Release notes

Keep up with latest product news

4950 notes
18 Nov, 2026 Joshua Brandt
OcuTalon
Version 1.0.13

Slack, on your own install

  • You can now talk to OcuTalon in Slack. Create a Slack app for your workspace, put its bot token and signing secret in .env, and point its Event Subscriptions at https://<your public address>/ocutalon/hook/slack/events. The guide's Slack section gives the steps. One install serves one Slack workspace.
  • An admin links each person's Slack member ID to their OcuTalon user in Settings > Channels > Slack. Only a linked member's direct messages to the bot reach OcuTalon, and a linked member who may not use the chat is refused. OcuTalon's answer comes back as a Slack message in that direct message, and the web chat marks the turn "via Slack".
  • Notices, pages and approval requests don't go to Slack; they stay on the web chat, the phone page, Telegram and email.
  • Slack replies were tested against a Slack API mock; they haven't yet been tried in a real Slack workspace.
  • Slack is off on the online demo.

On-call pages reach the backup, on your own install

  • Choose an on-call person and a backup in Settings > On call. When OcuTalon finds failures in a production it watches, it pages the on-call person in their OcuTalon chat and by Telegram or email, as set for them. Email pages go out from the mailbox set up under Channels.
  • The page in the web chat has a Got it button. If nobody presses Got it within the minutes set in Settings > On call, OcuTalon pages the backup, and Settings > On call says so. A practice Break the feed pages nobody.
  • The online demo never pages anyone.

Memories that aren't used fade

  • A saved memory that hasn't been used in an answer for 90 days now ranks below a fresh memory that matches as well. It is never dropped: when nothing else matches, it still comes back. A pinned memory never fades.

Triage

  • A triage reply that left out its confidence number and chose to tell you now or act is now used, instead of being replaced by a Tell now notice with no reasons; a reply that would ignore or digest an item without one is still not trusted.
  • A reply with one over-long piece of evidence is now kept, with that piece shortened, instead of the whole reply being set aside.

Fixed: who can use Telegram and email

  • A user without the chat role, or whose IRIS account is disabled or expired, could still start conversations through Telegram or email; that is now refused, as it is in the web chat. A disabled or expired account could also still approve, reject or press Got it from a paired Telegram chat; that is now refused too.

Fixed: a page's cause in Telegram is redacted

  • The cause in a Telegram page was sent as written. It now goes through the same redaction and length limit as the web chat's page card.

Fixed: Settings saves and their audit entries

  • An audited Settings save that failed after its audit entry was written could leave an entry for a change that never happened. The save now writes the setting first and the audit entry last, in one transaction, and if the audit entry can't be written the save is undone.

Fixed: the security overview's retention text

  • The security overview now states retention correctly: chat turns and the record of each notice sent are removed after 90 days, and a session record only once it is over 90 days old and nothing keeps it; the earlier text said no scheduled purge removed them.

OcuTalon's own MCP server (groundwork), on your own install

  • The install now adds /mcp/ocutalon, an MCP service with two read-only tools, check_status and recent_activity, for users who hold the new role OcuTalonMCPClient. Each call runs as the signed-in IRIS user, reads only that user's own OcuTalon session and writes an MCP audit entry. The security overview lists the tools, the roles and what they can read.
  • Connecting an MCP client to these tools isn't part of this release; it puts only the server side in place. None of it exists on the online demo.

Measured at this release

  • The README's Measured table and its agent-versus-fixed-rules comparison now come from this release's evaluation, with Claude Haiku 5.5 on November 17.
  • The agent got 21 of 24 routine attention decisions, 13 of 13 incidents and 11 of 12 held-out cases right. It sent one item that needed you now to the digest instead, and made 0 unsafe outcomes. The fixed rules alone got 6 of 24, 0 of 13 and 2 of 12. See With the agent and with fixed rules only.

Requires the InterSystems IRIS for Health AI Hub Early Access image, build 154 or build 109. Upgrading from 1.0.12 keeps your conversations, reminders, watches, approval requests, settings, memories, documents and Telegram pairing. The online demo now runs 1.0.13. Slack, the MCP server, paging, files, email, partner warnings by email, setting rollback, OAuth, webhooks and Telegram are for your own install and are off there.


Online demo: https://demo.ocutalon.com · Guide: https://github.com/jbrandtmse/OcuTalon/blob/main/GUIDE.md · Open Exchange: https://openexchange.intersystems.com/package/OcuTalon

17 Nov, 2026 Joshua Brandt
OcuPilot
Version 1.0.17

Security

  • Only an OcuPilot administrator can change agent definitions, including from code run directly in IRIS.
  • A user's proposals, conversations, preferences, screen-context sharing choice, background SQL queries and agent ledger entries can be read or changed only by that user or an OcuPilot administrator, including from code run directly in IRIS. A proposal you confirm is the one you were shown.
  • OcuPilot's install records can be changed only by an OcuPilot administrator, so uninstalling OcuPilot removes only what OcuPilot installed.
  • Changing a screen's access from code run directly in IRIS needs the same rights as the Screen permissions screen, and is recorded in the audit log the same way.
  • Upgrade to 1.0.17.

System Explorer

  • Compare namespaces can now promote. Check the classes and routines to copy, review the plan, and type the other namespace's name to confirm. Each document is written over its copy in the other namespace and compiled.
  • Promote works only between two namespaces of this instance. It never writes into %SYS or a document whose name begins with %, never deletes anything, and stops at the first document that fails. There is no rollback: documents written before a failure stay written.
  • New Unit tests screen. Run the tests of a package, a class or one method, up to 50 classes at a time, and read each recorded run and its methods, with the failure text. Production tests and OcuPilot's own tests are refused, and one run goes at a time.
  • New Command console and Class method console screens. Run one ObjectScript command, or call a class method with arguments, and see the output, the return value and any by-reference values. Output is cut at 32,768 characters.
  • Unit tests and both consoles run code as you, with your rights. They need %Development:USE and write access to the namespace's routines database, and every run asks you to confirm first.
  • Each console run is recorded in the audit log, without the command, the arguments or the output.

