Changelog¶
A release’s section is written in the release’s own pull request, from the
squash subjects since the previous tag, and no other pull request adds an
entry (section 9 of the organization standard and
CONTRIBUTING.md’s Pull requests). A section takes what a reader of
this repository would notice, grouped: what changed, why, and what it cost.
RELEASE_NOTES.md has the release notes, which say
what a user has to act on; this file is the record behind them, and is
where a claim in those notes can be checked.
Neither file states how many entries it holds: a stated number is one
more line to edit at every release. Naming a command that counts them is the
rejected alternative, and what rejects it is the shape of this file: a
### heading names one entry, and a theme several entries
share wherever the file already carried such a heading — section 9 of
the organization standard rules against writing them and
Nothing already written is rewritten keeps the ones that are there. A
- bullet at the margin is then one of an entry’s claims where the
heading above it names an entry and a whole entry where it names a
theme. Section 9 also lets an entry making a single claim have a
paragraph for a body, which no bullet pattern reaches. Counting the
entries is a reading of the file rather than a command.
v2026.11.2¶
The signet magic of a challenge past 0xffff bytes is derived¶
magic_from_signet_challenge writes the length prefix as Core’s
WriteCompactSize does, in all four forms. RELEASE_NOTES.md’s v2026.11.2
has what it and assert_chain answer (closes #708).
integration-bitcoind.yml asks Core 27.2, 30.4, 31.2 and 32.1¶
28.4 and 29.4, past their end of life, go, and 30.4 and 31.2 replace 30.3 and 31.1 (closes #711).
SECURITY.md links what changed since the security review¶
It names the reviewed commit and links that commit’s comparison with
main (issue btclib-org/.github#1795).
v2026.10.30¶
The empty signet challenge derives the magic Core uses¶
RELEASE_NOTES.md’s v2026.10.30 has what magic_from_signet_challenge and
assert_chain answer for it (closes #702).
A release is verified by repository id, with the signer named without a ref¶
SECURITY.md and RELEASING.md verify under --owner btclib-org and
compare the certificate’s repository id, which a rename keeps (closes #682).
The JSON-RPC reply decoding is fuzzed with ClusterFuzzLite¶
fuzz/fuzz_reply.py feeds arbitrary bytes to every reply reader, and
fuzz.yml runs it weekly and on dispatch (closes #691).
tests/chains_test.py compares the magic copy with btclib’s¶
btclib joins the test group at 2026.10.24, and chains.py says why the
copy stays (issue btclib-org/btclib#2707).
The descriptor test fails fstat for chains.py alone¶
pytest 9.2.0’s own capture calls os.fstat, and the test no longer fails
it (closes #713).
COMPARISON.md compares the transport and error handling with corepc¶
A row for redirects is added (closes #669, closes #687). The corepc commit
is linked on GitHub, and .lycheeignore excludes git.rust-bitcoin.org,
which answers the link check with a 403 (closes #689, closes #704).
integration-bitcoind.yml asks Core 32.1¶
32.1 is the newest entry, asked on every pull request beside 27.2 (closes #707).
The publish action, the uv floor and the lock move past advisories¶
gh-action-pypi-publish v1.14.3, uv >=0.13.4 and virtualenv 21.14.13
(issue btclib-org/.github#1844); setuptools 84.2.1 and filelock 4.2.4
(issue btclib-org/.github#1849).
Every uv_build specifier is <0.14¶
The check-sdist hook’s included (closes #679).
Development is LLM-based, and a contribution must be understood¶
CONTRIBUTING.md, REVIEWING.md, README.md and the pull request
template say so (issue btclib-org/.github#1846).
.github/dependabot.yml names the uv.lock exception¶
And says only uv_build has a ceiling in the dependency groups (issue
btclib-org/.github#1818, issue btclib-org/.github#1848).
.pre-commit-config.yaml says how a zizmor finding is answered¶
Whichever finding it is (issue btclib-org/.github#1813).
Dependabot and pre-commit.ci bumps¶
The actions group, uv-build 0.13.5 and the pinned hook revisions.
v2026.10.23¶
Refusals of a secret input raise the library’s errors and quote less¶
RELEASE_NOTES.md’s v2026.10.23 has the new exceptions and messages (closes #648).
tests/secret_free takes the organization’s grouped-secret check¶
assert_secret_free fails when a refusal’s text shows a secret split into
groups (issue btclib-org/.github#1697).
A report easy to rediscover is fixed as soon as possible¶
SECURITY.md says so, and publishes its advisory with the fix (issue
btclib-org/.github#1771).
verified-signatures.yml checks the signatures on main after each push¶
It reports whether every commit pushed is verified (issue
btclib-org/.github#1738).
The maintainer’s bypass is added back only in an emergency¶
main-self-merge has no bypass actor, and CONTRIBUTING.md takes the
organization’s shared half (issue btclib-org/.github#1740).
AGENTS.md replaces CLAUDE.md¶
CONTRIBUTING.md adds Worktrees and lets pull requests land while main
is red (btclib-org/.github#1777); REVIEWING.md, REPOSITORY.md and
.claude/commands/review.md follow (issue btclib-org/.github#1808).
The pyroma hook builds no dependency from source¶
A locked dependency with no wheel for the hook’s Python fails the hook instead of being built (issue btclib-org/.github#1763).
claude-review.yml declines zizmor’s dangerous-triggers on issue_comment¶
The trigger stays, with the ignore on its line and the reason above it (issue btclib-org/.github#1813).
uv.lock moves GitPython, aiohttp, certifi and virtualenv¶
GitPython 3.2.1, aiohttp 3.14.5 (closes #666), certifi 2026.10.19 and
virtualenv 21.14.12, past the advisories uv audit refuses (issue
btclib-org/.github#1835, issue btclib-org/.github#1824).
Dependabot and pre-commit.ci bumps¶
astral-sh/setup-uv and the pinned hook revisions.
v2026.10.17¶
SECURITY.md says why the client adds no TLS of its own¶
Its Limitations, not vulnerabilities links Core’s reasons and says what
keeps rpc credentials off the wire; README.md and the client’s docstring
point to it (closes #654).
The bounds on a shown name say why they are there¶
_MAX_NAME_SHOWN and _MAX_KEY_SHOWN say that what they let through is
shown even when it is a misplaced secret, and ASSURANCE_CASE.md’s
CWE-209 bullet names both (GHSA-cg87-q4qx-qw8p).
The workflows take the organization’s actionlint fork and api-break¶
The kjanat fork lints the workflows, and api-break.yml labels a pull
request that breaks the public API (issue btclib-org/.github#1698, issue
btclib-org/.github#1674).
tests/secret_free is the organization’s copy¶
Its self-test comes with it (issue btclib-org/.github#1697).
The review call sets cache-mode: none¶
It answers the finding the ignore in .github/actionlint.yaml silenced, and
that file is deleted (closes #649).
A merged pull request’s closed run of api-break queues¶
It waits behind the walk in flight instead of cancelling it (issue btclib-org/.github#1744).
claude-review.yml points at the called workflow’s tool list¶
It no longer copies that list (issue btclib-org/.github#1642).
RELEASING.md runs the dependents’ suites before the merge¶
It also reads the tag’s signature back from the API (issue btclib-org/.github#1647, issue btclib-org/.github#1660).
REPOSITORY.md says what tag-integrity refuses¶
A v* tag push that brings an unsigned commit, not an unsigned tag (issue
btclib-org/.github#1635).
The maintainer’s bypass is everyday practice¶
Until the review bot posts under an identity other than the author’s
(issue btclib-org/.github#1695); REPOSITORY.md records strict off on
main (issue btclib-org/.github#1712).
uv.lock moves to pydantic 2.14.0¶
Its pydantic-core has cp315 wheels, so the pyroma hook builds nothing (issue btclib-org/btclib#2685).
Dependabot and pre-commit.ci bumps¶
mypy, ruff, uv-build, coverage, actions/download-artifact,
actions/upload-artifact and the pinned hook revisions.
v2026.10.8¶
The release publishes the files the build job built¶
publish-testpypi, publish-pypi, github-release and test.yml’s dist
job on a release stop where the dist they download differs from the digests
reusable-build.yml outputs (issue #584).
The required checks gate a merge only when the landing waits¶
REPOSITORY.md,CONTRIBUTING.mdandRELEASING.mdsay a merge with--adminwaits for neither the approval nor the checks (closes #583):enforce_adminsis off, and auto-merge is the landing that waits for them.
The forms set a type, and the history files lose merge=union¶
Each form sets an issue type, no kind label (issue btclib-org/.github#1584); the history files have no merge driver (issue btclib-org/.github#1582); a release reviews the bestpractices.dev answers (issue btclib-org/.github#1589).
The post-release install refreshes the index¶
RELEASING.md installs the release with --refresh-package bitcoin-core-rpc,
because uv can answer from a cached index and miss a version just published
(issue btclib-org/.github#1595).
--admin waits for no required check¶
CONTRIBUTING.md’s emergency paragraph says --admin skips the
required checks too, and REVIEWING.md’s “hold the merge” excepts it
(issue btclib-org/.github#1597).
check-changelog refuses an entry added to an older release¶
A ### heading under a release older than the newest, absent from the file
at the merge base with origin/main, is refused
(issue btclib-org/.github#1614).
Refusals quote less of what a caller passed¶
RELEASE_NOTES.md’s v2026.10.8 has the new messages and what still shows (GHSA-cg87-q4qx-qw8p).
v2026.10.4¶
_resolve names the host it timed out on, even out of time before the join¶
_seconds_left, called insideThread.join, raised a bare “timed out” masking the “timed out resolving$host” message whenever the deadline was already spent before.join()ran (issue #564).
[tool.uv] required-version follows the uv dependabot-core bundles to 0.12.19¶
required-versionreads>=0.12.19, not>=0.12.17(issue btclib-org/.github#1438): a floor below the pin admits auvolder than the one Dependabot writesuv.lockwith.
pypi-install.yml installs the version the release published¶
The install names
bitcoin-core-rpc==<version>from the tagrelease.ymlpasses (issue btclib-org/.github#1456): a bare name let a lagging index serve the release before it.
CONTRIBUTING.md points a newcomer at good first issue¶
An issue carrying the label is small and self-contained (issue btclib-org/.github#1362): The issue tracker says so, and links the organization-wide search for the open ones.
The OpenSSF Baseline badge¶
README.md’s badge row ends with the OpenSSF Baseline badge, beside the Best Practices badge, which section 2 of the organization standard admits (issue btclib-org/.github#1460).
pypi-install.yml retries the install of the version the release published¶
Each install cell installs through btclib-org/.github’s
install_published_release.py, which retries only while the installer says the pin is not resolvable (issue btclib-org/.github#1458).
The release’s attestation bundle is attached as *.intoto.jsonl¶
RELEASING.md’s commands andSECURITY.md’s verification name the bundle<tag>.intoto.jsonl(issue btclib-org/.github#1468):reusable-github-release.ymlattaches it under that name.
codeql.yml’s aggregate runs check_run_jobs.py¶
The aggregate’s step runs
check_run_jobs.py, served frombtclib-org/.github, which reads the jobs listing again up to a deadline (issue btclib-org/.github#1463).
generate_sbom.py carries a not-affected list into the bill of materials¶
generate_sbom.pyreads.github/vex.toml, where the tree lists the vulnerabilities its release is not affected by, into the document’svulnerabilities; no list, no key (issue btclib-org/.github#1469).
SECURITY.md promises a response time¶
SECURITY.mdsays a report is acknowledged within 7 days, and a fix or a published advisory within 90 (issue btclib-org/.github#1460).
ASSURANCE_CASE.md states the lock and the reply’s numbers correctly¶
The assurance case names which commands resolve outside
uv.lock, and says floats becomeDecimalwhile integers remainint(closes #582).
The dependency census reads the whole package¶
The dependency census reads every
.pyfile of the package,__init__.pyincluded, and requires[project].dependenciesto be empty (closes #581).
Docs say a reply integer decodes as an int¶
README.md,docs/source/migrating.mdand docstrings inclient.pysay a floating-point number decodes as aDecimaland an integer as anint, as_parsed_json_bodydoes (closes #589).
release.yml audits the lock before it publishes¶
The
auditjob runsuv auditover what the wheel declares, by calling btclib-org/.github’sreusable-audit.yml, and both publish jobs wait for its success (issue btclib-org/.github#1466).
generate_sbom.py is btclib-org/.github’s¶
The
distjob writes the bill of materials with btclib-org/.github’sgenerate_sbom.py, served frommain, and the tree keeps no copy of it (issue btclib-org/.github#1478).
Dependency review is a required check¶
REPOSITORY.mdreadslint / Dependency reviewback with the other required checks (issue btclib-org/.github#1465).
[tool.uv] required-version is >=0.12.18¶
required-versionreads>=0.12.18, not>=0.12.19(issue btclib-org/.github#1482): the Dependabot service refused0.12.19withtool_version_not_supported.
SECURITY.md names the latest security review¶
SECURITY.md gives the date of the latest security review and links the issue
that records it (issue btclib-org/.github#1362).
A Signed-off-by: trailer on every commit of a pull request¶
Every commit of a pull request carries a
Signed-off-by:trailer (issue btclib-org/.github#1467): Pull requests says how to add it.
The primary-checkout section uses one form for the checkout¶
The section writes the checkout as
"${checkout:?}"throughout, says what<scratchpad>is and names the pull (issue btclib-org/.github#1500).
The python inventory has a copy kept in the tree¶
docs/source/_inventories/python.invis read whendocs.python.orgfails (issue btclib-org/.github#1508), so an outage of that site no longer fails the-n -Wdocs build.
public-api is red for a break RELEASE_NOTES.md does not name¶
release.yml and RELEASING.md say that a red public-api means
RELEASE_NOTES.md misses a name (issue btclib-org/.github#1517).
The release’s attestation is signed by reusable-build.yml, at SLSA Build L3¶
release.yml calls reusable-build.yml, which signs the files before the
publish jobs wait for approval (issue btclib-org/.github#1506).
CONTRIBUTING.md says the maintainer self-merges while the bot review is off¶
The review says no ack of record exists while claude-review.yml is off,
and that a local review of a named sha, by a reviewer other than the author,
stands in (issue btclib-org/.github#1527).
dependabot.yml does not say that every workflow passes --locked¶
.github/dependabot.yml says the workflows install from uv.lock with
--locked, bar the jobs that re-lock on purpose or install a published or
built package (issue btclib-org/.github#1538).
REPOSITORY.md reads back the web sign-off setting¶
REPOSITORY.md reads web_commit_signoff_required back, the organization
setting section 11 of the standard states (issue btclib-org/.github#1540).
The Sign-off check is required¶
CONTRIBUTING.md’s shared half says a pull request whose commits lack the
Signed-off-by: trailer cannot merge, and REPOSITORY.md lists
lint / Sign-off among the required checks (issue btclib-org/.github#1550).
CONTRIBUTING.md says the maintainer lands through the bypass¶
The review says the ack of record is a bot’s, so the criterion
two_person_review is unmet (issue btclib-org/.github#452).
The bypass is for emergencies, in CONTRIBUTING.md and this tree’s prose¶
CONTRIBUTING.md and REVIEWING.md say every pull request, the maintainer’s
included, lands with an approving review from somebody other than its author;
the bypass is for emergencies (issue btclib-org/.github#1362).
Earlier entries on how a pull request lands¶
Entries above that have the maintainer landing without another person’s approval describe the rule before issue btclib-org/.github#1362 (issue btclib-org/.github#1569).
REPOSITORY.md reads the review switch as the organization’s¶
CLAUDE_REVIEW_ENABLED is set on the organization, so this file records only
that the repository’s own variable store is empty (issue
btclib-org/.github#1560).
_checked_url’s refusals leave the url out¶
No refusal repeats the url or the scheme (closes #575):
http:user:pw@hostparses with the credentials as the path,user:pw@hostas the scheme.
_checked_url refuses what urlsplit cannot split and what holds whitespace¶
Such a url is a
BtcRpcValueErrorat construction (issue #578): it was a bareValueErrorforhttp://[user:pw@host/, a failure at the first call forhttp://host/a b.
A reply number whose exponent is past sys.get_int_max_str_digits is refused¶
1e1000000raisesFetchError(closes #576), the bound_decimal_jsonapplies to a parameter; a limit of 0 lifts it for both.
An RpcError’s str() escapes the node’s non-printable characters¶
A newline or an ESC in the node’s message is written as
\nor\x1b(closes #577), so it cannot forge a log line.args[0]keeps the text as it arrived.
The TLS key log is off¶
Both transports build their own TLS context, since ssl.create_default_context
writes the session keys to the file SSLKEYLOGFILE names (closes #580); a
replaced ssl._create_default_https_context is not consulted.
http_request refuses a header that would split the request¶
A header name that is no RFC 9110 token, or a value with a CR, an LF, a NUL or
a character latin-1 cannot encode, is a BtcRpcValueError, and a non-string
one a BtcRpcTypeError (closes #578).
The transports do not repeat a url they refuse¶
A url with whitespace or a control character, a scheme other than http(s), no
host, a port that is no number, or a login is refused without being repeated;
an error names scheme://host[:port] only (closes #616).
The TLS context is strict on Python 3.11 and 3.12¶
The context sets VERIFY_X509_STRICT and VERIFY_X509_PARTIAL_CHAIN, which
CPython 3.13 sets and 3.11 and 3.12 do not (issue #580).
A url with a non-ASCII character is refused¶
The client, http_request and the two transports refuse it, a host included,
not failing in http.client with a UnicodeEncodeError that holds the request
line. Write the host in ASCII, xn--bcher-kva.example (closes #619).
The fix for issue #564 above shipped in v2026.9.29¶
The v2026.9.29 tag names 7cf5ddb, the commit after that release’s own, so that release carries the fix; its entry sits in this section.
v2026.9.29¶
notice-rgx’s comment drops the hook this tree carries none of¶
The comment cited a
language: scripthook this tree carries none of; it now gives section 5’s own reason instead – a script run by path opens with its own shebang (issue btclib-org/.github#1294).
Python 3.10 leaves the supported range, and the floor is 3.11¶
requires-pythonis>=3.11, 3.10’s end of life being 2026-10:SessionTransport.__enter__returnsSelf, typing’s from 3.11.
The attestation’s signer is reusable-attest.yml from v2026.9.24 on¶
RELEASING.md,SECURITY.mdandsdist-rebuild.ymlname the called workflow as the signer, the path ofrelease.ymlkept for v2026.9.3 and earlier, which it signed (issue btclib-org/.github#1301).
The badge row ends with the OpenSSF Best Practices badge¶
Project 14815 at bestpractices.dev, after the Scorecard badge, where section 2 of the organization standard places it for a tree section 10’s
scorecardentry names, as it names this one (issue btclib-org/.github#350).
RELEASING.md’s griffe command finds the package under src/¶
It passes
-s . -s src, without which griffe searches only the-arevision’s root and stops atModuleNotFoundError(closes #522).
sdist-rebuild.yml stops passing attest-signer¶
The called workflow verifies against reusable-attest.yml alone, so the
input decides nothing (issue btclib-org/.github#1315).
ARCHITECTURE.md and ASSURANCE_CASE.md join the root¶
The architecture moves there from CLAUDE.md and README.md, which point
at it; the assurance case cites the tree for every claim (issue
btclib-org/.github#1321).
The three aggregate jobs no longer skip while draft¶
A skipped required check reads as passing, so codeql-passed,
integration-passed and test-passed now fail a first step on the
draft flag instead of skipping on it (issue btclib-org/.github#1327).
Python 3.15 is the pin, on its release candidate¶
A classifier and the platform sweeps name 3.15 too, and
.readthedocs.yamlstays on 3.14, read the docs offering nothing newer (issue btclib-org/.github#1324).
sdist-rebuild.yml’s comment names setup-python, not python-version¶
It named python-version, an input the called workflow does not declare, and
called the defaults what this file built before, a bare uv build; it names
setup-python and what the defaults build (issue btclib-org/.github#1346).
pre-commit.ci skips uv-lock¶
Its image does not carry the interpreter .python-version names, and with
no network there the hook fails downloading it; the lint workflow runs the
hook on that interpreter (issue btclib-org/.github#1348).
The rebuild of a release names the interpreter its tag pinned¶
RELEASING.md’s Rebuild a release from its tag reads it from the tag’s
.python-version rather than naming one (issue btclib-org/.github#1349).
The rebuild of a release runs in a worktree of its tag¶
RELEASING.md’s Rebuild a release from its tag builds in a worktree of
the tag, clean by construction, and drops the git archive export, which
has no .git for its git commands (issue btclib-org/.github#1352).
The rebuild of a release builds under the release’s own uv¶
RELEASING.md’s Rebuild a release from its tag builds under the uv
the published wheel names and verifies the sdist first: a wheel that
disagrees stops the chain after it (issue btclib-org/btclib-node#1063).
pypi and testpypi name three reviewers, not one¶
RELEASING.md named fametrano alone; pypi and testpypi now
require a review from any of fametrano, giacomocaironi and
pmazzocchi (issue btclib-org/.github#1355).
README.md and CONTRIBUTING.md link GOVERNANCE.md and ROADMAP.md¶
Both point a contributor at the organization’s one copy of each, in
btclib-org/.github (issue btclib-org/.github#1359).
REVIEWING.md lets a filed issue carry its fix¶
An issue filed from a review may now say the fix where one is known; the filing bar stands as it was (issue btclib-org/.github#1378).
call_batch refuses a malformed batch and reply with its own errors¶
A json array or object as a reply id is a FetchError, a generator a batch,
and a member that is no pair or a str or non-iterable calls is a
BtcRpcTypeError, None and "" included (closes #546) (closes #547).
SessionTransport sends a request once when the read fails¶
Where a reused connection closed after the whole request was written, the
request went out again, though the node may have executed it; that is now a
FetchError, and only a failed write is sent again (closes #541).
SessionTransport probes a kept connection with select.poll¶
select.select raised ValueError on a reused socket whose descriptor was
at or above FD_SETSIZE; the probe falls back to it only where poll does
not exist, as on Windows (closes #552).
The live smoke run asks about the batch, the session and the chain check¶
call_batch, call_raw, SessionTransport, RpcChannel and assert_chain
now meet a live node, whose short -rpcservertimeout each --protocol run
waits out once, for the session’s idle close (closes #548).
The fixtures are written from Core’s source, not recorded from it¶
ARCHITECTURE.md and rpc_smoke.py now say so, as tests/_data/README.md
does (closes #549).
tests/_data/README.md states no number of live Core versions¶
It points at the live workflow’s versions job instead (closes #550).
A Decimal parameter is written as the JSON number it is¶
call, call_raw and call_batch send one digit for digit where they refused
it, and COMPARISON.md says python-bitcoinrpc rounds it through float where
Core’s copy sends a string (closes #545) (closes #551).
SessionTransport reads each answer against the call’s max_body_size¶
http_request hands it to __call__ as it does to urlopen_transport, so it
widens or tightens the read where the constructor’s limit bounded it, and the
constructor takes none (closes #543).
The timeout bounds the name lookup, the status line and the headers too¶
A peer dripping headers held a call open under either transport; the call now
waits for no lookup, connect, handshake, send or recv past the deadline, which
a connection_factory’s connection is typed to carry (closes #542).
codeql.yml, REPOSITORY.md and dependabot.yml follow sections 10 and 11¶
The aggregate accepts analyze’s lagging rows (issue btclib-org/.github#1395),
signatures and SHA pinning are read back (issue btclib-org/.github#1409), and
pre-commit is an ecosystem left unused (issue btclib-org/.github#1391).
SessionTransport cuts an error body to MAX_ERROR_BODY_SIZE¶
It refused a 4xx or 5xx body over max_body_size, so a legacy 1.1 rpc error
became a FetchError with no status; the body is now truncated, as
http_request reads urlopen_transport’s HTTPError (closes #544).
The codeql aggregate re-reads a lagging row and judges analyze’s result¶
The listing is read again, three times at most, while an analyze row is
unfinished (issue btclib-org/.github#1416), and an analyze result other than
success or skipped fails the step on its own (issue btclib-org/.github#1424).
v2026.9.24¶
Section 9’s comment and placeholder rules land in this tree’s own docs¶
A trailing
#comment inside ashellfence ofCLAUDE.md,CONTRIBUTING.md,RELEASING.mdorREPOSITORY.mdno longer sits on the command line it followed (issue btclib-org/.github#771): an interactivezshleavesINTERACTIVE_COMMENTSunset, so it hands the comment’s own words to the command as arguments rather than treating them as a comment. What the comment said either became a sentence above the fence or moved to a line of its own that opens with#, which section 9 atbtclib-org/.github’s5712e05reads as harmless while it keeps to plain words, carrying neither an apostrophe nor a backtick.CHANGELOG.md’s two instances, inside a landed entry, stay: this file is append-only undermerge=union, and a landed entry is the record of what was said when it landed.CLAUDE.md’sWT=line loses its trailing worktree-name example (issue btclib-org/.github#786): the comment’s own words followed the line’s final>, so an interactivezshtook them as that redirection’s target instead of finding nothing there and failing at the parse before it was ever a command. The example moves into a sentence above the fence.REVIEWING.md’s two placeholders — thegh issue list --searchandgh issue create --title/--bodylines — go bare (issue btclib-org/.github#772): quoted,gh issue createtakes both flags as literal values and files an issue whose title is the placeholder itself, rather than failing at the shell.RELEASING.md’s--subject <title>loses its quoting and moves ahead of--body-file <path>, so a bare placeholder ends the command (issue btclib-org/.github#775): section 9’s unquoting bullet asks a placeholder standing as a whole argument to be bare, and its position bullet asks the bare one to end the command, which--body-file <path>already did and--subject "<title>"did not.
CLAUDE.md’s worktree fence states the editing and the gates once¶
The whole-line comment
# edit, gate and commit here, thenno longer sits between the fence’scdline and itsgit push(issue btclib-org/.github#793): the paragraph above the fence says that the editing, the gates and the commits all happen in the worktree before the push, and section 9’s One fact in one place argues against stating that fact a second time inside the fence.
CLAUDE.md’s worktree-removal fence guards its $WT expansion¶
The fence closes with
git worktree remove --force "${WT:?}", and the paragraph above it says what the:?is for (issue btclib-org/.github#790): section 9 of the organization standard reads a fence with nothing in it that fails at the parse as live — a paste of it alone is a command, and it runs with every value the reader was to set empty — so such a fence writes each of those values${name:?}, with prose beside it because a reader who is not told deletes it. The sentences saying so arebtclib-org/.github’s own, verbatim.
claude-review.yml’s fork comment no longer illustrates itself with bbt¶
The paragraph above the fork condition in
.github/workflows/claude-review.ymlno longer namesbtclib-org/bbtas an example of a repository that is itself a fork (issue btclib-org/.github#456): no repository of the organization is one, so the example did not instance the case it argued. The replacement reads.forkagainst the head repository having a parent rather than against this one, and points its own readback at this tree —gh api repos/btclib-org/bitcoin-core-rpc --jq .forkanswersfalse. The wording isbtclib-org/bbt’s own, landed there first.
[tool.uv] required-version is raised to the current Dependabot ceiling¶
pyproject.toml’s[tool.uv] required-versionmoves from>=0.12.5to>=0.12.7(issue btclib-org/.github#448): section 1 of the organization standard sets the floor at the ceiling rather than below it, and the ceiling — the newest uv Dependabot’s own bundled updater still reads — is0.12.7today, ahead of the0.12.5this tree carried.uv.lockis untouched: it carries norequired-versionof its own.
conftest_test.py covers the resolution asks_for_everything does¶
A case names one directory two ways — through a symlink and not — and
fails with either .resolve() removed from tests/conftest.py (issue
btclib-org/.github#806): resolved on one side only, the two spellings
compare unequal, so pytest tests reads as a subset of itself and the
run that measures the whole suite is gated at nothing. The link is made
under tmp_path rather than taken from the machine, so the case is
about the comparison and not about which directories an operating system
links, and a platform that will not create one skips.
CLAUDE.md’s worktree fence carries no uv sync¶
The fence’s
cd "$WT"line stands alone, with nothing chained to it by&&(issue btclib-org/.github#739): section 9 of the repository standard gives a line that writes a fence of its own, and a sync chained to thatcdwrites in the directory the shell is standing in — withWTunset,cd ""leaves/bin/zsh5.9 and thebash3.2.57 macOS ships where they were,bash5.3.15 refusing it, so for a reader ofCLAUDE.mdthat directory is the primary checkout.The block’s own parse is not the guard either. Read as a script with
-n, the fence unfilled is a syntax error to each of those shells, while the same lines with values filled in parse cleanly: what stops such a paste is the shape of<scratchpad>,<tracker>and<branch>rather than anything written to stop it, and an interactive paste is unmeasured.The environment is
CONTRIBUTING.md’s The environment and the gates, and the paragraph below the fence says so: that step is two commands,uv syncand theuv run pre-commit installthat puts the lint gate on a commit, where a line chained to thecdcarries only the first.
The badge row’s links open the runs page filtered to main¶
Every workflow-status badge link at the head of
README.mdcarries?query=branch%3Amain(issue btclib-org/.github#762): section 2 of the organization standard gives the link that filter in the spelling the runs page takes, the image’s?branch=mainbeing ignored there, so a click on a badge reaches the runs it answers for. The pre-commit.ci, Read the Docs and Scorecard badges are outside the rule, and their links reach no runs page.
.gitignore names its environment paths once¶
.env,.venvandENV/no longer appear in two blocks that disagreed on the rest of the list (closes #371): one namedvenv/andenvtest/, the othervenv*/, and git already took the union of both, so nothing was wrongly ignored — a reader looking for the rule on virtual environments found two answers instead of one. The surviving list keeps the literalvenv/,envtest/andvenv.bak/: nothing in this tree creates a virtualenv under a namevenv*/would catch andvenv/would not, and the one alternate-interpreter environmentCONTRIBUTING.mddocuments is dot-prefixed (.venv-3.10), which neither pattern matches.
The index wait declines a draft pull request too¶
pypi-install.yml’swait-for-indexjob takes the draft conditioninstall-publishedcarries (closes btclib-org/.github#811): the wait holds back the install job alone, which a draft skips, so a draft pull request runs neither. What the comment above the job refuses is anif:keyed on the release tag, which would leave a release the only thing that ever runs the wait; the schedule, a dispatch and a release call run it under the draft condition whatever a pull request is.
The uv_build floor is bounded rather than copied from a rev¶
[build-system] requiresmoves fromuv_build>=0.12.5,<0.13touv_build>=0.12.7,<0.13(closes #368): section 3 of the organization standard has a floor above the0.12.0boundary be alignment, and what the comment states is the range that alignment means – above the boundary, and no higher than the rev.pre-commit-config.yamlpins foruv-pre-commit(grep -A1 'astral-sh/uv-pre-commit' .pre-commit-config.yaml). Not an equality: pre-commit.ci moves that rev upward on this repository’s own weekly schedule with nothing moving this number, so a floor written equal to it stops being equal at the next bump where a floor under it stays right.The comment re-derives the
0.12.0boundary by calling the backend’s own hook at each version (closes #378), where it asked foruv buildunder a pinnedrequires. That is the way the standard names as the way not to measure it: a backend older than the one running is always the ceiling-below case, souv buildfalls back to the copy it bundles and only warns, answering for that copy rather than for the pin. The hook prints the archive’s name rather than its contents, so the block names what to read inside it, andbtclib-org/.github#143is where the per-version run sits.A sibling’s floor is no second reason for the same number. pre-commit.ci moves this rev on this repository’s own weekly schedule and a sibling’s on the sibling’s, so two trees taking their floor from their own rev sit at different numbers between the bumps.
check-sdistandpyromatake the specifier[build-system]declares, their hook environments being wherebuildfinds the backend when it does not isolate, andRELEASING.mdnames the range rather than repeating it.
The preamble names no command for counting entries¶
CHANGELOG.md’s preamble no longer names grep -c '^- ' as what
counts this file’s entries (issue btclib-org/.github#829): that
pattern answers about bullets at the margin – an entry’s separate
claims where a ### heading names an entry, whole entries where one
names a theme – and section 9 of the organization standard lets an
entry making a single claim have a paragraph for a body, which no bullet
pattern reaches. Counting the headings does not answer either, the two
kinds of heading sitting in one file. What the sentence keeps is its
point: neither file states a number, because a stated one is a line
every open branch has to edit under a driver that would keep both
sides’.
The smoke script’s pragmas name the case each one is¶
A
# pragma: no coverin.github/scripts/rpc_smoke.pycarries its own case after--, on the line it sits on (issue btclib-org/.github#838). Section 8 of the organization standard puts a pragma’s reason there, which is what makesgit grep -nE 'pragma: no cover$' -- '*.py'a gate rather than a census: it names every site giving no reason at all, where a reason written above the line is out of its reach. The reason all of them share stays where it is, in the module docstring – what is excluded takes a realBitcoinCoreRpcClienttalking to a real node, and mocking that node would be the recording this script exists to not trust – and what a line adds to it is what that function does with the node: the bitcoind it starts, the wallets it needs loaded, the reply it reads off the wire before the client classifies it.pyproject.toml’s[tool.coverage.report]comment says where a pragma’s reason goes (issue btclib-org/.github#838). A reason “beside” the pragma reads as either half, and the sentence this tree writes about the rule is what a reader here goes by, so it names the position, the grep that position buys, and the comment or docstring a reason too long for the line goes to as well.
CLAUDE.md’s worktree fence takes the standard’s converged form¶
The fence’s push is
git -C "$WT" push origin HEAD:refs/heads/<branch>, and thecd "$WT"line above it is gone (issue btclib-org/.github#824): acdbinds the shell that runs it, so a session that runs each line as its own command starts the next one in the directory it began in — the primary checkout — and a push there offers that checkout’sHEADinstead of the worktree’s. The paragraph the standard carries below its own fence comes with the change, giving the binding and its limit:git -C ""is documented to leave the working directory unchanged, so a-Cwritten against aWTthe session has already lost lands in the same wrong tree, exit 0 and no diagnostic.The create’s condition is stated rather than left implicit. The
>closing<branch>is reached only where the reader’s own directory already holds that name, the<before it having to succeed first, where the sentence it replaces promised a file the paste creates with no condition on it at all.The removal guard’s sentence names the empty value too.
${WT:?}fails on an unset or emptyWT, and the sentence said with no$WTset, which is narrower than the guard; the standard’s next sentence comes with it, that those are the only cases it catches and a$WTan earlier session left holding a path expands like any other.A paragraph below the removal fence names the sha the three converged at,
btclib-org/.github’s20ad654, so a later reader compares this fence against a tree rather than against an issue’s quotation of one. Those three paragraphs are that tree’s bytes; the paragraph about the environment, which sits between the push and the removal, is this tree’s own.The landed entry
CLAUDE.md’s worktree fence carries nouv syncabove describes a fence this one changes. Its first and third bullets speak of acd "$WT"line standing alone and of a sync chained to thatcd, and the fence has nocdnow. What they rest on survives — a line that writes goes in a fence of its own — and the reason below the fence is now unconditional rather than turning on a lostWT: no line of the block moves the shell,git -Cbinding the one command it is given. Its second bullet stands, re-measured on the three-line block: read as a script with-n, unfilled it is a syntax error to/bin/zsh5.9, to thebash3.2.57 macOS ships as/bin/bashand/bin/sh, and tobash5.3.15, and<scratchpad>filled alone, or<tracker>, or<branch>, leaves the same error at the same line.
.gitignore covers the environment the other-interpreter run creates¶
.venv*replaces the literal.venv(closes #375):CONTRIBUTING.md’s other-interpreter run givesUV_PROJECT_ENVIRONMENT=.venv-3.10, and a suffix makes a name the literal does not match, sogit check-ignore -v .venv-3.10found no line to name.venv/stays literal beside it, on the reason the comment above the list gives for it.git statusis silent about such a directory either way: uv writes a.gitignoreholding*inside the environment it creates, sogit check-ignoreis what answers whether this file covers a name.
The uv_build floor moves to the boundary it keeps, not above it¶
pyproject.toml’s[build-system] requiresfloor moves fromuv_build>=0.12.7,<0.13touv_build>=0.12.0,<0.13(issue btclib-org/.github#858): the comment beside it already measured0.12.0as where the sdist’s ownpyproject.tomlbecomes a normalized copy withpyproject.toml.origkept beside it, and the floor sat above that boundary anyway, aligned with the rev.pre-commit-config.yamlpins foruv-pre-commit— one of the two alignments section 3 of the organization standard now names as rejected. The comment is rewritten to say the floor sits at the boundary, and to record, without presupposing a floor above it, that both that rev and a sibling’s own number were weighed and declined. This supersedes the entry above for #368: the floor it sets and the reason it gives for sitting above the boundary are what move here..pre-commit-config.yaml’scheck-sdistandpyromahooks follow the same floor in theiradditional_dependencies: both are a copy of[build-system] requires, each carrying a comment saying so, and both move fromuv_build>=0.12.7,<0.13to the sameuv_build>=0.12.0,<0.13. Widening the range this way is a no-op for what either hook resolves —<0.13still caps both, and pip takes the newest version in range either way — so nothingcheck-sdistbuilds changes.[tool.uv] required-versiondoes not move: it answers what uv readsuv.lockwith rather than what range an isolated build may resolve for the backend, and section 1 of the organization standard already set it at the current ceiling.
