Mine Address Pr Issues

Address PR Issues

Triage and resolve everything blocking a PR from merging: unresolved review comments, merge conflicts, and failing CI checks. Works on both GitHub and Azure DevOps — detects the platform automatically.

When to Activate

• User asks to address, fix, or review PR comments
• User asks to check for unresolved PR feedback
• User asks to fix failing CI or merge conflicts on a PR
• User says "address PR issues", "fix PR", "make this PR mergeable", or similar
• User mentions Copilot comments on a PR

Usage

/mine-address-pr-issues [PR#]

If no PR number is given, auto-detect from the current branch.

Phase 1: Fetch & Detect

Platform detection

git-platform

Output is github, ado, or unknown. If unknown, tell the user the platform could not be detected from the git remote and stop.

PR metadata

GitHub:

gh pr view {PR} --json number,title,url,baseRefName,headRefName,mergeable,mergeStateStatus,statusCheckRollup,isDraft,reviewDecision

If mergeable is UNKNOWN, retry up to 3 times with backoff (3s, 6s, 12s) — GitHub computes mergeability asynchronously. If still UNKNOWN after retries, warn the user and continue.

ADO:

ado-api pr show {PR} --json

Returns pullRequestId, title, status, sourceRefName, targetRefName, repository.webUrl. URL: repository.webUrl + "/pullrequest/" + pullRequestId. Note: mergeStatus is optional and only present after a merge attempt.

Review threads & non-thread comments (MANDATORY — separate from metadata)

CRITICAL: gh pr view --json does NOT return review threads (inline comments from reviewers or Copilot). You MUST run the fetching command below. Do not conclude "no review comments" based on PR metadata alone — that field doesn't exist in the metadata response.

GitHub:

gh-pr-threads {PR} --json --all

Returns a JSON object with three surfaces — all three need triage:

• .threads — inline review threads (resolvable). Each has id (PRRT_…), isResolved, isOutdated, path, line, startLine, diffSide, and comments (with databaseId, body, author.login, author.__typename).
• .reviewComments — review-summary bodies carrying findings that are not inline threads. CodeRabbit posts substantial findings here ("Outside diff range comments", "Duplicate comments", "Nitpick comments") when it can't anchor to the diff. Not resolvable — no PRRT_ id; reply with a normal PR comment. Each has author, state, url, body.
• .issueComments — PR conversation comments (human comments, bot chat replies). Machine-generated status noise (walkthroughs, build reports) is already filtered out. Not resolvable. Each has author, databaseId, url, body.

Do NOT skip .reviewComments — it is the surface most often missed, and CodeRabbit routinely puts Major findings there.

ADO:

ado-api pr threads {PR} --json --all

Returns all threads as JSON. Threads with threadContext are inline comments; threads without are general conversation. ADO has no isOutdated concept. ADO carries general comments in the same thread list, so it has no separate .reviewComments/.issueComments split.

CI status

GitHub: From statusCheckRollup in metadata. Filter for conclusion in FAILURE, TIMED_OUT, ACTION_REQUIRED.

ADO:

az repos pr policy list --id {PR_ID} -o json

Filter for status in rejected, broken.

Pre-flight warnings

Check and display as informational warnings (NOT blockers):

• isDraft (GitHub) — "This is a draft PR. Changes can be made but it won't be mergeable until marked Ready for Review."
• reviewDecision == CHANGES_REQUESTED (GitHub) — "Reviewer requested changes. Even after fixing all comments, they'll need to re-approve."

Phase 2: Triage & Plan

Categorize all issues into three groups: review comments, merge conflicts, CI failures.

Review comments

Prerequisite: This section requires the gh-pr-threads / ado-api pr threads output from Phase 1. If you skipped that step, go back and run it now before triaging.

Exclude resolved threads: GitHub isResolved: true, ADO status != "active".

