“Cannot reproduce” to root cause and fix in one working day
How an agent team turned a ticket marked for closure into a root cause, a fix and proof
tr181-periodicfileupload: SIGSEGV (a crash) during … Transfer re-enable
“Unless you disagree, I will close this ticket as 'Rejected - Not a problem' in a week.”
One crash, seen once, and no way to make it happen again
I tried to reproduce the crash locally using the same 130-periodicfileupload.t Cram test on a prplOS freedom board setup.
can we [close] it, it hasn’t been seen lately, cannot be reproduced, the corrupted stack is corrupted.
Unless you disagree, I will close this ticket as 'Rejected - Not a problem' in a week.
This is what a careful engineer can do with one crash, no saved files and no way to trigger it. The limit was the process, not the person.
53 days of waiting against 7 hours 35 minutes
The agent bar, zoomed in 168 times
Calendar time, not working time. Nobody worked on the ticket every day for 53 days. The agents had a human lead for the whole day.
One working day, with a human at every gate
The plugin used memory after it gave it back
Think of a coat check. You leave a coat and get a ticket. Someone returns your coat early, and the hook gets a new coat. Then your ticket comes back, and the clerk hands out the wrong coat.
Turning off a transfer gave its record back while an upload request still pointed at it. When the server replied, the plugin read the slot. On the Freedom board, other data already sat there, so the plugin followed a garbage pointer and crashed.
The unsafe window is a few milliseconds. That is why the crash was so hard to repeat.
The crash now happens on demand, and the fix stops it
Models from different vendors check each other’s work
Why a backup?
AI models sometimes refuse tasks that look like attack research. Crashing a program on purpose and studying its memory errors is also the first step in writing an exploit. So a model can decline even when the work is defensive.
In an earlier project, one reviewer model refused to probe a live crash. We keep a backup with models from other vendors, so the work does not stall. The same rules and human gates apply to it. In this project, no model refused.
A human stayed in charge at every gate
- Before any code ranA planning interview of 36 questions over about 88 minutes, answered between other work. The answers became the rules of the campaign.
- 12:17, the first fix goes backTwo AI review rounds had passed it. The human lead read it and returned it as “overly defensive and overgenerated”.
- 14:05, the reworked fix is approvedOnly then did the agents open the merge requests.
- 15:27, the ticket comment is approvedThe text went out on the human lead’s word, not before.
- After the merge request opened, line coverageThe human lead asked for it. It showed 3 changed lines that no review had flagged.
The human catch
ReturnedLines of test code
The human lead also found internal campaign labels in the test comments. The rework removed them. The final product code change is 54 lines added and 4 removed, in 3 files.
Agent time for the whole case: about 171 USD, at most
At API list prices: 133.88 USD recorded for the Anthropic models, at most 33.19 USD for the OpenAI reviewer, and about 4.21 USD for a known gap in the counting.
Human time: the planning interview ran for about 88 minutes. The lead engineer answered between other work, often from a phone, and made each gate decision the same way. Active human time was not tracked.
The deck gives no cost for the human path. Nobody measured it, and we do not guess it.
A merge request is not the end
2 October 2026
Open
Draft
Draft
Devices do not have the fix yet. They get it only when a prplOS build includes the new version.
The review after the merge request found a gap
Human decision: the lead engineer asked for line coverage in the merge request, so reviewers can see which lines of the fix the unit tests run. A one-line change from the upstream project turned it on.
The 3 lines that no test reaches sit on two rare paths: a hook script (a helper program around an upload) that ends or times out, and a reset in the server reply handler.
No earlier review, by AI or by a human, had flagged them. They are now a review item. The work is not done when the merge request opens.
- PassBuilds
- PassUnit tests
Run this as a normal process, steered by Jira labels
A label is a tag on a Jira ticket. Anyone can filter tickets by it.
Start with the crashes that people would close
The cost of trying is known
At API list prices, this case cost about 171 USD of agent time, at most. A failed try costs the agent spend plus the time a maintainer needs to review it. The ticket then goes back with a written reason.
The gain is a bug that does not stay hidden
A crash closed as “cannot reproduce” stays in the product. Here, the virtual device crashed in up to 76 % of runs once the timing window was widened.
Widen step by step
Add ticket types when maintainers trust the results. Track the cost per ticket, the hand-back rate, and the share of fixes that maintainers accept.
What this case does not prove
- No run on a Freedom boardThe proof comes from QEMU, a virtual device, and from a PC. The original crash was on a Freedom board, and no hardware jobs ran.
- The crash rates are not field ratesThe test server held each reply to widen the timing window. In the automated tests on real hardware, the crash was seen once.
- Devices do not have the fixThe component merge request is open. The feed and prplOS merge requests are drafts.
- The fix has one known limitNo freed memory is read, and there is no crash. If a transfer is deleted and a new one reuses its memory slot, a forced upload started within milliseconds can show as finished early. Turning a transfer off and on again is not affected.
- Active human time was not trackedThe planning interview ran for about 88 minutes. The lead engineer answered between other work, often from a phone, and made each gate decision the same way.
- The cost is an upper boundThe total of about 171 USD uses API list prices. The OpenAI reviewer ran on a flat-rate plan, so its share of at most 33.19 USD cost no extra money.
- Unit tests do not yet reach 3 of the 30 changed lines.
The ticket and the three merge requests
To open these links, you need a prpl login for Jira and GitLab.
The ticket was marked for closure at 07:42 UTC. By 15:27 UTC the same day, it had a root cause, a fix in review and test results. Your turn: which tickets should get the agentic label first?
Build 9ab078e 2026-10-02