Four comments say what max-doc-length reaches¶
The
[tool.ruff.lint.pycodestyle]comment no longer calls the key “the rest – the half of a file that carries the reasoning” (issue btclib-org/.github#841): a comment following code on its line is part of that half too, and the key does not hold it to any width – section 9’s own ruling, landed there first. The comment keeps the MD013 anchor for the 80, and the W505-is-inert-without-the-key sentence; the URL exemption sentence beside it is left untouched – it is section 5’s own wording, transcribed, and itself wrong for a different reason (btclib-org/.github#866) that belongs to the standard to settle.The
"line-too-long"ignore reason drops the same claim stated the other way round: it no longer saysmax-doc-lengthmeasures “the prose the formatter never touches” instead of E501. This tree ignoresline-too-long, so with a trailing comment outsidemax-doc-length’s reach, nothing here measures one at any width..pre-commit-config.yaml’s yamllint comment carried the same claim in a second file: “the hook above holds markdown to 80 columns and ruff holds Python comments and docstrings to the same” now names what ruff holds – a docstring and a whole-line comment. That sentence is there to say why yaml needed a width of its own, and it says it with the same force once the claim about Python is true; the issue’s own census names this site in four trees, none of them counted when it was filed.The
toml-comment-widthcomment six lines below it carried it a fourth time, and no census names that one: it said “a Python comment has W505” and that[tool.ruff.lint.pycodestyle]“names 80 columns as the width of every comment in this tree”. W505 reaches a docstring and a whole-line comment, so it now names those two. The sentence is there to say that toml prose was the last kind held to nothing, and that is what it still says. Found by the review of this branch rather than by the issue, in the file this branch had already opened – which is why it is here and not filed.
The line-too-long ignore reason names what E501 reports here¶
The reason no longer offers a
# type: ignoreor a# ruff: ignorepushing an otherwise-short line past the limit as what E501 is left to find (closes btclib-org/.github#870): the rule measures a line up to the start of a pragma it recognises, and both of those are among them, so it passes over every line of that shape in this tree. The other shape the reason named, a url in the module’s own docstring, goes the same way: a line ending in a url that begins before the limit is passed over, in a docstring and in a trailing comment alike. No Python line here that holds a url runs past 88 columns anyway.What the reason names instead is what the rule reports:
DEFAULT_SIGNET_CHALLENGE’s recorded script, whose# pragma: allowlist secret– the secret scanner’s, and no pragma E501 recognises – carries each of its lines past the limit. The comment gives the command that lists them, as the entries below it do.Measured against
btclib-org/.github, which does not ignore the rule, with the pinned ruff: an ordinary trailing comment at the same width is the control, and it is reported. What decides is where the pragma begins and not that it ends the line – the same pragma with words after it is passed over just as well, and it is reported once the code ahead of it is itself past the limit.
The max-doc-length comment leaves the exemptions to section 9¶
The comment no longer states the URL amnesty unconditionally (issue btclib-org/.github#866): W505 passes a line over where its last word holds
://and everything ahead of that word fits the width, and reports one whose prose is already past the width before the link begins. The comment points at section 9 ofbtclib-org/.github, which states that condition and the pragma one beside it, rather than wording either a second time here.This supersedes Four comments say what
max-doc-lengthreaches where it leaves that sentence standing as section 5’s own wording: section 5 now calls the exemption conditional and points at section 9 for the condition, so what that entry left to the standard is settled.Measured with this tree’s own configuration: a whole-line comment ending in a link that begins past the width is reported, the same comment with the link beginning before it is not, and one carrying words after the link is reported again. The controls are an over-width comment holding no link and one whose link is spelt without the
://, both reported.ruff --isolateddrops thepreviewthis tree sets and measures a line whole rather than stopping at a pragma, which is the wrong instrument for that half.
toml-comment-width’s comment claims no other tool’s amnesty¶
The comment no longer gives the pattern’s exemption to MD013 and to ruff, as what those two give “a comment ending in a link that cannot be broken” (issue btclib-org/.github#843):
.{80}\S*[ \t]reports a line only where whitespace is left past column 80, so a comment whose overflow is one unbroken token is exempt whether or not that token is a link. What replaces the clause states the pattern’s own predicate and names no other tool, which is the wording btclib-org/.github#843 settled for the trees carrying this hook.The exemption is not W505’s, and neither rule’s exemption contains the other’s: at this tree’s
max-doc-length = 80, ruff reports a 94-column whole-line comment whose overflow is one unbroken non-link token and this pattern passes it over, and this pattern reports a 97-column comment opening with# type:that ruff passes over. A# abcd:of the same width and shape is reported by ruff, which is the control saying the pragma’s spelling is what does it, and an ordinary line of words at that width is reported by both. Measured at ruff 0.16.5, the rev.pre-commit-config.yamlpins and the versionuv.lockresolves.The sentence under the replaced clause stays: with no amnesty this hook would report a whole-line toml comment that MD013 passes over in a
.mdand that, where the unbroken token is a link, W505 passes over in a.py. Its “the rules it mirrors” reads back to the paragraph above, which names MD013, W505 and yamllint as the rules this hook was added beside. Measured at the markdownlint rev this tree pins: a 92-column paragraph line whose overflow is one unbroken token is passed over, and the same line with a space inserted at column 80 is reported, which is the control saying that silence is an absence.The hook’s
name:andentry:are untouched: the lexical amnesty the one states against the positional rule the other implements is the rest of that issue, one decision for every tree carrying the hook rather than this tree’s to take alone.
scorecard.yml’s actions: read carries its own reason¶
The job’s
actions: readsits under a comment naming what asks for it, andcontents: readmoves up to sit under the comment that already explained it (issue btclib-org/.github#871):ossf/scorecard’schecks/raw/github/packaging.gocallsClient.Actions.ListWorkflowRunsByFileNamefor a workflow file its Packaging check recognizes as publishing a package, whichrelease.ymlis by itspypa/gh-action-pypi-publishstep. The wording isbtclib-secp256k1’s own, landed there first.contents: readis granted here rather than inherited: a job’s ownpermissions:block replaces the workflow-level one, so a scope the block omits isnonefor that job, which one run oftest.ymlshows — thechangesjob, whose block namespull-requests: readalone, logs noContentsin itsGITHUB_TOKEN Permissionsgroup, while a job of the same run declaring no block logsContents: readunder the same workflow-level grant.
The smoke script’s docstring names the pragma rule rather than quoting it¶
.github/scripts/rpc_smoke.py’s module docstring writespragma: no coverin backticks and without its#(closes btclib-org/.github#890): section 8 of the organization standard asks prose quoting the rule to write it that way, the#form being what coverage’s own exclusion pattern matches, which reads the raw source a line at a time rather than tokenizing it first. What it costs is not the percentage — what such a match excludes is the statement the matched line belongs to, and a module docstring is no statement coverage measures — it is a pattern naming a line where nothing is excluded and nothing is wrong. The sentence’s verb iscarries: without the#its backticked text names the rule rather than quoting the comment those lines hold, and a function carries a rule rather than being one.
The token grants are read from the workflows rather than bounded in prose¶
REPOSITORY.md’s Token permissions gives the command that answers which jobs elevate (issue btclib-org/.github#891):git grep -n ': write$' -- .github/workflowsis anchored so that a comment naming a permission stays out of the answer, which here is no hypothetical — workflows declaring no write grant carry comment lines that name one. The sentence it replaces calledrelease.yml’s the only grants in the tree, whereclaude-review.yml,codeql.ymlandscorecard.ymleach hold their own.scorecard.yml’s workflow-level comment sends a reader to the job block instead of naming what the job does with its grants (issue btclib-org/.github#891). The comment it replaces named part of a block that the same file, below it, says is read as a whole: a scope the block omits isnonefor that job. The wording isbtclib-secp256k1’s, so a port of it into the remaining trees copies rather than decides.The job’s own
permissions:block is untouched. It is byte-identical tobtclib-secp256k1’s andbtclib-node’s, and btclib-org/.github#893 is where the clause quantifying over every scope is corrected for the three together.
scorecard.yml’s contents: read comment names the scope it is about¶
The comment says what omitting
contentsfrom the analysis job’s block would do rather than what omitting any scope would do (issue btclib-org/.github#893).metadatais where the wider reading fails:actionlint, which the lint gate runs, refuses it as a scope apermissions:block may name, and the analysis job’sGITHUB_TOKEN Permissionsgroup logsMetadata: readall the same.scorecard.yml’sactions: readcarries its own reason and The token grants are read from the workflows rather than bounded in prose above both state that consequence for every scope, and this entry supersedes each of those sentences; both hold ofcontents, which is the scope the run the first of them cites measured. That second entry also records the job’s block as matchingbtclib-secp256k1’s andbtclib-node’s, an identity each tree’s port of the issue moves as it lands.
A cross-repository reference in this tree’s prose names its owner¶
RELEASING.md’s merge step andREPOSITORY.md’s Token permissions citebtclib-org/btclib-secp256k1#288andbtclib-org/btclib-secp256k1#281(issue btclib-org/.github#642): section 9’s A reference to another repository is qualified is a general bullet of that section rather than one of itsCHANGELOG.mdsubsection’s, so it reaches a markdown paragraph the way it reaches an entry. Its one exemption is a pull request’s closing keyword, which the forge reads.CHANGELOG.md’s own unqualified references stay as they are: Nothing already written is rewritten binds what is written next, and.gitattributesputs the file undermerge=union.Neither paragraph is rewrapped: the qualifier leaves both lines inside the width their paragraphs already keep, so no word moves.
Token permissions names its own exceptions and reads a grant’s position¶
Two elevations on one job is stated as the exception to the one-elevation shape, with the reason for each pair (closes btclib-org/.github#891).
release.yml’sattest,scorecard.yml’sanalysis, andclaude-review.yml’sreviewandmention, each hold two write grants, which the shape sentence alone read as excluding. The wording isbtclib’s; every clause was re-derived against this tree’s own workflows before being taken, and the one clause not taken — the replace semantics of a job’s ownpermissions:block — stays at the key inscorecard.ymlrather than being repeated here.REPOSITORY.md’s Token permissions hands the readergit grep -nE '^ +[a-z-]+: write([[:blank:]]+#|$)' -- .github/workflows(issue btclib-org/.github#897): the: write$form it replaces drops a declaration carrying a trailing comment along with the comment lines it was there to exclude. The token grants are read from the workflows rather than bounded in prose above names the$as what keeps a comment naming a permission out of the answer, and that sentence describes a command this file no longer carries: what keeps one out now is the key’s own position, a comment line opening with a#where[a-z-]cannot match. This tree carries no declaration with a trailing comment today, so both forms answer the same twelve grants and the agreement proves nothing.The comment branch takes
[[:blank:]]rather than a space:actionlint, which the lint gate runs, accepts a tab between a grant and its comment, and the census reads that line rather than resting on the hooks that rewrite the tab.The shapes named beside the command as outside its answer are ones this tree’s gate leaves standing:
actionlinttakespermissions: write-all, a flow mapping and a quoted key or value each as a grant, andprettierkeeps each rather than rewriting it into what the pattern reads, as it does a double space or a trailing space.Folding a shape into the pattern would leave the next one out:
contents: >-withwriteon the line below is a grantactionlintaccepts andprettierhands back unchanged, and a line-oriented pattern reads a spelling where a permission is a value in a parsed document.
exclude_patterns names what this tree writes under docs/source/¶
docs/source/conf.pycarriesexclude_patterns: list[str] = [](closes btclib-org/.github#418): section 2 of the organization standard, under The documentation, is where that rule and its reason are. Nothing here writes underdocs/source— everysphinx-buildthis tree runs or documents,docs.yml’s and.readthedocs.yaml’s and the twoCONTRIBUTING.mdspells, is givendocs/build/htmlor$READTHEDOCS_OUTPUT/htmlto write into — so the list has no path of this tree’s to name._buildis dead here by where this tree builds, not by anything sphinx does: section 2 is where that distinction is drawn, and it is the one an entry calling the whole list inert would lose.Thumbs.dband.DS_Storeare dead for a different reason again,source_suffixhere declaring.rstand.md.The line keeps the
list[str]annotation: what the issue asks for is one line across the documentation trees, and this is that line. It is not there for a type check — mypy runs here oversrc/bitcoin_core_rpc,testsand.github/scripts, which leavesdocs/outside the only one this gate has.The three comment lines above the key went with it: they are
sphinx-quickstart’s own template text, describing the key — what it matches, and that it reacheshtml_static_pathandhtml_extra_path, neither of which this tree assigns — and saying nothing about this tree. The comment giving the reason there is nohtml_static_pathis this tree’s own and stays where it is.
conventions_test.py says what it reads and where it departs¶
The docstring of
tests/conventions_test.pysays what this copy reads and which of its departures are decided (closes btclib-org/.github#690): section 14 of the organization standard asks that header of every copy of the module, and the failure mode is why it asks — a defect in the parsing the shared job needs sits in every copy carrying it and turns nothing red anywhere, so a fix that is not about one tree’s rows or its root has to be carried to every copy in the same campaign.What it reads is
tests/README.mdbeside it, and a module a row names resolves against that same directory.The departure it records is
_CONVENTIONS, which transcribes section 7’s list rather than reading it off the standard: the standard is another repository’s file, so a copy is the only form the list takes here.The paragraph naming how the organization’s suites each call their convention tests is gone from that docstring:
tests/README.mdalready says it, so the docstring was the second statement section 9 refuses, and section 14 names that naming point as the thing the header is not about. The conclusion it reached — the audit reads a declaration rather than a directory — stays; only its premise now points attests/README.mdinstead of restating it.
conventions_test.py’s comment and docstrings say what is true here¶
The comment above
_NOT_TESTEDno longer saysnoneis the answer this repository gives (issue btclib-org/.github#903): theNot tested hereline oftests/README.mdnames conventions, so the clause was false here, and the sibling list it cited as the reason forre.DOTALLsent a reader to another repository for what this tree’s own answer shows, wrapping across lines as it does. The wording isbtclib-node’s, which keepsnoneexplained as the literal the code below special-cases without claiming it.test_the_table_is_not_empty’s docstring no longer says every assertion below it goes quiet on a table the row pattern stopped matching (issue btclib-org/.github#904): the assertions parametrized on the rows are skipped for an empty parameter set, and the two-halves assertion, which is not parametrized, fails on that same table naming every convention the table declared as accounted for by neither half. A retitled heading reaches neither, the section slice asserting while the module is imported. What this assertion adds is naming the table rather than the conventions, and the docstring now says so.The two-halves docstring states no count of what the table leaves silent (issue btclib-org/.github#905): the number it gave no longer matched section 7’s list, and any number there expires the next time that list grows.
btclib’s copy of the module writes “the rest”.
conventions_test.py’s module docstring names the ways and counts nothing¶
The last paragraph of
tests/conventions_test.py’s module docstring states no number of the assertions below it (issue btclib-org/.github#906): it said the four assertions below are the ones that fail on the ways a declaration actually rots, and the ways it listed left outtest_the_table_is_not_empty, which is what fails where a column added to the table or the backticks dropped stops the row pattern matching. An unmatched table is the first of the ways now, the ways standing in the order the assertions are defined in rather than one to an assertion — that table fails the two-halves assertion as well. The number is gone rather than corrected, section 9 of the organization standard refusing a stated total, a line every open branch has to edit. The wording isbtclib-node’s, landed there as btclib-org/btclib-node#903; btclib-org/.github#906 is filed against the copies of this module across the organization and stays open.
Token permissions states what caps a called workflow¶
REPOSITORY.md’s Token permissions says the caller’s grant is the cap on every job of the called workflow (closes btclib-org/.github#914): it said that grant replaces the callee’s default outright, a substitution refuted by a run of the very call it describes. This tree’sv2026.9.3release run 33753542084, at614a8720, whererelease.yml’stestjob grantscontents: readandpull-requests: readandtest.ymldeclarescontents: readat its own top level with a block onchangesalone: each job oftest.ymldeclaring no block of its own logsContents: readandMetadata: readin itsGITHUB_TOKEN Permissionsgroup, and not thepull-requeststhe caller grants. Under substitution each would hold the caller’s list. The control that could have failed ischangesitself, in the same run and through the sameuses:, which logsPullRequests: read, so the group does report that scope where a block asks for it.What naming
pull-requests: readalone would do to the other jobs oftest.ymlis no longer drawn (issue btclib-org/.github#912): the paragraph said it would leave them without evencontents: read, refusing their checkout steps, where what a run does with a called workflow’s top-level declaration outside the caller’s list is unmeasured. The refusal a job’s own declaration draws — the whole run refused before any job of it starts, btclib-org/btclib-secp256k1#281 — is not evidence for it, and the paragraph names the case as unmeasured instead.
conventions_test.py splits the Not tested here list at its separator¶
tests/conventions_test.pysplits the collapsed list at a semicolon and a space, and collapses no name after that (issue btclib-org/.github#911): the line above the split has already replaced every run of whitespace, newlines included, with one space, so collapsing a piece of it again isstrip()with nothing left to strip. The comment giving an eighty-column wrap that falls inside a name as the reason for that second collapse goes with it: the Not tested here list oftests/README.mdwraps at a semicolon, so no name in it is broken across lines here.The separator keeps its space rather than becoming the semicolon alone, and what the comment says now is why: the split is lossless —
sep.join(s.split(sep))issfor any non-emptysep— so a separator the declaration wrote some other way leaves the name whatever the split did not take, and the assertion that every name listed is one of section 7’s reports it. The semicolon alone would take a separator written without its space too, and nothing would report it.That assertion’s message quotes the names it read out of the declaration: a semicolon written with a space on each side is consumed by the split and leaves the name a trailing one, which unquoted reads as a name the same message goes on to list as known. The comments and this message are word-identical to
btclib’s, landed there asbd9e3e87, the issue asking one decision of every copy of this module rather than one per tree.
templates_path goes, this tree keeping no templates¶
docs/source/conf.pyno longer assignstemplates_path(issue btclib-org/.github#901): nothing is tracked underdocs/source/_templates,docs/holds no template of its own, and nohtml,jinja,j2ortmplfile is tracked anywhere, so the key named a directory this repository does not have. Section 2 of the organization standard is where the rule this converges on lives — “templates_pathnames the directory underdocs/source/where the tree keeps its own templates, and a tree keeping none does not carry the key” — and it names writing the key empty, the wayexclude_patternsbeside it is written, as the rejected alternative.No
# no templates_pathcomment goes in its place, though this file explains the absence ofsphinx.ext.todo,suppress_warningsandhtml_static_paththat way: section 2 carries the whole reason for this key, and section 9’s One fact in one place is then what decides it. The# no suppress_warningsblock above keeps the blank line it already had, so it still reads as that key’s absence comment rather than as a comment onexclude_patterns.
The declaration says what it is and leaves its assertions to the module¶
tests/README.md’s Convention tests sentence names no assertion ofconventions_test.py(issue btclib-org/.github#910): the list it carried left outtest_the_table_is_not_empty, which is what fails where a column added to the table or the backticks dropped from its second column stops a row parsing — either mutation failingtest_the_two_halves_account_for_every_conventionas well, the assertions parametrized on the rows being skipped on the empty parameter set. That module reads the heading, the row pattern and the Not tested here line rather than the sentence, so a list of its assertions there is a second statement of what it checks with nothing holding the two together, and completing the list would leave the next assertion free to go out of step the same way. What replaces it says what the declaration is — the table and the “Not tested here” line under it, accounting between them for the whole list — and sends a reader wanting what those assertions catch to that module’s docstring, section 9 of the organization standard asking that the second statement point at the first.The halves are named rather than dropped with the list: nothing else in the section says what the line under the table is, the paragraphs under it giving the reason each convention it names is absent rather than that line’s own place in the declaration. The tail clause is the one this copy already carried. The shape is the organization’s decision, taken on btclib-org/.github#910 and landed in
btclibas btclib-org/btclib#1869; btclib-org/.github#910 stays open for the copies in the other trees.
The convention-test entries’ open-issue sentences are superseded¶
The entry
conventions_test.py’s module docstring names the ways and counts nothing above ends by saying btclib-org/.github#906 stays open (issue btclib-org/.github#946): that issue closed on 2026-09-08.btclibtook the same change at btclib-org/btclib@385406f1,btclib-secp256k1at btclib-org/btclib-secp256k1@8d84015d,btclib-nodeat btclib-org/btclib-node@8e0a0a48 andbtclib-benchmarksat btclib-org/btclib-benchmarks@dc47a41f.The entry The declaration says what it is and leaves its assertions to the module above says btclib-org/.github#910 stays open for the copies in the other trees (issue btclib-org/.github#946): that issue closed on 2026-09-08.
btclibtook the same replacement at btclib-org/btclib@bd4a4654,btclib-secp256k1at btclib-org/btclib-secp256k1@d80e2c07 andbtclib-nodeat btclib-org/btclib-node@c8586937.Both sentences were true when they landed: each issue closed after the entry naming it did, the last copy of its change landing in another tree. A sentence naming the tree and the sha of a landing stays true as the other trees move, which is why the bullets above give those rather than a state.
The lint gate is not installed as a git hook¶
CONTRIBUTING.md’s The environment and the gates no longer asks foruv run pre-commit install(issue btclib-org/.github#949):pre-commit installwrites into the common git directory, which every worktree of this repository shares, so a worktree’s ownrev-parse --git-path hooksanswers with the primary checkout’s.git/hooksand one session installing the hook installs it for every other. The environment step isuv syncalone, and the paragraph saying why sits under Check exit codes, not filtered output, where.github,btclib-nodeandbtclib-secp256k1each put theirs.CLAUDE.md’s paragraph below the worktree fence calls that step one command. Its clause about a line chained onto the block carrying only the first of two goes with the second command, and what the paragraph rests on is untouched: a line that writes goes in a fence of its own, and no line of the block moves the shell.The landed entry
CLAUDE.md’s worktree fence carries nouv syncabove says that step is two commands,uv syncand theuv run pre-commit installthat puts the lint gate on a commit. It is one command now, and the fence that entry is about is unchanged.UV_PROJECT_ENVIRONMENT’s reason is what uv does to the environment, not what a hook does with it:uv run --python <version>removes.venv, builds it again on that interpreter and with the command’s own group set, and leaves it there. Both the paragraph in The environment and the gates and the line beside the matrix command in Reproducing what CI runs gave the reason as agit commitdying inside the hook this change declines to install.
.gitignore’s envtest/ is dropped¶
The
# Environmentslist no longer carriesenvtest/(closes #392):git log -S 'envtest' mainfinds only the commit that introduced the whole list and the one that later merged it with a second one, so nothing in this tree’s history has ever written a directory of that name.
The smoke script’s docstring stops claiming wait_for_rpc is untested¶
.github/scripts/rpc_smoke.py’s module docstring now nameswait_for_rpc’s retry policy among whattests/rpc_smoke_test.pycovers, and scopes the “mocking that node would be the recording this script exists to not trust” sentence towait_for_rpc’s own success path along with the rest of the file’spragma: no coversites (closes #382):wait_for_rpccarried no entry in either list, so it read as covered by neither claim, wheretests/rpc_smoke_test.py’s own docstring already named the retry policy as tested with a stub – a status that clears from one that does not, and the credential a node has already refused. The one branch a stub cannot reach is the answer that actually ends the wait, which keeps both the sentence and the pragma naming it.
The UV_PROJECT_ENVIRONMENT prefix names the interpreter, not 3.10¶
The prefix in The environment and the gates named a fixed
.venv-3.10regardless of the interpreter a--python <version>command selected (closes #386): a command run with--python pypy3.11and the same fixed prefix would build the pypy3.11 environment under a directory named for 3.10, so.venv-3.10on disk would not hold what its name says. The prefix now names the interpreter the command it precedes selects —.venv-<version>,.venv-pypy3.11for--python pypy3.11.The rule carried no exemption for
--no-project: that flag builds an ephemeral environment rather than rebuilding.venv, so a command carrying it needs noUV_PROJECT_ENVIRONMENTprefix, and the paragraph now says so.
.github/’s bare #1166 and #189, which #221 did not reach¶
test.yml:67,test.yml:251,dev-version/action.yml:2,release.yml:223andrelease.yml:339’sissue #1166are qualified asbtclib-org/btclib#1166(closes #398), matching whatCONTRIBUTING.mdandRELEASING.mdalready write for the same referent – the earlier fix, closing #221, reached only those two files, leaving.github/’s five sites bare. A bare#1166in this tracker resolves nowhere –gh api repos/btclib-org/bitcoin-core-rpc/issues/1166answers 404, this tracker’s own highest number sitting in the hundreds – and the five sites under.github/had stayed bare.test.yml:148’s#189is qualified asbtclib-org/btclib-secp256k1#189, and#145beside it stays bare. Read in this tracker, a bare#189resolves to this repository’s own closed issue aboutdocs/source/conf.py’s toctree comment undercounting its pages – closed, never merged since it was never a pull request, and sharing nothing with thechangesjob the sentence names, so the reference pointed at the wrong place with no 404 to signal it. Onlybtclib-org/btclib-secp256k1#189, the one the sentence means, is a merged pull request.#145is this repository’s own issue, about the permission that job needs on the caller’s side, and is correct bare.No arithmetic gate is added for a bare
#Nabove this tracker’s own highest number. It would have caught the#1166sites; it would not have caught#189, which sits inside the range and resolves to a real, closed (not merged) issue of this repository – the shape btclib-org/.github#899 found readingbtclib-node’s own.github/the same way, a number inside range resolving to the wrong issue in silence there too. Reading the sentence a bare reference sits in is what the arithmetic cannot stand in for.
regebro/pyroma’s pin returns to its latest released version¶
.pre-commit-config.yaml’sregebro/pyromaentry is pinned to5.0.1(closes #413): the weeklyautoupdatepull request had moved it to5.1b1, exactly what this file’s ownpinned-revhook exists to reject – “every pre-commit rev names a released version” –gh api repos/regebro/pyroma/tagsnaming5.1b1and5.1b2as pre-releases of an unreleased5.1and5.0.1as the latest actual release.
CLAUDE.md’s worktree name and advance clause take the standard’s wording¶
The worktree paragraph gives the argument for the order of
wt-<tracker>-<issue>-<repo>-<role>(issue btclib-org/.github#854): most general part first, an issue filed inbtclib-org/.github’s tracker being the key and the repository a detail of it. The sentence saying the name sorts every worktree of one issue together sits with that argument instead of inside therepoclause, sorting following from the order rather than fromrepobeing present. The two example issues are qualified, a bare#255and a bare#177each naming an unrelated closed item of this tracker.The clause on
refs/heads/mainnames the ref it may be moved onto (issue btclib-org/.github#919),origin/main, byte for byte asbtclib-org/.github’sCLAUDE.mdcarries it. What it replaces forbade advancingrefs/heads/mainwith work that is not yours, which is what the fast-forward the same section prescribes for bringing the primary checkout forward does.
Prose outside the shipped module stops naming the library downstream of it¶
Comments and docstrings across the workflows,
.pre-commit-config.yaml,pyproject.toml,COMPARISON.md,.readthedocs.yaml,CONTRIBUTING.mdandtests/conftest.pystop naming the library downstream of this package to justify a choice made here (closes #421): theno-downstream-name-in-packagehook only readssrc/bitcoin_core_rpc/, and section 9 of the organization standard reaches this repository’s own prose too, this repository being upstream of that library the same way the shipped module is. Each rewritten line keeps the fact it stated – a shared release schedule, a shared badge split, a shared URL key spelling, a cache exclusion, a documented incident, a type-stub cost – and drops the comparison. The hook’s ownentry:pattern, the JSON fixtures’ literal"id": "btclib", and the smoke script’s wallet and probe names are untouched: none of them is prose justifying a choice by comparison, and rewriting any of them would change what it tests rather than clean up a comment.
Three gate aggregates take btclib’s own allowlist shape¶
integration-bitcoind.yml’sintegration-passedstep decides in a shell loop overjoin(needs.*.result, ...)rather than in a booleanif:expression (issue btclib-org/.github#475): the boolean form is the shape btclib-org/btclib#1001 records failing – a matrix cell dying in “Set up job” leavesneeds.*.resultwithout'failure'in it, theif:then evaluates false, and a skipped step leaves its job green. The loop always runs, and an allowlist ofsuccessandskippedfails on anything else instead of only on the two named results. Section 10 of the organization standard names a different shape for this step – reading the run’s own job listing from the API – andrelease.ymlcalls this workflow, where that listing is the caller’s run rather than this workflow’s own (btclib-org/.github#474).test.yml’stest-passedstep takes the same shape, in place of acasedenylist over the same join (issue btclib-org/.github#537): a denylist passes any result outside its own patterns, and an empty join passed both forms alike; thecase ",$results," in *,,*)guard ahead of the loop is what btclib-org/btclib#1454 asked for.codeql.yml’scodeql-passedstep takes the same shape too, for the same boolean-if:reasonintegration-passeddid – found reading this tree’s own aggregates while porting the other two, and fixed alongside them rather than filed.
The coverage job’s pytest invocation drops its duplicate report flag¶
test.yml’scoveragejob runspytestwith nothing after it (issue btclib-org/.github#433):--cov-report term-missing:skip-coveredduplicated[tool.coverage.report]’s ownshow_missingandskip_covered, which already give the barepytestaddoptsruns the same report – measured identical, table for table, with and without the flag.CONTRIBUTING.md’s own reproduction of this job is updated to match.
The gate runs section 4’s local hooks¶
reasonless-coverage-pragmarefuses a#-commentpragma: no coverorpragma: no branchwith nothing after it on its line (issue btclib-org/.github#965): the hook’sentry:andtypes:arebtclib-org/.github’s ate1069d1, andtypes: [python]reaches the smoke script, whose pragmas carry their reason on the line already, so the tree is green under the hook at the run that adds it.unquoted-placeholderrefuses a placeholder standing as a whole argument in quotes, in every markdown file butCHANGELOG.mdandRELEASE_NOTES.md(issue btclib-org/.github#706): section 9 of the organization standard keeps the rule and section 4 names the quotes the pattern exempts and the three shapes it cannot see; the exclusion is the standard’s own, both files being append-only. The tree is green under the hook at the run that adds it.check-changelogrefuses a###heading repeated in this file’s open section, an issue(closes #N)by two entries, and a heading with no blank line above it (issue btclib-org/.github#21):.github/scripts/check_changelog.pyisbtclib-org/.github’s ate1069d1byte for byte, which section 14 owes every repository, and the hook sits ahead ofmarkdownlint-cli2, whose--fixrepairs the seam the third check names.pyproject.tomllets the script print as it lets the scripts beside it: what it prints is its report. The script’s_BLANK_LINEcarries the attribute docstringcheck-docstring-firstreads as a second module docstring, so that hook’sexclude:names the script beside the threesrc/modules it already names for the same shape, the script being owed byte for byte.toml-comment-width’sname:states the pattern’s own predicate, an unbroken final token past byte 80 (issue btclib-org/.github#843, issue btclib-org/.github#885): the width is bytes, pygrep matching the pattern against the raw line, and the comment above the hook says “byte 80” where it said “column 80”; theentry:is unchanged.toml-comment-width’s comment claims no other tool’s amnesty’s last bullet left thename:to a decision every tree carrying the hook takes together, andbtclib-org/.githubtook it at51f717e..yamllint.yaml’sallow-non-breakable-wordscomment states only the setting’s own predicate (issue btclib-org/.github#883): the clause comparing it to MD013’s exemption for a bare URL goes rather than being corrected, which is what btclib-org/.github#883 settled, and the file matchesbtclib-org/.github’s ate1069d1byte for byte, which is section 14’s comparison.
codeql.yml’s aggregate reads the run’s job listing¶
codeql.yml’scodeql-passedjob asks the run’s own job listing through the API rather than readingneeds.*.result(issue btclib-org/.github#982): section 10 ofbtclib-org/.github’sREADME.mdkeys the shape on whether something in the tree calls the workflow, and nothing here callscodeql.yml, whererelease.ymlcallstest.ymlandintegration-bitcoind.yml, sotest-passedandintegration-passedkeepneedsand onlycodeql-passedtakes the listing. The job elevates toactions: readfor the call, a read the write-only commandREPOSITORY.md’s Token permissions hands the reader does not list, as it does not listanalyze’s.Its allowlist names
successalone rather thansuccessandskipped(issue btclib-org/.github#990):analyzecarries no condition narrower than this job’s own draft/closed one, and no job of the workflow is conditional on the diff, so askippedrow would mean this job ran without its dependency having, which its ownif:rules out. btclib-org/.github#990 is where the standard settles the answer for every listing step; this tree’s is the narrowing, which can redden a check and never green one.The step’s comment cites btclib-org/btclib#1001 as what the listing carries and
needsdoes not, and no longer cites btclib-org/btclib#1454: the empty join that issue guards against has no counterpart in a listing, so the paragraph on thecaseguard goes with the loop. The entry above titled Three gate aggregates take btclib’s own allowlist shape describedcodeql-passed’s loop overneeds.*.result; that job reads the listing now, and what that entry says aboutintegration-passedandtest-passedstill holds.
REPOSITORY.md reads the wiki and the projects board back¶
has_wikiandhas_projectsare read back beside.visibilityandhas_issues(issue btclib-org/.github#550): section 11 ofbtclib-org/.github’sREADME.mdturns both off on every tree and asks a copy to read them back, so the sentence at the foot that placed them outside the perimeter, on the ground that section 11 stated no rule for either, goes. The call still answerstruefor both here; turning them off is the issue’s other half and not this change.
The hyphen hook reads Python and rst beside markdown¶
no-hyphen-at-end-of-linecarriestypes_or: [markdown, python, rst], underbtclib-org/.github’s own comment with itsREADME.mdread as the organization standard, the phrase the comments beside it use for the standard (issue btclib-org/.github#921): section 4 of the organization standard gives the hook the file types whose prose a build renders, and a docstring reaches that rendering through docutils, which leaves the source break inside the paragraph it builds and lets html collapse it to a space, as it does markdown’s join.The lines the widened hook refused hold their token whole:
docs/source/conf.py’s module docstring wrapped its URL atsphinx-and has it on a line of its own, and the docstring oftests/transport_test.py’stest_a_fresh_connect_failure_does_not_poison_the_poolwrappedevery-failure-is-at its last hyphen and carries the compound on one line.
The mention job’s pull-requests grant carries its reason¶
claude-review.yml’smentionjob says at itspull-requests: writeline what the grant is for (issue btclib-org/.github#915): the comment is# what posting the reply takes, the linebtclib-nodeandbtclib-benchmarkscarry above theirs, taken byte for byte so that one reason is not worded twice across the copies. Thereviewjob’s grant and theid-token: writebeside this one each carry a reason at their line, andREPOSITORY.md’s Token permissions says of the four two-elevation jobs that the reason for every pair sits where it is declared; this grant carried none. Section 14 ofbtclib-org/.github’sREADME.mdkeepsclaude-review.ymlout of the list itstests/verbatim_test.pycompares — the file “is owed by every repository section 11 governs, and section 15’s existence loop is what checks that — not this list” — so nothing red follows from the copies disagreeing at this line.
The issue-form hooks join the check-jsonschema block¶
check-github-issue-configandcheck-github-issue-formssit besidecheck-dependabotandcheck-readthedocsin.pre-commit-config.yaml’spython-jsonschema/check-jsonschemaentry, at itsrev: 0.38.0, underbtclib-org/.github’s own comment (issue btclib-org/.github#767): section 4 of the organization standard names the pair in its schemas bullet, both hooks carryingtypes: [yaml],config.ymlunder that spelling for the first and the directory’s yaml that is neitherconfig.ymlnorconfig.yamlfor the second, andcheck-hooks-applyfailing a hook that matches no file. This tree’s.github/ISSUE_TEMPLATE/holdsconfig.ymlfor the first andbug_report.yml,feature_request.ymlandquestion.ymlfor the second, and both hooks pass on those files, none of which changes here.
The token-grant entry’s deferrals to the other trees are superseded¶
This supersedes The token grants are read from the workflows rather than bounded in prose above, whose entry leaves
scorecard.yml’s workflow-level comment to be ported into the trees not yet carrying it (issue btclib-org/.github#946): no tree is owed it.btclib-secp256k1wrote that comment in btclib-org/btclib-secp256k1@e01d63f9,btclib-nodetook it in btclib-org/btclib-node@1cdaec6a andbtclibin btclib-org/btclib@30a05904.git grep -c 'each grant explained at its own line there' origin/main -- .github/workflows/scorecard.ymlanswers1in every repository of the organization carrying the file, withpermissions:in the same pathspec as the control the search reads.This supersedes that entry a second time, where it leaves the clause quantifying over every scope to be corrected on btclib-org/.github#893 rather than here (issue btclib-org/.github#946): that issue is closed, on 2026-09-07, and every repository carrying the file corrected its own copy — btclib-org/btclib-secp256k1@5716b6cc, btclib-org/btclib-node@d82ccc37,
5f5c280bhere and btclib-org/btclib@30a05904, which closed it. Their blocks read the same bytes out of eachorigin/main, which is the identityscorecard.yml’scontents: readcomment names the scope it is about above left to each tree’s port.Both sentences were true when
fee4f1f7wrote them: every landing they were waiting on is later than that commit.
