fix(tools): guard against catastrophic content loss in background review patches

[AlexFucuson9]

Summary

Adds a content-loss guard in _patch_skill to prevent the background review agent from silently corrupting skills via hallucinated patches.

Problem (P1 #55647)

The background review agent can call skill_manage(patch) on a skill it has not read via skill_view. When the model hallucinates old_string from the conversation transcript alone, fuzzy matching may find a partial match and the replacement silently deletes most of the file. Observed: 807 lines → 186 lines in a real session.

Fix

When _patch_skill is called from a background review context (is_background_review()), reject patches that would shrink the target file by more than 50%. The error message instructs the agent to read the skill first with skill_view before making targeted patches.

Changes

• tools/skill_manager_tool.py: Add content-loss guard after size validation in _patch_skill (26 lines added)

Fixes #55647

[alt-glitch]

This was generated by AI during triage.

Competing with #55657 for the same bug (#55647). This PR adds a >50%-shrink content-loss guard in _patch_skill for background-review context; #55657 enforces a read-before-write invariant (refuse autonomous patch unless the file was loaded via skill_view this turn). #55657 addresses the root cause (no read-before-write gate) while this PR is a defensive size heuristic. Both open against the same issue — flagging for a maintainer to pick the canonical fix.

[teknium1]

Closing in favor of #55906 (merged), which fixes the root cause of #55647 with a read-before-write invariant: the background-review fork must skill_view a skill before skill_manage can patch it. That prevents the corruption structurally rather than via a size-delta heuristic.

Your >50%-shrink guard was a reasonable backstop, but it can't catch sub-50% corruption and the arbitrary ceiling would also reject legitimate large condensations. The invariant approach supersedes it. Thanks for jumping on this fast.

#55906