Week 43 · 19–25 October 2026
The week's theme is what a caller may hold, and what it may learn.
Three strands of it. A permission key can now say, in its own declaration, that only the platform may hold it — and every door that could hand it to a person, an org, a connection or a role refuses, from substrat push down to the scope host at run time. A refusal whose text came from somewhere else — another tenant's deployment, a store, an upstream service — is now answered to you as a sentence and a reference, with the text itself kept where only the platform can read it. And a read addressed to a scope your deployment does not hold is refused outright instead of waking an empty store and answering "nothing here", which is the answer that looks like data.
Beside it, the week's other half is moving bytes without holding them. A connector can take a file from a provider into an attachment, and hand one back out, as a stream: memory stays bounded whatever the file's size, on both the platform's side and the deployment's. The same instinct reached the scope copy that preview create makes — it pages now, so a production-sized scope can be forked at all — and the schedule that sends mail, which went from one mail per occasion to an optional daily digest.
And a quieter third: keeping the record straight. Sixteen more admin writes now commit with their audit row in one transaction rather than beside it, usage is captured day by day so a past day can still be billed, and a vertical's ownership is kept as a history rather than a single current owner.
Attachments stream, so a file's size is not a memory limit
A connector could only move a provider's file into an attachment by holding the whole thing in memory first. A Worker has 128 MB and a Durable Object message has 32 MiB, so a large document simply could not be landed. Both directions now stream.
Going in, every seam on the way takes a stream plus the length its producer declares — ConnectorResponse gains a body, an attachment upload and the blob store take a sized stream, and a streamed connector upload travels on its own route with the metadata in a header rather than a multipart part. The length is declared rather than counted, because object storage needs it before the first byte, and it is enforced as the bytes pass: a stream that carries more or fewer bytes than it declared lands nothing. The hash and the size on the record come from what was actually written, not from what was promised.
Coming out, openAttachment streams at both ends — the deployment reads it out as a stream, and the platform relays it as one — so a connector that sends a sealed document to a signing provider or a document to a Teams folder never holds it either. Microsoft 365 and Scrive now send their attachments that way.
What is bounded and what is not, stated plainly: memory is bounded whatever the size. Time is not. The transfer runs under the connection's own timeout, so a file that cannot finish inside it lands nothing. Size has its own ceilings now too: the multipart upload route and the buffered client path are bounded before anything is read, a write permission is checked before and while an upload streams, and a streamed response body times out on idleness rather than on total transfer time — a slow 500 MB download is no longer the same thing as a stalled one.
A deployment that predates the streamed routes answers with a sentence telling you to re-push it, and nothing is stored. A buffered upload keeps the route every deployment already has, so buffered callers work against old and new deployments alike.
A key reserved to the platform, and nothing can grant it away
Some permission keys are not a person's to hold — an engine's own scheduled work, for instance, runs as a system principal and nobody should be able to assume it. That was a line in a lint tool's table. It is now a fact the key declares about itself: holder: 'system' on the permission entry in the module's manifest. A vertical cannot lift an engine's reservation by re-declaring the same key, because one module marking it is enough.
The point of moving it is that a declaration can be enforced everywhere, and now is:
substrat pushrefuses the surface before anything is derived or uploaded, if a role, an entity-grant shape or an owner grant carries a reserved key — in the same wordspnpm lint:permissionsprints.- The push boundary at the control plane refuses the same, read off the registry rows rather than trusting the client that built them.
- The scope host refuses at run time, from the modules it actually registers: every door a person's role goes through, every raw grant to a principal or an org,
ctx.grantwhoever calls it, a connection grant (a connection acts for the external provider holding its token, so it is judged like a person's), and a reconcile of declared grant shapes. Nothing already seated is revoked. - A host that defines or seats a role itself is held to it too:
defineRole, the role set a control-plane-less provision or a restore projects, the provisioning owner seat and the owner hand-over. - A staff restore or import judges the roles in the dump before it loads them, and a production promote judges a release's roles against the scopes it moves.
