Repository navigation
Conversation
…gine-us scorp-e2e.yml upgraded policyengine-us to 2.10.1+ on top of the dashboard requirements in both jobs, because passive mode was unsafe on older releases when #1199 added the workflow. #1411 backported the EITC investment-income fix (pe-us#9572) to every older release, so the override is no longer needed, and with it the workflow never ran the configuration the dashboard refresh uses: policyengine-us 2.6.17 plus that backport. Run both jobs on a two-release matrix instead: the dashboard environment as pinned, and the latest release. The pinned leg fails if the installed policyengine-us is not the pinned one. Each alignment report now records the policyengine-us and policyengine-core versions, and the workflow also runs when the S-corp classification, its reform or the pinned requirements change. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
|
The latest updates on your projects. Learn more about Vercel for GitHub.
|
…backport The interface tests had no case whose result depends on the EITC investment-income backport, so the dashboard-pin leg would pass even if the backport were skipped. Case 6 (wages 8,000, interest 12,500, S-corp -6,000) gets no EITC from TAXSIM; policyengine-us 2.6.17 without the backport nets the passive loss against the interest and gives 352.18. Compare v25 as well, so the EITC difference is reported directly. Checked locally: 6 of 6 pass on 2.6.17 and on 2.38.19; with the backport disabled on 2.6.17, the three passive-mode tests fail with fiitax -352.18 against TAXSIM's 0. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Review of 7d348f0 (GPT-6.1 Sol, APPROVE with non-blocking findings): - The pin check only compared policyengine-us, and the interface job installs test packages after the locked set. Add scripts/check_pinned_environment.py, which fails unless every requirements pin is installed at its version (it tolerates markers and hash continuations), and run it plus pip check, as the dashboard refresh does. Locally it passes on the pin with the test packages and lists 19 differing packages on the latest-release environment. - Run the workflow when scripts/refresh_dashboard.py (imported by the alignment script) or the new check changes. Runner, CLI, API and mapping changes stay covered per PR by the scorp-validation unit job, as the docs now say. - Fix the max-parallel comment: the cap is four jobs in total, not two per release. - Docs: the interface assertions now include the EITC; explain case 6's 352.18 (26 USC 32(a)(2) limitation, Rev. Proc. 2024-40 section 3.06), which the review had recomputed without the limitation. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
CI on 51ff5d3: every dashboard-pin job passed (alignment 2021-2025 and interfaces), and pip check passed on both releases. alignment (2021, latest) failed: the first 500 households of 2021 exceeded the 5 GiB worker limit on policyengine-us 2.39.1. The same batch peaked at 4,280 MiB on 2.17.0 (September 29 run) and 4,765 MiB on the pin today, and measures 5,599 MiB on 2.40.1 locally. Give the latest leg an 8 GiB limit through a new --max-memory-gb option (default 5, the refresh's setting); hosted Linux runners have 16 GB. The pin keeps 5 GiB. Second review (APPROVE, non-blocking): - Install the interface job's test-only packages under the locked versions as constraints on the pin, so a new modal or fastapi release cannot move a locked package. - check_pinned_environment.py now parses lines with packaging, evaluates environment markers, compares versions under PEP 440 and raises on any requirement line that is not an exact pin. Add a round-trip property test for the parser and marker, rejection and normalization tests. - Run the per-package check before pip check, and run pip check even when the per-package check fails, so both diagnostics appear. - Trigger on variable_mappings.yaml; correct the docs on which refresh settings the pin matches. The macOS 3.11 job failed on the known wall-clock flake (#1429) and a child-process segfault in the hash-seed test, which this PR does not touch; the segfault is split out as a separate task. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Third review (GPT-6.1 Sol, APPROVE at d54a9f2, non-blocking): - check_pinned_environment.py skipped every indented line and every "--" option, so an indented requirement or "--requirement" include went unchecked, and it rejected valid same-line hashes, tab-separated comments and "-i". It now strips comments and hash options, skips only options that choose a package source, checks indented requirements, and raises on includes, constraints and editables. Tests cover each. - Print the installed policyengine-us and policyengine-core with pip list, so a missing distribution cannot stop pip check from running. - The EITC amounts are in Rev. Proc. 2024-40 section 2.06 (SECTION 2, "2025 ADJUSTED ITEMS"), not 3.06. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Fourth review (GPT-6.1 Sol, APPROVE at 01fc36e, non-blocking): the hash regex could rewrite a package name containing "--hash" or a quoted marker string, silently dropping a pin, and a source option followed by -r on the same line was skipped. pins() now joins continuation lines, removes comments by pip's rule, and takes a line's requirement as its tokens up to the first one starting with "-", as pip's break_args_options does. A requirement may carry only --hash options; a line without one may carry only source options. Any other option, wherever it appears, raises. Tests cover both review cases and -r/-e/--global-option after other tokens. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
