The three 0.4.0 features landed in the order their PRs merged, which buried the
headline change (cohort predicates, 4.8x) below an operator config knob and a
documentation note. Reordered to: cohorts, pagination, DuckDB budget, docs,
Fixes -- with Fixes last, where it was already.
The stacked PRs each appended their own '## Fixes' heading, so resolving the
conflicts also folded two of them into the single section that belongs there.
Content is byte-identical to what merged; only section order changed. Verified
by diffing the sorted non-blank lines against the previous commit.
figures predate the row-group rechunk (cog_pipeline#93, published 2026-08-09)
and reported the mirrored column as 'local speed' with no number -- hiding the
largest difference available to a user.
Measured 2026-08-10, fresh R session per arm, against the live corpus at
pipeline_commit 3d28ddd. Madison WI, 16-core Linux workstation.
Three findings the old table could not express:
- A local mirror is 60-80x faster. A one-off question is ~12 s end to end
remotely against ~0.15 s mirrored. Stated outright now, because it is a
bigger and cheaper win for users than anything in the R code.
- Opening the session is the LARGEST remote cost (~7.5 s), bigger than any
individual query, and it lands on the user's first query rather than on
library(). The old table accounted for it nowhere, so every per-query figure
was quietly missing it.
- The remote cost is round-trips, not scanning: a repeat query over
already-touched partitions is ~1.5 s against ~4 s cold, and a full-history
query costs ~7 s whether it runs first or last (verified by running the arms
in both orders). This is why #93's 1.4-1.7x, measured through cog-api against
a local mount, does not show up on the remote path -- there, network latency
swamps scan time.
Corpus size corrected to ~201 MB: row-group chunking added ~3.4%, and 190.6 was
ambiguous between MB and MiB besides. Measured from the manifest and on disk.
The 0.3.0 NEWS section keeps 190.6 -- it was correct for that release.
Also documented HTTP 429: a burst of remote queries gets rate-limited by the
host. Hit while taking these measurements.
verbs were left materializing everything and slicing in R -- the pattern behind
the 2026-08-06 production incident. cog_gov_search() had no LIMIT at all, so an
unfiltered call returns the entire 40,336-row crosswalk.
Extracted the #39 machinery into R/pagination.R first (.validate_pagination(),
.paginate_sql(), .take_pagination_total()) rather than growing a third inline
copy: three definitions of what total_rows means is three places for it to
drift. Conflict refusals stay at the call sites because each verb's conflict
set differs. .verb_spendrev() now uses the shared helpers and is unchanged in
behaviour.
The empty-page fallback query is now passed as a thunk, so the unpaginated SQL
is only BUILT when an offset actually lands past the end instead of on every
paged call.
Two things #57 did not anticipate:
- cog_gov_search()'s ORDER BY was not a total order. population_acs DESC NULLS
LAST leaves ties -- and the whole NULL block -- in scan order, so two requests
can order them differently and a paged sweep duplicates one row while dropping
another. Added canonical_govid as tiebreaker. Unpaginated output changes only
in the relative order of already-tied rows.
- Basket mode returns one resolved row per requested name plus a sidecar
covering all of them, so a page of it is not a page of anything the caller
asked for. Refused with uscogdata_basket_pagination_conflict rather than
silently ignoring the arguments.
Both default to NULL, so cog-api adopts them behind its existing formals()
probe with no lockstep deploy.
Suite: 1067 passed, 0 failed, 0 warnings (2 pre-existing live-corpus skips).
cog_open() connected with a bare dbConnect() and set no resource pragmas, so
DuckDB claimed every visible core. Right for one interactive session on a
dedicated machine; wrong for a server, where cog-api runs two replicas on an
8-core host budgeted 4 and each replica independently claims all 8.
USCOGDATA_DUCKDB_THREADS and USCOGDATA_DUCKDB_MEMORY_LIMIT now resolve through
.cfg() -- inheriting the env var > option > default precedence USCOGDATA_URL
already had -- and are applied as pragmas when the connection is created.
Unset issues NO pragma, so an unconfigured session is byte-identical to before.
That negative property is asserted directly against a connection opened the
pre-change way rather than against a hardcoded core count.
.cfg() returns an env var as character, so both resolvers coerce and validate
rather than trusting the type: sprintf("SET threads TO %d", "4") would
otherwise abort inside the connection path with an error naming the pragma
instead of the setting the operator got wrong.
