Validate PR

validate-pr: Diff-Derived Test Plan

Generate and self-execute a validation plan matched to what actually changed in a PR. Replaces generic "tests pass" with area-targeted evidence and revert-test quality checks that prove tests catch regressions.

When To Use

• End of /fix-pr Step 5 (Validate), before Step 6 (Complete)
• Standalone after any PR fix, to generate targeted validation evidence
• When you need proof that revert-tests are genuine guards

When NOT To Use

• --scope minor with only formatting or doc changes (no logic changed)
• No diff available (clean branch, nothing changed)
• --skip-validate passed to /fix-pr

Algorithm

fetch diff -> group by area -> generate steps -> execute -> revert-test -> table

Step 1: Fetch Diff and Detect Areas

# Get changed file list from the PR
PR_NUMBER=<number from invocation or current branch>
CHANGED=$(gh pr diff "$PR_NUMBER" --name-only)
# Fallback when no PR number:
# CHANGED=$(git diff "origin/$(git rev-parse --abbrev-ref HEAD@{upstream})...HEAD" \
#   --name-only 2>/dev/null)

Group changed files into areas using ripgrep (grep if rg unavailable):

RUST_FILES=$(echo "$CHANGED"  | rg '\.rs$|Cargo\.(toml|lock)$' || true)
PY_FILES=$(echo "$CHANGED"    | rg '\.py$|pyproject\.toml$|requirements.*\.txt$' || true)
SH_FILES=$(echo "$CHANGED"    | rg '\.sh$|\.githooks' || true)
GRAMMAR_FILES=$(echo "$CHANGED" | rg '\.(lark|peg|g4)$' || true)

Area routing table:

Step 2: Generate and Execute Steps per Area

For each non-empty area, generate and run at least one verification step. Assign [E1], [E2], ... labels to each captured output.

Rust

# Build with default features
cargo build --workspace 2>&1
# Evidence: [En] → "0 errors, 0 warnings"

# Build with --all-features
cargo build --workspace --all-features 2>&1
# Evidence: [En+1]

# Per-crate test for each changed crate
# Extract crate directory from changed path, e.g. crates/token-types/src/lib.rs
CHANGED_CRATES=$(echo "$RUST_FILES" \
  | rg -o '(?:crates|src)/[^/]+' \
  | sort -u \
  | xargs -I{} basename {})
for CRATE in $CHANGED_CRATES; do
  cargo test -p "$CRATE" 2>&1
done

Python

# Targeted test per changed module
for PY_FILE in $PY_FILES; do
  MODULE=$(basename "${PY_FILE%.py}")
  TEST_FILE="tests/test_${MODULE}.py"
  if [[ -f "$TEST_FILE" ]]; then
    uv run pytest "$TEST_FILE" -v 2>&1
  fi
done

# Or project-specific runner if Makefile target exists
make test 2>&1 || uv run pytest tests/ -v 2>&1

Shell

for SH_FILE in $SH_FILES; do
  [[ -f "$SH_FILE" ]] && shellcheck "$SH_FILE" 2>&1
done

Build/config parse check

# YAML files
for YML in $(echo "$CHANGED" | rg '\.ya?ml$' || true); do
  [[ -f "$YML" ]] && python3 -c "import yaml; yaml.safe_load(open('$YML'))" \
    && echo "PASS: $YML" || echo "FAIL: $YML"
done

# JSON files
for JSON_F in $(echo "$CHANGED" | rg '\.json$' || true); do
  [[ -f "$JSON_F" ]] && python3 -m json.tool "$JSON_F" > /dev/null \
    && echo "PASS: $JSON_F" || echo "FAIL: $JSON_F"
done

Step 3: Revert-Test Quality Check

Prove at least one test is a genuine guard, not a dead assertion.

Safety: abort if the working tree has uncommitted changes.

if ! git diff --exit-code > /dev/null 2>&1; then
  echo "[RT] SKIP: working tree dirty: revert-test unsafe"
  # Mark INCONCLUSIVE and continue
fi

Algorithm (one representative fix):

1. From the changed source files, find one that has a corresponding test.
   • Rust: a #[test] in the same crate that exercises a changed function.
   • Python: tests/test_<module>.py for a changed <module>.py.
   • Shell: a test harness that invokes the changed script.
2. Identify the specific changed line or block from the diff.
3. Edit that line to revert the fix to its broken state.
4. Run the targeted test: confirm it FAILS with exit code 1 specifically. A pytest usage error (4) or an empty collection (5) is also non-zero, so a harness that only checks "not zero" reports a dead assertion as a genuine guard.
5. Restore: git checkout -- <file> (git-based restore, safe on interrupt).
6. Run the targeted test again: confirm it PASSES.
7. If any step cannot complete, mark INCONCLUSIVE with the reason.

Revert-test output format:

[RT-1] Target: <file>:<line>: <description of fix>
[RT-2] Broke fix: <edit description>
[RT-3] Ran: <test command> → <test name> FAILED (expected)
[RT-4] Restored: git checkout -- <file>
[RT-5] Ran: <test command> → <test name> PASSED
Result: PASS: test is a genuine guard

Reverting a test that guards document content:

Content tests assert on prose, and this repo wraps prose at 80 columns, so any anchor phrase long enough to be meaningful eventually straddles a line break. Collapse whitespace before matching. Otherwise a pure reflow turns the test red and tempts an author to "fix" it by unwrapping the line.

Normalizing reintroduces the hazard the revert test exists to catch: a rejoined anchor can also appear elsewhere in the file, so deleting the paragraph the test guards leaves it green. Anchor on a full clause that is unique to that paragraph, then delete the paragraph and confirm the test goes red. A DDD paradigm test passed its revert check this way in PR #612 while guarding nothing.

When no covering test exists:

Revert-test: INCONCLUSIVE: no covering test for <changed area>
Recommendation: add a test for <changed function or behaviour>