Context compression and output preservation

Scope: standalone Runtime. Model windows are finite while history and tool output grow. Compression reduces active history, Spill preserves oversized outputs, and consolidation retains reusable facts.

Separate responsibilities

Mechanism	Input	Goal	Retained form
Compression	Older active history	Make room for recent work	Summary and selected messages
Spill	Oversized tool output	Bound model input	Preview and retrievable original
Consolidation	Reusable information	Preserve facts and summaries	Scoped memory with provenance

Compression does not establish committed memory. Spilled output is not a finalized deliverable.

History compression

Near the effective context budget, older content is summarized while preserving necessary opening content, recent messages, and tool request/result alignment. Work happens on a copy; only a complete usable summary replaces active history.

Failed, truncated, or unavailable summaries retain the original content rather than saving an incomplete head-and-tail history. Preserving data does not guarantee an oversized request fits the model. Restore summarization, reduce input, or explicitly start a new session.

Output preservation

When enabled, Spill stores large tool results in dedicated workspace storage and supplies a preview and reference. Later reads retrieve relevant ranges or pattern matches without resending everything.

References retain access-scope checks. Guessing another reference does not authorize reading it. Capacity and retention determine availability, so references are not permanent document URLs.

Deliverables

When a user requests a document, create and verify a real artifact and provide the file or saved location. Logs, spilled output, and summaries support execution rather than replacing delivery.

See Configuration, Tasks and planning, and Memory.