Repository navigation
v1.6.0
MINOR: one new endpoint (GET /api/whoami), two new request parameters, the
organize token slot, and the Go side of the relay moving onto a toolchain that
still returns a verdict.
A thread read is now paginated, and that lands differently on the two interfaces
(#706, #649). Read both before upgrading.
1. A SIGNATURE break, service bindings only. MailboxService.thread() returns
Promise<Page<StoredMessage>> where it returned Promise<StoredMessage[]>.
Page<T> is { items: T[]; cursor?: string | null } (inbound/src/store.ts), so a
Workers caller that iterated the result must now read .items. That RPC is one of
the three documented ways to consume postern (README, docs/CONTRACT.md,
docs/INTEGRATION.md), so this is a real break, not an internal detail, and it is the
change most likely to fail your build. It is not quite the only one: MailboxService.search()
returns Promise<Page<SearchHit>>, so the snippet removal noted under Fixed can also fail a
TypeScript caller that referenced hit.snippet through the RPC. That field never had a producer
and always evaluated to undefined, so such a caller was reading dead code, but it will still
stop compiling.
2. A BEHAVIOUR change over HTTP, additive in shape. GET /api/threads/{id}
keeps messages with the same name, position and array type and adds cursor, so
nothing deserialises differently. But the default is now bounded at 20 messages
(THREAD_DEFAULT_LIMIT), where the route previously returned the entire thread with
no limit. An HTTP caller that ignores cursor therefore gets a silently truncated
conversation on any thread longer than 20. To read a whole long thread, follow the
cursor or pass an explicit limit. Our own webmail is such a caller and had to be
taught this in the same release: it now prints "first page" beside the count rather
than implying the page length is the thread length.
The Python client is unchanged for existing callers: get_thread() still returns a
list, with get_thread_page() added alongside. No other interface in this release
changes incompatibly.
No PROJECTION_VERSION or POSTERN_IMAP_UIDVALIDITY bump: the wire format and the
door projection are unchanged. One declared-but-producerless field is removed, noted
below.
Operator action: none for a deployment; one for a service-binding caller
POSTERN_API_TOKEN_ORGANIZEis a new, OPTIONAL worker secret slot (#692, #694).
It issues theorganizescope on its own:POST /api/messages/seen,/flagsand
/move, and nothing else. It cannot read, send, hard-delete, reach admin, or write
through theimapseam, and it does NOT coverGET /api/folders, which is a read.
With the slot unset nothing changes, which is the state every existing
deployment is in: animaptoken still satisfiesorganize, and areadtoken is
still refused on those three routes. Set it only if you want a bearer that can
manage read state, flags and placement without holdingimap. SeeDEPLOY.mdand
docs/AUTH-CONTRACT.md.
Added
GET /api/whoami(#701). A Bearer caller can ask what its own scope is instead
of discovering it from a 403. Correctlyread-scoped: a caller with no read
capability gets the estate-level answer rather than a member view.countOnlyonGET /api/messages(#702). Returns the match count under the
rows' own predicate, so a count and a page agree by construction rather than by a
caller reimplementing the filter.seen,flagsandmoveas organize-scoped MCP tools (#645, #700). They
register only when an organize-capable bearer is configured, and they are mutating
and opt-in. Shipped on thepostern-mcp-v*train, which keeps its own ledger.
Fixed
- A thread read is bounded, and the truncation is visible (#706). An unbounded
thread fetch could return an arbitrarily large body; it is now capped and says when
it dropped something, rather than silently returning less than was asked for.
This is the change described at the top of this section, on both interfaces: the
RPC signature becomesPage<StoredMessage>so the bound and the cursor are
expressible at all, and over HTTP the shape is unchanged while the default becomes
bounded. snippetremoved from the wire type (#699). It was declared and no producer
ever set it, so every consumer reading it gotundefined. A field that cannot be
populated is worse than an absent one, because it reads as empty rather than missing.- The IMAP seam presents the imap-scoped token on the organize routes (#686, #703).
It was sending a token that those routes refuse, so door-side organize failed on
scope rather than on anything the caller did. fieldoutsidesubstrmode is refused instead of echoed (#651, #698). The MCP
search acceptedfieldin modes that ignore it and reflected it back in the
response, which read as "applied".- The live smoke probe is pinned to one deployed version (#692, #695). It could
otherwise probe a mixed state during a rollout and report a failure that belonged to
neither version.
Relay and CI
- Go toolchain pinned to
go1.26.9andgolang.org/x/netraised to v0.60.0
(#707, #713). v0.60.0 is the only release that clears the outstanding advisories.
Thegodirective moved to 1.26.0 with it, which retiredGOTOOLCHAIN=go1.25.13as
a negative control;relay/README.mdrecordsgo1.26.0as the lowest toolchain that
still returns a verdict, and that a lower one exits 1 rather than going red. - CI asserts the scanned Go equals the shipped Go (#704, #711). The vulnerability
scan and the Dockerfile could previously disagree, so a clean scan said nothing about
the binary that ships. mypynow type-checks the server and threaded bodies, with
check_untyped_defson for clients and imap (#680, #696, #697, #708).tag-preflightassertsinbound/src/version.tsand counts its pins truthfully
(#693)./healthreports that constant, so a release could previously ship a
version string that disagreed with its own tag.- The IMAP test waits for the server side to close before the reactor check (#715).
Test-only; it was a race in the harness, not in the door.