mocchibird/ascend-agent-kit/pr-feedback

Process GitHub PR Review Feedback

Accept a GitHub PR URL as the argument. If no URL is provided, ask the user for one.

Step 1 — Fetch all review comments

Use gh CLI (must be authenticated) to fetch:

gh api repos/{owner}/{repo}/pulls/{number}/comments \
  --jq '.[] | "\(.id) | \(.user.login) | \(.path):\(.original_line // .line) | \(.html_url)\n\(.body)\n---"'

Also fetch top-level issue comments:

gh api repos/{owner}/{repo}/issues/{number}/comments \
  --jq '.[] | "\(.id) | \(.user.login) | \(.html_url)\n\(.body)\n---"'

And PR metadata:

gh pr view {number} --repo {owner}/{repo} --json title,state,body,reviewDecision

Parse the owner, repo, and PR number from the URL argument. The URL format is https://github.com/{owner}/{repo}/pull/{number}.

Step 2 — Prioritize and present

Categorize each comment by importance:

1. Critical — Algorithm/kernel/correctness changes, performance issues, data-integrity bugs
2. Important — Benchmark configuration, test coverage, memory layout, API design
3. Minor — Naming, style, documentation, formatting

Present a numbered summary to the user showing for each comment:

• Priority tier
• Reviewer name
• File:line (if inline comment)
• The exact comment text
• The GitHub comment URL

Then ask the user to confirm proceeding, or let them reorder/skip comments.

Step 3 — Process each comment

Work through comments in priority order. For each comment:

1. Read the relevant file(s) to understand the current code
2. Make the code changes to address the feedback
3. Compile and verify:

• If the project has a build system, compile the changed code
• If tests exist, run them
• If this is an NPU kernel project (has bisheng compilation, torch-npu tests), follow the npu_kernel_general skill's mandatory requirements: compile with bisheng, execute on NPU, verify numerical correctness

1. Write a step report as Step{N}_{short_title}.md in a claude_tasks_reports/ subdirectory relative to the changed files. The report must contain:

# Step {N}: {Title}
 
## PR Comment Addressed
 
**Reviewer:** {reviewer}
**File:** `{file}:{line}` (if inline)
**Comment URL:** {github_comment_url}
 
> {exact comment text, blockquoted}
 
## Problem
 
{Brief description of what the reviewer identified}
 
## Changes
 
{What was changed and why}
 
## Files Changed
 
-`{file1}`: {description}
-`{file2}`: {description}
 
## Verification
 
-{Compilation result}
-{Test results}
-{Any numerical checks}

1. Mark the task complete before moving to the next comment

Step 4 — Final verification

After all comments are addressed:

1. Run the full test suite to check for regressions
2. Summarize all changes made across all steps
3. List the report files created

Important notes

• Always read files before modifying them
• Do not batch multiple comments into a single step — one step per comment
• If a comment requires investigation (e.g., "can this be done in FP16?"), investigate first, then implement if feasible
• If two comments conflict, flag this to the user before proceeding
• Use task tracking (TaskCreate/TaskUpdate) to show progress