The tightened behavior below is available in v0.24.0+. Review new removal signals when upgrading from v0.23.0; numeric scoring and policy defaults are unchanged.
diff and compare now reuse the instruction text captured by each underlying
scan for removal and directive-polarity checks. They do not reopen instruction
paths after scanning. Previously that second read could silently skip an I/O
failure, see different text, or follow an alias replaced after its validated
scan read. A report could consequently miss a reversal or invent a removal.
The private scan helper shares captured text only within the comparison. Raw
source is not attached to Report or RegressionReport, their dataclass
representations, or their exported JSON. Existing finding messages and paths can
still contain repository details; ordinary reports are not privacy-minimized.
Initial discovery, resolution and source-read failures remain operational errors:
exit 2, no success JSON or newly written success report. Sources still use the
instruction-file boundary.
Reusing captured text does not weaken initial source validation.
This is not an atomic repository snapshot. Base and head are scanned sequentially, files are read at different times, and configuration, ignore rules and path-existence checks have their own reads. A file changed or deleted after capture does not retroactively alter the report's captured instructions. Keep both trees stable during comparison and rerun after changes. No complete input snapshot, concurrent-change detector or security attestation is claimed.
instruction-file-removed is an error-level, regression-only review signal.
A disappeared supported filename is exempt only when:
- A newly added supported filename in the same repository-relative parent directory has the same nonempty captured decoded text.
- That new file has not already paired with another removal. Matching is one-to-one, with removed paths processed in sorted order.
Normal text reading converts line endings to LF, so CRLF versus LF alone does not prevent a match. Spaces, indentation, blank lines and literal/code-block whitespace are otherwise significant. This compares decoded text, not raw bytes; instruction decoding retains the scanner's existing UTF-8 replacement behavior.
An existing duplicate is not a move destination. Neither is a modified filename that already existed in the baseline. One new file cannot excuse deletion of two old files. Moving a root instruction into a subtree, moving between sibling directories, or moving a nested instruction to the root still reports the old filename's removal. Changed directory applicability must be reviewed, not assumed equivalent merely because the words match.
This remains a filename-coverage heuristic, not full agent-scope analysis.
Same-directory root CLAUDE.md to AGENTS.md migrations retain their existing
exact-text exception; agent consumers, format semantics, override precedence and
frontmatter/glob applicability are not certified equivalent. Legitimate
cross-directory migrations or content merges may therefore need a reasoned,
expiring, baseline-reviewed suppression. The candidate cannot grant its own
exception. No new rule ID, numeric penalty, suppression format or policy default
is introduced.
The scan scores for both sides may still be 100 while a removal or exact
directive reversal appears in new_findings. Regression-only rules do not
change standalone numeric scores. Enable policy.fail_on_new_errors in the
trusted baseline or pass --fail-on-new-errors to block them with exit 1.
init generates that baseline policy; the Action uses it and defaults to new-error
blocking when no policy exists. The bare CLI without a policy remains score-only
by default; this compatibility default has not silently changed.
Tests include synthetic read-error/replacement cases, external-link replacement, literal-whitespace controls, one-to-one move matching, scoped moves, retained same-directory moves, privacy and baseline suppression. Installed-package acceptance checks scoped removal and existing-copy deletion with the actual baseline gate. These are development regressions, not independent real-world accuracy measurements or user feedback.