Skip to content

Stewardship log while the maintainer is away #131

Open

Description

@PastaPastaPasta

The maintainer (kwvg) is away. Until he is back, pull requests in this repository are reviewed and merged by a steward: an AI agent (Claude) acting for PastaPastaPasta. Merges appear under that account, and each one has a comment below. This issue is the record for kwvg to read on his return.

Rules

  • The steward reviews and merges. It writes no feature code and does not push to PR branches. Authors never approve or merge their own PRs.
  • The standard is kwvg's reviews on sdk%fix: range-check decoded lengths, apply Dash Core's decode-time caps #57 to primitives%fix!: check operator keys under the payload's BLS scheme #129, applied as written. A finding he raised once is a blocker the second time. His reviews are never dismissed or edited.
  • Class A: makes the SDK accept, refuse, decode or encode what the reference implementation does, inside existing types and APIs. Steward decides.
  • Class B: a new typed message, payload or primitive that a named consumer needs, in the style of the existing ones. Steward decides.
  • Class C: a new public abstraction (trait, macro, wrapper, feature flag, crate, dependency) or a reshaping of existing types. Needs a written proposal and a unanimous council of three models (Claude Fable 5.2, Claude Opus 5.5 and GPT 6.1 Sol). So does any change to operative crypto in dash-pkc, a PR he never reviewed that adds more than about 300 lines of library source, and a thread of his that the author disputes. If the council is not unanimous, it does not merge and the question is listed below.
  • Class D: conflicts with one of his rulings (state or tunables, devnets or builders, a BLS scheme chosen at run time, a consensus-level choice made inside the SDK, check-order parity, debugging allowances, handler policy, V1 transport). Nothing in class D merges.
  • Before a merge: just comb passes on every commit, CI is green on the exact head including the Codecov patch checks, and a model from a different family than the author has reviewed the diff. Merges are made with contrib/git_merge.py.
  • The steward does not publish, tag or bump versions, does not change AGENTS.md, the style guide, the PR template, lint or CI policy or the MSRV, and does not force-push develop.
  • To undo a merge: git revert -m 1 <merge sha>. Each entry below gives the command.

Questions for kwvg

  1. Operator signatures on legacy payloads. Signable::verify on ProUpServTx and ProUpRevTx picks the BLS scheme from the payload version (OperatorKey::verify; it was verify_operator in primitives%feat!: decode v24 special transactions as the reference implementation does #88, and primitives%fix!: check operator keys under the payload's BLS scheme #129 moved it without changing the rule). The reference implementation verifies under the scheme active at the block (CheckHashSig, VerifyInsecure). A version 1 ProUpServTx mined on mainnet after v19 (for example at block 1991126) is legacy-encoded and signed under the basic scheme, and verify refuses it. The same happens when the masternode's stored key is legacy-encoded and the payload is version 2 or 3. Found while reviewing primitives%feat!: add payload sign hashes, read signatures of any length #121, whose corpus holds such transactions. Not changed by any merged or open PR. The operator key itself is read under the payload version in the reference, so the key check added in primitives%fix!: check operator keys under the payload's BLS scheme #129 does not depend on this. Question: how should the caller name the scheme here? Until answered, any proposal goes to a council as class C.

  2. Check-table rows that pin which error is reported first (case::x_before_y). develop has them in several modules (proregtx.rs, proupsharetx.rs, netinfo.rs), and primitives%fix: check net info entries after purposes and the reward #109 says the order of checks is not something to match. Question: are new rows of this kind wanted? Until answered, new PRs add none, and the ones on develop are left alone.

  3. The lockfile pins dtolnay/rust-toolchain@stable, a branch that upstream rebuilds on top of master, so the pinned commit leaves the ref and CI fails until the pin is refreshed (sdk%chore(ci): refresh actions.lock after rust-toolchain@stable moved #133). Question: keep the branch ref and refresh it with --accept-moved when it breaks, or pin a ref that does not move? Until answered, the steward refreshes it with your recipe, as in sdk%chore(ci): refresh actions.lock after rust-toolchain@stable moved #133.

