Draft-05 and sunlight mirror support - #10
Merged
Merged
Conversation
Draft-05 changes no Merkle-tree, signature, certificate, or checkpoint bytes: TBSCertificateLogEntry, MTCProof, CosignedMessage, the checkpoint and log-origin encoding, the OIDs, and the ID derivation rules are all unchanged. Three things do change. maxSerial (§5.5) is the one breaking wire change. It is a required field appended to the MTCCertificationAuthority SEQUENCE, so -04 and -05 encodings are mutually unparseable and any issued CA certificate has to be reissued. §7.1 correspondingly revokes [0, minSerial) *and* [maxSerial+1, 2^64). That upper range is not representable as a half-open uint64 range, so RevokedRange is now closed rather than half-open. Clamping the exclusive end to 2^64-1 would have silently left the largest serial unrevoked, which no existing test would have caught; TestInitialRevokedRangesUpperBound pins it. Note a zero-valued MTCCertificationAuthority now revokes almost everything, which is the fail-closed direction. The ACME link relation "enhancement" becomes "acme-optional-alternate", registered in the new §9.1. That section standardises exactly what 314dc76 arrived at independently — an optional alternate that returns 202 + Retry-After and must not block issuance — so only the relation string changes; the status code, headers, and media type stay put. (draft-04 had specified 503 here.) §7.2 step 2 now requires verification to fail if signatureValue is not a whole number of octets. encoding/asn1 accepts a non-zero unused-bits count and reports it via BitLength, so SplitCertificate has to reject it explicitly rather than parse a proof out of a partially-used final byte. Also adds the four Appendix C "accumulated" test vectors, which are new in -05. Each is one rolling SHA-256 over the formatted output of every valid input for trees up to size 130, covering subtree hashes, subtree inclusion proofs, subtree consistency proofs, and covering-subtree selection. They pass unmodified, which independently confirms the §4 algorithms in tlogx — including that FindSubtrees followed the reference Python rather than -04's buggy prose, one of the two things -05's change log calls out as fixed. Section-number citations throughout the tree still refer to -04 and are renumbered separately; -05 renumbers §6.x and inserts a new Appendix B.2. Claude-Session: https://claude.ai/code/session_01LgyCGiwxZCkSXxgy3PqfPC
cactus no longer acts as a mirror or witness. Mirroring becomes push-based against an external c2sp.org/tlog-mirror (Sunlight), which inverts the relationship: a tlog-mirror never fetches from the log, so the log has to push to it. The pull follower and the /sign-subtree server we exposed are the wrong half of that protocol. The two halves went together. mirror/server.go signed subtrees of the log copy that mirror/follower.go had fetched and verified, so keeping the server without the follower would leave it attesting to a log it no longer independently checks — worse than not having it. Both go, along with the mirror.* config section, the sign-subtree listener, the four mirror-mode metrics, and config-witness-example.json. What stays is the CA *client* side: ca_cosigner_quorum.mirrors[] and cert.RequestCosignatures still request cosignatures from external mirrors. That is the direction cactus keeps. The quorum tests previously stood up real followers and servers to answer sign-subtree. They now use an in-package stub witness that parses just the subtree range and hash and returns one ML-DSA-44 cosignature line. Both tests came out stronger: the three-cosigner test had a no-op verification loop that only counted recognised cosigner IDs and now runs a real cert.VerifyMTCSignature on each, and the quorum test now also asserts that the slow witness is absent from the collected signatures while still having been contacted, which is what makes "fan out to all, wait only for the minimum" observable rather than merely plausible. Coverage genuinely lost with the feature: independent tree rebuild from tiles, follower restart resume, halt-on-bad-key, and the require_ca_signature_on_subtree DoS gate. Those tested the deleted code. docs/test-instance.md was premised on one process being both CA and witness and pointed at the deleted example config, so it is trimmed to a CA-only guide rather than left dangling. Note for anyone with a deployed config: because config.Load uses DisallowUnknownFields, a leftover "mirror" block is now a load error rather than an ignored key. Claude-Session: https://claude.ai/code/session_01LgyCGiwxZCkSXxgy3PqfPC
A tlog-mirror never fetches from the log it mirrors, so replication is
the log's job: drive add-checkpoint to advance the mirror's pending
checkpoint, upload entries with add-entries, and collect the mirror's
cosignature on the resulting checkpoint. The new mirrorpush package
implements that client and cmd/cactus drives it from OnFlush. With no
mirror_push targets configured, behaviour is unchanged.
