Compression is a request-time view; the compressed form is text

The compress extension replaces finished turns that fall out of the keep window with one short standing-in message in outgoing LLM requests. The session file is never rewritten. The request list is re-derived every time in the context hook from the session entries plus the span cache (ADR 0011's projection pattern). Each compressed form is computed once, in the background after the turn ends, and persisted as a compress-span custom session entry. A span's identity is its list of session entry IDs: it is stable across requests, survives restarts, and a restored cache means a paid compression is never re-run for a span that already has a form.

The compressed form is plain text, not soft tokens or a provider-native context edit. The hosted chat APIs pi talks through (OpenAI Responses, Anthropic Messages) give no path to hand the model a reference to the original messages, so the standing-in message must be self-contained text any model can read. It is one synthetic user message carrying a fixed lossy-view frame plus the form, which keeps the prefix stable for the provider's prompt cache.

Rejected alternatives: a compaction entry cannot hold one form per finished turn - it replaces a whole span with one summary and rewrites what a later request sees from that point on. Rewriting the append-only session tree to shrink turns in place breaks fork identity and the Usage scan's once-per-distinct-line counting (ADR 0009). A sidecar store for the forms duplicates data the session file already can hold in a custom entry.

Cost accepted: a failed compression leaves the span raw, warns once, and retries on the next turn end; the outgoing request until then carries the full span. A span that fails three calls in a row stops earning calls and stays raw for the session, and a span whose serialized input does not fit the compression model's context window never earns a call, so a span that cannot compress cannot pay for it forever.