Gate pkgType = 'source' on ref, own it in val_build() only (fixes #181) by aclark02-arcus · Pull Request #182 · pharmaR/val.pipeline

aclark02-arcus · 2026-09-14T14:57:47Z

Closes #181.

Root cause recap

Two coupled problems in val_pipeline() -> val_build() regressed logrx covr_coverage from >90% (bare val_build()) to ~62% (through val_pipeline()):

val_pipeline() pushed options(pkgType = "source") into the parent session unconditionally, outside the ref gate.
val_build() itself also set pkgType = "source" when ref == "source". This is the intuitive setting for a "source-tier assessment", but it's backwards. Confirmed empirically: entering val_build(ref = "source", ...) under getOption('pkgType') == 'source' reproduces the ~62% number; entering under "both" or "binary" gives >90%.

Likely mechanism (`pkgType = "source"` breaks source-tier covr)

val_build() calls configure_riskmetric_offline_if_requested() near entry, which uses utils::available.packages() to build an offline cache. available.packages() defaults type = getOption("pkgType"), so under "source" the cache is filtered to source-only. Any test-time dependency that exists in the caller's PPM-style mirror only as a binary drops out of the offline cache. Downstream, when covr instruments the target package and its test subprocess needs those deps, the install probe fails silently and the test suite short-circuits — covr reports the resulting partial coverage as the package's "real" number.

pkgType = "both" keeps the cache complete (binary-preferred, source fallback) while ref = "source" still drives val.pipeline's own source-tarball assessment path in val_pkg(). That's what this PR now enforces.

What this PR does

Extract apply_val_build_options(ref, opt_repos) into R/utils.R. Sets repos (always) and, when ref == "source", sets pkgType = "both" (NOT "source"). Returns the prior slot values in an options()-shaped list ready for on.exit(options(old), add = TRUE).
Drop the unconditional pkgType = "source" at val_pipeline.R:286. val_build() owns pkgType now.
Fix broken on.exit(function() options(old)) in val_build(). It constructed a function value and threw it away, never restoring options — and lacked add = TRUE, silently wiping the earlier on.exit(options(old_cfg), add = TRUE).

Regression tests

Six tests in tests/testthat/test-apply_val_build_options.R:

Two unit tests for the base contract (ref = "source" sets repos + pkgType = "both"; ref = "remote" sets repos only, leaves pkgType alone).
One restore-round-trip test (options(old) fully undoes the mutation).
Directional regression guard: asserts apply_val_build_options(ref = "source", ...) NEVER sets pkgType = "source". Forces any future attempt to flip this back to the intuitive-but-wrong direction to fail loudly.
Two source-scan tests on the installed R sources:
- val_pipeline.R post-prep options block does NOT set pkgType.
- val_build.R sets options AFTER configure_riskmetric_offline_if_requested() (kept AFTER for scope isolation, though the value now — not the ordering — is what actually carries the fix).

What this PR intentionally does NOT touch

val_prep_pipeline.R:210 and val_decision.R:287 also set pkgType = "source" unconditionally, but their side effects are contained by val_prep_pipeline()'s own on.exit(options(old), add = TRUE) guard (see R/val_prep_pipeline.R:142). They don't leak into val_pipeline()'s post-prep scope, so they aren't the root cause. Can be revisited in a follow-up if desired — but the same "prefer both" logic likely applies wherever pkgType is set in a val.pipeline scope.

Verification

devtools::test(filter = "apply_val_build_options") → 23/23 pass.
devtools::test(filter = "val_") → 219/219 pass.
Full devtools::test() (last full-run baseline) → 1088 pass, 0 fail, 1 pre-existing skip.

User-visible expected impact

Existing val_pipeline() callers get higher covr_coverage on any package whose test-time deps included binary-only mirror entries. Aaron's 2026-08-25 run had ~600 packages below the 65% threshold; some will now land above threshold on the next run.
Bare val_build(ref = "source") behavior CHANGES: previously it set pkgType = "source", now it sets pkgType = "both". Callers who invoked val_build() directly and depended on the old pkgType = "source" side effect for downstream install behavior will see the same > 90% coverage improvement.
ref = "remote" runs are unaffected — pkgType is left at the caller's value in either case.

