Widen the notif: match filter and fetch scope for a session's own
generic developer/app-user subscribe: today it pins ref to the
session's own uuid, so a row naming an app (handler-suspension
notices, app-bound worker deploys) never matches live and never
replays on reconnect, even though the audience predicate already
grants the holder every such row it owns. The predicate is the
authority and already reruns per row/page after the match, so
widening the filter to it (account is unaffected — it never names an
app) adds no exposure.
An actor holding an app reads the `app-user` rows naming that app, plus its
`developer` rows when the holder owns it. `account` rows reach no app, and a
slice an actor may not see comes back empty rather than refused. The audience
predicate becomes the enforced read path in the same change that lifts the
blanket app-actor 403, layered behind an audience/app_uid SQL scope; two-segment
`notif:` subjects expand server-side from the actor's own app, so an app can
never name another app's uid.
No feature flag: `audience` defaults to 'account', so every pre-registry row is
default-denied to app actors and the backfill can only narrow.
* feat: presence and cross-region event forwarding (PUT-1679)
* fix: fan cache bumps to sibling nodes and stop the forward shed cascading (PUT-1679)
`outer.events.generationBumped` and `outer.events.presenceBumped` rode
`outer.*`, which the broadcast service only webhooks to peer regions;
only `outer.pubsub.*` also fans over Redis to a region's other nodes.
Both caches are per-process maps with no expiry, so a bump landing on
one node left its siblings stale until that user's next transition.
Renamed onto `outer.pubsub.events.*`; the listeners already accept the
`from_outside` copy the Redis re-emit carries.
`PeerForwardQueue.push` called `onOverflow` synchronously and the
handler pushed markers straight back, each of which re-tripped the
bound and shed the next item: one item over a 5000 bound recursed ~2200
deep, threw a RangeError, and turned ~2200 queued deliveries into gap
markers. It also re-summed `bytes` over the whole queue per drop. The
handler now returns its markers and the queue appends them past the
bound check, sheds deliveries before markers, keeps one pending marker
per (peer, subscription), and subtracts bytes per dropped item.
`compileMatch` turned every `*` into an unbounded `[^/]*`, so a subject
like `fs:~/x/*a*a*a*a*a*a*a*a*a*a*z` tested against a 24-char filename
with no `z` — both named by the same subscriber — cost the regex engine
C(34,10) splits before failing: 1.0 s per event at ten stars, 3.7 s at
eleven, on the fs-write dispatch path, inside the 256-character cap.
With one `*` per segment the delimiters pin each star and a wrong split
dies in one step; the single `**` is the only choice point left, so the
worst allowed shape is O(depth × length): 0.4 ms at depth 60 of 200-char
segments. Every documented pattern (`*.png`, `**/build.log`, `**/*.png`,
`dir/**`, `report-?.csv`) stays valid; `*a*`, `a**b` and `**/x/**` are
refused with `invalid_subject_pattern`. The rule is stated where the
syntax is introduced, on the limits page, and in onLocal's error table.
App payloads carried only the /app-icon endpoint URL, which 302s to the
icons hosting subdomain. Networks that mangle that redirect render no icon
at all, and every icon load pays a round trip for the hop.
Ship the direct subdomain URL as `iconCdnUrl` alongside it (taskbar items,
installedApps, recent/recommended launch apps, suggested apps), and have the
GUI load that first with the endpoint URL as a one-shot retry - desktop
taskbar, start menu, dashboard app grid and recents. Only rows whose `icon`
column is already an http(s) URL get one: a data: column means the resize
pipeline has not written anything to the subdomain yet.
Also folds the four copies of the generated-size list into one exported
APP_ICON_SIZES.
Covers PUT-1708, PUT-1709 and PUT-1743.
Twelve routes, every one setting requireUserActor -- that option is what
installs requireAuthGate, requireVerifiedAccount and requireNonAccessTokenGate,
because server.ts derives `needsAuth` from the route options. Reads need it as
much as writes: without an auth option a route gets no suspension check and
admits access tokens, so a just-disabled member could still read the roster and
a scoped third-party token could read the audit log.
Authority is checked before anything observable. Validating the body first made
POST /members answer 400 before 403, and resolving :username first turned the
member routes into a global username-existence oracle.
Provisioning applies the same username and email rules as signup rather than
its own -- USERNAME_REGEX, USERNAME_MAX_LENGTH, RESERVED_USERNAMES and
validator.isEmail, now exported from AuthController. Without them a workspace
could mint accounts signup would refuse, claim unregistered reserved names, and
mail arbitrary unvalidated addresses.