Activity

  1. PastaPastaPasta commented on Nov 14, 2026

    @PastaPastaPasta
    MemberAuthor
    #126 primitives%fix!: refuse governance object signatures over 96 bytes  merged 2026-11-14 at ed326129
    class: A   reviewers: GPT 6.1 Sol, high
    decided by: steward
    kwvg's review: reviewed, changes made since
    his open threads: three, all settled in the code (below)
    judgment calls: three (below)
    revert: git revert -m 1 f480538feb4e6f5b3e2a0221a97e773fc5a25a4c
    

    His threads, all from the review at 7b552ac:

    • Shorten the GovObject doc and drop the layout block: the doc is "Governance object." and the block is gone.
    • The tautological test: the GovObject-level tests are gone. codec_base_bounds_a_limited_field in codec.rs checks the field: limit form on a stand-in struct, with an empty field, a field at the limit and a count over it.
    • No changelog entry for an unreleased crate: no commit touches CHANGELOG.md. The wording is in the description's "Breaking Changes".

    The threads are left unresolved on GitHub and the review is not dismissed, so both read as he left them.

    Judgment calls:

    1. codec_base! is exported and gains a field: limit form. I classed that as A, not as a new public abstraction. It only passes the limit on to impl_vec!'s form from p2p_core%feat!: bound inventory, locator and address counts #119, whose description named this field as the follow-up, and his review of this PR asked for a test of the form and did not question it. He did not see the final head.
    2. No test or corpus row decodes a GovObject with a 97-byte signature. The bound on GovObject itself rests on the macro test plus the one line sig: Self::MAX_SIGNATURE_SIZE. That is what his thread asked for; the corpora hold signatures of 0 and 96 bytes.
    3. The history was folded from six commits to two after his review (the macro form with its test, then the bound). The tree is identical to 8ad5a62, the head that answered his threads.

    Proof: just comb passed on both commits, run by the steward. CI and every Codecov patch and project check were green on ed32612. The independent review had no findings.


    🤖 Posted autonomously by Claude on behalf of pasta.

  2. PastaPastaPasta commented on Nov 14, 2026

    @PastaPastaPasta
    MemberAuthor
    #132 primitives%fix!: correct the names of revocation reasons 1 to 3  merged 2026-11-14 at 8b25917b
    class: A   reviewers: GPT 6.1 Sol, high
    decided by: steward
    kwvg's review: never reviewed
    his open threads: none
    judgment calls: two (below)
    revert: git revert -m 1 a4ea791e64e67d4053c6b328b43fe49df6f7f99d
    

    What it is: one commit, ten lines in pkgs/primitives/src/support.rs. RevocationReason named codes 1 to 3 KeyCompromise, ChangeOfKeys, ViolationOfService. The reference names them termination of service, compromised keys and change of keys (providertx.h#L396-L401). They are now TerminationOfService, CompromisedKeys, ChangeOfKeys. The codes, the codec and the check on an unknown reason are unchanged. The consumer is a block explorer that shows the names.

    Judgment calls:

    1. ChangeOfKeys keeps its name and moves from code 2 to code 3, so code written against the old enum compiles and means a different code. The alternative was a name that differs from the old one for code 3 as well, which would have turned the silent change into a compile error at the cost of not using the reference's name. I took the reference's names; the description's "Breaking Changes" states the move first. The independent review found that the description had missed it.
    2. No test pins the names. A review bot asked for one and the author declined it as a restatement of the table, which follows the tautological-test rule. It is also the reason this error was on develop unnoticed: nothing but a reading against the reference checks a name.

    Proof: just comb passed on the commit, run by the steward. CI and every Codecov patch and project check were green on 8b25917. The independent review's two findings were both on the description and were corrected before merge.


    🤖 Posted autonomously by Claude on behalf of pasta.

  3. PastaPastaPasta commented on Nov 14, 2026

    @PastaPastaPasta
    MemberAuthor
    #129 primitives%fix!: check operator keys under the payload's BLS scheme  merged 2026-11-14 at 1193b7e3
    class: A, with a reshaping asked for in review   reviewers: GPT 6.1 Sol, high
    decided by: steward
    kwvg's review: reviewed, changes made since
    his open threads: five, all settled in the code (below)
    judgment calls: three (below)
    revert: git revert -m 1 f748a78f34053c73602a0ba663d03e51def55eed
    

    What it is: the ProRegTx, ProUpRegTx and ProUpSharedRegTx checks refused an operator key only when it was null. They now also refuse a key that is not valid under the scheme the reference reads it with: legacy for a version 1 ProRegTx or ProUpRegTx, basic otherwise (providertx.h#L188, #L354, #L542, bls.h#L495-L515). The validity rule is dash-pkc's is_canonical; nothing in dash-pkc changes.

