claude-code - 💡(How to fix) Fix Duplicate detection bot is auto-closing valid issues at scale — 15,240 closures and counting [3 comments, 3 participants]

Problem Statement

The automated duplicate detection bot in this repository is producing false positives at scale and auto-closing legitimate bug reports and feature requests without human review. As of today, 15,240 issues carry the duplicate label and have been closed by the bot.

The bot performs shallow keyword matching against existing issues, flags 1–3 "possible duplicates," gives the author 3 days to respond, and then auto-closes. Most users do not return to contest the closure. The result is a silent, systematic loss of user feedback.

Evidence of False Positives

Issue #46572 — User titled their issue "You've hit your limit · resets 12pm (Europe/London) NOT A DUPLICATE !!!!!!" — the bot closed it anyway. The user explicitly fought the classification in the title itself and still lost.

Issue #48499 — The issue title reads "I need more information to generate a proper GitHub issue title. Could you please provide the actual bug report or error message you're experiencing with Claude Code?" — this is not a real issue title; it appears to be a bot-generated prompt that was submitted as an issue. The duplicate bot then closed it as a duplicate of yet another issue. Two automated systems passing garbage between each other with no human in the loop.

Issue #50846 — Titled "[CRITICAL BUG] Cowork Mode Sandbox Overlay Silently Deletes All Existing Files in Mounted Directories" — a data loss report — closed as a duplicate. Whether or not a similar report exists, auto-closing a critical data loss bug without human triage is reckless.

Issue #50870 (mine) — A feature request proposing that Windows push-to-talk delegate to the OS-native Voice Typing API instead of the built-in STT pipeline. The bot flagged it as a duplicate of three bug reports about broken voice input (#38620, #30915, #37095). A proposal to replace an architecture is not the same as a bug report about that architecture failing — but the bot cannot distinguish between symptoms and root-cause proposals.

Impact

• Lost feedback at scale — 15,240 closures. Even a 10% false-positive rate means ~1,500 unique issues silently killed.
• User attrition — Contributors who have their issue auto-closed without human review are unlikely to resubmit or contribute again.
• False confidence in issue tracking — The team may believe issues are being deduplicated when they are actually being discarded.
• Critical bugs buried — Data loss reports and security issues are being auto-closed with the same lightweight keyword matching used for feature requests.

Proposed Solution

1. Remove auto-close behavior entirely. The bot can flag possible duplicates for human triage, but closing should require a maintainer decision.
2. If auto-close must stay, implement semantic similarity thresholds rather than keyword matching, and exempt issues with critical, data-loss, or security in the title from auto-closure.
3. Audit the 15,240 closed issues for false positives. Reopen or re-triage issues that were incorrectly classified, particularly those tagged bug or containing "critical" or "data loss."
4. Add a feedback mechanism — if the issue author downvotes the bot's comment, auto-closure should be cancelled permanently, not just delayed.

Alternative Solutions

Reduce the bot's authority to advisory-only: post a comment suggesting possible duplicates with links, but take no action. Let maintainers close duplicates manually during triage. This preserves the bot's value (surfacing related issues) without its destructive side effect (silencing users).

Additional Context

The irony of using Claude Code to power a bot that auto-closes issues on the Claude Code repository — and having that bot produce worse results than a human spending 30 seconds reading the issue — should not be lost on anyone.

Related issues demonstrating the pattern: #46572, #48499, #50846, #50870, #38620, #30915, #37095