Home / Context and cost

The four levers of a long session

A session that runs long rests on just four levers: the cache, which reuses a stable prefix, clearing out stale tool outputs, compaction, which summarises the history, and partitioning, which delegates bulky work to isolated sub agents.

A Claude Code session that runs for several hours does not endlessly swell the same block of text sent to the model with every message. Four distinct mechanisms limit what that session drags along behind it: the cache, which reuses an identical start of conversation rather than recomputing it, clearing out tool outputs that have become useless, compaction, which replaces the history with a summary, and work delegated to sub agents, separate conversations that handle part of the task and return only their conclusion.

The cache, the sort, the summary, the delegation

The prompt cache retains the computation done on a start of conversation, called the prefix, for as long as that prefix stays identical down to the byte from one message to the next. As soon as anything changes at the very start of that prefix, everything that follows has to be recomputed. That is why position matters: general instructions and project context are best kept at the front, unmoving, while questions that change come afterwards.

Clearing out stale tool outputs, what this course calls masking, clears away the results of old calls that no longer serve the current reply, a real mechanism even though no public Anthropic documentation names it in exactly those words. Compaction, meanwhile, summarises the message history into shorter text once the conversation nears its limit, a topic covered in detail in compacting, clearing, reading just what is needed. Partitioning, lastly, hands off a portion of the work, a long search, reading several files, to a sub agent that works in its own conversation and returns only its conclusion, which avoids loading the main conversation with the detail of its search.

One single move is enough to activate the first lever

These four mechanisms do not all switch on in every session. A short conversation, that stays well clear of the context limit, triggers no compaction, and a task that requires no extensive search delegates nothing to a sub agent. The cache, by contrast, switches on from the second message of the same conversation, with no particular setting to configure, on one condition: that the start of the message stays identical.

Message 1: project instructions (identical every time) + question A
Message 2: same project instructions + question B

The second message reuses the computation already done on the project instructions, provided they have not moved by a single character between the two messages.