This branch was successfully deployed
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
1 participant
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
What this changes
.github/workflows/scorp-e2e.ymlinstalled the dashboard requirements and then upgraded policyengine-us to 2.10.1 or later in both jobs. #1199 added that upgrade because passive mode was unsafe on older releases. #1411 backported the EITC investment-income fix (pe-us#9572) to every older release, so the upgrade is no longer needed. With it in place, the workflow never ran the configuration the dashboard refresh uses: policyengine-us 2.6.17 plus that backport.Both jobs now run on a two-release matrix:
dashboard-pin: the locked dashboard refresh environment (scripts/dashboard-refresh-requirements.txt, policyengine-us 2.6.17) with the refresh's 5 GiB worker memory limit. On the pin, the interface job installs its test-only packages with the locked versions as constraints. The newscripts/check_pinned_environment.pythen fails the job if any of the 86 pinned packages is installed at another version, andpip checkruns as it does in the refresh.latest: the same requirements with policyengine-us upgraded to the newest release, as before. Its alignment workers get an 8 GiB memory limit (see below).The PR also:
tests/test_scorp_e2e.py(wages 8,000, interest 12,500, S-corp −6,000, TY2025) and comparesv25. Before this, no end-to-end case depended on the backport, so the pinned leg would have passed even if the backport were skipped. The expectation comes from the bundled TAXSIM binary, which gives no EITC.scorp.py,scorp_reform.py,scripts/refresh_dashboard.py, the new check script, the pinned requirements andvariable_mappings.yaml. Backport the EITC investment-income fix to older policyengine-us and default scorp to passive everywhere #1411 changed the first two, and the workflow did not run on it. Runner, CLI and API changes stay covered on every PR by the main CI'sscorp-validationjob, which runs on the pin. The docs say so.--max-memory-gbtoscripts/validate_scorp_alignment.py, defaulting to the refresh's 5.docs/scorp-validation.mdto say which policyengine-us versions are covered, and how.CI evidence
Run 38193095171 at 51ff5d3 (before the last round of fixes):
The 2021 latest failure comes from policyengine-us using more memory, not from this PR. The same 2021 batch peaked at 4,280 MiB on 2.17.0 in the September 29 run, and 2022–2024 grew by about 1.1 GiB between 2.17.0 and 2.39.1. Locally on macOS, the first 2021 batch peaks at 4,895 MiB on 2.6.17 and 5,599 MiB on 2.40.1. Hosted Linux runners have 16 GB, so the latest leg now gets 8 GiB. The pin keeps the refresh's 5 GiB. Its 2021 workers have the least headroom: 4,765 of 5,120 MiB.
That run's
test (macos-latest, 3.11)failure is not from this PR. It is the wall-clock flake in #1429 plus a child-process segfault intests/test_deterministic_ordering.py, which is being looked at separately.Local evidence (macOS, Python 3.11, bundled TAXSIM)
SCORP_E2E=1 pytest tests/test_scorp_e2e.py:EITC_BASKETS_MIN_PE_VERSION = (0, 0, 0))test_real_cli[passive],test_stdin_default_is_passive,test_real_http[False-passive], eachfiitax−352.18 vs TAXSIM 0 on case 6Without the backport, case 6's $352.18 EITC follows the statute once the S-corp loss nets investment income down to $6,500:
The credit is min(7.65% × 8,000 = 612, 649 − 7.65% × (14,500 − 10,620) = 352.18) = 352.18. PE 2.6.17's
eitc_phased_in(612),eitc_reduction(296.82) andeitc(352.18) match. The first reviewer recomputed it as $315.18, and the second review accepted this correction.check_pinned_environment.py: 14 tests pass on Python 3.11 and 3.10, including Hypothesis round-trip and mismatch properties. It passes on the pinned environment after the constrained test-only installs (86 pins), and lists 19 differing packages on the latest-release environment.Reviews
Invariants
scripts/dashboard-refresh-requirements.txtis installed at its pinned version, checked after every install step.pins()round-trips: any set of exact pins written in uv's format (markers, hash continuations, comments) parses back to the same mapping, and pins for another interpreter are skipped (property test).mismatches()flags exactly the packages that are changed or missing (property test).axiom: n/a: CI, test and documentation only, no policy change.
🤖 Generated with Claude Code