Pr Validity

DLC: PR Validity Analysis

Detect duplicate or redundant code introduced by a PR. Read-only analysis — no code modifications.

Before running, read ../dlc/references/ISSUE-TEMPLATE.md now for the issue format, and read ../dlc/references/REPORT-FORMAT.md now for the findings data structure.

Step 1: Resolve Target PR

Determine the PR to check and fetch all data needed for subsequent steps in a single call:

# If PR number provided as argument
PR_JSON=$(gh pr view <PR_NUMBER> --json number,title,url,headRefName,state,additions,changedFiles,files,body)

# If no argument — detect from current branch
PR_JSON=$(gh pr view --json number,title,url,headRefName,state,additions,changedFiles,files,body)

# Display PR summary
echo "$PR_JSON" | jq '{number, title, url, headRefName, state, additions, changedFiles}'

If no open PR is found, abort with: "No open PR found for the current branch. Push your changes and open a PR first."

PR Description Quality Check

Extract .body from PR_JSON and evaluate against a minimal structure rubric:

Evaluation order: Check #1 first. If the body is empty or under 50 chars, emit only the High finding and skip checks #2 and #3 (they are logically entailed and not independently actionable on an empty body). When the body is ≥ 50 chars, evaluate checks #2 and #3 and emit all that match (unlike spec-quality's precedence rule — each missing element is independently actionable).

For the required file field, use PR#<number> (same convention as spec-quality PR-level findings).

Large PR gate: If the PR has 500+ additions, use AskUserQuestion before proceeding:

• Present the addition count and file count (use changedFiles for the accurate total — .files | length caps at 100)
• Options: "Analyze everything" / "Only new files" / "Abort"
• If "Abort", stop and report: "PR validity analysis aborted by user."
• If "Only new files", limit Steps 2-3 to files with status added or renamed (skip modified)

Step 2: Fetch Diff & Extract Additions

Note: If the user selected "Only new files" in Step 1, restrict all diff parsing and construct extraction in this step to files where status == "added" or status == "renamed" (skip status == "modified").

Retrieve the full diff and extract new code constructs:

# Get the full diff (raw text — cannot be combined with the JSON call)
gh pr diff <PR_NUMBER>

# Get per-file status from the cached PR_JSON (no extra API call)
echo "$PR_JSON" | jq '.files[] | {path: .path, status: .status, additions: .additions}'

Parse the diff for + lines (excluding +++ b/ headers) and extract declarations:

For each extracted construct, record:

• name: identifier name
• file: file path where it appears in the PR
• line: line number in the new file
• kind: function, class, component, method, constant
• signature: parameter list and return type (if available)
• body_snippet: first 5 lines of the body (for similarity matching)

Edge case: If no code additions are found (e.g., the PR only modifies docs, configs, or deletes code), create a single Info finding: "No new code constructs detected in PR diff — only non-code or deletion changes." Then skip to Step 7 (Report).

Step 3: Codebase Search

For each extracted construct, search the existing codebase for matches. Exclude files that are part of the PR diff.

Before targeted searches, launch an Explore agent with the following structured prompt:

Explore the codebase structure to identify where the following constructs might have existing implementations. For each construct, report:

1. Directories most likely to contain similar code
2. Any files with matching or similar names
3. Related test files that might reveal expected behavior

Constructs to locate: (list each construct name and kind extracted from Step 2)

Exclude these PR files from results: (list each file path from the PR diff)

Use repomix-explorer (if available) for large codebases. Use the Explore agent output to prioritize directories for Steps 3a–3d. Then use Grep and Read for the targeted searches below.

3a. Name-Based Search

Use Grep to find constructs with the same name:

Grep: pattern="(function|const|class|def|fn|func|type)\s+{name}\b"

Exclude: node_modules, dist, build, .git, vendor directories, and the PR's own files.

3b. Signature Comparison

For each name match found in 3a:

1. Read the matched file around the match location (20 lines of context)
2. Compare parameter lists, return types, and overall structure
3. Score similarity: exact match, compatible (same params different order), or different

3c. Pattern-Based Search

Search for structural matches beyond exact names:

• Export names: Grep for the same export identifier
• Component structures: similar prop types, render patterns
• API endpoints: same route path or handler pattern

3d. Body-Based Search (Code Movement Detection)

For constructs where the PR also deletes code (file has both + and - lines):

1. Extract the body of the deleted construct
2. Grep for distinctive lines from the deleted body in the new location
3. If the new construct's body closely matches a deleted construct, classify as Code Movement (not duplication)

Self-match guard: When the deleted and added constructs are in the same file with overlapping line ranges, classify as Update (not Movement). Only flag Movement when code migrates between different files.

