Skip to content
Sectr

Compaction

Every turn re-invokes your handler with the full transcript — that’s what makes runners stateless and sessions portable. Conversations that outgrow the model’s context need compaction, and sectr handles it without breaking either property.

The mechanism

Section titled “The mechanism”

Your code (or your framework) computes the compacted context — sectr journals it as a TranscriptCompacted event carrying the full post-compaction canonical messages. From that point the transcript is:

  • everything after the snapshot, plus the snapshot itself standing in for everything before it.

The snapshot is closed context by definition: ctx.messages positionally resets its fold at it — user turns derive from INVOCATION_STARTEDs after the snapshot, assistant messages from the post-snapshot MESSAGE_* events, and so on. Identity without snapshots holds (the fold needs no state), so crash recovery and approval resumes derive the same context deterministically from the journal.

Who computes it

Section titled “Who computes it”

Frameworks with built-in memory management (langgraph) compute compaction naturally; the adapter journals the result. Hand-rolled agents can emit TranscriptCompacted themselves:

yield TranscriptCompacted(messages=compact(ctx.messages))

After that point, ctx.messages in the next turn folds through the snapshot — the conversation continues with the compacted history.

Guidance

Section titled “Guidance”
  • Compact at a turn boundary when you can (end of a turn, before the next user message): the snapshot then covers a complete exchange and no in-flight streaming needs special handling.
  • The snapshot must be canonical messages — the same OpenAI-style shape ctx.messages produces — because the fold restarts from it.
  • Don’t try to journal a diff: the event carries the full context so that replay, resume, and rehydration need no history beyond it.