Replaces cog-api's getFromNamespace(".ensure_session", "uscogdata") workaround,
which depended on a private name and on the session already being open.
cog_spending(), cog_revenue() and cog_balances() gain optional state/type
arguments. Both default to NULL, so every existing govid-based call is
unchanged.
The verbs took a cohort only as a govid vector, which .sql_lit_chr()
rendered into a quoted IN list and .verb_spendrev() embedded into 5-8
separate statements per call: the scope check, the main aggregate, the
per-capita join, the harmonization block, and the suggestion and
suppression queries. For type = "city" that list is 301,589 characters,
parsed and planned from scratch every time it appears.
Passing state/type instead expresses the cohort as a subquery against
canonical_fips_xwalk, so its size never enters the SQL string at all.
Measured on the production corpus, same FY2022 aggregate over the
20,106-government city cohort, DUCKDB_THREADS=2, median of 5:
IN (20,106 literals) -- 0.3.0 432 ms
join against a temp cohort table 132 ms
predicate on canonical_fips_xwalk 102 ms
no cohort filter at all (the floor) 105 ms
The predicate reaches the no-filter floor: the cohort restriction is
now free. End to end through cog_spending(category = "Police"),
1080 ms -> 271 ms, 3.99x -- larger than the single-query saving,
because the repetition across statements is what actually cost.
Design decisions, both made explicitly rather than left implicit:
- govid AND state/type INTERSECT. "These ids, narrowed to that
state/type" is a real query, and an error here could never be
relaxed later without breaking callers.
- A predicate cohort has no id list to report, so
provenance$scope$govids_found/govids_missing stay empty and a new
scope$cohort block carries state, type and n_governments. Resolving
the ids just to report them would put 20,000 govids in every
fleet-scale response body -- the cost this change removes. A
govid-named cohort's provenance is untouched.
state/type are coerced with .coerce_state_to_fips()/.coerce_type(), the
same helpers cog_gov_search() uses. That is load-bearing: the argument
is a postal abbreviation ("WI") while fips_state holds a FIPS code
("55"), and a predicate on the raw parameter matches nothing and returns
an empty result indistinguishable from "reported nothing". cog-api hit
exactly this trap optimizing the same path.
.attach_per_capita() now keys its population lookup on the govids present
in the result rather than the requested cohort. Those are the only ones
its LEFT JOIN can match, so the output is identical -- but it needs no id
list, and on a paginated call it looks up one page instead of the fleet.
Fixes uscogdata#58.
gsub() in regex mode treats backslashes in the REPLACEMENT string as
escape sequences and silently drops them. A Windows corpus path is full of
them, so C:\Users\RUNNER\AppData\... was substituted into the view SQL
as C:UsersRUNNERAppData... and every DuckDB read failed with 'No files
found that match the pattern'.
Effect: uscogdata could not read a LOCAL corpus on Windows at all -- the
bundled fixture included, so the entire test suite failed there, and any
cog_mirror() copy was unusable. Remote https URLs were unaffected, having
no backslashes, which is part of why it stayed hidden.
The bug predates the {long_files} token; it lived in the original {url}
substitution since that code was written. Nothing ever ran on Windows
until the GitHub mirror's check matrix existed, which found it on its
first run: 14 of 15 Windows failures were this, the 15th a downstream
consequence of view registration failing.
fixed = TRUE treats pattern and replacement as literal text. The
regression test reproduces on any platform -- it is string handling, not
filesystem behaviour, so it needs no Windows runner.
A plain non-force git push rather than Gitea's built-in push mirror. A
push mirror force-updates the refs it owns, so a Merge clicked on a GitHub
PR would be silently overwritten on the next sync -- the PR still reading
'Merged' while its commit became unreachable. A non-force push is rejected
instead, which turns that into a red CI run.
Secret is PAT_GH, not GITHUB_MIRROR_PAT: Gitea reserves the GITHUB_ and
GITEA_ prefixes for its own injected variables and refuses secrets using
them.
The canonical Gitea runner is Linux-only, and this package hard-depends on
duckdb and httr2 -- both compiled, both with real platform variance --
while having never been checked on Windows or macOS. A large share of the
audience is on Windows.