Three parts of the wire format are easy to get subtly wrong, so each has
a test that fails if the detail is reversed:
- An add-entries package proves the subtree [rounded_start + i*256,
end), not [start, end). When upload_start is unaligned the client
transmits only its own entries but must prove the whole 256-aligned
bundle; the mirror reconstructs the leading entries from its own
storage. Proving [start, end) instead yields either an invalid
subtree or a valid-but-wrong one that 422s.
- The recovery ticket is raw bytes in the request and base64 in the
202/409 response. It is decoded on receipt and stays raw thereafter.
- add-entries and sign-subtree use the MTC §4.4 *subtree* consistency
proof, while add-checkpoint uses an RFC 6962 *tree* consistency
proof. These are different algorithms over the same tree, so the two
are now separate named methods rather than one ambiguous one.
Also fixes two ways the existing sign-subtree client violated the
current c2sp.org/tlog-witness spec, both of which a strict mirror
rejects outright:
- It emitted a CA subtree cosignature line. C2SP deleted subtree
cosignature lines from the grammar; the line is not base64 and fails
to parse as a consistency proof hash, giving 400. Removed, along
with the now-dead CASignature/CACosignerID/CACosignerKey fields.
- It presented a CA-only-signed checkpoint. The spec now requires the
reference checkpoint to carry the *responding mirror's own*
cosignature or the request is refused with 403. That cosignature
only ever comes from an add-entries 200 response, which is why the
push client has to exist before sign-subtree can work at all: the
pool now feeds its retained cosigned checkpoint back in.
Since a signed note carries many signature lines and each cosigner
ignores lines that are not its own, every mirror's cosignature is
appended to a single note, so the CA still fans one body out to all
mirrors rather than building one per mirror.
State advancement is deliberately conservative: the next entry index
moves only on a 200 or a value the mirror explicitly advertised, a 202
that fails to advance is an error rather than a spin, and the ticket is
persisted with the pending size it belongs to. 422 is treated as an
integrity signal — a proof failed to verify — so it is fatal rather than
retried, as is 400.
Supporting additions: cert.MarshalSignatureInputAt takes a timestamp
(checkpoint cosignatures require a non-zero one, subtree cosignatures
require zero); TileWriter.ReadEntries and Log.Entries read entries back
out for upload; Log.TreeConsistencyProof supplies the RFC 6962 proof.
FuzzAddEntriesFraming found a builder/parser asymmetry (the parser
accepted a zero-length origin the builder refuses) and an unbounded
canonical sequence: header fields are uint64, so a corrupt header can
describe 2^55 packages. Parsing is now lazy and bounded by bytes
received. Regression corpus committed.
Claude-Session: https://claude.ai/code/session_01LgyCGiwxZCkSXxgy3PqfPC
The two sides of a mirroring relationship want the same key in different encodings. A tlog-mirror's log list identifies our log by a c2sp.org/signed-note vkey; our own mirror config wants a PEM public key. Previously only the PEM direction existed, so wiring the two together meant hand-assembling a vkey. -vkey emits one, -from-vkey converts one back. The vkey name is the COSIGNER's OID name, not the log's, and -from-vkey recomputes the key ID from the name and key rather than trusting the one in the string. Both details are load-bearing. A cactus checkpoint's origin line is the log ID (<CA-ID>.0.<log number>) but its signature line is labelled with the cosigner name, which is the CA ID (§5.4); a verifier matches on that name, so a vkey generated under the log ID is silently unmatchable and the mirror answers add-checkpoint with a bare 403. Recomputing the key ID likewise turns a truncated or mistyped vkey into a startup error instead of a cosigner whose signatures never match anything. The flag is named -cosigner-id to make the distinction hard to get wrong, and consumers are expected to carry the log origin separately (sunlight's log list has an `origin` line for exactly this). Note the vkey's key material is standard base64, whose alphabet includes '+', so parsing splits on only the first two separators. Claude-Session: https://claude.ai/code/session_01LgyCGiwxZCkSXxgy3PqfPC
`make docker-up` brings up cactus, Sunlight as a c2sp.org/tlog-mirror
and ML-DSA-44 cosigner, and skylight serving the mirror's read side.