Analytics

  • Term lists can be exported, imported and copied. Export writes the classic Term List Manager's CSV to a file on the server or a download. Import and Copy always create a new term list and never replace one.
  • Folder items can be exported, imported and deleted. Import takes OcuPilot's export or the classic Folder Manager's, and refuses a file that holds classes, routines or anything other than folder items, so it never loads code. It creates every item or none.
  • Deleting a folder item asks you to type its name. A locked item cannot be deleted.

Agent

  • The agent can propose a promote, and every term list and folder item change, for you to confirm. Promoting and deleting a folder item start switched off; exporting, importing and copying start switched on.
  • Running unit tests, running a command and calling a class method are for people only. The agent is never offered them, and what they print never reaches it. All three start switched off for everyone; an administrator switches each on under Agent co-pilot → Governance policy.
  • The agent can answer questions about recorded unit-test runs.

Requires InterSystems IRIS or IRIS for Health 2026.2 or later. Upgrading from 1.0.16 keeps your agent definitions, conversations, settings, proposals and audit records.

17 Nov, 2026 FFan Ji
iris-vector-search
Version 1.3.0

IRIS 2026.2 Update
Improved large-load performance — Faster vector embedding generation and storage during bulk data inserts.

Quince
Version 1.0.1

Multilanguage available (English, French and Spanish).

17 Nov, 2026 SShengzhi Wu
PaymentGuard AI
Version 1.1.0

Learning layer — an approved human review now becomes a rule. When an analyst approves a gray-zone payment, the (customer, device) pair behind it is trusted for 30 days: the next payment of that customer on that device is decided by the rule engine (LEARNED_DEVICE_FAST_PATH) with zero LLM calls. Boundaries: scoped per pair, amount cap $250, expires, demoted by any later BLOCK, and Learn() refuses pairs with BLOCKED history. Hard rules (sanctions, blacklists, caps, velocity, duplicate claims) always run first and can never be learned away. Learned data survives demo reseed. Also: Tests/TestLearned (5 cases), suite runner fixed for this build, OPERATIONS note on reseed vs learned data.

17 Nov, 2026 Joshua Brandt
OcuTalon
Version 1.0.12

The setting change behind an incident, on your own install

  • When a setting of a watched production changed in the 24 hours before its failures began, the incident now names the setting, when it changed and how long before the failures, for example "The feed broke four minutes after this setting changed". The setting change is also listed on the incident's evidence panel.
  • When putting the setting back would plausibly fix the incident, OcuTalon offers that as a fix with its own dry run, which shows the setting's value now and the value it would go back to. Nothing changes until you approve. Once it has run, you can ask OcuTalon in the chat to undo it the same way, also from an incident that is already closed.
  • Setting values are shown redacted everywhere OcuTalon shows them: the web card, the evidence panel, the phone page and Telegram.
  • Putting a setting back needs one new permission on your own install, which the security overview names: OcuTalonServiceRole holds %Ens_ProductionConfig:W, only for the rollback and its undo. Setting history and rollback are off on the online demo, where every visitor shares the Orders production.

Partner contacts and partner delay notices, on your own install

  • An admin sets each production's partner contacts in Settings > Partner contacts.
  • Ask OcuTalon in the chat to warn the partner about an incident's delay. It drafts a short notice from the incident's count, times, affected flow and cause, never from a message's contents, and raises it as an approval request that shows the exact text and every recipient. Only an OcuTalon admin can approve it, and approving sends it by email from the mailbox connected in Settings > Channels > Email, only to that production's partner contacts. With no mailbox connected, the approved notice is kept as a draft and nothing is sent.
  • If OcuTalon can't tell whether the mail server took the notice, the card and the chat say "The delay notice may have been sent. Check the mailbox's Sent folder before sending it again.", and the incident's timeline adds the line "{time}: the delay notice may have been sent. Check the mailbox's Sent folder before sending it again."
  • On the online demo the partner contacts stay simulated, and an approved notice goes to the Simulated mailbox, not to a real inbox.

Paging the on-call person, on your own install

  • With an on-call person set in Settings > On call, OcuTalon pages them once, in their paired Telegram chat, at an incident's first notice, with a Got it button. It pages only for failures in the productions it watches, not for a practice Break the feed. It doesn't page by email or escalate to the backup, and Settings > On call says so. Paging is off on the online demo.

Send a file in the chat, on your own install

  • Press the paperclip beside the message box, or drop a file on the chat, to send a PDF, a Word (.docx) file, or a JPEG or PNG photo of up to 10 MB. In Telegram, send it to your paired chat.
  • OcuTalon reads it and offers to add a PDF or Word file to your documents, and it saves a photo's caption to memory. A file it can't read gets a short reply that says so. Sending files is off on the online demo.

A production that stopped is looked into

  • When a whole watched production has stopped, OcuTalon now treats it as a fault and looks into it, instead of only telling you, as it does when one part of a production was disabled on purpose.

Fixed: resend counts

  • Earlier versions could count a resent message as done while it had in fact failed again and Ens still held it, or while a business process that waits for a reply was still handling it. A fix's result could then report more messages resent than really went through. 1.0.12 counts a copy as done only when it is.
  • A suspend or resume that IRIS saved but couldn't apply to the running production said "Nothing was changed". OcuTalon now puts the earlier value back and says "Nothing was changed" only when that is true. If it can't put it back, it says the change was saved but not applied.

Fixed: signed URLs are masked

  • OcuTalon's redactor now also masks the signature parts of signed URLs, such as sig= and X-Amz-Signature=, keeping the parameter's name. A signed URL in a setting no longer shows its signature in what OcuTalon shows, logs or sends to the model.
    Measured at this release
  • The README's Measured table and its agent-versus-fixed-rules comparison now come from this release's evaluation, with Claude Haiku 5.5 on November 16. The incident set now has 13 incidents, up from 11, with the setting-change incidents added.
  • The agent got 22 of 24 routine attention decisions, 12 of 13 incidents and 12 of 12 held-out cases right. It sent two items that needed you now to the digest instead, and made 0 unsafe outcomes. The fixed rules alone got 6 of 24, 0 of 13 and 2 of 12. See With the agent and with fixed rules only.

Requires the InterSystems IRIS for Health AI Hub Early Access image, build 154 or build 109. Upgrading from 1.0.11 keeps your conversations, reminders, watches, approval requests, settings, memories, documents and Telegram pairing. The online demo now runs 1.0.12. Setting history and rollback, partner contacts and partner email, paging, files, MCP servers, sources and email are for your own install and are off there.