Inert here: Gitea reads .gitea/workflows, GitHub reads .github/workflows,
so both configs coexist and the Gitea CI stays authoritative for deploys.
This goes live the moment the mirror repo exists (#44).
The matrix needs no credentials -- setup.R points at the bundled fixture --
which is why that fixture must never be .Rbuildignore'd.
Removing ^vignettes$ was right; removing ^doc$ and ^Meta$ with it was
not. Those are devtools::build_vignettes() artefacts, not sources -- R CMD
build regenerates inst/doc/ from vignettes/ by itself, and shipping the
local copies earned a 'non-standard file/directory found at top level'
NOTE.
R CMD check --as-cran is now 0 errors, 0 warnings, 0 notes.
Moves developer, testing and release instructions out of the README,
minus the fixture-stripping advice, which was wrong.
Explains that a GitHub PR closes itself as merged once the mirror syncs,
because the merge preserves the contributor's SHAs -- so a PR closing
without a visible Merge click reads as success rather than rejection.
Documents the cog-api dependency: its CI clones this package at
USCOGDATA_REF, defaulting to main with no pin, so anything merged here
reaches the API's next build. Includes the commands to run its suite
against a branch first.
NEWS described changes relative to states no user had ever seen --
'Breaking: corpus schema_version 4', 'the package now requires...' --
across the whole pre-release development. To someone deciding whether to
depend on this, that reads as instability.
0.3.0 is written as an announcement: what it covers, the verbs, that
reading the corpus now works out of the box, four things to know before a
first query, and the known limits. The 0.2.0 changelog is kept verbatim.
The 0.1.0 development log is dropped; that history is in git.
cog_explain() now documents what provenance actually holds, since the
README points readers at it -- in particular why series_break_refs and
corpus_break_refs are separate fields rather than one list.
Reordered around a new user: what the data is, where it comes from,
install, a quickstart that runs with no configuration, then the
full-dollars warning and the concepts that decide whether a published
number is right.
Adds a 'Where the data comes from' section linking the API documentation
site, the live API, the Hugging Face corpus and the Census source, so
attribution and provenance are reachable from the top rather than implied.
Drops the sibling-repo path, the commented-out install line, the Status
block, and the release advice telling you to strip the fixture -- which
would break the vignette and leave public CI unable to check without
credentials.
The quickstart passes years=; cog_spending() has no full-history default,
so the obvious one-liner errors on a reader's first call.
The reference index covered 6 of 14 exports, so pkgdown errored on the
eight missing topics and the docs site did not build at all. Adds a
Comparison & aggregation section and a Corpus metadata section, lists both
vignettes as articles, and sets url so canonical links and search resolve.
build_site() now completes clean: reference metadata ok, no problems.
.Rbuildignore excluded ^vignettes$, ^doc$ and ^Meta$, so an installed
uscogdata had no vignettes at all -- while the README instructed users to
run vignette("total-spending"), which failed for every one of them.
Both build offline: total-spending points USCOGDATA_URL at the bundled
fixture, population-denominators is eval = FALSE. Confirmed present in the
built tarball as both source and rendered inst/doc/.
The test also pins the fixture as never-excluded -- it is what lets
R CMD check run with no credentials on r-universe and GitHub Actions.
LICENSE held only the two-line stub and no LICENSE.md existed, so the
repo carried no license text for a human browsing it or for GitHub's
license detector.
usethis::use_mit_license() writes LICENSE.md but leaves an existing
LICENSE alone, so the stub kept saying 'Civilytics' while the full text
said 'Civilytics Consulting LLC'. Corrected by hand, with a test pinning
the two together.
Authors@R was an org with no human, so citation() and the r-universe
maintainer page had nothing to render and ORCID could not collate this
with merTools. The given-name vector c("Jared", "E.") matches merTools
exactly; person("Jared", "E. Knowles") would render the same but put the
middle initial in the family-name slot.
MaxCorpusSchema claimed 5 while .validate_schema() accepts 4-7 and the
published corpus is 7 -- metadata contradicting code by two versions.
0.3.0 rather than 0.2.0: remote reads go from broken to working and the
default URL from placeholder to live, which is user-visible behaviour.
The remote-read defect survived because every test path used a local
corpus, and so did the API in production. This is the only test that runs
the package the way a new user does: no USCOGDATA_URL, no option, no
fixture -- just install and call a verb.