The anchor depth, the free-threading reading and the allowlist converge¶
docs/source/conf.pysetsmyst_heading_anchors = 6, and the comment gives the standard’s reason in place of this tree’s own heading depth (issue btclib-org/.github#715): section 2 ofbtclib-org/.github’sREADME.md, under The documentation, makes six “every level markdown heads at, which makes the number a fixed point rather than a value re-derived from files that move”, and names a depth read off the tree’s own headings as the rejected alternative – which is the reason the comment it replaces gave for three. Nothing here links into a heading deeper than###, so no link is repaired: whole html trees built from this tip at 3 and at 6 differ in no byte, with a planted####heading as the control thatdiff -rcan see one. What the wider setting takes away is the failure a deeper heading would reintroduce – a link into a####heading ofREADME.mdfailssphinx-build -n -Wat 3 and builds at 6.tests/interpreters_test.pyreads the free-threading biconditional’s second side off the jobstest: every job passedwaits on, rather than offtest.ymlas a file (issue btclib-org/.github#634): section 3 of the standard makes that side “the jobs the required check waits on – the aggregate’s ownneedsclosure”, and names reading the workflow file as the rejected alternative. The aggregate is located by the namemain’s rule requires, whichREPOSITORY.md’s Required checks on main reads back from the endpoint, and the test refuses a closure that does not hold it rather than comparing against an empty one. The two readings answer alike here, every job oftest.ymlsitting insidetest-passed’sneeds:; what changes is what the answer rests on, measured against a copy of the workflow carrying a report-only job at3.14toutside that closure – the reading this replaces asks for aFree Threadingclassifier there and this one does not..github/workflows/codeql.yml’s listing allowlist namessuccessandskipped(issue btclib-org/.github#990): section 10 of the standard fixes both names “whatever the workflow’s own jobs can report today”, and names an allowlist following the jobs as the rejected alternative. This supersedescodeql.yml’s aggregate reads the run’s job listing above, whose second bullet has the allowlist namingsuccessalone and calls this tree’s the narrowing. What that bullet gave as its reason holds unchanged – no job of this workflow is conditional on the diff and none carries a condition narrower than the aggregate’s own – and the comment now writes it as the condition it is, which a job added later with a condition of its own ends.
REVIEWING.md and .gitattributes are the standard’s copies byte for byte¶
REVIEWING.md’s shared half isbtclib-org/.github’s at1058ef3, and everything from## This repository in particulardown is this tree’s, unchanged (issue btclib-org/.github#353): the half is replaced whole rather than difference by difference,tests/verbatim_test.py’s comparison being of bytes. What it says that the copy it replaces did not: a finding about the wording of prose no user reads is named at the foot of the review rather than filed, on btclib-org/.github#976’s authority; what makes a suggestion the right shape is the signature GitHub’s own commit carries; and Re-review reads the old sha off the previous round’s verdict, an amend or a rebase leaving it off the branch..gitattributescarries what the union driver charges at the seam, and the rejected alternative of not setting the driver at all (issue btclib-org/.github#1026): the paragraphs are the standard’s own words, and section 14 compares this file whole, so a wording of this tree’s own is one that comparison refuses.check-changelogis the hook the seam paragraph names, and.pre-commit-config.yamlhere runs it ahead of the markdownlint autofix for the reason that paragraph gives.
CONTRIBUTING.md reproduces the docs job whole¶
The
docs.ymlblock under Reproducing what CI runs carries the unresolved-linkgrepand the public-surface check beside thesphinx-buildcall (issue btclib-org/.github#973): that section’s opening sentence is that each command in it is the one a CI job runs, and a contributor reproducing the build alone leaves the checks over its output unrun — green locally, with the job red all the same. Each of them reads what the build wrote, so they stand in one ordered block, and the paragraphs under it say what each is for.The gate section The environment and the gates says that build is not the whole of the
docsjob, and points at that block rather than repeating it: what a contributor runs before committing is the job, and the commands stay in one place.
A bill of materials is published with the release, and attested with it¶
test.yml’sdistjob writes a CycloneDX 1.6 document intosbom/and uploads it as an artifact of its own (issue btclib-org/.github#144):.github/scripts/generate_sbom.pyis btclib’s atd68c1e8a, and it reads the built wheel’sMETADATArather thanpyproject.toml, so a rehearsal describes the.devversion it actually built.componentsis an empty list here,dependencies = []being what this distribution declares, and a signed statement of that is what the document is for: a consumer reading the organization’s releases gets the same artefact from each rather than learning which of them describes itself and why.release.yml’sattestjob carriessbom/*among its subjects, andgithub-releaseattaches the document beside the distribution files: one attestation covers the wheel, the sdist and the document, so the document verifies against the bundle the wheel does. Recovering a release by hand therefore takes the run’ssbomartifact as well asdistandattestation.RELEASING.mdreads the document instead of explaining its absence, and its pointer at btclib-org/.github#24 goes with the paragraph that carried it: that issue watches an exemption this repository no longer takes, and stays open forbtclib-secp256k1, which has not adopted.“Rebuild a release from its tag” exports
SOURCE_DATE_EPOCHrather than prefixing it ontouv build: a prefix binds the one command, and both scripts the fence runs after it read the variable out of their own environment and refuse to run without it. The tag is a${version:?}placeholder there now, which is what lets the fence verify a document only a release that carries one has.The submodule half of the script is ported unrun. This repository vendors no submodule, so
GITLINKand the functions around it are a no-op here; stripping them would put a silent divergence between the copies,generate_sbom.pybeing no part of section 14’s list. Its two bare issue references are qualified asbtclib-org/btclib#1280andbtclib-org/btclib#1194, which resolve to nothing in this tracker.
The gate’s needs: closure reads a block list, and the comments in it¶
tests/interpreters_test.pyreads a job’sneeds:in each of the three shapes GitHub takes, and takes a comment or a blank line among the items (closes btclib-org/.github#1031, issue btclib-org/.github#1038): the reader this replaces matched a flow list or a scalar on the key’s own line, so a block list under the key named no job at all, the closuretest: every job passedwaits on narrowed with nothing red to say so, and the free-threading biconditional that reads it measured a gate it had not read. No job oftest.ymlwrites itsneeds:under the key, so the closure holds the same jobs either way; what changes is what a job written that way would do to it.The spelling is the organization’s rather than this tree’s (issue btclib-org/.github#1038):
btclib,btclib-secp256k1andbtclib-nodecarry the same module, and that issue stays open for the three of them to take the same spelling. This tree wrote it because it is the one that reads the job blocks with their comments still in them: it meets a# a noteline at the indent the items take, where the others meet the whitespace their strip leaves there, so a run tolerating the whitespace alone passes there and fails here. An item’s own trailing comment goes with the two – an item written- changes # the gateis the shape the tolerance drafted in that issue’s body still dropped, and an item is read up to it.A step is not an item, and the test says so. A
- name: Setup uvstep line sits at the indent the items take, and what keeps it out is that an item is the whole line up to its own comment; a run widened to the rest of the line reads its first token,name:, as a job and every item below it with them. What ends the run ahead of a real job’s steps is thesteps:key at the indent a job’s keys take, so what separates the two readings is a step line written where an item goes.
The testpaths resolution takes a case that asks for no symlink¶
tests/conftest_test.pyasks the coverage gate about atestpathsentry whose..leaves the directory it names, and the case fails with the.resolve()removed fromwantedintests/conftest.py(closes btclib-org/.github#1022):tests/../srcissrc, which a command line namingtestsis not above, while unresolved it reads as a directorytestscontains – so a run collecting nothing ofsrcis handed the whole suite’s ratchet.test_a_symlinked_spelling_of_one_tree_is_still_the_whole_suiteis the other case that kills that call, and its assertions sit behind apytest.skip, so on a platform that refuses a symlink nothing else in the suite kills it.--no-covis what makes the pair reproducible. WithPath.symlink_tomade to refuse and the call removed,uv run pytest --no-covexits 0 without this case and 1 with it, that case being the failure. The same pair under the coverage floor cannot tell the two runs apart: the skipped case’s own assertions go unexecuted either way, so an identical shortfall fails both.An entry reaching the same directory with no
..in it leaves the removal undetected, which is what assigns the kill to the segment rather than to the method. A..that re-enters –tests/../tests, on the entry and on the command line alike – answers the same with the call and without it.The symlinked spelling is what reaches the resolution on
given: with the.resolve()removed there the symlink case fails and this one passes, so neither case stands in for the other.
The command line’s parent segment takes the case the family already has¶
tests/conftest_test.pygainstest_a_parent_directory_segment_names_the_whole_suite_too, which names the whole suite astests/../testsand as../testsfrom insidetests/, and fails with the.resolve()removed fromgivenintests/conftest.py(closes btclib-org/.github#1045): a..segment survives into the path object where.and a trailing separator are collapsed, so an unresolved command-line path compares unequal to thetestpathsentry it names and a run that takes the whole suite in is gated at nothing. Each of the two spellings fails on its own.The case, its name and its wording are
btclib-benchmarks’s, carried byte for byte out of itstests/conftest_test.pyatac260ed6and placed where it sits there, betweentest_a_path_is_read_against_where_pytest_was_startedand the symlinked spelling. Deriving a spelling here instead is how one question gets two answers in the family.--no-covis what makes the pair reproducible. WithPath.symlink_tomade to refuse and the call removed,uv run pytest --no-covexits 0 without this case and 1 with it, that case being the failure. Under the coverage floor both runs fail and the exit code tells them apart no better; the totals differ there, but only because a case that fails on its first assertion leaves a line of its own unexecuted, which reports the failure rather than the mutation.The same
..on thetestpathsside leaves the removal undetected, which is what assigns the kill to the command line, and so do./testsandtests/, which carry no..for the join to keep.What The
testpathsresolution takes a case that asks for no symlink says about the symlinked spelling reaching the resolution ongivenis superseded: this case reaches it too, with no symlink and no privilege. That entry stands as written, nothing landed being rewritten. Its own case still answers the call onwantedwhere this one does not – with the.resolve()removed fromwantedthat case fails and this one passes – and its sentence about a..that re-enters answering the same with the call and without it is about the call onwanted, which is the call the entry is written of.The same claim in
test_a_testpaths_entry_is_the_directory_its_parent_segment_reaches’s docstring sits in a file this diff opens, so it is replaced rather than superseded: that case now carriesbtclib-benchmarks’s wording whole, which namestest_a_parent_directory_segment_names_the_whole_suite_tooas what defends the call ongiven. A sentence derived here rather than taken from the family is what went false.
The symlink case’s pragma takes the case, not the handler alone¶
The
# pragma: no coversits on the case’sdef(issue btclib-org/.github#1042): an exclusion on a line that introduces a block takes the whole block, so it reaches the assertions after the skip as well. On theexceptit reaches the handler and thepytest.skipalone, which are the lines that do not run wherever the link is made, and a platform refusingos.symlinkthen meets the skip and a coverage floor it cannot reach in the same run – the exit code the floor’s and the failure naming a percentage rather than a symlink. Measured with a plugin makingPath.symlink_toraiseOSError: with the pragma on theexceptthe documenteduv run pytestexits 1 with the lines after the skip named missing, and with it on thedefthe same run meets the floor and the case reportsSKIPPED.The comment above the line says why coverage can ask nothing of the case, and what the exclusion costs: the body is reachable only where the platform makes a symbolic link, so a floor over a
sourcenamingtestsasks about the runner rather than about the suite, and dead code inside the case stops being flagged in exchange. A reason above the line with an inline half naming the case is what[tool.coverage.report]’s comment asks of a reason too long for the line it belongs to.The docstring’s sentence – a platform that refuses says so as a skip, which
-rareports – is what the move makes true of the run: the case skips either way, and with the pragma on theexceptthe run it skips in fails the floor.What The
testpathsresolution takes a case that asks for no symlink and The command line’s parent segment takes the case the family already has say about the coverage floor is superseded: each says the floor cannot tell its reproduction pair apart, and the skipped case’s unexecuted assertions are what make that true – withPath.symlink_torefusing and one.resolve()removed fromtests/conftest.py, the run with that removal’s own case deleted fails on the shortfall alone. The exclusion takes the shortfall away, so the documenteduv run pytestexits 1 with that case in the file and 0 with it deleted. The--no-covhalf of each pair answers as those entries say it does, and they stand as written, nothing landed being rewritten.btclibcarries the same shape, so the issue stays open on this landing and the citation above isissuerather thancloses.
The tree gains the deps-oldest sentinel, and the badge with it¶
.github/workflows/deps-oldest.ymlruns the suite weekly with every direct dependency at the oldest releasepyproject.toml’s own specifier allows (issue btclib-org/.github#323): section 10’s Which trees carry which sentinel names this repository for that row, andgit grep -l 'lowest-direct' -- .github/workflows/answered nothing here withuv lockindeps-latest.ymlas the control that the path was read – the tree resolved upward and nothing resolved down, leaving every>=it declares a claim no run had installed. The cron is12 3 * * 4, read off section 10’s two tables: the workflow table’sdeps-oldestrow gives Thursday and hour 03, the repository table’sbitcoin-core-rpcrow gives minute 12. It gates nothing – no aggregate job and no branch protection rule – andREADME.md’s badge row gains it where the calendar puts it.The cell is 3.10 and there is no matrix:
requires-pythonnames that floor, where.python-versionpins 3.14 as the newest interpreter the package supports, and a floor run asks the other end. One cell because a floor is one claim to verify rather than a range to scan, the range beingdeps-latest.yml’s and the platform sweeps’.UV_RESOLUTION: lowest-directis declared on the job and not passed to the step that resolves:uv lockrecords a non-default mode insideuv.lock, and a later uv command under the defaulthighestreads that lock as stale, so--lockedrefuses the lock the step above it just wrote. Measured with uv 0.12.7 on a probe project holding one specifier: against alowest-directlockuv run --lockedexits 2, and 0 with the variable set.Neither the lint job nor the dist job of
deps-latest.ymlis mirrored. That lint job is there because mypy is a local hook shelling out to uv, so a new mypy release adds a check this code base fails; a floor resolution moves mypy backward instead, which is the tooling’s question rather than the floor’s. The same reading refuses a dist job from the other end: thecheckgroup declares no floor, so an old twine, check-wheel-contents and pyroma would rate metadata this resolution does not move, and the floor a distribution does declare –[build-system]’suv_buildrange – is not inuv.lockfor the step to reach.The resolution fails on a group entry rather than at any declared floor, so the sentinel goes red before it reaches the suite:
uv lock --resolution lowest-direct --dry-run, uv 0.12.7, warns on each[dependency-groups]entry that declares no lower bound and then fails buildingcoverage==3.0.uv lock --helplists one option naming a group,--upgrade-group, and none that selects or excludes one, whereuv sync --helpnames several, so a lock is over every group there is and the resolution cannot be narrowed to the entries carrying a bound. What bound each entry owes is #446, filed by the review of this branch; each landed copy points at the same question in its own tree –btclib-org/btclib-node#761, since closed by that tree giving every entry a bound, andbtclib-org/btclib-benchmarks#320.
A run coverage’s configuration never reached is refused¶
tests/conftest.pyrefuses a run coverage read no configuration for (issue btclib-org/.github#443). coverage looks for its configuration in the directory the process started in, so a run started fromtests/finds nofail_under, nosourceand nobranch = trueand is held to no floor, while pytest walks up and readspyproject.tomlall the same. That asymmetry is what the guard reads, and it needs no copy of the floor’s number. Section 8 of the organization standard leaves a tree with a local floor to point such a run at its configuration or to make it say it is ungated, and this is the second limb:pytest.UsageError, which pytest prints without a traceback and exits4for – an exit of its own, so the code says the run measured nothing rather than that something in the tree failed. Measured fromtests/: exit 4, the hook running ahead of collection.The message names the directory the run started in, the configuration pytest read, the rootdir to run from, and
--cov-config.--cov-configis named without a caveat because in this tree it restores everything: fromtests/, a run given the tree’s ownpyproject.tomlthat way reports a per-file table byte-identical to the run from the root and meets the floor. Theomitis what would otherwise be left behind, and is not here – coverage makes a pattern absolute against the process’s own directory unless it opens with a wildcard, which this tree’s one pattern does.sourceis not what carries this tree either: coverage takes asourceentry naming no directory as a package, sotests, a directory at the root, is an import name fromtests/that reaches the same files, andbitcoin_core_rpcis an import name from either.The exemptions are an enumeration and not a rule:
--no-cov, an explicit--cov-fail-under,--helpand--collect-only, each measured fromtests/at exit 0, where--markersand--fixturesare refused knowingly, measured there at exit 4. What would decide the question is whether pytest-cov would have gated the run, which is no property to read at this hook, so a longer list is the same guess under more names. Every run this tree makes on purpose starts from the repository root – no step of any workflow gives a pytest command a working directory – and the mutation session’stest-commandpasses--no-covbesides.coverage_fail_underis not what does this, and is not repaired. pytest-cov fillsknown_args_namespace.cov_fail_underfrom the coverage configuration inpytest_load_initial_conftests, beforepytest_configureruns, so fromtests/the hook is handed a floor of zero and hands it back. The guard runs in front of it and raises before it writes.pyproject.toml’sruntime-cast-valuecomment no longer says the tree holds nocast():tests/conftest_test.pycasts its stand-ins to the types they stand in for, in the quoted form that rule asks for, and the comment says so where it said there was none to find.The justifications this guard falsified name the run that still demonstrates them: the
.coverageentry of[tool.uv.build-backend]’s exclude list andasks_for_everything’s docstring each cited a barepytest .fromtests/, and what leavestests/.coveragebehind, and what reachesasks_for_everythingfrom a directory the rootdir is not, is that run given--cov-config.btclib-nodeandbtclib-benchmarksare owed the same guard, btclib-org/.github#443 having taken that decision for the family;btclib-secp256k1landed it atbd71d7c8andbtclibatda3e0d15. What lands with the last of them is the sentence recording which limb of section 8 the family took, which is why this cites the issue rather than closing it.
asks_for_everything says why it resolves both sides¶
The docstring gains a paragraph on the two .resolve() calls (closes
btclib-org/.github#810): either side can arrive carrying a symlink –
pytest builds the rootdir with os.path.abspath, which leaves one in the
path alone, and a positional argument is whatever was typed – so one
directory reaches the comparison under two spellings unless both sides
are resolved. The reason goes in the docstring rather than in a comment
beside the calls, which stand either side of an early return: one comment
would sit at one of them, and the docstring is where this function’s
contract is stated.
The gate’s closure is asserted to walk, and both consequences are named¶
tests/interpreters_test.pyasserted nothing about_gatingreaching a job named only by another job (issue btclib-org/.github#1053):test.yml’stest: every job passednames each job it waits on directly, so a_gatingreading the aggregate’s ownneeds:and stopping answers the real gate, and the shape cases above read one job block each and ask for no second hop. The case added for it is written against a dict in which the aggregate nameschangesandchangesnamescoverage, so it asserts its own closure and nothing about a shape. Measured over the whole module by replacing that walk with a single union over the same_needed: every assertion above the case passes and it is the one that fails.The comment above
_NEEDSgave one of the two things a reader blind to the block shape does (closes btclib-org/.github#1057): it named the biconditional passing on a gate it has not read, which is what happens where the jobs the narrowing keeps still name an interpreter. Where the narrowing leaves the aggregate alone, the aggregate’s own job names none and theno job the merge gate waits on names an interpreterassertion ahead of that biconditional fires instead. Both halves are measured against this tree’s owntest.yml:test-passedwrites itsneeds:as a flow list on the key’s own line, so a reader narrowed to that line answers the gate’s closure unchanged, and with thatneeds:written under the key instead the narrowed reader collapses the closure totest-passed, no job of it names an interpreter and that assertion fires; with a job naming a free-threaded interpreter listed below a comment line among those items, the reader as written puts it in the tuple the biconditional reads, where a reader carrying no whole-line alternative answers3.14alone and passes on a gate it has not read. The comment’s third paragraph ends in a full stop, and_NEEDSand_ITEMare untouched.This bears on The gate’s
needs:closure reads a block list, and the comments in it above. That entry has the narrowed closure leave the free-threading biconditional measuring a gate it has not read, which is the half of the consequence that holds where the jobs the narrowing keeps still name an interpreter; where the narrowing leaves the aggregate alone the assertion ahead of that biconditional fires instead, and the comment above_NEEDSis where both halves are written.
What enforces the lint gate is a local run, and not a commit¶
.pre-commit-config.yaml’s header says CI enforces exactly what a local run of.pre-commit-config.yamlenforces (issue btclib-org/.github#966): the lint gate is not installed as a git hook here, so a commit enforces nothing, whichCONTRIBUTING.md’s The lint gate is not installed as a git hook paragraph states with therev-parse --git-path hooksmeasurement behind it. A local run is the nounbtclib-benchmarkssubstituted atc2c0b2a2, where one entry gating a commit and CI at once became one entry gating a local run and CI at once.Naming the command in
.pre-commit-config.yaml’s header is the rejected alternative.CONTRIBUTING.mdhere documentsuv run pre-commit run --all-fileswherebtclib-org/.github’s documentsuvx pre-commit run --all-files, so one spelling written into text the trees share is false in one of them, and a spelling per tree reopens the divergence the shared text closes..pre-commit-config.yaml’s skip recipe keeps itspre-commit runline alone.SKIP=mypy git commit -m "foo"and the--no-verifypair were instructions for skipping a hook this repository does not install, written in the file a session reads while it is committing. Keeping them for a reader who installed the hook anyway is the rejected alternative: that reader is acting against this tree’s own instruction, and pre-commit’s own documentation is where an interface this tree does not use belongs.CONTRIBUTING.md’s The editor paragraph says a local run in both of its halves: what installing the extensions changes nothing about, and where a finding is met instead of while typing. The decision on btclib-org/.github#966 quotes the first half; the second named the commit that trips over a finding, which is the same claim about a hook that is not installed..vscode/extensions.jsonnames no commit. What the editor reports is what a local run of.pre-commit-config.yamlwould report, what formats yaml and jsonc on save is the tool that formats them on a local run, and what writes back what a second formatter on save wrote is theruff-formathook.
The hook preambles say why the hook is there, in the present tense¶
.pre-commit-config.yaml’s yamllint preamble said the workflows were the one place prose could grow without a limit (issue btclib-org/.github#880): a sentence about the tree before the hook, where section 9 of the organization standard asks a comment for why the code is as it is in the present tense. The reason survives the tense, so the sentence names what holds each format instead: markdownlint holds markdown,max-doc-lengtha Python docstring and a whole-line comment,toml-comment-widtha toml comment, and the prose in a yaml file is this hook’s or nothing’s.prettier reads the same files two hooks above, so the sentence says what it does with a comment. Measured at the rev this tree pins: it explodes an inline sequence that runs past the width and hands a long comment back byte for byte, and the second is what leaves a yaml comment’s width to this hook. The exploded sequence is the control saying the file was rewritten at all.
the hook abovenamed markdownlint rather than taplo, the hook it sat under. Each tool is named.The
toml-comment-widthpreamble carried the same tense: a toml comment’s width was the last prose here held to nothing, and nothing read toml for it. It opens on that width being prose and names MD013, W505 and yamllint as what holds the other three, the sentence below it already giving, in the present tense, the reason no tool off the shelf does toml.This supersedes one sentence of Four comments say what
max-doc-lengthreaches above, which has that preamble there “to say that toml prose was the last kind held to nothing, and that is what it still says”. What the same entry says about W505’s two kinds and about the 80[tool.ruff.lint.pycodestyle]names is untouched.
A changelog entry is its title and at most three lines¶
REVIEWING.md,CONTRIBUTING.mdandcheck_changelog.pyfollowbtclib-org/.github(issue btclib-org/.github#1075): what is filed is one test, and an entry’s body past three lines is refused.
The lint gate’s prose names a local run, not a commit¶
lint.ymlsays what a local run enforces (issue btclib-org/.github#966).
links.yml calls the organization’s reusable workflow¶
The lychee job lives in
btclib-org/.github(issue btclib-org/.github#35): this file keeps the trigger, the schedule and the files it checks.
The needs: shape test reads hyphenated job keys¶
A block list of
- test-passedand- free-threadedis asserted beside its rows (issue btclib-org/.github#1063):\w+in place of\S+for the item token fails it.
The wiki and the projects board are off¶
REPOSITORY.mdrecords both asfalse(issue btclib-org/.github#550), the settings having been turned off.
sdist-rebuild.yml verifies the latest release’s sdist weekly¶
The tag is rebuilt as
RELEASING.mdrebuilds it (issue btclib-org/.github#523), andgh attestation verifyover the rebuilt sdist fails the run where no attestation of the release carries its digest.
release.yml’s called workflows take a concurrency-suffix¶
release.yml’s remaining called workflows gain theconcurrency-suffixtest.ymlandlint.ymlalready had (closes btclib-org/.github#1083): a calledgithub.refis the caller’s, so a rehearsal sharedmain’s group.
codeql.yml calls the organization’s reusable workflow¶
The matrix and its two CodeQL steps live in
btclib-org/.github(issue btclib-org/.github#35): this file keeps the trigger, the schedule, the concurrency group and the aggregate job.
docs.yml calls the organization’s reusable workflow¶
The checkout, the build and the link check move to
btclib-org/.github’sreusable-docs.yml(issue btclib-org/.github#35):main’s required check renames todocs / Build the documentation.
REPOSITORY.md reads the required checks back in the rule’s order¶
The required-checks table, the
PATCHexample and the sentence naming this row takedocs / Build the documentation(issue btclib-org/.github#35), the endpoint’s own order and name.
lint.yml calls the organization’s reusable workflow¶
The checkout and the pre-commit run move to
btclib-org/.github’sreusable-lint.yml(issue btclib-org/.github#35):main’s required check renames tolint / Lint and type-check.
REPOSITORY.md puts the lint rename’s check at the rule’s tail¶
The required-checks table and the
PATCHexample takelint / Lint and type-check’s tail position (issue btclib-org/.github#35), the endpoint’s own order.
conf.py’s version-source comment no longer claims an absent install¶
The comment’s false clause is replaced with the reason that survives either install form (issue btclib-org/.github#1098): a version should not depend on whether the environment installs the package.
[tool.mypy] exclude is anchored to the build/ directory it means¶
"build"was an unanchored regex, dropping any path merely containing that substring (issue btclib-org/.github#1102): the readthedocs-wait script and its test, un-excluded, type-check clean.
links.yml’s targets: reaches every tracked markdown file¶
targets:skipped.claude/and.github/, hidden to lychee’s glob walker without a glob naming them (issue btclib-org/.github#1104): it becomes"**/*.md" ".github/**/*.md" ".claude/**/*.md" "docs/**/*.rst".
CONTRIBUTING.md’s local docs-gate grep matches the widened pattern¶
The documented docs-gate grep still read
href="#\./, not the widenedhref="#\.\.\?/(issue btclib-org/.github#1105): a broken../destination passed the local check and failed the actual gate.
release.yml calls the organization’s reusable workflows¶
version-check,public-apianddocumentedbecome calls tobtclib-org/.github’s reusable workflows, at@main(issue btclib-org/.github#35): none is a required check onmain.
CONTRIBUTING.md’s docs-gate line passes check_api_page.py the directory¶
The documented last line passed the built page itself, where the script wants the directory it lives under (closes #467): run as written, it raised rather than checking anything.
deps-oldest.yml calls the organization’s reusable workflow¶
The floor sentinel’s own job becomes a call to
btclib-org/.github’sreusable-deps-oldest.yml, at@main(issue btclib-org/.github#35): none is a required check onmain.
interpreters_test.py reads a caller’s with: beside the block¶
_PYTHONS_CALLERreads the JSON-encoded interpreter list a caller ofreusable-os-suite.ymlwill carry, beside_PYTHONS’s own block sequence (issue btclib-org/.github#1119): no such caller exists yet.
The os-* sweeps call the organization’s reusable workflow¶
Each
os-*sweep’ssuite-*job now callsreusable-os-suite.yml(issue btclib-org/.github#35): none is a required check onmain, and this bears oninterpreters_test.pyreads a caller’swith:beside the block above.
CONTRIBUTING.md gains the command naming an open section’s order¶
CONTRIBUTING.mdgains theawkcommand naming the open section’s headings, andREVIEWING.mda question asking whether the branch’s own entry is last (issue btclib-org/.github#1097): no gate reads it.
test.yml’s changes job calls the organization’s reusable workflow¶
The event handling, the pagination and the decision move to
btclib-org/.github’sreusable-changes.yml(issue btclib-org/.github#35): this file keeps which files are its own prose.
check-changelog’s stanza loses files: and gains always_run: true¶
A rebase that eats a heading’s blank line stages nothing, so the hook’s old
files:key left it unreachable exactly then (issue btclib-org/.github#1138):pass_filenames: falseselected nothing.
check_changelog.py’s docstring names the comparison, not a heading¶
The citation to
CONTRIBUTING.md’s Committing and rebasing named a heading no tree holds (issue btclib-org/.github#1137): the docstring now states the rebase discipline in its own words instead.
mutation.yml calls btclib-org/.github’s reusable workflow¶
The mutation job becomes a call to
btclib-org/.github’sreusable-mutation.yml(issue btclib-org/.github#35): the counter’s exit code decides the job now.
deps-latest.yml calls the organization’s reusable workflow¶
lint-latestandsuite-latestbecome one call tobtclib-org/.github’sreusable-deps-latest.yml, at@main(issue btclib-org/.github#35):dist-lateststays, uncovered by the callee.
scorecard.yml calls the organization’s reusable workflow¶
The
analysisjob becomes a call tobtclib-org/.github’sreusable-scorecard.yml(issue btclib-org/.github#35): the caller grants every scope the callee needs, its pin moving tov4.38.0.
sdist-rebuild.yml calls the organization’s reusable workflow¶
The
rebuildjob becomes a call tobtclib-org/.github’sreusable-sdist-rebuild.yml, at@main(issue btclib-org/.github#35): this workflow stops building the wheel it never read back.
pypi-install.yml’s wait-for-index calls the organization’s reusable workflow¶
The
wait-for-indexjob becomes a call tobtclib-org/.github’sreusable-wait-for-index.yml, at@main(issue btclib-org/.github#35):packageisbitcoin-core-rpc, read off the script’s own argument.
claude-review.yml becomes a caller of reusable-claude-review.yml¶
claude-review.ymlbecomes one job, a call tobtclib-org/.github’sreusable-claude-review.yml,@main(issue btclib-org/.github#35): fixes the stale-verdict jq (btclib-org/btclib-secp256k1#394), pinsef8bb1e4.
claude-review.yml’s comments describe the file, not the conversion¶
The before/after account of the step-level closed guards, of the concurrency key and of
claude_argsgoes (closes btclib-org/.github#1178): section 9’s No history in the prose.
The mutation profile states no totals and keeps its judgement¶
bitcoin_core_rpc.toml’s counts go and its survivor prose is re-derived from the last run (issue btclib-org/.github#1158): the workflow prints those counts, andRPCErrorCode’s members now dominate.
The ready_for_review comments name the workflow the draft condition is in¶
docs.yml,links.ymlandlint.ymlname the workflow each calls as where the draft condition is (closes btclib-org/.github#1177): the caller’s own job carries none.
release.yml’s attest/github-release call the organization’s workflows¶
Both become calls to
btclib-org/.github’sreusable-attest.ymlandreusable-github-release.yml,@main(issue btclib-org/.github#35): the signer moves at the next release.
wait_for_pypi_release.py takes the body the four trees converge on¶
The docstring gains the
(btclib-org/btclib#1165)citation it never carried, the usage example passes"$PACKAGE", and theexceptbecomes two clauses (issue btclib-org/.github#1160).
scorecard.yml’s analysis job documents its permission grants¶
The three grants
zizmor --persona=auditorflags take a trailing comment, the leading prose kept (issue btclib-org/.github#1164): the auditor reads a permission’s own line, not the one above it.
codeql.yml’s two permission blocks document their grants too¶
The three flagged grants, in
analyzeandcodeql-passed, take a trailing comment, the leading prose kept (issue btclib-org/.github#1164): the auditor reads a permission’s own line.
no-downstream-name-in-package’s pattern drops its btclib-node exception¶
The hook’s negative lookahead no longer whitelists
-node|_node(closes #422):btclib-nodedepends on this package directly, the relationship the hook exists to guard against.
interpreters_test.py reads .python-version where dist names no interpreter¶
A gating job whose Setup uv step names no
python-version:runs on whatever.python-versionpins, and_PINnow stands in for it there (closes #435):dist’s own step is written that way.
RELEASING.md’s provenance fence takes a ${version:?} placeholder¶
The “Verify the provenance of an asset” fence no longer hard-codes
v2026.8.8(closes #440): it takesbtclib-node’s placeholder-and-${version:?}shape.
generate_sbom.py gains the test btclib’s own copy already carries¶
tests/generate_sbom_test.pyisbtclib’s test of the identical script, its fixtures renamed to this distribution’s own wheel and sdist (closes #439).
Every [dependency-groups] entry declares a lower bound¶
Every entry but
ruff, which already had one, gains a floor set to the version its own consumer runs at (closes #446):uv lock --resolution lowest-directused to fail oncoverage==3.0first.
wait_for_readthedocs_build.py takes the body the four trees converge on¶
The usage example passes
"$SLUG", andUSER_AGENTbecomes a templateunservedformats with the project, taking it as a third parameter (issue btclib-org/.github#1186).
mutation_counts.py gains the test btclib’s own copy already carries¶
tests/mutation_counts_test.pyisbtclib’s test of the identical script, unchanged (closes #496): a synthetic Cosmic Ray session names no distribution.
regebro/pyroma’s recurring pre-release proposal is accepted, not fixed¶
.pre-commit-config.yamlwrites down whypinned-revstays the backstop against it (closes #494): pre-commit.ci excludes no repo from autoupdate, so nothing stops a third proposal.
zizmor’s auditor persona becomes the gate, every finding answered¶
.pre-commit-config.yaml’s zizmor hook gains--persona=auditor, every flagged grant and the shared gate step now answered (issue btclib-org/.github#1164, issue btclib-org/.github#1198).
check_changelog.py gains the fifth check .github already carries¶
This tree’s own copy is replaced with
.github’s, byte for byte above the marker, and_GRANDFATHERED_ENTRIESbelow it is set to this tree’s own count of entries above the rule (issue btclib-org/.github#1215).
The check-changelog hook’s comment names no count¶
.pre-commit-config.yaml’s comment abovecheck-changelogno longer countscheck_changelog.py’s checks (issue btclib-org/.github#1225): its own docstring already enumerates them.
[tool.uv] required-version moves to the ceiling dependabot-core now bundles¶
pyproject.toml’s[tool.uv] required-versionmoves from>=0.12.7to>=0.12.17(issue btclib-org/.github#1229): the old floor admitted auvolder than the onedependabot-core’s updater writesuv.lockwith.
source-exclude excludes a shape, not a list of caches¶
[tool.uv.build-backend] source-excludeexcludes any dot-prefixed path underdocs/,src/bitcoin_core_rpc/ortests/, replacing named cache entries with the shape (issue btclib-org/.github#1070).
source-exclude gains the two script-loading tests it missed¶
generate_sbom_test.pyandmutation_counts_test.pyjoin the six already-anchored entries (closes #508): each loads a script out of.github/scriptsby path, the shape the comment already covers.
links.yml and sdist-rebuild.yml take the conditional in place of true¶
Both workflows’
concurrency:block takes the conditional the standard section 10 names, in place oftrue(issue btclib-org/.github#1226): each takesclosedwith nopushtrigger.
REPOSITORY.md marks its two observations, stable in its own subsection¶
The
stabletag and the plan name carrya fact about a changing worldand a read date (issue btclib-org/.github#1017):stable’s mark sits apart from the hook reading beside it, which stays gated.
The pyroma hook moves out of autoupdate’s reach, as a local hook¶
.pre-commit-config.yaml’spyromahook isrepo: local, bound now bycheck’spyroma>=5.0.1(issue btclib-org/.github#1199):autoupdatecan no longer propose it a prerelease revision.
mypy_exclude_test.py reads [tool.mypy] exclude on one line too¶
_EXCLUDE_LIST’s regex reads[tool.mypy] excludewhichever shape it is written in (closes #510):uv_build’s sdist writes a single-entry array on one line, which the previous pattern missed.
notice-rgx takes an optional shebang line ahead of the notice¶
pyproject.toml’sflake8-copyrightnotice-rgxnow reads^(#![^\n]*\n)?ahead ofCOPYRIGHTtranscribed (issue btclib-org/.github#1294): one shebang line may precede the notice.
v2026.9.3¶
Repository¶
REPOSITORY.md’s Security settings stated six values and read none of them back (issue btclib-org/.github#468). The block now carries the three commands that answer them –.security_and_analysisfor the five keys of that object,/vulnerability-alertsfor the one whose endpoint has no body and says so with its status, and/private-vulnerability-reporting– and every row is traceable to one of the three or to the Required checks on main readback the CodeQL row already points at.The same table named three of the five keys its own command answers with (issue btclib-org/.github#468).
secret_scanning_non_provider_patternsandsecret_scanning_validity_checksnow have rows of their own, both readingdisabled, with the reason they are off left where it is rather than repeated: a table shorter than its own command is the defect that issue names, not a tidier one.The topics are read back and compared with the keywords (issue btclib-org/.github#468), the comparison run with both sides confirmed non-empty before its empty output was believed – a sibling’s copy of the same shape matches nothing on the right and so reports an agreement it never measured.
REPOSITORY.md’s Tag protection readback queriedrulesandbypass_actorson the ruleset list endpoint (closes #317), which answers neither: the command now reads the ruleset’s id off the list first and reads those two fields from the single-ruleset endpoint, the same endpoint the Merge methods readback already uses.The Merge methods section’s
main-self-mergereadback asked the reader to fill in a ruleset id by hand (closes #320). It now derives the id from the ruleset’s name the same way Tag protection does, rather than leaving that one lookup unresolved beside it.REPOSITORY.mdopened by claiming to be the whole of what is set outside the tree (issue btclib-org/.github#551) – a promise the file did not keep, the default branch among the settings it does not record. Section 11 of the organization standard rejects that claim and bounds the file at the settings the standard asks about, which is what the opening paragraph now says. A What this file passes over section at the foot carries the other half: the repository document’s URLs and counts, the switches no section of the standard states a rule for, the facilities that answer empty, and the default branch as a gap rather than an exclusion.RELEASING.md’s release pull request opens the next cycle’s “work in progress” sections ofCHANGELOG.mdandRELEASE_NOTES.md, above the two it retitles (issue btclib-org/.github#528). Opened in a pull request of their own after the release’s, the sections were missing frommainfor as long as that pull request took to land, and a branch landing in between filed its entry under a release it is not in, with nothing to say so. The version bump stays in the pull request that opens the cycle,version-checkcomparing the tag against whatpyproject.tomldeclares. Section 12 of the organization standard is the decision; this is its port.