Companion dev script

dev/dev_rerun_low_covr.R (added on this branch): surgical driver for a caller's low-covr cohort — filters a qual_assessments × qual_metadata-merged snapshot to packages likely to benefit from the fix (default covr_coverage > 40 AND runtime <= 15 min), reruns them via val_build(replace = TRUE, prep = NULL, finalize = FALSE), and prints a before/after covr_coverage delta. config_path defaults to NULL (resolves to packaged inst/config.yml) so the script runs cleanly outside a launcher project.

Version + NEWS

DESCRIPTION 0.1.59 -> 0.1.60 (patch).
NEWS.md: one bullet under new # val.pipeline 0.1.60 heading, describing both prongs of the fix.

Note: PR #180 also bumps to 0.1.60. Whichever merges first wins; whichever merges second needs a rebase-bump to 0.1.61.

val_pipeline() used to push options(pkgType = 'source') into the parent session before dispatching val_build(), unconditionally and outside the ref gate. That pre-set state changed how configure_bioc_repositories_if_requested() -- which val_build() calls near entry -- resolved its available.packages() cache, so the same val_build(pkg_names = 'logrx', ref = 'source', ...) returned >90% covr_coverage when called directly but ~62% when called through val_pipeline(). Also fixes val_build's broken on.exit(function() options(old)) -- which constructed a function value and threw it away, never restoring the option snapshot, and silently wiped an earlier on.exit(options(old_cfg), add = TRUE) because it lacked add = TRUE. Ownership of the pkgType slot now lives in exactly one place: val_build(), via the extracted apply_val_build_options() helper, gated on ref == 'source' and set BEFORE the bioc-config helper runs. Verified: * devtools::test(filter = 'apply_val_build_options'): 17/17 pass. * devtools::test() full: 1088 pass, 0 fail (1 pre-existing skip). Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com>

Aaron flagged that essentially every real val.pipeline run uses ref='source' (that's what riskmetric needs to produce R CMD check + covr_coverage), so the ref gate is effectively always-true. That means the previous version of this PR -- which moved the pkgType set to BEFORE configure_bioc -- would have regressed bare val_build to the same 62% coverage that was reported through val_pipeline. The historical bare-val_build ordering (configure_bioc first, then set pkgType) is the state that produced the accurate >90% number on logrx. Preserve that. The fix for the val_pipeline path stays: drop the outer-seam pkgType pre-set so val_pipeline enters val_build with pkgType at the caller's default, matching what bare val_build has always done. Test suite: 1088/1088 pass (0 fail, 1 pre-existing skip). Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com>

aclark02-arcus · 2026-09-14T15:08:02Z

Push 6d3d7a5 — course-correction after Aaron pointed out that ref = "source" is essentially always set in real runs (it's what riskmetric needs to produce R CMD check + covr_coverage), so the ref gate is effectively always-true.

That means the earlier commit 2eb3fc6 -- which moved the pkgType set to BEFORE configure_bioc_repositories_if_requested() -- would have regressed bare val_build() to 62% too, since every source-tier bare val_build() call would now run bioc-config under pkgType = "source". Same failure mode Aaron's original one-liner experiment exposed.

Corrected ordering: configure_bioc runs FIRST (under whatever pkgType the caller's session carried at entry -- typically unset), THEN apply_val_build_options() sets pkgType = "source" when ref = "source". This matches the historical bare-val_build() code path that produced the accurate >90% covr_coverage on logrx.

val_pipeline() fix stays intact: the outer seam's unconditional pkgType = "source" pre-set is still dropped, so val_pipeline() -> val_build() now enters val_build() with pkgType at the caller's default, matching bare-val_build() state.

Verified: full devtools::test() -- 1088 pass, 0 fail, 1 pre-existing skip.

Reads dev/low_covr.rds (gitignored, sensitive) and drives val_build() against a filtered high-yield / low-cost subset (default: covr > 40% AND runtime <= 15 min) with replace = TRUE. Post-run diffs against the prior covr_coverage values and reports which pkgs cleared the 65% threshold. Companion .gitignore entries keep the sensitive rds inputs + generated compare tables out of the tree. Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com>

Copilot

🟡 Changes recommended

The rerun utility cannot read post-run coverage correctly, has a broken default config path, and the documented option ordering contradicts the implementation.

Get a fresh assessment by requesting another Copilot review.

Pull request overview

This PR centralizes pkgType handling in val_build() to address inconsistent coverage results and option leakage.

Changes:

Adds conditional option setup and restoration.
Removes unconditional pkgType = "source" from val_pipeline().
Adds tests, release notes, and a low-coverage rerun utility.