Handle problems are 400 or 409 rather than a bare Error, which the server turns
into a 500 and a deduped critical alarm -- an uppercase handle should not page
on-call.
Disable drops sessions through SessionStore.removeByUuid rather than a raw
DELETE. The store invalidates every composite cache key; without that a
disabled member kept authenticating from cache for the session TTL, which is
exactly the "takes effect on the next request, not after a cache TTL" property
disable is supposed to have. Revoking also preserves last_ip/last_user_agent,
which the member-facing audit view reads.
Audit writes live in TeamService at the point of each action rather than in the
route, so a caller reaching the service directly cannot skip them, and the SQL
lives in TeamStore. Audit reads map internal user ids to usernames, and remain
readable by the owner after the workspace is soft-deleted -- otherwise the
delete_team entry was written and immediately unreachable.
teams_enabled gates route registration through an optional isEnabled() the
server honours, so with it off the paths do not exist rather than existing and
refusing. It does not gate DDL.
TeamIsolation.http.test.ts asserts the negative the feature rests on: the
workspace manages accounts and cannot read them, including through a
full-access token and after the member is disabled. It asserts outcomes rather
than the absence of an implicator.
Covers PUT-1705. The master account supplies { username, email }; the account
is created with no password, gets the default filesystem tree, joins with
org_owned = 1, and receives a one-shot activation link.
Activation reuses password recovery rather than new token machinery: the same
pass_recovery_token, the same one-hour purpose-scoped JWT, the same
/action/set-new-password link. No team_activation table, no new token type,
and no unauthenticated endpoint on the team surface. Activation state needs no
column either -- an unactivated account is one with no password.
Applies the same username and email rules as signup rather than its own:
USERNAME_REGEX, USERNAME_MAX_LENGTH, RESERVED_USERNAMES and validator.isEmail,
now exported from AuthController. Without them a workspace could mint accounts
signup would refuse -- the username becomes the /username home-directory
segment -- claim unregistered reserved names, and send activation mail to
arbitrary unvalidated addresses at the route's daily limit.
Usernames come from Puter's global pool, so a taken one is refused with free
alternatives rather than silently modified: a suffixed name would appear in
every share dialog that person ever sees, and they never agreed to it. The
check runs before any write, so a rejected provision leaves no orphaned user
row -- asserted by a test on the workspace's member count.
The new account carries requires_email_confirmation, since the address came
from the administrator rather than its holder.
Adds a team_account_activation email template stating what the workspace can
and cannot do -- including that it can reset the password, which the design
requires be said rather than only claiming files are private.
free_storage stamping and the billing event are phase 3.
Covers PUT-1704 and PUT-1707: creating a workspace, admitting the master
account, and the whole of offboarding.
`createWorkspace` admits the creator with org_owned = 0, which is what makes
the master pay for itself and stay an invalid target of every member route.
`checkOwnerInvariant` asserts the rule no dialect can express -- the owner is
a member with org_owned = 0 and the only such member -- and a test breaks it
deliberately, since the schema cannot refuse a second one.
Three authority checks: 404 to a stranger so the endpoint is not an existence
oracle, 403 to a member who is not the master, and the master refused as a
target of member routes.
Handle problems surface as 400 (unusable) or 409 (taken), including the
unique-index race. `TeamStore` throws a bare Error, which the server would turn
into a 500 and a deduped critical alarm -- an uppercase handle should not page
on-call.
Disable writes `user.suspended` as well as suspended_at and suspended_reason.
PUT-1707 named only the latter two, but those are siblings added by 0061 and
0063 -- `userProtected` rejects on `if (user.suspended)` and reads neither.
Setting only the timestamp and reason would have recorded a disable that never
took effect, and disable is the whole of offboarding here.
Sessions are dropped through SessionStore.removeByUuid rather than a raw
DELETE. The store invalidates every composite cache key with its double-delete
pattern; without that a disabled member keeps authenticating from cache for the
session TTL, which is exactly the "next request, not after a cache TTL"
property disable is supposed to have. Revoking also preserves last_ip and
last_user_agent, which the member-facing audit view reads.
Files are untouched and re-enable restores the account.
Adds team_not_found, not_the_master_account and not_an_org_account to the
HttpError legacy codes, which the controller also needs.
Billing events, invalidateActorSubscription, audit rows and the GUI push are
deliberately not here -- they belong to phase 3 and PUT-1708.
Membership management for workspaces: addMember, removeMember,
getMembership, isMember, listMembers and listTeamsForUser. The permission
scan is untouched -- readUserGroupPerms already joins jct_user_group and
resolves group grants; this is the management side.