Tests¶
chains.py,client.py,errors.pyandtransport.pyeach declare their own__all__(closes #309). Section 7 of the organization standard asks it of every public module of a published package, not only of the facade__init__.pyalready named one for;btclib-org/.github ISS 79settled the question for this package before the split of #292 gave it four more modules the old answer did not reach. Each list is exactly the names the module defines, none of it imported from a sibling, and__init__.py’s own__all__is their union –census_test.pynow states both as tests rather than as two facts that merely happen to agree.A static, per-module check pins what
chains.pyanderrors.pyimport, closing #313’s fix on every interpreter (closes #315). The baselinetest_importing_chains_or_errors_adds_no_socket_layerreads is taken afterhashlib, which already holdssocketon PyPy, so a future directimport socketin either module would pass that test there without ever being added to anything. The AST-level check reads the source instead of a subprocess’ssys.modules: the two modules either namesocket,sslorurllibor they do not, on every interpreter and without opening one.A bare
uv run -p pypy3.11 pytestfails the coverage floor, at an unstable number (closes #314). Two tests that raise a genuineRecursionErroron purpose –client_test.py’stoo deepcase andtest_a_reply_nested_too_deeply_to_parse– trip coverage’s pure Python tracer under PyPy, either alone:CoverageWarning: Trace function changed, data is likely wrong.pytest-covstops coverage from apytest_runtestloopwrapper, before theatexit-registered callback that coverage’s own PyPy suppression for this warning waits for ever runs; that one check fires once, at the end of whatever run included either test, judging whatever fraction had already run by then – unstable with pytest-randomly’s seed for that reason.CONTRIBUTING.md’s Reproducing what CI runs now says whypypy3.11needs--no-covbeyond matching the platform sentinel, which already runs that way and measures nothing of this.
The verbatim files are the standard’s copies, byte for byte¶
.gitattributesstates the union price as section 9 of the standard does (issue btclib-org/.github#423): the driver is a checkout’s and the forge does not apply it, so a pull request whoseCHANGELOG.mdorRELEASE_NOTES.mdoverlaps its base is reportedCONFLICTINGhowever cleanly the pair merges locally, and a rebase on a checkout is what clears it..markdownlint.jsoncpoints at section 14 of the standard for who carries it (issue btclib-org/.github#316), in place of an enumeration of trees.CONTRIBUTING.md’s shared half is btclib-org/.github’s (issue btclib-org/.github#281): the half is replaced whole rather than each change applied by hand, a hand-written list of them being what comes up short. Among them, The landing queue points atREPOSITORY.md’s Plan-gated settings for the ceiling’s figure (issue btclib-org/.github#412).
REPOSITORY.md’s perimeter is section 11’s¶
The scope statement carries section 11’s three limbs in the standard’s own words (issue btclib-org/.github#582), the third – the settings a behaviour the standard describes rests on – being what admits
allow_auto_merge,has_issuesand.visibilitybelow.The opening no longer claims that nothing in the file is recoverable by reading the tree (issue btclib-org/.github#571): the topics and
.homepagehave a copy inpyproject.toml, which Topics and Publishing already said, so the clause names those two and holds of the rest..visibilityandhas_issuesare read back (issue btclib-org/.github#584, issue btclib-org/.github#550): section 10’sscorecardsentinel reads a public repository and nothing else, andCONTRIBUTING.md’s tracker rule rests on the issues switch.has_wikiandhas_projectsare outside the perimeter (issue btclib-org/.github#550), by section 11’s sentence rather than as a question this copy held open.allow_auto_mergeis read back with the other merge settings (issue btclib-org/.github#566), the landing Merge methods describes resting on it.The file no longer counts how many times it names auto-merge (closes btclib-org/.github#578): the sentence at the foot that carried the count is replaced by the read-back above, and section 9 refuses a count nothing checks in any case.
.default_branchis read back (issue btclib-org/.github#549), under Branch protection, and the foot stops naming it as a gap; the Pages read-back was already in place.The Read the Docs project is read back from its own API (issue btclib-org/.github#564):
latesta branch followingmain,stablethe taggit tagsorts highest, the active tags the automation rule’s result. A hook found on the repository is stale and deleted, which is now section 11’s rule; the sentence saying the deletion implied nothing for a sibling is gone.The Actions and Dependabot secret stores answer empty by the standard’s decision, not for want of use (issue btclib-org/.github#572): the token is the organization’s, at
visibility=all, so the two stores leave the facility bullet and get the organization store read back beside them.
The badge row carries the qualifier section 2 asks of it¶
Every workflow-status badge at the head of
README.mdcarries?branch=main(issue btclib-org/.github#579): each answers formainor answersno status, where the unqualified badge falls back to another branch’s run whenmainhas none. The pre-commit.ci, Read the Docs and Scorecard badges are outside the rule.
The free-threading classifier is a claim about the merge gate¶
tests/interpreters_test.pygates theFree Threadingclassifier on the merge gate running a free-threaded build (closes #332, issue btclib-org/.github#577). Section 3 of the organization standard declares the classifier where the gate exercises that build and names the enforcement: the biconditional that gates the PyPy classifier, with the gate’s own interpreters as its second side. That side istest.ymland not the platform sweeps, whose3.14tcell runs beside a landing and blocks nothing — a sweep passing is the ground the standard declines. Neither side holds in this tree, so the test lands green, and it goes red on whichever side moves alone.
The release’s publish jobs opt back in past a red public-api¶
publish-testpypi,publish-pypiandpublishedopen theirif:withalways()and name the results they do require (closes btclib-org/.github#484).public-apiexits non-zero on any break in the public surface since the last tag, which is what a cycle with breaking changes produces by design, and a bareneeds:refuses to start a job whose listed dependency failed: the publish jobs list it, so under a guard of the event and the repository alone a redpublic-apileaves the upload unstarted and, behind the upload, everything guarded on its success.publishedlistspublish-pypialone and still needs analways()of its own, a bareneeds:reading back through the listed job’sneeds:chain, so the widening onpublish-pypidoes not reach it. Section 12 of the organization standard is the rule, with the rejected alternatives beside it.RELEASING.mdreads the release run job by job forskipped, as a step of its own, and its griffe step names thepublic-apijob (closes btclib-org/.github#484). A skipped job carries no step and turns nothing red, so a run that lost its post-publish check to the wiring above reads as a run that finished; the step names the command, which jobs a tag and a rehearsal each skip by design, and how a skip is recovered,gh run rerunreaching none.
links.yml asks lychee for the fragment too¶
.github/workflows/links.ymlpasses--include-fragments(issue btclib-org/.github#583). A link into a heading is checked as an anchor and not only as a page, so a heading renamed in the tree a link here points into is red in this run rather than nowhere. Run over the tree with the workflow’s flags and the token, the flag adds no request and is red on the anchorsdocs/source/migrating.mdmakes into the_link.mdstubs: the heading each names belongs to the file the stub includes, so the anchor resolves in the built site and not on the forge, whose page for the stub carries no heading — that is #330’s, and the run stays red on those four until it is settled. With the token thegithub.com/btclib-org/.github#<heading>anchorsCONTRIBUTING.mdcarries are not caught by it; btclib-org/.github#630 is where that is weighed.
The index wait is a script with a test¶
.github/scripts/wait_for_pypi_release.pyis whatpypi-install.ymlwaits with, in await-for-indexjob the install matrixneeds:, andtests/wait_for_pypi_release_test.pyis what drives it (closes btclib-org/.github#509). A schedule and a dispatch pass no version and take the empty-tag branch, so the retry, the per-request timeout and the::error::of the loop this replaces were reachable on a release call and nowhere else. The test substitutes the transport and the clock and advances the clock past the deadline, which is what section 10 of the organization standard asks of a step waiting on something outside the run, btclib-org/.github#466 having decided that shape for this class.btclibandbtclib-secp256k1carry the same script and the same test, and this is the last of the three trees the issue named.The budget is a deadline and not a count of attempts times an interval, so what the job’s
timeout-minutesis compared against is the one number the script states, and every request is bounded by what is left of that deadline as well as by its own timeout. No repair is claimed in it: the attempts and intervals of the loop this replaces multiplied out to well inside the timeout of the job that ran them. The deadline itself is the five minutes that loop named as its budget, chosen rather than measured – nothing here times how long the index takes to serve an upload.One job ahead of the matrix rather than a step inside the one that installs. The question is about the index and not about a platform, so it is asked once; the install job still checks nothing out, which is what makes the import its own step verifies resolve to what pip installed. The wait job’s checkout is sparse and reaches
.github/scripts, and the script takes the standard library alone underuv run --no-project, so nothing of this project is built or installed to run it.[tool.ruff.lint.per-file-ignores]exempts the script fromT201, as it does the two other scripts under.github/scriptsthat print: what it prints is what the log of the step running it shows, which is the case that rule is not about.
The documentation wait is a script with a test too¶
.github/scripts/wait_for_readthedocs_build.pyis whatrelease.yml’sdocumentedjob waits with, andtests/wait_for_readthedocs_build_test.pyis what drives it (closes btclib-org/.github#573). A tag push is the only trigger that reaches the wait, so a loop kept in the workflow file is first executed on the day a release depends on it, the same reasoning that moved the index wait beside it. The job checks out.github/scriptsalone and runs the script under--no-project, the shapepypi-install.yml’s wait job already has.btclibandbtclib-secp256k1carry the same script and the same test, and this is the last of the three trees the issue named.The request carries a
User-Agent. The Cloudflare zone in front of read the docs bans the interpreter’s own default (Python-urllib/3.14) outright, so a request sent without one is a 403 against a tag that is in fact served, and the wait would burn its whole deadline and annotate an error on every release.curl, which the loop this replaces called, is not banned, so the loop never needed one.The budget is a deadline rather than a count of attempts, compared against the job’s own
timeout-minutesas one number against one number. The loop this replaces already fitted inside that timeout, so the deadline is chosen against the job’s figure rather than to repair an overrun.This takes no empty tag as nothing to wait for, which is where it parts from the release wait beside it: that one runs on every schedule and dispatch and takes an empty version as nothing to do, while
documentedis guarded by the event and a branch ref would spend the whole deadline on a URL that answers 404 by construction.No trigger reaches this from a GitHub runner.
documentedhas no rehearsal, so what the script does against read the docs was measured from outside a runner, and Cloudflare can weigh address reputation alongside the user agent on a request the runner actually sends.
The sweep comparison passed without reading every sweep¶
A platform sweep whose
python:block stopped matching_PYTHONSleft_declared()’s dict instead of disagreeing with it, so the comparison narrowed to the survivors and the dropped sweep went unchecked (closes #333)._SWEEPSnow names the three platform sweep workflows –os-macos.yml,os-ubuntu.yml,os-windows.yml– and the test asserts_declared()read all three before comparing what they name: agreement among however many sweeps were found is a claim that gets weaker the fewer there are, and vacuous at one. btclib closed the identical defect in its own copy of this module first (btclib-org/btclib#1251).
The undocumented-public-init comment gave the class as the place¶
pyproject.toml’signorecomment said__init__is documented by its class (issue btclib-org/.github#588), where section 5 of the organization standard puts the constructor’s documentation in__init__’s own docstring, after PEP 257. The comment now gives that section’s reason for the entry: the rule checks that a docstring exists and never that it says anything, so declining it declines the presence check and never the documentation – an argument’s meaning, a raised exception, an invariant the constructor establishes still has nowhere else to go. It names the section, as the comments beside it carrying an organization-wide decision already do. It also says the convention sits below theignoretable, where it said above.
The Topics reason names the [project] table¶
REPOSITORY.md’s Topics gave thispyproject.toml’s[build-system]as whytopics_test.pyholds this repository to the topics/keywordscomparison (issue btclib-org/.github#665). The clause names the[project]table instead: that is whatkeywordedselects on, and section 3 of the organization standard states it under its own heading – the rule turns on that table and not on the index. Both keys are declared here, so it was the reason and not the conclusion that was wrong, and a reader carrying it to a tree with a[project]table and no build backend exempts a repository the suite asks the same question of.
A convention name wrapped in its middle is read as it is written¶
tests/conventions_test.pycollapses the whitespace of the Not tested here list and then of each name in it (issue btclib-org/.github#651).strip()takes whitespace off a name’s ends and leaves what an eighty-column wrap puts in its middle, so a name the break splits matched none of section 7’s and was reported as a convention the declaration invented – a red test on a declaration that reads correctly in the file.tests/README.md’s line here breaks at a semicolon, so nothing was red; what goes is the dependence on where the column falls, which nothing enforces.btclib-benchmarksreads its own line this way, its wrap falling inside a name.
The .gitattributes comment names the driver’s sides and one anchor¶
The comment above the two
merge=unionlines isbtclib-org/.github’s wording (issue btclib-org/.github#646):unionkeepsoursfirst and thentheirs, each of merging and rebasing is named for which side it callsours, and the comment reads the driver as resolving two branches writing an entry at one anchor rather than a bullet appended to one of a few changelog groups.
migrating.md’s links into README.md and COMPARISON.md name the root file directly¶
The anchors
docs/source/migrating.mdmakes intoreadme_link.md’s andcomparison_link.md’s own headings are now written../../README.md#...and../../COMPARISON.md#comparison, the shape a reader on the forge resolves directly (closes #330): the stub those anchors used to name carries no heading of its own until Sphinx’s{include}has run, so a forge reader following the old link landed on a page with nothing to anchor to – measured againstREADME.md’s own rendered page, which does carry it.docs/source/conf.py’sRootFileLinkstransform now also resolves a link written this way outside a shim, against the document that wrote it rather than against the repository root, so the same text still reaches the shim page that carries the heading once this build’s own{include}has run;sphinx-build -n -W --keep-goingandlychee --include-fragmentsboth read it clean, with no.lycheeignoreentry for either stub.local-link-prefixaccepts a../destination alongside a./one. The hook refused../on the premise that a broken link written that way reaches myst’s silent, ungreppable fallback (href="#../page.md"), which the transform above now falsifies: a broken../link is left unresolved for myst’s own resolver instead, whichsphinx-build -n -Wreports directly, with neither lychee nor the grep in docs.yml in the loop.
CLAUDE.md’s worktree block puts the removal line in a fence of its own¶
git worktree remove --force "$WT"no longer sits under a line that ends in a placeholder (issue btclib-org/.github#676): the assignment above it fails on its own redirection, and a shell that discards that failure as a parse error reads the next line as a fresh command, so a paste made before the placeholders were filled reached the removal against whatever$WTa previous run of the same block had left set. The removal now stands in a fence of its own, with prose between the two fences saying why the first closes early, so a reader reaches it only by pasting it on purpose.
REPOSITORY.md reads both variable stores back for the review switch¶
The file records
vars.CLAUDE_REVIEW_ENABLED, the switchclaude-review.ymlguards its jobs with (issue btclib-org/.github#682). Both variable stores are read back: the repository’s because a variable set here would take precedence over one of the same name set on the organization, and the organization’s for the empty name list section 11 reads as the switch’s off state – with atotal_countbeside it, a store that prints nothing at all when it answers needing one to show the call reached it. The repository’sactions/variablesleaves the loop of facilities whose empty answer records no decision, that zero now being half of what records this one.
REPOSITORY.md names is_template among the switches it passes over¶
is_templateis named among the fields of the repository documentREPOSITORY.mdpasses over (closes btclib-org/.github#691): it is in that document, no section above reads it back, and no section of the standard states a rule for it, which is the class that block enumerates; a field of that class the block does not name is silent in both places at once, and reads as one nobody looked at rather than as one weighed and left out.
CLAUDE.md’s git worktree add ends in the branch placeholder¶
-b <branch>sits after the path and the commit-ish (issue btclib-org/.github#687), so the>closing the placeholder has no target and the line is a parse error before it is a command. With the flag first that>takes"$WT"instead, and the removal that finishes a session leaves no directory at that path to fail on, so a paste made withWTstill set writes a file where the nextgit worktree addwants none. Section 9 ofbtclib-org/.githubis the rule, and the reason now sits under the fence.
The documentation build passes -W without --keep-going¶
sphinx-buildruns with-n -Wand no--keep-going(issue btclib-org/.github#347) indocs.yml, in.readthedocs.yamland inCONTRIBUTING.md’s gate commands, section 2 of the organization standard being where the flag is ruled out. The comments beside the command credited it with reporting every problem in one run, which is what-Wdoes on its own.
RpcError.code gets a name through RPCErrorCode¶
errors.pygainsRPCErrorCode, anIntEnumtranscribed from Core’sRPCErrorCodeinsrc/rpc/protocol.h(closes #352) – every member of Core’s enum, values included, its reserved and deprecated codes carried rather than trimmed.RpcError.codestays a plainint: a code the enumeration does not name is still an answer the node gave, not a defect of this package, and typing the field to the enumeration would turn a code Core has not yet defined into aValueErrorhere instead of a value the caller reads.RPCErrorCode(exc.code)is the caller’s own line for the name.
RELEASING.md guards a pasteable block with a fence and ${name:?}¶
A block whose unfilled paste would otherwise run something stands in two fences: one holding the placeholders, with nothing under them to reach, and one that writes each value
${name:?}and chains its lines with&&(issue btclib-org/.github#745). The tagging block spelled its tag as a literal example version, so the lines under its placeholder parsed cleanly and reachedgit push origin v2026.8.6.Both guards are needed. A
${name:?}failure is a run-time failure, which the chain propagates; a parse error takes its own trailing&&with it, so a chain under a discarded placeholder line guards nothing. That is why the placeholders sit in a fence of their own rather than above a chain.A placeholder inside quotes, or inside a path, is no parse error at all, so the fence around it is the whole of its guard. Section 9 of the organization standard is the rule the split answers to.
The __main__ guard test’s docstring gave a reason coverage does not act on¶
Two docstrings said a subprocess would leave
if __name__ == "__main__":uncovered, and this project measures neither of the scripts they describe (closes #355).[tool.coverage.run]’ssourceisbitcoin_core_rpcandtests; a script under.github/scriptsis in neither, so nothing under its guard changesfail_underwhichever way the test runs it.tests/wait_for_readthedocs_build_test.pyandtests/wait_for_pypi_release_test.pynow say what the test does reach:sys.argvand the guard itself, the one path through the file no other test in it takes, the rest callingscript.main([...])or the script’s own serving check directly instead.
client.py gains BitcoinCoreRestClient, Core’s -rest interface¶
A second client, beside
BitcoinCoreRpcClientand sharing nothing with it but the transport and the chain vocabulary (closes #351), decided by btclib-org/btclib#1193:-restis off by default, on the same port as JSON-RPC, and authenticates nobody who reaches it, so a method on the rpc client would have made “authenticated” a property of the call rather than of the object.get_binfor a.binpath,get_jsonfor a.jsonone..hexis the same octets as.bin, one more decode away, and is not offered.get_jsonreads its reply through the same_parsed_json_bodycall_rawreads its own envelope with – aDecimalfor every number,NaNand the two infinities refused – and a non-200 status isHttpError,-restcarrying no error object of its own for either method to read.A path is a path, and this client passes any. There is no
get_tx, noget_block, noget_utxos:pathis built by the caller from Core’s own documentation of-restand appended after/restunread./getutxosis the reason a per-resource method is refused rather than merely undone – it reads the UTXO set, so an output already spent and one never created answer the same way, and a wrapper turning that answer intoNonewould read as telling more than/getutxosdoes..github/scripts/rpc_smoke.pystarts the regtest node with-rest=1and checks-restagainst the same chain it already generates for the rpc client:.bincompared byte for byte with the rpc client’s own serialization,.jsonread for the field that names what was asked for, and/getutxosasked about an outpoint that chain has already spent and one it has not –tests/having no live node to ask either question of.
v2026.8.29¶
Added¶
call_batchsends several rpc methods in one HTTP request (closes #294). Each member is built the waycallbuilds its own request – the 2.0 marker, an id of its own, the same params validation, refused by position where it fails – and the answer is a list aligned with the input: position i holds member i’sresult, or itsRpcErroras a value, matched to its request by id rather than by where the reply array happened to put it. Only a failure of the whole exchange raises.COMPARISON.md’s Batching has what changed the earlier refusal: a batch member’s reply is read by the same version discrimination a lone call’s is, so there is no second parsing branch to establish against a live node first, where the round trip a batch amortises stays worth it beside a node reached over a WAN or a TLS link and not beside the loopback one this client targets.call_rawhands back the envelope instead of extracting aresultfrom it (closes #295). The same authenticated requestcallbuilds, with the protocol marker itself an argument – a string sent verbatim,Nonesending none at all,"2.0"the default – and the answer is the HTTP status and whatever_parsed_json_body’s safe json reading parses the body into, with neither the id, the error member nor the shape read any further: an array or a bare scalar comes back exactly as parsed rather than being refused the waycall’s own reply is. For a caller testing a server’s own handling of the protocol rather than asking a node something, which is deliberately not this client’s job for a request it would refuse to build in the first place.SessionTransport, a keep-aliveHttpTransportbuilt onhttp.client(closes #293).urlopen_transportopens a socket per call and sendsConnection: close, urllib’s own handler setting the header unconditionally; this one keeps one connection per(scheme, host, port)and reuses it across calls, passed astransport=like any other. The timeout is a deadline over the whole exchange, read the same way_read_boundedreads it for the default transport, andmax_body_sizebounds a reply the same way too. No redirect is followed,http.clientnot following one on its own. A connection the node closed while idle is caught before it is reused:select.selectprobes the kept socket for the peer’s close or reset, so a stale one is replaced and nothing is ever written into it. A drop landing in the narrower window between that probe and the write is still caught after the fact, reconnecting and sending once more. A response that had already started arriving and then broke is not: something came back, so the request already reached a node that read it, and it is never re-sent. One instance is safe to share between threads by serializing the whole exchange under one lock, documented as the trade it is.README.mdandCOMPARISON.md’s One connection per call say what is now true.docs/source/migrating.md, a migration guide fromAuthServiceProxy(closes #296). Every pattern a caller of Bitcoin Core’s own copy or of python-bitcoinrpc’s uses now has a counterpart in this client –call_batch,call_rawandSessionTransportare what made the last three complete – so the guide covers construction, invoking a method, wallet endpoints, batching, errors, amounts, timeouts, connections, the envelope and testing, each as theAuthServiceProxyline and the line that replaces it, labelled by file where the two no longer agree with each other.COMPARISON.mdkeeps the case for switching and is linked rather than repeated.README.md’s own section now carries that case and one example, pointing to the new page for the rest.
Changed¶
The documentation is built with
furoand with-nbeside-W(issue #256). Section 3 of the organization standard picks the theme for the shape a reference generated from docstrings has, and section 5 asks for nitpick mode:-Wnever sees a cross-reference that resolves to nothing, so a renamed class in a:class:role is a green build and a dead link.sphinx.ext.intersphinxlands with it and comes first, without which every annotation namingpathlib.Pathorcollections.abc.Callablewould be reported as this tree’s own broken link.nitpick_ignoreis empty, which is what the build answering clean means.Chain’s docstring no longer opens with a phrase napoleon reads as a type (issue #256).Core's five chain names: what -chain= takeshas the shape napoleon parses astype: description, so the built page carried aType: Core's five chain namesfield naming a class nothing declares — invisible while-Walone ran, and the first thing-nreports. The docstring says the same in a sentence that is not that shape.The sdist ships no test module it cannot run (closes #220).
tests/rpc_smoke_test.pyandtests/normalize_sdist_test.pyload a script out of.github/scriptsby path andtests/interpreters_test.pyreads the workflow matrices, and.githubis not in the archive — it is oncheck-sdist’s own default ignore list. Unpacked and run, those three were aFileNotFoundErrorand a failed assertion rather than a test, so a packager building from the sdist read a red suite that said nothing about this package.[tool.uv.build-backend]source-excludenow names them.A run narrowed by
--deselect,--ignore,--ignore-globor--lfis reported and not gated (closes #268).tests/conftest.py’scoverage_fail_underread-k,-mand the paths alone, souv run pytest --ignore=tests/transport_test.pywas handed the whole suite’s floor with part of the suite left out and ended inRequired test coverage of 100.0% not reached— the shortfall being the tests that did not run, printed the way one of the tree’s own would be. Section 8 of the organization standard names the set, and it is the set the hook now reads.--lfcounts wherever it appears rather than only where the cache holds a failure to rerun: what decides is what the invocation asked for, and the price is the--lfthat finds nothing to rerun and so is the whole suite ungated. The hook takesconfig.optionrather than one parameter per flag, so what a run asked for is read in one place.The convention declaration names section 7’s socket bullet (issue btclib-org/.github#458).
tests/conventions_test.py’s_CONVENTIONSis the vocabularytests/README.md’s declaration is written in, and the suite opens no socket was missing from it — a name outside the tuple fails the check, so that convention could be put in neither half of the declaration, neither the table nor Not tested here. It is declared not tested, with the reason:rpc_smoke_test.pybinds a loopback port of its own and listens on it to test the port probe.github/scripts/rpc_smoke.pyuses, so a walk oversocket()call sites — which is how section 7 asks for the convention to be driven — finds those constructions, and a test of it here would carry that file written into it as the exemption the bullet refuses. The sentences that asserted the opposite in passing, inCONTRIBUTING.md’s environment section and inpyproject.toml’s-n autocomment, say instead what holds: no test here reaches the network.The client is a package of four modules, not one file (closes #292).
errors.pyis the exception hierarchy,chains.pythe chain and network vocabulary,transport.pythe urllib layer,client.pythe RPC client built on the three beneath it; a module imports only the ones before it in that order.__init__.pyis a facade that re-exports today’s__all__unchanged,TYPE_CHECKINGimportingclient.pyandtransport.py’s names for a type checker while__getattr__resolves them at runtime only when asked for, so importingbitcoin_core_rpc.chainsorbitcoin_core_rpc.errorsalone still leavesurllib.request,sslandsocketout ofsys.modules.tests/census_test.pyreplacestests/standalone_test.py’s single-file walk with the same properties read across the four, plus the import order and the socket-layer isolation as tests of their own. The constraint this keeps is “nothing but the standard library behind it”; vendoring a single file is no longer what the shape is for. Every exception’s qualified name gained.errors,bitcoin_core_rpc.errors.RpcErrorwhere a traceback readbitcoin_core_rpc.RpcError–except,isinstanceand unpickling all match on the class and do not see it.
Tests¶
assert_chain’ssignet_challengeand_read_bounded’struncatestay keyword-only (closes #298). A mutation run turned the bare*,before each into/,and the suite still passed: every call already giveschainanddeadlinepositionally and the keyword-only parameter by keyword, which the mutant accepts the same way.test_signet_challenge_is_keyword_onlyandtest_truncate_is_keyword_onlycall each positionally instead, joining the twotests/README.md’s near-miss note already named.The import-weight census measures against the interpreter’s own baseline, taken in the same subprocess (closes #313). The former test read a fresh interpreter’s whole
sys.modulesafter importingchains.pyorerrors.pyand failed all seven pypy3.11 jobs of the release rehearsal:hashlib,chains.py’s own dependency forsha256, loadssocketon PyPy as an implementation detail of its cffi OpenSSL hash backend, unconnected to any networking. The baseline is now taken afterhashliband before either import, so the assertion is about a networking module the package’s own code adds, which is what the test polices.
Repository¶
claude-review.yml’s header points at the ceiling’s home (closes #269). Section 10 of the organization standard gives the ceiling on concurrent jobs one home per tree,REPOSITORY.md’s Plan-gated settings, beside the command that re-derives it and GitHub’s own table. This header namedCONTRIBUTING.md, which states the ceiling unnumbered and points on to that section: a pointer to a pointer, where the other workflow comments that need the same reasoning name the section themselves.git grep -n 'Plan-gated' -- '*.yml'
One word:
'concurrent jobs'and'Plan-gated settings'each answer five, and a different five. It answers six —claude-review,codeql,mutation,os-macos,os-ubuntu,os-windows— which is every workflow comment here that states the ceiling. The command is what finds them the next time the figure’s home moves; this entry is what one such move left behind.The badge row is a function of the tree and its order is section 2’s (closes #262, issue #256). The licence badge was
img.shields.io/badge/license-MIT-blue, which rendersMITbecause the URL says so rather than becauseLICENSEdoes; the derived form replaces it. Read the Docs is named atapp.readthedocs.org, the other spelling answering307to it. What the property rule adds iswheel,implementation,github/v/releaseand a badge per sentinel, in the order section 10’s calendar gives them. The rows are grouped by a blank line asbtclibandbtclib-secp256k1group theirs, which is what makes the block’s own comment true of what renders: two hard line breaks split it into a row the comment did not describe.trailing-whitespacekeeps no markdown derogation now, nothing here spelling a line break that way.ruff selects every family it ships (issue #256).
select = ["ALL"]rather than a hand-picked list, which is a thing that rots: nothing forces a second edit the day ruff ships a family nobody has looked at. What each newly reached family costs is inignore, one entry per rule and the reason beside it — the formatter’s own conflict list cited from ruff’sdocs/formatter.md, the declines this tree argues for itself, and the families whose construct this tree does not contain.PYI034is answered where it fires instead, with a# noqathatRUF100retires the day the interpreter floor reachesSelf.codeql.ymlruns on a pull request (closes #233, issue #256). The OpenSSF Scorecard’sSASTcheck walks a merged pull request’s own commits and scored the analysis as run on none of them, the workflow producing no run there. The same trigger is what makescodeql: every job passeda name a branch rule could require, which is what an aggregate job is for: without it the job produced a context no rule could ask for. Whether the rule asks isREPOSITORY.md’s, which reads the endpoint.scorecard.yml, andgriffe checkin the release path (issue #256).gh api repos/btclib-org/bitcoin-core-rpc --jq '.fork, .private'answersfalseandfalse, which is the property section 10 keys the sentinel on; its calendar row already exists and its minute is this repository’s. Thepublic-apijob compares the tag being cut against the tag before it and the publish jobs wait on it: the public-surface census asserts that__all__is declared and that what it names exists, which never asks whether this release took something the last one gave.The public surface is a census (closes #235). Section 7 stops its escape clause short of a repository publishing an importable package, and this one publishes.
standalone_test.pywalks the module’s own body and fails when a name it defines and does not underscore is missing from__all__, or when__all__names something nothing defines — the second being a failure a caller writingfrom bitcoin_core_rpc import *would otherwise be the first to see.tests/README.mdmoves the bullet from Not tested here to the table.The ceiling on concurrent jobs has one home (issue btclib-org/.github#412). The figure was stated in four workflow headers, in
CONTRIBUTING.mdand twice inREPOSITORY.md, none of them beside a command. It now lives inREPOSITORY.md’s Plan-gated settings, withgh api orgs/btclib-org --jq .plan.nameand GitHub’s own table beside it; every other statement gives the reasoning with the ceiling unnumbered and points there. A date beside the number was the rejected alternative: it says when the figure was true and never that it still is.pypi-install.yml’s index-wait step runs on every trigger (issue btclib-org/.github#49). Guarded byif: inputs.version != ''a tag was the script’s only execution path, and a step a release is the first to run is a step whose defect ships with that release — which this script has already done, in more than one repository of the organization. The empty case is a condition around the loop rather than an earlyexit 0, and that is the whole point of the change: bash parses a script as it runs, so an exit at the top leaves what follows unparsed and the weekly runs would go on not reading it. Measured — a syntax error below an earlyexit 0prints and exits0, the same error inside a branch the run does not take exits2.The queue measurements that nothing re-derives are gone (closes #213, #216, #217).
os-windows.yml’s header weighed “the fourteen Windows cells” against “ubuntu’s twenty-eight”, where the second contained the first, and named a job pair42e600ehad already superseded;os-macos.ymloffered a command readingtest.yml’s latest run, which carries no macOS cell, for a figure taken from a matrix the gate no longer has. Both headers now say why those rows are off the gate, which is the ceiling above, and state no count of their own.REPOSITORY.md’s code-quality figure keeps its number and gains the command that answers it — the check runs of a commit of that era, since the run-list name it pointed at is gone, andper_pageis load-bearing there.CHANGELOG.mdsays what section 9 says it holds (closes #218). The head claimed “every change of a release, in full”, which a workflow-comment change refutes by landing without an entry. The rule is the standard’s andCONTRIBUTING.md’s: an entry for anything a reader would notice.CONTRIBUTING.mdandREVIEWING.mdtake the standard’s shared half byte for byte (closes #246). Everything above This repository in particular is now identical tobtclib-org/.github’s copy, which is what section 14 compares. What arrives with it: The landing queue, the two citation spellings named rather than pointed at, theshapin on the merge call, andREVIEWING.md’s verdict section, whose ack of record is a review rather than a comment — whichclaude-review.ymlalready does,e57f9a6having converged it ongh pr review --commentbefore this branch existed, so the shared half arrives describing a workflow this tree already runs rather than one it owes.tests/conftest.pyno longer claims to keep btclib’s docstring in full (closes #265). It does not, and cannot: each docstring names an example path out of the suite it sits in. What is true is that this tree took the reasoning rather than deriving a second one, which is what the sentence says now.tests/__init__.py’s claim is about the clients (closes #248). No test builds a client that could reach a node, which is the propertytransport=on every construction buys. Sockets are opened here:rpc_smoke_test.pybinds and listens on a loopback port to test the port probe, so section 7’s stricter “the suite opens no socket” is not what this suite answers, and the docstring says so rather than leaving a walk oversocket()call sites to find it.hypothesis stays out, and
_read_boundedis why that was measured (closes #253). Its arithmetic is over indices, which is the space a generator searches and a case list does not — but a chunk boundary cannot split a character here, the body being accumulated in one bytearray and decoded once after the loop, and both places a body is cut —_read_bounded’struncate=True, and theMAX_ERROR_BODY_SIZEslice on a caller’s own transport, which_read_boundednever sees — are reached only under a non-200, where every parse failure is answered with the status. What is left is two indices,max_body_sizeand one past it, whichtests/transport_test.pyalready asserts in each direction. The exemption’s comment carries that measurement.The coverage floor holds for
uv run pytest tests, the spelling that names the whole suite (closes btclib-org/.github#430).testpathsistests, so that path collects exactly what a bare run collects, andtests/conftest.pyread it as a selection from the suite and handed the run a threshold of zero: the same command passed with the path named and failed without it, over the same tests and the same total.asks_for_everythingdecides it by containment againsttestpathsnow — a path at or above an entry takes the suite in,./testsand the absolute path being that same directory — and a file under it is still the subset the floor comes off for. The hook is handed two directories because pytest reads a positional argument against the one it was invoked from andtestpathsagainst the rootdir.--help, whereconfig.option.file_or_dirisNonerather than[], names no path either and is folded into the same answer.btclib-node’stests/conftest.pyis where that function comes from. Whether--lf,--deselectand--ignoreshould relax the floor too is a separate question, btclib-org/.github#424;-kand-mare unchanged.claude-review.yml’smentionjob refuses a missing credential in the words of the job it guards (issue btclib-org/.github#402). That job answers an@claudecomment and reviews nothing, so its step isRefuse to answer without a credentialand its message endswithout it this workflow answers nothing. The comment above the step gives the reason by pointing at the review job’s own rather than restating a measurement made there (issue btclib-org/.github#410).claude-review.yml’sclaude_argscomment names thegh prsubcommands this file uses (issue btclib-org/.github#398):diff,reviewandview, whichgrep -n 'gh pr ' .github/workflows/claude-review.ymlreads back. Spelling the three out in--allowedTools, in place ofBash(gh pr:*), is what would put that line past the 100 columnsyamllintholds the file to.claude-review.ymlconverges tobtclib-org/.github’s current mechanism (issue btclib-org/.github#340) (issue btclib-org/.github#385). The job carries a job-levelCLAUDE_REVIEW_ENABLEDswitch, so a review left disabled organization-wide skips cleanly instead of a step failing loudly; the guard step readsapi_error_statusandstop_reasonfrom the SDK’s own execution file, which the action’s log otherwise drops; and the verdict is posted as a pull request review of typeCOMMENT— one ofACK <sha>,CHANGES REQUESTED <sha>orNACK <sha>— with the verification step readingpulls/<n>/reviewsin place ofissues/<n>/comments. The header keeps this tree’s own reasoning aboutREPOSITORY.md’s required-checks rule and its own accounting of the concurrent-job ceiling.claude-review.yml’sallowed_botscomment cites the organization standard by name for what account produces the ack of record (closes #234), landing with the convergence above: it namedREVIEWING.md, which does not state which account produces one.CONTRIBUTING.mdno longer saysclaude-reviewruns beside the other gates. TheCLAUDE_REVIEW_ENABLEDswitch the convergence above adds is an organization variable that is currently unset, so neither of the workflow’s jobs runs in this tree at all while it stays that way; the sentence claiming otherwise is the same onebtclib-org/.github’s own copy ofCONTRIBUTING.mdcorrected in the commit that introduced the switch.Every fixable hook now fixes, and this file’s lint derogation is gone, closing issue #249.
