Skip to content

Week 38 · 14–20 September 2026 ​

The platform can now answer "what did this request do". Every event, every refusal and every consumer delivery records the call it happened in, and the dashboard can open everything one call recorded from any event in it. It also lists the deliveries that gave up across a whole app, not only one record's. Several fixes this week tighten security, and some of them change behaviour. On the dashboard, a viewer can no longer promote, delete or preview a vertical. Module code can no longer enqueue the platform's own sweep bookkeeping. An erasure now also reaches the second copy of an event that the spine kept for connector deliveries. The kernel gained two surfaces that verticals have been asking for: a scope can push change notifications to the people allowed to see them, and a long import can stop and resume from where it left off. ticket0 now takes email: inbound mail becomes a conversation, a reply joins its thread, and a desk can block a sender and clear out conversations nobody picked up.

Every call leaves a trail you can follow ​

Until now the spine could say which operation emitted an event and which event caused another, but not that two events came from the same call. Now the platform mints an id per invocation and stamps it on every event that call emits, including the events its in-scope consumers emit in the same tail. It writes the same id into the invocation log line, so the events say what happened and the log line says how long the call took and how it ended. readHistory and the facet reads carry it, and invocation is a facet dimension, so "which calls did the most" is a query you can run.

Two more records carry the id. A denial records the call it happened in, because a refusal emits nothing and so used to be the one thing a request did that nothing could tie back to it. A delivery records the call that made its latest attempt. For an executor that retries, that is a different call from the one that emitted the event, and the dead-letter list used to imply they were the same. Rows written before these columns existed keep a null, which means unrecorded.

On the dashboard, every event in a record's history has a third button beside Why? and What did it do?: Same call. It lists everything the request recorded, oldest first, including siblings that neither causal walk can reach, such as two events one operation raised side by side. The kernel read behind it is readInvocation. An app's Flow view gained a Deliveries that gave up panel. It lists every consumer delivery in the app that exhausted its retries, newest first, paged, with the error. Until now a dead letter was visible only to someone who already knew which record to open. The kernel read is readDeadLetters.

All of this appears for versions pushed with a current release. Older rows and older versions say they have no id rather than showing an empty trail.

Security fixes, three of which change behaviour ​

Three things change behaviour.

  • A dashboard viewer can no longer change a deployment. The dashboard's deployment write routes (promote or roll back production, delete a vertical, create or remove a preview, bind a domain to one) checked that the vertical was the team's and never checked the caller's role, so a viewer could do all of them. They now require the same permission as provisioning an app. Owners, admins and members keep what they had.
  • Module code cannot enqueue sweep-runs. ctx.requestPlatform accepted any kind, and sweep-runs is the kind the platform's scope sweeper uses to record what it did. A vertical could enqueue one for its own scope and turn a stale schedule or freshness expectation green. That kind is now refused with a forbidden error. Every other kind, model-usage included, is accepted as before.
  • An erasure reaches the connector copy of an event. On a host without a control plane, which is every hosted vertical, each connector delivery is held as a platform request carrying the whole event. shredSubject redacted the outbox and stopped there, so the subject's name survived in those rows and in every export and backup taken afterwards. Those payloads, and the provider error message beside them, are now replaced with a redaction marker. An intent still pending is settled as failed, because it can no longer be delivered.

Three more fixes. Restoring an archived scope now refreshes its copy of the tenant's permissions before the scope goes active. Before, someone whose access was revoked while the scope sat archived kept it on the restored scope until the next tenant-level change. The dashboard's connection to the control plane is now limited to one tenant by the credential it holds, and the control plane refuses a request outside that tenant. Before, the dashboard made that promise itself on every call. The builder studio's file routes are now confined to the current project. Reads could reach the whole checkout, .dev.vars included, and a symlink could step outside it in either direction.

The repository has a SECURITY.md with a private reporting channel, a weekly dependency audit, and Dependabot. The shared hono version moved to 4.13.8, which clears seven advisories against it.

Live reads: a scope can announce its own changes ​

A scope can now hand out a subscription to its own changes, and it announces a change only to subscribers allowed to read the entity that changed. A module declares which of its entity types can be watched, and with which read key:

ts
liveTargets: [{ entityType: 'conversation', readPermission: 'conversation:read' }]

An entity type with no entry reaches nobody, so a manifest written before this announces nothing, as before. A notification says only that something changed, never what it changed to. The client re-reads through the ordinary operation, so nothing bypasses the declared read surface or its permission checks. Notifications go out after the transaction commits and after in-scope consumers run, so nothing a rollback could take back is announced.

This is the platform half, and it runs on the Cloudflare host only: it needs something that outlives a request, which is the scope's own Durable Object. The pure SQLite host declares the surface as never, so code that assumes it exists fails to compile rather than silently doing nothing.

