agentsop-context-scope-discipline

Context dilution causes coding agents to edit wrong files when the editable working set grows too large. Enforcing a read-vs-edit split and keeping write-permitted files under a ~25k token budget maintains edit accuracy during multi-file refactors.

Manage token budget for multi-file editing by classifying files into read-only context and a smaller edit working set. Use a read-only repo-map for breadth, add only write targets, monitor token usage, and drop completed files to stay under ~25k tokens.

The read vs edit separation pattern delegates breadth to a read-only repo-map of file signatures while reserving full-text, write-permitted files only for actual changes. This keeps the working set focused and prevents context dilution.

Split tasks during API signature changes when the token budget cannot be recovered by dropping completed files. If adding write targets for rename or extract-hook operations pushes the working set over ~25k tokens, divide the refactor into smaller subtasks.

Yes, the context scope discipline approach works without dependencies by applying an operational SOP of classify, locate-then-add, watch, drop, and split-task patterns to manage the working set within any compatible coding agent environment.

To stop diffs landing in wrong files near the dilution threshold, enforce a working-file budget by dropping completed files from the edit working set and splitting tasks when token usage cannot be recovered through file removal.