Verified end to end: cactus pushes its log, Sunlight mirrors the entries
and cosigns the checkpoint, and an issued certificate carrying that
cosignature passes full §7.2 verification.
The cactus image does not build from source. cactus needs Go 1.27 for
crypto/mldsa and no golang:1.27 image exists yet — a released Go cannot
even download the toolchain — so the binaries are cross-built on the
host with gotip and copied in. Sunlight has no such constraint (go 1.25,
filippo.io/mldsa rather than the standard library) and builds in its own
image, but does need cgo: its checkpoint lock backend is
crawshaw.io/sqlite, and CGO_ENABLED=0 builds only sunlight-keygen.
Four things about running Sunlight as a mirror cost real debugging time
and are documented where they bite:
- submissionprefix MUST NOT carry a port. Sunlight registers its
handler on net/http.ServeMux as "<host>/", and ServeMux strips the
port from the request Host before matching, so a pattern containing
one matches nothing and every endpoint 404s with nothing logged. A
portless pattern matches any port. The https scheme is still
required — Sunlight rejects anything else — but is never used for
transport, since without an acme section it serves plain HTTP/h2c.
skylight has the same trap.
- Sunlight cannot run as a mirror alone. main() ends on
sequencerGroup.Wait(), and sequencer goroutines are started only per
entry in `logs:`; with none, Wait returns immediately and the process
shuts down and exits 1 without logging anything, despite the
witness/mirror machinery being fully initialised. A throwaway
keepalive CT log we never submit to keeps it alive. That log needs
notafterstart/notafterlimit, and an explicit empty roots file —
otherwise CCADBRoots defaults to "trusted" and it fetches the CCADB
root list over the network at startup. Its inception date must be
today when first created, so the config is rendered at init time.
- A log listed in both loglists and mirrorloglists logs a single
"adding new log" line with mirror=false; the mirror upgrade path
sets the flag and continues silently. mirror=false in the log is not
evidence the mirror is unregistered — probe add-entries instead,
where 422 means known and 404 means unknown.
- Sunlight refuses to create its own checkpoint lock table, as a guard
against pointing a log at the wrong backend, so init creates it.
Key exchange is three init containers because neither side can start
without the other's public key: cactus exports its log vkey, Sunlight
derives its keys and builds the log list, then the vkey is converted
back to the PEM form cactus consumes. All three are idempotent, so
bringing the stack up twice does not rotate keys out from under a log
that has already been mirrored; docker-down deletes the volumes and
therefore the keys.