codespellgains--write-changes, next tomarkdownlint-cli2andtypos, which already fixed in place. Withmarkdownlint-cli2repairing a joined file rather than only reporting it,CHANGELOG.md’s two-comment directive disabling MD022 and MD032 has nothing left to guard: a rebase that drops the blank line between two joined###sections is now the hook’s own repair on its next run, so the directive is gone and both rules apply to this file again.codespell --write-changesover the whole tree rewrote nothing, which is the measurement behind turning the flag on.ignorenames no half of a pairconvention = "pep257"settles (btclib-org/.github#178). Under a declared convention ruff disablesincorrect-blank-line-before-class(D203) andmulti-line-summary-second-line(D213) whether or notignorenames them, so an entry for either reads as an enforced choice and decides nothing. Measured with the ruffuv.lockpins, over a tree carrying a module written to violate both:ruff check --no-cache .answersAll checks passed!with the entries and without them, and ruff’s incompatible-pair warning appears only where the convention itself is taken out. Section 5 of the standard is the rule, and it now sits at the key that decides it.[tool.mypy]sets noshow_error_codes(btclib-org/.github#191). mypy has no such option:Optionscarrieshide_error_codes,Falsebefore any config file is read, andconfig_parser.py’s genericshow_/hide_inversion is the whole of why the key parses.mypy --helpunder the locked mypy writes--hide-error-codesas the flag that changes the default and gives--show-error-codesas its inverse, which is what section 6 of the standard reads as a setting mypy has already; the same erroring file reports the same message, error code included, with the line and without it.show_column_numbersstays – its default isFalse, and section 6’s sample carries it.check-sdistandpyromabuild with the backend[build-system]admits (btclib-org/.github#197). check-sdist’s default installer isuv build, and that is not PEP 517 for this backend: givenbuild-backend = "uv_build"and arequiresthe running uv satisfies, it builds with the copy bundled in whichever uv is importable or onPATH, so the backend that packed the archive the gate compares against git was not one this file declares – measured with arequireswhose ceiling excludes the running uv, whereuv buildwarns that the declaration does not contain its own version and builds the sdist anyway.args: [--inject-junk, --installer=pip]puts check-sdist onpython -m build --no-isolation, which does read the hook environment, andadditional_dependenciesnames[build-system]’s own specifier; the hook then fails withERROR Missing dependencieswhere the two disagree, which with the manifest’s args it passed. What that does not keep is the two specifiers equal: arequireswidened past the hook’s line leaves it satisfied and green, the half btclib-org/.github#145 leaves open.pyromareads the metadata throughbuildas well, without isolation where the environment satisfiesrequiresand falling back to an isolated build otherwise – the environment pre-commit.ci cannot create – and here it was the fallback that ran: in the environment that hook had,build.util.project_wheel_metadata('.', isolated=False)raisesBuildBackendException: Backend 'uv_build' is not available.The question form names the Slack link beside it rather than
CONTRIBUTING.md. The sentence pointed atCONTRIBUTING.md, which carries no such link: the channel is a contact link in.github/ISSUE_TEMPLATE/config.yml, shown to the reader as Chat with the developers on the page that offers the form (btclib-org/.github#101).REVIEWING.mdsays the reviewer runs the whole suite, every time. The organization’s copy, shared half byte for byte (section 14): the test suite is run whole on the sha under review, never a subset and never relied on from the author’s run.mypy’s optional error codes are the organization’s list, section 6 of the standard, the same in every repository that runs mypy (btclib-org/.github#165).
explicit-override,possibly-undefined,redundant-exprandunused-awaitablejoin;narrowed-type-not-subtypegoes, the locked mypy enabling it by default. Whatexplicit-overrideasked for:@overrideon the test doubles, fromtyping_extensions, declared in thetestgroup – and a named ignore on the client’s own overriding methods,typinghaving the decorator from 3.12 while the one-file module holds to a 3.10 floor and the standard library alone.REVIEWING.md’s The gates are the evidence excepts no gate from the run a reviewer may rely on, the test suite included. The organization’s copy, shared half byte for byte (section 14): a run is whole whoever makes it — never a module on its own, a-k, a--lf, a deselect or a marker in its place — and one that was narrowed or cut short is reported as no run (btclib-org/.github#168).REVIEWING.mdis the organization’s copy. A review reads the prose that stays in the tree, treats a commit message or a pull request’s body as a finding only where it decides something, and asks a stated count, a measurement nothing re-derives, or the history of the code told in a comment to go — section 14 of the standard, the shared half byte for byte.Every source file opens with
COPYRIGHT’s three lines, andnotice-rgxis that file transcribed (btclib-org/.github#119). The headers carried the MIT permission notice in full, on the argument that the copied-out module has noLICENSEbeside it; that issue decided section 5 states one notice for every tree, the pointer form, and that a self-contained one in one tree is a second notice – the third line ofCOPYRIGHTnames the URL of the license text, and a copy losesAUTHORS.mdandpyproject.tomlwith theLICENSEanyway.git ls-files '*.py'names every file rewritten, anduvx ruff check --select CPY .passes on them;notice-rgxis byte for byte whattests/copyright_test.pyofbtclib-org/.githubderives fromCOPYRIGHT, where before it was the long notice that test refused.CLAUDE.md, the README’s Vendoring, the module docstring’s own vendoring paragraph anddocs/source/conf.py’s year comment each argued the embedded notice and now describe the pointer; the entry further down this section saying the headers “still carry the MIT permission notice in full” describes the tree between that landing and this one.The lint gate runs
pretty-format-json,toml-comment-widthanddecoded-subprocess-encoding(btclib-org/.github#130 and btclib-org/.github#134).pretty-format-jsonwas left out on the claim that every.jsonhere was undertests/_data, so that an exclude would leave it matching nothing;git ls-files '*.json'on the tree before this answers.claude/settings.jsonas well, and the hook now reads that file, with.vscode/andtests/_data/excluded for the two reasons beside it –tests/_data/README.mdalready said the hook was configured to leave that directory alone, and is true now. The two local pygrep hooks are section 4’s, in the text btclib carries: the first holds a toml comment to the 80 columns[tool.ruff.lint.pycodestyle]already names for every other comment in the tree, and every comment inpyproject.tomlwas under it by hand; the second refusestext=Trueon a subprocess, andtests/standalone_test.pypassed it once, nowencoding="utf-8", the same text mode with the decoding named.[project.urls]carries the seven links section 3 ofbtclib-org/.githubnames, keyed as the other two published packages key them (btclib-org/.github#133). v2026.8.20 is on the index with six –Homepage,Download,Documentation,GitHub,IssuesandPull Requests, aspypi.org/pypi/bitcoin-core-rpc/jsonreports them – and no changelog, which is the one link from a version already published back to what changed in it.GitHubisrepository,changelogpoints atRELEASE_NOTES.mdfor the reasonbtclib-secp256k1gives beside its own, and every key is lowercase with an underscore, so that a check reading the set back across the three packages keys on one spelling.docs/source/conf.pyread theGitHubkey for the base of its links to the root files and now readsrepository; the documentation build with-Wis what would have caught the rename without it.links.ymlaccepts every status lychee would accept unasked, and stops passing a cache no step keeps (btclib-org/.github#110 and btclib-org/.github#111).--accept 200,206,429replaced lychee’s default range rather than adding to it –lychee --helpgives that default as100..=103,200..=299– so a host answering201or204was a dead link, to name a206the default already covered; the argument is now that range with429added, a rate limit being an answer from a host that is alive.--cache --max-cache-age 1dis gone with the clause that credited it: on the tree before this,grep -c 'actions/cache\|lycheecache' .github/workflows/links.ymlanswered0, so the file lychee wrote at the end of a run was read by no later one, and the comment’s “the retries, the timeout and the cache are what keep a slow or throttling host from being reported as a dead one” named a mechanism that was not there. Dropped rather than made true with anactions/cachepair, because the issue’s own measurement is that lychee collapses repeated URLs without it.This file stops asking for the blank line a union merge drops. MD022 and MD032 are off for
CHANGELOG.mdalone, by a comment at its head and not by an edit to.markdownlint.jsonc, which is section 14’s verbatim copy: a rebase of two branches that each appended a section joins them without the blank line between, and the rule would then fail the gate on a file that never conflicted. btclib-org/.github#138 is the record, and the comment says when it goes back on.claude-review.ymlreports red on anything but an ack of this head. The job ended at a step testing whether the action had run at all, so aCHANGES REQUESTED, a run that posted no comment, and an ack naming a sha the branch had moved past each left the check green – btclib-org/.github#146 measured that guard and nothing past it in every copy of the workflow. The step btclib-org/.github carries at18e6c64is here now, with its comment: it reads the last verdictclaude[bot]posted on the pull request and fails unless it is anACKwhose sha is a prefix of the head. Run outside the workflow against this repository’s own pull requests, withREPO,NUMBERandHEAD_SHAin the environment: on PR 222 at its head it printsthe review acked 18affcac...and exits 0; with any other head it exits 1 naming both shas; on PR 219 and PR 215, where noclaude[bot]verdict was posted, it exits 1 saying so. Still not a required check, and red there gates nothing. The two copies differ beyond this step – the header, the checkout comment,allowed_botsand the prompt are each this repository’s – and those stay as they were.CODE_OF_CONDUCT.mdis gone, and the inherited copy is what GitHub shows. Section 14 ofbtclib-org/.githubno longer lists it and section 2 no longer asks for one: the file was a pointer to the PSF code of conduct, and the copy inbtclib-org/.githubis displayed for a public repository carrying none (btclib-org/.github#123). The entry further down this section that calls it one of three verbatim files describes the tree between that landing and this one.test.yml’sprosepattern stops naming it, a name nothing matches being a rule that has quietly stopped applying. Nothing else linked it: on the tree before this,git grep -n CODE_OF_CONDUCTanswered with that pattern, the file itself and this file’s earlier entries..gitattributesis the organization’s file (btclib-org/.github#102). Section 14 of the standard names it as the same file in every repository, andtests/verbatim_test.pythere compares the copies it finds; the difference here was in the comment and never in the twomerge=unionlines. It is byte for byte the copy inbtclib-org/.github, on the digest of the raw file the contents endpoint serves. This tree has no attribute of its own to keep under## This repository in particular, so it carries no such heading.git check-attr mergestill answersunionfor both files.COPYRIGHTleaves the sdist.license-fileshere was already["LICENSE", "AUTHORS.md"], so the wheel never carried it, andsource-includenamed it by hand. btclib-org/.github#135 decides that every package of the organization ships the same set and the file is not in it – the holder a consumer needs is inLICENSE, andCOPYRIGHTis the source of a header rather than a statement to a consumer. Measured onuv buildbefore and after: the sdist’s root losesCOPYRIGHT(andCODE_OF_CONDUCT.md, above) and nothing else, and the wheel’sdist-info/licenses/isAUTHORS.mdandLICENSEboth times.[tool.check-sdist]git-onlygains the name, that hook otherwise reporting a tracked file the archive does not carry.The build backend is
uv_build, andMANIFEST.inis gone with the include list it was. btclib-org/.github#118 settles that a pure-Python project uses that backend, andbtclibis where the pin comes from:uv_build>=0.12.5,<0.13. What the floor has to clear is 0.12.0, where the sdist’s ownpyproject.tomlbecame a normalized copy with the verbatim one kept beside it aspyproject.toml.orig– which this tree’s sdist now carries, matching the file in the checkout byte for byte. Measured by replacing the specifier with==<version>and reading the member list,0.11.31answers with no.origand every0.12.xwith one, so a floor below0.12.0would be a different archive rather than an older one and0.12.5is not that boundary. It sits there because0.12.5is the rev.pre-commit-config.yamlpins foruv-pre-commit, which keeps the backend no older than the uv whose lock format that hook writes, and because it isbtclib’s floor – two reasons to hold it rather than constraints, which is what makes it the number to lower first. The ceiling is the next minor, where uv’s versioning policy puts a breaking change.What
MANIFEST.insaid in include and exclude lines is now the glob patterns of[tool.uv.build-backend], beside the rest of the configuration. Built both ways and the archives compared:uv build --sdist -o dist-before # on origin/main uv build --sdist -o dist-after # here diff <(tar tzf dist-before/*.tar.gz | sed 's|^[^/]*/||' | sort) \ <(tar tzf dist-after/*.tar.gz | sed 's|^[^/]*/||' | sort)
The only tracked file that leaves is
MANIFEST.in, which this diff deletes. Everything else the command prints is backend metadata or the spelling of a directory member:setup.cfgand the four files ofbitcoin_core_rpc.egg-info/go,pyproject.toml.origarrives, and where setuptools wrote seven directory members, every one with a trailing slash in the name,uv_buildwrites six: the root keeps its slash, the five nested ones lose it, andbitcoin_core_rpc.egg-info/is the seventh, gone with the backend that made it. So the listing comparesdocs/againstdocs. It does not count them; this does:python3 -c 'import sys, tarfile print(sum(m.isdir() for m in tarfile.open(sys.argv[1]).getmembers()))' \ dist/*.tar.gz
The wheel answers the same question the other way round: it loses
dist-info/top_level.txt, which is the only file entry the two builds differ on, and it gains three directory entries where setuptools wrote none —bitcoin_core_rpc/,dist-info/anddist-info/licenses/.py.typedis among the members it keeps, with nopackage-dataentry left to name it.What the backend reads is the working directory, walked through those globs, and not git: an untracked file matching one of them is packed like any other and moves the digest of the archive built from a given commit.
check-sdistturns that into a red gate rather than a shipped archive,source-excludenames the caches a linter or a type checker is known to leave undertests/ordocs/— btclib shipped a nested one whole (btclib-org/btclib#985) — and RELEASING.md’s rebuild section now carries it as the first of three bounds on what a digest comparison proves, with the clean-export command beside it.check-sdistreplacescheck-manifestin the lint gate, asking the same question – what is tracked and not in the sdist, and what is in the sdist and not tracked – against the patterns of[tool.uv.build-backend]rather than against an include list no backend reads any more. It registers a plugin for that backend, so it builds with the uv it installs and needs neither--no-build-isolationnor asetuptoolsadditional dependency.[tool.check-sdist]git-onlyholds what no include pattern adds in the first place, and it was read off the tool rather than carried over: with the list emptied the hook reports.gitattributes,.gitignoreand the tracked files under.claude/and.vscode/, and nothing else —.githubis on check-sdist’s own default ignore list, and.lycheeignore, which[tool.check-manifest]ignored, is not tracked here at all.The sdist normalizer stays, and its reason changes. It existed because setuptools tarred a staging directory whose sub-second
mtimeSOURCE_DATE_EPOCHdid not reach. That backend is gone and the nondeterminism with it:uv_buildignoresSOURCE_DATE_EPOCHand writes every sdist member at epoch 0, mode 0644 or 0755, owner root with no names and no PAX records, the gzip header at 0 and the wheel’s zip entries at 1980-01-01 – a build with the variable exported and one without give one sha256 per artefact. So the script now runs over an sdist already deterministic without it, and it runs for the reason btclib’s copy of it states: epoch 0 is uv’s choice and uv may revisit it, while a release rebuilt from its tag has to give the bytesrelease.yml’sattestjob vouched for however many backends later that is.What it does there is a rewrite and not a check, and the difference is the digest this repository publishes: every member’s
mtimegoes from uv’s 0 toSOURCE_DATE_EPOCH, which moves the archive.uv build --sdist -o dist && shasum -a 256 dist/*.tar.gz SOURCE_DATE_EPOCH=$(git log -1 --pretty=%ct) uv run --no-project \ --python 3.14 .github/scripts/normalize_sdist.py dist/ shasum -a 256 dist/*.tar.gz
So the step in
test.ymlis what decides the bytesattestsigns, and dropping it would publish others. Its docstring carries that argument and the measurement under it,SOURCE_DATE_EPOCHstays because the script is what reads it, and the mode is rewritten on the same footing as the timestamp – there the line really is a no-op, 0644 and 0755 being whatuv_buildalready writes – rather than against a umask no backend now consults.claude-review.ymlpoints at the file that states the ceiling, instead of at a measurement taken of something else. Its header promised that “REPOSITORY.mdmeasures what a commit at that ceiling costs in wall clock”.REPOSITORY.mddoes measure a wall clock against that ceiling, but one job’s and not one commit’s – the seconds of a slot a singleAnalyze (python)held – while what it says a commit asks for is a count of jobs, which42e600esuperseded. Following it does land on a measurement, which is what makes the wrong generality worse than a dangling reference: the reader arrives at a number and takes it for the one that was named. The ceiling isCONTRIBUTING.md’s to state, andCONTRIBUTING.mdis also where what the platform sweeps ask of it is accounted for, so the sentence names that file. Restating the number here would be one more copy of what the tree already repeats:git grep -n 'concurrent jobs' -- '*.yml' '*.md'
That form rather than
'twenty concurrent', which misses two of them:CONTRIBUTING.md, where the sentence wraps between the two words, and this very workflow, whose header stops repeating the number here. The narrower pattern answers with neither the file the ceiling belongs to nor the file this entry is about.A bare
uv run pytestis the coverage gate, and the prose that said otherwise is gone.CLAUDE.mdcarried “uv run pytestis not the coverage gate” and a longer command to use instead;--covhas been inaddoptssince it moved there so that the ratchet is what a bare run measures, andpyproject.tomlsays as much beside the flag. The run answers withRequired test coverage of 100.0% reached, which is the gate. What is true, and is what the moved prose now says, is that a selective run is reported and not gated:uv run pytest tests/transport_test.pyprints the whole tree’s coverage and passes,tests/conftest.py’spytest_configurebeing what makes the difference.CLAUDE.mdholds only what no document written for a human can. The commands, the gates and how they lie areCONTRIBUTING.md’s last section, and the review rules areREVIEWING.md: a human had to open an agent’s file to learn how to run a gate, and an agent’s file was the second place a review rule was written. What is left is what those two cannot say because it is about a session rather than about the tree – what the module is made of, the worktree rule, the model, and the failure modes that otherwise cost a session,cancellednot beingfailureand a draft pull request running nothing among them.REVIEWING.mdis the organization’s file down to the same heading, and.claude/commands/review.mdis its copy of the invocation. A review that means one thing here and another in a sibling is no standard, which is section 14’s reason for comparing the two halves that are meant to agree. What the shared half says is gone from the tail: what is under review, what to look for, what a finding contains, what becomes of a collateral finding, the gates as evidence, the verdict and the re-review. The tail keeps what a review of this tree checks and a generic one does not – a dependency or a second module, both JSON-RPC dialects, the boundary the layers draw, and whether the suite is still hermetic..claude/commands/review.mdhad drifted into naming this tree’s facts, which is not what a verbatim file can carry: those facts are in the tail above, and the command is the invocation again.CONTRIBUTING.mdis the organization’s file down to## This repository in particular. Section 14 ofbtclib-org/.githubhas it verbatim to that heading and this tree’s below it, andtests/verbatim_test.pythere compares the halves that are meant to agree. What the shared half already says is gone from here rather than said twice in this repository’s words: where an issue is filed, the prose style, what a pull request is and answers, one subject per pull request, how a review is exchanged and how a change lands. Each of those now points at the section of the standard that decides it.What stayed is what only this tree can say: the constraint that this is one source file with nothing but the standard library behind it, the commands that build the environment and run the gates, the editor files, the command each CI job runs verbatim, the cadence table, the mutation session, the live node and the secrets baseline. The gates moved here from
CLAUDE.md, a human having no reason to open an agent’s file to learn how to run one, andREVIEWING.mdreads which checks are required from this file’s last section.The toolchain badges move under that heading with them, the shared half being byte-compared and a badge line being this repository’s.
.gitattributespointed here for the union-merge reasoning and now points at section 9, which is where it went;docs/source/conf.pynamedCONTRIBUTING.md’s links toCODE_OF_CONDUCT.mdandRELEASING.mdas the reason the root files are not copied into the documentation tree, and namesREADME.md’s link toLICENSEinstead, that being the one such link left..yamllint.yamlis the organization’s copy, byte for byte. It is one of the files section 14 ofbtclib-org/.githubcalls verbatim, and the copy there has moved:document-startis raised from the default set’s warning toerror, the pre-commit hook running plainyamllintwith no--strict, where a warning exits 0 – so at the inherited level the rule reported the convention and gated nothing. It costs this tree nothing today:git ls-files '*.yml' '*.yaml' \ | xargs -I{} sh -c 'head -1 {} | grep -q "^---" || echo {}'
answers with nothing here. The hook’s own comment no longer lists what the file departs from the default set in, that list being one more thing to keep in step with the file itself.
AUTHORS.mdnames this repository’s authors. It pointed at this repository’s contributor graph while calling them btclib’s, which is the shape section 14 ofbtclib-org/.githubgives for why the file cannot be verbatim: a shared pointer is accurate only while one graph stays a superset of the others, and the first person to contribute somewhere else goes uncredited in silence.LICENSE,COPYRIGHTandCODE_OF_CONDUCT.mdare the organization’s copies, byte for byte. Section 14 ofbtclib-org/.githubcalls the three verbatim, andtests/verbatim_test.pythere compares them across the repositories, so a copy that reads better here is a copy that fails everywhere.LICENSEgains theMIT Licensetitle and loses the2017-2026range – a range is a line nobody updates, andCOPYRIGHTstates the holder without one, so the two disagree the first January nobody remembers.CODE_OF_CONDUCT.mdis the file GitHub falls back to for a repository of the organization carrying none, which is why a repository that carries one carries that one.COPYRIGHTis now the three-line notice the other trees point at, and the headers here still carry the MIT permission notice in full: what the ruffCPYrule requires of a source file isnotice-rgxinpyproject.tomland never this file, and the argument for the long form – a vendored copy has noLICENSEbeside it to point at – is aboutbitcoin_core_rpc.py, which is copied, and not aboutCOPYRIGHT, which is not. That comment says so where the rule is, and names the section it departs from.A workflow’s file name says its own subject, and a shared prefix groups a family.
ubuntu.yml,macos.ymlandwindows.ymlare nowos-ubuntu.yml,os-macos.ymlandos-windows.yml: the platform sweep is a family, and the prefix is what says so in a directory listing that is read alphabetically.latest.ymlisdeps-latest.yml,latesthaving named an adjective without the noun it qualified – what that workflow moves is the dependency resolution, theos-*sweep being the one that holds it at the lock.published.ymlispypi-install.yml, which says what it asks of the published package rather than that there is one: it installs it from the index and round-trips a call through it.integration.ymlisintegration-bitcoind.yml, integration naming neither of the two sides it joins, where the other side is a live node the workflow downloads and starts.No job name moved, and that is the constraint the renames were written under: a job name is the context a required status check carries, and
integration: every job passedis one of them, so renaming it would drop a check and add another under a rule that blocks every open pull request in between – REPOSITORY.md’s Required checks on main has the cost in full. Each workflow’sname:key and its concurrency group follow its file, the group being a literal this repository writes for itself rather than a name anything else reads.What the rename leaves owed is a row apiece in section 10’s calendar. That table gives a workflow a day and an hour under its file stem,
tests/grid_test.pyofbtclib-org/.githubreads it against the trees, and a stem it does not name is reported as a cron on no calendar. No cron moved – this repository’s minute is12and each still lands on the slot its old name held – so what is owed is the table, in the repository that holds it:grep -l 'cron:' .github/workflows/*.yml
against section 10’s workflow table. The rows this repository leaves behind stay answered by btclib, whose
.github/workflowsstill schedules every one of the old names, so none of them is left dangling by this.A
TODOthat opens a comment no longer passes lint (btclib-org/.github#87). Ruff’sFIXfamily joins[tool.ruff.lint] select, so a comment beginningTODO,FIXME,XXXorHACKis a finding, on its own line or after code. Unfinished work belongs in an issue, where whoever might do it can see it and where it can be closed; a marker in the tree is visible only to somebody already reading that file, and it outlives the branch that put it there without anyone noticing.What the rule does not reach is worth knowing before trusting it: a marker further into a comment, one inside a docstring or a string literal, and any file that is not Python. It cost nothing to adopt – the tree measures zero findings under it, so this is a ratchet on something already true rather than a cleanup, and no code changed.
The gate runs one image and one interpreter, and the sweeps moved to the calendar (btclib-org/.github#85).
test.yml’ssuitejob is gone: three images by seven interpreters, minus one cell, asked before every review on an organization whose plan gives it twenty concurrent jobs across every repository. Measured on one run of the workflow carrying it, 24 jobs did 5.8 minutes of work and spent 36.5 minutes queueing. What waits for a review isubuntu-lateston the version.python-versionnames, and the images that left areubuntu.yml,macos.ymlandwindows.yml— the last renamed fromwindows-arm.yml, which had stopped naming what it runs the daywindows-latestjoined it.Each of those runs its matrix whole, the cell the gate already covers included. A sweep that subtracted the gate would be a matrix with a hole in it, and whoever asked what ran would have to re-derive the hole from another file.
integration.ymlgains the same shape and aversionsjob to give it one: the compatibility boundary — the last Core major that has never heard of the"jsonrpc": "2.0"marker, and the newest — before a merge, every supported major on the schedule. Its chain cells go from four to five,regtesthaving been left out on the grounds that the protocol cells start one in passing; a matrix row that reads as itself outlives the comment explaining where the fifth went.Every schedule here is on the organization’s grid, which gives a workflow its day and its hour and this repository its minute. Seven crons moved;
published.ymlgoes from monthly to weekly with the rest. The grid is section 10 ofbtclib-org/.github, which is why the cadence table in CONTRIBUTING.md no longer names a day: one calendar covering six repositories is one thing to remember, and six copies of it are six things to keep true.The interpreters this package claims are the ones it runs on (btclib-org/.github#83).
requires-python, the per-versionProgramming Language :: Python ::classifiers and the interpreter list the workflows run are one fact written three times, and nothing compared them.tests/interpreters_test.pydoes: the floor is the lowest classifier, the classified set and that list’s CPython set are each other, and the PyPy classifier is present exactly when a PyPy interpreter runs.The drift it catches misleads somebody who is not reading this repository — PyPI shows a classifier to whoever is choosing the package, so a version left behind when a floor moves is an interpreter advertised and never touched.
It does not encode the calendar. The organization standard’s rule is that a library covers every Python still in support, which moves twice around each October; python.org keeps that schedule, and a date written here would be one more thing to move. The claim held is the weaker and checkable one: whatever the three say, they say the same thing.
Which of section 7’s conventions this suite tests is declared, and a test says the declaration is true (btclib-org/.github#32). A new
tests/README.mdnames each convention the organization standard lists that this repository tests and the module that tests it, then names the ones it does not.Both of the ones it does test are answered by
standalone_test.py, which is the shape section 7 has in mind where it says what must not be aligned is where these live. This package is one file, so what elsewhere needs a walk over a tree is here a property of that file, and stronger for it: the documentation is every name of__all__carrying a docstringautomodulewill render, againstdocs/source/api.rst’s claim to list the whole public surface; the import graph is the source importing nothing outside the standard library, and a copy of it running underpython -I -S.Among those it does not, two are near misses the file names so that “absent” is not read as “overlooked”.
__all__is declared and read, but to ask whether each name is documented rather than whether the module’s public names are all in it — section 7 asks for a census that fails when a new public name appears, and nothing here fails. Two tests assert a keyword-only signature, each about one function, where section 7 asks for it as a rule over the package.tests/conventions_test.pyis what keeps the declaration from being prose, and its assertions were checked by making the declaration wrong and watching the suite go red.A Dependabot pull request can be reviewed (btclib-org/.github#77).
claude-review.yml’s review job failed on #207, the first Dependabot pull request opened since the credential guard landed, and it failed twice for two different reasons. The first was the credential: apull_requestrun whose actor isdependabot[bot]reads from GitHub’s Dependabot secret store rather than from Actions secrets, and that store was empty, so the guard fired exactly as designed. The second is this change: the action refuses a run a bot initiated unless the bot is named — “Workflow initiated by non-human actor”.allowed_bots: "dependabot[bot]", named rather than*: the input’s own description warns that on a public repository*lets an external App invoke the action with a prompt it controls. The mention job takes no such input, what triggers it being somebody writing@claude.Without this a Dependabot pull request is the one class that can never carry the ack of record
REVIEWING.mdrequires, which is the class whose whole value is landing promptly.The sibling repository is named
btclib-secp256k1in the seven places here that named it otherwise, under either spelling. Hyphenated, in.markdownlint.jsonc,.github/dependabot.ymlandRELEASING.mdtwice; with an underscore, inREADME.md,CONTRIBUTING.mdandCHANGELOG.md, where what is named is the sibling that “carries the same three lines” of badges and “the same CONTRIBUTING.md run” — which a repository does and a distribution does not.grep -rn libsecp256k1over the tree is what says whether any is left..markdownlint.jsoncis the one that reaches past this repository: §14 of the organization standard holds that file identical in every repository, andshasumover the five copies is what says whether it does. With this one it is four of five,btclib-nodebeing the last.A markdown line does not end inside a word (btclib-org/.github#71). Markdown joins two source lines with a space, so a word wrapped at its own hyphen renders with the hyphen and then a space inside it – “read-any-” and “number-of-times” on consecutive lines come out as “read-any- number-of-times”. The source looks correct, which is why reading a diff does not find one, and nothing here read the output: markdownlint has no rule for it, the width rules read a line rather than what two lines become, and
sphinx-build -Wis not asked whether a token means anything.A
pygrephook, for the reasonlocal-link-prefixbeside it is one: nothing off the shelf does it and the pattern is one line. The rule and what the hook cannot see – a code span whose content breaks at a/or a.renders the same way and has no hyphen to match – are in section 4 of the organization standard, which the comment points at rather than restates. This tree was already clean under it, so the hook costs nothing today and exists to keep the next one from being written.claude-review.ymlstops counting the required checks (btclib-org/.github#22). “The four required checks stay what they are” is a count, and a count in a workflow comment is checked by nothing:actionlintreads the workflow,zizmorreads it for injection, and prose beside them reads as authoritative for sitting there. It happens to be right here and is wrong in two sibling repositories, which is the same sentence in the same file – the shape that made it worth removing rather than confirming.What the paragraph is for survives without the number: this job is not a required check and must not become one, a model’s judgement not being a branch rule. Which contexts are required is
REPOSITORY.md’s to record and the endpoint’s to answer, and the comment now says so. Found by the workflow-comment sweep that issue asks for.release.yml’sconcurrency:comment stops counting the workflows it calls (issue #175). “this one calls three” was true when it was written –git log -S 'this one calls three'finds the one commit, andrelease.ymlcalledtest,lintanddocsat that point. Four more were added under it since,macos,windows-arm,integrationandpublished, none of them revisiting a comment far enough above to be off the screen – the nearest of those fouruses:lines is hundreds of lines below it, which is the distance that made the sentence safe to leave alone.The count carried nothing the reader needed. What the comment is for is why the group is named literally –
github.workflowinside a called workflow is the caller’s name, so every one of those calls would otherwise share this workflow’s group – and that argument is the same whether the number is three or seven. It now says “several” and gives thegrepthat answers it in the tree that is asking, which is what stays true the next time one is added.The yamllint hook’s comment says what the hook reads (btclib-org/.github#68). The change that turned the default rule set back on renamed the hook and left the paragraph above it describing a width check – “the two rules enabled and the two left off”, which was true of the
.yamllint.yamlthat change replaced. That is the same defect the change existed to fix, one file over: a comment naming a narrower scope than what runs, with nothing to notice. Caught by the review of the sibling change inbtclib-benchmarks#148, after this repository’s had already landed. The paragraph now says what nothing else here reads – width being the half a sibling gate could in principle have covered, where a key written twice or a block indented under nothing is read by no other hook at all..yamllint.yamlturns the default rule set back on (btclib-org/.github#68). The file listedline-lengthanddocument-startunderrules:and extended nothing, and yamllint enables no rule a configuration does not name – so those two were the only rules that had ever run here. Indentation, trailing whitespace, duplicate keys, colon spacing and the rest of the default set were off, while the file’s own comment namedcommentsandtruthyas the exceptions and read as though the rest were on.Three things kept it out of sight, and each generalizes past this file: the comment said otherwise; the gate stayed green, because removing a check cannot make a conformant tree fail; and the file is shared byte-for-byte, so the defect travelled by being copied rather than by being written twice.
extends: defaultis back, and the two exceptions are named underrules:as explicitdisables where a reader sees them instead of inferring them from what is missing.The hook’s own name follows it: “yamllint (line width and document start)” was an accurate description of a defect, and is now “yamllint (the default rule set, less two)”.
It costs this tree nothing:
git ls-files '*.yml' '*.yaml' | xargs uvx yamllintreports the same nothing before and after. That is a fact about how the yaml here is written, not an argument for having left the set off – what a rule is worth is what it catches the day somebody writes the line..taplo.tomlstops carrying one tree’s furniture (btclib-org/.github#40). The file is shared byte-for-byte across the organization’s repositories, deliberately, and this copy justified its two settings by naming things that are one tree’s:reorder_keysbecause “[project]reads top to bottom, and the ruff rule sets are grouped by what they cover”, the indent because “the arrays inpyproject.tomlalready use” it. Both readings are true here and neither is the reason – the reason is that a table’s order is an argument rather than an accident, and that one indent across every toml in the organization is the point of having a formatter. A shared file arguing from local furniture is a file the next repository inherits an untrue sentence with.It is now byte-identical to the copy
btclibandbtclib-benchmarksalready carry, which argues from the rule and says outright that it is shared and that nothing in it may be true of one tree only. Neither setting moves:indent_stringis the same four spaces andarray_auto_collapsethe samefalse, which the formatter leaving every toml here untouched is what shows.docs/source/conf.py’s reason for not copying the root files names the links thisCONTRIBUTING.mdhas (issue #188). The paragraph rejecting a build-time copy said the file “links to CODE_OF_CONDUCT.md and tests/_data/README.md”. It does not link the second and never did:git log -S 'tests/_data/README.md' -- docs/source/conf.pyreturns the commit that opened this repository, andCONTRIBUTING.mdat that commit namedCODE_OF_CONDUCT.mdand nothing else outside the documentation. The sentence isbtclib’s, where it is true — that tree’sCONTRIBUTING.mddoes link./tests/README.md— carried over with the path bent to a file this tree happens to have. It is the shape.yamllint.yaml’s survey had: a claim that reads as measured because it was, somewhere else.The conclusion never depended on it. One destination outside the documentation is enough to reject copying, and there are two here as the sentence always said:
CODE_OF_CONDUCT.md, andRELEASING.md, whichCONTRIBUTING.mdlinks and no shim indocs/source/includes. The comment now names that second one, and says what decides the question — whether a shim includes the file — so the next reader checks the thing that settles it rather than a list. What the wrong name cost was never a build: it was a reader running the check the comment invites, watching it fail, and having no way to tell whether the decision rested on a misreading or the tree had moved under a decision that was right.CONTRIBUTING.md sends a contributor to the organization standard (btclib-org/.github#52).