Resolves the ticket's "do not leave two writers" by splitting domains and
enforcing the split in SQL rather than by convention. Every existing caller
of GroupStore targets a seeded system group -- ADMIN_GROUP_UID,
default_user_group, default_temp_group -- never a team, so the two stores
were already disjoint in practice. GroupStore.addUsers/removeUsers now carry
`AND kind IS NULL`, making a team uid a no-op there, which costs no extra
query because it folds into the existing subquery and matches how addUsers
already treats an unknown username. TeamStore's writes select group_id from
a kind-filtered subquery, so neither store can reach the other's rows.
org_owned is written here but never accepted from a request; TeamService
sets it at provisioning and workspace creation only.
listMembers is keyset-paginated on id per doc/pagination.md, using the
shared cursor and limit helpers and fetching one row past the limit to
decide whether a cursor is warranted.
Passes 1/0 for org_owned rather than db.booleanValue, which yields a real
boolean on postgres and is rejected by the smallint column there -- sqlite
accepted it silently.
`GroupStore` has only addUsers/removeUsers; nothing creates, reads back or
lists a group at runtime. `TeamStore` is that missing half, scoped to rows
with `kind = 'team'`.
A workspace is addressed by `uid`, which `group` has carried as NOT NULL
UNIQUE since 0015. `handle` is a mutable display label with no addressing
role, so a rename invalidates nothing and a stale reference can never
resolve to a different workspace.
Soft delete releases the handle and keeps `name`. Nothing points at a
handle, so the name returns to the pool instead of being reserved forever
by a global unique index that cannot exclude dead rows -- mysql has no
partial indexes, so that exclusion was never available.
Handles validate to ^[a-z0-9]+(-[a-z0-9]+)*$, 3-64 chars, against a
reserved list. The charset is deliberately narrower than the column so the
engines' collations cannot disagree: mysql's utf8mb4_unicode_ci also folds
accents and eszett, which sqlite's NOCASE and postgres's lower() do not.
Every read filters `kind = 'team' AND deleted_at IS NULL`, which is what
makes the seeded admin/system groups unreachable rather than merely absent.
Handle lookups compare lower(handle) on postgres, where the index is on
that expression rather than the column.
Insert-only record of what a workspace administrator did to an account,
shaped like \`audit_user_to_group_permissions\` after 0019: nullable FK
beside a NOT NULL \`_keep\` column. The FKs are ON DELETE SET NULL, never
CASCADE, so hard-deleting an account cannot erase the record of the
resets performed on it.
Two indexes rather than one. The member's own view is the only place a
reset becomes visible to the account it was performed on, so
(user_id_keep, id) is a read path, not an optimisation.
\`share.holder_group_id\` mirrors \`holder_user_id\` from 0067. The existing
unique index does not constrain team shares at all -- it leads with
\`holder_user_id\`, which is NULL on every team share, and NULLs are
distinct -- so the group-scoped unique index is what prevents duplicates.
Drops the \`role\` column from the specified DDL: it contradicted the
settled single-administrator model.
Review follow-ups on the dedup migration.
mysql and postgres track no per-file applied state and re-execute every
migration on each boot, so the unguarded DELETE self-joined the whole table at
every process start, forever. Both now sit behind the same index-existence
check that guards the ALTER, which also stops a rolling deploy deleting on one
instance while another adds the index.
The dedup test was a false green. `targetVersion: 67` never applies 0071 -- the
loop breaks on `threshold + 1 >= targetVersion` but stamps the target anyway --
so the fixture asserted the current schema version on a database missing a
migration. It now replays the real 0072 file against a fully migrated database,
and a second test pins the off-by-one so nobody builds a fixture on it again.
\`jct_user_group\` had no unique constraint on (user_id, group_id) and
\`GroupStore.addUsers\` had no conflict clause, so re-adding a member
inserted a second row. \`readUserGroupPerms\` joins the junction table on
group_id alone, so each duplicate returned another copy of every group
permission the user holds.
Deduplicate keeping the lowest id, add the unique pair index, and make
\`addUsers\` ignore conflicts via the existing \`insertIgnoreInto\` helpers --
without that last part the index turns a re-add into a raised error, which
five call sites would log as a failed signup step.
mysql cannot delete from a table it reads in a subquery (error 1093), so
it uses a self-join with the same lowest-id-wins semantics.
Review follow-ups on the team columns.
NOCASE moves onto the `handle` column itself, not just the index. Index-only
NOCASE makes uniqueness case-insensitive while leaving `WHERE handle = ?`
case-sensitive, so the same lookup would match on mysql (utf8mb4_unicode_ci)
and miss on sqlite. Postgres still needs lower(handle) at the call site.
