Skip to content

Konclave v0.14.0

Pre-release
Pre-release

Choose a tag to compare

@github-actions github-actions released this 23 Oct 02:29
19a3ef1

The app

Changed

  • The device keeps the network upgrade a transaction was signed under (#638). Its record of what
    it signed now stores the consensus branch it read in the transaction, beside the expiry, so a later
    upgrade is counted when it decides whether signing again could pay twice. Nothing changes until the
    next upgrade's row is added: an older record without a branch reads as NU6.3, which is true of
    every record that exists. When the earlier attempt is ended by an upgrade rather than by its own
    expiry, the confirm names the last block it can be mined in and when that is expected, and when
    the device has no turning point to give (a record that never stops counting) it says so instead of
    naming a time. Seven new texts (English and Portuguese), approved by the maintainer as written:
    signing.againUntilEpoch, signing.againNoTime, error.needsNewSignature,
    error.fundsHeldUntil, error.upgradeHold, error.upgradeHoldCrossing and
    error.coordinatorNotUpdated. The app words the coordinator's refusals by the reason it sends,
    also on /net, in both languages. None can appear yet.
  • The public roadmap is current as of 18 October 2026. Step 1 says what is already measured
    (making one transaction's proof in a phone's browser, with the median times on three phones and a test anyone
    can run) and what is still to measure. The page no longer says a step starts only once the one before
    it is done: a step can start earlier, and the page says what each one waits on. The network upgrade
    (NU7, #638) is listed as current work outside the numbered order, and #567 carries its title as it
    reads on GitHub now.

Fixed

  • The site names the 0.13.2 desktop pre-release. The landing and the help named 0.13.1 after
    the 0.13.2 installers were published. The download buttons stay off until the desktop app is
    validated on each platform (#212).

The desktop shell and shared parts

Services deployed when this was cut (2026-10-23): coordinator 404a8bb, relay f8dc88d0970337a3.
The web app is this release's own commit. The two services are deployed by hand and may be
older than it - recording what they were ANSWERING is the point, not what they should have been.

Security

  • A client could use up another client's request budget at the relay. The relay refuses a
    sender that makes too many requests in a short window (429). That budget was counted under a key
    built from a string the requester supplied, and the per-address and per-sender budgets were kept
    in one table, so a supplied string could name someone else's budget. A request that named another
    address, or another member's tag, was counted against that other budget, and once it was used up
    the other party's own requests were refused for as long as this went on. Anyone who knew an
    address could do it to that address, and anyone who had a room's id could do it to a member's tag
    there. It affected whether requests were answered, not what the relay carries. The code keyed the
    per-sender budget this way from 2026-07-15 and the per-address one from 2026-08-22, the earliest
    dates it could have been exposed. We have no report that it was used; we did not check. What a
    member can notice: nothing, unless a ceremony stalled with no message to explain it. Every budget
    is now scoped to the address the relay attributes to the request, and a poll whose tag is longer
    than a post's limit is refused. The hosted relay has run this since 2026-10-19. The coordinator
    never had a per-sender budget, so the mechanism above did not exist there; it now counts under the
    same typed address, which the hosted coordinator has done since 2026-10-22.
  • The relay had no limit on how many bytes it held, so one sender could fill its memory. The relay
    kept at most 512 messages per room and 512 rooms, with messages up to 128 KiB, and counted nothing
    in bytes. At those bounds that is 32 GiB. A single address at the per-address request limit could
    post about 37 MiB every ten seconds, so it could reach a few GiB in minutes (this is arithmetic
    from the limits, not a run against the hosted relay). When a process runs out of memory it is
    killed and restarted, and every room is lost, so a ceremony in progress would stop. Separately, a
    poll of a room full of large messages built its whole answer while holding the relay's one lock,
    so every other request waited behind it (measured at 2.1 seconds in a test build, not measured in
    a release build). It affected whether the relay answered, not what it carries: nothing in this
    gave access to the messages. The code had no byte bound from 2026-07-15, the day the relay was
    added, which is the earliest it could have been exposed. We have no report that it was used; we
    did not check. What a member can notice: nothing, unless a ceremony stopped with no message to
    explain it. The relay now holds at most 4 MiB per room and 256 MiB in all, drops the oldest
    messages first, and builds a poll's answer outside the lock. The hosted relay has run this since
    2026-10-19 22:48 BRT (2026-10-20 01:48 UTC).