Claude-Session: https://claude.ai/code/session_01LgyCGiwxZCkSXxgy3PqfPC
Follows the code changes across the last four commits. The draft-05 rename of the ACME link relation had left stale references behind: README, MTC.md, and test-instance.md all still described `rel="enhancement"`, which no longer exists. They now say `acme-optional-alternate` and cite §9.1, which is also a better explanation than the old text — the non-blocking behaviour is now specified rather than something cactus decided on its own. test-instance.md was trimmed to a CA-only guide when the mirror mode was deleted, because its premise (one process being both CA and witness) had stopped being true. Now that there is a working mirror again it points at the compose stack for that case, rather than describing a by-hand setup that would mean reimplementing log replication. Also corrects two things that were quietly wrong. CLAUDE.md said integration tests live behind the `integration` build tag; there are no build tags in that package, so plain `gotip test ./...` has always run them and `make integration`'s `-tags=integration` selects nothing — it is really just the `-race -count=1` runner. And landmark sequences are §6.4 in draft-05, not §6.3. The remaining ~330 in-code §x.y citations still refer to draft-04 numbering and are a separate mechanical pass. Two hazards for whoever does it: draft-05 inserts a new Appendix B.2, so stale appendix references point at a real but wrong section rather than dangling; and `acme/` mixes MTC and RFC 8555 citations whose numbers collide (§6.3 is Landmark-Relative Certificates in one and POST-as-GET in the other), so it cannot be swept mechanically. Claude-Session: https://claude.ai/code/session_01LgyCGiwxZCkSXxgy3PqfPC
TestBulkIssuanceStress issues 800 certificates concurrently through the full ACME flow, verifies each against the log, and then checks the log is internally consistent. Behind the `stress` build tag so it stays out of the default suite; size and concurrency are tunable via CACTUS_STRESS_CERTS / CACTUS_STRESS_CONCURRENCY, so the same test serves as a quick local run and a heavier soak. 800 is past both the 256-entry tile boundary and the existing 100-certificate TestParallelIssuance, so it exercises multi-tile growth and partial tiles under concurrent load. The load is not really the point — the post-conditions are. The log is a single writer with no cross-process coordination, so a sequencing bug surfaces as a duplicated or skipped index rather than as an error any individual request would observe. Asserting only that every request returned 200, or only that the tree ends at the right size, would miss two entries swapping or an index being handed out twice while the count stays correct. So the test requires the assigned indices to be exactly a permutation of [0, n). It also reloads the published entry tiles, replays them through tlog.StoredHashes, and compares the recomputed root against the signed checkpoint. The per-certificate inclusion proofs cannot catch a tile written inconsistently under concurrent flushes, because they are served from the same in-memory hashes the checkpoint was built from. Measured: 800 certs in ~0.6s (~1300/sec) plain, ~1.5s under -race, and 5000 certs in ~3.1s. Clean under the race detector. I checked the permutation assertion is not vacuous by mutating certLogIndex to collide two indices; it fails with "log index 0 was assigned to 2 certificates" and "log index 99 was never assigned". Also fixes two stale claims in the docs found while writing this: the integration package has no build tags (so `gotip test ./...` has always run it, and `make integration`'s -tags=integration selects nothing), and README cited integration.TestRelyingPartyFastPath, which does not exist — the landmark fast-path case is TestLandmarkRelativeCertConstruction. Claude-Session: https://claude.ai/code/session_01LgyCGiwxZCkSXxgy3PqfPC
Runs integration.TestBulkIssuanceStress under -race on a schedule rather than on every push. The test takes seconds, so cost is not the reason — its value is in repetition. A sequencing race that survives one run may not survive fifty, and the assertions it makes about log index assignment are exactly the kind that fail intermittently or not at all. Running it nightly samples that; running it once per push would not, and would slow CI for little added signal. It also runs on pull requests that touch the issuance path (log, ca, acme, cert, tlogx), since that is where a regression it could catch would originate, and on workflow_dispatch with overridable size and concurrency for a deliberate soak. Scheduled and pull_request runs pass empty inputs, which the test reads as "use the defaults" (800 / 64). CI proper gains a vet step for the stress-tagged file. It is behind a build tag, so nothing in the normal suite compiles it, and without this a break would go unnoticed until the next nightly run — landing on whoever reads that failure rather than on the pull request that caused it. Claude-Session: https://claude.ai/code/session_01LgyCGiwxZCkSXxgy3PqfPC
Comment-only, plus one user-facing validation string.
Far less moves than the citation count suggests. Sections 4, 5, 7 and 8
are identical between the drafts, and 9, 12 and 13 only gain new
subsections without renumbering existing ones. Two families shift:
§6.x draft-05 inserts a new §6.1 "Certificate Inputs", pushing
Certificate Format to §6.2, Standalone Certificates to §6.3,
Landmark-Relative Certificates to §6.4 (and its subsections to
§6.4.x), and Size Estimates to §6.5.
