merge-review-json-and-submit-pr-review

name: merge-review-json-and-submit-pr-review description: Use when multiple review-output-format JSON files must be merged into one result and submitted as a single GitHub PR review where each finding becomes one inline comments[] item.

Merge Review JSON And Submit PR Review

Overview

Merge multiple JSON files produced by review-output-format, create one GitHub PR review payload, and submit it in one API call.

This skill guarantees:

• semantic-similarity decisions are made by the agent (not by script heuristics)
• the script performs deterministic transforms only (exact dedupe, sorting, anchor parsing, payload build, submit)
• merged findings are sorted by severity (Blocker, Major, Minor, Info, Nit)
• each severity level is rendered with a dedicated emoji for fast scanning
• summary starts with an explicit AI-generated attribution and manual follow-up note
• findings summary is grouped by severity to avoid repeated risk labels on every line
• one finding maps to one comments[] object when an inline anchor is available
• summary body stays short and includes rollup counts
• any unanchored findings stay in summary body instead of being dropped

Inputs

• pr_link (required)
  • full GitHub PR URL: https://github.com/<owner>/<repo>/pull/<number>
• input_files (required)
  • each file must follow review-output-format: top-level { "findings": [] }
  • pass with repeated --input flags or --input-glob
• merged_output (optional)
  • default: merged-review-output.json
• payload_output (optional)
  • default: github-review-payload.json

Required Semantic Merge Step (Agent, Not Script)

Before running the script, the agent must consolidate semantically similar findings across input_files manually:

1. read all findings from all input files
2. group findings that describe the same issue
3. merge info inside each group into one final finding:
   • keep the highest severity
   • preserve the best anchor (path + line) when available
   • merge why, risk_if_unchanged, evidence, and change_request content without losing unique details
4. write a deterministic pre-merged JSON file (same review-output-format)
5. run this script using that pre-merged file as input

The script intentionally does not perform semantic matching or semantic merge.

Inline anchor resolution for each finding:

1. use path + line directly if present
2. otherwise parse from scope with <path>:<line>
3. if anchor cannot be resolved, keep finding in summary body

Command

python3 ./merge-review-json-and-submit-pr-review/scripts/merge_and_submit_review.py \
  --pr-link "https://github.com/pingcap/tidb/pull/12345" \
  --input semantic-merged-review.json \
  --merged-output merged-review-output.json \
  --payload-output github-review-payload.json

Use globs when needed:

python3 ./merge-review-json-and-submit-pr-review/scripts/merge_and_submit_review.py \
  --pr-link "https://github.com/pingcap/tidb/pull/12345" \
  --input-glob "outputs/reviews/*.json"

Dry Run

Generate merged JSON and payload without posting:

python3 ./merge-review-json-and-submit-pr-review/scripts/merge_and_submit_review.py \
  --pr-link "https://github.com/pingcap/tidb/pull/12345" \
  --input-glob "outputs/reviews/*.json" \
  --dry-run

Authentication

Submission requires GitHub CLI authentication:

gh auth status

If unauthenticated, run:

gh auth login

Output

The script writes:

• merged findings JSON (merged_output)
• final GitHub payload (payload_output)

Then it prints one JSON result object with counts and submission status.