`requires_password_change` becomes NOT NULL DEFAULT 0 on all three dialects,
matching the three sibling `requires_*` flags. Left nullable, any query written
as `= 0` would silently exclude every pre-existing user.
Schema foundations for teams. Ships dark — nothing reads these columns yet.
group: kind, name, handle, plan_id, deleted_at
jct_user_group: org_owned
user: requires_password_change
`kind = 'team'` marks a workspace; the seeded system groups keep it NULL so a
team query can never surface them. `org_owned` distinguishes accounts the
workspace created from the master account, deciding who pays rather than who
may read. `requires_password_change` is a fourth `requires_*` flag for
assertVerifiedAccount, needed by the phase 5 reset flow but shipped here to
avoid a second three-dialect migration.
`handle` uniqueness is case-insensitive, and the dialects disagree by default:
mysql gets it from utf8mb4_unicode_ci, sqlite needs COLLATE NOCASE (as
0055 does for usernames), and postgres indexes lower(handle). Without this the
same migration would accept `Design-Team` next to `design-team` on sqlite and
postgres while mysql rejected it. Postgres handle lookups must therefore
compare lower(handle) to use the index.
idx_group_owner is sqlite-only: mysql and postgres already index that column.
idx_jct_user_group_group is composite, unlike the existing single-column keys.
The Errors section claimed a `{ message, code }` body and a single
ten-minute wait. The body also carries `error`, `errorCode`, `provider`,
`upstreamCode` and, for pre-job rejections, `upstreamStatus`, and the
Sora provider still waits five minutes and folds refusals into
`upstream_failed`. Document the fields and the exceptions.
Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
A video job that outlived its poll window, an SDK request that timed out,
or a Veo operation that finished with an error all reached the HTTP error
handler as plain Errors. Each became an unhandled 500 with critical
severity and paged on-call for what is the provider's pace or the
provider's fault.
Video providers now share one poll loop that gives up with a 504
`upstream_timeout`, treats a transient poll failure (timeout, dropped
connection, 408/429/5xx) as a missed poll rather than a failed job, and
stops polling with a 400 `client_aborted` when the caller disconnects, so
nothing is metered for a clip nobody will receive. The driver controller
exposes the disconnect as an `abortSignal` on the request context. The
window is ten minutes for every provider; Together and BytePlus move up
from five.
Failed jobs are classified: content-filter refusals become a 400
`bad_request` with `errorCode: moderation_flagged`, rejected parameters a
400 `upstream_bad_request`, and anything else a 502 `upstream_failed`,
each carrying the provider's own code. Veo's filtered output keeps
`disallowed_value` and gains the same `errorCode`. The sanitizer and
content-filter pattern move from the Replicate provider into a shared
util so image and video agree.
Status-less SDK connection timeouts are translated to a 504
`upstream_timeout` at the driver boundary, and the chat driver records
them per attempt so an all-timeout chain is a 504 and a mixed chain is
`upstream_failed` instead of an `internal_error` 500. The Together chat
client gets the same ten-minute request timeout as the other providers.
The OpenAI video provider is left alone beyond an import path: its API is
scheduled to shut down on 2026-09-24.
Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
* feat: events metering and quotas (PUT-1683)
* fix: drop the standing subscription charge and price single deliveries at 100 µ¢
An idle durable row costs nothing worth billing; the plan quotas bound how
many an account holds. Removing the daily line also removes the global
day-claim, the whole-table scan and the timer it rode on.
* test: stop asserting on documentation pages
The limits test read rate-limits-and-quotas.md and grepped it for numbers,
so every rewording of the page failed the backend suite. Docs are kept in
step by the PR and checked in review; AGENTS.md now says so.
* feat: delivery re-check cache, revocation and anchor settle (PUT-1677)
* fix: authorize re-anchors, settle each row once, purge revoked backlog (PUT-1677)
- A path-form row whose anchor is deleted only climbs to an ancestor its
holder may still watch under the mode it subscribed with; otherwise it ends
with `anchor_deleted`. It used to land on any surviving ancestor (a guest's
row on the owner's home), where the re-check denied every delivery but the
row still held an anchor slot and a filter evaluation there.
- After a climb the new anchor is re-verified and the climb repeated if a
recursive delete took that level too, instead of leaving the row on a dead uid.
- suspend() is one conditional write per row and reports which rows it was the
one to suspend; concurrent settles of the same grant (an unshare revokes
several strings) no longer each purge, forget and notify the same rows.
- One "subscriptions ended" notification per holder and app, carrying the count
and subjects, instead of one per row.
- A revoke that removed nothing no longer announces; the sweeper purges (not
defers) the backlog of a permission_revoked row; the reap purges pending
entries with the row.
