context-budget

Long tasks rarely fail at the start; they fail at 80%, when the window is stuffed with re-read files and the original plan has scrolled out of view. context-budget treats the context window as a finite budget: read the head first to learn the shape, then the specific section, then targeted search. A full read is a deliberate purchase, made only when the head read proves the middle matters. Established facts, decisions, and file paths land in a working notes file one line each. At phase boundaries, restate the plan and current state in three lines. Together, these habits let an agent stay grounded after compaction instead of running on fumes.

The core capability boils down to three mechanical habits: graduated reads (head, section, search), notes to disk (notes beat memory), and checkpoint restatement. The notes file survives compaction; the conversation does not, making it a reliable anchor across runtime resets. Checkpoints keep direction intact at phase transitions.

Best suited for agent workflows that span many steps and files. For tasks under roughly ten steps, the small write overhead of the budget machinery outweighs the savings, and the plugin can be safely skipped.