lnurlcashkernel

Verify spends of LUD-25 notes with Bitcoin Core's own script interpreter, unmodified (libbitcoinkernel), from Python.

Every LUD-25 note is a taproot output key Q (cp1<Q>). A redeemer opens it with a ck1 (a key-path signature by Q) or a cw1 (a script-path spend of one of its leaves). The plain bearer note is a single OP_SHA256 <h> OP_EQUAL leaf under BIP-341's NUMS key, spent by revealing the preimage.

import lnurlcashkernel as k

spend = k.decode_spend(k1)   # a ck1, a cw1, or a bearer note's 64-hex preimage; None if none
q = spend.output_key         # the note to look up in the mint's records
tpl = k.verify_spend(
    output_key=q,            # 32-byte x-only Q, as recorded by the mint
    domain="mint.example",   # this mint's own host: signatures are bound to it
    spend=spend,
    now=now,                 # THE MINT'S clock, Unix seconds - passed in, never read
    locked_at=locked_at,     # when the mint locked this note
)                            # returns None, or raises SpendRejected

What this verifies - and what it does not

It verifies scripts. Whether a key-path signature is valid for Q, whether a revealed leaf and control block really commit to Q, and whether the leaf's conditions (a signature, a hash preimage, a multisig, a timelock comparison) are satisfied, is decided by Bitcoin Core's own code, compiled from an unmodified pinned tree. This package adds no verification logic of its own; the only arithmetic it does itself is recomputing Q from a control block (taproot.py, public data only) so a mint knows which note to look up.

It does not verify time, and cannot. There is no chain here. Bitcoin splits timelock enforcement in two: the script compares its number to the spending transaction's nLockTime/nSequence, and separate consensus code checks that those fields are acceptable against the chain tip. With no chain, the second half falls to the mint's own clock (policy.check_time_claim).

A timelock "verified" by this library means the mint asserts its own clock. It is a custodial policy decision, not a consensus guarantee: nothing stops a mint from lying about the time, and nobody can independently re-check it. That is no new trust beyond what a custodial mint already asks for - but it must not be described as trustless, in UI copy or anywhere else.

The library never reads the system clock; now and locked_at are always arguments, and tests/test_purity.py enforces this.

The redeemer signs the time claim

A tapscript CHECKSIG signature commits to the transaction's nLockTime and nSequence (even under ANYONECANPAY). So the redeemer chooses those values, signs them, and presents them; the mint cannot pick "now" after the fact. The script check (Core) and the clock check (mint) are then independent, exactly as on a real node.

The canonical spend transaction (version 2)

Core verifies an input of a transaction, so a fixed minimal one is assembled. A signer must sign the BIP341 sighash of exactly this shape (see verify.py; it is normative and versioned):

field	value
nVersion	2 (CSV requires ≥ 2)
vin[0]	prevout (tagged_hash("LNURLcash/mint", domain), 0), empty scriptSig, nSequence = claimed
vout[0]	value 0, empty scriptPubKey
nLockTime	claimed
spent output	(OP_1 <Q>, 0)

domain is the mint's lowercase host, so a signature seen by one mint cannot be replayed at another. The spent amount is always 0: the mint enforces a note's value from its own records, and leaving it unsigned means one key has exactly one signature for its note, reproducible from seed. A key-path spend always signs locktime 0 and the final sequence 0xffffffff.

Version 1 (zero prevout, amount_msat committed) is gone; signatures made for it do not verify.

This package never signs. For a signer in Python, key_path_sighash(Q, domain) and script_path_sighash(Q, domain, leaf, locktime=, sequence=) return the one hash that will verify (SIGHASH_DEFAULT), checked against LUD-25's own test vector 3.

Wire format

value	payload
cp1	Q (32 bytes) - 61 characters, fits a LUD-12 comment
ck1	key-path spend: Q ‖ sig(64) - 163 characters
cw1	script-path spend: u32 locktime ‖ u32 sequence ‖ (u16 len ‖ item)* over script, control block, witness (bottom first)
64 hex in k1	a bearer note's preimage - decode_spend builds the OP_SHA256 <h> OP_EQUAL spend under NUMS
64 hex for a cp1	a bearer note's h - decode_note returns its Q

Supported spends

Any key-path signature by Q, and any leaf script: what a leaf may do is decided by Bitcoin Core's consensus rules alone, not by a list of shapes. The plain bearer note, OP_SHA256 <h> OP_EQUAL under BIP-341's NUMS key, is just one such leaf.

