pr-assistant

name	pr-assistant
description	Analyzes git changes, drafts localized PR titles and bodies, and assists with creating comprehensive pull requests. Use when user wants to create a PR, review changes before PR, or needs help drafting PR descriptions. Triggers on phrases like 'create PR', 'make a pull request', 'draft PR description', 'what changed in this branch', 'prepare PR'.

PR Assistant

Assists with creating well-structured pull requests by analyzing changes and generating reviewer-friendly PR descriptions.

Core Philosophy

Default: human-in-the-loop. Prepare a draft, show it to the user, and ask for approval before creating the PR.

Explicit fast-path. If the user clearly asks you to create the PR immediately, you may generate the draft, run sanity checks, and create the PR without a separate approval turn.

Requested language is binding. If the caller or user specifies a language, the final PR title and body must be written in that language. Treat generated commit-derived prose as draft material and rewrite it when needed.

Workflow

1. Analyze Changes

First, gather context about the current branch:

# Get current branch and target branch
CURRENT_BRANCH=$(git branch --show-current)
TARGET_BRANCH=${TARGET_BRANCH:-main}

# Get diff summary
git diff $TARGET_BRANCH...$CURRENT_BRANCH --stat

# Get commit messages
git log $TARGET_BRANCH...$CURRENT_BRANCH --oneline

# Get detailed changes by category
git diff $TARGET_BRANCH...$CURRENT_BRANCH --name-status

Use scripts/analyze_changes.py to categorize changes into generic buckets:

• Frontend
• Backend
• Application
• Tests
• Documentation
• Configuration
• CI
• Infrastructure
• Scripts
• Assets
• Data
• Other

When issue references are present, preserve them in three buckets when possible:

• issue_references: every detected issue reference
• closing_issue_references: issues that should auto-close on merge
• related_issue_references: issues that should stay linked without auto-closing

If classification confidence is low, do not force category-based prose in the PR body. Fall back to diff stats and top-level area summaries.

2. Determine PR Type

Prefer the branch name first when selecting the template:

• Feature: New functionality or enhancements
• Bugfix: Bug fixes or corrections
• Docs: Documentation updates
• Refactor: Code restructuring without behavior change
• Chore: Dependencies, config, tooling updates

If the branch name is inconclusive, use commit prefixes and the change mix. Do not switch to a bugfix template just because a feature branch contains a single fix: commit.

3. Prepare Output Files

Store generated artifacts inside skill-local output storage:

mkdir -p .opencode/skills/pr-assistant/outputs
BRANCH_SLUG=$(git branch --show-current | sed 's#/#-#g; s#[^A-Za-z0-9._-]#-#g')
STAMP=$(date +%Y%m%d-%H%M%S)
ANALYSIS_FILE=".opencode/skills/pr-assistant/outputs/pr-analysis-${BRANCH_SLUG}-${STAMP}.json"
PR_BODY_FILE=".opencode/skills/pr-assistant/outputs/pr-body-${BRANCH_SLUG}-${STAMP}.md"

Always save both the analysis JSON and the generated PR body.

4. Generate PR Draft

Use scripts/generate_pr_body.py with the appropriate template from assets/templates/.

If a requested language is provided:

• Choose the matching localized template when available, for example pr-template-feature-ja.md, pr-template-bugfix-ja.md, or pr-template-docs-ja.md
• Run generate_pr_body.py --language <requested-language> so auto-generated prose and checklist items match that language
• Treat the generated body as a draft and rewrite any remaining commit-derived text into the requested language before presentation or creation
• The final PR title must also be written in the requested language

Auto-fill sections:

• Summary: Commit themes, capped to a short list
• Changes breakdown: Diff stats plus category or top-level area summaries
• Related issues: Parse issue references from commit messages and render close-safe refs as Closes #123 or Fixes owner/repo#456; render non-closing refs as Related to #123
• Checklist: Auto-populate based on change types

Only emit closing keywords when the PR targets the repository default branch. Otherwise, downgrade those references to non-closing Related to ... lines.

After auto-generation, edit the body if necessary. Avoid dumping every changed file into the PR body.

5. Quality Checks

Run scripts/quality_checks.sh to check for:

• Uncommitted changes
• Merge conflicts
• TODO/FIXME comments in changed files
• Missing test files (for significant code changes)
• Large file additions (>1MB)

Run validation commands sequentially when a repository has generated artifacts or known typegen/test interactions. Avoid parallel validation unless the repository is known to support it.