README.mdinbtclib-org/.githubstates the toolchain, the lint gate, the workflow set and the branch rules once for this repository and its siblings, and it claims to be linked from each repository’s CONTRIBUTING.md. Nothing here named it. Every mention of that repository in this tree cited an issue filed there, to justify the rule that issue records; not one of them told a reader the document itself exists. A sibling repository has one mention that is a link rather than a citation, into a subsection of the standard to justify a setting, and it makes no difference to this: a destination reached only by somebody already reading the paragraph that carries it is not a signpost either.git grep -n 'btclib-org/\.github'is what re-derives them, a list here going stale the next time one is written. So a contributor following CONTRIBUTING.md to REPOSITORY.md to CLAUDE.md was never told a document above them existed — and a rule stated only there was one they could not find. The pointer is in the opening rather than in REPOSITORY.md or CLAUDE.md: the audit and the normalizing checklist that standard carries are performed holding it, so the reader who arrives without it is the contributor, and this is the file that reader is already in — and the one of the three the documentation build renders,docs/source/contributing_link.mdincluding it. Hence the absolute github.com url, the shape a sibling repository is linked with elsewhere here: a relative destination resolves against the rendered site.RELEASING.md records why no SBOM is attached to this repository’s releases. This repository declares
dependencies = [](pyproject.toml:62), andpublished.ymlalready asserts that against the installed package on every run (assert not requires(name), monthly and on dispatch). btclib’s RELEASING.md carries the reasoning for why that makes a bill of materials here redundant with what CI already checks directly, in its “Read the bill of materials attached to the release” step; see that and btclib-org/btclib#1159 for the evaluation.That step names the trigger it shares with its siblings (btclib-org/.github#24).
dependencies = []is this repository’s own premise, and the sentence already said what would unseat it; what it did not say is that the decision rests on a second premise it holds in common with the siblings — btclib’s generator builds itscomponentsfromRequires-Dist, so a generator that learned to describe a component that metadata cannot express would reopen the question everywhere the decision is recorded, here included. The step now names that second trigger and the open issue watching it, where btclib-org/btclib#1159, cited beside it for the evaluation, watches nothing.The Slack badge is gone from README.md and CONTRIBUTING.md. The badge block’s own comment states the criterion — a badge reports state, which is why “we use ruff” is not one — and this badge reported a route, to a channel of the course workspace the sibling library was first taught in. What answers a question here is the issue tracker and a pull request, which leave the answer where the next reader of these files can find it. This repository carried the badge alone, with no bullet sending a contributor there, so the block loses a line and nothing else changes. btclib does the same in its own pull request, and the organization profile no longer names Slack either.
check-wheel-contentsgains apackageconfiguration, closing btclib-org/btclib#1160’s divergence for this repository: the two other publishing repositories in this organization each carry a form of a wheel-contents check the third lacked.btclibhas a hand-written script comparing the built wheel and sdist member for member against an allowlist;btclib-secp256k1configurescheck-wheel-contentswithignoreentries for the members its compiled wheels carry that a pure-Python one does not. This repository rancheck-wheel-contentswith no configuration at all. The comment above[tool.check-wheel-contents]inpyproject.tomlhas what changed, the measurement behind it, and why no script was ported from either sibling — one place for that reasoning, pointed at rather than repeated here.test.yml’s andlint.yml’s concurrency groups take aconcurrency-suffixinput, andrelease.ymlpasses-releaseat both call sites, part of btclib-org/btclib#1158 (the issue lives inbtclib; this repository was already right about the other half of it).release.ymlcalls both workflows, and inside a called workflowgithub.refis the ref of the release run itself: aworkflow_dispatchrehearsal frommaincomputesrefs/heads/main, the same group a push tomainholds, withcancel-in-progress: true, so one cancelled the other. Not exercised directly here — reproducing it costs an actual rehearsal race — butbtclib-secp256k1had already fixed it with the same mechanism this copies, and this repository’s own concurrency key (github.event.pull_request.number || github.ref) already carried the other half of the issue’s fix, so this pull request changes only the suffix.CLAUDE.mdis corrected in the same pull request: it still named the pre-#1152test-${{ github.ref }}form; the actual group had already moved on.release.yml’sgithub-releasejob gains an explicitif, closing issue #149:always() && needs.publish-pypi.result == 'success' && needs.attest.result == 'success'. Without it the job was skipped on every tag sinceattestentered itsneeds— v2026.8.12, v2026.8.13 and v2026.8.20 each published to PyPI and cut no release, and each was recreated by hand from the run’s owndistandattestationartifacts. The cause is not a failure anywhere in the run, which is what it looked like:attestneedspublish-testpypi, a tag push skips that job, and a job standing behind a skipped ancestor is skipped whatever its own needs report, unless its own condition begins withalways()— an opt-out that holds for the job saying it and not for whoever needs that job. The two.resultchecks withoutalways()are the fix that is not one, measured on btclib-secp256k1’s v0.8.0.2 (btclib-secp256k1#141, then #151); the condition here is that repository’s, which cut v0.8.0.3 and v0.8.0.4 from the workflow. It keeps the property the absentifwas relied on for: aworkflow_dispatchrehearsal skipspublish-pypi, so the condition is false and no release is cut. RELEASING.md’s troubleshooting entry for the symptom names the cause instead of owing a report to GitHub support, and its post-release check now asksgh release viewfor the release’sauthor—github-actionsis the workflow having cut it.RELEASING.mdgains an explicit decision point for a dependency’s drift, and the actual command that lands a release pull request here, both findings from cutting btclib-secp256k1’s 0.8.0.4 (btclib-secp256k1#288, #289, #290): whether to act onuv lock --upgrade’s drift or leave it belongs in the release pull request rather than defaulting by omission, andgh pr merge --squashalone refuses that same pull request client-side on a solo-maintainer repository —--adminclears it, measured on btclib’s #1111, #1113, #1114 and #1133, naming the release commit’s title and body explicitly rather than leaving them to this repository’sCOMMIT_MESSAGESsquash default.RELEASING.md, REPOSITORY.md, CONTRIBUTING.md and README.md stop stating counts nothing checks (#150), the prose rule CONTRIBUTING.md already carried but 22 places in these four files hadn’t caught up with: an unguarded number restating a job count, a permission count or a structural claim, removed or replaced by naming the things directly, so two files describing the same fact can’t say it two different ways. One of the 22 was already wrong rather than merely fragile: CONTRIBUTING.md’s workflow table read
integration’s matrix as “2 Core versions, 4 chains”, which next to theN platforms × M interpretersrows around it reads as a cross product, 8;integration.yml’s actual matrix is 6 — twoprotocolcells plus fourchaincells, additive — matching what REPOSITORY.md’s own “six matrix cells” already said correctly.N platforms × M interpreterstable rows are otherwise left untouched, and GitHub’s platform-limit numbers stay as the dated historical constants they already were.RELEASING.md’s environment step named the wrong holders of
id-token: write(#161): “publish-pypiandpublish-testpypiare the only holders” was false,attestdeclaring it too for its own Sigstore exchange, which REPOSITORY.md’s “Token permissions” section already got right. Fixed to nameattestand explain why the gate still covers it:attestonly runs vianeeds: [publish-pypi, publish-testpypi], after one of the two reviewed jobs has already succeeded, so nothing bypasses the environment review the sentence was arguing for.A
tag-integrityruleset now enforces the other half of issue #139, requiringrequired_signaturesonrefs/tags/v*, with no bypass actor. RELEASING.md’s tagging step already produces a signed tag by default (v2026.8.20’sgit tag -schange), so nothing about the release procedure changes; what the ruleset adds is that an unsignedv*tag is now refused outright rather than merely undocumented. It carries nodeletionornon_fast_forwardrule on purpose: RELEASING.md’s own recovery path deletes and re-tags a release that failed beforepublish-pypi, and either rule would block that. Existing tags are unaffected — the ruleset applies to pushes going forward, not retroactively — closing the issue now that both options it named are done. REPOSITORY.md records it under a new “Tag protection” section.test.ymlgains achangesjob, the cheapest one in the workflow and the one that decides whether the rest of it runs at all: a pull request that edits only prose —CHANGELOG.md,RELEASING.md, the rest of the root’s own documentation,.github’s markdown, the mutation configuration, the detect-secrets baseline — changes nothing the matrix reads, and was spending the whole matrix to prove it.suite,coverageanddisteach skip when it answers nothing to check, andtest-passedstill fails if it answers wrong: a failure inchangesshows up intest-passed’s own results directly rather than being read as everything downstream correctly declining nothing to do.README.mdis deliberately not on the prose list — thedistjob’stwineandpyromasteps check it, so an edit to it is a change the matrix can fail on. Mirrors btclib-secp256k1’s ownchangesjob (#189), closing issue #145:release.yml’stest:job — the one callingtest.ymlwithuses:— now grantspull-requests: readalongsidecontents: readexplicitly, a called workflow’s jobs being capped at what the caller grants rather than at what the callee’s own top-levelpermissions:declares. Left ungranted, thechangesjob’s static declaration alone would refuse the whole release run before a single job started, the way it did in btclib-secp256k1’sv0.8.0.3(btclib-secp256k1#281) — even thoughchangesnever actually reads a file list on that path, answeringcode=trueunconditionally off anything that is not apull_requestevent, the check GitHub runs at startup being against what the job declares rather than against what it goes on to do.version-checknow asks where the tag is, closing issue #153:release.yml’sCheckout codestep in that job gainsfetch-depth: 0, and a new step readsgit merge-base --is-ancestoragainstorigin/<default branch>before anything is built. The three checks already there ask whether the version is right; none of them asked whether the commit carrying it is reachable from the branch a release is supposed to ship — a tag pushed from a stale worktree, a branch whose pull request is still open, or a commit squashed away on its way intomainwould otherwise publish a tree no review approved, and PyPI accepts no file name twice, even after a yank. Ported from btclib’s own form of the check rather than btclib-secp256k1’s: btclib verifiesorigin/<default branch>is present first and fails with the reason —origin/<default branch> is not in this checkout— where btclib-secp256k1’s baregit merge-base --is-ancestorwould fail withNot a valid object name, correct but silent about why. The placeholder-tag guard btclib carries beside this check is not ported: this repository’sCheck that the tag matches the declared versionstep already refuses a two-component tag.The
builddependency-group is renamedcheck(issue #156): it heldcheck-manifest,check-wheel-contents,pyromaandtwine, the tools that inspect a distribution, not the ones that build it, and--only-group buildinstalled the opposite set of packages in btclib-secp256k1, which already has both abuildgroup for building wheels and acheckgroup for inspecting them. A step copied between these trees — repository btclib’s issue #1154 is about to do more of that copying — installed the wrong four packages under the old name.pyproject.toml, itsdevgroup’s reference,uv.lock,CONTRIBUTING.mdand every--only-group buildcall site intest.yml,latest.ymlandrelease.ymlmove tocheck; nothing in btclib-secp256k1 changes, and the analogous rename in btclib is tracked separately, as repository btclib’s issue #1155. The rename’s actual return:--only-group buildhere now fails —Group `build` is not defined in the project's `dependency-groups` table— instead of quietly installing the wrong four packages, so a step copied in by mistake stops rather than runs.[tool.uv]’srequired-versionfloor was above what Dependabot’s own updater can satisfy (#159), and not merely a wrong comment: the updater independabot/dependabot-corerunsuv lockwith the uv it bundles (uv/Dockerfile’sFROM ghcr.io/astral-sh/uv:...,0.12.1at the time of writing) and refuses instead of upgrading itself when this key asks for more, turning every uv-ecosystem update it attempted, security ones included, into a silent no-op rather than a pull request — this repository’s.github/dependabot.ymlhas run theuvecosystem weekly since its first commit and produced none, whileruffandmypysat one point release behind both sibling repositories on packages all three share.>=0.12.3was the floor doing this;>=0.11.31replaces it, matching repository btclib’s issue #485, the same failure caught there first, and its ownrequired-version. Read, not assumed:uv lock --checkunder uv 0.11.31 resolves this repository’s existing lock (schema revision 3) without complaint, so the floor drops with no re-lock needed. The comment above the key is rewritten to name the actual ceiling — a number nobody in this repository controls, that moves only when a human bumps that Dockerfile — in place of the uv-lock pre-commit hook’srev, which was never it: that rev moves on pre-commit.ci’s own schedule and happened to read close to the old floor by coincidence.release.ymlgains adocumentedjob, closing issue #154: on a tag push it pollshttps://bitcoin-core-rpc.readthedocs.io/en/<tag>/at a fixed interval, and fails the run if that URL is never served. Read the Docs activates and builds a release tag from an automation rule of its own, and a build that never starts fails nothing else in the pipeline — a site answering 200 keeps serving the last build that succeeded for as long as it takes somebody to notice, which is the failure mode btclib’s own RELEASING.md records in its aggravated form: a webhook refusing every delivery with a 400 for three years while the site kept answering 200. The job asks the rendered URL rather than the v3 API: Read the Docs throttles an unauthenticated caller and blocks cloud-provider addresses without a token, so asking the API from a runner would mean storing a credential for a check that gates nothing, where/en/<tag>/is CDN-served, 404s for a version with no successful build, and is the URL a reader follows anyway. It runs afterversion-checkand nothing depends on it: a late documentation build is not a reason to withhold a wheel, and the fix for a red run here is a build on Read the Docs’ side, never a moved tag. Landed after issue #149’sgithub-releasefix, deliberately, so that a job able to fail the run could not import a new way to lose the GitHub release before that fix existed. Ported from btclib’s owndocumentedjob inrelease.yml;release.ymlcannot run on a pull request, so the polling loop was run verbatim from the shell against a tag Read the Docs did build (v2026.8.20, served) and one it did not (v2026.8.12, 404) before landing, rather than trusted on the strength of the diff alone.A re-run of a finished TestPyPI rehearsal now gets a version of its own, closing issue #157: the suffix was
.dev<run number>, and a re-run keepsgithub.run_numberand only raisesgithub.run_attempt, so re-running a rehearsal rebuilt the same version and TestPyPI refused the upload. RELEASING.md’s answer had been a warning to dispatch a fresh run instead of re-running one — a rule to remember at the moment a failed rehearsal is least likely to leave room for it. Fixed the way btclib-secp256k1’srelease.ymlalready was (repository btclib-secp256k1’s issue #32, PR #58): the suffix is.dev<run*100+attempt>, the multiplier keeping every attempt of every run both unique and ordered — attempt 2 of run 7 still sorts after every attempt of run 6 — and the step refuses outright past the two digits it reserves for the attempt rather than wrapping into another run’s range. RELEASING.md’s warning goes with it, the rule it stated no longer being true.release.yml’sbuildjob now runstwine check --strict,check-wheel-contentsandpyroma --min 10on the files it is about to publish (issue #155). Those three already gated every pull request, intest.yml’sdistjob, butbuildrebuilds the distribution files from the same tree rather than reusingdist’s artifact, and ran none of the three on its own copy — so a tag that passed every check still published files nobody had checked, “usually” identical to the checked ones being the whole of the guarantee. The three new steps run afterUpload the distribution filesrather than before it: each check installs a dependency of its own, and the existing comment on that upload step already establishes that nothing gets installed beforedist/is frozen for the publish jobs to download, a boundary these steps keep rather than move.publish-testpypiandpublish-pypialready hadbuildinneeds, so a failing check now stops both. Left undone, and deliberately, the same way repository btclib’s issue #1154, PR #1164 left it:release.ymlstill builds the distribution files twice, once in thetestjob it calls and once inbuild— reaching the shape with no second build, asbtclib-secp256k1’srelease.ymlalready has, needstest.yml’s artifact published instead of rebuilt, a larger change than this one.test.yml’s comment on thedistjob and RELEASING.md’s rehearsal description, both of which read as thoughdistwere the only place checking a distribution, are corrected alongside the workflow.Coverage is now measured on a bare
uv run pytest, and names itssource, closing issue #6 of the organization’s.githubrepository (the issue lives in another repository, so no keyword here closes it).--covmoves intoaddopts, right after-raand not last: it takes an optional value, so as the final token it would be handed whatever the command line goes on to say —pytest tests/transport_test.pybecomes--cov=tests/transport_test.py, the whole suite runs regardless, and the run reports againstfail_underhaving measured a directoryomitexcludes rather than the one path asked for. Moving it needs the suite to stop failing on the tree’s coverage when it is asked for less than the tree:tests/conftest.pygains apytest_configurehook, ported from btclib’s owncoverage_fail_under, that drops the threshold to zero for a run selecting paths,-kor-m, and leaves--lf,--deselectand an early-xat the full gate, those being an iteration’s flags and not a selection. The threshold is written toconfig.known_args_namespacerather thanconfig.option: pytest-cov reads only the former, filling it beforepytest_configureruns, so the latter changes nothing and fails silently. The hook is itself code the 100% gate now covers, sotests/conftest_test.pycarries the tests btclib has for it that apply here — this tree has no hypothesis profile and no golden-file fixture to port alongside it.[tool.coverage.run]gainssource = ["bitcoin_core_rpc", "tests"]: unnamed,--covmeasures whatever the run imports, which on a pristine checkout can differ from what thecoveragejob intest.ymlmeasures, and a local gate stricter than the one it reproduces is the worse of the two directions. That job’s own--cov=bitcoin_core_rpc --cov=testsis now the redundant copy —addoptsandsourcealready say what and how — so the step drops back topytest --cov-report term-missing:skip-covered, the same measurement as the bare command above it in CONTRIBUTING.md.[tool.mypy]gainswarn_unreachable = true, which the three siblings already set; unlike the other optional error codes this one is not a ratchet with nothing to catch — the survey found six sites, each a runtimeisinstanceorcallableguard against a caller that skips type checking, on a value whose annotation already promises what the guard re-checks. Each site keeps its guard and takes its own# type: ignore[unreachable], naming the code, rather than a blanket exemption that would silence the check everywhere the finding is legitimate too.addoptscarrying--covalso instruments every cell oftest.yml’ssuitematrix,macos.yml,windows-arm.ymlandlatest.ymlunless told otherwise, which broke pypy3.11 outright rather than merely costing wall clock: coverage.py has no C extension built for pypy, so it falls back to a pure-Python tracer there, and that tracer does not surviverpc_smoke_test.pyexercisingmain()’s severalsys.exit()paths in-process — it warns “Trace function changed” partway through the run and stops recording, so a suite that passed every test reported 71% against the 100% gate. Each of those four workflows’Run pyteststep now passes--no-cov, restoring the invarianttest.yml’s owncoveragejob comment already stated: coverage is measured once, on one interpreter, in that job alone. A fifth bare-pytestcall site missed that pass:.github/mutation/bitcoin_core_rpc.toml’stest-command, which shells out topython -m pytestonce per mutant rather than through a workflow file, and whose own comment already argued coverage does not belong there — a mutant that removes a statement fails the ratchet for a reason unrelated to whether the suite noticed the mutation. It now carries--no-covtoo.branch = trueandexclude_alsoare deliberately not part of this change — issue #7 of the same repository is that, across every repository in the organization, and the two are meant to be read together.Coverage measures branches and not statements alone,
branch = truein[tool.coverage.run], which the organization standard asks of every repository under it (btclib-org/.github#7). Statement coverage calls a half-testediffully covered: the line ran, one of its two ways out never did, and in a module whose bodies are guard clauses — the credential already supplied, the cookie file already read, the reply that is not an error — the way nobody took is the refusal nobody exercised. What the sweep found is that the client itself was already whole under the stricter measure, and that the one partial branch was in the suite:tests/standalone_test.py’s scan for the docstringsautomodulerenders never met a module-level definition without one, every definition of the client carrying a docstring. Closed by a test rather than by apragma: no cover, since the branch is reachable and a scan nothing has ever made report is a check whose passing says nothing — the scan is now_documented_namesand is read against a source written for it that holds a definition of each kind, documented and bare, assignment with a string literal after it and assignment without. No pragma was added anywhere.pydoclintjoins the lint gate, overbitcoin_core_rpc/and configured in[tool.pydoclint](btclib-org/.github#7): ruff’sDrules ask that a docstring exists, and this asks that the arguments and the return it documents are the ones the signature declares — the half that goes wrong in silence when a signature changes and the prose above it does not. It reports nothing today, and one of its options is why:skip-checking-short-docstringsstays at its default, where btclib-secp256k1 turns it off, so a docstring carrying no section at all is skipped whole. Turning it off here would ask for anArgs:section in every docstring of a file whose paragraphs already name their parameters in sentences, which is the same fact written twice in the one file a consumer vendors and reads end to end. Issue #172 holds that question open, with the output the stricter setting produces and the command that reproduces it; the hook meanwhile gates the case the sections exist for, a docstring that does list arguments and lists them wrong.published.ymlcarried a comment for anif:it never had, andsetup-uvwas two minor versions behind both siblings (issue #160, divergences 19 and 21 of btclib#1152). “The second half filters theworkflow_runtrigger” sat aboveif: ${{ !github.event.pull_request.draft }}, a single condition with no second half and noworkflow_runtrigger anywhere in the file — onlyworkflow_call.git log -Son the phrase finds one commit, #76, which introduced both theworkflow_calltrigger and this sentence together: the sentence never matched theif:it sat above, and the scenario it described — aworkflow_dispatchrehearsal ofrelease.ymlreaching this workflow — cannot happen either,publishedbeing called only vianeeds: publish-pypi, which a rehearsal always skips. Aspirational from the moment it was written, not a leftover from a removed trigger. Deleted rather than corrected: there is no accurate version of “the second half” to write, there being only one half.setup-uvwas pinned tov10.0.0(ae62891f) throughout, wherebtclibandbtclib-secp256k1are onv10.0.1(20cfd1bf, read from a sibling and verified against the tag rather than re-resolved); brought level across all 14 pins.RELEASING.md’s “only
github-releasefailed” bullet names both of a release’s artifacts, not one (btclib#1150). “create the release by hand from thedistartifact of the run” left outattestation, which the job downloads besidedistbefore it writes the release; a release built fromdistalone would carry the wheel and the sdist with no signed statement beside them, and “Verify the provenance of an asset”, earlier in the same file, already assumes one is there to--bundleagainst. This repository attaches no bill of materials — RELEASING.md’s own “No CycloneDX bill of materials is among them, on purpose” paragraph has the reasoning — so the fix is two artifacts,distandattestation, and a pointer at theskippedcase’s own script below rather than a second copy of it, the two bullets already differing only in whethergh run rerun --failedis worth trying first.docs.yml’s unresolved-link grep now carries both known shapes, one of two companion fixes for btclib-org/btclib#1157, which btclib-org/btclib’s own PR closes. MyST renders a link itsRootFileLinkstransform cannot resolve ashref="#<target>"verbatim, so what the grep matches depends on how the target was written —href="#\./for./CONTRIBUTING.md, which every internal link in this repository’s own root files uses, with no exception, andhref="#[A-Za-z0-9_.-]*\.md"for a bareSECURITY.mdwith no./, the shapebtclib-secp256k1’s root files currently carry alongside its own./links — not a settled convention there, its present drift. This repository ran only the first grep, which is blind to a bare-shape link breaking here the way one already can there. Built the documentation with a deliberately unresolved link of each shape and confirmed the added grep reports it before removing it;sphinx-build -W --keep-goingpasses on both unchanged, since neither is the broken reference-Walready catches — a link myst resolves happily and is dead anyway is the whole reason this second check exists.REPOSITORY.md records that this repository pins its own
default_workflow_permissionsrather than inheriting the organization default (issue #165). The REST API has no field distinguishing an override from an inheritance and no value that clears one — checked against the current documentation rather than assumed — so a future move of the organization default will not reach this repository unless whoever moves it also moves this one by hand, with the command the new section gives.btclibandbtclib-secp256k1are recorded as untested rather than known good: both were already at the target value before the organization default moved, so neither could have been observed following it.Every
run:step that can land on a Windows runner declaresshell: bash(btclib-org/btclib#1148).suite-latestinlatest.yml,suite-windows-arminwindows-arm.yml, andsuiteintest.ymleach carry awindows-latestorwindows-11-armmatrix cell, where the default shell is PowerShell and"$GITHUB_ENV"and the rest of the POSIX spelling are literals rather than variables — the same defect btclib-org/btclib#1141 shipped in btclib’s ownpublished.yml, and here in the two remaining scheduled workflows and intest.yml’s own merge-gate job. All four affected steps are one plain command each and run identically under either shell today, so nothing shipped broken.latest.ymlandwindows-arm.ymlrun on neither a pull request nor a push tomain, so the hazard is the next conditional, substitution or loop landing on one of them, first executed on a schedule or a tag;test.yml’ssuitejob runs its Windows cell on every pull request, where a shell-less step still executes PowerShell rather than the intended bash, going green on the wrong interpreter’s syntax rather than red on the right one.Enumerated by parsing every workflow file’s
jobs.*.strategy.matrix(osandinclude) andruns-onwith PyYAML, and flagging arun:step whose job’s runner set containswindowsand which declares noshell:of its own and inherits none from a job-leveldefaults.run—actionlint, the pinned linter in the pre-commit gate, reports nothing for this shape, measured on a minimal workflow with a POSIXforloop and noshell:.btclibandbtclib-secp256k1had the samelatest.yml/windows.ymldefect in their own copies, fixed in a pull request of their own;btclib-secp256k1’stest.ymlcarried two further instances of its own.release.ymlno longer has abuildjob, andtest.yml’sdistjob is now the one build a release publishes (issue btclib-org/ btclib#1166, filed from btclib’s own copy of the same seam). The entry above (issue #155) closed the correctness hole — both builds checked — and left the seam standing, in the terms it used itself:diststill built a copy that was checked and thrown away,buildstill built a second, separate copy — with noSOURCE_DATE_EPOCHand no sdist normalization — checked a second time and published.distnow pinsSOURCE_DATE_EPOCHfrom the commit, normalizes the sdist, uploads thedistartifact before installing anything, and only then runs the three checks and the wheel smoke test;release.yml’spublish-testpypiandpublish-pypijobs download that artifact unchanged, andattestsigns the digests the upload already fixed. A rehearsal’s version suffix moves too, from an inline step inbuildto a.github/actions/dev-versioncomposite actiondistruns before its ownuv build, computed once inversion-checkand passed down as aworkflow_callinput so every job the release dispatches during one run agrees on it — carrying over therun*100+attemptscheme issue #157 above already fixed here. Nocheck-newest-bindings-style second smoke test, unlike btclib’s own fix: this package declares no runtime dependency to ask that question about.btclib-secp256k1has had the one-build shape from the start.dist’s ownSetup uvstep carriesbuild’s caching-off forward the same way, through a newdisable-dist-cacheinput rather than a flatenable-cache: false: a cache entry written by one run is read by another through GitHub’s branch/PR scoping, so the build whose output actually reaches an index must fetch its dependencies afresh, where an ordinary pull request’s build is thrown away and caching it costs nothing. Found and fixed the same way as btclib#1180 (dd198d83), the twin of this pull request, whose first review round caught the same regression.A
local-link-prefixhook refuses a local markdown link destination that does not begin with./, anddocs.yml’s unresolved-link check is one grep again (btclib-org/.github#20). What that grep can match is decided by how the link was written, myst rendering the destination of a linkdocs/source/conf.py’sRootFileLinkstransform cannot resolve verbatim:./SECURITY.mdbreaking giveshref="#./SECURITY.md", a bareSECURITY.mdbreaking giveshref="#SECURITY.md". The links here were already written the first way, and nothing said they had to be — the hook is what says so, and with it the second grep, added for the bare shapebtclib-secp256k1was writing rather than for anything this tree carried, has nothing left it can catch. That grep was no superset of the first even so: a character class of name characters stops at the#of an anchor, sohref="#README.md#build"is a shape both patterns pass over, which is the argument for moving the rule upstream rather than writing a third pattern.The rule is the prefix and not the extension, and the table in btclib-org/btclib#1175 is what decides that:
DOES_NOT_EXIST.txt,sub/DOES_NOT_EXIST.md,DOES_NOT_EXISTand../DOES_NOT_EXIST.mdeach reach MyST’s fallback and each is missed by the union of both greps, so an.md-scoped rule would leave all four writable.../is the row with nothing downstream behind it at all: the transform declines a target normalizing above the repository root, so that shape reaches the fallback by design and rendershref="#../X.md", which the surviving grep does not match and should not be widened to reach — the hook is the only place it can be caught. A link reference definition,[label]: page.md, renders the same fallback and carries no(, so the pattern has a second branch for it, anchored at the start of a line so that a reference use followed by a colon is not mistaken for one.Measured here as well, by building this documentation with an unresolvable link written each way:
./page.md,./page.md#anchor,./page.txt,./sub/page.mdand an extensionless./pageeach render#./followed by the destination, so the one grep sees every one; the same destinations without the./render the destination alone, which nothing catches without also matching the autodoc anchors these pages carry.README.md’s licence badge was cited here as the live case for that, and the citation was wrong — recorded rather than swapped out, because the rule this entry lands was argued from a destination the rule did not reach.
[](./LICENSE)is a badge: its destination sits behind a nested image, and a link text written[^]]*stops at the]that closes the alt text, so the first version of the pattern checked the imagesrcand never the badge’s own href. Every badge destination in this repository was unchecked, and neitherdocs.ymlgrep, the removed one or the surviving one, would have seen one lose its./either. The link text is now(?:[^]]|\]\([^)]*\))*— a character that is not], or a whole](…)group — which steps over the image and reaches what is behind it, and by backtracking still checks thesrcon the way. Measured on the built html: a badge href renders exactly what a plain href renders,#./NOPE.mdwith the prefix and#NOPE.md,#NOPE.txt,#sub/NOPE.md,#NOPE,#../NOPE.mdwithout it. The corrected hook reports nothing against this tree, badges included.RELEASING.md’s link torelease.ymlwas the one destination in this repository’s markdown that did not begin./, and now does.uv run --locked --only-group lint pre-commit run local-link-prefix --all-filesre-derives the whole of it.docs/source/conf.py’sRootFileLinkscomment lists every root markdown file the toctree pulls a page from, in place of a count (issue #189). The sentence introducing the shims said how many pages carried them and named some of the pages but not all of them, so the number and the list disagreed with each other as well as with the toctree indocs/source/index.rst: COMPARISON and REVIEWING were missing from both. btclib’s ownconf.pycarries the same paragraph with no count at all, naming its files instead, and that is the shape adopted here too — a page missing from the list is visible on the next read, where a stale number was not.REVIEWING.mdasks a review to run what a diff decides with. A regex, a grep, a pattern in a hook, a script or a query decides an outcome by matching or computing, and a review of one executes it rather than reading it — against the shapes the diff’s own prose claims to cover, and against the shapes the tree actually holds. A claim the prose makes about the tree takes the same treatment, “every link here is already./-prefixed” being onegit grep’s worth of evidence and the reason a change is offered as safe.What earned the rule is this repository’s own review of the
local-link-prefixpygrep hook, pull request #192, the same pattern having been proposed across the organization byte for byte. The leading clause\[[^]]*\]\(cannot cross the]that closes an image’s alt text, so on the badge shape[](./href)it examines the image source and never the destination the link itself carries — which is the destination that hook’s own comment cites as its motivating case, left free to lose its./unreported. That review ran the clause againstREADME.md’s licence badge and quoted what it printed; the sibling reviews reasoned about the same pattern and reported it correct. The rule generalizes what the run did rather than leaving it to whoever thinks of it.The section says what it is not,
claude-review.yml’s prompt telling a review here not to re-run the gates: those run what the rest of the tree already exercises, where a pattern a diff adds has been run against nothing until a review runs it. The prompt is unchanged — it namesREVIEWING.mdrather than restating it, so the standard moves without the workflow being edited.REVIEWING.mdsays when a gate already run is relied on rather than repeated, andCLAUDE.mdsays which checks a pull request is gated on. The condition is the sha and the record: where the gates have been run on this very commit — the required checks running beside a review, or an author handing over a branch they gated and stated the result of — the review relies on that run and names whose it is, and where there is no such run it runs them itself. A run on another tree is not a run on this one, so a rebase voids it, the branch having been gated before the tree moved under the gate.What earned it is
claude-review.yml’s prompt, whose reason for not running the gates was that a second run of them would cost a runner slot. That is an argument about price, and a reviewer reading it generalizes it into “running costs, so do not run” — which is the disposition that acked thelocal-link-prefixpattern of the entry above without running it. The prompt now carries the instruction and points atCLAUDE.mdfor why the reliance is sound, which is also what keeps its hunk small: a pull request editing that file cannot be reviewed at all, the action refusing to run under a workflow differing from the default branch’s copy (btclib-org/.github#58).It leaves the entry above where it was. The gates exercise what the tree already held, so what a diff decides with has been run by nothing whether they ran or not, and a review runs that either way.
latest.yml’s pre-commit cache is keyed on the runner image and the interpreter rather than on the config hash alone (btclib-org/.github#25). What that cache holds is virtualenvs, so the config hash alone survives anubuntu-latestrotation and restores environments built on the previous image — not a graceful miss but a hit, surfacing as whichever hook touches the broken environment first.lint-latestis the job with the most to lose from that: it runs to report what a new release of a dependency breaks, so a failure the cache caused arrives looking exactly like the one the workflow exists to find, and is chased in the tree. TheIdentify the runner imagestep readingImageOSfrom a shell, the key overrunner.os, that output andhashFilesof.pre-commit-config.yamland.python-version, andrestore-keysstill naming the image so a partial restore cannot cross an image boundary, are alllint.yml’s, which has carried them since before this was noticed. The comment above the bare key claimed it was keyed “as inlint.yml”, which was false where it stood and offered that file as the justification; what replaces it is this job’s own reason for the key, withgit grep -n 'pre-commit-\${{' -- .github/workflows/beside it as the command that re-reads both keys together.REPOSITORY.mdstops naming which sibling repositories are untested against the organization workflow-permissions default (btclib-org/.github#23). That paragraph was a roster of other repositories’ settings kept in this one, and it had already gone incomplete —btclib-benchmarkswas missing from it. Each sibling now records its own status in its ownREPOSITORY.md, which is where a reader of that repository meets the question; what stays here is the reason they share, that each already heldreadwhen the organization default moved on 21 August 2026 and so could not be observed following it. This repository’s own entry is unchanged: it is pinned, and the read-back beside it still answersread..readthedocs.yaml’s header stops describing a sibling’s tree, and starts saying what the file does not decide (btclib-org/.github#26). The line aboutreadthedocs.ymlbeing honored but deprecated arrived with a clause aboutinclude *.ymldragging the file into the sdist beside the Jekyll configuration. Nothing serves a site from this root, there is no such configuration to sit beside, andMANIFEST.innamesinclude *.yamlwith its own reason written beside it — so the clause was true of btclib and of nothing here. It is now attributed to btclib rather than deleted, a clause named as another repository’s being harder to copy back than a gap.What is added is the fact this tree cannot answer anywhere: which versions run this file.
latest,stableand each release tag are versions on read the docs’ side, and an automation rule is what activates a new tag there — so a release that keeps a permanent URL of its own and one that does not differ in a setting rather than in the tree. btclib’sREPOSITORY.mdcarries a “Read the Docs” section recording exactly those settings; this repository’s has none, and the comment says so and names btclib-org/.github#26, which carries the commands that read the state back. Writing that section here is a change toREPOSITORY.mdand left to one.documented’s wait can no longer spend the whole job timeout, and names what it last saw when it gives up (btclib-org/.github#18). The loop’s worst case andtimeout-minuteswere the same budget — an attempt costs itscurl --max-timeplus thesleep, and that sum taken over every attempt reached the job’s own limit exactly — so a response slow enough to spend its--max-timewithout ever answering ran the loop into that wall, and the runner killed the job before the::error::line, the one line the job exists to print. A red run then said only that it had exceeded its maximum execution time, which is a fact about this workflow rather than about the documentation. The ordinary case never came near it, Read the Docs answering a 404 in well under a second while a build has not started; it is the slow or unreachable case that lost the pointer, and that is the case a reader most needs it in. The loop now leaves a margin it cannot spend into, and the last attempt no longer sleeps before giving up.The same wait stops discarding what it already knows.