File summaries

File	Description
`R/utils.R`	Adds the option-management helper.
`R/val_build.R`	Applies and restores build options.
`R/val_pipeline.R`	Removes unconditional `pkgType` setup.
`tests/testthat/test-apply_val_build_options.R`	Tests helper behavior.
`dev/dev_rerun_low_covr.R`	Adds targeted reassessment tooling.
`.gitignore`	Excludes rerun artifacts.
`DESCRIPTION`	Bumps version to 0.1.60.
`NEWS.md`	Documents the fix.

Review details

Suppressed comments (1)

dev/dev_rerun_low_covr.R:214

The comparison reads _meta.rds, but val_pkg()'s metadata bundle does not contain covr_coverage (R/val_pkg.R:934-1061); the raw scalar is stored in _assess_record.rds. Consequently this condition is false for every successful rerun, so covr_after, every delta, and the threshold count are always NA/zero. Read the assessment record instead.

  m <- tryCatch(readRDS(meta_path), error = function(e) NULL)
  covr_after <- if (!is.null(m) && "covr_coverage" %in% names(m)) {
    as.numeric(m$covr_coverage)

Files reviewed: 7/8 changed files
Comments generated: 3
Review effort level: Balanced

💡 Add a code-review agent skill or configure MCP servers for context-aware, tailored reviews. Learn more in the docs.

+# state between `val_pipeline() -> val_build()` and bare `val_build()`
+# and silently regressing covr_coverage on some source-tier packages.
+
+test_that("apply_val_build_options(ref = 'source') sets repos AND pkgType", {


+  # Establish `options(repos, pkgType)` AFTER
+  # `configure_bioc_repositories_if_requested()` on purpose: the bioc
+  # helper's `available.packages()` probe must run under whatever
+  # `pkgType` the caller's session carried at entry -- pre-poisoning
+  # it with `pkgType = "source"` produces a different bioc-repo cache


Three adoptions from Copilot's review of #182: 1. R/val_build.R: rewrite the comment above apply_val_build_options() to accurately name which helper's available.packages() call is pkgType-sensitive. Copilot was right that `configure_bioc_repositories_if_requested()` doesn't call available.packages() -- it only installs a session-scoped `assignInNamespace('repositories', ...)` shim on BiocManager. The offender is the NEXT helper we run, in the same block: `configure_riskmetric_offline_if_requested()`, which does call available.packages() to build its offline cache (R/bioc.R). The fix (setting options AFTER both helpers) is unchanged; only the rationale needed correcting. 2. tests/testthat/test-apply_val_build_options.R: add two source-scan regression tests. The pkgType leak was a call-path bug, not a helper-signature bug -- the pre-existing unit tests can't catch a re-introduction of unconditional `options(pkgType = 'source')` at the val_pipeline() -> val_build() seam. New tests grep the installed R sources and assert: - val_pipeline.R's post-prep options() block does NOT set pkgType. - val_build.R sets options AFTER configure_riskmetric_offline_if_requested(). Test file: 17 -> 22 assertions, all pass. 3. dev/dev_rerun_low_covr.R: fix config_path fallback. Previously defaulted to `file.path(getwd(), 'config.yml')` which errors when the script is run outside a launcher project. Now defaults to NULL, which val_build() resolves to the packaged `inst/config.yml`. Also updated the config_path echo to print '<val_build default>' when NULL, so operators can eyeball what resolution they're getting without opening val_build(). Tests: `devtools::test(filter = 'val_')` -> 218/218 pass. Refs #181, #182. Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com>

aclark02-arcus · 2026-09-14T17:21:07Z

Adopted all 3 Copilot review suggestions in 0548949, plus a PR body refresh:

1. R/val_build.R comment — name the correct helper. Copilot was right on the factual sub-claim: configure_bioc_repositories_if_requested() doesn't call available.packages() — it only installs a session-scoped assignInNamespace('repositories', ...) shim on BiocManager. The offender is the NEXT helper in the same block, configure_riskmetric_offline_if_requested(), which builds an offline available.packages() cache (see R/bioc.R). Rewrote the comment above apply_val_build_options() to name the correct helper. The fix itself is unchanged — options still set AFTER both helpers.

2. tests/testthat/test-apply_val_build_options.R — source-scan regression tests. Copilot was right that unit tests on apply_val_build_options() alone can't catch a re-introduction of the call-path bug at the seam. Added two grep-level assertions on the installed R sources:

val_pipeline.R post-prep options block does NOT set pkgType.
val_build.R sets options AFTER configure_riskmetric_offline_if_requested().

