Semantic SZZ Analyzer

This skill is a delta over → szz-bug-identifier. Run classic SZZ first; this skill filters and re-ranks its candidates using semantic understanding instead of line-level blame.

Classic SZZ's precision problem: git blame is textual. It tells you who last touched a line, not who last changed its meaning. A variable rename, an indent, a refactor that moves a line unchanged — all of these become false-positive bug-introducers.

Semantic filters — applied on top of classic SZZ output

Re-ranking

When multiple candidates survive filtering, rank by:

1. Semantic distance: how much did the candidate change the behavior of the fixed line? A commit that added the line beats one that renamed a variable in it.
2. Temporal proximity to bug report: a commit 2 days before the report beats one from 2 years before.
3. Author signal: if the fix author is also the candidate author, slight boost — they probably know what they broke.

Worked example

Fix: Adds a null check on user.profile before dereferencing.

Classic SZZ candidates:

• a1b2c3d — "Rename p → profile" (6 months ago) — blame hits this because the dereference line was touched
• e4f5g6h — "Remove unused profile validation" (3 weeks ago) — blame misses this; it touched a different line
• i7j8k9l — "Add profile feature" (1 year ago) — the original dereference

Semantic pass:

• a1b2c3d: AST-diff → rename only, no behavior change → dropped. Re-blame through it.
• Re-blame lands on i7j8k9l. But: bug was reported 3 weeks ago. i7j8k9l is 1 year old — survived a year without a report?
• Bug-pattern match: the fix adds a guard. Did any recent commit remove a guard? → e4f5g6h removed validateProfile() which checked non-null.

Verdict: e4f5g6h is the true introducer. Classic SZZ missed it entirely because the fix and the removal touch different lines.

Edge cases

• The fix and the introduction touch disjoint code: Classic SZZ is blind here; semantic SZZ can catch it only if the def-use chain connects them. If not, neither algorithm finds it.
• Introduced by omission: "The bug is that a check was never added." No commit to blame — the closest you get is the commit that added the code around where the check should be.
• Squash-merge repos: All bugs blame to the squash commit. You need the pre-squash branch history, or you're stuck.

Do not

• Do not use semantic SZZ as your first pass. It's an order of magnitude slower. Classic SZZ → filter → semantic re-rank on survivors.
• Do not override a classic-SZZ hit with a semantic-SZZ miss. Semantic SZZ adds candidates the textual blame missed; it shouldn't silently drop direct hits.

Output format

Same as szz-bug-identifier, with an additional semantic_evidence field per candidate:

fix: <sha>
  candidates (semantic-filtered):
    <sha>  confidence=<high|med|low>
      semantic_evidence: <AST change summary / def-use link / pattern match>
      classic_szz_hit: <yes|no — found via semantic reroute>