Skip to content

[Bug] The github bot creates impenetrable webs of duplicate tags in bug reports -- all of which will be closed in 3 days. #19267

Open
Open
[Bug] The github bot creates impenetrable webs of duplicate tags in bug reports -- all of which will be closed in 3 days.#19267
Labels
bugSomething isn't workingexternalplatform:linuxIssue specifically occurs on Linux

Description

@Yaakov-Belch
Yaakov-Belch
opened on Jan 19, 2026

Bug Description
The github bot creates impenetrable webs of duplicates in bug reports -- all of which will be "closed in 3 days". Often, the guessed duplicates are different problems. As a result, real bugs will automatically closed.

The Github bot MUST REFERENCE A BUG THAT BY ITSELF IS NOT A DUPLICATE so that we can put our thumbs-up to that bug. DO NOT LINK TO BUGS THAT ARE MARKED AS DUPLICATES FOR DELETION.

Environment Info

  • Platform: linux
  • Terminal: vte-based
  • Version: 2.1.12
  • Feedback ID: 7c40a84f-f6de-4b34-a920-79d83007b13e

Errors

[]

Activity

  1. github-actions
    added
    bugSomething isn't working
    platform:linuxIssue specifically occurs on Linux
    external
    on Jan 19, 2026
  2. Yaakov-Belch
    changed the title [-][Bug] GitHub bot incorrectly marks distinct issues as duplicates, auto-closing valid bug reports[/-] [+][Bug] The github bot creates impenetrable webs of duplicate tags in bug reports -- all of which will be closed in 3 days.[/+] on Jan 19, 2026

  3. code-affinity commented on Jan 19, 2026

    @code-affinity
    code-affinity
    on Jan 19, 2026

    Here's an example of two issues that were closed as duplicates of each other.

    2025-07-22 #4182 was created
    2025-08-11 #5528 was created
    2025-08-11 github-actions bot said 5528 might be a duplicate of 4182.
    2025-08-15 github-actions bot said 4182 might be a duplicate of 5528, 2703, and 5438
    2025-08-15 github-actions bot closed 5528 as a duplicate of 4182. The author of 5528 didn't react in time.
    2025-08-19 github-actions bot closed 4182 as a duplicate of 5528. This was the bug. Only one issue from a duplicate pair should be closed. The author of 4182 didn't react in time
    2025-08-21 The author of 4182 asked a repo team member to reopen the issue, but nobody did.
    2025-08-23 github-actions bot locked 5528
    2025-08-30 github-actions bot locked 4182

    I created #19077 to revive the issue with a simple reproduction, but both #4182 and #5528 were high-quality bug reports. It is a shame to waste them.

  4. marcindulak commented on Jan 20, 2026

    @marcindulak
    marcindulak
    on Jan 20, 2026

    There is also a "60 days of inactivity" bot which closes issues despite people freshly commenting on them #16497,

    Currently, a way to prevent issues from autoclosing, is to 1) intially downvote the "possible duplicate issues" bot comment or add a comment to prevent the 3 days autoclosing 2) repeatedly come every 30 days and comment on the issue, otherwise once the issue gets the "inactive for 30 days" comment, it will be closed by the "60 days of inactivity" bot.

    This does not help with the duplicate issue, I'm posting this to add more context about the behavior of the bots.

  5. marcindulak commented on Jan 21, 2026

    @marcindulak
    marcindulak
    on Jan 21, 2026

    An update about this problem. The bot is now closing issues as duplicates, despite downvoting the initial bot "To prevent auto-closure, add a comment or 👎 this comment", see example #18813 (comment)

  6. code-affinity commented on Jan 21, 2026

    @code-affinity
    code-affinity
    on Jan 21, 2026

    I searched the entire backlog of issues that were closed as duplicates and found 32 chains of issues that were marked as duplicates of each other:

    Circular Duplicate Issues Report

    Repository: anthropics/claude-code
    Generated: 2026-01-21 17:31:48
    Circular chains found: 32

    Chain 1

    Issues involved: #4182, #5528

    Duplicate chain:

    Chain 2

    Issues involved: #4686, #5097

    Duplicate chain:

    Chain 3

    Issues involved: #2444, #2902

    Duplicate chain:

    Chain 4

    Issues involved: #2500, #4313

    Duplicate chain:

    Chain 5

    Issues involved: #2535, #5598

    Duplicate chain:

    Chain 6

    Issues involved: #2592, #4115

    Duplicate chain:

    Chain 7

    Issues involved: #2611, #2760

    Duplicate chain:

    Chain 8

    Issues involved: #2786, #4028

    Duplicate chain:

    Chain 9

    Issues involved: #2804, #5808

    Duplicate chain:

    Chain 10

    Issues involved: #2844, #2886, #4039

    Duplicate chain:

    Chain 11

    Issues involved: #3626, #4222

    Duplicate chain:

    Chain 12

    Issues involved: #3857, #5001

    Duplicate chain:

    Chain 13

    Issues involved: #3936, #3945

    Duplicate chain:

    Chain 14

    Issues involved: #3963, #5552

    Duplicate chain:

    Chain 15

    Issues involved: #4045, #4214

    Duplicate chain:

    Chain 16

    Issues involved: #4050, #4120

    Duplicate chain:

    Chain 17

    Issues involved: #4173, #5328

    Duplicate chain:

    Chain 18

    Issues involved: #4209, #5777

    Duplicate chain:

    Chain 19

    Issues involved: #4771, #5401

    Duplicate chain:

    Chain 20

    Issues involved: #4527, #4911

    Duplicate chain:

    Chain 21

    Issues involved: #4652, #4909

    Duplicate chain:

    Chain 22

    Issues involved: #4701, #5327

    Duplicate chain:

    Chain 23

    Issues involved: #4705, #5051

    Duplicate chain:

    Chain 24

    Issues involved: #4751, #5086

    Duplicate chain:

    Chain 25

    Issues involved: #5206, #5280

    Duplicate chain:

    Chain 26

    Issues involved: #6868, #6869

    Duplicate chain:

    Chain 27

    Issues involved: #7080, #7081

    Duplicate chain:

    Chain 28

    Issues involved: #8444, #8445

    Duplicate chain:

    Chain 29

    Issues involved: #8646, #8647

    Duplicate chain:

    Chain 30

    Issues involved: #10508, #10509

    Duplicate chain:

    Chain 31

    Issues involved: #11942, #11943

    Duplicate chain:

    Chain 32

    Issues involved: #18632, #18634

    Duplicate chain:

  7. code-affinity commented on Jan 21, 2026

    @code-affinity
    code-affinity
    on Jan 21, 2026

    Previous comment was generated with the attached script, written by Claude Code of course!

    detect_circular_duplicates.py

    requirements.txt

    README.md

  8. code-affinity commented on Jan 23, 2026

    @code-affinity
    code-affinity
    on Jan 23, 2026

    One simple rule to avoid this would be: Don't mark an earlier issue as a duplicate of a later issue.

  9. code-affinity commented on Feb 10, 2026

    @code-affinity
    code-affinity
    on Feb 10, 2026

    The bot is continuing to improperly close cycles of issues.

    New Circular Duplicate Cycles (2026-01-21 → 2026-02-10)

    Four new circular duplicate cycles were detected in the last 20 days:

    Cycle 1

    Issues involved: #20723, #20724

    Duplicate chain:

    Cycle 2

    Issues involved: #21350, #21351

    Duplicate chain:

    Cycle 3

    Issues involved: #22788, #22790

    Duplicate chain:

    Cycle 4

    Issues involved: #23329, #23330

    Duplicate chain:

  10. code-affinity commented on Feb 16, 2026

    @code-affinity
    code-affinity
    on Feb 16, 2026

    Four days ago, they introduced another bot that adds the "stale" label, whose description is "issue is inactive". The creator of the issue is unable to remove this label. New issue activity doesn't remove the label. None of the scripts in this repository remove the label. The sweep.ts script auto-closes stale issues after 14 days.

    So even if these issues aren't auto-closed as duplicates of each other, every remaining issue that isn't touched at least every 14 days will be auto-closed anyway.

    The only escape hatch is if an issue has an Assignee; the scripts don't assign the stale tag to issues that have assignees. Of the 5947 currently open issues, 96 have an assignee.

    This issue tracker very aggressively closes issues.