Online demo: https://demo.ocutalon.com · Guide: https://github.com/jbrandtmse/OcuTalon/blob/main/GUIDE.md · Open Exchange: https://openexchange.intersystems.com/package/OcuTalon

16 Nov, 2026 OOleg Naryzhnykh
iris-inbox-triage
Version 0.5.8

README: the online demo was closed after the contest, and its link is removed from the app page. No code changes.

17 Nov, 2026 Mikolaj Salata
Shutter
Version 1.5.1

Shutter 1.5.1 fixes what a review of 1.5.0 found. Nothing changes in the gate, its policy or the stored data.

  • A trust's own scenarios: retiring or resuming a scenario is proved on the pattern actually in force; one approver can suspend a scenario while a change to it waits; the approval dialog shows the proof the approval stands on.
  • Set up the site: the check on typed answers catches typographic dashes and invisible characters, and an overlong receiver name is refused before it can tie up the server.
  • A draft handed over from Set up the site keeps its reviewed answers when the page refreshes.
  • This site's guide prints each rule as Shutter actually applies it at your site, and masks long numbers.
  • Tested: the full CI on GitHub against IRIS, the upgrade from 1.5.0, the web suite and the full browser suite.

Full change log: https://github.com/MxSalata/shutter/blob/main/CHANGELOG.md

16 Nov, 2026 Joshua Brandt
OcuPilot
Version 1.0.16

OS management

  • Three new Backup screens: Backup tasks, Backup history and Backup database list. Backup tasks lists each task with its last run and status, and shows a banner while database writes are frozen. Backup history lists the 50 most recent backups, newest first.
  • Run a backup task from its row. It runs in the background, and its result appears in Backup history. A run is refused while another backup is running or while writes are frozen.
  • Freeze writes for an external backup, then Thaw writes when it is done. A freeze always takes a suspend limit from 1 to 600 seconds, and writes resume by themselves when it passes. A freeze is refused if writes are already frozen, or if the instance cannot enforce the limit.
  • Add a database to the backup database list, or remove one. A database already on the list, one the instance does not have, and IRISTEMP are refused.
  • As in the classic portal, the Backup screens need write access to the IRISSYS database.
  • Seven new settings screens: Startup settings, Memory settings, Compatibility settings, SQL settings, SQL Shell settings, TSQL settings and ISQL settings. Each lists the instance's settings with their values. Change one with Set value, whose dialog says what the setting accepts. The memory and startup settings from the classic Memory and Startup page are rows here.
  • A setting that takes effect only at the next restart says so before and after you save it. OcuPilot never restarts the instance.
  • Changing one of five SQL settings clears the cached queries in every namespace, as the classic page does. The dialog says so before you save.
  • Some settings are shown but never changed here. The default superserver port and the web server settings are read-only, because a change could cut OcuPilot off at the next restart. Settings that hold a file path, the SQL Shell's escape condition (it runs as code), the default schema and the SQL settings the classic page does not offer are read-only too.
  • On an instance with no TSQL dialect set, the first change to a TSQL option also sets the dialect to Sybase, as the classic page does.
  • The agent can propose any of these changes for you to confirm. Running a backup, freezing and thawing writes, and adding or removing a database on the backup database list start switched off. The seven settings changes start switched on. Change any of them under Agent co-pilot → Governance policy.
  • New read-only Mirror monitor screen. It lists each mirror member with its role, status, journal transfer latency and dejournal latency. On an instance that is not a mirror member it says so, and also says when the license does not allow mirroring. Like the Backup screens, it needs write access to the IRISSYS database.
  • OcuPilot refuses to remove a mirror together with taking its databases out of mirroring, and the Guardrails page lists this refusal. OcuPilot does not change mirroring.
  • Two new screens, System DDL mappings and User DDL mappings, list the datatype class each DDL type maps to. Set value changes a mapping's datatype, and Delete removes a mapping after you type its name. New mapping adds a user mapping. System mappings cannot be added, as in the classic portal.
  • Refused, with the reason shown: a name the instance already has, and a name or datatype in a shape the instance does not accept.
  • The agent can propose these mapping changes for you to confirm. Deleting a system or user DDL mapping starts switched off; adding and changing mappings start switched on.

System Explorer

  • Two new read-only screens. Code metrics counts the lines in the classes and routines that match a pattern, such as MyApp.*.cls: code, test code, comments and blank lines, with percentages. Document hashes lists a SHA-256 hash and the last-modified time of each matching class and routine, so you can see whether two instances hold the same code.
  • New read-only screen, Compare namespaces. Choose another namespace of this instance and a pattern, and it lists the classes, routines and include files that are only in this namespace, only in the other one, or whose content differs. Documents that are the same in both are not listed.
  • All three can include generated routines. A pattern of * on its own, or one that matches more than 1,000 documents, is refused with a hint to narrow it to a package.
  • The agent can answer questions from all three screens.

Analytics

  • Term lists can now be created, edited and deleted, and their entries added, changed and removed. New term list sets a list's name, caption, description, pattern, subject area, extra fields and Hide value. Edit term list changes these and never touches the entries.
  • Deleting a term list takes all its entries with it, and you type the list's name to confirm. Removing an entry asks you to type its key.
  • Refused, with the reason shown: a name the namespace already has (names ignore case), a key the list already has, and removing a field that some entries still use. A change is also refused when the database that holds the term lists is read-only or not journaled, so that a change is never left half done.
  • Each change is recorded in the audit log. Changing term lists needs the %DeepSee_PortalEdit resource, as the classic Term List Manager does.
  • The agent can propose any of these for you to confirm. Deleting a term list and removing an entry start switched off; the others start switched on.

Interoperability

  • Productions now shows whether each production starts with the instance, its startup priority and its shutdown group.
  • Start with the instance sets a production to start when the instance starts, at a startup priority, or stops it doing so. A namespace starts at most one production with the instance, so this replaces any production set before, and the dialog says so.
  • Shutdown group sets the group a production stops in when the instance stops. The default group is 2.
  • Changing a startup priority or a shutdown group needs read access to the settings of every namespace that runs interoperability, as the classic pages do.
  • The agent can propose both changes for you to confirm. Both start switched on.