A role is now stored with its holder, so a dropped role leaves a tombstone and a system-held one reads back as what it was defined to be rather than as a person's role that happens to hold odd keys.
Two things you will see. PERMISSIONS.md §1 gains a Reserved to a system principal column, so dropping a reservation shows up as a line in the reviewed diff. And the promote diff itemises a reservation that moved, a role that moved between a person's and the platform's, and a system-held role that changed — in the promote dialog, the Permissions tab, the Deployments ledger, the console and substrat promote.
One consequence worth knowing: a scope host now takes the permissions value your vertical declares, and refuses to register a module that value leaves out. That closes the gap where the list a reviewer reads was not the list the host runs. pnpm lint:host-permissions makes forgetting to pass it red.
A failure's text no longer tells one tenant about another
A promote, a provision or a store mint can fail for a reason that comes from somewhere else: another tenant's deployment, an object store, an upstream service. That text was going back in the response.
Now a confined caller — a tenant credential, a builder — is answered a fixed sentence and a reference. The text goes to the platform's own log line and to a staff-only row under the same reference. Staff keep the detail, and the reference is what ties them together, so an operator can find the text from anything you quote them. Whether a caller is confined is decided by the request's principal alone, never by a tenant or scope the request names — the one comparison with the path is on a provision reconcile, and it only ever gives text back: a failure about your own install is still your own answer about your own scope.
The reference you are given is always the platform's own id. Where an upstream service's trace handle appears in the text, it is kept on the staff row only, at the very start, so it survives the row's length bound — and a staff search finds both rows by either id, while your own search finds only your own row. Text that spells something looking like that marker is neutralised, so a forged one cannot be made to match.
The same rule now reaches the three places that were still relaying text: the host-level writers that file tenant rows (scope lifecycle delivery, the platform drain's give-up rows), the platform-intent journal, which withholds another deployment's or tenant's text, and the whole control plane rather than only the promote.
A read of a scope you do not hold is refused, not answered empty
Thirteen introspection reads — facets, event causes and effects, dead letters, entity history, the lifecycle flow, operation series, applied migrations, bookmarks, the table list, undrained events — addressed a scope by id alone. For a scope the deployment does not hold (never provisioned there, reaped, carried away) they woke an empty store and answered empty, which reads as "this app has no history" rather than "this is the wrong place to ask". They now answer conflict, with the same sentence the two size reads already used. The pure adapter refuses from the directory before any store file is opened; on the hosted side the held check and the read are one call.
What changes for you: a dashboard or console panel for a scope that has moved shows a failure rather than an empty panel. GET /scopes/:id/health still answers 200, with the role count at zero and the projection marked empty, so health stays a health answer.
The case that mattered most was invisible: the Tier-2 event drain lists active scopes, and an active scope whose deployment did not hold it read as "nothing to ship" — silently, forever. It is now judged against the directory as it stands. Reaped, archived or gone is nothing to ship; a scope whose route moved mid-read is picked up at its new home next tick; anything else is a standing condition, said in the sweep's log line and filed once a week rather than mailed every fifteen minutes.
Daily digests, so a busy topic is one mail a day
engine-notifications lets a person take any topic that is not required as a daily digest instead of one mail per occasion. Setting a cadence is one optional field on the existing preference operation, and no row means immediate, so every existing person and preference keeps meaning exactly what it meant.
The engine holds each occasion's rendered message, records its delivery as skipped with the reason digest, and a schedule sends each person one mail of everything held for them once their oldest held message is a day old — oldest first, at most fifty per mail, each person on their own clock rather than on the scope's. It composes no English of its own: a digest of one message is that message, and a digest of several puts each under its own subject as one document. A mute still wins, at hold time and again at send time; switching a topic off withdraws what was held for it, unsent.