Gated on USCOGDATA_LIVE_TEST so offline CI skips rather than fails.
Measured against the live corpus while writing this: cog_spending for one
government is 3.9s for a single year and 5.9s across 2000-2022. Well above
the 1.5-2.8s raw parquet scan, because the verbs also join crosswalks,
resolve categories and assemble provenance.
The default was a REPLACE_WITH_SHARE_TOKEN sentinel and no document in
the package supplied a working URL, so a new user installing uscogdata
had no path to a session at all -- just an actionable-looking error with
nothing actionable behind it.
The default is now the public HuggingFace mirror: CC-BY-4.0, no
credential, CDN-backed, and it keeps the origin's uplink out of the read
path. USCOGDATA_URL and options(uscogdata.url=) still override, so
Nextcloud and cog_mirror() copies are unaffected.
The sentinel guard stays for half-edited configs; the two tests covering
it set the URL explicitly, so they only needed renaming to stop calling
it 'the default'.
DuckDB cannot expand a glob over generic HTTP -- there is no directory
listing, and allow_asterisks_in_http_paths only forwards the literal
'**/*' as a filename, which 404s. So every remote corpus read failed.
Only local paths worked, which is how the API (a host mount) and the test
fixture run, so nothing ever caught it.
Measured against the published corpus: the explicit list returns the same
46,148,034 rows the hf:// glob does, with hive_partitioning still
recovering year from the paths. Building it from the manifest keeps the
reader host-agnostic rather than binding it to one vendor's protocol.
Also extracts .render_view_sql(). Four test sites had hand-rolled the
{url} substitution -- one commented as doing it 'exactly as
.register_views() does' -- and all four broke on the second token. They
now share the one function that knows the vocabulary, and a new test
renders every SQL file to prove no token survives.
Spec and plan were written against a branch 25 commits behind main, where
the package still read 0.1.0. It is 0.2.0, with a real 0.2.0 changelog in
NEWS that the plan would have deleted.
0.3.0 rather than 0.2.0 because remote corpus reads go from broken to
working and the default URL from placeholder to live -- user-visible
behaviour, so a minor bump. Not 1.0.0: types 4 and 5 remain out of scope.
Task 9 now prepends a 0.3.0 section, keeps 0.2.0 verbatim with a diff
check to prove it, and drops only the 0.1.0 development churn. Task 4
gains the Version bump.
Ten tasks, 58 steps, TDD throughout. Tasks 1-3 fix the P0 (manifest
enumeration, working default URL, and the live-corpus test whose absence
let the defect survive); 4-7 are metadata and packaging; 8-10 rewrite
README, NEWS and CONTRIBUTING.
Distribution mechanics stay out of scope -- r-universe publishes check
results on registration, so it comes after final verification is green.
Covers the P0 finding that the package cannot read the corpus remotely at
all -- no working default URL, and Hive globs are unsupported over generic
HTTP by DuckDB 1.5.5. Fix is manifest-driven file enumeration (measured:
46,148,034 rows over plain https, identical to the hf:// glob) plus a
working public default.
Also: seven release-readiness fixes, a README restructured for a stranger,
NEWS rewritten as an initial release rather than a pre-release churn log,
and the Gitea-canonical/GitHub-mirror/r-universe distribution mechanics.
Picks up seven new catalogued series breaks, all break_year 2017,
recording that the employee-retirement X-codes (X21, X30, X42, X44, X47,
Z77, Z78) were last collected in the annual finance file at FY2016 before
those systems moved to the Annual Survey of Public Pensions.
series_breaks.parquet 202 -> 209 rows; nothing removed. manifest built_at
and pipeline_commit updated, and the series_breaks sha256 with them.
cog-api's paginate() sliced an already-fully-materialized result: every
page of a deep sweep re-ran the whole cog_spending()/cog_revenue() query
and re-listified every row, just to keep up to 1000 and discard the
rest. A 193,105-row/194-page fleet-wide sweep (cog_explorer's Southern
guide, corpus summary build) repeated that full cost 194 times and
wedged the production server for hours on 2026-08-06 -- single request,
CPU-bound, single-threaded plumber process, no other request could get
through, not even /health.
cog_spending()/cog_revenue() gain optional limit/offset, pushed into
.build_verb_sql() as SQL LIMIT/OFFSET behind the existing (already
deterministic) ORDER BY. The full unpaginated row count rides along via
COUNT(*) OVER() in the same scan -- exposed as a total_rows attribute --
so a caller walking pages never needs a second round trip to ask how
many there are. A page now costs O(limit), not O(full result).