Agent

  • The agent can now follow interoperability sessions. It finds sessions by time, status, source, target, message type, session id, super-session or related session, and finds the sessions that use a given class.
  • It reads a session's summary, timeline and message headers, and a message's detail and body. All fourteen of these are read-only.
  • A search by time looks back 24 hours unless you ask for more, and at most 30 days. Reading a message's detail or body also needs the %Ens_MessageContent resource, as viewing message contents does in the classic portal.

Security

  • The agent switches, including the kill switch and read-only mode, the users switched off from the agent, the provider connection settings and the governance policy can now be changed only by an OcuPilot administrator, including from code run directly in IRIS. Upgrade to 1.0.16.

Fixes

  • The Task Manager mail password is now hidden in the audit log, on screen and to the agent. If you set it, change it.
  • After a change to SQL settings, some lists, such as Upcoming tasks and the data servers on ECP settings, could fail to load once and then work when tried again. They now load reliably.
  • A pasted HL7 message keeps characters outside Latin-1 in an item test and a transformation test, instead of turning them into "?".
  • A message body longer than 1,000,000 characters now shows a note that it was cut, instead of losing the rest without a word.
  • Editing a business partner now saves only the fields you changed, so a change made elsewhere after the list loaded is kept.
  • A screen reader now announces when an item test ends.
  • Empty interoperability screens no longer read "Or ask the agent: Ask the agent".

Requires InterSystems IRIS or IRIS for Health 2026.2 or later. Upgrading from 1.0.15 keeps your agent definitions, conversations, settings, proposals and audit records.

16 Nov, 2026 JJoseph John
SynapseHealer
Version 1.0.2

SynapseHealer 1.0.2

Fixes

  • IPM/ZPM install: zpm "install synapse-healer" no longer fails in the Reload phase with ERROR #5021 ("Directory .../src/cls/Synapse does not exist"). The package used to ship src/cls/Synapse as a symbolic link, which ZPM extracted as an empty file. The 23 class files now live directly under src/cls/Synapse, and the package contains no symlinks.

Changes

  • Source layout: classes moved from src/Synapse to src/cls/Synapse, matching the SourcesRoot in module.xml. Docker (iris.script loads the whole src tree) and docker-compose work unchanged.
  • module.xml version bumped to 1.0.2.
  • README links and the IPM badge updated for the new paths and version.

Upgrade
zpm "install synapse-healer"

16 Nov, 2026 Mikolaj Salata
Shutter
Version 1.5.0

Shutter 1.5.0 lets a trust add scenarios of its own, set up its site profile one receiver at a time with Ask Shutter, and print its own site guide.

  • A trust's own scenarios: stop entries a site adds to the catalogue, which send a matching case to a person with the trust's own instructions. They can never allow a resend or a closure, the gate is unchanged, and each entry needs one person to prove it on stored cases and a second to approve it.
  • The scenarios have their own place on the Site profile page: "Make this one of our scenarios" from a case, an editor with a live count of the cases it would catch, Try it and Prove, approval, suspend, retire and resume.
  • Set up the site: a new page that walks through one receiver at a time and shows what Shutter will do there before anything is saved. Where a site switches it on, Ask Shutter can draft a new receiver's description, kind and tier from a person's own words, for review; every policy and the recovery mode stay a person's own answer.
  • This site's guide: Export this site's guide (PDF) prints, from the install's own answers, every receiver, the gate's twelve checks and what Shutter does with every scenario at your site, stamped with the versions in force.
  • Upgrade note: outage resend sets approved before 1.5.0 and not yet sent must be read and approved again.
  • Tested: the full CI on GitHub against IRIS, plus the web and browser suites, the upgrade from 1.4.9 in every install shape, and an overnight soak of CI's IRIS job.

Full change log: https://github.com/MxSalata/shutter/blob/main/CHANGELOG.md

IRIS Chronos
Version 1.1.2

IRIS Chronos 1.1.2 — verified clarification chat correction

Fixes a real browser/server contract mismatch: after native Workflow clarification, the server's informational CLARIFICATION_SOURCE_REQUIRED response could prevent the chat from opening. The interface now accepts that specific error explanation while preserving identity, binding, unknown-error and dispatch checks. No new feature or approval authority is added.

Verified on IRIS AI Hub Community build 154

A clean installation exercised a real HOLD, a native Workflow task and human clarification, a new model decision for B, and three English chat questions. The corrected branch was observed before approval. The browser prepared an exact preview and explicitly confirmed one B recovery, cost 6 within budget 8. The same conversation resumed through READ after switching panels. “Send it again” explained the existing acceptance and did not cause another submission. Desktop and mobile views reopened the same operation, attempt and receipt.
Observed: two live decision calls, three live chat calls, one neutral warmup, one confirmation and one accepted recovery effect. Logout and all four service shutdowns were verified; temporary accounts and the route were disabled. No page/console errors or blocked requests were observed.
466 web tests passed, including nine new regression tests; an 11-check cross-language comparison reproduced the original rejection and verified the narrow correction. The 324 Python tests remain the previously verified 1.1.1 run; Python runtime sources are unchanged.

Downloads and scope

  • iris-chronos-1.1.2-application.zip: complete application, the exact archive checked by this clean run.
  • iris-chronos-1.1.2-native-ipm.tgz: native module only; the same 52 ObjectScript sources, descriptor version 1.1.2. Follow the Compose guide for the full application. Prior actual IPM loading evidence concerns 1.1.0.
  • iris-chronos-1.1.2-verification.zip: sanitized summary, answers, logs and real screenshots. Its summary binds the exact application and manifest hashes.
  • iris-chronos-1.1.2-SHA256SUMS.txt: download checksums.
    All shipments are synthetic. Acceptance does not prove physical delivery. This run used one Windows machine and did not interrupt IRIS; earlier interruption evidence retains its own version and scope. Images, model weights and credentials are not distributed. The 62-second video remains a clearly labelled historical overview. Version 1.1.1 is preserved.
    Download 1.1.2 and its verification report · Jury acceptance checklist
15 Nov, 2026 rreet singh
Finish With Me
Version 1.0.2

Public IRIS preview image quick start (scripts/quickstart.sh), fixed workflow approval checks, Vector Search precedents, new demo video.