Changed

  • Groundwork for the next network upgrade (NU7): one table, one margin, and almost nothing different
    for you until the upgrade's height is known.
    The device (the browser signer) and the coordinator each
    decided when an earlier transaction could no longer be mined from its expiry alone, each with its
    own copy of a ten-block margin, and the device picked the network upgrade it was in from its clock.
    Neither knew that a transaction signed for one upgrade cannot be mined once the next one is active,
    so across an upgrade a payment would have been held until its expiry (days or weeks for a
    cross-pool one) and the device would have warned "this pays twice" about a transaction the network
    already refuses. Both now read the same rules from konclave-seal::epochs: the table of upgrades,
    the margin (one copy), the height estimate and the header acceptance. A source scan fails the build
    if a second copy appears. The table has one row today (NU6.3), and the window of heights, the lapse
    times and the verdict on a sent payment are the numbers they were (pinned by tests written against
    the old code). The new rules act on an upgrade only once a row for NU7 is added, which is done when
    its activation height is decided. Three things are different today, and are named so that "nothing
    changed" is not read as more than it is: a device record with no expiry (none can be written, a
    header with no expiry is refused) used to count as dead at once and now never stops counting; the
    coordinator already refuses to sign a transaction its engine built on any branch but NU6.3 (it
    could not happen, and says "the coordinator has not been updated for the network upgrade"); and a
    header whose expiry is below the NU6.3 activation block, with the clock in the first days after
    it, is now refused as being for another upgrade (not reachable now). The device's record of what
    it signed keeps the consensus branch it read from the transaction; an older entry without one reads
    as NU6.3 (true of every one, an older bundle cannot sign a later upgrade), while the coordinator's
    record of a send with no branch is treated as UNKNOWN and is never ended by an upgrade.
    Confirmed by the maintainer: K (24 blocks: the 600 s ceremony ceiling plus 600 s to build and prove and
    600 s to be mined, at 75 s a block) and K_c (288 blocks, about six hours, for cross-pool
    payments). The wasm export approvalLapsesAt is gone: the time on the confirm now comes with the
    decision, from the same rule.
  • The commit list in the release notes reads hash [scope] description. One line per commit,
    the short hash first and linked to the commit, then the scope in brackets and the summary as written,
    with the issue number when the summary ends with one. It used to put the scope in bold and the hash at the end.
    The grouping by type (Security, Added, Fixed, and so on) is unchanged (cliff.toml).

Fixed

  • A refusal from the relay is shown in words. When the relay refused a request the coordinator
    makes (it was full, it was limiting this server, the message was too large), the operator read
    "failed to parse relay: post resp: missing field seq". It now reads the relay's reason in a
    sentence after that prefix. The hosted coordinator has run this since 2026-10-22.

Known limits

  • This change alone does not deliver the cap on a transaction's life. It stays off until the
    engine move (the next commit) adds the konclave-signer next-upgrade command: the coordinator
    turns the cap on only when its table and the engine agree on the next upgrade's height, and the
    engine that runs today reports none (the boot log says so). Until then a cross-pool payment signed
    for NU6.3 and caught by the upgrade is held until its own expiry, as before. Turning it on also
    needs the PR that adds the NU7 row to make the verdict depend on the branch the chain reports, not
    only on the two copies of our own constant (#638).
  • For about a day after the upgrade the device and the coordinator disagree. The device judges by
    its own clock with a margin that is about a day wide at that distance from its anchor, so it keeps
    asking "this could pay twice" while the coordinator, which sees the chain, already says the payment
    needs a new signature. The release that re-anchors the clock after the upgrade shortens this.
  • The local bridge sends without the upgrade rules. The desktop's local bridge (orchestrate_send)
    builds a payment with no build hold, no check of the branch the engine built, and no record of the
    attempt, so the protections above cover the hosted coordinator only. It is a path an operator drives
    at a terminal, and it is on the list for the change that adds NU7's row (#638).
  • The rules are inert until NU7's row exists. The wallet reports the tip it
    knows and not the height it scanned, so a transaction is called dead because of an upgrade only
    100 blocks after the margin. A cross-pool payment caught by the upgrade keeps its funds reserved
    in the wallet until its own expiry, which can be weeks away; the build hold exists to keep new
    ones out of that window.
  • The address the relay and the coordinator count a request under is only as good as the
    forwarding header they are given.
    Both take it from the first hop of X-Forwarded-For. Where the
    platform in front of one of them keeps a value the client sent, a client can still choose which
    address its requests count under, and the fix under Security reaches no further than that. Tracked
    in #400.