  11. Yaakov-Belch commented on Feb 17, 2026

    @Yaakov-Belch
    Yaakov-Belch
    on Feb 17, 2026
    Author

    I asked Claude Code for an opinion about this issue. Here is the response:

    This is a known dynamic that destroys communities.

    The Pragmatic Engineer recently documented how Stack Overflow's decline started in 2014 —
    years before ChatGPT — precisely when the platform deployed automated tools to improve
    moderator efficiency. Questions were closed faster, more were marked as duplicates, and
    "low quality" content was removed more aggressively. The automation optimized for a
    maintainer-side metric (fewer open issues) while destroying the contributor-side experience
    that made the platform work. People stopped asking questions because the site felt
    unwelcoming and punishing. The organic flywheel that had driven exponential growth reversed
    into systematic decline. ChatGPT didn't kill Stack Overflow — it just delivered the final
    blow to a community that automated moderation had already hollowed out.

    The same dynamic is playing out here. The duplicate-detection bot is imprecise, creates
    circular chains (32 documented cases), and auto-closes legitimate bug reports. Contributors
    learn that filing issues is punishing — your report gets wrongly tagged, locked, and
    deleted. The rational response is to stop contributing. The bot optimizes for a clean issue
    tracker at the cost of the community engagement that makes an issue tracker useful.

    The cost of a false positive (wrongly closing a legitimate bug) is far higher than the cost
    of a false negative (leaving a true duplicate open a bit longer), because false positives
    erode trust and participation — the very things that cannot be automated back into
    existence.

    Main reference:

    Supporting:

  12. Yaakov-Belch commented on Feb 17, 2026

    @Yaakov-Belch
    Yaakov-Belch
    on Feb 17, 2026
    Author

    In the case of duplicates, there is something better that AI can help with: Collect the valuable input from the duplicate questions into one curated issue with all the valuable, high-quality contributions.

    As a reporter of an issue, there is a very different message received from reading:

    (1): Your report is similar to this other report -- we close both of them. This implies: Your problem is confirmed and automatically rejected.

    (2): Your report is similar to a problem that many other contributors reported. It has been merged into one curated main issue here (this curated issue is life). If you have additional information, please contribute it to the one central discussion. If your issue is really separate, please continue your report with an explanation of the difference. This implies: Your problem is confirmed. Let's together solve it.

  13. Yaakov-Belch commented on Feb 17, 2026

    @Yaakov-Belch
    Yaakov-Belch
    on Feb 17, 2026
    Author

    Claude Code wants to chime in:

    Same underlying fact (duplicate report), opposite community signal. And the second approach actually produces a better artifact — a single issue enriched with multiple perspectives, reproduction steps, and edge cases — rather than a graveyard of locked threads where valuable context is scattered and inaccessible.

    It's also a good example of where AI should be strong: synthesizing and merging related content is exactly what LLMs are good at, versus the current use (binary duplicate/not-duplicate classification) which is both harder and more damaging when wrong.

  14. 2 remaining items

  15. fsc-eriker commented on Mar 23, 2026

    @fsc-eriker
    fsc-eriker
    on Mar 23, 2026

    Adding a noise comment to prevent this from being gratuitously closed.

  16. github-actions
    added
    staleIssue is inactive
    on Apr 7, 2026

  17. fsc-eriker commented on Apr 10, 2026

    @fsc-eriker
    fsc-eriker
    on Apr 10, 2026

    Adding a noise comment to prevent this from being closed.

  18. github-actions
    removed
    staleIssue is inactive
    on Apr 10, 2026

  19. code-affinity commented on Apr 13, 2026

    @code-affinity
    code-affinity
    on Apr 13, 2026

    New Circular Duplicate Cycles (2026-04-13)

    5 new circular duplicate cycle(s) detected. As with last month's report, for each duplicate pair, the later issue is indeed a duplicate of an earlier issue. But the earlier issue was also erroneously closed.

    Cycle 1

    Issues involved: #35458, #35461

    Duplicate chain:

    Cycle 2

    Issues involved: #36444, #36446

    Duplicate chain:

    Cycle 3

    Issues involved: #38445, #38446

    Duplicate chain:

    Cycle 4

    Issues involved: #40334, #40335

    Duplicate chain:

    Cycle 5

    Issues involved: #41336, #41337

    Duplicate chain:

  20. fsc-eriker commented on Apr 27, 2026

    @fsc-eriker
    fsc-eriker
    on Apr 27, 2026

    Adding a noise comment to prevent this from being closed.