Mutually exclusive with complete = TRUE (which fills a grid over the
FULL requested (year, category) space -- pagination over a partial slice
of already-grouped rows has no defined meaning for the cells it would
fill) and with recipe (whose result comes from a separate, not-yet-wired
query path). Both abort with a clear classed condition rather than
silently ignoring the parameter.
limit/offset default to NULL; every existing call site is unaffected.
.build_suggestions()'s recipe-candidate sub-select was keyed on
`WHERE category IN (<category>)`. The reserved pseudo-category
"All Categories" is never itself a row in summary_categories.category, so
in all-categories mode `candidates` always came back empty and coverage
signposting (uscogdata#9) was structurally impossible for the one mode
whose entire premise is "you cannot sum the wrong scope" -- measured on
Los Angeles County FY2011: category = "Public Welfare" reports 2
suggestions (incl. $271,589,000 excluded E68), category = "All Categories"
reported 0, silently losing that same signal.
Apply the branch's own design principle: the concept boundary is subtype,
not category. .build_suggestions() now accepts all_categories/subtype_col/
subtype_scope (all optional, default off, so no other caller's behaviour
changes) and, when all-categories mode is active, scopes the candidate
sub-select by `<subtype_col> IN (<subtype_scope>)` instead -- symmetric
with .build_verb_sql()'s own WHERE predicate. The M/L recipe exclusion and
the is.null(category) early return are unchanged.
After the fix, LA County FY2011 "All Categories" reports 5 suggestions,
including welfare_cash_e68_wide for the exact $271,589,000 gap.
Adds two covering tests to test-all-categories.R using the bundled fixture
(AL state gov, FY2011, "Corrections"): one end-to-end (per-category and
all-categories both signpost the same recipe) and one direct on
.build_suggestions() proving the subtype-vs-category branch is what
changes the query. Updates the 0.2.0 NEWS entry.
- .detect_direct_suppressed() keys on (year, canonical_govid, category);
all-categories mode collapses category to one literal value, so the key
collides and the detector silently reports FALSE instead of "unknown".
Report NA there instead, and stop isTRUE() in .build_provenance() from
collapsing that NA back to FALSE. Schema widened to allow null.
- Refuse complete = TRUE + category = "All Categories": the completion grid
has no per-category cells left to fill once categories are collapsed,
so the prior silent 0-rows-filled result was never actually checked.
- cog_balances(category = "All Categories") returned zero rows with no
error. .validate_verb_inputs() gains allow_all_categories (default
FALSE); .verb_spendrev() passes TRUE, cog_balances() does not, so the
three verbs share one place to reject it instead of drifting again.
- Fix the false `subtype = "operations"` argument claim (no such argument
exists) in NEWS.md and an internal spending.R comment.
Adds three covering tests to test-all-categories.R for the three
behaviour changes above.
Bumps the minor version because 'All Categories' adds public surface
without breaking any existing call.
The bump is load-bearing, not cosmetic: cog-api installs this package
with install_local(), which no-ops when the version already matches.
Without it, Phase 1 would silently test against the 0.1.0 reader and
pass while proving nothing.
Closes uscogdata#36. It counts governments with rows for the requested
category, so a surveyed government that genuinely spends nothing there
is indistinguishable from one never surveyed. In FY2022, a complete
census year, Georgia reports 393 of 567 cities for Police -- the gap is
cities that contract to the sheriff.
Documents the comparison that IS valid: same category, census year vs
sample year.
Geographic totals are the expensive case cog-api#37 was filed about --
without this a caller issues one rollup per category and sums them.
The pass-through was expected to work by construction; this asserts it
rather than assuming it, including under per_capita and inflation
adjustment.
A reserved value nobody can discover is a trap, and this is the view
the API's /categories endpoint is built from. Emitted for the two flow
vocabularies only -- cog_balances() returns a stock and has no concept
to sum within.
Also fix test-categories.R to exclude pseudo-category rows from
crosswalk-specific assertions (one row per (category, subtype) pair,
non-empty item_codes, valid subtypes).