The one exception is tapscript's upgrade hooks, which consensus accepts unconditionally and on-chain nodes only refuse by policy - a policy libbitcoinkernel does not expose. script.check_leaf refuses them itself, before Core is called:

• a leaf version other than 0xc0
• any OP_SUCCESSx opcode in the leaf (pushed data is not scanned)

Otherwise such a leaf would be spendable by anyone who sees it, and no new opcode could ever be given a meaning. BIP-342's third hook, a signature check against a key that is neither empty nor 32 bytes, can't be spotted statically (the key may come from the witness) and is left to consensus: wallets must never write such keys.

Time claims are accepted only in forms that mean something with no chain: a non-zero nLockTime must be a Unix time (≥ 500,000,000) and an enabled relative nSequence must carry BIP68's time flag (512-second units). Block-height and block-count claims are refused, not guessed at.

Upstream

vendor/bitcoin is a git submodule pinned to one release tag and never edited. upstream_version() reports which (v31.1 (9be056a)) - log it; it is what tells an operator which script verification they actually run. See UPSTREAM.md and SECURITY.md.

Core builds this as libbitcoinkernel, which Core itself labels "experimental." Its predecessor libbitcoinconsensus was removed in Core 28.

Development

git clone --recurse-submodules --depth 1 <this repo>
python -m venv .venv && . .venv/bin/activate
pip install cmake ninja scikit-build-core pytest
pip install .            # builds Core's kernel library (~3 min cold) into the wheel
pytest

Core's CMake requires Boost headers (≥ 1.74) unconditionally. The distro's may be too old (manylinux_2_28 ships 1.66), so fetch a known version:

scripts/fetch_boost_headers.sh            # -> build/deps/boost-shim
pip install . -Ccmake.define.Boost_DIR=$PWD/build/deps/boost-shim

Tests include Core's own BIP341 vectors driven through the binding, and real script-path spends built from scratch and judged by Core.

Releases

Publishing follows lnbits/electrum-client's flow: pushing a tag vX.Y.Z runs .github/workflows/release.yml, which uploads to PyPI with uv publish over OIDC trusted publishing (no API token is stored anywhere) and then creates a GitHub release.

git tag v0.1.0 && git push origin v0.1.0

It differs from electrum-client in one way: that package is pure Python, this one ships Bitcoin Core's native library. So wheels are built per platform with cibuildwheel in manylinux containers (a wheel built on a developer machine is tagged linux_x86_64, which PyPI rejects - the one built during development needed manylinux_2_39, far too new for real servers), plus an sdist that bundles Core's sources. Only the final publish job holds id-token: write.

One-time PyPI setup

Trusted publishing must be registered on PyPI before the first tag - it cannot be done from this repo. On pypi.org, add a pending publisher for lnurlcash-kernel with the GitHub owner/repo, workflow release.yml, and no environment (the workflow does not use one, exactly like electrum-client's).

The GitHub Actions workflows in .github/ are written but have not been run

• they need a real runner. Treat the first tag as the test. What was verified locally: the version-from-tag behaviour, the dev-build guard, the Boost fetch script, and a full wheel + sdist build.

Versioning

The version comes from the git tag (vX.Y.Z -> X.Y.Z) via setuptools-scm, as electrum-client does with hatch-vcs. PyPI rejects PEP 440 local versions, so none is ever emitted (local_scheme = "no-local-version"); an untagged build gets a .devN version, and the publish job refuses to upload any 0.0.0/.dev artifact, so a build that could not see its tag can never reach PyPI.

The bundled Bitcoin Core release is deliberately not in the version string (0.1.0+core31.1 would be a local version, which PyPI rejects). It is exposed by upstream_version(), stated in each GitHub release's notes, and enforced against the submodule pin by a test.