Two parts of this are worth reading if you run it yourself. A scope may hold only so many platform requests waiting for the drain, so a pass sends to at most the room left and keeps a reserve for mail that must not wait — a backlog drains across passes instead of meeting the cap. And a held message that no longer matches the engine's published shape is set aside on its own, with the person's readable messages still going out in the same pass; a row set aside is read again once a day and released when it reads, so a rollback does not lose it.
A person can also now read their own deliveries, and the engine is published.
A scope of any size can be carried
preview create, and every carry onto another version, moved a scope as one value: one Durable Object message, 32 MiB, inside two isolates of 128 MB each. A hosted production scope is larger than that, so the copy simply failed — and the CLI's payload was never the problem.
The export now answers about 4 MiB of cells at a time, and the destination stages the pages in its own untyped tables and commits them through the same load path a whole dump takes, with the same fences and refusals. Pages are idempotent and must arrive in order; a replaced stage's commit is refused. If the source is written to between pages the copy starts again, and after three attempts it answers 409 with nothing moved. A scope that fits in one page, or a deployment on either end that predates the paged verbs, still moves whole — that is the old ceiling, and only there.
On top of that, a copy that outlasts its request now runs as a background job, with the carry stage staying owned while concurrent copies run and the keep-alive conditioned on pending directory state rather than on a timer.
Your app's API reference serves itself
An app could serve /openapi.json while the dashboard's API docs link answered 404 unknown route: /api/docs. mountOperations now registers the reference page and a bundled, self-hosted renderer automatically. An app with a specification adopts it by updating @substrat-run/vertical-host and redeploying — no route to add and no asset build step. The page reads the specification through the app's existing session, and says so plainly when authentication fails, when there is no specification, or when the document does not parse. A reference you registered yourself before the operation mount keeps precedence, and docs: false opts out.
Sign in with GitHub, without registering a client
Every Auth Server install now gets platform-provided social sign-in, offered as a switch that starts off. Nobody registers an OAuth client by hand: one reconciler reads the relay's live providers, builds the callbacks for each of the install's hostnames, ensures the install's client at the relay and delivers it. GitHub is live; Google and Apple go live by giving the relay their credentials.
Two decisions behind it. The platform owns the credential and the operator owns the decisions — a re-delivery keeps enabled, create accounts and trusted email as the operator set them, and a platform row cannot be removed, only switched off, because the platform would deliver it again. And a custom domain's callback is registered only once Cloudflare reports the domain verified, asked at that moment rather than read off its row: a callback on a domain nobody verified would put the platform's shared sign-in, with the platform's name on the consent screen, on a site that is not on the platform.
Connections land by declaration
A connection created through Connect, a connect link or the dashboard held no permission key and nothing bound it — so a company that approved a link landed nothing, and the sweep row said ok anyway.
A module manifest can now declare how a provider's connections land: the provider, the operation to call and the key to grant, with options. The platform seats it on each connector pass, which is the only clock that sees a connection arrive — a company's administrator approves days after the deploy, when nothing is being provisioned or reconciled. Consent does not confer the key; the running version's declaration does, which means it is versioned and reviewed like any other permission change: it is lifted into the permission registry, so a promote that changes it needs an acknowledgement, and PERMISSIONS.md renders it in a Provider connections section.
The sweep row is honest now too — failed naming each scope it could not land in, skipped when there was nothing to poll, ok otherwise — and a bug that fell out of the end-to-end test is fixed: with two companies connected to one vertical, the sweep opened the credential by vertical rather than by company, so every company's sweep failed.
The dashboard's reads are declared and team-scoped
Every read-only observability read — published-vertical metrics and logs, installed-app metrics, logs, request volume, facets, log patterns, tenant metrics, connector calls, business volumes, team traffic — moved to the declared /api/teams/:teamId surface. Each request selects the team named in its path, so two browser tabs can read different teams at once instead of fighting over one shared team cookie. The old cookie-scoped URLs for these reads are removed, and a path naming a team you are not a member of answers 404 before any team scope or control-plane read.
