Self-audit a PR against patterns Copilot has historically flagged in this repo. Run AFTER staging the diff but BEFORE pushing. Covers the 23 LESSONS themes accumulated across PR1-PR7 of the Portfolio Backtest Engine refactor. Use when you're about to push a branch, open a PR, or address Copilot review comments.
インストール
Codex または Claude でインストール この Prompt をコピーして Codex、Claude、または他のアシスタントに貼り付けると、Skill ページを確認してインストールできます。
Self-audit a PR against patterns Copilot has historically flagged in this repo. Run AFTER staging the diff but BEFORE pushing. Covers the 23 LESSONS themes accumulated across PR1-PR7 of the Portfolio Backtest Engine refactor. Use when you're about to push a branch, open a PR, or address Copilot review comments.
PR self-review
You are about to push a branch / open a PR. Run through this checklist
against the staged diff before pushing. Each item references a LESSONS.md
theme; if you fail any, fix it now — Copilot would have flagged it
post-push and the round-trip is wasteful.
The full distilled rules live in CLAUDE.md at the project root. This
skill is the operational checklist — short, scannable, with a per-item
diagnostic command where applicable.
Step 1 — Survey the diff
git diff --stat# files touched
git diff main...HEAD # full content diff vs main
git log main..HEAD --oneline # commit list
Note any new file > 200 lines, any cross-cutting change (signature
changes touching 3+ files), or any new public symbol. Those are where
the patterns below tend to bite.
Both must exit 0. The CI lint job runs the same commands on every
push — fixing locally avoids the round-trip.
Step 3 — Run the suite
.venv/Scripts/python.exe -m pytest -q
Must end with N passed, 0 warnings. A new RuntimeWarning means the
code path emits expected-but-noisy output that should be silenced
(warnings.catch_warnings() + targeted filter, like src/fire.py
already does for the All-NaN slice case).
Step 4 — Walk the 23-theme checklist
Mark each ✅ if you've actively verified, ⏭️ if not applicable to this
PR. Don't ⏭️-tag without thinking — the cost of false-skipping is the
exact pattern that bit us in PRs 1–7.
Marketing copy vs. capability
README claims match the code's actual delivered surface in
this PR. No "users can build any portfolio" if --portfolio only
accepts the default preset. (Theme 1, PR15/PR16/PR18.)
Streamlit st.title / st.markdown / help= strings
describe behavior the user can actually invoke. (Theme 1.)
Module / class / method docstrings match the symbol's
current signature, return shape, and exception list. (Theme 17,
PR20.)
When deleting a public helper, every from x import Foo
callsite is also deleted or # noqa: F401-marked with reason.
(Theme 11.)
When deleting a feature, the corresponding test file is
deleted too — don't leave tests asserting against stubs.
(Theme 11.)
Function signature changes
Transitive callers audited. Grep for the symbol in every
module under src/, tests/, root entry points. Any helper
reaching the symbol via a None-default fallback to a global has
been threaded with the override. (Theme 21, PR21.)
Branch decisions read from CLI/config spec, not display
names. A custom Portfolio with name="Four Umbrellas" doesn't
bypass any "is this default?" check. (Theme 4, PR16.)
Type hints match defaults.field: Foo = None is wrong;
field: Foo | None = None is right. (Theme 5.)
Per-field type coercion at the parser. String for float
rejected. bool for int rejected (bool is an int subclass).
List of wrong length rejected. (Theme 22, PR21.)
Allowlist of permitted fields maintained as a constant.
Unknown fields produce ValueError naming the offending key.
Don't rely on dataclasses.fields() as the schema — that includes
fields with no runtime effect that shouldn't be persistable.
(Themes 3, 23, PR16/PR21.)
All expected exceptions caught at the CLI entry point.FileNotFoundError, ValueError, tomllib.TOMLDecodeError,
json.JSONDecodeError, etc. Map to sys.exit(2) + clean stderr
message. (Themes 3, 11, 15.)
Help text matches argparse default. If argparse default
is None and downstream resolves it, say that in help=.
(Theme 6, PR16.)
Source of truth = live data, not metadata cache. When the
same fact is in both a manifest and the bundle, derived
computations read from the bundle. (Theme 12, PR19.)
Streamlit changes
Computed values are applied, not just displayed. A yellow
warning showing effective_start must mean the engine call uses
it. (Theme 7, PR18.)
UI options filtered to backend surface. Asset pickers
intersect with bundle.monthly_returns_eur.columns. Submit-time
re-validation present. (Theme 9, PR18.)
st.session_state["last_run"] snapshots run-time params.
Post-run renders read from the snapshot, never from live widget
state. (Theme 14, PR19.)
New widgets are end-to-end-wired. Grep for the consuming
function. There's a test reading back what the engine actually
used. (Theme 8, PR18.)
No-mapping fallback handled. Loading a Portfolio whose
rebalance_months doesn't match a UI radio choice falls back to
a default + emits st.warning. (Theme 20, PR20.)
No bare tracebacks. Engine calls wrapped in
narrow-scoped try/except mapping to st.error /
st.warning. (Theme 15, PR19.)
CLI changes
python scripts/check_ascii_print.py ... exits clean for
every CLI module touched. (Theme 18, PR20/PR21.)
Exit codes: 0 for success, 2 for input errors, 1 reserved
for runtime failures. (Theme 11.)
Test changes
Test docstrings describe what the assertion enforces —
not what you wished the engine did. (Themes 10, 16.)
Wrong-type-per-allowed-field matrix present for new
parsers (string for float, bool for int, list of wrong length,
etc.). (Theme 22.)
No hardcoded display names in assertions where the
runtime label is dynamic. (Theme 16, PR19.)
Step 5 — Self-summary
After running the checklist:
State which themes you verified and which you skipped (with
reason).
State whether the diff preserves the byte-identical guarantee on
python backtest.py --synthetic. If you touched src/rebalance.py,
src/options_overlay.py, or src/portfolio.py, run:
