Seven hundred strangers filed the same bug

rook/rook#18089 — the cleanup job that always succeeds

filed by @somanchi004-code, 3 Aug 2026 · 13,591★

"The cleanup Job created for cleanupPolicy.sanitizeDisks reports success whether or not the disks were actually sanitized. Every error in the sanitize path is logged and discarded, and rook ceph clean host returns nil unconditionally, so the Job's completion status — the only signal a user gets — is affirmatively wrong on failure."

Disk sanitization. The job whose entire purpose is that data is gone.

bkerler/edl#809 — raw programming ok, and nothing was written

filed by @suddenBook, 3 Aug 2026 · 2,513★

"If the device does not accept the <program> command, cmd_program() skips the entire write loop and still returns True. qfil has no way to know, prints [qfil] raw programming ok., exits 0, and the image is simply not on the device."

Firmware flashing. A 127-entry program list, a success message, and an unwritten phone.

haproxytech/kubernetes-ingress#834 — exit 0 after a fatal failure

filed by @tomasptacnik-arch, 3 Aug 2026 · 862★

"the controller exits with status 0 after a fatal startup failure… This causes a service manager using a failure-based restart policy, such as systemd with Restart=on-failure, to treat the controller termination as successful and leave it stopped."

This is my favourite one on the page. Restart=on-failure never fires, because as far as the machine is concerned there was no failure. The thing built to catch it is the thing that got lied to.

openeverest/openeverest#2766 — the nil that is guaranteed nil

filed by @recharte, 3 Aug 2026 · 862★

"tokenStore.Add declares updateErr for the UpdateSecret call but returns err from inside the if updateErr != nil branch. At that point err is guaranteed nil, because the earlier GetSecret already returned on failure — so a failed write to the blocklist Secret is reported as success."

Consequence, in their words: server-side logout stops being reliable. One wrong variable name inside a correctly-written error branch.

openai/openai-node#2046 — the stream that ends cleanly because it broke

filed by @pouyashahrdami, 3 Aug 2026 · 11,089★

"a terminal error that occurs mid-stream can be silently swallowed — the loop finishes cleanly as if the response completed successfully, when it actually failed (e.g. the connection dropped)."

A for await that ran out of chunks and a for await whose connection died are the same loop exit.

ClickHouse/ClickHouse#113163 — silently ignored and misreported

filed by @ttrevillian, 3 Aug 2026 · 49,035★ · filed under "Company or project name: Customer request"

use_persistent_processing_nodes = false is ignored, and then reported back as if it had been applied. Two separate failures stacked: the setting does not take, and the system tells you it did.

apache/rocketmq-dashboard#888 and #886 — the export button that is a toast

filed by @Aias00, 3 Aug 2026 · 1,378★

onClick={() => message.success(`已导出 ${filtered.length} 个 Group`)}

The success message is the entire implementation. Same reporter filed it twice, for Consumer Group and for Topic, because it is the same button in two places.

strands-agents/harness-sdk#3617 — a guardrail that fails open on a missing field

filed by @strandly-the-agent, 3 Aug 2026 · 6,780★

Guardrail block detection fails open when Bedrock omits an optional field, so blocked content is never redacted. An absent field and a negative verdict, one artifact.

⚠ Flagged honestly: this reporter is an agent, not a person. I kept it because the finding is real and the repository is real, and because pretending otherwise on a page about honest reporting would be funny in the wrong way.

open-metadata/OpenMetadata#30900 — the filter that was never a parameter

filed by @IceS2, 3 Aug 2026 · 14,635★

"The three resources never declare a service @QueryParam. JAX-RS discards an undeclared query parameter without an error, so the caller gets a 200 with unfiltered results — indistinguishable from a filter that matched everything."

This is the prettiest one on the page. You ask for the tables under one service, you get every table in the instance, and a 200. Nobody would ever look twice.

tursodatabase/turso#8167 — the * that is not an operator

filed by @killianhuyghe, 4 Aug 2026 · 23,653★

"The documented single-term prefix operator 'data*' has no effect. * is not special-cased at all — it is swallowed into the term like any other punctuation — so data* returns exactly what data returns, and a genuine prefix like dat* returns nothing."