Test file: 17 → 22 assertions, all pass.

3. dev/dev_rerun_low_covr.R config_path fallback. Was file.path(getwd(), 'config.yml'), which errors when the script is run outside a launcher project. Now defaults to NULL, which val_build() resolves to the packaged inst/config.yml. Echo now prints <val_build default> when NULL so operators can eyeball what resolution they're getting.

PR body refresh. Copilot's third comment on the same file noted the PR description contradicted the current code (it read as if options were set BEFORE the bioc helpers — that was the pre-course-correction narrative). Rewrote the body to reflect: (a) the current AFTER-the-helpers sequencing, (b) the 6d3d7a5 course-correction rationale, (c) the correct offender helper name, (d) the regression tests added post-review.

Verification:

devtools::test(filter = 'apply_val_build_options') → 22/22
devtools::test(filter = 'val_') → 218/218

Aaron's empirical result: entering val_build under getOption('pkgType') == 'source' regresses covr_coverage on some source-tier packages (logrx: >90% -> ~62%). Under 'both' or 'binary' the number holds at >90%. Intuition (source-tier assessment => pkgType='source') is exactly wrong. Likely mechanism: configure_riskmetric_offline_if_requested() runs near the top of val_build's body and calls utils::available.packages(), whose default 'type = getOption("pkgType")' filters the offline cache. Under 'source', binary-only pkgs from a PPM-style mirror drop out of the cache and downstream test-time installs fail silently -- the test suite short-circuits and covr reports the resulting partial coverage. 'both' keeps the cache complete (binary-preferred, source fallback) while 'ref = "source"' still drives val.pipeline's own source-tarball assessment path in val_pkg(). Changes: - R/utils.R: apply_val_build_options() sets pkgType='both' when ref='source', not pkgType='source'. Updated roxygen w/ the full rationale. - R/val_build.R: reworded the comment above apply_val_build_options() to explain the 'both, not source' direction and the offline-cache mechanism. - tests/testthat/test-apply_val_build_options.R: updated the primary source-tier test to start from pkgType='binary' (so the helper's mutation is observable) and assert pkgType=='both' post- call. Added a dedicated regression test asserting the helper NEVER sets pkgType='source' under ref='source' -- forces any future attempt to flip this back to the intuitive-but-wrong direction to fail loudly. - NEWS.md: expanded the 0.1.60 bullet to describe both prongs of the fix (removal of val_pipeline's unconditional set + the correction from 'source' to 'both' in val_build). Tests: devtools::test(filter = 'val_') -> 219/219 pass. Refs #181. Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com>

aclark02-arcus · 2026-09-14T18:26:45Z

Course correction in a57bdc9: the true fix is pkgType = 'both', not 'source'.

You reported empirically that entering val_build under getOption('pkgType') == 'source' reproduces the ~62% regression, and 'both' / 'binary' don't. That inverts the direction of the original fix on this branch. Previously:

apply_val_build_options(ref = 'source', ...) set pkgType = 'source' (the intuitive-but-wrong direction).
The PR argued the bug was about val_pipeline pre-poisoning parent-session state before val_build's own set.

New model:

Regardless of who sets it, pkgType = 'source' at val_build entry breaks the offline available.packages() cache that configure_riskmetric_offline_if_requested() builds. Any dep only present as a binary in a PPM-style mirror drops out; downstream test-time installs fail silently; covr short-circuits and reports the partial coverage as the package's "real" number.
'both' keeps the cache complete while ref = 'source' still governs val.pipeline's own source-tarball assessment path in val_pkg().

Applied in a57bdc9:

R/utils.R: apply_val_build_options() now sets pkgType = 'both' when ref == 'source' (was 'source').
R/val_build.R: reworded the comment above the call to explain the 'both, not source' direction and the offline-cache mechanism.
tests/testthat/test-apply_val_build_options.R:
- Updated the primary source-tier test to start from pkgType = 'binary' (so the helper's mutation to 'both' is observable) and assert getOption('pkgType') == 'both' post-call.
- Added a directional regression guard: apply_val_build_options(ref = 'source', ...) must NEVER produce pkgType == 'source'. Forces any future attempt to flip back to the intuitive-but-wrong direction to fail loudly.
NEWS.md: expanded the 0.1.60 bullet to cover both prongs.
PR body: rewritten to match the new root-cause model.

