Merge Reviews

merge-reviews

Aggregate N drop files from /review-capture into one canonical plan. Verify findings against HEAD, dedup across reviewers, detect conflicts, schedule waves, then clean up the drops.

When to Use

• All review terminals for the current branch have finished capturing.
• Want one consolidated plan before execution.
• Need to detect phantom findings, duplicates, and conflicts before any

code change.

When NOT to Use

• Captures still in progress in other terminals. Wait for them.
• Single review with a handful of findings. Execute directly.
• You haven't reviewed yet. Run /review-capture first.

Invocation

/merge-reviews

No arguments. Operates on the current branch's drop directory.

Procedure

Follow every step. Do not skip verification or cleanup.

1. Locate drop files

BRANCH=$(git rev-parse --abbrev-ref HEAD)
DROP_DIR=".claude/review-drops/${BRANCH}"

If ${DROP_DIR} doesn't exist or is empty, abort: "No review captures found for branch <branch>. Run /review-capture first."

List every *.yaml in ${DROP_DIR}. Read all of them into memory. Parse each envelope.

2. Check snapshot consistency

For each drop, compare head_sha against current git rev-parse HEAD.

• All drops match current HEAD: proceed cleanly.
• Drops span multiple SHAs but all are ancestors of HEAD: proceed,

but note drift in the final report per drop.

• Any drop has a SHA that is NOT an ancestor of HEAD: something

destructive happened (rebase, branch reset). Report this to the user and ask whether to proceed or abort before continuing.

3. Verify every finding against HEAD

For each finding across all drops:

Anchored findings (have file, line_start, line_end, code_quote): 1. Confirm file exists at HEAD. 2. Read the file and confirm code_quote appears at or near the cited lines. Allow ±20 lines of drift (code may have moved). 3. If code_quote appears elsewhere in the file, update line_start/line_end to the actual location and record a drift_lines field. 4. If code_quote is absent entirely, mark the finding as STALE. 5. Ignore whitespace-only differences when comparing.

Unanchored findings (benchmarks, design concerns):

• No anchor verification possible. Mark them UNANCHORED and pass them

through to the merged plan as-is. The user decides whether they still apply.

Test-gap findings: verify the cited file still exists; treat missing coverage as perpetually valid unless the file was deleted.

Do NOT run cargo check or test suites here — verification is a purely textual check against HEAD. The execute phase runs quality gates.

Produce three buckets: verified, stale, unanchored.

4. Deduplicate across reviewers

Group verified findings that refer to the same code region:

• Same file AND line ranges overlap (within ±5 lines of each other)

AND claims are semantically similar → candidates for merge.

• Use claim similarity as a tiebreaker: if two findings on the same

region claim fundamentally different issues (one says "mutex held too long", the other says "wrong error variant"), they are not duplicates — keep both.

Merge rules:

• Retain the finding with highest severity.
• Combine suggested_fix values into a single suggested_fix list

(preserve both if they differ).

• Record merged_from: [<id1>, <id2>, ...] on the survivor.
• Record sources: [<source_slug1>, <source_slug2>, ...] listing all

reviewers that flagged this region — a concurrence signal.

5. Detect conflicts

For each surviving finding and for each other surviving finding on the same file with overlapping line ranges, check whether suggested_fix values contradict.

Contradictions to flag:

• One says "inline X"; the other says "extract X".
• One says "hold the lock"; the other says "release the lock early".
• One says "add fallible error return"; the other says "make this

infallible".

• One adds a field/parameter; the other removes it.

Heuristic: read both suggested_fix strings and ask "can a single implementation satisfy both?" If no, flag as a conflict.

Record conflicts as a top-level list in the merged plan:

conflicts:
  - region: "src/cache.rs:412-418"
    finding_a: <id>
    finding_b: <id>
    nature: "inline-vs-extract"
    summary: "perf-01 wants inlining; test-07 wants extraction for mocking"
    resolution: null    # user fills this in

Do NOT discard conflicting findings. The user resolves them before execute.

6. Schedule execution waves

Group findings by file, then assign each finding a wave such that no two findings in the same wave touch overlapping regions within the same file.

Algorithm:

• Sort findings by file, then by line_start.
• For each file, greedily assign findings to waves: put each finding in

the earliest wave whose existing findings for this file do not overlap with it.

• Findings on different files can share a wave freely.

This matches the parallel-feature-development pattern — waves can be executed concurrently where file regions don't collide.