15 Nov, 2026 Mikolaj Salata
Shutter
Version 1.4.9

Shutter 1.4.9 keeps a site's own settings over an upgrade, says plainly when an answer never came back, and builds with docker compose up on the contest's build 154.

  • An upgrade keeps the settings a site gave Shutter's production items, such as a pool size or a switched-off item, where it used to put the shipped values back.
  • When a write's answer never reaches the page, the approval and outage dialogs say "No answer came back" instead of reading as a refusal, and a retry is answered with the first record.
  • Settings has Your tasks: one line per task (inspect, investigate live, deliver an escalation, recover) saying whether it is ready now and the next step.
  • A case closed by another scenario run's reset has its own Cases tile, Closed by a reset.
  • docker compose up, the Dockerfile and .env.example default to build 154; on build 109 set SHUTTER_IMAGE. ./install.sh still picks whichever build is loaded.
  • The 1.4.3 fair test can be re-graded by command from its published blind exports.
  • A live run's reservation holds its cost until the run is saved, the FHIR items are tidied, and the Doctor says when their sign-in is refused.
  • Tested: the full CI on GitHub against IRIS, plus the web and browser suites, and an overnight soak of CI's IRIS job on two sandboxes.

Full change log: https://github.com/MxSalata/shutter/blob/main/CHANGELOG.md

14 Nov, 2026 Mikolaj Salata
Shutter
Version 1.4.8

Shutter 1.4.8 keeps a site an install of its own over an upgrade, leads a stale approval to the current plan, and says more of the case in plain words, with the keyboard's focus kept where the page changes.

  • An upgrade with no -Ddemo keeps the install's own demo setting, so a site's install no longer becomes a demo. A first install with no -Ddemo is still the demo, and an explicit -Ddemo=0 or -Ddemo=1 always wins.
  • A review whose plan or case changed under it says what changed and offers Review current plan, which opens a fresh review that needs its own press.
  • A case closed when another scenario run reset the receiver says so, closed by the system, not by a person.
  • The case, the review and the Scenarios page say more in plain words, and keyboard and screen-reader users keep their place as the page changes; three short task recipes are in the README, the install guide and First steps.
  • The fixes from two rounds of model jurors walking the judges' path, and the install documentation findings from the 1.4.5 and 1.4.7 receipts.
  • The fair test's pages say who wrote the held-out cases.
  • On the online demo, the judges' share is 4 USD an hour for the last day of voting.
  • Tested: the full CI on GitHub against IRIS, plus the web and browser suites, a fresh install from the public IPM registry on AI Hub builds 154 and 109, and an overnight soak of CI's IRIS job.

Full change log: https://github.com/MxSalata/shutter/blob/main/CHANGELOG.md

IRIS Chronos
Version 1.1.1

Five audit corrections for the jury installation and review path:

  • Accept the hash-verified IPM descriptor in the complete application installer.
  • Wait for read-only console readiness before the single installation command; retry only the exact temporary startup refusal.
  • Repair line endings in four jury documents and make native Workflow setup mandatory for the recommended HOLD journey.
  • Explain uncertain chat results without promising resolution or retrying a question.

Validation: 9 bundle-integrity tests, 9 startup failure-injection tests and 457 web tests passed. The public console-wait function was exercised on IRIS AI Hub 154 without account changes or model calls. A separate complete local instrumented case demonstrated live decision/chat, exact approval, one process cut, portable recovery and the original UI receipt with one submit and one accepted effect. That full case preceded the final startup and wording additions; they were checked separately. The earlier Python discovery attempt reached its time limit and remains incomplete. A subsequent run passed all 324 distributed Python tests in 24 isolated modules on Windows (402.247 seconds), with zero failures, errors or skips; all 278 tested source files matched release 1.1.1. These local component tests are separate from native IRIS and web-test evidence; the nine bundle and nine startup tests are included in the 324. See the current evidence index for the sanitized results and logs.
The native IPM archive contains the same 52 ObjectScript source classes; the version is 1.1.1. Prior actual IPM loading evidence concerns 1.1.0, not a newly executed 1.1.1 install. The complete application also requires its documented Compose components. All shipment data is synthetic; acceptance does not establish physical delivery. The existing 62-second video is explicitly a historical overview.
GitHub tag: v1.1.1. Source, complete application ZIP, native IPM archive and SHA-256 checksums are versioned separately.

15 Nov, 2026 José Pereira
InteropMCP
Version 1.0.6

AI agent as a production item

  • Agent inside the production: a new custom inbound adapter
    (br.musketeers.InteropMCP.Agent.Adapter) turns an AI HUB agent into a
    production item, published on the MCP endpoint next to the production's own
    capabilities — no bespoke endpoint required.

  • Production-agnostic catalogue: the registry now distinguishes capability
    from agent entries, carries each tool's description into the generated
    method, and supports long-running services (AIGATEWAYASYNC) that report
    their own outcome through GatewayInvocationId.

  • Traceability and identity: direct invocations now write a native
    Ens.MessageHeader request/response pair, and the caller's identity
    propagates through nested agent tool calls so guardrails classify by the
    original agent.

  • Demo: agent REST service, admission policies, sample PDF invoices with
    embedded-Python extraction, and an async batch service as reference
    implementation.

  • Operations: bridge tool refresh disabled (it was answering 401 and dropping
    tool registrations), CSP response timeout raised to 300s, single source of
    truth for the restricted CNPJ list, and non-ObjectScript assets excluded from
    the Docker build context.

  • Docs and tests: new article — the agent running directly in the production —
    operations-guide sections for async services and agent attachment, plus
    ObjectScript unit tests and a Python contract test pinning the published MCP
    tool surface.

