feat: smarter diff filtering for minDiffLines + pass cache (hook v1.2.0)

[darrylmorley]

Fixes #1

Problem

Two sources of wasted/misleading QA runs:

1. The minDiffLines skip threshold counted every non-empty line of the diff — including diff --git/index/@@ metadata and unchanged context lines — so it barely reflected meaningful change. Import reordering could trigger a review while a one-line logic fix could still be dominated by noise.
2. Every Stop re-reviewed the full uncommitted diff even when nothing had changed since the last verdict — e.g. asking Claude a question while sitting on an already-reviewed dirty tree paid for a full Ollama inference again.

Changes

hookqa-hook.ts (v1.0.0 → v1.2.0)

Smarter diff counting (#1):

• Replaced countDiffLines with countMeaningfulDiffLines, which counts only +/- change lines (file headers +++/---, metadata, and context lines are ignored).
• Filters out noise before counting: blank lines, comment-only lines (//, /* */, * continuations, # style, <!-- -->), and import-only lines (JS/TS/Swift/Python/Java import, Python from … import, JS re-exports, CommonJS require, Ruby require, C/ObjC #include/#import, C# using, Rust/PHP use, Swift @testable import).
• Whitespace-only edits and pure line moves cancel out: within each hunk, removed/added lines with identical whitespace-normalized content are matched as multisets and contribute 0. Moves across hunks or files still count.
• Careful edges: #include is filtered as an import, but #define and #! shebangs still count as meaningful; *ptr = x; is not mistaken for a block-comment continuation.
• SKIPPED log entries now read Diff too small (N meaningful lines, min M).

Pass cache (skip redundant reviews):

• After a review that doesn't block, the hook stores a SHA-256 fingerprint of the diff plus every config field that can change the outcome (model, temperature, weights, custom instructions, blockOnWarnings) in a per-project file under /tmp (same lifecycle as the existing retry files).
• On the next Stop, an identical fingerprint skips the Ollama call entirely and logs SKIPPED: Unchanged diff already reviewed — skipping.
• Any config change or diff change invalidates the cache; ERROR and blocking reviews never populate it, so a FAIL is always re-reviewed after fixes.

Structural:

• The entry point is now guarded by import.meta.main so the pure functions are importable by tests without executing the hook. Runtime behavior is unchanged and the never-block-Claude-Code exit-0 safety path is preserved. Still fully standalone — Bun built-ins only.
• Version banner bumped to v1.2.0, so installed copies show the update prompt in the app's Hook tab.

tests/hookqa-hook.test.ts (new)

26 bun test cases: realistic multi-file diffs, per-language import filtering, whitespace cancellation, move semantics within/across hunks, multiset handling of duplicate lines, comment/preprocessor edge cases, fingerprint determinism and sensitivity (changes for diff/model/temperature/weights/instructions/blockOnWarnings; stable for unrelated fields like timeout and logging), and cache-path hygiene.

BehaviourTab.swift

Updated the "Min Diff Lines" slider description to reflect the new semantics (imports, comments, and whitespace don't count).

Behavioral note

Meaningful-line counts are lower than raw counts, so an unchanged minDiffLines value is now effectively stricter — more trivial stops will skip QA. Combined with the pass cache, repeat stops on an unchanged tree no longer hit Ollama at all. Both are the intent of #1, but existing configs will notice the shift.

Testing

• bun test tests/ — 26 pass, 0 fail.
• Smoke-ran the hook end-to-end via stdin (exits 0 cleanly with no model configured).
• Verified importing the script has no side effects (guard works).
• Swift change is a one-line string edit; not compiled here (no macOS toolchain in this environment) — worth a local xcodebuild before release.

🤖 Generated with Claude Code

https://claude.ai/code/session_01GaLnfPXe58Gthpy6RoCwMJ