Behavior change for bare val_build(ref = 'source') callers: previously it set pkgType = 'source' on their session; now it sets 'both'. If your workbench default is 'source' (as yours appears to be — no Rprofile.site override), this is the change that actually delivers the >90% covr on logrx.

Tests: devtools::test(filter = 'apply_val_build_options') -> 23/23; devtools::test(filter = 'val_') -> 219/219.

Ready for one more workbench A/B on the low-covr cohort before merge.

Companion helper for the low-covr re-run driver (dev/dev_rerun_low_covr.R) and any other val_build(replace = TRUE) cohort: reports done / remaining / throughput / ETA by filtering assessed/<pkg>_<ver>_meta.rds by mtime relative to a caller-supplied run_started timestamp. The mtime filter is necessary because replace = TRUE overwrites metas in place -- a naive file-count is fooled by pre-existing metas from an earlier run. Inputs (via globals in caller's env): - pkgs : the cohort vector passed to val_build(). - run_started : POSIXct or 'YYYY-MM-DD HH:MM:SS' string; the kickoff moment (or Sys.time() - elapsed_hr if only the duration is known). - out, val_date: standard val_build layout knobs; the script derives val_dir = <out>/R_<Rver>/<YYYYMMDD>/. Outputs: - Console summary (cohort size, done, remaining, throughput, ETA). - `progress` data.frame in caller env, one row per cohort pkg, status in {done, pre_existing_meta_not_touched, no_meta}. - `remaining_pkgs` vector for follow-up. - Peek: 20 most-recent completions + first 20 remaining. Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com>

Companion to dev/dev_probe_run_progress.R + dev/dev_rerun_low_covr.R. Motivating scenario: the #181 fix (pkgType='both') lifts covr on some packages but may regress others; we need a thorough per-pkg delta to inspect regressions individually and, ideally, cluster them into a mechanism. Inputs (globals in caller env): - old_qa / new_qa: data.frame OR path to qual_assessments RDS. new_qa may also be a val_dir/ path, in which case the script walks assessed/*_meta.rds and reconstructs a qual_assessments- shaped frame -- important for runs kicked off with finalize = FALSE (top-level qual_assessments.rds not regenerated at end of assessment loop). - old_qm / new_qm: (optional) qual_metadata for decision-category diff. - pkgs: (optional) restrict diff to a cohort. - out_diff: (optional) output path; defaults to dev/covr_diff_<YYYYMMDD_HHMMSS>.rds (gitignored). Outputs (console): - Cohort membership (old-only / new-only / shared). - Metric column presence diff (catches silent config drift). - Coverage-band shift table (NA/0 through 90-100). - Improved/regressed/unchanged counts w/ +/- 5 pp thresholds. - Top-20 regressions + top-20 improvements. - Decision-category shift table (if qm supplied). Outputs (globals): diff_df, regressed_df, improved_df. .gitignore: add dev/covr_diff_*.rds pattern (output RDS may echo sensitive pkg names / assessment data). Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com>

Redesigned dev/dev_diff_run_assessments.R to support two workflow needs Aaron surfaced while inspecting a mid-run cohort: 1. Visual + shareable output. In addition to the console tables, the script now writes a self-contained HTML report to <out_dir>/covr_diff_<run_tag>.html -- inline CSS, no external assets, safe to email or drop in Slack. Green/red-highlighted delta cells; band-shift and decision-shift matrices rendered as HTML tables; per-section notes explaining what each list is for. run_tag defaults to a timestamp; can be overridden. 2. Rev-dep-weighted top-20 lists on top of the raw delta ranks: - UNLOCKERS: pkgs that crossed the 65% threshold UPWARD, ranked by rev-dep count. Highest rev-dep count = largest downstream slice of the qualified set unlocked by this pkg entering the qualified pool. These are the run's leverage wins. - BLOCKERS: pkgs still BELOW 65% after the run (regressed across the line OR were already below and stayed below), ranked by rev-dep count. Biggest downstream blockers surface first -- highest-value investigation targets for follow-up since unblocking them propagates. Rev-dep source priority: reverse_dependencies column on new_qa -> same column on old_qa -> fallback derived from qm's Depends/Imports/LinkingTo columns -> NA w/ friendly warning. Also added: - crossed_up / crossed_down / below_after boolean flags on diff_df so callers can slice however they want. - covr_status categorical ({improved, regressed, unchanged, lost_metric, new_only, both_na}) for easier filtering. - version_changed flag preserved from prior draft -- caller can eyeball which regressions coincide with upstream bumps vs which are pipeline-side. .gitignore: added dev/covr_diff_*.html to the existing rds pattern. Smoke test with synthetic frames: 50 shared pkgs, forced unlocker/regression rows -> HTML renders cleanly (13.4 KB), all five global data.frames populate, band-shift matrix and top-20 tables all print. Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com>