Present warnings to the user but don't block PR creation unless there is a hard failure or the user says otherwise.

6. Present Draft or Create Immediately

By default, present the generated PR body in a clear, editable format and explicitly ask the user to review and approve it.

Example presentation:

I've prepared a PR draft. Please review the content below and let me know if you'd like any changes:

---
[Generated PR body here]
---

Would you like me to:
1. Create the PR with this content
2. Make edits to the description
3. Change the target branch (currently: main)
4. Add/remove reviewers

If the user explicitly requested immediate PR creation, or the calling workflow explicitly says to create the PR without another approval turn, briefly summarize the generated draft and then create or update the PR.

If the draft contains closing keywords, explicitly call out which issues will auto-close when the PR is merged.

7. Create the PR

Once the user approves, or when the user explicitly requested immediate creation, ensure the branch is published and then use gh CLI.

Before creating or updating the PR:

git push -u origin "$CURRENT_BRANCH"

If an open PR already exists for the current branch, update it instead of creating a duplicate:

EXISTING_PR=$(gh pr list --head "$CURRENT_BRANCH" --state open --json number --jq '.[0].number')

if [ -n "$EXISTING_PR" ]; then
  gh pr edit "$EXISTING_PR" \
    --title "PR Title" \
    --body-file "$PR_BODY_FILE"
else
  gh pr create \
    --title "PR Title" \
    --body-file "$PR_BODY_FILE" \
    --base main \
    --head current-branch
fi

If no open PR exists, create it with gh pr create:

gh pr create \
  --title "PR Title" \
  --body-file "$PR_BODY_FILE" \
  --base main \
  --head current-branch \
  [--reviewer user1,user2] \
  [--label label1,label2]

Only suggest reviewers or labels when repository conventions are obvious or the user asks.

Preserve any Closes ..., Fixes ..., or Resolves ... lines in the final PR body so GitHub can auto-close the linked issues on merge.

Generic Classification Strategy

• Use coarse categories that travel well across repositories.
• Prefer confidence-aware output. If the analysis reports low confidence, summarize top-level areas instead of pretending the category split is precise.
• Keep Other as an escape hatch instead of overfitting repository-specific path rules.
• When the category split is weak, prefer diff stats plus the major top-level directories that changed.

Branch naming conventions:

• feature/* → Feature template
• feat/* → Feature template
• bugfix/* or fix/* → Bugfix template
• hotfix/* → Bugfix template
• docs/* → Docs template
• refactor/* → Refactor template
• chore/*, build/*, ci/* → Chore template
• Other → Feature template (default)

Best Practices

See references/pr-best-practices.md for detailed guidelines.

Key points:

• Keep PRs focused (single concern)
• Write clear, descriptive titles
• Explain the "why" not just "what"
• Include testing evidence
• Link to related issues
• Self-review before requesting review
• Prefer concise, reviewer-friendly summaries over exhaustive file inventories

Interactive Editing

If the user wants to edit the draft:

1. Ask which section to modify
2. Update the specific section
3. Re-present the full draft
4. Repeat until approved

Common Commands

User might say:

• "Create a PR" → Full workflow
• "What changed?" → Just analysis (step 1)
• "Draft PR description" → Steps 1-3, present draft
• "Update PR description" → Edit existing draft
• "Check if ready for PR" → Run quality checks

Error Handling

If issues arise:

• No commits ahead of target → Inform user, suggest checking branch
• Merge conflicts → Show conflicts, suggest resolving first
• No GitHub CLI → Inform user, offer to create PR body for manual submission
• API rate limits → Suggest waiting or manual creation
• Requested language not directly supported by a bundled template → fall back to the nearest template, then rewrite the final title/body into the requested language before creating the PR

Output Format

Always save the generated analysis and PR body to skill-local output files for easy editing:

mkdir -p .opencode/skills/pr-assistant/outputs
BRANCH_SLUG=$(git branch --show-current | sed 's#/#-#g; s#[^A-Za-z0-9._-]#-#g')
STAMP=$(date +%Y%m%d-%H%M%S)
ANALYSIS_FILE=".opencode/skills/pr-assistant/outputs/pr-analysis-${BRANCH_SLUG}-${STAMP}.json"
PR_BODY_FILE=".opencode/skills/pr-assistant/outputs/pr-body-${BRANCH_SLUG}-${STAMP}.md"

The local .gitignore in pr-assistant/ should ignore outputs/ so these artifacts stay out of version control.