7. Write the merged plan

Path: ${DROP_DIR}/_merged-<HEAD_SHORT>-<TIMESTAMP>.yaml

Use a leading underscore so ls-sorting puts it above the per-reviewer drops, and so a future /review-capture won't accidentally treat it as a reviewer output.

merged_at: <ISO-8601 UTC>
merged_from_head: <current HEAD SHA>
branch: <branch>
sources:
  - source_slug: ce-review
    head_sha: <sha>
    findings_total: 12
    findings_verified: 10
    findings_stale: 2
  - source_slug: asm-forge
    ...

summary:
  total_raw_findings: <N>
  verified: <N>
  stale: <N>
  unanchored: <N>
  after_dedup: <N>
  conflicts: <N>
  waves: <N>

stale_findings:
  # Preserved for audit. Not executed.
  - id: <id>
    source_slug: ce-review
    reason: "code_quote not found at file:line and absent from file"
    original: {...}

conflicts:
  - region: "src/cache.rs:412-418"
    finding_a: <id>
    finding_b: <id>
    nature: "..."
    summary: "..."
    resolution: null

waves:
  - wave: 1
    findings:
      - id: <id>
        kind: code_review
        file: ...
        severity: ...
        claim: ...
        suggested_fix: ...
        sources: [ce-review, cache-correctness-review]
        merged_from: [ce-review-003, cache-correctness-review-001]
      - id: <id>
        ...
  - wave: 2
    findings:
      - ...

8. Cleanup — delete the individual drop files

After the merged plan is written and verified readable, delete every per-reviewer drop file under ${DROP_DIR}. Keep only the merged plan and any other _merged-*.yaml files (from prior runs — do not delete those either, they're history).

# Delete only the per-reviewer drops, not the merged plans.
for f in "${DROP_DIR}"/*.yaml; do
  base=$(basename "$f")
  case "$base" in
    _merged-*) continue ;;   # keep merged plans
    *) rm -v "$f" ;;
  esac
done

Verify with ls "${DROP_DIR}" that only _merged-*.yaml files remain.

If the merged plan write failed for any reason, do NOT delete anything — leave all drops in place and report the failure so the user can re-run.

9. Report to user

Print a structured summary:

Merged <N> drop files into: .claude/review-drops/<branch>/_merged-<sha>-<ts>.yaml

Sources (N reviewers):
  - ce-review            : 12 findings (10 verified, 2 stale)
  - asm-forge            :  5 findings ( 5 verified, 0 stale)
  - cache-correctness    :  8 findings ( 7 verified, 1 stale)

After dedup: 18 findings (4 merges)
Conflicts:    2 — RESOLUTION REQUIRED BEFORE EXECUTE
Waves:        3

Severity:
  critical: 1
  high:     6
  medium:   8
  low:      3

Conflicts requiring your decision:
  1. src/cache.rs:412-418 — perf-01 vs test-07 (inline vs extract)
  2. src/fs_cache_handlers.rs:240-262 — codegen-02 vs correctness-04 (alignment vs layout)

Stale findings (2) preserved in merged plan for audit; not scheduled.

Individual drop files cleaned up: <N> files deleted.

Next: resolve conflicts in the merged plan, then run
/execute-review-findings .claude/review-drops/<branch>/_merged-*.yaml

Rules

Do NOT execute any findings

This skill produces a plan. It does not write code. If the user asks "and now execute them" after running /merge-reviews, redirect them to /execute-review-findings with the merged plan as input.

Do NOT delete drops on failure

Cleanup is the last step. If any prior step aborts (consistency failure, parse error, write failure), leave the drop files untouched. The user can fix the issue and re-run without losing work.

Preserve stale findings in the plan

Stale findings are discarded from execution but recorded in stale_findings: with their reason. This is the audit trail that tells the user which reviewers were working against outdated code.

Preserve merged plans across runs

Never delete _merged-*.yaml files. Each merge run produces a new one. The directory accumulates history across branches and re-reviews.

File-region overlap, not file-only

Two findings on cache.rs do not conflict if one touches lines 1-50 and the other touches lines 800-820. Only overlapping line ranges within the same file count.

Failure modes

• Drop directory missing: abort cleanly, recommend /review-capture.
• YAML parse failure on a drop: isolate the bad file, skip it with

a warning in the report, continue with the rest. Do NOT delete the bad file; the user can inspect.

• Current HEAD unrelated to drop SHAs: stop, explain, ask.
• All findings stale: still produce a merged plan (with empty

waves:) so the stale-finding audit trail is preserved. The user can re-capture against current HEAD.