Context Model

Agent Context uses a simple lifecycle:

sources -> context_surface -> context_selection -> context_budget -> context_assembly -> context_injection -> context_evidence
                              |                     |
                              v                     v
                         missing_context       context_compaction

Objects

Object	Purpose
context_envelope	One portable snapshot for a turn, task, session, tool call, UI review, or peer-agent handoff.
context_surface	The context that was available before selection.
context_item	A unit of context, either inline or by reference.
context_source_ref	Origin, selector, digest, authority, and native id mapping for an item.
context_selection	Candidate, selected, omitted, ranking, and rationale records.
context_budget	Token, byte, item, time, and visibility limits.
context_window	Ordered segments and reserved space for the final context window.
context_assembly	Ordered blocks assembled for a target.
context_injection	Where context became visible: system prompt, message history, tool argument, tool-visible resource, UI panel, runtime metadata, or peer-agent message.
context_compaction	Replacement of original context with a shorter summary or structured digest.
context_summary	Summary content plus source coverage, loss notes, and refresh policy.
missing_context	Known missing inputs, blockers, questions, and requested sources.
context_event	Lifecycle event for surfaces, selections, budgets, injections, compactions, redactions, and exports.

Context kinds

Initial context_kind values:

• user_message
• system_prompt
• developer_instruction
• session_history
• thread_summary
• working_memory
• durable_memory
• knowledge_pack_item
• document_excerpt
• file_excerpt
• artifact_ref
• tool_result
• browser_state
• runtime_state
• team_memory
• policy_note
• external_resource
• manual_note
• computed_summary
• peer_agent_message
• retrieval_result
• custom

Visibility targets

Context may be visible to one or more targets:

• model
• tool
• runtime
• ui
• user
• reviewer
• peer_agent
• evidence_export

A compatible implementation should not assume that context injected into one target is visible to every target.