Find Issues

Analyze the repository at $REPO_URL and find open issues that match your profile.

Your Profile (from config.json)

Profile information is read from ~/.claude/config.json:

• Website: From profiles.website
• Blog: From profiles.blog
• GitHub: From profiles.github

Step 1: Discover Interests (Single Parallel Call)

Launch these two MCP calls in one parallel batch — nothing else:

mcp__github__get_me()
mcp__github__search_repositories(query="user:{username}", sort="updated")

Extract from results:

• Bio, languages, topics from repos
• Primary tech stack (infer from top 10 most recent repos)

Do NOT fetch website/blog — it's slow and rarely adds signal beyond GitHub data.

Step 2: Fetch Issues (Use GraphQL API, NOT Search API)

CRITICAL: Use mcp__github__list_issues (GraphQL), NOT mcp__github__search_issues (REST). The REST search API triggers secondary rate limits (403) when making multiple parallel calls. GraphQL list_issues is more reliable and supports server-side filtering.

mcp__github__list_issues(
  owner="{owner}",
  repo="{repo}",
  state="OPEN",
  since="6 months ago ISO 8601",
  direction="DESC",
  orderBy="UPDATED_AT",
  perPage=50
)

If the result is too large (>50KB), delegate filtering to a sonnet subagent with this prompt pattern:

"Read the file at {path} in chunks. Find issues that are: unassigned, no linked PRs, have labels like bug/feature-request/documentation/open-to-a-pull-request, and match the user's profile: {tech_stack_summary}. Return top 10-15 candidates."

Label strategy: Do NOT filter by labels in the API call. Many repos (including large ones like microsoft/playwright) don't consistently use "good first issue" or "help wanted". Fetch broadly, filter locally by reading labels from results.

Priority labels (in order of contribution-friendliness):

1. open-to-a-pull-request — explicit maintainer invitation
2. good first issue, help wanted — standard contribution labels
3. bug, feature-request, documentation, enhancement — general work
4. No labels — still consider if the issue is well-defined and unassigned

Step 3: Verify No Open PRs (Batched)

Use mcp__github__search_pull_requests to verify candidates have no linked PRs.

CRITICAL constraint: GitHub search allows max 5 AND/OR/NOT operators per query. Batch issue numbers in groups of 4 max:

# Batch 1 (max 4 issue numbers with OR)
mcp__github__search_pull_requests(
  query="repo:{owner}/{repo} is:open {num1} OR {num2} OR {num3} OR {num4}",
  owner="{owner}",
  repo="{repo}"
)

# Batch 2
mcp__github__search_pull_requests(
  query="repo:{owner}/{repo} is:open {num5} OR {num6} OR {num7} OR {num8}",
  owner="{owner}",
  repo="{repo}"
)

Run batches in parallel. If total_count > 0, read results and exclude matched issues.

Step 4: Compile Results

Select max 5-7 issues (one per distinct topic/area). Present as a table:

Difficulty heuristics:

• Easy: Single-file change, UI fix, add a field/property, documentation
• Medium: Multi-file change, new tool/API, requires understanding architecture
• Hard: Cross-cutting concern, protocol changes, complex state management

Step 5: Present & Await Approval

• Highlight top 3 with one-sentence reasoning each
• If no suitable issues found, explain why and suggest alternatives
• Await user's pick before taking any action

Execution Pipeline Summary

[1] get_me + search_repositories  ──parallel──>  Profile ready
[2] list_issues (GraphQL, 50 items)  ────────>  Raw issues
[3] Subagent filters if large  ──────────────>  10-15 candidates
[4] search_pull_requests (batches of 4)  ──parallel──>  PR verification
[5] Compile table + top 3  ──────────────────>  Present to user

Total API calls: ~4-6 (vs 8-12 in naive approach). No rate limit risk.

Rules

• ALWAYS use GitHub MCP for GitHub operations
• NEVER make more than 2 search_issues/search_pull_requests calls in parallel — triggers 403
• Do not suggest issues that already have an open PR
• Do not suggest issues assigned to someone else
• Prefer issues where maintainer has responded (comments > 0)
• If an issue is ambiguous, note it rather than skipping silently
• If working directory IS the target repo, skip the "is it maintained?" check

Usage

/find-issues REPO_URL=https://github.com/microsoft/playwright