context-budget-discipline

To manage context budget in long-running AI sessions, use four operations: SELECT to load context just-in-time, WRITE to persist decisions for durable handoffs, COMPRESS to summarize older turns, and ISOLATE to delegate sub-work to fresh subagents, preventing token inefficiency and context bleed.

Agents lose decisions in long sessions because context windows shift and drop earlier rationale. Persisting decisions and rationale to disk for durable handoffs ensures that critical context survives token compression and can be reloaded just-in-time when needed.

The best way to prevent context bleed between sub-tasks is to isolate them by delegating sub-work to child agents in fresh context windows. This ISOLATE operation prevents contamination from unrelated context, keeping sub-tasks focused and reliable.

To handle large files without exceeding the context window, apply just-in-time reads by selectively loading context only when needed. This SELECT operation avoids reading everything at once, preserving the token budget for long-running sessions spanning many features.

You should summarize older turns when the context window tightens and token inefficiency threatens session reliability. The COMPRESS operation summarizes older turns to free up context budget while retaining essential information for ongoing work.

Yes, you can maintain context budget across sessions spanning many features by selectively loading context, persisting decisions, and isolating sub-tasks. This approach enables durable handoffs and fresh-context subagents to prevent context bleed under token constraints.