App B draft-05 inserts a new B.2 "Subtrees", pushing Inclusion Proof
Evaluation to B.3, Consistency Proof Structure to B.4, and
Consistency Proof Verification to B.5.
Done in a single pass per line, so a citation cannot shift twice — a
naive sequence of replacements turns §6.1 into §6.2 and then §6.2 into
§6.3.
Three things made this not purely mechanical.
Appendix B fails silently rather than loudly. Because -05 *inserts* B.2
rather than appending, a stale "Appendix B.2" on ProofLen still names a
real appendix — just "Subtrees" instead of "Inclusion Proof Evaluation".
Nothing would have flagged it. Same for the two §B.3 consistency-proof
references, which are Consistency Proof Structure and are now §B.4.
(An earlier reading of these had them down as RFC 9162 citations to
leave alone; the cited content matches the MTC appendices exactly, and
RFC 9162 has no subtree proofs in this form.)
§6.x collides with two other specs cited in the same files, sometimes
within a few lines. RFC 8555 numbers its own §6.2 (POST content type),
§6.3 (POST-as-GET), §6.4 (url header) and §6.5 (nonces) — acme/handler.go
cites MTC §6.1/§6.3/§6.3.1/§6.3.2 alongside all four. trust-anchor-ids
has a §6/§6.1 for the property-list PEM layout. Both are excluded by
explicit file:line rather than by pattern, since no pattern distinguishes
them.
Five sites were already mis-cited and a mechanical shift would have
preserved the error under a fresh-looking number. cert/landmark.go and
cmd/cactus-cli/cert.go cite §6.3.3 while describing how a
landmark-relative certificate is *constructed*; §6.3.3 is "Publishing
Landmarks" and construction is §6.3.4. They are corrected to §6.4.4, not
the §6.4.3 a shift would have produced.
Also drops a phantom citation: integration/multi_mirror_test.go referred
to "§6.2/9.4", and §9 has no subsections in draft-04 and only 9.1/9.2 in
draft-05.
The `draft-04` labels become `draft-05` except in cert/cacert_test.go,
where they deliberately name the old encoding the maxSerial test rejects.
Not touched, and worth knowing about: several §6.4.1 citations (formerly
§6.3.1, "Landmark Tree Sizes") sit on code handling the published
/landmarks list, whose format is really §6.4.3 "Publishing Landmarks".
The shift is correct; whether §6.3.1 was the right citation to begin with
is a pre-existing question across ~30 sites, and sorting it out means
judging intent case by case rather than renumbering.
Claude-Session: https://claude.ai/code/session_01LgyCGiwxZCkSXxgy3PqfPC
Two issuance-authorization gaps in the CSR validator: 1. buildExtensionsFromCSR copied the subjectAltName extension value verbatim while Validate only cross-checked dNSName/iPAddress entries against the order. A CSR whose SAN carried an rfc822Name, URI, or otherName alongside an authorized dNSName smuggled an unvalidated identity onto the leaf certificate. Reject any GeneralName type other than dNSName/iPAddress (sanContainsOnlyDNSAndIP). 2. Validate enforced only CSR ⊆ order (every CSR SAN authorized). RFC 8555 §7.4 requires the CSR to indicate the exact same set of identifiers, so also enforce order ⊆ CSR. Adds regression tests for both. Claude-Session: https://claude.ai/code/session_01LgyCGiwxZCkSXxgy3PqfPC
Several RFC 8555 §6.3 access-control gaps in the ACME handlers: - handleOrder / handleAuthz / handleChallenge fetched objects by ID and returned them without checking the authenticated account. Any account holding an ID could read another account's order, authz, or challenge (and in http-01 mode drive validation of a foreign challenge with the caller's key authorization). Trace ownership order/authz/challenge → account and report a foreign resource as 404. - handleFinalize discarded the account entirely, so an account that knew a victim's ready order ID could finalize it with its own CSR (§6.3, §7.5). Bind the order to the caller before the ready→processing claim. - §11.1: the CSR public key was never compared to the account key. Reject a CSR whose key equals the account key. Also: emit Content-Type: application/json on new-account/new-order responses, add the §7.1 rel="index" Link relation on resource responses, and exclude invalid orders from the account orders list (§7.1.2.1). Adds cross-account and account-key-CSR regression tests. Claude-Session: https://claude.ai/code/session_01LgyCGiwxZCkSXxgy3PqfPC
