[Lanre Shittu (Shizoqua)]

Summary

Fixes #372.

A step-level replace or delete edit whose target starts in editable skill text and runs into a protected region's start marker was applied. That removed <!-- SLOW_UPDATE_START --> (or <!-- APPENDIX_START -->) and left the protected content in the editable body. An insert_after with such a target wrote into the protected region. These edits are now skipped with skipped_protected_region, the same as edits that target text inside the region.

User impact: without this fix, a training run can silently lose its slow-update or appendix guidance: the marker disappears, the protected text becomes editable body text, and later steps can rewrite or duplicate it.

Proof: the regression test fails on upstream eb959c3 and passes with this change.

Root cause

_is_in_protected_region() in skillopt/optimizer/skill.py only tested whether the target's start index was inside a region:

if start_idx <= target_idx < region_end:

A target that begins before the start marker and extends into it fails that test. _apply_edit_with_report() then runs skill.replace(target, ..., 1) across the marker.

Changes

• _is_in_protected_region() now treats a target as protected when its span [target_idx, target_idx + len(target)) overlaps [start_idx, region_end). A target that ends exactly where the region starts does not overlap and still applies.
• _apply_edit_with_report() is the only caller, so this covers the trainer path (apply_patch_with_report) and the exported apply_edit / apply_patch.
• Regression test in tests/test_skill_aware_reflection.py next to the existing protected-region tests, covering both regions and all three targeted ops, plus an adjacent edit that must still apply. It is also registered in the file's standalone main() runner.
• CHANGELOG entry under Unreleased / Fixed.

Reproduction

from skillopt.optimizer.skill import apply_patch_with_report
from skillopt.optimizer.slow_update import (
    has_slow_update_field, inject_empty_slow_update_field, replace_slow_update_field,
)

skill = replace_slow_update_field(
    inject_empty_slow_update_field("# QA Skill\n\n## Rules\n- Answer briefly."),
    "Prefer exact spans.",
)
edit = {"op": "delete", "target": "- Answer briefly.\n\n<!-- SLOW_UPDATE_START -->"}
new, report = apply_patch_with_report(skill, {"edits": [edit]})
print(report[0]["status"], has_slow_update_field(new))

Before: applied_delete False, and the result is '# QA Skill\n\n## Rules\n\nPrefer exact spans.\n<!-- SLOW_UPDATE_END -->\n'.

After: skipped_protected_region True, and the skill is unchanged.

Testing

• python -m pytest tests/test_skill_aware_reflection.py -q: the new test fails on upstream eb959c3 with AssertionError: delete across <!-- SLOW_UPDATE_START --> must not change the skill, and passes with the fix.
• python tests/test_skill_aware_reflection.py: 7/7 passed.
• python -m pytest tests/test_skill_aware_reflection.py tests/test_semantic_density.py tests/test_sleep_engine.py -q: 158 passed, 1 skipped.
• python -m pytest -q: 2034 passed, 12 skipped, 4 xfailed.
• python -m mkdocs build --strict: passed.
• python -m ruff check on the changed files: no new findings. The 6 import-order findings are on lines that are unchanged from main.

Risk and compatibility

Only targets that overlap a protected region change behavior, and those previously broke the region. Edits fully outside the regions, append, and insert_after fallbacks are unchanged; the existing byte-identical and coexistence tests still pass. No backend, role, config key, file format or CLI change.

[Yifan Yang (Yif-Yang)]

Merged, thank you. _is_in_protected_region now rejects any target that overlaps a protected region, not only one that starts inside it, so delete/replace/insert_after can no longer remove a SLOW_UPDATE/APPENDIX start marker or write inside the region. I reproduced #372 on main and confirmed all three ops are now skipped with the skill unchanged. The new test fails on main and passes here, and the full suite passes. Closes #372.

Note: your branch and #375 each added an entry at the same place in CHANGELOG.md. After #375 merged, I pushed one maintainer merge commit to your branch that keeps both entries and changes nothing else, then merged.