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
The @param category text on cog_spending()/cog_revenue() told users to
"Combine with subtype = ..." but neither verb has a subtype argument.
Replace with accurate guidance: filter the returned frame's
spend_subtype/revenue_subtype column.
Returns one summed row per (year, govid, subtype) across every
category in the requested concept's subtype scope, so a caller never
sums categories client-side and cannot sum the wrong scope.
Combining it with other category names is an error rather than a
silent partial sum.
The concept boundary in this package is subtype, not category, so a
total is the existing query with the category dimension collapsed and
no category predicate applied. subtype is deliberately kept in the
grouping: subtype=operations plus all-categories is 'operating
expenditure', which is the measure a fiscal comparison wants.
Named 'All Categories' rather than 'Total' because category='Total'
would sit one argument from expenditure_concept='total' and mean
something different.
R/suggestions.R crossed the project's 400-line limit (424 lines).
Pure move of .suppressed_components() and its roxygen block per the
plan's Task 5 Step 2 remedy; no logic, SQL, or wording changed.
Scoped re-review measured .needs_suppression_query() against the fixture
and found it doesn't pay for itself: it skips the round trip on ~3% of
healthy candidate-bearing calls, ~0% of the multi-govid batch shape
(cog_geographic_rollup()/cog_peer_compare()) it was meant to help, and
reaching the gate costs an unconditional metadata query that on its own
roughly cancels the expected saving -- net slower on the fixture. The
gate was also correct (0 unsound skips) but left an untested exactness
invariant (result$codes_included and the anti-join sharing the harmonized
item_code space) whose silent violation would kill signposting, which is
the exact failure class uscogdata#9 exists to prevent.
Owner's call: revert it and keep the code simple. A batch-aware
optimization, if warranted, is a separate issue.
Removes .needs_suppression_query() entirely (function, roxygen, call
site, comp_rows/flow_components), restoring .build_suggestions() to call
.suppressed_components() directly -- unchanged from f77adb6 except that
it still threads flow_prefixes through (I1, kept). Also removes the two
tests that existed solely to exercise the gate (the five-branch synthetic
test and the local_mocked_bindings call-counter test); no test asserting
real signposting behavior was touched.
I1 (flow_prefixes filter), I2 (schema wording + ig_recipe_id required),
I3(a) (restated NOT EXISTS literals for partition pruning), and M4
(reworded scope claims) are all untouched by this revert.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
Final whole-branch review fix wave for the partial-coverage signposting
feature:
- I1: .suppressed_components() now filters measured recipe components to
the calling verb's own flow_prefixes. Without this, a candidate recipe
from the OTHER flow family was always absent from the verb's own view by
construction and so was always reported as "suppressed" -- fabricating a
dollar claim across flow families (cog_revenue(category = "Corrections")
claimed $3.63B excluded that cog_spending() actually reports in full).
- I2: reworded provenance-v1.json's trigger/suppressed_amount descriptions
to describe what the code actually measures (the verb's underlying long
view, not "the result"), and to note suppressed_amount can be negative.
Added ig_recipe_id to the suggestions items' required list, matching the
key's always-set/nullable runtime behavior.
- I3(a): restated the year/govid literals inside .suppressed_components()'s
NOT EXISTS subquery so DuckDB can partition-prune that side too (verified
via EXPLAIN: Scanning Files 1/4 instead of an unfiltered full scan;
all.equal(old, new) results confirmed unchanged).
- I3(b): added .needs_suppression_query(), a free, exact pre-check reusing
the verb's own already-computed result$codes_included to skip the anti-
join round trip on the common fully-covered path, without weakening the
"suppression can fire with zero gap years" guarantee.
- M4: corrected the overbroad "confines every fire to 2011" scope claim in
R/suggestions.R and NEWS.md -- the suppressed-dollar measurement is now
flow-scoped (post-I1), but the empty_year trigger itself is not, and can
still fire in modern years for a mis-scoped cross-flow-family category.
Added a regression test for I1 plus direct unit-test coverage for the new
flow-family filter and the I3(b) pre-check.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
devtools::document() should not change DESCRIPTION when only JSON/Markdown/test files are edited. Restore the original RoxygenNote: 7.3.3 and remove the Config/roxygen2/version auto-generated line that resulted from running devtools::document() locally.
jared
