bluecollarContext
What the model reads on each step, and what happens when it grows.
The system instruction is fixed at the start of the run. Each request carries the requester, the company, the visible conversation, memory facts, the active goal, the plan, a step budget and the observations so far, replayed as native tool calls and results.
A tool result larger than its share of the context has its middle elided. A host that registers a ToolResultSpillStore receives the full output and the model is told where to read it. The share comes from the declared context window.
Compaction starts when the estimated prompt passes 60 percent of the context window, or 96,000 tokens when no window is declared. Long tool results are pruned first. If that is not enough, older observations are summarized into a pinned TaskContextSummary (goal, completed steps, artifacts, key decisions, exhausted routes, active failure debt, next plan), keeping the ten most recent, and a new summary needs at least six new observations and 20,000 new characters.