Step 4: Check Issue Reference & Spec Quality

Check whether the PR references any GitHub issues:

# Extract PR body from the cached PR_JSON (no extra API call)
echo "$PR_JSON" | jq -r '.body'

Scan the PR body for GitHub issue references in any of these forms:

• #N (e.g., #123)
• owner/repo#N (e.g., myorg/myrepo#456)
• Full issue URLs (e.g., https://github.com/owner/repo/issues/789)

If no issue references are found, produce a finding:

Deduplicate the extracted issue references before fetching (a PR body may reference the same issue multiple times). Preserve both the owner/repo and the issue number from each matched reference. For each unique referenced issue:

# Shell variables don't persist across separate Bash tool calls — REPO is
# fetched fresh here (Step 1 never set it).
REPO=$(gh repo view --json nameWithOwner -q .nameWithOwner)
if [ -z "$REPO" ]; then
  echo "ERROR: gh repo view returned no repository — check gh auth status." >&2
  exit 1
fi
gh issue view <ISSUE_NUMBER> --repo <OWNER/REPO> --json number,title,state,labels,body

If the reference is plain #N, use $REPO as the default <OWNER/REPO>.

Record the issue state (open/closed) and labels. This is informational only — produces Info-level findings if issues are referenced.

Body Inspection (Spec Quality)

For each fetched issue, inspect the issue body for specification quality:

When multiple conditions match for a single issue, emit at most one spec-quality finding for that issue, choosing the highest-severity match using this precedence order: Medium > Low > Info.

Spec-quality findings are metadata-only signals and must not be counted toward the redundancy-focused issue-creation threshold in Step 6; only findings with type: redundancy are eligible to trigger a "{n} redundancies" issue.

File Field for Spec-Quality Findings

Spec-quality findings describe PR or issue metadata rather than repository files. For the required file field, use these conventions:

• PR-level findings (e.g., no linked issue): PR#<number> (e.g., PR#128)
• Issue-level findings (e.g., missing acceptance criteria): issue#<number> (e.g., issue#99)

Step 5: Classify & Build Findings

For each construct extracted in Step 2, assign a classification based on the search results from Step 3:

Classification findings from this step use type: redundancy.

Severity mapping (reinforced here for defense-in-depth):

Step 6: User-Gated Issue Creation

Threshold: Create an issue only if there is any high redundancy finding OR 3+ medium redundancy findings. Findings with type: spec-quality or type: pr-description are excluded from this threshold.

If the threshold is not met, skip issue creation and proceed to Step 7.

If the threshold is met, use AskUserQuestion:

• Present the finding counts by severity
• Options: "Yes, create issue" / "No, skip" / "Show details first"
• If "Show details first", display each finding with file, line, classification, and matched location, then re-ask with the first two options
• If "No, skip", proceed to Step 7 without creating an issue

If the user approves issue creation, proceed:

Read ../dlc/references/ISSUE-TEMPLATE.md now and format the issue body exactly as specified.

Critical format rules (reinforced here):

• Title: [DLC] PR Validity: {n} redundancies in PR #{number}
• Label: dlc-pr-validity
• Body must contain: Scan Metadata table, Findings Summary table, Findings Detail grouped by severity, Recommended Actions
• In the Scan Metadata table, set Project Type to PR analysis (this skill is PR-focused and does not detect codebase project types)

Additional section — add after Findings Detail:

## Change Classification Summary

| Classification | Count | Files |
|---------------|-------|-------|
| New | {n} | {comma-separated file list} |
| Duplicate | {n} | {comma-separated file list} |
| Duplicate (divergent) | {n} | {comma-separated file list} |
| Override | {n} | {comma-separated file list} |
| Update | {n} | {comma-separated file list} |
| Trivial Overlap | {n} | {comma-separated file list} |
| Code Movement | {n} | {comma-separated file list} |
| **Total Constructs** | **{n}** | |

## Referenced Issues

| Issue | Title | State |
|-------|-------|-------|
| #{n} | {title} | {open/closed} |

> Omit this section if no issues were referenced in the PR body.

Raw Output: This skill has no CLI tool output to capture. Omit the Raw Output section from the issue body.

Acquire BRANCH for the Scan Metadata table — the PR's own head branch, not the local checkout's, since this skill analyzes a PR's diff rather than scanning the working tree. Step 1 already fetched headRefName, but that value doesn't survive into this separate Bash tool call, so re-fetch it here:

BRANCH=$(gh pr view <PR_NUMBER> --json headRefName -q .headRefName)
echo "BRANCH=$BRANCH"   # the Write step below is a separate tool call and can't see this shell's variables

Then follow ISSUE-TEMPLATE.md's Issue Creation Command lifecycle exactly — it acquires REPO itself — substitute {skill-name} = pr-validity, {Type} = PR Validity, {type} = pr-validity, and the title with the {n}/{number} counts from this skill's findings. Delete the {additional-required-sections} token entirely from the for section in ... line — do not paste any text in its place — since this skill's body has no Raw Output section.

If the user declines, skip issue creation and proceed to Step 7.

Step 7: Report

Print a summary:

PR validity analysis complete.
  - PR: #{number} ({title})
  - Constructs analyzed: {n}
  - Classifications: {n} new, {n} duplicate, {n} divergent, {n} override, {n} update, {n} trivial overlap, {n} code movement
  - Findings: {n} high, {n} medium, {n} low, {n} info
  - PR description: {n} findings ({n} high, {n} medium, {n} low)  [only if pr-description findings exist]
  - Issue: #{number} ({url})  [only if created]
  - Referenced issues: #{n1} (open), #{n2} (closed)  [only if found in Step 4]