qual_metadata frames written by val_build() key on `pkg`, not `package` (that's qual_assessments's convention). The diff script hard-coded 'package' when subsetting qm for the decision-category diff, producing: Error: Can't subset columns that don't exist. Column `package` doesn't exist. Fix: new `.qm_key_col()` helper checks c('package', 'pkg', 'pkg_name') in order and picks the first hit. Applied at the qm merge site; the built decision-diff frame is normalized to 'package' internally so the downstream merge onto diff_df still keys correctly. Also drop a dead `names(counts) <- qm_src$package` line in the rev-dep qm-derivation fallback that likewise assumed 'package' and was never actually used (counts was unpopulated; dep_tbl is built independently from all_deps). Smoke tested with a qm frame keyed on 'pkg' + final_decision col: decision-shift table populates cleanly. Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com>

Bakes the covr-vs-pkgType mechanism write-up into the HTML report so it travels with the tables when shared with colleagues. Contents (styled as a highlighted callout box below the summary cards, above the delta tables): - Explicit statement that the ONLY pipeline-side change between the two runs is options(pkgType) flipping from 'source' to 'both' -- no metric defs, thresholds, decision rules, or {riskmetric} version changes. - Why covr doesn't care about the tarball: it instruments the install and runs the test suite in a subprocess, so target code is byte-identical either way. - Where the delta comes from: dependency install success. Comparison table of source vs both vs binary and their common failure modes. - Two amplifiers specific to val.pipeline + PPM: 1. available.packages() type filter narrows the offline cache. 2. Binary-only mirrors mean 'source' can render packages invisible even though they exist. - Why the number swings 20+ pp, not just a little. - Diagnostic tips for inspecting a specific regressor: read assess_record.rds's covr slot for 'no package called X' errors, grep val_pipeline.log for compilation failures. CSS: added a .context styleset (yellow callout, warm border, inner table with muted grid) so the write-up reads as 'meta / interpretation' rather than as another data table. Smoke-tested: HTML contains the section, all downstream tables still render. Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com>

aclark02-arcus and others added 2 commits September 14, 2026 07:57

aclark02-arcus requested a balanced review from Copilot September 14, 2026 17:06

Copilot started reviewing on behalf of aclark02-arcus September 14, 2026 17:06 View session

Copilot AI reviewed Sep 14, 2026

View reviewed changes

aclark02-arcus and others added 5 commits September 14, 2026 18:31

aclark02-arcus mentioned this pull request Sep 16, 2026

Graduate BioC blocklist helper into R/; expand write_qualified_pkg_lists() by default (#183) #184

Merged

Navigation Menu

Gate pkgType = 'source' on ref, own it in val_build() only (fixes #181) - #182

Gate pkgType = 'source' on ref, own it in val_build() only (fixes #181)#182
aclark02-arcus wants to merge 10 commits into
mainfrom
ac-181-gate-pkgtype-on-ref

aclark02-arcus commented Sep 14, 2026 •

edited

Loading

Uh oh!

aclark02-arcus commented Sep 14, 2026

Uh oh!

Copilot AI left a comment

Uh oh!

Uh oh!

aclark02-arcus commented Sep 14, 2026

Uh oh!

aclark02-arcus commented Sep 14, 2026

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

2 participants

Conversation

aclark02-arcus commented Sep 14, 2026 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Root cause recap

Likely mechanism (pkgType = "source" breaks source-tier covr)

What this PR does

Regression tests

What this PR intentionally does NOT touch

Verification

User-visible expected impact

Companion dev script

Version + NEWS

Uh oh!

aclark02-arcus commented Sep 14, 2026

Uh oh!

Copilot AI left a comment

Choose a reason for hiding this comment

🟡 Changes recommended

Suppressed comments (1)

Uh oh!

Uh oh!

aclark02-arcus commented Sep 14, 2026

Uh oh!

aclark02-arcus commented Sep 14, 2026

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

2 participants

aclark02-arcus commented Sep 14, 2026 •

edited

Loading

Likely mechanism (`pkgType = "source"` breaks source-tier covr)