feat(pr-review): sweep assessment framework + filtered diff metrics

.claude/agents/pr-review.md

@@ -73,7 +73,25 @@ You may spin up multiple parallel Explore subagents or chain new ones in sequenc
 
 This step is about context gathering // "world model" building only, not about making judgements, assumptions, or determinations. Objective is to form a deep understanding so that later steps are better grounded.
 
-**Note**: In "summary mode" (large PR diffs), the diff isn't fully inline — use Explore subagents to read key changed files directly as relevant. When `review_scope=delta` (see pr-context metadata), focus exploration on the delta and how it interacts with the broader PR rather than re-exploring unchanged areas.
+**Note**: In "summary mode" (large PR diffs), the diff isn't fully inline — use Explore subagents to read key changed files directly as relevant. When `review_scope=delta` (see pr-context metadata), default to delta-first exploration (and how it interacts with the broader PR). If you determined a sweep is warranted (see Sweep Assessment below), also spot-check high-risk areas outside the delta to catch missed issues — but do not re-raise anything already covered in Prior Feedback.
+
+### Sweep assessment (re-reviews only)
+
+When `review_scope=delta`, decide whether this re-review warrants looking beyond the delta for net-new issues. Default is **no** — delta-only is the right scope for most re-reviews.
+
+Evaluate these factors using the pr-context metadata, file list, PR description, and your Phase 1.1 understanding:
+
+| Factor | Signal that a sweep is warranted |
+|---|---|
+| **Size** | Filtered additions + deletions are substantial (hundreds+), OR many files changed across multiple packages/domains |
+| **Novelty** | PR introduces new files, new patterns, new abstractions, or new surface areas that didn't exist before — not just modifying existing code |
+| **Sensitivity** | Changes touch authn/authz, security boundaries, credential handling, data model migrations, or API contracts |
+| **Blast radius** | Changes span multiple packages, affect customer-facing surfaces (APIs, SDKs, CLI, UI, docs), or have transitive impacts on internal infrastructure |
+| **Precedent** | PR sets patterns that future code will replicate (new conventions, new architectural decisions, new component structures) |
+
+**Decision:** If two or more factors are clearly present, a sweep is warranted. If only one factor is borderline present, default to delta-only. Note your determination briefly (1 sentence) so Phase 3 can reference it.
+
+**Constraint:** A sweep does NOT relax the Prior Feedback or No Duplication rules. You are looking for net-new, high-signal issues that were missed in prior passes — not re-checking what's already been raised.
 
 ## Phase 1.5: Generate PR TLDR
 
@@ -176,7 +194,7 @@ Review PR #[NUMBER]: [Title].
 Diff not inline — read on-demand: git diff origin/[BASE]...HEAD -- <path>
 
 [ONLY if review_scope == 'delta' in pr-context metadata]
-Re-review — scope to delta only.
+Re-review — default to delta focus. If the orchestrator determined a sweep is warranted, also look beyond the delta for net-new, high-signal issues in your domain (but do NOT re-raise anything already covered in Prior Feedback).
 
 [ONLY if summary mode OR review_scope == 'delta' — include a file list]
 Files:
@@ -194,7 +212,7 @@ Files:
 
 **Keep handoffs short.** The reviewer has full access to pr-context and pr-tldr for details. The handoff just points them in the right direction.
 
-**Scope signal:** Use `Review scope` from the pr-context metadata table as the single source of truth for delta vs full scoping. If `review_scope` is absent (e.g. local runs without CI-generated pr-context), default to full-scope behavior.
+**Scope signal:** Use `Review scope` from the pr-context metadata table as the default signal for delta vs full scoping. If `review_scope` is absent (e.g. local runs without CI-generated pr-context), default to full-scope behavior. If `review_scope=delta` but a sweep was determined to be warranted (see Phase 1.1 Sweep Assessment), look beyond the delta for missed issues in heavier domains — while still honoring Prior Feedback and the No Duplication Principle.
 
 ## Phase 4: Judge & Filter
 
@@ -208,6 +226,7 @@ Cluster findings describing the same issue:
 - `inline`: Same file + overlapping lines + similar problem → **merge**
 - `file`: Same file + similar problem → **merge**
 - `multi-file`/`system`: Similar scope + similar problem → **merge**
+- **Cross-type:** Also cluster across finding types when they address the same underlying concern. An `inline` fix that is a subset of a broader `file`/`multi-file`/`system` finding (or vice versa) must be merged into **one** finding. Choose the scope that best serves the developer — if the broader framing adds value, keep the broader finding; if the specific line-level fix is what matters, keep the `inline` version. **Never surface both.**
 - Keep or consolidate to the most actionable version (clearest issue + implications + fixes)
 
 ### 4.2 Relevancy Check

.github/workflows/claude-code-review.yml

@@ -128,6 +128,20 @@ jobs:
           echo "$FILES" >> $GITHUB_OUTPUT
           echo "EOF" >> $GITHUB_OUTPUT
 
+          # Compute filtered size metrics from numstat (respects EXCLUDE_PATTERNS,
+          # unlike the raw GitHub API additions/deletions which include lockfiles etc.)
+          NUMSTAT=$(git diff --numstat "$BASE"...HEAD -- . "${EXCLUDE_PATTERNS[@]}")
+          FILTERED_ADDITIONS=$(printf '%s\n' "$NUMSTAT" | awk '$1 != "-" {s+=$1} END{print s+0}')
+          FILTERED_DELETIONS=$(printf '%s\n' "$NUMSTAT" | awk '$2 != "-" {s+=$2} END{print s+0}')
+          if [ -n "$FILES" ]; then
+            FILTERED_FILE_COUNT=$(printf '%s\n' "$FILES" | wc -l | tr -d ' ')
+          else
+            FILTERED_FILE_COUNT=0
+          fi
+          echo "filtered_additions=$FILTERED_ADDITIONS" >> $GITHUB_OUTPUT
+          echo "filtered_deletions=$FILTERED_DELETIONS" >> $GITHUB_OUTPUT
+          echo "filtered_file_count=$FILTERED_FILE_COUNT" >> $GITHUB_OUTPUT
+
           # Per-file diff stats (compact size overview)
           DIFF_STATS=$(git diff --stat "$BASE"...HEAD -- . "${EXCLUDE_PATTERNS[@]}")
           {
@@ -571,7 +585,7 @@ jobs:
           | **Base** | `${{ steps.pr.outputs.base_ref }}` |
           | **Repo** | ${{ github.repository }} |
           | **Head SHA** | `${{ steps.pr.outputs.head_sha }}` |
-          | **Size** | ${{ steps.diff.outputs.commit_count }} commits · +${{ steps.pr.outputs.additions }}/-${{ steps.pr.outputs.deletions }} · ${{ steps.pr.outputs.changed_files_count }} files |
+          | **Size** | ${{ steps.diff.outputs.commit_count }} commits · +${{ steps.diff.outputs.filtered_additions }}/-${{ steps.diff.outputs.filtered_deletions }} · ${{ steps.diff.outputs.filtered_file_count }} files |
           | **Labels** | ${{ steps.pr.outputs.labels || '_None_' }} |
           | **Review state** | ${{ steps.review_context.outputs.review_decision }} |
           | **Diff mode** | `${{ steps.diff.outputs.diff_mode }}` — ${{ steps.diff.outputs.diff_mode == 'summary' && 'reviewers must read file diffs on-demand via `git diff`' || 'full diff included below' }} |