The §7.2 verifier accepted the optional TBSCertificate tail fields (issuerUniqueID [1], subjectUniqueID [2], extensions [3]) in any order and canonicalized them into a fixed-order log entry. Two byte-distinct certificates — one valid DER, one with the tail tags swapped (invalid DER per X.690 SEQUENCE ordering) — therefore rebuilt to identical entry contents and both satisfied the same inclusion proof and cosignatures, a draft-05 §12.6 certificate-malleability violation. Enforce strictly ascending tail tags. Adds a regression test. Claude-Session: https://claude.ai/code/session_01LgyCGiwxZCkSXxgy3PqfPC
VerifyCertificate checked only the outer Certificate.signatureAlgorithm. draft-05 §7.2 step 1 refers to the TBSCertificate's signature field (the inner AlgorithmIdentifier, which RFC 5280 §4.1.1.2 requires to equal the outer one). RebuildLogEntryFromTBS parsed that field only to discard it. Validate it as id-alg-mtcProof with absent parameters so both the inner and outer algorithm IDs are checked. Claude-Session: https://claude.ai/code/session_01LgyCGiwxZCkSXxgy3PqfPC
discover() seeded next-entry from the mirror's advertised checkpoint size via an unauthenticated GET (no signature verified). A forged or buggy oversized value made every pushEntries take the "mirror is ahead" bail-out, which returned nil with no log line — so add-entries was never sent, the mirror-info correction loop never ran, and the mirror silently dropped out of cosignature collection until process restart. Clamp the discovered seed to our own checkpoint size (a genuinely-ahead mirror is corrected by the authenticated add-entries loop) and log a warning both when clamping and in the ahead-of-us bail-out. Adds a regression test. Claude-Session: https://claude.ai/code/session_01LgyCGiwxZCkSXxgy3PqfPC
Three smaller correctness fixes in the push client: - The add-checkpoint 200 path advanced pendingSize without clearing the ticket, so the next add-entries sent a ticket issued for the previous size paired with the new upload_end. Clear the ticket on advance. - do() read exactly maxResponseBytes via LimitReader, so a truncated oversized body was indistinguishable from a complete one and could be escalated as a fatal mirror fault. Read one byte past the cap and return a distinct "response too large" error. - VerifyCosignatures accepted any timestamp; reject values above 2^63-1 (tlog-cosignature MUST) rather than retaining and republishing them. Claude-Session: https://claude.ai/code/session_01LgyCGiwxZCkSXxgy3PqfPC
- log.pool_size was parsed, defaulted, and validated but never used. Wire it as log.Config.MaxPoolSize: Append signals the sequencer for an early flush once the pool reaches it, bounding pool memory and issuance latency. main passes cfg.Log.PoolSize. - log.shortname was required but unused; make it optional and attach it to the structured logger context. - A listener that failed to bind was only logged while the process kept running with a dead endpoint. startServer now reports on an error channel and run() shuts down and returns the error. - acme.tls_cert / tls_key: require both or neither (a lone one silently served plaintext), and resolve their paths relative to data_dir like every other configured path. - ca_cosigner_quorum.mirrors[].algorithm was only checked non-empty; require mldsa-44 to match the profile (and mirrorpush.New). - cactus-keygen -vkey hardcodes the ML-DSA-44 signature-type byte; reject -alg mldsa-65/87 instead of emitting a mislabeled vkey. Adds config and log regression tests. Claude-Session: https://claude.ai/code/session_01LgyCGiwxZCkSXxgy3PqfPC