curl -sfthrew the response away, so every attempt printed “is not served yet” and the final::error::named only the URL and the builds page. A status code and a transport failure are now told apart, each attempt names which of the two it got, and the verdict readslast seen: HTTP 404orlast seen: curl exit 28 (no response within 10s)before pointing at the builds page — named as where the reason is written, this job saying only that there is one. What the wait still cannot do is tell a build that has not started from one that never will:/en/<tag>/answers 404 for both, and the v3 API that separates them stays ruled out for the reason the job’s own comment already gives. Dropping-fneeded a second fix the diff does not show: under the shell’s defaultset -ea barestatus=$(curl ...)aborts the step on curl’s own non-zero exit, so a hung server would have ended the run on the first slow response — precisely the case the margin exists for — and the assignment is the condition of anifinstead.release.ymlcannot run on a pull request, so the step’s literal script was extracted and driven against a mockedcurlandsleep: immediate success, success on a later attempt, a fast failure every attempt, and every attempt hanging to--max-timeeach reach the intended verdict, and the last attempt is observed not to sleep. Ported from btclib’s own fix, which found theset -edefect the same way..yamllint.yamlsays what is true of every repository carrying it, and re-derives the rest (btclib-org/.github#40). The file is copied byte-for-byte across the organization, and its comment carried a survey run once, inbtclib: how many findings each disabled rule produced, what share of the action pins sat at or past 80 columns, what a limit of 80 would report and how much of that was pins, the width of the longest pin, how many yaml files the tree held, and two over-length shell lines in arelease.yml. Re-measured with the pinned yamllint against this tree and against the others sharing the file, not one of those figures still held anywhere —btclibincluded, where the survey was taken — and therelease.ymlclaim names a file two of the repositories carrying the comment have never had. The survey is replaced by the commands that answer for whichever tree is asking, the default set at the configured width and what a limit of 80 would cost, beside the reasons that do not vary between repositories: dependabot writes one space before a pin’s trailing tag,on:is the boolean true in yaml 1.1, and a 40-character SHA with its tag has spent most of an 80-column line before the action is named. This repository’s.taplo.tomlalready differs from the shared one and carries none of this, so it is untouched.CLAUDE.md’s primary-checkout paragraph names the read that cannot go stale (btclib-org/.github#255). It said reading the checkout was fine and so wasgit fetch, without sayinggit fetchmovesrefs/remotes/origin/mainand leaves the work tree where it was, so agrepor aReadagainst the checkout answered for whenever it was last brought forward. The paragraph now namesgit show origin/main:<path>as the read that does not go stale, and gives the fast-forward that brings a clean checkout forward without working in it.The shipped module’s exception-naming comment no longer names
btclib(btclib-org/.github#81). This package is zero-dependency and vendorable, sobitcoin_core_rpc/__init__.pyis read by people who copied only that file and have never heard ofbtclib; the comment explaining theBtcRpcprefix stated a collision withbtclib’s ownBTClibValueErrorthat is not that reader’s. It now gives the reason a distinct prefix is worth having without naming the dependent: a consumer already carrying an error hierarchy of its own should not have to tell twoValueErrorsubclasses apart. Ano-downstream-name-in-packagepygrep hook holds the shipped module to it, the organization spelling and the copyright line’s “The btclib developers” excepted.CLAUDE.md’s worktree recipe named the worktree after the issue alone,wt<issue>(btclib-org/.github#292). A worktree’s administrative directory lives in the.gitof the repositorygit worktree addwas run from, one per repository, so two repositories cannot collide there; what the recipe left uncovered was a same-repository collision, between two worktrees of different work sharing a generic basename, and a path collision across repositories, since the workers of one session share one scratchpad directory and a session carrying one issue into several repositories computed the same target path for each. The recipe now names the worktreewt-<tracker>-<issue>-<repo>-<role>, most general part first:trackerbecause an issue number is unique only within one tracker,issueagainst the same-repository collision,repoagainst the cross-repository path collision, androleagainst a coder and its reviewer holding a worktree at once.README.mdends with the line naming who supports the work. The organization standard states the line as tier 1’s, for the reasonSECURITY.mdis: the archive leaves github.com, and a reader who has it and not the repository meets the project with no organization beside it (btclib-org/.github#98).no-downstream-name-in-packagematchesBtclibandBtclibError, not onlybtclib(btclib-org/.github#299). The hook took noargs, so it matched case-sensitively and let a downstream name that opens a sentence or sits in a class-name-style compound through where the same name mid-sentence would not;btclib-secp256k1’s copy of the same hook already carriesargs: ["-i"], so the two enforced different rules under one name.args: ["-i"]here now closes that gap.REPOSITORY.mdandRELEASING.mdstop citingbtclib’s own state to justify this repository’s choices (btclib-org/.github#81). A required-check name, a GitHub Actions concurrency count, an--adminmerge precedent and a CycloneDX bill-of-materials decision each reasoned from whatbtclibdoes or measures, standing in for reasoning that holds on its own; the bill-of-materials paragraph also sent its reader tobtclib’s ownRELEASING.mdfor “the full reasoning” instead of stating it. Each now reasons from this repository’s own configuration alone. Theno-downstream-name-in-packagehook does not reach either file, only the package directory that ships, which is why this half needed a reading rather than a red hook; with it, this repository carries the last boxbtclib-org/.github#81had open.bitcoin_core_rpc/moves undersrc/bitcoin_core_rpc/(issue btclib-org/.github#313): section 2 of the organization’s standard puts every repository’s package undersrc/, and this repository had not yet converged.[tool.uv.build-backend] module-root = ""is deleted rather than changed: that key existed only to override the backend’s own default, which issrc/. Every path this repository’s own configuration and prose named the package by moves with it –.pre-commit-config.yaml’s twofiles:patterns and oneexclude:pattern, the mypy invocation’s arguments,.github/mutation/bitcoin_core_rpc.toml’smodule-path,[tool.check-wheel-contents] package, and the citations inCLAUDE.md,CONTRIBUTING.md,README.mdand.vscode/settings.json.[tool.coverage.run] sourceand everyimport bitcoin_core_rpckeep naming the package unchanged, since an import name is not a path..pre-commit-config.yaml’styposhook converges tobtclib-org/.github’srepo: localshape (issue btclib-org/.github#399).autoupdatewalks everyrepo:entry exceptlocalandmeta, so thecrate-ci/typosmirror kept receiving the movingv1alias thatpinned-revthen refused on every run;additional_dependencies: [typos==1.49.0]now carries the pin in place ofrev:, and the comment explaining why isbtclib-org/.github’s own, taken across unchanged. Two comments this repository carries on its own claimed every pre-commit hook’s revision is kept current either by pre-commit.ci or, formypyalone, bydeps-latest.yml’s own upgrade – neither reachestyposany longer, and.github/dependabot.ymlanddeps-latest.ymlare corrected in the same diff..gitignoreno longer namesdocs/_build/(issue btclib-org/.github#411).build/, already present above it, ignoresdocs/build/html, the directorydocs.yml’ssphinx-buildstep writes to;docs/_build/matched no path in this tree and decided nothing.README.md’s badge row drops the link to the repository (issue btclib-org/.github#381). Section 2 of the organization standard refuses it now: the badge renders the repository’s name because the URL says so, and the row is an audit, so the item that measures nothing is the one that does not belong in it.[project.urls]’srepositorykey already carries the same link for the reader who meets this file as an index’s long description or as an unpacked sdist’s README.pyproject.toml’s comments are corrected against measurement (closes #226) (closes #242) (closes #230) (closes #239) (closes #247). Thelicense-filescomment quotes the noticeLICENSEcarries, “The btclib developers”, rather than a holder no notice in the tree names. Thewarn_unreachablecomment drops a count of sibling repositories that no longer describes the organization. Theskip-checking-short-docstringscomment states btclib-secp256k1’s current reason for turning the option off, an Args/Returns form on almost every docstring there, rather than its retired one. The docstring-ignorecomment adds the fact that thepep257convention leavesundocumented-magic-methodandundocumented-public-initenabled, so naming neither entry is what asks a docstring at every such site.[tool.check-sdist] source-excludegains.coverage,.coverage.*andcoverage.xml(closes #264). coverage.py writes its data file beside the tree it measured rather than under a cache directory, socd tests && uv run pytest .lefttests/.coveragebehind, gitignored and swept into the sdist by thetests/**include; the three new entries are unanchored, the same shape as the three cache entries already there.[tool.uv] required-versionrises to the ceiling Dependabot’s own uv-ecosystem updater allows (closes #250), and the comment beside it points at the organization standard’s own argument and its command to re-derive the ceiling rather than restating either.Ruff’s
unspecified-encoding(PLW1514) is now selected (closes #225). It is a preview rule, unreached by thePLfamily already selected, and asks thatopen,read_textandwrite_textname an encoding rather than take the locale’s – the defectdecoded-subprocess-encodingalready refuses one layer out, at a subprocess call. No source in the tree fails it, which makes the change a ratchet rather than a cleanup.CONTRIBUTING.mdandRELEASING.mdqualify a bare#1166asbtclib-org/btclib#1166(closes #221), the three sites citing btclib’s own issue rather than one of this repository’s.This release’s
check-sdist/pyromaentry overstates whatpyroma’s non-isolated build call checks (closes #238).pyroma’swheel_metadata()calls the backend’s PEP 517 hook directly throughbuild.util.project_wheel_metadata(..., isolated=False), which never reads[build-system]’srequires, and falls back to an isolated build only when that call raises aBuildExceptionorBuildBackendException– in this tree’s own hook environment the backend-import failure that entry’s own next sentence already quotes, not arequirescheck. That entry is not rewritten, being already landed; this corrects the record beside it.Every
cron:here fires at the instant section 10’s calendar gives that workflow and this repository (issue btclib-org/.github#480). The calendar’s rows sit in the order of what they ask about – the depth a tree’s suite is tested to, what it does against software it does not ship, what it depends on and what it publishes, its platforms, its own health and its security – so a workflow’s day and hour follow its family. The minute is this repository’s row in that section’s second table, and it does not move.dependabot.ymlis untouched, section 10 keepingdeps-lateston the day before Dependabot’s own.README.md’s badge block is section 2’s three groups (issue btclib-org/.github#480). What the software is opens with the release identity as a pair, the PyPI version besidegithub/v/release, which is how section 2 says those two are read: where they disagree, a release reached the forge and not the index. The licence andwheelsit in that group, this tree having aLICENSEand publishing a wheel, and section 2 derives the row from such properties rather than curating it. The CI group is the gates in the order a commit meets them, Read the Docs among them because it answerspassing,failingorunknownas the workflow badges around it do, and then the sentinels in the calendar’s order – so the badge order is the calendar order over that subset, and the two move together or not at all. The Scorecard badge is the OpenSSF group, and section 2 reads it as the last of the sentinels without being among them,scorecardbeing the calendar’s last row.The calendar sentence points at section 10 and states no number (closes #276). It gave the calendar six repositories, where section 10’s second table names more. Section 9 of the organization standard asks prose never to state how many of anything a file holds: a total is exact on the day it is written and stays grammatical after it stops being true. The clause beside it already says the calendar is not this file’s to restate, and a count of what it covers restates that table. The phrase is now the one
btclib-secp256k1’s copy carries.pypi-install.ymlsays why itspull_requesttypes leave outclosed(closes #273). Section 10 of the organization standard namesclosedamong the types, and its reason is the merge landing in the pull request’s own concurrency group to cancel the run still holding it. This workflow’s group carriescancel-in-progress: false, so a closed event cancels nothing whichever group it lands in; and the job’s condition declines a draft and says nothing about a close, so the run such an event added would install from the index on every cell of the matrix, for a pull request nobody is on. The types are unchanged and the divergence states its reason where it is, which is what section 9 asks of a deviation.The downloads badge’s link target is the plural
pepy.tech/projects/bitcoin-core-rpc(closes #281). The singular answers308, matching the reason section 2 of the organization standard already gives for the Read the Docs spelling it fixed the same way: a link that works only because a third party keeps forwarding it is a dependency nobody here recorded. The badge’s image,static.pepy.tech/badge/bitcoin-core-rpc, is section 2’s own spelling and is unchanged; only the href moves.claude-review.ymltakesclosedamong itspull_requesttypes, and declines it at step level rather than on the job (closes #278). Section 10 of the organization standard givesclosedthe reason of landing a merge in the pull request’s own concurrency group to cancel a run still holding it, and this job’s group is job-level rather than workflow-level: the job itself, not only the workflow run, has to be scheduled on a closed event to enter that group, so itsif:stays open to the event rather than declining it the waylint.yml’s workflow-level group lets its job do.github.event.action != 'closed'sits on the review step and on each step that reports a verdict instead, so closing or merging a pull request cancels the run still holdingclaude-review-<number>and concludes green, without starting a review of a pull request nobody is looking at any more.claude-review.yml’s credential check declines a closed event (closes #283). The review job is scheduled on a close so that it entersclaude-review-<number>and cancels the run still holding it, and the review step – the only other step of that job readingCLAUDE_CODE_OAUTH_TOKEN– declines that event, as does each step that reports a verdict. So on a close the check asserted its invariant over a credential the run never reaches, where the empty secret that fails it is a red check on a pull request already merged or closed. It carries the samegithub.event.action != 'closed'its neighbours do, and no!cancelled()beside it: anif:naming no status check function is evaluated assuccess() && ..., which is what a precondition wants, where the step reporting a review that never ran has to fire after a failure and says so where it sits.claude-review.ymlchecks the repository out only where a step reads it (closes #285). The review job is scheduled on a close so that it entersclaude-review-<number>and cancels the run still holding it, and every other step of it declines that event: the review step, which is the only one that opens the tree, and the steps that report a verdict, which read the API with the repository from the event context and the action’s own execution file. SoCheckout codefetched the repository onto a runner for a job with nothing left to look at it, against the ceiling on concurrent jobsREPOSITORY.md’s Plan-gated settings measures. It carries the samegithub.event.action != 'closed'its neighbours do: what contests a job-level concurrency group is the job being scheduled, which its ownif:still is, so the cancellation is unchanged and the job concludes green with every step skipped.normalize_sdist.pystates what the step decides (closes #289). Its docstring presented the rewrite as a step the backend had made unnecessary, and sent the reader to btclib’s copy of the file for an argument that copy does not make.uv_buildwrites anmtimeof0into every member of the sdist and ignoresSOURCE_DATE_EPOCH, so this script moves every one of them to the commit’s second and the archive’s sha256 with it: the digest the attestation vouches for is the step’s output, and dropping it as inert publishes other bytes. Section 12 of the organization standard states the property a published sdist owes and leaves the mechanism to the tree, so the reason now sits beside the command inRELEASING.md, beside the step intest.ymland beside the block inCONTRIBUTING.md.[project.urls] homepagenames this tree’s own documentation site (issue btclib-org/.github#533). A releasing tree provides documentation, and its home is that documentation rather than a sibling’s project page: the field readhttps://btclib.org, whichdocumentationalready pointed past, and now carries the identicalhttps://bitcoin-core-rpc.readthedocs.io/.documentationstays, an index showing the two fields as one link being cheaper than the field tools read for that purpose specifically.docs.ymlchecks the built api page against__all__(closes #299).automoduledrops a data member with no attribute docstring in the module being documented, silently –-n -Wsees no malformed cross-reference and no failure, and a facade re-exporting a constant through__getattr__is exactly that shape, measured on the tree that introduced it: 20 of 30 names on the built page, every function, class and exception, and no constant or type alias..github/scripts/ check_api_page.pyreads the page the job already built and refuses one missing a name ofbitcoin_core_rpc.__all__, in this job rather than in the suite: the project is already installed here for autodoc’s own sake. This job runs no suite at all; the suite would need thedocsgroup installed in every environment that does run one –test.yml’scoveragejob on every pull request, andos-macos.yml’s,os-windows.yml’s,os-ubuntu.yml’s,deps-latest.yml’s andmutation.yml’s besides – for a second sphinx build the suite cannot even ask for correctly,docs/build/htmlbeing this job’s own output and not a fixture the suite carries.tests/check_api_page_test.pyproves the comparison itself catches a missing name, against a page invented for the purpose. The stale example anchor at.pre-commit-config.yaml’s local-link-prefix hook, naming a pre-split id the page has not carried since the split, is fixed alongside it.docs/source/conf.py’s theme comment cites section 2 of the organization standard, correcting a citation to section 3 (closes #287). Section 3 ispyproject.tomlis the configuration; the theme decision – furo weighed against shibuya – is section 2’s The documentation. The commented-out"m2r2"line openingextensionsnamed a package in neither the dependency groups nor the lock, and is removed.This release’s own entry for issue #256 also cites section 3 for the theme decision (closes #304). Section 3 is
pyproject.tomlis the configuration; the decision is section 2’s The documentation. That entry is not rewritten, being already landed; this corrects the record beside it.REPOSITORY.mdrecords what connects this repository to Read the Docs (closes #291). Anactive: truewebhook whose deliveries 404ed is deleted rather than repaired: the documentation kept building and serving throughout, because the connection that actually carries it is the organization-wideread-the-docs-communityGitHub App (repository_selection: all), and the repository now carries no hook of its own. A dead hook reads as a fallback and is not one.
v2026.8.20¶
Repository¶
RELEASING.md’s tagging step signs the tag it creates.
git tag -ais annotated and unsigned;git tag -sis the same object with a signature on it, and nothing else about the procedure or aboutversion-checkchanges. Every commit reachingmainalready carries a verified signature,main-integrityrequiring it with no bypass actor, but the tag that turns one of those commits into a PyPI release —release.ymltriggers onpush: tags: ["v*"], and thepypienvironment is restricted to that pattern — carried none. Thepypienvironment’s required review before the publish job runs is already a real compensating control. A tag cannot be signed retroactively without moving it, and moving a released tag is worse than leaving it unsigned, so existing tags stay as they are — this applies from the next release onward. Atagruleset requiring signatures onv*, enforcing the same thing at the repository-settings level rather than documenting it, is a further step left for a separate, maintainer-authorized change (sibling repository btclib landed both halves as issue #1022).claude-review.ymlreads a pull request againstREVIEWING.md. Two jobs: one on every non-draft pull request, whose prompt names that file rather than restating it, so the standard moves without the workflow being edited; one answering@claudein a comment, carrying no prompt of ours on purpose — the action reads the comment that triggered it. It gates nothing and must not:main’s required contexts are named outside the repository. The gates are not re-run,test.yml,lint.yml,docs.ymlandintegration.ymlrunning them beside it on the same sha.Three things it refuses to do silently, each measured in btclib before being asked not to: without
CLAUDE_CODE_OAUTH_TOKENthe action reviews nothing and reports success; withoutid-token: writeit dies before authentication, minting a GitHub OIDC token at startup whatever the Anthropic credential is; and it refuses to run at all when the workflow file differs from the copy on the default branch, reporting that refusal by skipping, green. It fails on an empty secret and on an emptyexecution_file, which is exactly when no review was written — so on a pull request that adds or edits this file the job is red until the change is onmain.It is the one workflow in
CONTRIBUTING.md’s table taking noworkflow_dispatch, both jobs reading the pull request or the comment that triggered them; thegrepthat paragraph cites says so.The landing convention says what is in force, and it is the button.
CONTRIBUTING.mddescribed two landings, both pushes, with the merge button “pressed for neither” and one signer down the whole ofmain;REPOSITORY.mdcarried the pushes;RELEASING.mdsaid the button “stays permanently disabled”;CLAUDE.mdkept a fast-forward exception for a stacked pull request. Themain-self-mergebypass is inpull_requestmode across the organization, so it excuses the approving review a solo-maintainer repository cannot produce and excuses nothing else:mainis reachable through a pull request and through nothing else, a direct push refused for everyone. The ruleset namessquashas the only merge method it accepts. What the fast-forward bought — a stacked child keeping its base — is paid for with a rebase, a button recreating rather than moving whatever the count of commits, GitHub’s documentation having rebase-and-merge “always updates the committer information and creates new commit SHAs”. That the commit carries GitHub’s web-flow signature rather than the maintainer’s costs nothing:main-integrityasks for a valid signature and not for a particular signer, with no bypass actor for anyone.REVIEWING.mdis the standard a review is written against, the reviewer’s half ofCONTRIBUTING.md: what a review establishes before it gives an ack, what a finding must contain and how it labels its severity, and what becomes of everything a reviewer notices that the diff is not about — every collateral finding is filed as an issue rather than asked for in a comment. It states no new rule. Registered the wayCONTRIBUTING.mdis: a page of the sphinx toctree through areviewing_link.mdshim, named from the README and fromCLAUDE.md, which is the file a repository-aware reviewer reads, with.claude/commands/review.mdas the/reviewcommand. The body is deliberately the text btclib carries, one section excepted: the questions a review of this tree asks — the single-module and no-dependency propertytests/standalone_test.pyenforces, and the two JSON-RPC dialects_legacy_resultand_v2_resultkeep apart.A pull request of one commit is fast-forwarded onto
main, where every pull request was squashed. A squash writes a new commit: the sha the branch was written on is gone, so a pull request stacked on that head is asked for a rebase and for a re-review of a diff nobody edited, and the commit that lands is not the commit the checks ran on. Two commits or more are still squashed into one, the steps of a review being the pull request’s record rather thanmain’s. A verified signature lands either way – the branch’s own where the head is pushed as it stands, the maintainer’s where a squash is composed in a worktree, the merge button being pressed for neither and GitHub’s web-flow key therefore signing nothing here – and the ruleset takes no unsigned commit, so an unsigned single commit is squashed and signed rather than fast-forwarded. The subjectmainreads is the branch’s own where nothing composes one at the button, so a one-commit pull request names its own number, which is an amend and a force-push on work no reviewer has read yet. CONTRIBUTING.md states the rule, REPOSITORY.md the pushes that do what the merge button does not, and RELEASING.md lands a one-commit release branch this way. The push is the whole landing: GitHub reads the head as reachable frommain, marks the pull request merged, closes what it says it closes and deletes the head branch, which isdelete_branch_on_mergetreating the merge it detects as the merge it performs.The generated Code Quality analysis is off.
Analyze (python)ran on every pull request and every push tomainfrom adynamic/github-code-scanning/codeqlworkflow no file in this tree declares – some 44 seconds of a slot each time, out of twenty shared with every other repository in the organization, where the same setting was on. What it produced cannot be read from outside a browser: there is nocode-quality/alertsendpoint and nocode-quality/analyses, both 404, the alert list is empty and every analysis carriescodeql.yml’s own category. REPOSITORY.md gains the section and the endpoint that reports and sets it,code-quality/setup: notcode-scanning/default-setup, and not the Actions API, which answers 422 for a workflow this repository does not own.A pull request asks for thirty-three jobs instead of forty-four, and the number that decided it is a ceiling rather than a wall clock: GitHub Free gives an organization twenty concurrent jobs, shared across every repository in it, so a commit here and a commit in btclib compete for the same twenty. Measured on one pull request run, this workflow set asked for forty-four jobs and 16.9 runner-minutes for work that finishes in 212 seconds. Three changes, and each moves an answer off the review path rather than dropping it:
the seven
windows-11-armcells becomewindows-arm.yml, weekly on Saturday and called byrelease.yml, exactly asmacos.ymlalready holds the macOS ones. The fourteen Windows cells were 9.4 of the 16.9 runner-minutes, and that image is the slower half of them, 40 to 77 seconds a cell against 29 to 34 onwindows-latest.windows-lateststays: the datadir table indefault_datadiris the one platform-conditional branch here, the suite drives every row of it but the running one by patchingsys.platform, and a real Windows runner is where%APPDATA%and thePath.home()under it are asked for real.codeql.ymlloses itspull_requesttrigger and keepsmainand its Tuesday schedule, socodeql: every job passedis no longer one of main’s required checks – four now, and REPOSITORY.md carries the rule and thegh apipatch that dropped the context.zizmoris a pre-commit hook, solint.ymlstill audits these workflows for an injected expression on every pull request.the
(3.14, ubuntu-latest)cell oftest.yml’s matrix is excluded, because thecoveragejob beside it is that cell: same image, same interpreter, the same suite, and the only difference is the instrumentation.
RELEASING.md reads what is open before it tags. A pull request touching
release.ymlor a workflow it calls is one the tag is about to run, so leaving it in review runs the defect it fixes on the release itself – and nothing catches that, every such workflow being green on the pull request that fixes it.published.ymlis the case named, its failures arriving after PyPI has accepted the files.A release branch is thrown away and redone when
mainmoves under it, never rebased and never merged into. CHANGELOG.md and HISTORY.md aremerge=union, so a change that opened a### Repositorygroup where the release opens its own is fused into one section carrying that heading twice, with no conflict reported. Thegit diff --cachedat the landing step reads the same hazard one step too late, the fused headings being what is committed by then.The GitHub release is checked by asking for the release,
gh release view v<version>, rather than by reading the run’s conclusion: the two have disagreed on the last two tags,github-releasereportingskippedwhile its needs and the run itself reportsuccess. “If something goes wrong” records that as the expected outcome until it is fixed, and names the report to GitHub support as owed.The TestPyPI rehearsal says when it is worth its run: when the publish path or what travels it has moved –
release.ymlor a workflow it calls, the packaging metadata,MANIFEST.in,normalize_sdist.py, the trusted publisher registration, a new file the distribution has to carry. A cycle that touched the module and the prose is one the tag’s own run judges as well, since every job up topublish-pypiis the same job; what the file asks is that skipping it be stated in the release pull request rather than left to be inferred.test-passedandintegration-passedskip a cancelled run instead of failing it. Both carriedalways(), so a run the next push’s concurrency group already superseded – every dependency reportingcancelled– still reached the “Fail unless every job above succeeded” step and turned that into a required check with no defect behind it, red beside the newer run’s green on the same head sha.always()becomes!cancelled()in both jobs’ condition; a skipped required check satisfies branch protection the same as a passing one,integration.ymlalready recording that for its ownpathsfilter. The step-level condition inside each job, which fails the gate on a dependency cancelled on its own without the run itself being cancelled, is unchanged.A closed pull request’s run no longer lands in its merge’s own push run’s concurrency group.
test.yml,lint.yml,docs.ymlandintegration.ymlgroup bygithub.refalone, and every one carries a comment saying apull_requestrun lands in a different group from apushrun and so cannot cancel it – true foropened,synchronizeandready_for_review, not forclosedon a merged pull request, wheregithub.refresolves to the base branch’s ref instead ofrefs/pull/N/merge. The two events fire within about a second of each other on every merge, and after #136 landed, its owndocspush run for the merge commit was cancelled two seconds after being created, before any job started – required checks readingcancelledfor a commit the run that got to run never tested. The group is nowgithub.event.pull_request.number || github.ref: a pull_request run of any action, closed included, groups by the pull request’s own number instead, which cancels that same pull request’s earlier run exactly asclosedwas added for, and cannot equal any push’sgithub.ref.links.ymlgets the same fix though it has no push trigger to collide with, itsscheduleandworkflow_dispatchtriggers resolving to the same ref a closed run does.HISTORY.mdisRELEASE_NOTES.mdnow, mirroring sibling repository btclib’s issue #1011 (PR #1039). Its own H1 always read# Release notes; the filename did not, and in common usage the two words that were split here name the same file, where a project that does split them puts it the other way round — Keep a Changelog defines CHANGELOG.md as the curated, human-facing list, which is what this file already is.CHANGELOG.mdis unchanged: it is the file whose name and contents already agree, and every entry it has ever made about the old name stays written as it was true then.The load-bearing references moved with it:
release.ymllifts the GitHub release notes out of the tag’s own section by filename, andversion-checkrefuses a tag whose heading is not retitled, in both files it reads by name..gitattributeskeepsmerge=unionunder the new name, so a parallel release-note bullet still resolves without a conflict.docs/source/history_link.mdmoves todocs/source/release_notes_link.md, its toctree entry relabelledRELEASE NOTESto match the other entries’ pattern. This repository has no website config of its own to redirect through, so the oldHISTORY.mdpath 404s rather than forwarding anywhere; git history is where the old name stays reachable.
v2026.8.13¶
Added¶
assert_chain, the checkverify_chainmakes, as a method a caller can invoke at a moment of its own: a client built against an explicit url – a node on another host, one behind a proxy – gets whatfrom_chainwould have asked for it.from_chain(verify_chain=True)is now this method, so there is one implementation and not a second copy in whatever library wraps this one.signet is identified by its challenge, which is the half of that check a chain name cannot make. Core reports
signetfor the default signet and for every custom one alike, so two nodes sharing nothing but the shape of a challenge answer the same string;assert_chaincompares the p2p magic the challenge derives, taking the node’s from thesignet_challengemember ofgetblockchaininfoand the caller’s from the newsignet_challengekeyword, or from the default signet when none is given. A challenge off signet is refused before the round trip, there being nothing a reply could settle; one passed withverify_chainoff is refused as well, being the one argument that would otherwise quietly do nothing – every signet answers on 38332 and keeps its cookie in the same subdirectory, so the challenge changes nothing but the check. A node too old to report the member, or reporting one that is not hex, is aFetchError: the check could not be made, which is not the same as making it and passing.magic_from_chainandmagic_from_signet_challenge, the two facts that check rests on, published for a caller with a use of their own – a p2p handshake, a comparison against another implementation’s table._MAGIC_FROM_CHAINispchMessageStartper chain from Core’ssrc/kernel/chainparams.cpp, in the byte order Core writes it there, and keyed by Core’s chain names like the port and the datadir subdirectory beside it.magic_from_signet_challengeisSigNetParams’ own rule – the first four bytes of the sha256d of the challenge script serialized with its CompactSize length – and takes hex or the bytes it spells, refusing anything else:bytes(71)is 71 zero bytes rather than an error, so a challenge that arrived as a number would have been hashed as a script of that length. The table’s signet entry is a copy of the constant Core states, with a test deriving it fromDEFAULT_SIGNET_CHALLENGE– so the constant and the rule are held to agreeing, rather than the entry being a computation nothing would catch being wrong. A challenge above 65535 bytes is refused rather than serialized: consensus caps a script at 10000, so the longer CompactSize forms cannot arise for one, and refusing keeps the two forms written here the whole of the encoding.DEFAULT_SIGNET_CHALLENGE, the block challenge of the signet everyone means by “signet” – thebina node started without-signetchallengeuses, fromSigNetParams. What makes it worth publishing is that it is the one signet a caller can name rather than describe.
Repository¶
The README says who uses this client: btclib, which declares it a dependency in its own
pyproject.toml. A reader deciding on a package that depends on nothing has its tests and its badges to judge it by, and neither of those says whether anything runs it.The index-wait step of
published.ymlnames its shell,bash, where it named none and a Windows runner picked pwsh:for attempt in $(seq 20); dois aParserErrorthere, before a single curl runs, so all five Windows cells of the install-check matrix failed the same step on v2026.8.12’s release run while every Linux and macOS cell passed. The two steps after it already carriedshell: bash; this was the one the pattern had not reached.CONTRIBUTING.md states what a merge requires: an open issue the pull request closes with
Closes #N, and an approving review from somebody other than the author, GitHub allowing no self-approval. It also states the four commit-level guarantees themain-integrityruleset enforces on every push with no bypass actor – a verified signature, linear history, no force push, no branch deletion – so a contributor learns that an unsigned commit is rejected before it is something to review, rather than from the push that bounces.RELEASING.md lands the release by a signed local squash, where it said “Squash and merge” as if the button were available. It is not: the review branch protection requires cannot come from the author, there is no second maintainer, and a bypass would still have GitHub compose the commit and sign it with the web-flow key rather than the maintainer’s. The steps are the ones every maintainer-only commit here already takes – squash in a worktree, sign, push to
main, then close the pull request and delete its branch by hand, which is what the button would otherwise have done.Two release-day failures RELEASING.md had no answer for are written down with theirs: a job left
queuedfor tens of minutes onubuntu-latestis the org’s shared Actions concurrency, not this repository’s to fix, and is waited out rather than cancelled into a race with itself; agithub-releasejob reportingskippedwhile both of its needs succeeded is rebuilt by hand from the run’s own artifacts, digests compared against PyPI’s, sincegh run rerun --jobrefuses a skipped job and re-running the workflow would ask PyPI for a second upload.
v2026.8.12¶
Added¶
RpcChannel, an opt-in attribute-style façade over a client’scall:channel.getblockcount()forclient.call("getblockcount"), positional arguments the sequence form ofparamsand named ones the mapping form, never both at once.request_timeoutandmax_body_size,call’s own keyword-only controls, are reserved and forwarded tocallrather than reaching the node as a named parameter, and every name starting with_– dunders included – is anAttributeErrorinstead of a request: without that guardcopy.copyandcopy.deepcopyturn into a bound call for__setstate__or__deepcopy__, verified against a transport that raises on any request rather than assumed.BitcoinCoreRpcClientitself keeps its explicitcall(method, params)surface unchanged; see COMPARISON.md’s “Dynamic dispatch” for why the façade lives here rather than in a vendoring caller’s own copy.verify_chain, an opt-in keyword onfrom_chain: after building the client, callsgetblockchaininfoand compares itschainfield to the onefrom_chainwas given, raisingBtcRpcValueErroron a mismatch. Off by default, sincefrom_chainotherwise asks the node nothing; a cookie or a datadir authenticates the node it names, not the chain it is running, and-chain=testunder a cookie or a datadir carried over from amainsetup is what this catches instead of a wrong-network call succeeding silently.cookie_path_from_chain, the path of the cookie file a chain’s node writes: a datadir, the subdirectorydatadir_subdir_from_chainnames, and.cookie, which isCOOKIEAUTH_FILEin Core’ssrc/rpc/request.cpp. The datadir defaults todefault_datadir, asked at the call asfrom_chainasks it, and where there is no absolute base to answer with,BtcRpcValueErrornamingdatadirrather than a path resolved against the working directory.from_chainderives its owncookie_pathwith it; what it is for is the callerfrom_chaincannot serve – a node started with-datadir=somewhere else, or one reached at a url of the caller’s, which is the constructor and derives nothing – who had the two tables published and the name of the file to write out, that being the one of the three facts no function here stated.CookieNotFoundError, aFetchErrorfor a cookie file that is not there. bitcoind writes one while it runs with its rpc server enabled, so an absent cookie is a node that is not running – orbitcoin-qtwithoutserver=1– where the four other cookie failures (unreadable, oversized, non-ascii, malformed) are a file to go and look at. Since it derives fromFetchError, anexcept FetchErrorcatches it as before; what the class buys is the caller that says “start the node” for this one and names the file for the rest, instead of reporting both as a file to inspect.cookie_authraises it forFileNotFoundErroralone: ENOTDIR stays the unreadable failure, a datadir that is no directory being configuration the node starting does not fix.
Changed¶
An
HTTPErrorwith no headers is a status and a body, wherehttp_requestleft it asAttributeError: 'NoneType' object has no attribute 'get'– from reading the announcedContent-Lengthoff headers that are not there.HTTPError.headersanswers thehdrsit was built with, so the shape comes from a transport of a caller’s own, which the module docstring invites onto that path: a test double standing in for a busy node is the ordinary way to write one, andNoneis what gets passed for a field the double has no opinion about. The announced size is now read only where there is something to read it from, the bounded read being unchanged and absentContent-Lengthalready the ordinary case for a chunked reply. So the promise that everything below the HTTP status is aFetchErrorholds for that exception too – the same one carrying anHTTPMessagewas always read as intended, which is why nothing in the suite reached it.urlopen_transportchecks the scheme, the timeout and the limit before it opens anything. It is public and it takes aRequesta caller built, wherehttp_requestchecks the url it is handed and then builds one – so a caller reaching the transport directly had no scheme boundary at all:urlopen_transport(Request(Path("LICENSE").resolve().as_uri()), ...)read the local file and answered with its bytes,urlopenspeakingfile:anddata:as well as http(s). A url that reaches configuration or untrusted input could therefore turn this transport into a local-file reader, and atimeoutof zero or a negativemax_body_sizewas refused after the resource had been opened rather than instead of opening it. The refusals arehttp_request’s own, in the same words; nothing abouthttp_requestchanges.verify_chainreads thegetblockchaininforeply rather than indexing it, so a result that is no mapping, or one whosechainis not a string, is aFetchErrornaming the unreadable reply:result["chain"]on an array isTypeError: list indices must be integers or slices, not strand on a mapping without the member aKeyError, both arriving from underneath a library whose every other unreadable answer from a backend is aFetchError, and neither insideexcept FetchError. The mismatch of a well-formed reply is still theBtcRpcValueErrornaming both chains: that one is the caller’s configuration, where this is what the node sent.for_walletrefuses a client that is already a wallet endpoint, where it appended a second one:client.for_wallet("hot").for_wallet("cold")builthttp://127.0.0.1:8332/wallet/hot/wallet/cold, a path Core does not serve, so the mistake surfaced as anHttpErrorfrom the node about a path rather than as a refusal at the line that made it. The message names the client to call it on – “call for_wallet on the client it was derived from” – that arrangement, one client per wallet derived from the one built for the node, being whatfor_walletrecommends and what deriving from a wallet client is the mistaken form of. A url a caller wrote by hand ending in/wallet/hotis refused by the same check, the constructor taking the endpoint of a node; a wallet namedwalletis still a name to add,/wallet/walletbeing the endpoint of a wallet a filesystem allows. Replacing the trailing segment instead of refusing is rejected on the grounds the rest of this class refuses on: it would makefor_wallet("hot")andfor_wallet("hot").for_wallet("cold")two spellings of one endpoint, and nothing tells a caller who meant that from one who lost track of which client they were holding.default_datadiranswers Core’s datadir for the platform it runs on, not Linux’s on all of them:%APPDATA%\Bitcoinon Windows,~/Library/Application Support/Bitcoinon macOS,~/.bitcoinon everything else, which isGetDefaultDataDirin Core’s src/common/args.cpp – the#elsebranch there being what a platform absent from the table takes here. Sofrom_chainderives a cookie path Core actually writes on the three, where on two of them it derived one no node was ever going to have written and failed with a “no such file” that reads as a node that is down.Noneis still the answer where there is no absolute directory to hang the datadir off, and Windows has one more way to be in that state than a home directory does:APPDATAunset, whichPath.home()’s fall back to the passwd database has no equivalent of.from_chain’s refusal names it. The argument the old docstring made against this – “guessing per platform would put two branches here that no test on a third can reach” – is answered by the shape rather than waved away: the three directories are a table keyed onsys.platform, read at the call, so one run of the suite patches its way through all three rows and the coverage gate stays at 100 on a single operating system. A chain ofsys.platform ==comparisons would not have done: mypy narrows those to the platform it is aimed at and stops type checking the rest, and a table is what keeps every row in front of the checker as well as the runner.An absent cookie file is reported as absent:
no rpc cookie file <path>: bitcoind writes one while it runs with its rpc server enabled, where it wasunreadable rpc cookie file <path>: [Errno 2] No such file or directory– a message that says the file was found and would not open, for the case where nothing is at the path at all, and that leaves the reader to work out from an errno what a node writing the file would have meant.CookieNotFoundErrorabove is the class it now carries; a caller matching on the text of the old message is what this breaks.default_datadiris public, no longer_default_datadir.from_chainalready called it to find the liveHOMEat the moment of the call rather than the oneDEFAULT_DATADIRfroze at import; a caller deriving its own datadir-relative path – a wallet directory, a second cookie underdatadir_subdir_from_chain– needed that same answer and had no way to ask for it short of importing the underscored name or copying the function. No behavior changes; the name in__all__is the only diff.COMPARISON.md’s “Dynamic dispatch” no longer calls the façade a vendoring caller’s own to write:
RpcChannelabove is the reason, and the section now says why it is offered here instead.The module’s docstrings and comments are shorter, and no reason is stated twice: each now sits in one place and is pointed at from the others – the retry policy in the module docstring, which
HttpErrorandcallrefer to; the four obligations of a caller’s transport inHttpTransport, where the comment above it repeated them; the truncation of a failure’s body inMAX_ERROR_BODY_SIZE; the datadir and the platforms it is right on indefault_datadir, whichDEFAULT_DATADIRandfrom_chainrefer to.FetchErrornames its two subclasses once, rather than each of them repeating that it is one. Two statements were also wrong: the notifications paragraph named a_requestfunction this module does not have, and the client’s class docstring said nothing here asks the node which chain it is on, whereverify_chainabove does exactly that on request.The upstream url in the module docstring is right again: it named the repository
btclib-bitcoin-core-rpc, which this one was renamed from, and themasterbranch, which is nowmain. Both spellings still redirect, so nothing was broken by them – but that url is what a vendored copy is asked to record beside itself, and a copy is not a thing to leave depending on a redirect. The raw url for a release, on the same line, carried the same stale repository name.
