#19658: feat(skills): add overlap-check — search before filing issues or PRs

Agents don't check what already exists before filing issues or opening PRs. This leads to duplicate threads, competing PRs for the same fix, and comments that rehash solved discussions. Example: anthropics/claude-code#13738 has 16+ comments and community workarounds for WSL2 clipboard paste. Meanwhile #25935 filed a one-line fix for the same root cause and got auto-closed as a duplicate. #14635 covers overlapping ground too. Three threads, same problem, nobody searched first. This skill teaches agents to search before creating. It fires when the agent is about to run `gh issue create` or `gh pr create`, runs two `gh search` queries (issues + PRs), and shows any matches. The agent then decides: comment on existing, or proceed with a new thread. **What it does:** - Extracts keywords from the proposed title - Searches the target repo for open issues and PRs with those keywords - Shows top 5 matches with comment counts - Lets the agent (or user) decide what to do **What it doesn't do:** - No LLM calls, no semantic matching — just keyword search via gh - No blocking — warns, doesn't prevent - No heavy analysis — that's what upstream-recon is for (separate, manual tool) Two gh search calls, under 5 seconds, zero cost. I looked into delivering this as a PreToolUse hook (enforced gate instead of behavioral skill), but OpenClaw's hook system fires on gateway events (command:new, gateway:startup), not tool calls. Issue #1733 proposed tool:pre but it was closed without implementation. So this ships as a skill — agents are instructed to check, not forced to. If tool:pre lands later, this could upgrade to an enforced gate. Refs #19606 Written by Claude Code, guided by SeMmy <!-- greptile_comment --> <h3>Greptile Summary</h3> Adds a new `overlap-check` skill that instructs agents to search for existing GitHub issues and PRs before creating new ones, reducing duplicate threads and competing PRs. - The skill is well-structured: it identifies the target repo, runs two `gh search` queries (issues + PRs), evaluates matches, and lets the agent/user decide whether to proceed or contribute to an existing thread - Follows existing skill conventions (frontmatter with `name`, `description`, `metadata`; concise markdown body) - One bug: `gh pr view NUMBER` on line 50 is missing the `--repo OWNER/REPO` flag, inconsistent with the `gh issue view` command on the same line — this would cause failures when the agent is not inside the target repo's directory - Lightweight addition (single markdown file, no scripts or dependencies beyond `gh`) <h3>Confidence Score: 4/5</h3> - This PR is safe to merge with one minor fix needed — it's a behavioral skill (markdown instructions only) with no code execution risk. - Single markdown file addition with no code changes. The skill is well-written and follows project conventions. One command inconsistency (missing --repo flag) could cause the agent to view the wrong PR when working outside the target repo, but this is a non-blocking behavioral instruction, not executable code. - skills/overlap-check/SKILL.md — line 50 has a missing --repo flag on gh pr view _{Last reviewed commit: 545c372} <!-- greptile_other_comments_section --> _{(2/5) Greptile learns from your feedback when you react with thumbs up/down!} <!-- /greptile_comment -->