Triage outdated threads (GitHub only): For each thread with isOutdated: true:

1. Read the comment body to understand the concern
2. Read the current code at that location
3. If the location was entirely deleted: auto-categorize as "location removed — likely addressed by refactoring"
4. If the concern is addressed: categorize as "already addressed" ONLY if you can cite the specific line that addresses it
5. Otherwise: categorize as "needs manual review"

Do NOT blanket-exclude outdated threads. Do NOT confidently dismiss concerns without citing evidence.

General comment assessment: For non-inline comments, determine if actionable vs. discussion/approval/acknowledgment. Include only actionable, unaddressed comments.

Already-addressed detection: For each unresolved thread, read the current code at the referenced location and compare against what the reviewer requested. If already present, categorize as "already addressed."

Assign investigation depth

For each issue, assign a depth based on the comment's nature:

Group logically

Group related comments for efficient execution — e.g., all error-handling comments together, all comments on a single feature together.

Merge conflicts

GitHub: mergeable == "CONFLICTING" or mergeStateStatus == "DIRTY" ADO: mergeStatus == "conflicts" (if present)

CI failures

Fetch failure logs and categorize: test failures, lint/type errors, build errors, other.

GitHub: gh run view <run-id> --log-failed ADO: ado-api logs read <build-id> --failed --issues — reads the failed step's log and filters to error/warning lines. Run ado-api logs read --help for full usage.

Present the plan

Print the plan as a numbered list before the AskUserQuestion. Each entry must include:

1. The reviewer's concern — one-sentence summary of what they asked for
2. Proposed fix — concrete description of the code change (name the function, the file, what will change). "Fix error handling" is not a plan; "wrap fetch_user() in try/except for ConnectionError in auth.py:42" is.
3. Investigation depth — light / medium / deep
4. Resolution policy — resolve (bot or self-review) or reply-only (human reviewer)

Mark items that need a user decision with [DECISION NEEDED] — these are comments where:

• The reviewer suggests two or more valid approaches
• The concern is about design/architecture with no single obvious answer
• You disagree with the reviewer's suggestion (state why)
• The requested change would conflict with another reviewer's comment

For [DECISION NEEDED] items, state the options and your recommendation.

Also include:

• Pre-flight warnings from Phase 1
• "Already addressed" items (with evidence) — listed separately so the user can verify

AskUserQuestion:
  question: "Here's my plan for PR #{N}. Items marked [DECISION NEEDED] have options listed — choosing 'Looks good' accepts my recommendation for those. Review and tell me if anything should change."
  header: "PR Plan"
  multiSelect: false
  options:
    - label: "Looks good — address all"
      description: "Proceed with the full plan, using the proposed recommendations for [DECISION NEEDED] items"
    - label: "Adjust items"
      description: "I'll tell you what to change, skip, or decide differently"
    - label: "Cancel"
      description: "Exit without making changes"

If "Adjust items": ask which numbered items to skip or change. For items the user wants changed, ask what they want instead and update the plan entry before proceeding.

Merge conflict strategy

If conflicts exist, ask the user:

AskUserQuestion:
  question: "Merge conflicts detected. How should I resolve them?"
  header: "Conflicts"
  options:
    - label: "Merge (Recommended)"
      description: "git merge origin/<base> — creates a merge commit, preserves history"
    - label: "Rebase"
      description: "git rebase origin/<base> — rewrites history, requires force-push"

Phase 3: Execute

Create temp directory

get-skill-tmpdir mine-address-pr

Fix each logical group (serial)

For each group from the plan, launch a standard-worker subagent with:

• The review comment(s) to address (bodies, file paths, line numbers)
• The approved proposed fix from the plan (what the user agreed to)
• For [DECISION NEEDED] items: the resolved decision
• The investigation depth (light, medium, or deep)
• Output path: <tmpdir>/group-N/result.md

Subagent prompt template:

You are addressing PR review feedback. Your output goes to {output_path}.

Comments to address: {comment_details}

Approved fix: {proposed_fix_from_plan}

Investigation depth: {depth}

If light: Read the target file. Apply the fix. Run the project's test suite. If tests fail: fix or escalate. Max 3 retries.

