Example
.macroscope/check-run-agents/api-review.md:
Output
Each Check Run Agent reports in GitHub check details. Withmodify_pr, it can also post inline review comments and conversation comments. Without that tool, it cannot post either kind of comment.
Instructions control report structure. For example: “List findings by severity, with file, line, impact, and suggested fix.” showToolCalls: false hides the tool-call log from check details.
Importing files
Write@path/to/file.md in the instruction body to insert that file’s contents. Imported front matter is stripped. Imports do not expand inside front matter, inline code, or fenced code blocks.
Paths resolve relative to the importing file unless they start with
/. Any text file can be imported. A directive must contain / or an extension; handles, decorators, and email addresses are left alone.
Imports support 4 hops and 50 imports per prompt. Paths must stay inside the repository and outside .git; symbolic links are not followed. Keep shared Markdown outside .macroscope/check-run-agents/ if you do not want it discovered as another check.
Missing paths, cycles, or exceeded import limits leave the directive as literal text and add a warning to check details. The check still runs. Imported content has no separate size cap, but the complete prompt must fit the model’s context window.
Imports come from the same commit as the definition, including the default-branch exception for external forks.
Context from prior runs
Every run starts a fresh model conversation with context from prior PR activity:- Its own review threads, including replies and resolution state.
- Recent comments from others when it has little discussion history of its own.
- Its prior actions, including comments, Slack messages, labels, and reviewer assignments.
Context window
Instructions and imports are included in full. The remaining space is shared by tool schemas, PR context, discussion, and diff or commit messages.
A context failure is titled Prompt too large for the selected model. Choose a larger-context model or shorten instructions. Narrowing
include/exclude does not shrink a full_diff prompt.