15 Nov, 2026 Joshua Brandt
OcuTalon
Version 1.0.11

MCP tools as sources, on your own install

  • An admin can now make a Read-only MCP tool a source in Settings > Tools and MCP servers: give it a name, the tool's arguments as JSON, and how often to check it (every 5, 15 or 30 minutes, every hour, every 6 hours or once a day). Only a Read-only tool can be a source.
  • Each poll is one call through the same tool gate as a call from the chat, with the same audit entry, and shows as its own step in Visual Trace. The poll itself asks no model.
  • The first answer sets a baseline. A later answer that differs from the 50 newest different answers becomes one untrusted item, and OcuTalon makes an attention decision on it. The whole answer is compared, so a change anywhere in it makes a new item, even past the point where a long answer is cut in the item.
  • If the tool is classified Risky or Unclassified, or its description or inputs change, the source stops polling and makes no call, and every admin is told once.
  • Your digest lists each source as checked or not checked, with the reason: the server was unreachable, the check timed out or stopped with an error, the tool is no longer Read-only, or a check didn't run on time. A source that wasn't checked never counts as passed.
  • The worked example now ends with a source: Poll new tickets assigned to you as a source. Sources are off on the online demo.

A tool that changes goes back to risky

  • When a server's tool changes its description or inputs, it is Unclassified again and counts as risky until an admin classifies it again. Settings names the tool and its server above the Tools table.

Fixed: an email's approval request

  • An email to any address but the connected mailbox's own now shows as its own approval request, "Send an email", on the web page, the phone page and in Telegram. It says why it needs an approval, its dry run shows the email's recipients, subject and text, and approving sends exactly that email once. In 1.0.10 the web page labelled this request "Repair and resend".
  • The request keeps the email's recipients, subject and text out of the trace. The trace still keeps the chat's messages as written, so whatever you type in the chat is in it. Email is off on the online demo.

Measured at this release

  • The README's Measured table and its agent-versus-fixed-rules comparison now come from this release's evaluation, with Claude Haiku 5.5 on November 15.
  • The agent got 23 of 24 routine attention decisions, 10 of 11 incidents and 12 of 12 held-out cases right. It sent one item that needed you now to the digest instead, and made 0 unsafe outcomes. The fixed rules alone got 6 of 24, 0 of 11 and 2 of 12. See With the agent and with fixed rules only.
  • The README's test counts are updated: over 600 test classes with over 4,500 tests, and over 100 browser test files.

Requires the InterSystems IRIS for Health AI Hub Early Access image, build 154 or build 109. Upgrading from 1.0.10 keeps your conversations, reminders, watches, approval requests, settings, memories, documents and Telegram pairing. MCP servers, sources and email are for your own install and are off on the online demo, which still runs 1.0.9.


Online demo: https://demo.ocutalon.com · Guide: https://github.com/jbrandtmse/OcuTalon/blob/main/GUIDE.md · Open Exchange: https://openexchange.intersystems.com/package/OcuTalon

IRIS Chronos
Version 1.1.0

Native Workflow clarification, grounded live case chat and clearer jury routes.

  • Ambiguous intent can be clarified through a native human Workflow task; recovery approval remains separate and explicit.
  • The case chat has message bubbles, a multiline composer and a real request-pending indicator. It explains checked evidence without approval or dispatch authority.
  • A model-free rehearsal provides four fixed cases and four exact questions with current IRIS checks and authentication.
  • Compact desktop/mobile layout, visible case context and exact approval values.
  • Reliability fixes for preview preparation, prompt line endings and bounded startup readiness.

Verified on IRIS AI Hub Community 2026.3.0AI.154.0: native clarification and separate approval; fresh UI decision/question with native E1/C2 counts 1/1; original receipt reopened after one native recovery, one submission and one effect. IPM 0.10.10 installed 1.1.0 and compiled all 52 expected source classes. 457 Node UI tests passed.

The strict portable interruption observer required additional cleanup/restart and remains experimental. IPM class installation is separate from the full Compose application. All shipment data is synthetic; acceptance is not physical delivery. The 62-second video shows an earlier interface.

Start the jury review · Evidence and limits · Verified download packages

15 Nov, 2026 rreet singh
Finish With Me
Version 1.0.1

Packaged for IPM/ZPM: module.xml with Interoperability production, Workflow Engine approval loop and Embedded Python agent.

15 Nov, 2026 IInunanba One
Storm Link
Version 1.0.5

1.0.5 — Optional Java PEX handoff

  • New: an optional Java Business Operation built with InterSystems PEX. Storm.JavaHandoff (EnsLib.PEX.BusinessOperation) runs storm.pex.SafeHandoffOperation through the %Java Server gateway and turns each newly saved root cause into a short on-call handoff card, shown in the storm's own Visual Trace session.
  • Off by default. A standard install behaves exactly as 1.0.4; the README section "Java PEX (optional)" shows how to switch it on, what to look for, the recovery step and how to switch it off again: https://github.com/inunanba/storm-link#java-pex-optional
  • The FINAL result is returned before Java is called. In fresh isolated IPM installs on IRIS 2026.3 AI Hub build 154 with Java 11, storms A1 and A2 gave the same results and Managed Alerts with Java PEX off and on, and the native suites passed (12/72, Workflow 9/64): https://github.com/inunanba/storm-link/blob/main/docs/java-pex-1.0.5-ipm-run.txt
  • Java source and build script are in pex/; the shipped jar rebuilds to the same class files.
15 Nov, 2026 Joshua Brandt
OcuTalon
Version 1.0.10

MCP servers in Settings, on your own install

  • An admin can now add an MCP server in Settings > Tools and MCP servers, with an optional secret kept in the IRIS Wallet, and classify each of its tools as Read-only or Risky. Until a tool is classified, it counts as risky. Only admins add or remove a server or classify its tools, and only from the Settings page.
  • Every call of a server's tool goes through OcuTalon's tool gate, and each call that runs writes an audit entry. A risky call waits for a person's approval.
  • The approval request shows the tool, its MCP server and each of its inputs on a line of its own, with secrets masked, instead of one line cut at 200 characters. You can answer it in the web chat, on the phone page or in Telegram. In Telegram, a request too long for one message shows only its start, with no Approve button.
  • Removing a server cancels its pending approval requests.
  • The guide has a worked example, A Jira MCP server, step by step. MCP servers are off on the online demo.

Email from the chat, on your own install

  • Ask OcuTalon in the chat to send an email. All email goes out from the one mailbox connected in Settings > Channels > Email.
  • The admin who connected the mailbox can have OcuTalon email the mailbox's own address without an approval.
  • An email to any other address needs an approval, and only an OcuTalon admin can have OcuTalon email another address. Without the admin role, OcuTalon tells you so and sends nothing.
  • No email's recipients are in the trace. The trace does keep the chat's messages as written, so an address you type in the chat is in it. Mail is off on the online demo.
    Less of your mailbox in the trace
  • Saving the mailbox in Settings > Channels > Email no longer puts its address and user name in the trace, so a user with the trace role can't read them there.

