Resolve Comments

Resolve PR/MR Review Comments

Work through all unresolved review comments from a pull request or merge request on GitHub, GitLab, Bitbucket Cloud, or Azure DevOps.

Step 0: Graphify Availability (mandatory if present)

Before anything else, check whether the Graphify skill is available in the running agent for this project.

Check the list of available skills surfaced by the harness for a graphify skill (it may appear as graphify or a plugin-namespaced <plugin>:graphify). If present, Graphify is available; invoke it via the Skill tool when required below.

If Graphify is available, using it is mandatory for this run. Wherever a later step requires understanding code, mapping relationships across files, or making a design decision, you MUST first run Graphify on the relevant input (the affected file(s), the PR/MR diff, or the surrounding module) and use the resulting knowledge graph to inform your work. This applies in particular to:

• Step 4 — understanding each comment in the context of the surrounding code.
• Step 4b — every deferred item: build the graph for the item's scope before choosing and executing a design route.
• Step 6 — self-review: graph the diff to check cross-file impact.

If Graphify is not available, proceed normally — do not block on it.

Record availability as a transient session fact — do NOT persist it; skill availability can change between sessions.

Step 1: Platform Detection & CLI Verification

1.1 Check Memory

Check the agent's persistent memory (if available) for a previously saved pr-comments-resolver-platform configuration. If found, run a quick auth verification using the recorded Auth method and the platform module's Auth verification section.

If the memory exists and auth is healthy, skip to Step 1.6 (Load Platform Module).

If the memory is stale (auth fails) or not found, continue with detection.

Back-compat: memories from older versions may not contain an Auth method field. Treat a missing Auth method as cli.

1.2 Detect Platform

Get the remote URL:

git remote get-url origin

Match the URL against known patterns:

If the URL does not match any known pattern (e.g., a self-hosted instance), ask the user which platform this repository is hosted on. Offer the options: GitHub, GitLab, Bitbucket Cloud, Azure DevOps, Other. Use AskUserQuestion to present the options.

If the user picks "Other", inform them: "This skill currently supports GitHub, GitLab, Bitbucket Cloud, and Azure DevOps. Other hosting platforms are not yet supported." Then stop.

1.3 Select Auth Method (Bitbucket and Azure only)

For bitbucket and azure, if memory does not already record an Auth method, ask the user which method to use. Use AskUserQuestion to present the options.

Options:

• Bitbucket: CLI (acli) or MCP (Atlassian Remote, OAuth).
• Azure DevOps: CLI (az + azure-devops extension) or MCP (@azure-devops/mcp, PAT-based).

For github and gitlab, Auth method is always cli — no prompt is shown.

Remember the chosen value as Auth method (it gets written to memory in Step 1.5).

1.4 Verify Auth

Use the platform module's Auth verification section, following the path that matches the chosen Auth method (Path A for cli, Path B for mcp). If auth is not healthy, follow the module's stop-with-instructions guidance.

1.5 Persist Platform (if supported)

After successful detection and CLI verification (and, for bitbucket/azure, auth-method selection), save a project memory:

---
name: pr-comments-resolver-platform
description: Detected hosting platform, auth method, and identifier for this repository
type: project
---

Platform: <github|gitlab|bitbucket|azure>
Auth method: <cli|mcp>
CLI tool: <gh|glab|acli|az>           # only when Auth method = cli
MCP server id: <ado|atlassian>        # only when Auth method = mcp
Token env var: $<ENV_VAR_NAME>        # only when MCP auth uses a PAT (e.g. azure mcp); omit for OAuth (atlassian) and all cli paths
Org / Workspace: <name>               # bitbucket workspace or azure organization; omit for github/gitlab
Repository: <platform-native identifier>
Reply/Resolve permission: <a|b|c>     # a = post & resolve, b = post only, c = none

Back-compat: memories saved by older versions of this skill may lack Auth method. When that field is missing, treat it as cli.

Back-compat: memories without Reply/Resolve permission trigger the one-time Step 4.0 prompt.

1.6 Load Platform Module

Read the file platforms/<platform>.md (relative to this SKILL.md). It contains the platform-specific instructions for auth verification, repository identifier, PR/MR number detection, comment fetching, posting replies, and resolving threads. Use it as the authoritative source for those operations throughout the rest of this skill.

Platform modules may invoke the MCP Config Writer for setting up an MCP server. The writer's procedure is:

MCP Config Writer

When a platform module needs to add an MCP server entry, follow this procedure exactly. Do not silently overwrite existing configuration.