Documentation only, no behavior change: - README/CLAUDE.md: the ca_cosigner.id example used the absolute OID form; cactus stores trust anchor IDs relative to the 1.3.6.1.4.1 base, so use 44363.47.1.99. Document shortname, pool_size, tls_cert/tls_key, gate the pprof note on a loopback metrics listener, and fix the prove example's missing log-number prefix. - docs/disk-layout.md: drop the nonexistent log/subtrees/ directory and the false O_EXCL tile-write claim; add state/landmarks/sequence.jsonl and mirrorpush/. docs/threat-model.md: the temp file is not fsynced either. storage.Put doc: the exclusive path is not reader-atomic. - MTC.md: checkpoint_period_ms (not flush_period_ms); the mirror role lives in mirrorpush/, not a mirror/ package; per-subtree signatures are in-memory only; "seven" indices, not "three". - docker/README.md: cactus-keygen -cosigner-id (not -log-id); init step counts are 1/2/3-of-3. landmark handler: the list format is §6.4.3. Makefile .PHONY: vet exists, lint does not. Bump remaining draft-04 doc comments to draft-05. Claude-Session: https://claude.ai/code/session_01LgyCGiwxZCkSXxgy3PqfPC
Non-functional tidy-ups: - cactus-cli: remove the hand-rolled min() (Go builtin) and the hexs() reimplementation of hex.EncodeToString. - cmd/cactus: drop the pointless pemDecode wrapper; call pem.Decode. - mirrorpush: pushEntries never used its note parameter; remove it. Claude-Session: https://claude.ai/code/session_01LgyCGiwxZCkSXxgy3PqfPC
flush() derived the covering-subtree range from the tile writer's pre-append size. When a flush advanced the tile writer (persisting the treeSize file) but failed before committing the checkpoint — or a crash landed in that window, which loadCheckpoint tolerates (tw.Size() may exceed the checkpoint size) — the next flush saw newSize == tree size and minted no covering subtree for the gap entries. Those entries became committed with no signed subtree, so buildIssued fell through to the signature-less whole-tree path and issuance for them failed (the issuer rejects a zero-signature Issued; quorum mode hung until the deadline). Cover [committedSize, newSize) using the last committed checkpoint size, which re-covers the gap while remaining identical in the steady state, and gate the no-op early return on the actual tree size. Also drop the now-unused collectMirrorSigs parameters. Adds a regression test that injects a checkpoint-write failure and confirms the recovery flush produces a signed covering subtree. Claude-Session: https://claude.ai/code/session_01LgyCGiwxZCkSXxgy3PqfPC
The §7.1.2.1 invalid-order filter only skipped when the order still existed; also skip an ID whose order is gone. Currently unreachable (OrderIDsForAccount only returns existing orders) but the guard now reads as intended. Claude-Session: https://claude.ai/code/session_01LgyCGiwxZCkSXxgy3PqfPC
Go 1.27 release candidates are out, so CI no longer needs to bootstrap and cache a gotip devel toolchain. Both workflows now install go1.27rc2 with a single actions/setup-go, which also handles the module and build caches -- dropping the separate gotip SDK and Go cache steps. The version is pinned in a GO_VERSION env var in each workflow so moving to a later RC, or to '1.27' once 1.27.0 ships, is a one-line edit. GOTOOLCHAIN=local stays: the RC satisfies the `go 1.27` go.mod requirement on its own. The Makefile still defaults to GO ?= gotip for local development, so the CI steps pass GO=go explicitly. Verified against go1.27rc2 locally: gofmt, vet, build, `go test -race ./...`, the integration suite, the stress-tagged vet, and `make stress` all pass. Claude-Session: https://claude.ai/code/session_01LgyCGiwxZCkSXxgy3PqfPC
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
1 participant
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
This updates to draft-05, and switches to using external mirrors.
Add integration testing with sunlight.
Various bugfixes from review.
this is all vibecoded.