A faster first load

  • Settings, History and Usage now load when you open them. The files a browser downloads before OcuTalon's first page shows went from 991,989 bytes in 1.0.9 to 888,402 bytes.

Upgrades clean up after themselves

  • An upgrade now removes the earlier versions' web files that the new version doesn't use, and the install log names each file it removes.
  • A browser tab left open across an upgrade reloads itself once when you version doesn't use, and the install log names each file it removes.
  • A browser tab left open across an upgrade reloads itself once when you open Settings, History or Usage, instead of failing to open the page.

Measured at this release

  • The README's Measured table and its agent-versus-fixed-rules comparison now come from this release's evaluation, with Claude Haiku 5.5 on November 14.
  • The agent got 22 of 24 routine attention decisions, 11 of 11 incidents and 11 of 12 held-out cases right. It sent one item that needed you now to the digest instead, and made 0 unsafe outcomes. The fixed rules alone got 6 of 24, 0 of 11 and 2 of 12. See With the agent and with fixed rules only.

Requires the InterSystems IRIS for Health AI Hub Early Access image, build 154 or build 109. Upgrading from 1.0.9 keeps your conversations, reminders, watches, approval requests, settings, memories, documents and Telegram pairing. The online demo still runs 1.0.9.


Online demo: https://demo.ocutalon.com · Guide: https://github.com/jbrandtmse/OcuTalon/blob/main/GUIDE.md · Open Exchange: https://openexchange.intersystems.com/package/OcuTalon

14 Nov, 2026 ZZulqarnain Abubakar
Knos
Version 1.0.0

The online demo is offline during voting: its GitHub codespace's free hours ran out on 14 Nov. Watch the 43 s video (https://raw.githubusercontent.com/drexthealpha/Knos/main/iris/media/knos-iris.mp4) or run Knos with one command on your machine (iris/README.md). The package is the same as 0.1.14.

14 Nov, 2026 ZZulqarnain Abubakar
Knos
Version 0.1.14

The online demo is offline during voting: its GitHub codespace's free hours ran out on 14 Nov. Watch the 43 s video (https://raw.githubusercontent.com/drexthealpha/Knos/main/iris/media/knos-iris.mp4) or run Knos with one command on your machine (iris/README.md).

14 Nov, 2026 Mikolaj Salata
Shutter
Version 1.4.7

Shutter 1.4.7 adds a FHIR R4 receiver, proved against InterSystems IRIS for Health's own FHIR server and playable on the online demo as scenario S35.

  • A failed transfer whose receiver is a FHIR server can be repaired as one ward-location update to an existing Encounter, sent with If-Match after a person's approval, and verified from the Encounter's own history. Its own gate needs the server's measured capability, not a declared one.
  • S35 plays it on the online demo in Site policy: the case waits for a person's approval before the one update.
  • Budget: a cap or share below one run's estimate admits a run while nothing is spent; the install says plainly that OpenAI runs are unpriced, so the caps don't limit them.
  • Two unit tests no longer race the running production.

Tested: the full CI on GitHub against IRIS, including the FHIR end to end check, plus the web and browser suites, and a fresh install on the AI Hub build.

Full change log: https://github.com/MxSalata/shutter/blob/main/CHANGELOG.md

13 Nov, 2026 Mikolaj Salata
Shutter
Version 1.4.6

Shutter 1.4.6: each case says what the agents contributed, the online demo shows its pathology work, the first screen leads to a short route, and the site profile can ask Shutter about its key answers.

  • What the agents contributed, on the case: who proposed the plan (the coordinator agent, one agent, the rules or the site's policy), the plan's recorded reason, what it cites as links to the evidence, and how the specialists stood to it. No agent is named for a plan no agent made.
  • Pathology on the online demo: P1 sends a result the results system refuses, and its case goes to a person with no resend; P2 takes the results system down, and one approval has the engine send the three failed results again, unchanged.
  • The short route: the trial's welcome and Home both start S2 in Board mode with Run ready; the tour follows S2, S3 and S8.
  • Ask Shutter about this, on the site profile: beside a receiver's kind and tier and beside enabling recovery, the Explain agent on IRIS AI Hub explains that answer; asking changes nothing.
  • The review on a phone shows what releasing a plan is expected to leave before the confirm button comes into view.
  • Hide resolved on Cases; every evaluation figure names its cohort and release; budget admission counts the next run's estimate in one locked step; OpenAI can be configured as a second provider in the installer (not yet measured).
  • Docs: two guides (How Shutter works, How to integrate Shutter at your trust), a worked example of a proposed local stop rule, and a two-minute video.

Tested: the full CI on GitHub against IRIS, plus the web and browser suites.

Full change log: https://github.com/MxSalata/shutter/blob/main/CHANGELOG.md

14 Nov, 2026 Joshua Brandt
OcuPilot
Version 1.0.15

Interoperability

  • Stop, start or restart one item of the running production from its row on Production items, as the classic per-item Actions tab does. A stop ends the item's jobs until you start it again: the item stays enabled, messages sent to it wait in its queue, and the next production restart ends the stop. Each one updates the production to apply it, after a warning that says so.
  • Copy an item from its row: you name the copy, and it is added disabled, with the source's class and settings.
  • Change an item's class, after you type the item's name to confirm. Settings the new class does not have are dropped.
  • Business Partners can now be added, edited and removed. A partner that an item still names cannot be removed.
  • The agent can propose any of these for you to confirm. Changing an item's class and removing a business partner start switched off; the others start switched on. Change any of them under Agent co-pilot → Governance policy.

Analytics

  • A Cube manager screen lists each cube with its facts, its last build and last update, whether it can be synchronized, and the cubes that depend on it. From it you can build or synchronize a cube, and each one is recorded in OcuPilot's audit.
  • A build empties the cube, and every cube that depends on it, until the build finishes. A cube that is already updating is refused by name, and so is a synchronize of a cube that was never built.
  • The agent can propose a build or a synchronize for you to confirm. Its build action starts switched off and synchronize starts switched on.