Target file selection. Ask the user where to write the config:

• User scope: ~/.claude.json (default)
• Project scope: .mcp.json in the repository root

Use the AskUserQuestion helper.

Procedure:

1. Read the target file. If it does not exist, treat the starting content as {"mcpServers": {}}.
2. If mcpServers.<id> does not exist: insert the new entry and write the file back. Inform the user where the entry was added.
3. If mcpServers.<id> already exists:
   1. Compute the diff between the existing entry and the entry we intend to write.
   2. If identical: no-op. Inform the user "MCP entry already present and matches; nothing to write."
   3. If different: show the diff to the user and ask via AskUserQuestion: Overwrite / Keep existing / Abort. Apply the chosen action.

Secret handling.

• Never write a token literal into the config. Use ${ENV_VAR_NAME} placeholders in the env section.
• Verify the named env variable is set in the current shell environment before writing. If unset, instruct the user to add it to their shell profile (e.g. export AZURE_DEVOPS_PAT=... in ~/.zshrc), then stop the skill until they confirm.

Restart notice.

After writing, tell the user: "MCP server configured. Restart Claude Code to load the new server. Re-run /resolve-comments afterwards." Stop the skill in this session.

1.7 Get Repository Identifier

See the platform module's Repository identifier section.

Step 2: Determine PR/MR Number

If $ARGUMENTS is provided, use it as the PR/MR number.

If $ARGUMENTS is empty or not provided, detect automatically from the current branch. See the platform module's PR number (or MR IID) section.

If no PR/MR exists for the current branch, ask the user for the PR/MR number. Use AskUserQuestion to present the options.

Step 3: Fetch & Display Unresolved Comments

Use the platform module's Fetch unresolved … section to obtain the list of unresolved review threads/discussions and map them to the uniform {file, line, author, body, thread_id} shape. Then render them via the Display Unresolved Comments block below.

Display Unresolved Comments

Show unresolved comments as a numbered list:

## Unresolved PR Comments (X of Y total)

1. **path/to/file.ts:42** - @author
   > Comment text here...

2. **path/to/file.ts:108** - @author (outdated)
   > Another comment...

The (outdated) marker means the code has changed since the comment was written — check relevance before acting.

If there are no unresolved comments, inform the user and stop.

Step 3.5: Classify & Confirm

Classify each unresolved comment into one of two buckets:

• deferred: the comment flags a missing critical or larger piece of work — typically (a) a missing module/feature, (b) a security or correctness gap, (c) a refactor that touches more than two files, or (d) a reviewer-applied marker such as blocker, critical, or must-fix.
• normal: everything else (in-place edits within existing logic).

Use these signals as a heuristic — the user confirms the final split:

• Scope: would the fix introduce a new file/module, or change one existing function?
• Keywords: presence of missing, not implemented, should also handle, security, race, architecture, refactor entire, add support for, or localized equivalents.
• Cross-file: would the fix touch >2 files or require a design decision?
• Severity markers: explicit blocker / critical / must-fix labels or words from the reviewer.

Present the proposed split via the Classification block below.

Display the proposed classification as two numbered lists. Use the comment numbers from the Step 3 unresolved-comments display so the user does not need to look up identifiers.

## Classification (proposed)

Deferred (<N>):
  <i>. <path>:<line>  — @<author> — "<short excerpt>"
       reason: <one short phrase, e.g. "security + new module needed">
  ...

Normal (<M>):
  <comma-separated indices>

Reply with one of:
  OK
  move <i>→normal
  move <i>→deferred
  exclude <i>

Use AskUserQuestion when the harness supports it; otherwise emit the block as plain text and read free-form input. Re-render the block after each edit until the user replies OK.

Repeat the presentation, accepting move <N>→normal, move <N>→deferred, exclude <N>, or OK, until the user types OK. Excluded comments are dropped from both buckets (matches the existing exclude behaviour).

Step 4: Create Tasks & Resolve Comments

Create a TodoWrite task for each unresolved comment.

4.0 Reply/Resolve Permission (one-time per project)

Read the project memory pr-comments-resolver-platform. If the field Reply/Resolve permission is absent, ask the user once via AskUserQuestion:

• a — post replies AND resolve threads on my behalf
• b — post replies only (you resolve manually)
• c — no, I'll handle posting and resolving myself

Persist the answer in the same project memory written by Step 1.5 (pr-comments-resolver-platform). On later runs, skip this prompt unless the user explicitly resets.

Apply the chosen permission in every reply/resolve interaction that follows in Step 4 and Step 4b, including the no-code-change path (post a short justification → a: reply + resolve, b: reply only, c: show the justification to the user).

