Context Management

Coverage

The working discipline that controls what enters, stays in, and exits an active agent session. Intake triage that sorts every candidate context source into a four-bucket classification (must-have / useful soon / durable background / noise) before any large file is read. The six-step context-management loop: state the active question in one sentence, name the minimum evidence needed to answer it, load the cheapest sources first (index → search → narrow file slice), collapse confirmed facts into a checkpoint, drop disproven assumptions from the active thread, re-check whether the question changed before reading more. Working-set shaping rules — what to keep active vs what to push out — and the distillation pattern that converts a 300-line log into a 2-line summary, a whole file into a function name plus slice plus invariant, a long conversation into current-state-blocker-next-step. Drift detection signals (re-reading the same file, ideas changing every turn, search-space unbounded, the agent forgetting what was proven) and the anti-drift rules (one active hypothesis at a time, one primary question, one verification target). The compaction-ready handoff format with five required fields (task / question / proven facts / rejected paths / next step) and the under-thirty-seconds resume test. The selective-rebuild recipe for recovering after the thread is lost.

Philosophy

Context management is the practical layer between having the right information available somewhere in the workspace and having it active in the agent at the right moment. The goal is not to load more context — it is to keep the smallest working set that still lets the agent act correctly. Without this discipline, agents speculate from stale assumptions, re-read files they already processed, and lose the decision trail at the moment of compaction. Every context slot occupied by noise is a slot unavailable for the evidence that would actually resolve the current question.

The hardest part is not what to load. It is what to drop. Disproven hypotheses, raw logs after the key pattern is extracted, full files after the needed lines are identified, alternative hypotheses that have already been falsified — all of these continue to occupy context until they are deliberately removed. The working set is what the agent is actively reasoning over, not everything it has ever seen.

1. Outcomes