    His threads, from the review at 75dab82:

    • Cryptographic tests do not belong in dash-primitives, and the same tests were repeated in three structs (three threads): all removed. The check tables keep each payload's null_operator row, and one row in proregtx.rs where a version 2 payload from the chain is set to version 1.
    • The KeyEdit enum, flagged before in primitives%feat!: decode v24 special transactions as the reference implementation does #88: gone, with nothing in its place.
    • Wrap the operator key in a struct with a version-based is_valid as an inherent function, and add a commit that folds in verify_operator: OperatorKey sits beside OperatorSig, with is_valid(version). The second commit moves verify_operator into OperatorKey::verify, and ProUpServTx and ProUpRevTx take an OperatorKey as their Signable::Key.

    Judgment calls:

    1. The reviewer asked for a row showing a key that is valid under the legacy scheme refused under the basic one, for ProRegTx and for ProUpSharedRegTx. I did not ask the author for it. It needs a scheme-specific key fixture, which is the test his first thread removed as dash-pkc's to own. The corpus rows run the check on version 1 and version 2 payloads, so both branches execute, but no test in this crate shows that direction refusing.
    2. is_valid(version) picks the scheme from the payload version. The reviewer first read that as a scheme chosen at run time and withdrew it on re-check: the type carries no scheme, the choice ends in a compile-time scheme parameter as verify_operator already did, and it is the function his thread asked for.
    3. OperatorKey::verify keeps the rule that the signature scheme follows the payload version. That rule is question 1 in this issue's list and is still open. This PR moves it without changing it; the key check does not depend on it, because the reference does read the key under the payload version.

    Proof: just comb passed on both commits, run by the steward. CI and every Codecov patch and project check were green on 1193b7e. The independent review ran three times: on the first ready head, on its own findings after I disputed three of them, and on this head. The first ready head (60cedb5) went back for a test that could not fail, a duplicated row and a history fold.


    🤖 Posted autonomously by Claude on behalf of pasta.

  4. PastaPastaPasta commented on Nov 15, 2026

    @PastaPastaPasta
    MemberAuthor
    #122 p2p_core%feat!: type the BIP37 and BIP157 messages  merged 2026-11-15 at 7ce7bcf2
    class: B, with two class-C questions   reviewers: GPT 6.1 Sol, high; Claude Fable 5.2, high; Claude Opus 5.5, medium
    decided by: council, unanimous: is `BloomFlags` open or closed, and may `dash-p2p-core` depend on `bitcoin-units` for `start_height`; steward for the rest
    kwvg's review: reviewed, changes made since
    his open threads: one, settled by aligning with upstream (below)
    judgment calls: six (below)
    revert: git revert -m 1 d8cfd142dbc7518022142c44bb52ea8100387ffd
    

    What it is: the ten BIP37 and BIP157 messages, which dash-p2p-core carried as raw bytes, become typed messages, and dash-primitives gains TxMerkleNode. This undoes the compact filter regression listed in #40.

    His review at 31a4a44 offered two options: defer the PR for bitcoin-p2p-messages and undo #39, or align with upstream semantics and leave a TODO for the substitution. The second was taken. Names, fields and field order follow rust-bitcoin at 94c6f791. Each module has one TODO(pasta) naming the types upstream's can replace, PartialMerkleTree, FilterHash and FilterHeader. The description says why the others stay.

    Council (Claude Fable 5.2, Claude Opus 5.5, GPT 6.1 Sol, all high; unanimous in round 2, 2026-11-14; recorded in #122 and below):

    • BloomFlags is an open enum_map! enum with Unknown, so it reads every byte, as the reference does. Upstream's enum is closed. That is the one place this PR departs from upstream, and it is why BloomFlags and FilterLoad stay this SDK's types.
    • bitcoin-units becomes a direct dependency of dash-p2p-core, for start_height: BlockHeight in getcfilters and getcfheaders. Your earlier file (228e54c) had it, and upstream types the field the same way.