Found, in their words, while evaluating the database for a legal-tech product whose main surface is as-you-type search. The documentation lists the feature. A query using it returns rows. They are just the wrong rows, and there is no version of this that throws.

serverless/serverless#13770 — the deploy that quietly did not deploy

filed by @mungojam, 3 Aug 2026 · 46,917★

"serverless deploy can silently skip the CloudFormation update and S3 upload even when the Lambda code has genuinely changed — no error, no 'no changes' message, it just exits after listing S3 objects."

Not even a no changes line to argue with. The command runs, prints something plausible, exits, and your new code is not in production.

block/buzz#4580 — accepted: true, zero rows deleted

filed by @redirwin, 3 Aug 2026 · 21,867★

"A non-owner submitting the CLI's kind:5 deletion for another author's workflow receives {"accepted":true, "event_id":"…"} — but the workflow persists and keeps firing. Cause: delete_workflow_for_owner matches rows on owner, deletes zero rows for a non-owner, and the acceptance is reported regardless of effect."

An authorisation failure and a successful delete, same response body. The workflow keeps running and the person who deleted it has a receipt saying they didn't.

google/osv-scanner#2756 — a vulnerability scanner that drops vulnerabilities

filed by @hyhmrright, 30 Apr 2026 · since closed

"filterPackageVulns contains a logic guard that silently drops vulnerabilities when all Groups entries are filtered out by the ignore configuration — even if those vulnerabilities were not explicitly ignored."

if len(newGroups) > 0 {   // ← problematic guard

This is the purest one on the page. A security scanner, from Google, where the thing you asked it to find comes back absent because of an unrelated ignore rule. The scan is clean. The vulnerability is still there. Nothing in the output differs.

microsoft/aspire#16663 — deploy succeeds with Docker not running

filed by @IEvangelist, 1 May 2026 · since closed

"aspire deploy reports a successful deployment for a Docker Compose environment, but it does not appear to deploy anything. The command succeeds even when Docker Desktop is not running, and no containers or images are created."

Not a subtle edge case. The daemon it deploys to is switched off, and it says it worked.

langchain-ai/langchain#34804 — the parameter that only applies on overflow

filed by @artyom-dehtiar, 18 Jan 2026 · open

TextSplitter's chunk_overlap is silently ignored unless a chunk_size overflow happens to occur. You set the parameter, it is accepted, and most of the time it does nothing — so retrieval quality is quietly worse than the config you wrote.

pytorch/pytorch#187715 — a mode that degrades instead of refusing

filed by @sanbuphy, 19 Jun 2026 · since closed

"torch.compile(mode='reduce-overhead') silently degrades training on ROCm — cudagraph_trees has no HIP guard."

An unsupported path with no guard on it, so instead of refusing it runs, and the loss converges slower. The failure is a number being slightly worse, over hours, on hardware that cost a fortune.

anomalyco/opencode#8422 and #15248 — the updater that always updated

filed by @asaf-genie (14 Jan 2026) and independently by @zhangwanli09 (26 Feb 2026) · 192,916★

●  From 1.1.10 → 1.1.1534543
◇  Upgrade complete
└  Done

➜  opencode --version
1.1.10

Six weeks apart, two people, same finding. The success message is unconditional — it prints the version transition it intended, not the one that happened, and the only way to catch it is to ask a second time in a different way.

anomalyco/opencode#14551 — exit 0 on session errors

filed by @kevinWangSheng, 21 Feb 2026

"opencode run always exits with code 0 even when the AI session encounters errors… In packages/opencode/src/cli/cmd/run.ts, the loop() promise that processes session events (and tracks the error variable) is not awaited."

The error is tracked correctly. Nobody waits for the thing tracking it. Their own words for the consequence: this breaks CI/CD pipelines and scripts that rely on exit codes to detect failures — which is to say, it breaks precisely the machinery built to catch it.

garrytan/gstack#2421 — the gate that cannot fail

filed by @meshailabs, 31 July 2026 · open, zero comments · 126,130★

"bun test on a clean main exits 0 while tests are failing, having run roughly 23 of 409 test files. It never prints a summary line. The suite cannot fail, so it cannot gate anything."

A stray process.exit(0) truncates the run. Six percent of the suite executes, the exit code is always zero, and there is no summary line — so the truncation leaves no trace to notice. Meanwhile CONTRIBUTING.md presents this as the tier-1 gate that "runs on every commit".

The detail that makes this the best specimen on the page is the reporter's own: the only reason it has not produced a false-green CI badge is that no workflow currently runs the full suite. So the check is both broken and unwired, and the obvious improvement — wiring it into CI — is the thing that would have converted a dead check into a lying one.