If medium: Read the target file fully. Grep for call sites of the function/class being modified. Read at least one call site to understand usage. Apply the fix. Run the project's test suite (follow the test execution discovery order from references/common/testing.md). If tests fail: fix the code. Max 3 retries, then escalate to user.

If deep: Read the target file fully. Grep for call sites — read ALL callers. Read related test files. Read adjacent modules in the same package/directory. Apply the fix. Run the project's test suite. If tests fail: fix or escalate. Max 3 retries.

CRITICAL: Never explain away a test failure or CI error. If tests fail after your fix, the fix is wrong — revise it. Do not suggest the test is outdated, flaky, or testing the wrong thing. Do not suggest skipping or marking the test as expected failure. Fix the code until tests pass, or escalate to the user after 3 attempts.

Write a one-line summary as the first line of your output file, then details below.

Read only the first line of each result file for the summary — do not read full subagent output into main context.

Code review loop

After all subagents complete:

1. Run code-reviewer agent on modified files
2. For each CRITICAL or HIGH finding: auto-fix when unambiguous, defer to user when judgment is needed
3. If any auto-fixes applied, re-run code-reviewer (max 3 iterations)
4. Stop when no CRITICAL/HIGH issues remain or 3 iterations reached
5. Run integration-reviewer once on the final diff

Do NOT commit until both reviewers pass. If CRITICAL/HIGH findings remain after 3 iterations, present them to the user before proceeding.

Commit and push

Commit per logical group with descriptive messages:

fix(auth): use logging instead of print per review
fix(config): add LOGIN_REDIRECT_URL to test settings

Push once after all commits.

Thread replies and resolution

After push is confirmed, reply to threads. For each addressed thread:

1. Idempotency check: Search the thread's comment history (fetched in Phase 1) for ANY comment containing <!-- addressed-pr-issues -->. If found, skip the reply.
2. Post reply with the <!-- addressed-pr-issues --> marker in the body. Keep replies concise and professional:
   • Code change: "Fixed — [brief description of what was changed]. "
   • Already addressed: "This was addressed in a previous commit — [cite specific evidence]. "
   • Outdated/removed: "The code at this location was refactored and this concern no longer applies. "
3. Resolve per policy:

Bot detection:

• GitHub: author.__typename == "Bot" is the primary signal. Fall back to [bot] suffix check on author.login if __typename is unavailable.
• ADO: Check for [bot] suffix on author.uniqueName. Also check if uniqueName matches service account patterns (no @ domain, or matches the project's build service identity).

GitHub resolution: Use gh-pr-reply {PR} {comment-database-id} "{body}" --resolve {thread-id} for combined reply+resolve.

ADO resolution: Two calls: ado-api pr reply {PR} {thread-id} "{body}" then ado-api pr resolve {PR} {thread-id}.

Rate limiting: 1-second delay between mutative API calls.

Phase 4: Summary

Present a structured summary:

## Summary

### Review Comments
- Resolved (bot threads): N threads [replied & resolved]
- Replied (human threads): M threads [reply posted, awaiting reviewer]
- Already addressed: K threads [replied]
- Skipped: J threads [reason]

### Merge Conflicts
- Resolved: N files — [merge/rebase] origin/<base> into <head>

### CI Failures
- Fixed: N checks — [brief description of each fix]
- Still pending: CI will re-run on push

### Commits
- [list each commit with its message]

### Needs Manual Review
- [any items that could not be resolved automatically]

Helper Scripts

IMPORTANT: Use these helper scripts instead of inline commands. They handle authentication, pagination, and output formatting.

• GitHub: gh-pr-threads, gh-pr-reply (with --resolve), git-platform — run --help on each for usage
• ADO: ado-api pr (show/list/create/update/threads/reply/resolve/resolve-pattern), ado-api logs read (CI failure logs), ado-api work-item — run ado-api --help for usage
• Platform: git-platform — prints github, ado, or unknown

Error handling

• Auth failures: suggest gh auth login (GitHub) or az login (ADO)
• Rate limiting: inform user and suggest waiting
• No PR found: ask user for a PR number
• GitHub GraphQL permissions: suggest gh auth refresh -s repo