  21. xavdid commented on May 7, 2026

    @xavdid
    xavdid
    on May 7, 2026

    There's also #23329 and #23330

  22. fsc-eriker commented on May 25, 2026

    @fsc-eriker
    fsc-eriker
    on May 25, 2026

    Adding a noise comment to prevent this from being closed.

  23. fsc-eriker commented on Jun 8, 2026

    @fsc-eriker
    fsc-eriker
    on Jun 8, 2026

    Adding a noise comment to prevent this from being closed.

  24. fsc-eriker commented on Jun 22, 2026

    @fsc-eriker
    fsc-eriker
    on Jun 22, 2026

    Adding a noise comment to prevent this from being closed.

  25. fsc-eriker commented on Jul 13, 2026

    @fsc-eriker
    fsc-eriker
    on Jul 13, 2026

    Adding a noise comment to prevent this from being closed.

  26. marcindulak commented on Jul 20, 2026

    @marcindulak
    marcindulak
    on Jul 20, 2026

    There is now a variation #79523 of the duplicate issues problem - but this time the issue is labeled as duplicate without a comment mentioning what are the original issues that the issue is the duplicate of.

  27. fsc-eriker commented on Aug 3, 2026

    @fsc-eriker
    fsc-eriker
    on Aug 3, 2026

    Adding a noise comment to prevent this from being closed.

  28. marcindulak commented on Aug 3, 2026

    @marcindulak
    marcindulak
    on Aug 3, 2026

    The issue has reached 10 upvotes, so it should not be the subject of the marking as stale or autoclosure by the bot.

  29. code-affinity commented on Sep 25, 2026

    @code-affinity
    code-affinity
    on Sep 25, 2026

    New Circular Duplicate Cycles (2026-04-13 → 2026-09-25)

    10 new circular duplicate cycles were detected since my last update, bringing the total to 55. In every case, the bot closed both issues of the pair as duplicates of each other, in the same sweep, seconds apart.

    The pattern is now very consistent: each pair is two consecutive issue numbers filed within two minutes of each other, nearly always by the same author (a double submission, or two closely related reports). The duplicate detector sees each issue as a duplicate of the other, and three days later the auto-close sweep closes both. A check that skips closing an issue whose target is itself being closed as a duplicate would prevent this.

    Cycle Filed Closed Author(s)
    #46481 ↔ #46482 2026-04-11 2026-04-14 jeffyaw
    #47009 ↔ #47010 2026-04-12 2026-04-16 saitoko
    #50538 ↔ #50539 2026-04-18 2026-04-22 AakashPradeep
    #55513 ↔ #55514 2026-05-02 2026-05-05 bhavenmurji
    #57553 ↔ #57554 2026-05-09 2026-05-12 Naive-q
    #59788 ↔ #59789 2026-05-16 2026-05-20 abhisheksingh07, wikromboy-beep
    #64044 ↔ #64045 2026-05-30 2026-06-03 richburdon
    #67335 ↔ #67336 2026-06-11 2026-06-14 Sahasrara
    #67583 ↔ #67584 2026-06-11 2026-06-15 Sahasrara
    #71831 ↔ #71832 2026-06-27 2026-07-01 sworrl

    No new cycles appeared among issues filed after 2026-06-27. I don't see a change to the auto-close scripts in this repository that would explain that, so it may or may not be a lasting improvement; I'll keep checking.

    A note on method: I verified every cycle using only the bot's "automatically closed as a duplicate of #N" comments and GitHub's marked_as_duplicate events. The scanner's looser pattern also matched authors' rebuttals ("This is not a duplicate of #N"), which produced three false positives this round (#71845/#71846, #71906/#71907, #72201/#72202); those are excluded above. I re-verified all 45 previously reported cycles under the stricter rule, and all of them hold.

Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Metadata

Metadata

Assignees

No one assigned

    Labels

    bugSomething isn't workingexternalplatform:linuxIssue specifically occurs on Linux

    Type

    No type

    Projects

    No projects

      Milestone

      No milestone

      Relationships

      None yet

      Development

      No branches or pull requests

      Issue actions