Repository¶
normalize_sdist.pynormalizes the member modes too, 0644 for a file and 0755 for a directory, beside the timestamp and the ownership it already rewrote. The mode came from the working tree, so the umask of the checkout reached the published archive: a verifier following RELEASING.md’s “Rebuild a release from its tag” from a checkout whose files carry group write got a digest mismatch on an archive whose content matches byte for byte, and is told to read a mismatch as a rebuild that ran a newer setuptools, or as tampering. It is digest-preserving for what CI publishes, a runner’s checkout carrying those two modes already, so the digests already attested stay the digests a rebuild produces. Nothing in the sdist needs the executable bit – no source file here opens with a shebang, which.pre-commit-config.yamlrecords as a decision – so one mode for files and one for directories is the whole of it.tests/normalize_sdist_test.pyasserts the property the script exists for, on archives staged by hand: two whose members differ only in a mode normalize to the same bytes, which no build of this tree can check on its own, carrying the modes of the checkout it ran in.pyproject.tomlsays where the version is declared and stops there. The comment aboveversionalso restated the convention for the placeholder between releases – “the version just released, with its trailing component bumped by one, or.1appended if it had none”, which would be2026.8.9where the declared value is2026.9, the month – and went on to say the placeholder is shaped like a release on purpose, whererelease.yml’s version-check job says the opposite in the job itself and refuses a tag of that shape. RELEASING.md’s “Which version string is which” states the convention correctly and is now the one place that answers it. The calendar-versioning note and the no-leading-zero note are true of this file and stay.Three stated counts are gone from
pyproject.toml, the command that re-derives each of them staying: the parallelism note’s test count, which moves with every merge and said 298 against 306, and the site counts of thePLR2004andTRY003ignore entries, which said 47 and 37 against 48 and 38.PLR0913’s said 4 and still measures 4, and goes with them for the same reason: what argues for those three waivers is the shape of the finding, and a number nothing checks is a claim waiting to be false.The gate that talks to a live node is
integration, the name btclib gives its own, where it wasrpc-smoke:integration.yml, anintegrationjob, anintegration: every job passedaggregate and a required context of that name.rpc-smokedescribed two of the six cells – the two that grow a regtest chain and read the replies off the wire – while the other four start Core on the chains regtest is not and check only that-chain=<name>is accepted, so a reader holding the two repositories side by side had two names for one question. What the workflow runs keeps its own name,.github/scripts/rpc_smoke.pybeing a smoke script and not what the rule names. Renaming a required check is the change a pull request cannot make alone: REPOSITORY.md has the order, and the rule moved firstThe weekly sentinels run on btclib’s days.
latest.ymlandmacos.ymlmove from Thursday to Wednesday and dependabot from Friday to Thursday, keeping the day between them that made the pairing worth having: the sentinel reports, and the pull requests open the morning after, so each is a diff whose result is already known. What changes is which weekday that is, and only so that the same question is asked on the same day across the three repositories instead of on two calendarsREPOSITORY.md’s branch protection is read back from the rule rather than remembered. It said
enforce_adminswas on – “the rule holds for an administrator too, which is what makes it a rule” – and that no approving review was required, calling that omission the deliberate half of the setup.gh api repos/btclib-org/bitcoin-core-rpc/branches/main/protectionanswersenforce_admins: falseand one approving review withdismiss_stale_reviews: the opposite arrangement, where the review is required and the administrator bypass is what lets the maintainer’s own pull request merge past a review nobody can give it, GitHub not allowing self-approval. The paragraphs that argued from the old shape argue from this one, and the two that reasoned “withenforce_adminson there is nothing to override” – the deadlock warning and the rename procedure – say what the bypass costs instead. The rest of the file was checked the same way and holds: the five contexts with their app bindings,delete_branch_on_merge, thepypienvironment’sv*tag policy, release.yml’s one elevation per job, and the six security settingsThe suite runs serially again.
-n auto --dist workstealcost more than it saved here: nine interleaved runs each put the medians at 0.50s serial against 0.90s parallel, and 0.65s against 1.23s under coverage, with CI agreeing on ubuntu and on the coverage job. 298 tests that open no socket finish before ten workers have started, which is the opposite of the condition the flag exists for – checksig’s suite starts a node per test and gains about 4.5x from the same setting. pytest-xdist stays installed because the mutation session passes-n0, and without the plugin pytest rejects the flag rather than ignoring itCode scanning is a workflow in the tree,
.github/workflows/codeql.yml, rather than GitHub’s default setup. It was the only required check whose definition a diff could not review: the setting generates a workflow it does not show, where every other check onmainis a file here with its actions pinned to commit SHAs. What the setting held is reproduced rather than re-chosen –gh api repos/btclib-org/bitcoin-core-rpc/code-scanning/default-setupanswerslanguages: [actions, python],query_suite: defaultandschedule: weekly, and those are the matrix, the unsetqueriesinput and the Tuesday cron. One job per language, so a failure names the language, andcodeql: every job passedaggregates them for a rule that has to name an outcome rather than a matrix cell. The two are exclusive at the upload and not at the start –initsucceeds andanalyzeis refused, “CodeQL analyses from advanced configurations cannot be processed when the default setup is enabled” – so this cost a five-step switch a person performed, the analysis red until step 2 of it. That switch is done: default setup answersnot-configured, the rule ends incodeql: every job passedatapp_id15368 where it named thegithub-advanced-securitycontextCodeQL, and both analyses report green. REPOSITORY.md’s “Turning default setup off without deadlocking” is the order it went in and the order to repeat if the setting comes back; dropping theCodeQLcontext before disabling the setting is what keeps the rule from waiting on a check whose result is not the tree’s to produce, that context being thegithub-advanced-securityapp’s. It does not stop with the setting: it still reports on a pull request’s head,neutraland summarised1 configuration not found. A generateddynamic/github-code-scanning/codeqlworkflow outlives it too, uploading code quality results rather than security ones – a separate setting thecode-scanningendpoint does not reportCONTRIBUTING.md’s workflow table names what
releasecalls instead of counting it. The count said six wheregrep -n 'uses: ./.github/workflows' .github/workflows/release.ymlfinds five, and a row added above it would have made “the six above it” wrong a second way. Theworkflow_dispatchsentence beside it went the same way: it excepted the gates, andgrep -c workflow_dispatch: .github/workflows/*.ymlreports every workflow taking itThe
PATCHREPOSITORY.md documents for the required-check list runs, where the spelling it carried could not:gh api -fsends every value as a string and that endpoint typesapp_idas an integer, so-f 'checks[][app_id]=15368'earns422 Invalid request. For 'properties/app_id', "15368" is not a null or integer. The body is JSON on stdin instead,--input -. Where that command is used is what makes an unrun one expensive: renaming a required check cannot be done in a pull request, so the rule moves first and by hand, and a 422 at that moment invites the whole-objectPUTthe same file warns drops the signatures. The list in it is the rule’s own, read back by thegh apinow beside the table –rpc-smoke: every job passedwas in that table and missing from the command, so the command as written would have dropped a required check rather than moved oneA pre-commit rev that is not a released version fails the gate. The weekly autoupdate pull request has twice offered the same two moves this configuration must not take – typos onto
v1, a floating major tag its repository keeps beside the versioned one, and pyroma onto5.1b1, 5.1 having no released tag – and twice they were caught by hand, the second time after the merge, which is why the commit is reverted here. A local pygrep hook now names them by line instead. pre-commit itself warns that a mutable reference “is not supported” and exits zero anyway, which is the difference between the warning and this hook. A commit SHA stays acceptable: the pattern requires a prerelease marker to end the valueA merge no longer depends on bitcoincore.org answering. The bitcoind archive
rpc-smoke.ymldownloads is cached across runs, keyed by the sha256 the matrix already pins rather than by the version, so a key can only ever hold the bytes it names. What it saves is nine seconds a cell, measured; what it removes is five fetches of Core per pull request from a server nobody here runs, which arrived with that workflow becoming a required check. The digest check and its negative control still run on a restored archive exactly as on a downloaded one – verified rather than trusted is the property, and a cache is a mutable storeA pull request waits about three minutes instead of eighteen, and the wait was never work. One run of the full matrix is 45 jobs, 17 minutes of compute and 212 minutes of queueing: no cell runs longer than a minute, while the two macOS images wait 15.7 and 13.1 minutes on average for a runner against 0.1 to 0.3 for ubuntu and windows. So
test.ymlkeeps all seven interpreters and drops to four platforms, and the newmacos.ymlruns the two macOS images against the same lock weekly, on demand, and from a release – which is what keeps a macOS regression from being published while nobody waits for one. It is scheduled the same morning aslatest.yml, half an hour before, so the two read as a difference: red in both is the platform, red inlatestalone is the upgraderpc-smoke.ymlgates a merge, where it gated only a release: the reason it did not was a cost the run log does not support, 12 to 24 seconds a cell and about 90 seconds for the workflow, live bitcoind downloads included. Itspathsfilter had to go with the promotion – a required check that never runs blocks a merge where a skipped one satisfies it – so every pull request now pays those 90 seconds, and gets the one claim a recorded reply cannot make. The rule names its aggregate,rpc-smoke: every job passed, and not six cellspublished.ymlis called byrelease.ymlonce PyPI has accepted the files, with the tag’s version, and waits for the index to serve that version before installing – so it can no longer pass by testing the release before this one, which is what the dispatch RELEASING.md asked for by hand could do. Not aworkflow_runtrigger, which zizmor rates dangerous and rightly: that runs the default branch’s copy on a push nobody reviewed, where a call runs inside the release that gated it. Its schedule goes from weekly to monthly, the release path now answering the question the weekly was standing in forEvery CI job has a name, and the names say what the job answers. The ids lose a suffix that distinguished nothing –
test-py,coverage-pyanddist-pybecomesuite,coverageanddist, there being one language and one package here – while the suffix that does name a variant stays,suite-latestandinstall-publishedamong them. Every job oftest.ymlandrelease.ymlnow declares a display name, where eleven of them showed their id in the checks list beside the sentences the other workflows wrote. The aggregate gate istest: every job passed, which carries its workflow because branch protection keys a context by name alone and two workflows with a job named the same thing produce one ambiguous check; REPOSITORY.md has thePATCHthat moves the rule and the reason a rename cannot be done in a pull requestrelease.ymlhas a concurrency group, the last workflow without one and the only one whose runs have side effects: two runs at once are two attempts at the same publication.cancel-in-progress: false, against every other workflow here, and the rule behind both is now written down – a superseded run is cancelled where its subject is the commit, and kept where its subject is a version on an index, an attestation or a releaseThe distribution files are reproducible: a rebuild of a released tag is the same bytes as what was published, so the provenance attestation can be verified against a file the verifier built rather than one they downloaded.
release.ymlexportsSOURCE_DATE_EPOCHfrom the commit date, which is the whole of it for the wheel, and runs.github/scripts/normalize_sdist.pyfor the sdist – setuptools stages that archive in a directory it creates at build time and tars it as the first member, whose sub-second timestampSOURCE_DATE_EPOCHdoes not reach and whose PAX record therefore changes length between two builds of one commit. The script rewrites member metadata and no content. RELEASING.md has the rebuild command, and the two bounds on the guarantee: the build backend is resolved rather than pinned, and a TestPyPI rehearsal is a different version by constructionThe documentation build is a workflow of its own,
docs.yml, where it was the second job oflint.yml: a failed sphinx build and a failed hook are two different verdicts, and a workflow each is what gives them a badge each and a line each in the checks list. The job keeps its name,Build the documentation, so the required check onmaindid not have to move – a context is matched by name and not by the workflow that reported it, which is what makes moving a job free where renaming one is not.release.ymlcalls the new workflow alongsidetest.ymlandlint.yml, so a tag still cannot publish docstrings that read the docs would fail to renderThe distribution files attached to a GitHub release carry provenance, where until now only the copies on PyPI did: the publish action generates PEP 740 attestations for what it uploads to the index, and the byte-identical wheel and sdist on the releases page carried nothing, so whoever pinned to a release asset url or mirrored the page had no way to check where the files came from.
release.ymlgains anattestjob –actions/attest, one SLSA build provenance statement covering both files, signed with a short-lived Sigstore certificate – andgh attestation verify <file> --repo btclib-org/bitcoin-core-rpcis what checks it. The signed bundle is attached to the release too, so--bundle <tag>.attestation.jsonlverifies the same signature without asking the attestations API for it. The digests are the index’s own, the job downloading thedistartifact rather than rebuilding it. A job of its own and not two more permissions ongithub-release:id-token: writeandattestations: writestay off the job that writes releases, and further off the job that runs the build backend. It runs after whichever publish job ran, so a dispatch from an arbitrary branch signs nothing thetestpypienvironment approval did not already let through – and the TestPyPI rehearsal exercises it, which on the release path would otherwise happen for the first time after PyPI has the files and the tag can no longer be moved.mainis the only branch, renamed frommasterand now the whole of the model:devis gone, every change reaches the trunk through a pull request – a contributor’s, a maintainer’s, Dependabot’s and pre-commit.ci’s alike – and a release is a tag onmainrather than a branch merged into it. What two branches bought was somewhere bot commits could land with no required check in front of them; what they cost was a permanently draft release pull request, a rebase merge whose new SHAs left the two branches equal in content and unequal in identity, and the realign that then had to force-updatedevbehind a rule blocking force pushes that an administrator is not exempt from. One branch pays none of that, and gates every commit on the same four checks. Dependabot declares notarget-branchand pre-commit.ci noautoupdate_branch, both taking the default branch; the draft exemption the release pull request needed (|| github.base_ref == 'master', in five workflows) goes with it, so a draft runs no CI at all and nothing is exempt;test.ymlandlint.ymlpush-trigger onmain.REPOSITORY.mdholds what the branch rule requires and why,CONTRIBUTING.mdhow a change lands,RELEASING.mdhow a release is cut from one branch.main’s branch rule requires no approving review, and applies to administrators. The four checks are unchanged andstrictwith them, as are required signatures, linear history, blocked force pushes and blocked deletions; what goes is the one approving review, which GitHub does not let an author give themselves – on a solo-maintainer repository it either stops every merge or is waved through on every merge, and a rule whose normal operation is its own bypass gates nothing while reading as though it does.enforce_adminsis on in its place, so what is left holds for everyone.Squash is the only merge method the repository enables. The merge commit was refused by
main’s required linear history already, so turning it off takes away a button that could not have worked; the rebase merge could have, and what it would have done is replay a branch’s commits onto a trunk whose rule is one commit per landed change. What a single method takes away is not the choice on a pull request in front of somebody: GitHub preselects whichever was used last and offers the same dropdown in the dialog that switches auto-merge on, hours before anything merges and with nothing asking again – so “read the button before pressing it” was advice three files gave and none could enforce. REPOSITORY.md gains the section, with the two fields that shape the commit a squash writes; CONTRIBUTING.md and RELEASING.md said all three methods were enabled and say what holds instead.
REPOSITORY.md’s recorded answers stop eliding, glossing and sharing¶
REPOSITORY.md’s elision, gloss and shared comments recorded other than a command’s own output (issue btclib-org/.github#1262) (issue btclib-org/.github#1263) (issue btclib-org/.github#1264).
v2026.8.8¶
Added¶
COMPARISON.md, the case for this client against
AuthServiceProxy: the comparison table row by row, the three rows that carry the weight – amounts, credentials and errors, enforced by construction rather than left to discipline – and the consequence of the typed errors that nothing announces, a refused connection arriving as aFetchErrorand no longer as theOSErroranexceptclause was written for. Then the three features this client does not have and what decides each: dynamic dispatch, which absorbs typos of the client’s own surface and returnsAnywhere the package shipspy.typed; batching, whose correlation and partial failure JSON-RPC 2.0 section 6 settles, and whose cost is a timeout covering several node operations, amax_body_sizethat stops mapping onto an answer, and a third parsing branch – reachable throughhttp_requestandauth_headerfor the WAN link a batch pays on; and one connection per call, which is CPython’sdo_opensettingConnection: closerather than a choice made here, negligible per call on loopback and socket churn in aggregate. The README’s “Migrating fromAuthServiceProxy” is the line-by-line rewrite; this is why the rewrite is worth doing. Linked from the README, and fromdocs/source/index.rstalongside the other root documents.
Changed¶
The timeout bounds the body of a failure too, where
MAX_ERROR_BODY_SIZEand the socket timeout bounded it alone – and a socket timeout is per recv, so a peer sending an octet just inside it holds the call for as many packets as 64 KiB takes.http_requesttakes amonotonic()deadline before the transport call and reads theHTTPErrorthrough_read_bounded: anHTTPErrorforwardsread1to the response it wraps and answersheaderswith the ones it carries, so it is read as the response it is. A page still arriving at the deadline goes back as its status with an empty body, which is what a page that cannot be read at all already did – the status is the part a caller has a policy for. A transport of a caller’s own is held to that deadline for the error body alone, that being the only part of such an exchange this module does the reading of._read_boundedgrows atruncatekeyword, the two bounds differing in what they do at the limit: an answer overmax_body_sizeis refused, a diagnostic overMAX_ERROR_BODY_SIZEis cut to it and answered, and an announcedContent-Lengthover the limit is not grounds for refusal undertruncateeither._read_boundedaccumulates into onebytearraynow, not alistof chunks joined at the end. Alistkeeps every chunk alive as its own object until the join, so a response nearmax_body_sizesat in memory twice over – once as the pieces, once as the joined result – for as long as both stayed in scope; a single growing buffer replaces the pieces instead of standing beside a second copy of them. The one copy this function cannot avoid isbytes(buffer)at the end, since abytearrayis not the immutable value it promises: the docstring now says so directly,max_body_sizebounding what is read and not the memory reading it costs, which is that bound plus the one copy the return type is worth.http_request’stimeoutis validated at the boundary now, and not onlymax_body_size.BitcoinCoreRpcClientalready refused a0, a negative number,Trueor aNaNfor its owntimeoutandrequest_timeout, buthttp_requestis public on its own, so a direct caller with a transport of its own reached the socket layer with whichever of those it passed._assert_valid_timeoutruns alongside_assert_valid_max_body_sizenow, before the request is built, so every public timeout argument is refused the same way and the transport is never reached with one that cannot work.Three documented claims narrowed to what the code does. The named parameter form was one “no attribute lookup can express”, which a
**kwargsfaçade expresses; what no attribute lookup can carry is both parameter forms together with the client’s own per-call controls, and that belongs to COMPARISON.md’s account of the non-goal rather than to a migration example, so the claim is gone and the example stands on its own. A loop overcallwas the replacement for a batch without qualification, and it is an equivalent beside the node and not over a link where the round trip costs something, which is where a batch pays – the README and the module docstring both say so now, the second because a vendored copy has no README beside it. And a connection per call “costs nothing” beside the node per call, while a great many of them are socket churn: RFC 9112 section 9.6 has the server initiating the close onConnection: close, so the node is what holds the sockets in TIME_WAIT, and the docstring now sends a caller polling in a loop to its owntransporton loopback too and not only overhttps.The between-releases placeholder is the month,
2026.9, and no longer the last release with its trailing component bumped. The old placeholder was shaped exactly like a release on purpose, which left one guard in front of it:version-check’s heading check against HISTORY.md and CHANGELOG.md. A version with no day cannot be published at all – the sameversion-checkrequires three components of anything tagged, and always did – so the month-only shape puts a second, independent guard behind the first, and a checkout ofdevstops reporting itself as a release it is not.pyproject.toml,uv.lockand the two work-in-progress headings move with it, andrelease.yml’s comment on the three-component check says which guard it now is.RELEASING.md takes back eight steps from btclib’s copy of it. The rehearsal from
masterafter the version bump, and the annotated tag that names the release commit and is read back withgit showbefore the push, the argumentlessgit tagbeing onecdaway from tagging the commit before the bump; the local gates as the evidence they are here,test.ymlandlint.ymltriggering onpull_requestand a push tomasteralone, so a commit pushed straight todevruns neither; the read-the-docs builds page, which a rendered page answering 200 does not stand in for; the 100-commit ceiling on “Rebase and merge”, past which the button readsThis branch can't be rebasedand the command line it wraps is what is left; the fast-forward that ceiling’sgit merge-base --is-ancestoralso detects, which makes the realign step below it moot and is now the question that step opens with; the 404 aworkflow_dispatch-only workflow answers until its file reaches the default branch;git tag --sort=v:refname, without whichv2026.10lists beforev2026.7; and deleting a botched tag locally in every worktree, a tag being per-repository where a branch is per-worktree.latestgets the short form of btclib’s paragraph, there being no sibling pin here for the long one.keywordsand the GitHub repository topics name the same things. The keyword list had neitherzero-dependencynorvendorable, which the topics already carried and which are two of the three things the README leads with;bitcoin-rpcis new to both, and is what somebody looking for this searches for. The order is by relevance rather than alphabetical, PyPI showing keywords as the metadata gives them; GitHub sorts its own, so there the order decides only which topic goes when the twenty it allows are full, which nine is well short of. Three candidates are left out, and the comment beside the list says why:python, which names no feature and repeats theProgramming Language ::classifiers,type-hints, which repeatsTyping :: Typed, and a chain name,from_chaincarrying Core’s defaults for all five. Thepythontopic goes for the first of those reasons, leaving the two lists identical.
Repository¶
The README’s migration section pairs each
AuthServiceProxycommand with the one that replaces it, where it listed lines and left the reader to pair them. One block held three changes, each of them a comment followed by two or three lines in no marked order: theAuthServiceProxyspelling and the one here were told apart only by the class name where a line happened to carry one, so the twoclient.calllines underrpc.getblockread as a sequence to run rather than as the same call twice, and thefor_walletchange had no “before” line at all. It is now one block per change, each opening# AuthServiceProxyand continuing# this client, under a sentence of prose saying what moved and why – connecting, invoking a method, a wallet command, a batch and an error, the last two of which were prose with no command in them.batch_gains the loop that replaces it, spelled out. The claim that an unknown method “arrives at the node, not an AttributeError here” is gone with it:AuthServiceProxy.__getattr__builds a proxy for any name too, so the sentence named a difference that is not one, where what does differ is that a method name is data here and cannot collide with the client’s own attributes. No count of the changes either, the old “four changes” having counted three blocks and a paragraph. Every# this clientline was run against a stub transport before the section was rewritten around them, which is what the section had never been.The README gained the badges it was missing –
status,downloads,docsandpre-commit.ciamong them – and now carries only the ones that can turn red. None of the additions claimed something new:pre-commit.ciis the check already configured in.pre-commit-config.yaml’sci:block and already running againstdev; what was missing was the badge saying so. They read on three lines, in the order the reader asks for them: version, downloads, development status, license and supported interpreters first; test, lint, pre-commit.ci and the documentation build second; the repository and theslackchannel third, “where is the code” and “where do I ask” being the two questions a reader has once the first two lines have answered theirs. Version and downloads are adjacent, both being PyPI reading back the same project, andtestprecedeslint, a red suite and a red linter not weighing the same. The badges that report no state –cal_ver,uv, ruff, mypy, markdownlint-cli2 andpre-commit enabled– name a choice rather than measure anything, so they open CONTRIBUTING.md, which is the file that says how each choice is enforced and what the command for it is; ruff is three of them, its formatter, its linter and its docstring rules being three gates with three documentation pages and three ways to fail, where one badge announced them as one.pre-commitcloses that run because it is what runs the others, and the repository andslackbadges close the line, a contributor wanting both. Notwitterbadge, which the sibling projects dropped too. The alternative text says what a badge means – “PyPI version”, “supported Python versions”, “test workflow status” – rather than naming the site that serves the image: it is the accessible name of the link, and a flat list has nothing else to carry the meaning. One badge per source line, which puts a badge change in one line of a diff and needs nomarkdownlint-disable MD013: the rule’s 80 columns bind only where a space follows them, and in a bare URL none does. btclib and btclib-secp256k1 carry the same three lines and the same CONTRIBUTING.md run, which is what makes the three comparable; the badge sets differ only where the projects do.The GitHub repository is
btclib-org/bitcoin-core-rpc, where it wasbtclib-org/btclib-bitcoin-core-rpc: the distribution took the shorter name in v0.1.0 already, and the repository read as the wrong thing to clone for a client that is not part of btclib for exactly as long as it kept the old one. GitHub redirectsgit clone, the web UI and most of the REST API from the old path, so nothing with it cached breaks outright; every URL this tree spells out regardless –release.yml’sgithub.repositoryguard on the publish jobs,pyproject.toml’s[project.urls], RELEASING.md’s and REPOSITORY.md’sgh apirunbook commands, the README’s badges and Links section, CONTRIBUTING.md, SECURITY.md, AUTHORS.md and both issue templates – now names the new one, so none of it depends on the redirect staying up. PyPI’s Trusted Publisher entry lives outside this tree and does not follow a GitHub rename automatically; RELEASING.md’s setup step now records the new repository name, and the entry on pypi.org needs updating to match before the next release, or that release’s OIDC handshake fails.
v2026.8.7¶
Added¶
rpc_port_from_chainanddatadir_subdir_from_chain, the port and the subdirectory Core gives a chain, whichfrom_chainreads and a caller could not derive:mainkeeps its cookie in the datadir itself andtestlives undertestnet3, neither of them a name the chain name gives away. A node on another host, or one started with-datadir=elsewhere, is a url and acookie_paththe caller assembles, and these two are what it assembles them out of instead of a table copied from here. Functions rather than the dicts behind them, for the reason the vocabulary pair are functions: an unknown chain is refused where it is named, and a published dict is a table a caller can write to.ChainandNetwork, aLiteraleach for the five names of each vocabulary, and whatchain_from_networkandnetwork_from_chainare now annotated as returning. Not what they take: an argument arrives from a config file or fromgetblockchaininfo, as astrno annotation narrows, so a parameter of that type would only mean a cast at the call site — the runtime refusal is what checks a name either way. Not anEnumeither: both vocabularies arestrin every place they are spoken,-chain=and a json body included, so an enum would be an island every caller converts at, and this file has to import on 3.10, whereStrEnumdoes not exist and thestr, Enumthat stands in for it formats asChain.MAINinside an f-string.USER_AGENT, whatcallnow sends as theUser-Agentheader:bitcoin-core-rpc, where urllib’s default named the interpreter and identified neither this client nor the program running it. What a node’s access log and any proxy in front of it record, the requestidalready marking the same call in the node’s debug log. No version in it, for the reason there is none anywhere here – the release tag is the version, and a copied file would carry whichever one it was copied from forever. Published so that a caller writing a transport of their own can send the same thing.
Changed¶
every public name reaches the documentation.
docs/source/api.rstsays the page lists the whole public interface, and it listed fourteen of the twenty-three names in__all__.automodulewith:members:documents a class or a function by its docstring and a module-level assignment by the string literal that follows it; a#comment is neither, so a name carrying one is absent from the built page rather than undescribed on it. The nine wereCOOKIE_USER,DEFAULT_DATADIR,DEFAULT_MAX_BODY_SIZE,DEFAULT_TIMEOUT,MAX_ERROR_BODY_SIZE,USER_AGENT,Chain,NetworkandHttpTransport– the last of which is the type a caller has to implement to pass a transport of their own, so the interface the module tells them to use was the one they could not look up.Each has a docstring now, and the
#comment above it keeps the reasoning that is not a caller’s business: the docstring states the contract, the comment says why not the alternative, which is the split CONTRIBUTING.md already draws between the two.test_every_public_name_carries_a_docstringfails on the next name that arrives without one.a read asks for a chunk, not for the whole limit.
read1allocates what it was asked for, and the response narrows that request only where it knows how –HTTPResponse.read1caps n at the remainingContent-Length, or at the rest of the current chunk, and does neither for a body the close of the connection delimits. Against a peer like that,max_body_sizewas paid in full on every reply however small, which is the limit behaving backwards: widening it to allow one large answer cost that much on each little one. Measured withtracemallocagainst a local HTTP/1.0 server sending a fifty-octet reply with no announced length, peak allocation for onecall:max_body_sizebefore
after
8001024 (the default)
8.27 MB
0.34 MB
1024
0.14 MB
0.20 MB
Core announces a length, so a call to a node was never the case that paid it; something in the way that does not is. Not a regression from the
read1of the deadline change either – theread(max_body_size + 1)before it allocated the same, measured at the same 8.27 MB on the commit before.test_a_read_asks_for_no_more_than_a_chunkis what keeps the request bounded.three arguments are checked where they are written, which is
_checked_url’s rule and the three that escaped it. Atransportthat is not callable built a client and failed at the firstcall, from inside urllib; acookie_paththat is no path left through pathlib’s TypeError about__fspath__; awallet_namethat is not a string left through urllib’s aboutquote_from_bytes– andfor_wallet(b"hot")did not fail at all,quotetaking bytes, so it built an endpoint from a name nobody spelled. Each is aBtcRpcTypeErrornaming the argument now.cookie_pathis annotatedstr | PathLike[str] | Nonewith that, where it saidPath | str | None. The runtime check is whatPath()on the next line accepts, and an annotation narrower than the check is the disagreement that matters: aPurePosixPathworked and a type checker said it would not. Widening a parameter breaks no caller –Pathis aPathLike[str]– and what the attribute holds is aPatheither way.for_walletbuildstype(self), where it named this class. A subclass kept its type throughfrom_chain, which builds withcls, and lost it at the one call whose whole subject is carrying this client’s configuration over.the timeout bounds the whole exchange, where it bounded each socket operation. A socket timeout is reset by every packet, so a peer sending one octet just inside it held a call open until
max_body_sizeoctets had arrived – with the default that is eight megabytes at a byte a time, and no number of seconds a caller could write said otherwise.urlopen_transporttakes amonotonic()deadline before the connect and_read_boundedstops at it, so the wait is the timeout plus the one recv in flight when it passes. Refused asstill arriving when the timeout expired, a FetchError like the other ways an exchange does not produce an answer. What it costs is a reply that legitimately takes longer than the timeout to arrive – a largegetblockover a slow link – andrequest_timeoutis what that call passes.The read is
read1and notread, which is the half without which the deadline is decoration: the response reads through aBufferedReader, whoseread(n)blocks until it holds n octets, so the whole drip happened inside one call and no check ran between packets. Measured against a server sending an octet every 0.3s under a one-second timeout: thereadspelling returned only when that server stopped, theread1one at the deadline.FakeResponsein the tests answers the two calls differently for the same reason – a fake whosereadreturned early cannot tell a bounded read from an unbounded wait.Two bounds are unchanged, and deliberately. A caller’s own transport gets no deadline: two positional arguments have nowhere to carry one, and
HttpTransport’s comment lists this with the rest of what such a transport owes. And the body of a failure, whichhttp_requestreads off theHTTPError, is still bounded byMAX_ERROR_BODY_SIZEand by the socket timeout alone – a drip there is 64 KiB of diagnostic rather than eight megabytes of answer, and threading the deadline into that path is a change to the exception handling rather than to the read.a refusal on size names
max_body_size, in all three places one is raised, where they saidmore than the 8001024 allowedand left the reader to find which knob carried that number.core_chain_from_networkandnetwork_from_core_chainarechain_from_networkandnetwork_from_chain. In this modulechainis Core’s vocabulary andnetworkthe BIP one –from_chainis named for it, and so are the parameters of both functions – socore_was spelling a distinction the names around it already carry. What it costs is a rename in a caller: the old names are gone rather than aliased, one release being a short enough life for a name that a second spelling of it is the worse thing to publish.an unknown chain is refused as
unknown Core chainwherever it is refused,from_chainincluded, where that one saidunknown chain: ... These are Core's names, not the BIP ones. The check there is nowrpc_port_from_chain’s, so the three functions a chain name reaches say the same thing, and the three words say what the sentence did.the class docstring says that a call opens its own connection. urllib holds none, so every
callsendsConnection: closeand connects again. Beside the node that is loopback and costs nothing; to a node overhttpsit is a TLS handshake each time, which is a caller polling one in a loop wanting atransportof their own. Documented and not fixed: keeping a connection alive means a pool, its own thread-safety and its own eviction, none of which one bounded request needs, andHttpTransportis the seam a caller who wants all three already has.
Tests¶
the five chains are checked against every table that takes one. They are written down four times – a port, a datadir subdirectory, a network name and a
Literaleach way – and only theLiterals failed anything when one was forgotten, at a type check rather than here. A port with no subdirectory beside it derives a cookie path under a directory that is not the node’s.
Repository¶
the smoke script stops waiting on a credential the node refused.
wait_for_rpcretried everyFetchErroruntilSTARTUP_TIMEOUT, and a 401 is aFetchError– so a wrong cookie was two minutes of polling followed byno rpc answer in 120.0 s, which names the symptom. It is the caseHttpError.statusexists for, and the script already asserted the property from the other side incheck_credentials_refused. Every other status is still retried: a 503 from a full work queue is the one that does clear on its own, and the rpc error -28 of a node reading its index arrives as anRpcError.except (FetchError, RpcError)isexcept FetchError, there and instop.RpcErrorandHttpErrorare both aFetchError, so the pair read as two families where there is one.RELEASING.md asks griffe whether the breaking-change list is complete, as a release step carrying the command. HISTORY.md promises that a break is announced there, the calendar version promising nothing, and until now nothing read that promise back: the suite judges the code, and review weighs what the notes say rather than what they leave out. griffe reads the public API at the previous release and at the tree and reports the breakage between them – against v2026.8.6 it names
core_chain_from_networkandnetwork_from_core_chain, which this cycle removed and HISTORY.md announces, so the first cycle the step applies to is one it passes. A step and not a hook, for the reason RELEASING.md gives: the comparison is against the previous release, so a deliberate break stays a finding until the release that announces it, leaving every pull request in between red for something no branch introduced.
v2026.8.6¶
The first release.
Added¶
the distribution ships
py.typed, so a consumer’s type checker reads these annotations with no configuration at all. That marker has to sit inside a package directory, which is why the source isbitcoin_core_rpc/__init__.py: measured against a built wheel of each layout, a top-level module leavesmypy --strictreporting “missing library stubs or py.typed marker” and treating every name asAny– and apy.typedor a.pyiplaced beside such a module changes nothing, mypy looking for the marker under<module>/. What it costs is that vendoring is now a copy and a rename, which the module docstring and the README both say.
Changed¶
BTClibValueError,BTClibTypeErrorandBTClibRuntimeErrorareBtcRpcValueError,BtcRpcTypeErrorandBtcRpcRuntimeError. The old names were this file’s while it lived inside btclib, and btclib declares three of its own spelled exactly that way – so a consumer holding both had two same-named classes and anexcept BTClibValueErrorthat reads correct at every call site while being the wrong one at some of them. btclib hit that taking the dependency, in anexceptaround a function of this module. Nothing catches these names inside this package, so the change is a rename; nothing has shipped, so no caller has to act.a request id is prefixed
btcrpc-, where it saidbtclib-. The prefix reaches a node’sdebug.log, and naming a library the caller may not be using was a claim this package had no business making.BitcoinCoreRpcClient, one Bitcoin Core JSON-RPC endpoint and the credentials to reach it.callinvokes any one method with positional or named parameters,from_chainbuilds a client for the local node of one of Core’s chains from Core’s own port and datadir tables, andfor_walletderives the/wallet/<name>endpoint with the name percent-encoded.JSON-RPC 2.0 with 1.1 read back, so a node older than v28 — which does not know the
"jsonrpc": "2.0"marker and replies 1.1 to it — is answered correctly, an rpc error under an HTTP 500 included.FetchError,HttpErrorandRpcError: one exception for a backend that did not answer, one for an exchange that failed with a status, one for an error the node computed with its code and itsdata. All three areFetchError, so oneexceptcatches the lot, and the fields are what a caller’s own retry policy reads.HttpErrorandRpcErrorhand every constructor argument toBaseException.__init__and compose their message in__str__, which is what makes them picklable —BaseException.__reduce__rebuilds an exception fromself.args, and a class carrying two or three fields is not rebuilt by calling it with the one composed string those two used to leave inargs— and what aProcessPoolExecutorneeds to send one back from a worker rather than report the pool broken.Amounts as
Decimalin both directions: a number in a reply is parsed exactly, aDecimalparameter is refused rather than rounded throughfloat, andNaNandInfinityare refused either way.cookie_authandDEFAULT_DATADIR, for the credential bitcoind writes and rotates at every restart — read at each call rather than held.core_chain_from_networkandnetwork_from_core_chain, the two vocabularies written down once:mainwhere BIP32 and BIP173 saymainnet.HttpTransport,urlopen_transportandhttp_request: the seam that lets code calling a node be tested without one, and the bounded urllib implementation behind it.
Security¶
No credential in the url: a url carrying
user:password@is refused, that string being what ends up in configuration files and tracebacks. The class has no generated__repr__for the same reason.No redirect followed, and no proxy taken from the environment: a request already carries the
Authorizationfor the host it names, and both would send it somewhere else.httpandhttpsonly, so a url from configuration cannot make this read the local disk throughfile:.Every response body is read under a bound,
Content-Lengthchecked first and then not believed.
Repository¶
The project standards of btclib-org: the lint gate is
.pre-commit-config.yamlandlint.ymlruns that very file; every action is pinned to a commit SHA; coverage is a 100% ratchet rather than a report;mutation.ymlasks weekly whether a line the suite executes is a line the suite checks;rpc-smoke.ymlasks two live bitcoind versions whether the recorded replies are still what Core sends.Calendar versioning,
YYYY.M.D: a release is named by the day it was cut, and what it breaks is read in HISTORY.md rather than inferred from the number. The month carries no leading zero, PEP 440 normalizing2026.08to2026.8andrelease.ymlcomparing the tag to the declared string as written.release.yml’s version-check job also refuses a tag with fewer than three components: without it a tag naming only the month,v2026.8, would pass every other check and publish a version indistinguishable from the placeholderpyproject.tomldeclares between releases, which names the same month and is never tagged..github/scripts/rpc_smoke.pyrestructured fortests/rpc_smoke_test.pyto cover: every function with no live node behind it —check,port_is_free,check_legacy_reply,check_v2_reply,check_cookie,print_log_tail, andmain’s own argument parsing — now has a test with nobitcoindin the loop, and every function that does talk to one carries# pragma: no cover, that half stayingrpc-smoke.yml’s to monitor against Core itself.mutation.yml’s session runs undertimeout --signal=INTrather than the default SIGTERM, with agit diff --quietcheck afterwards that restores the tracked source if anything is left changed. The local distributor applies a mutant and restores it from afinallyblock that only runs if the process can unwind; SIGTERM kills it outright, mid-mutant, before that block runs, where SIGINT is what Python turns into a catchableKeyboardInterrupt. A session cut at its budget could otherwise leavebitcoin_core_rpc.pymutated for whatever ran next.The MIT permission notice in full at the head of every source file, and
COPYRIGHT— the text a hook requires them all to carry — is that notice rather than a pointer toLICENSE:bitcoin_core_rpc.pyis meant to be copied out, and a copy has noLICENSEbeside it to point at. No year in it, so a vendored copy nobody has touched does not look out of date every January.