- The delivery auth cache indexes entries by subscription so forget() is not a
scan of the whole cache.
* feat: pending event deliveries and delivery-class invariants (PUT-1676)
* fix: make pending delivery claims and drains atomic, keep the region under its ceiling (PUT-1676)
- claim() and the drain-time reindex run as Lua over the subscription's own
{subId}-tagged keys. Two claimers can no longer both lease the head, and a
drain that finds the queue empty deletes it in the same step it checks, so a
concurrent enqueue is never wiped between the two.
- An append writes the entry and its queue position in one MULTI (same slot),
with the index seeded before it and corrected after, so an entry is never
visible without its position and never left out of the sweeper's index.
- Pipelines no longer mix slots (index/counter vs. per-subscription keys), so
the store works on a multi-shard cluster, not only a single-shard one.
- Region shedding counts the marker it leaves behind; it used to stop one over
the ceiling and convert a real event into a marker on every enqueue after.
- A claimed or suspended subscription moves to the back of the sweeper's index,
so a delivery nobody settles cannot hold the head against every other backlog.
- `single` rows must carry a `worker` target: with sockets exhausted and no
handler, an unacknowledged delivery would sit at the head forever.
- A gap marker for a row with no socket target is dropped rather than counted
as a delivery of nothing.
- Backlog keys carry a 7-day TTL, refreshed by every claim, as a backstop for
keys a purge/enqueue race left unindexed.
* feat: durable event subscriptions store, cache, and routes (PUT-1673)
* fix: durable subscription hardening (PUT-1673)
- Expired rows stop delivering at dispatch time and no longer count toward
the per-account cap, instead of waiting for the sweep.
- The expiry sweep runs hourly with a jittered first pass shortly after boot;
a 24 h interval never fired on a fleet that redeploys more often than that.
- Only a durable generation bump marks peer regions cold. A session
subscribe/unsubscribe in one region used to force a primary read in every
other region on its next dispatch.
- `subject`/`anchor_path` widen to varchar(4096) to match `fsentries.path`,
and subjects longer than that are refused with `invalid_subject` rather than
failing the insert on MySQL/Postgres.
- The dispatch and durable integration suites wait for the specific delivery
they expect and assert only within their own folder; the old any-delivery
`settle()` let a late event from a previous test satisfy or pollute the
next one under CI load.
* feat: puter.events SDK module (PUT-1675)
* fix: end event subscriptions on a server-side disconnect and retry budget refusals (PUT-1675)
- A disconnect socket.io will not retry (`io server disconnect`) now fails
every subscription with `events_connection_failed` and drops the socket, so
the next onLocal() starts a fresh connection instead of reusing a dead one.
- The connection closes when the last subscription lapses on re-subscribe.
- A re-subscribe refused with `too_many_requests` is retried after 10 s
rather than ended for good.
- The API suite no longer passes when it cannot connect; every runtime it
runs on (node, browser, workerd) carries the socket.
OpenRouter and Together reject a request whose prompt plus max_tokens
overflows the model's context window. Both providers retried by deleting
max_tokens, which threw away the output cap the credit gate had sized to
the caller's remaining balance and let the retry run to the model's full
output limit with no second gate and no new hold.
The retry now goes through a shared helper that sizes a new cap from the
window and input count the rejection reports, falling back to the model's
declared context and a doubled prompt estimate, and never exceeds the cap
the gate set. When no window can be determined or no output fits, the
original rejection is rethrown instead of retrying uncapped. The rejected
params are copied rather than mutated, so the first attempt's record is
not rewritten after the fact.
Also corrects the estimator's own comment, which described the mean of
two approximations as a deliberate halving.
Co-authored-by: Claude Fable 5.1 <noreply@anthropic.com>
A Replicate prediction that ran and ended `failed` reaches the provider as a
plain Error with no HTTP status, so the driver-boundary translator could not
classify it and it surfaced as an unhandled 500, a critical alarm, and an
on-call page. Most of these are the model's content filter refusing the
user's prompt.
Wrap the run call and classify the failure: content-filter refusals become
a 400 with `errorCode: moderation_flagged` (the code chat refusals already
use); anything else becomes a 502 `upstream_failed`, which the alarm gate
skips. Status-bearing SDK errors pass through untouched so the boundary
translator keeps handling them. Upstream messages are stripped of markup and
bounded so an HTML error page can no longer ride into a response body or an
alarm signature.
Documents the codes callers can now act on in the txt2img reference.
Co-authored-by: Claude Fable 5.1 <noreply@anthropic.com>
Daniel Salazar