Alongside it the dashboard's own operations are declared, so each is held to the key it checks; its routes answer a declared plane value outside a known set, and an output that fails to parse answers 500 rather than shipping a wrong shape; six list reads page; enum-indexed lookups are Maps; an app is found by scope id with one read instead of a walk of every page; Settings team deletion is scoped to the path; and app lists come back in stable alphabetical order.
AI explanations for findings and process maps
A finding in Observability now has an AI explanation toggle: three or four sentences written from that finding's own evidence, saying whose side the problem is on — your app, the platform, or a supplier. The process map gets a note under the diagram from the same replay it draws. There is an Ask entry in the ⌘K overlay, on a model host that can run tools.
The interesting part is what does not happen. A page load never calls a model; opening the block reads what was filed, and writing happens only when you ask. What reaches the model is a field allowlist — kinds, severities, codes, counts, timestamps, versions, up to three messages — never a row, and never a tenant, scope, person or record id. Error messages are redacted first: keys, URL credentials, bearer tokens, JWTs, provider token prefixes, email addresses, phone numbers. Every paragraph carries an AI marker and a provenance line naming how many records it read and linking each source, computed by the dashboard rather than by the model. There are daily limits per person and per team, a unit is reserved before the call so parallel requests at the last unit make one, and a failed call still counts. Texts are kept in your own team's scope for 90 days.
This is on in production now, and the privacy policy lists Workers AI as a sub-processor for that traffic.
Refusals are problem documents, and three more rules hold
The control-plane worker's own routes, the CLI login token exchange and three dashboard 401s answered bare { error } as application/json; they are application/problem+json now, with statuses and the error sentence unchanged and error duplicated as detail. The builder answers refusals the same way, and the hosted adapter's live-read refusals do too. Three other rules became mechanical this week:
- An operation that declares a permission and never calls
ctx.checkis recorded by both scope hosts — awarnline in production — and fails every node test run, rolling the operation back. A check counts wherever it was made: in a helper, caught after a refusal, insidectx.atomic, or inside a kernel verb that re-checks the caller. Derived handlers,narrowswalks over nothing and modules with no declared surface are not held, each for a reason its declaration states. - A Worker that imports a Node builtin its compatibility flags do not give it is refused before a push, a preview or
push --check, with the import path in the message — statically, dynamically and through CommonJS, transitively. Such a bundle passes every local check and then fails to start. - A scope host that registers modules must be given the
permissionsvalue its vertical declares, or the lint refuses it.
The scope hosts also refuse a duplicate module id, a duplicate operation name and a module that withdraws its own operation, rather than registering something ambiguous.
Hash-only invites have a cutoff
createInvite now requires the become capability id — a missing, null or empty one throws before anything is written, so no new hash-only invite row can be created at all. The explicit seam for one, used by fixtures, is a separate function that is not a method of the identity directory, so the directory's RPC surface has no hash-only write left.
Rows that already exist stay claimable until the first @substrat-run/vertical-auth release published on or after 9 January 2027. Before that release, the platform counts and withdraws any hosted-scope invites still pending, which have to be sent again.
Separately, the shared identity directory no longer bundles a credential provider. It statically imported an unused email/password provider, which pulled Better Auth and Drizzle into every OIDC app that binds the directory. The provider, its schema, its HTTP surface and the doAuthProvider export are gone; use oidcAuthProvider or oidcRpAuthProvider. Existing credential tables are untouched, and the separate auth server's own credential handling is unchanged.
Eight changes to notice
createInviterequires a capability id. A missing,nullor empty one throws. Migration: pass thebecomecapability id. A fixture that needs a hash-only row usescreateLegacyInvitefrom@substrat-run/vertical-auth/member-directory.doAuthProvideris removed from@substrat-run/vertical-auth.Migration: useoidcAuthProvideroroidcRpAuthProvider. Your credential tables are left as they are.- A scope host that registers modules must be passed
permissions. A module your declared surface leaves out is refused at registration. Migration: passdefinePermissions(...)'s value in the host's options — the same constant your seed already exports.pnpm lint:host-permissionsfinds the sites. - An operation that declares a permission and checks nothing fails your test run. Migration: call
ctx.check— or, if it genuinely checks nothing, say so withnarrows: { reason, checks: [], unchecked: true }. - A role or grant carrying a key reserved to a system principal is refused at push, at the push boundary and at every run-time door. Migration: a role that legitimately needs one declares itself system-held. Nothing already seated is revoked.