    Judgment calls:

    1. The bitcoin-units dependency. The council accepted it, but on a strict reading the "consumer is blocked without it" test for class C is not met: the crate could use its own height type.
    2. CFCheckpt is typed although no node code reads its fields yet. A reviewer first raised it under p2p_core%refac!: gate all inner values of encodable types, apply getter policy to crate, restrict make_num! behind {get,set}ters #92 and withdrew it, for three reasons: the variant was already public as bytes, Known Issues #40 lists the stubs as a regression, and your file typed it.
    3. filterload refuses a filter over 36,000 bytes, and filteradd an element over 520 bytes, while reading. The reference refuses both while deserialising (bloom.h#L76-L78, net_processing.cpp#L5584-L5589), so I treated them as the same fault at the same stage, not handler policy. The two bounds are public associated constants, FilterLoad::MAX_FILTER_SIZE and FilterAdd::MAX_DATA_SIZE, in the form addr.rs uses. The node needs them to build a filter, since new takes any length.
    4. Hash function count, tweak and filter types are not checked. The reference checks the count and the filter type after reading, and never checks the tweak.
    5. Reviewers could not verify without a node that the block hashes and filters in the corpora are what mainnet returns, or that the programs print these bytes. That rests on the author's runs. They did check by hand that both merkleblock rows hash to the merkle root of block 2,000,000, and that the cfheaders hashes and the header chain recompute from the filters in the rows.
    6. Two nits were left in contrib:
      • contrib/cfilter.py:31 lacks the two blank lines before def field. CI's lint does not flag it.
      • parse_args and main are the same twelve lines in both scripts.

    Proof: just comb passed on all three commits, run by the steward. CI and every Codecov check were green on 7ce7bcf.

    Reviews:

    • Full reviews at 9f937a5 by GPT 6.1 Sol (high) and Claude Fable 5.2 (high). Both found no conformance fault; Sol withdrew the two findings I disputed.
    • The range 9f937a5..7ce7bcf holds only the nine changes the steward asked for. Claude Opus 5.5 (medium) checked it. The Sol re-check of that range failed on a rate limit and was not repeated.
    • The disclosure names both Claude Opus 5.5 and GPT 6.1 Sol as authors, so no reviewer came from a family that wrote none of the code.

    One mistake of mine: deleting the branch closed #128, which was stacked on it. I restored the branch at 7ce7bcf and reopened #128. Its base stays on that branch until the author rebases it onto develop.


    🤖 Posted autonomously by Claude on behalf of pasta.

  5. PastaPastaPasta commented on Nov 16, 2026

    @PastaPastaPasta
    MemberAuthor
    #133 sdk%chore(ci): refresh `actions.lock` after `rust-toolchain@stable` moved  merged 2026-11-16 at ec78d9e3
    class: A (CI upkeep; no library code)   reviewers: Muse Wonder Full 1.0, high
    decided by: steward
    kwvg's review: never reviewed
    his open threads: none
    judgment calls: three (below)
    revert: git revert -m 1 59464246886a3d7790f8b1099705353b5b690690
    

    What it is: develop was red at d8cfd14 for a reason outside any PR. dtolnay/rust-toolchain rebuilds its stable branch on top of master, so b725eda5, the commit actions.lock pinned it to, was no longer on the ref (compare). The stable jobs failed at "Set up job" and Lint failed its lockfile check (run 43672876347). The commit is the output of your gh actions-lock recipe, plus --accept-moved. The tool refuses to write without that flag when a pin has left its ref.

    Judgment calls:

    1. I treated a pin refresh made with your own recipe as upkeep, not a change to the CI layout, and merged it as steward. Your merge of the dependabot bump in sdk%chore(ci): bump the actions group across 3 directories with 5 updates #99 is the precedent.
    2. The tool also moved two other pins and renamed a repository, and they were kept rather than removed by hand:
      • taiki-e/install-action@v2 and Mic92/niks3-action@v1 moved forward along their tags, made by their maintainers, and each old pin is an ancestor of the new one.
      • jlumbroso/free-disk-space is now lumbroso-commons/free-disk-space in the lockfile and in six workflows. It is the same repository (ID 506133999) at the same commit, transferred to an organisation of which jlumbroso is a member.
    3. There was one reviewer from another model family, not two. The change is 19 lines of lockfile and uses: references with no library code. The reviewer checked every new pin against GitHub and found nothing beyond two nits. One was that the command adds --accept-moved to your recipe. The other was that the pin will break again the next time upstream rebuilds stable, which is question 3 in this issue's body.

    Proof: CI on ec78d9e had 52 checks passing and 3 skipped, Lint's lockfile check included. There is no Rust change, so the CI run on the head was the proof.


    🤖 Posted autonomously by Claude on behalf of pasta.

Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Metadata

Metadata

Assignees

No one assigned

    Labels

    No labels
    No labels

    Type

    No type

    Projects

    No projects

      Milestone

      No milestone

      Relationships

      None yet

      Development

      No branches or pull requests

      Issue actions