4.0.1 Clarification Routine

When a comment is unclear, ask one short clarifying question. Before asking, scan prior user answers in the current session — if the answer is already implied, state the inferred assumption and continue instead of asking. Keep clarifications terse; expand only on explicit request. Add a one-line example when the question itself is ambiguous.

Then work through each task sequentially:

1. Read the affected file at the referenced location
2. Understand the comment in the context of the surrounding code (if Graphify is available per Step 0, graph the affected file(s) first and use the result to understand the surrounding code)
3. If the comment is marked (outdated), check whether the feedback is still relevant before acting
4. If unclear: Ask the user what exactly is expected. Use AskUserQuestion to present the options.
5. Implement the change according to the reviewer's feedback
6. Mark the task as completed

Important:

• Work through comments sequentially, not in parallel — changes may overlap in the same file
• Follow conventions from CLAUDE.md (or AGENTS.md as fallback) if present in the project
• Respect existing code patterns and architecture in the project

Step 4b: Resolve Deferred Bucket

If the deferred bucket from Step 3.5 is empty, skip this step entirely.

Otherwise announce: "Now handling N deferred items." Process items strictly in the order shown in the Step 3.5 list, one at a time.

For each deferred item:

0. If Graphify is available (Step 0), build the knowledge graph for this item's scope first and use it to inform the route choice and execution below.

1. Detect which design skills are visible to the running agent:

   Check the list of available skills surfaced by the harness. Note whether superpowers:brainstorming and either sdd or superpowers:sdd are present. Record this as a transient session fact — do NOT persist it; skill availability can change between sessions.

If neither is available, only the built-in plan mode ([p]), inline ([d]), and skip ([x]) options are real choices; show [s] and [b] as unavailable with a one-line note ("requires the SDD / Brainstorming skill — not installed").

2. Ask the user how to handle this specific item. Present five options; unavailable skills are shown but marked unavailable:

   • [s] use SDD (spec → plan → tasks)
   • [b] use Brainstorming skill
   • [p] enter plan mode (built-in)
   • [d] just do it (no design skill — agent proceeds inline)
   • [x] skip this item (recorded as deferred-skipped in the summary)

3. Execute the chosen route. s/b invoke the corresponding skill; p activates plan mode and waits for ExitPlanMode; d follows the normal Step 4 inline path; x records the skip and continues.

4. After the route completes, apply the Reply/Resolve permission from Step 4.0 exactly as in the normal flow.

Step 5: Verification

After all comments have been addressed, run project-defined verification commands.

5.1 Find Verification Commands

Check the project's CLAUDE.md (or AGENTS.md as fallback) for defined verification commands (linting, static analysis, tests). Look for sections like "Commands", "Scripts", "Testing", or similar.

Examples of what you might find:

• PHP: task cs-fixer, task psalm, task test
• JavaScript/TypeScript: npm run lint, npm test
• Python: ruff check ., pytest
• Go: go vet ./..., go test ./...

5.2 Run Verification

If verification commands are found in CLAUDE.md (or AGENTS.md as fallback), run them in order.

If no verification commands are found, ask the user (offer a "skip verification" option as well): Use AskUserQuestion to present the options.

"What verification commands should I run for this project? (e.g., lint, tests, type checks)"

5.3 Fix Issues

If any verification step fails:

1. Analyze the error output
2. Fix the issue
3. Re-run the failing verification command
4. Repeat until all checks pass

Step 6: Self-Review

Review all changes made:

git diff

If Graphify is available (Step 0), graph the diff and use the result to check cross-file impact and relationships before the checks below.

Check for:

Reusability

• Is existing code reused instead of duplicated?
• Are new abstractions justified or unnecessary?

Code Quality

• Is the code clear and understandable?
• Does it follow existing patterns in the project?

Security

• No SQL injection, XSS, command injection, or other OWASP Top 10 vulnerabilities?
• Are inputs validated at system boundaries?
• No secrets or sensitive data in the code?

Project-Specific Conventions

Check any project-specific conventions and patterns defined in the project's CLAUDE.md (or AGENTS.md as fallback). The checks above are universal — defer to whatever the project documents for language- or framework-specific rules.

If you find issues during self-review, fix them immediately and re-run verification from Step 5.

Step 7: Summary

Show the user a summary of the work done:

• Comments addressed: List each comment with file path, line number, and what was changed
• Files modified: List each file with a brief description of changes
• Verification results: Which commands ran and their pass/fail status
• Self-review results: Any issues found and fixed during self-review
• Unresolved comments: If any comments could not be addressed, state the reason (e.g., unclear intent, requires architectural change, blocked by external dependency)