- The dashboard's observability reads moved to
/api/teams/:teamIdand the cookie-scoped URLs are gone. Migration: none if you use the generated client; it is re-emitted. - Introspection reads answer
conflictfor a scope the deployment does not hold, where they used to answer empty. Migration: none, unless you read an empty answer as "no data". - A Worker's Node builtin imports are checked before push. Migration: add the compatibility flag the message names, or drop the import. Run
pnpm lint:worker-node-importsto find yours.
For hosted operators
- Usage is captured day by day, because it cannot be recovered later. One row per (UTC day, tenant, vertical, scope) for requests, split by status class with rate-limit refusals counted, every count an estimate and marked as one. One snapshot per (UTC day, tenant) of the meters — status, scopes by effective status and by vertical, every entitlement with its expiry and whether it was live — taken by the first pass of the day in one directory transaction. Daily attachment bytes beside the storage gauge, with an internal size read. And a vertical's ownership history, kept as periods rather than a current owner, so a slug deleted and registered again by someone else still bills the builder who owned it then. Three new staff routes read them back (
/meters/requests,/meters/days,/verticals/:slug/owners). The retention horizon for billing rollups is deliberatelynull— nothing is pruned — until a period is decided; ownership history is kept for good and survives a tenant reap. - Sixteen more admin writes commit with their audit row. The tenant and vertical directory writes now write state and admin row in one transaction on both adapters, so a refused row can no longer leave a state change unrecorded. Two behaviour changes come with it: a promote's own row commits with the channel move, so it lands before the rebind rows of the scopes an embedded promote moves; and a hosted tenant reap clears its directory rows in the same unit as the status flip, as the pure adapter always did. Scope writes, and the tenant-node role, grant and org writes, landed the same way.
- The platform sweep stopped double-spending attempts. The same-pass second drain re-ran every intent the first drain left pending, and each run counted toward the attempt ceiling — so a transiently failing intent was given up after about twelve hours instead of about a day. The second drain is now told what the first attempted, totals count each intent once, and "no drain reached this scope" means that again. A scheduled pass kicks the platform drain, purge-horizon runs report the requests their purges enqueue, a purge head the pending cap refuses is deferred rather than failed, and the supplied scope sweeper is quiet on routine passes again — a line per successful pass buried the failures it exists to surface.
- Wire signing is being built, and nothing is wired yet. One-way trust between the platform and a deployed vertical: the signers, the uploader, a fleet walk that brings every deployed script to the key lists and state the control plane gives now, and an advisory readiness report that lists what blocks moving an environment to signed-only, dropping a key, or withdrawing the shared credential. Unknown is never read as fact — a script with no record, a held record or an unverified one blocks every gate, and the confirming-walk read fails closed. Nothing is enforced on the flip yet, and the runbook says to run a confirming pass immediately before any flip or drop.
- Both adapters read every applied migration, not the first hundred. A scope past a hundred migrations was being judged on a truncated list.
- Smaller operational fixes: a reap keeps the storage days of the scope it reaps; an empty vertical slug is refused at provisioning and one unparsable scope row no longer stops a whole pass; a custom-hostname create adopts the zone's existing object on a 409; a sealed file's size is checked before it is downloaded and a too-large one is held; the sign-in issuer is discovered for default MCP mounts; the logs-incomplete header is forwarded and shown in the staff client, console and dashboard proxy, and the tenant-grain logs report incompleteness when a line is filtered out.