Agent

  • Investigate a log entry, an alerts.log row included, from the log hub, the error log, the audit log or the log viewer: choose "Investigate this entry", and the agent works through a guided investigation from that entry, returning a recap and its likely causes, ranked, with the rows it used. Like Explain this entry, it changes nothing.

Fixes

  • Clicking a name in a list always does something useful. On 17 lists, a name used to lead back to the same list; it now opens that entry's own page, or shows as plain text where there is nothing to open.
  • Compiling, deleting and importing in System Explorer, by you or by the agent, now need write access to that namespace's code, as saving and creating already did.
  • On Journal records, Newest first now lists the newest records first, and the order you chose is kept when you come back to the page.
  • A save sent with a malformed or oversized request now gets a clear error instead of a server error.

Requires InterSystems IRIS or IRIS for Health 2026.2 or later. Upgrading from 1.0.14 keeps your agent definitions, conversations, settings, proposals and audit records.

14 Nov, 2026 ZZulqarnain Abubakar
Knos
Version 0.1.13

The demo hours are one window: online without a break until 05:00 UTC on 16 Nov, when voting ends. A Measured table: the AI Hub agent and the fixed rule on 12 held-out cases, two runs, every failure kept. The one-command edition replays the agent's answers recorded in the online demo. In the online demo a visitor answers only the approval task of a sample they sent.The demo hours are one window: online without a break until 05:00 UTC on 16 Nov, when voting ends. A Measured table: the AI Hub agent and the fixed rule on 12 held-out cases, two runs, every failure kept. The one-command edition replays the agent's answers recorded in the online demo. In the online demo a visitor answers only the approval task of a sample they sent.

13 Nov, 2026 Mikolaj Salata
Shutter
Version 1.4.5

Shutter 1.4.5 leads each case with what it proposes and opens its review in one press, lets the Cases list hide what is closed, and says how each scenario run executes.

  • A waiting case's action button opens the review of the exact plan from any tab, named for what it does ("Approve the resend"). The review leads with the plan's action in plain words and keeps the plan digest, the case version and Cancel in view at the confirming press.
  • The Cases list has a Hide closed switch, filtered on the server (GET /cases?open=1), so the total and the pages count the same rows. An empty list fits the install: the trial and the demo offer a scenario, a site says what it watches.
  • The Source message link opens the message in the Management Portal: every link names the message header's class.
  • Scenarios shows each run mode's words in its own card, says how a run executes apart from who proposes, and leads a finished run to its own case. Home starts a short route: S2, then S3, then S8.
  • Provenance in plain words: the trial's runs say "Browser recording: no new inference", and a run is called live only when its own case shows a model call. S8's "every check passed" says that Rules only is exempted there on purpose and that the gate holds the resend on a real case.
  • The site profile's questions line up, and the Orders and results card says what a case gets and how the outage resend works.
  • The repository checks run in their own workflow on every push to main, documentation-only pushes included.
  • The README puts S8 side by side under the opening, has one block for judges, gives each figure its cases, runs and release, and ends with a note on how Shutter was developed. The install guide makes enabling recovery its own step after the conformance and stop tests.

Tested: the full CI on GitHub against IRIS, plus the web and browser suites.

Full change log: https://github.com/MxSalata/shutter/blob/main/CHANGELOG.md

14 Nov, 2026 Joshua Brandt
OcuTalon
Version 1.0.9

The agent and the fixed rules, measured side by side

  • The README now sets OcuTalon's agent beside its own fixed rules, the fallback it uses when it has no model key, on the same recorded evaluation cases: 24 routine items and 11 incidents, plus 12 held-out cases (8 routine items and 4 incidents) whose expected answers were committed before either side ran on them. See With the agent and with fixed rules only.
  • With Claude Haiku 5.5 at this release's evaluation, the agent got 22 of the 24 routine items, 11 of the 11 incidents and 12 of the 12 held-out cases right, and it put one item that needed you now in the digest instead. The fixed rules alone got 6 of 24, 0 of 11 and 2 of 12; they tell you about everything, so they missed none.
  • The README's Measured table now comes from this release's evaluation too.

Routine attention decisions

  • When an item of a watched production is disabled, OcuTalon now decides Tell now instead of Act. In this release's evaluation it got 22 of 24 routine attention decisions right; the README's earlier figure, from release 1.0.4, was 21 of 24.

Sign in to your mailbox with Google or Microsoft 365, on your own install

  • An admin can now connect the mailbox in Settings > Channels > Email by signing in with Google or Microsoft 365, instead of saving a password. OcuTalon then reads the mailbox over POP3 and sends over SMTP with the sign-in, and IRIS renews the sign-in's access token when it expires.
  • If the sign-in is withdrawn at the provider, the next check of the mailbox says "The mail sign-in is no longer valid." and Settings offers the sign-in again.
  • The guide has the steps for both providers: Gmail and Microsoft 365. The online demo has no real mailbox and no sign-in.

Your daily digest by email, on your own install

  • With an outgoing mail server saved, Settings > Daily digest can also send your digest to the connected mailbox's own address. OcuTalon doesn't read its own digest back as a new email.
  • Mail is off on the online demo.

A rejected request leaves the Status section at once

  • When you reject an approval request, it now leaves the Approval requests list in the Status section at once, instead of a second or two later. Try it on the online demo: after the guide's incident, More to try, Reject a fix.

The README names the platform parts

  • The judges' table now names the AI Hub and Interoperability parts OcuTalon is built from: its three %AI.Agent agents, the outbound adapter, the AI Hub classes behind the tool gate and Visual Trace, the RAG knowledge base, the Config Store entries, the EnsLib.Workflow approvals and the resend that a fix uses.

Requires the InterSystems IRIS for Health AI Hub Early Access image, build 154 or build 109. Upgrading from 1.0.8 keeps your conversations, reminders, watches, approval requests, settings, memories, documents and Telegram pairing. The online demo runs 1.0.9.


Online demo: https://demo.ocutalon.com · Guide: https://github.com/jbrandtmse/OcuTalon/blob/main/GUIDE.md · Open Exchange: https://openexchange.intersystems.com/package/OcuTalon