needs; the connection owner is notified, and two approvers sign off before it is wired up.
Per-tenant session limits. System Settings gains controls for how many AI sessions can run at once and how many per day; a session over the limit waits and resumes automatically once capacity frees up.
Smarter recovery from a stuck plan. When a plan step fails, the platform now decides automatically whether to retry it, replan, or ask a person, instead of ending the run.
A new agent capability: compute. An agent can write and run a short piece of code over the masked results of a record, or over a whole table as a background job, and can propose a fix to its own configuration for a person to review on the same list used for rule change proposals.
Entra audit trail. The Microsoft Entra directory audit log is now readable, so an access incident can show both the state of an access request and its cause.
Record sessions can read your service catalogue and CMDB. A session deciding on a record can now consult your own services and configuration items, which was blocked by mistake before.

Review after upgrade

The Applications page will show an Update badge for Incident Management 1.25.0, Request Management 1.18.0, Change Management 1.5.0, Agile Project Management 1.6.0, Asset Management 1.4.0, License Management 1.1.0, Entra Access Management 0.11.0, and Entra door 1.2.0.
If any approval policy has auto approval turned on, review it: a change to a group now counts as evidence toward auto approval, which it did not before this release.
Check the new per-tenant session limits under System Settings; they ship with a default that may be lower or higher than what your tenant needs.
If you connected the Entra Access agent through its own trigger rows, those rows are now deactivated in favor of a single selection rule; review the Session tab setting on the affected tables.
If anything outside FlowLakes writes directly into an incident, request or change table through the bulk import loader or the external token API, it will now be refused by name; move it to the governed import path instead.

Compatibility

Editing an approval policy after a plan built under it was approved, but before that plan runs, now blocks the plan at the moment it tries to run; propose it again under the edited policy.
The bulk import loader and the external token API no longer write directly into a table that has tasks (incidents, requests, changes, and the rest); use the governed import path instead, which keeps the record’s number, timeline and inherited fields intact.
The Agent Designer no longer sets which records reach an agent, or its autonomy mode, directly; that now lives on the table’s Session tab as a selection rule.
The trust an agent connection has earned toward auto approval resets to zero on upgrade; plans approved before the upgrade are not counted, and the streak rebuilds as new plans run and are confirmed.
The Entra access agent does not perform a multi-factor reset or a credential reset; those stay with your IT staff.
Analytics queries an agent runs now respect the same read permissions as its ordinary record queries; a query that relied on broader access will be refused.
Known gaps in this release: there are no public docs pages yet for the new compute capability, the automatic retry and replan behavior, or the self-configuration proposals (a docs release will cover them); the Entra blueprint does not grant the compute capability by default, so an administrator has to add it; and the background job that runs code across a whole table has no page of its own yet.

Pages in this documentation carry a min_version marker; everything documented here is available on 1.30 unless the page says otherwise.

What a release entry contains

From the next release onward, each entry lists:

What’s new. Features and improvements, described by what you see and configure.
Review after upgrade. Settings a release adds or changes that an admin should check.
Compatibility. Anything deprecated or changed in behavior, with the migration path.