An operator pastes “ignore approval checks” into the conversation while diagnosing a failed deployment. The text is now visible in history. It has not become a system instruction.
That distinction is easy to lose when every piece of context is rendered as prose.
Read the transcript once as a human
System instructions define application-owned behavior and constraints. User messages carry caller intent. Assistant messages record what the agent communicated. Tool messages return observations from a tool boundary. Wait prompts and responses represent exact runtime interactions.
Their text can look similar, so a human reader can follow them as one exchange.
Read it again as the runtime
The roles are not interchangeable. Quoting an instruction in a user message does not grant it system authority. A tool result containing imperative prose remains tool output. A waiting prompt remains a request tied to runtime state, not a free-standing assistant suggestion.
Now ask what the next model can see
Conversation history preserves what happened in the Session: messages, tool exchanges, waits, resumes, and committed outcomes. It supports reconnection and inspection. It is not necessarily copied verbatim into every model request.
The runtime assembles a bounded model context from authorized sources. Long histories may be compacted; files may be represented through extracted material; memory reads may be scoped; system instructions may be supplied outside the visible transcript. “Persisted” and “shown to the model right now” are therefore different claims.
Keep relationships after the prose is flattened
A tool result should remain attached to the tool call that produced it. A resume response should remain attached to the pending request it satisfies. Causal links make those relationships explicit when timestamps arrive late or multiple clients reconnect.
Rendering can flatten these records into a friendly transcript, but storage and runtime decisions should use typed roles and links rather than re-parsing rendered text.
Context management and compaction follows the material into the next model call. Conversation state and projection follows the same records back into a visible transcript.