Hosted search_issues silently returns incorrect keyword results

先針對託管的 MCP 端點重現列出的 search_issues 查詢，並將其與 GitHub REST GET /search/issues 進行比較。接著追蹤 search_issues 的實作與託管請求路徑；當關鍵字查詢保留 GitHub 搜尋語意，或回傳明確的不支援查詢／錯誤訊號，而不是靜默回傳錯誤的成功結果時，即視為完成。

Bug description

The hosted GitHub MCP search_issues tool silently returns materially incorrect keyword search results while reporting incomplete_results: false.

This reproduces when calling the hosted endpoint directly, without an MCP aggregator.

Environment

• Endpoint: https://api.githubcopilot.com/mcp
• Server: github-mcp-server
• Reported version: github-mcp-server/remote-34ac9c16f8fd3e2a5af74929de89de3a11e98f9b
• Protocol negotiated: 2025-11-25
• Date verified: 2026-08-31

Reproduction

Call search_issues with ordinary GitHub issue search syntax already scoped to issues.

The three returned hooks results were titled plugins, trap, and Hook before Exit, which also does not match normal GitHub full-text search behavior.

Passing owner and repo separately with a keyword-only query produced the same incorrect results, so this is not limited to embedding repo: in query.

Qualifier-only searches such as repo:modelcontextprotocol/python-sdk label:bug return plausible counts, which makes the keyword failure easy for an agent to trust silently.

Expected behavior

search_issues should preserve GitHub issue search semantics and return counts/results consistent with the GitHub REST Search API for the same effective query, or return an explicit error/unsupported-query signal.

Actual behavior

The tool returns a successful response with incomplete_results: false, but keyword matches are missing or unrelated.

Impact

Agents can incorrectly conclude that an issue, bug report, or prior discussion does not exist. This is a correctness failure rather than only a ranking-quality difference.