Also landed
- A per-PR preview's MCP endpoint is registered at its team's auth server, so an MCP client pointed at the preview gets a login page rather than
invalid_target— and un-registered by the same reap that removes the preview's client, through every door that reaps one. Separately, MCP mounts sendAllowon a 405, warn when an audience turns delegation off, and have their mount check hold under a HonobasePath. - The dashboard can manage a vertical's email sending permission, and holds the platform key that verifies a vertical's connect links.
- A pushed registry's holder marks are verified against the module declarations, rather than trusted as the client built them.
@substrat-run/contractsgains provider-neutral file change and write schemas, re-exported from the Microsoft 365 files vocabulary; its paged-list gate now looks through wrappers and refuses az.customoutput.- Store driver releases: an RFC proposing that a release becomes runnable through a person's approval of its uploaded bytes rather than a pin in the platform's source, the graph store driver's registration prepared without a pin, and a Python driver release allowed to carry its dependencies.
- Tooling: version labels are written from a whole id rather than its tail, outside the test kit too, and
lint:testsrefuses the spelling that collides; the load-sensitive cases got explicit timeouts; a one-MiB body is compared byte for byte rather than through a deep equal. - Docs: a privacy policy and terms of service; the trust boundary of a pushed vertical stated; a guide to the changeset PR preview workflow; the decisions of 23 October recorded (legacy invites, runtime extensions, file sources, semantic search, billing); and the third word meter and the semantic index written up as they stand.
- Dependencies: a
dompurifyadvisory patched so the audit is green, and the minor-and-patch group across eighteen updates. CI's Python anduvsetup actions moved to v7. - Upkeep: last week's changelog entry.
Released
| Package | Span |
|---|---|
@substrat-run/contracts | 0.151.0 → 0.162.0 |
@substrat-run/kernel | 0.151.0 → 0.162.0 |
@substrat-run/adapter-sqlite | 0.151.0 → 0.162.0 |
@substrat-run/adapter-cloudflare | 0.151.0 → 0.162.0 |
@substrat-run/vertical-host | 0.151.0 → 0.162.0 |
@substrat-run/control-plane-api | 0.151.0 → 0.162.0 |
@substrat-run/contract-tests | 0.151.0 → 0.162.0 |
@substrat-run/control-plane-client | 0.3.2 → 0.3.13 |
@substrat-run/vertical-auth | 0.25.0 → 0.28.2 |
@substrat-run/cli | 0.47.0 → 0.49.6 |
@substrat-run/attachment-extractors | 0.2.12 → 0.2.23 |
@substrat-run/model-emit | 0.10.7 → 0.10.18 |
@substrat-run/model-view | 0.4.0 → 0.6.9 |
@substrat-run/model-providers | 0.5.41 → 0.5.52 |
@substrat-run/dev-issuer | 0.2.33 → 0.2.44 |
@substrat-run/engine-notifications | 0.3.1 → 0.3.7 |
@substrat-run/engine-workorder | 0.12.35 → 0.12.46 |
@substrat-run/engine-protocol | 0.13.36 → 0.13.47 |
@substrat-run/engine-invoicing | 0.11.35 → 0.11.46 |
@substrat-run/engine-booking | 0.8.35 → 0.8.46 |
@substrat-run/engine-invites | 0.9.28 → 0.9.39 |
@substrat-run/engine-absence | 0.6.35 → 0.6.46 |
@substrat-run/engine-metering | 0.6.35 → 0.6.46 |
@substrat-run/connector-scrive | 0.14.57 → 0.14.68 |
@substrat-run/connector-fortnox | 0.4.53 → 0.4.64 |
@substrat-run/connector-planima | 0.2.48 → 0.2.59 |
create-substrat | 0.12.1 → 0.12.12 |
vertical-auth's three minors are the required capability id on an invite, the removed credential provider, and the type tightening that follows. cli's two are the Worker import refusal before push and the reserved-key refusal on the surface. model-view's two are the reservation column and the registry table reader the bump classifier and the console now share. engine-notifications is published from this week. In 0.x semver a minor is where a breaking change goes, so the fixed group's 0.162.0 carries every "change to notice" above.