Commits

Added

  • 9ce5c9a [seal] the network-upgrade rules in one module, with the NU7 row absent (#638)
  • 5e3a033 [orchestrator] the epoch cap in settle, the build hold, the branch check and the boot check, all inert until the next row lands (#638)
  • 6507b59 [wasm] Second names whether an upgrade ends the earlier transaction and no screen calls the lapse time with an expiry alone (#638)
  • 05841e9 [orchestrator] upgrade refusals reach the screen with a reason and no 'may be on the network' state (#638)
  • fd8a390 [ui] the signed record keeps the branch the device read, Second carries its own time, three upgrade texts as drafts (#638)
  • 8fcc406 [relay] bound the relay's memory by bytes (#400)

Fixed

  • 5b806c6 [ui] the site names the 0.13.2 desktop pre-release
  • 155a1dd [orchestrator] an unknown branch is never capped, a stale proposal is not told to sign again, refusals carry a stable reason, the boot check has a deadline (#638)
  • 11cf0dd [wasm] Second gives no time when an earlier record never stops counting, names the last block, and the branch test uses a non-NU6.3 header (#638)
  • 81e16e6 [ui] the upgrade endings are decided by the coordinator's reason, the record refuses a missing branch, the confirm names the last block or says the time is unknown (#638)
  • ace8ea0 [orchestrator] the stdin of the boot check is tested whatever the harness has, held funds follow the wallet's rule, a trap test guards the one-row table (#638)
  • af81dca [relay] scope the request-limit keys to the caller's address (#400)
  • a34fcd5 [orchestrator] show the relay's refusal in words (#400)

Refactored

  • 2de9f7f [wasm] the device reads its epoch rules from konclave-seal and keeps the branch of what it signed (#638)

Documentation

  • 503edd8 [changelog] the epoch rules are in place and inert until the next upgrade's row lands (#638)
  • 93c6238 [changelog] say what changed today, that the cap needs the next-upgrade command, and the day of disagreement after the upgrade (#638)
  • 882ba91 note the request-limit key change in the changelogs
  • dd5463f [relay] classify the request-limit key change as a security fix (#400)
  • 3c7a0ec [relay] say the coordinator never had the per-sender budget and name both servers in the known limit (#400)
  • 5156597 note the relay byte ceilings in the changelogs
  • 8872fda [relay] say the byte ceilings have run on the hosted relay since 2026-10-19 (#400)
  • 2117411 [ui] the public roadmap is current as of 18 October, with the median proving times and NU7 alongside the steps (#638)
  • 713c2e3 [relay] classify the byte ceilings as a security fix and say what was open (#400)
  • d8ff0e0 [relay] fix three points of the byte-ceiling security entry (#400)

Maintenance

  • 8c97473 [engine] record the v0.13.2 coordinator deploy in the lock
  • 054679c [wasm] pin the device's height window, lapse times and decision turn as they are today (#638)
  • 435b60d [seal] a source scan keeps the margin and the epoch table in one home (#638)
  • 6332f1f [wasm] rebuild the committed wasm with the epoch rules and the branch in the record (#638)
  • ee5df2b [release] the release-notes commit list reads hash [scope] description, with the issue when the summary has one
  • 849f9b4 [engine] record the 0.14.0 coordinator deploy in the lock
  • 5d28005 [release] v0.14.0

Desktop installers (Windows / macOS / Linux) are attached below.