Long work that resumes where it stopped ​

The kernel gained a fourth way to move work off a request, for jobs that do not fit the other three. Walking a hundred thousand objects in an external system takes an hour, and it has to survive an eviction, a deploy and a failure along the way. A run is a durable record in the scope with a resume cursor. Each pass does a bounded chunk and hands the cursor on, and the named steps inside a pass are committed as they finish, so a step that succeeded is not run again after a retry or a crash. Asking for a second run while one is in flight joins the first. startJobRun starts one. Step names must be a pure function of the payload and earlier results, and reusing a name inside one pass is refused, because the second call would skip its own work.

Engines publish their event maps ​

Every engine now exports a typed map of the events it emits, for example ProtocolEvents and WorkorderEvents, and every emit site in the engine goes through it. A consumer written against the map gets typed payloads. If it handles one half of a completion group and forgets the other, it does not compile:

ts
consumers: consumersFor<[ProtocolEvents]>()({ 'protocol.signed': …, 'protocol.countersigned': … })

The typed consumer seam has existed in the kernel for some time, but no engine implemented it. All seven engines ship a minor release this week for it. Nothing existing changes.

Refusals carry their own error codes ​

Five kinds of refusal (an unknown vertical, an unknown version, a push that breaks the sandbox contract, a rejected version, and a vertical that still backs installs) now carry their error code from where they are thrown. Before, the code was guessed by matching the message text. The response you get is the same 404, 403 or 409, with the same sentence, with one exception that was a bug. Deleting a vertical whose only remaining installs are archived used to answer a bare 500, because the message pattern matched the live-installs sentence and missed the archived one. It now answers 409 and tells you to reap or restore them.

The dashboard: a status band, a time cursor, and bound installs ​

  • An app's Overview opens with a status band: health across its schedules and freshness expectations, the 24-hour error rate beside its request count, the running version, and the owner seat. Each tile links to the page that explains it. Below it is a small sparkline of the last day.
  • The Observability chart marks what explains its shape: deploys and go-lives, applied migrations, failed schedule runs, recorded failures, and shaded spans where a freshness expectation was stale. Click a marker to open the view that explains it.
  • The chart's time axis is a cursor. Click a bar and Logs and Events narrow to that bucket. Click a failure marker and they narrow to five minutes either side of it. The window is in the URL, so a shared link opens on the minute that matters.
  • A connection says when it was last used. A live connection with no recorded use in the retained window reads unused 14d on the Integrations card and is drawn dashed on the Flow map. A lapsed one is not also called idle, and a record that could not be read is shown as unknown rather than as idle.
  • A vertical's page lists the installs it still backs, which are the ones the "still backs N scope(s)" refusal counts. You can Move them to another vertical your team owns, or Retire them. Retire needs you to type the number of installs, takes a backup first, and stops at the first failure.
  • A hook your vertical does not declare no longer shows as an error. An owner-seat or configure hook the running version lacks answers 501, and the log view showed each one as a red error row on every page load. It now reads → 501 capability absent at info level, and the dashboard asks only once per running version.
  • The left menu is grouped into Build, Operate and Configure, and every page resolves the signed-in person in one directory read instead of two per team. That chain was why most dashboard routes took three to four seconds.

ticket0: email in, threads, a blocklist and a reaper ​

  • Inbound email works. Mail to the desk's support address arrives through Resend's inbound webhook, the signature is checked, the message is re-read from Resend by id, and it becomes a conversation. The route stays switched off until the desk has both RESEND_WEBHOOK_SECRET and RESEND_API_KEY. It has not yet been run against a live inbox.
  • A reply joins its thread on In-Reply-To, but only if the sender's address belongs to that conversation's contact. Someone else quoting the same message id gets a new conversation. A reply to a closed thread opens a follow-up, as it already did for the widget.
  • A desk can block a sender by address, by domain (sub-domains included) or by contact, at both public doors. The check runs before anything is written or any model call is made. Blocked mail is acknowledged and dropped, so the provider stops retrying.
  • Abandoned conversations are closed. An hourly schedule closes conversations that are still new (no reply, no assignee) and have been silent for thirty days. Silence is measured from the last message, not from when the conversation opened. A desk can set its own window, from one day to ten years, in Desk settings.
  • A colleague can be put on one conversation with a revocable per-record grant: ctx.grant(principal, 'conversation:read', conversationRef(id)). For now that gives a direct-link read only; the follower does not see the thread in the inbox.
  • The desk's own assistant is now refused as an assignee by the API, not only hidden from the picker.

The external example signs in for real ​

The example vertical built only from published packages no longer names its caller with a header. It signs in through @substrat-run/dev-issuer, the same way npm create substrat does, answers a denial with 403 instead of 400, and fails closed when it is deployed without an issuer. A weekly and post-release job now installs, typechecks and dry-run deploys it from the registry, so it cannot stop working again without anyone noticing. The scaffold's worker now flags platform requests on its responses by default, so an intent it raises during a request is drained within seconds instead of waiting for the fifteen-minute sweep.

For hosted operators ​

  • Events drain to a lake. The control plane binds a Pipelines stream and ships every scope's events to an Iceberg table in R2. Batches are sized by UTF-8 bytes, each row carries its size for per-tenant volume, and the drain stamps rows only after the stream confirms. If the stream is missing, it stamps nothing, so rows catch up once the binding is fixed. A dropped lake table can be rebuilt without losing history, using pnpm lake:redrain (--dry-run reports how many rows it would reopen). Provisioning runs as your own wrangler login rather than a stored admin token. An erasure does not reach an event that drained before its subject was erased. That limit is now documented.
  • The committed Wrangler configs name no account-specific id. Database and stream ids come from the environment and are substituted at deploy time, and a deploy refuses a config with an unresolved placeholder. A stale stream id now fails before the deploy starts, with a message naming the variable.
  • A failure digest by email: set STAFF_ALERT_EMAIL and each control-plane sweep sends one message listing the failures recorded since the last pass. Unset, nothing is sent.
  • OIDC_REQUIRE_EMAIL_VERIFIED, off by default, makes the staff roster, dashboard invite acceptance, the roster repair step and the support-desk identity accept an address only when the issuer says it is verified.
  • A scope read that a hosted vertical serves now leaves an access-log row naming what was read, the same row a co-located read leaves.
  • A schedule and a freshness expectation are now told apart by a column rather than a name prefix. A schedule named freshness:… used to share a row with the evaluator and read its verdicts as its own run times.
  • A hosted issuer upgraded from Better Auth 1.6 could not be signed into again. Its signing key's alg read back as the literal text "alg". The missing columns are now added on boot, and no key rotation is needed.
  • A platform-held social sign-in relay for Google, GitHub and Apple exists as a worker. An issuer would federate to it as an ordinary OIDC upstream and never hold the provider credential. Nothing uses it yet.

Also ​

  • CLI: substrat push and push --check print whether the push carries an entity model, as entity model: 4 entities (model.json) or a note saying there is none. Before, you found out from an empty Model tab after the deploy.
  • Builder studio: a Model tab renders the project's model.json with the same viewer the dashboard uses, and the model is emitted after every turn.
  • Callout mounts the shared invite routes instead of its own copy, so a body that fails to parse gets a 400 naming the problem.
  • Docs: the book has a second edition, with new chapters on observability, the audit trail and the lake, and metering and billing. It is offered as an ebook from the landing page instead of the top nav. A new figure shows how committed work leaves a scope: the kick, the sweep and the drain. Reference pages caught up with the event readers, the sweep phases, requestConnectUrl, freshness, the refused SUBSTRAT_ binding prefix, emailVerified, federated logout, the scaffold's file list, the console's Operations views, and the dashboard's four tabs.

Released ​

PackageSpan
@substrat-run/contracts0.112.0 → 0.116.0
@substrat-run/kernel0.112.0 → 0.116.0
@substrat-run/adapter-sqlite0.112.0 → 0.116.0
@substrat-run/adapter-cloudflare0.112.0 → 0.116.0
@substrat-run/vertical-host0.112.0 → 0.116.0
@substrat-run/vertical-auth0.14.1 → 0.14.5
@substrat-run/control-plane-api0.112.0 → 0.116.0
@substrat-run/contract-tests0.112.0 → 0.116.0
@substrat-run/model-emit0.9.4 → 0.9.8
@substrat-run/model-view0.2.13 → 0.2.17
@substrat-run/model-providers0.5.2 → 0.5.6
@substrat-run/psl0.2.5
@substrat-run/dev-issuer0.1.27 → 0.1.31
@substrat-run/engine-workorder0.11.11 → 0.12.0
@substrat-run/engine-booking0.7.19 → 0.8.0
@substrat-run/engine-invoicing0.10.2 → 0.11.0
@substrat-run/engine-protocol0.12.17 → 0.13.0
@substrat-run/engine-invites0.7.11 → 0.8.0
@substrat-run/engine-absence0.5.24 → 0.6.0
@substrat-run/engine-metering0.5.18 → 0.6.0
@substrat-run/connector-scrive0.14.17 → 0.14.21
@substrat-run/connector-fortnox0.4.14 → 0.4.18
@substrat-run/connector-planima0.2.9 → 0.2.13
@substrat-run/cli0.32.2 → 0.33.0
create-substrat0.9.7 → 0.9.8

The seven engine minors are the published event maps, and the CLI's minor is the entity-model line on push.