src/compact/qa_summarizer.rs

Truncation’s information loss gets a real alternative here, in the QaSummarizer. It is a compactor that preserves the facts a future turn will need.

Each pass runs up to three non-streaming calls on a caller-supplied client, and a dedicated cheaper model is the documented production shape.

The first call summarizes the about-to-drop portion, and the second questions gaps by asking the model what it would need that the summary does not carry.

A third call answers from the full pre-compaction transcript before it is discarded.

Recovered facts are folded under ## Facts recovered, and N. UNKNOWN answers are dropped so nothing hallucinated enters your summary. This file is always compiled and has no feature flag.

Key items

QaSummarizer::new(client, config) takes a QaSummarizerConfig, and its budget is reserve-then-cap: min(max_summary_tokens, target × summary_budget_pct).

That fraction is clamped to [0.05, 0.95] and held as an exact per-mille integer, and transcript_max_chars bounds the input prompt for your small local models.

PriorSummary is the running summary, meaning text plus pass count, and it accumulates across your passes behind a recovered plain mutex. prior_summary() and reset() expose and clear it.

CompactionSummary is one pass’s parsed result.

Behavior notes

Any ApiError fails the pass through CompactionOutcome::failed with your original messages intact, which is the FallbackCompactor seam.

A blank response or an over-budget summary fails the pass the same way, strictly > at budget, and a sufficient-summary [] answer skips step 3.

Your pre-compact hook guidance reaches the prompts through instructions and additional_context. The previous pass’s summary message is filtered out of the transcript so accumulate mode never presents you the same summary twice, and a leading system-role message is pulled back unsummarized.

Since 0.4.1 the recognition is two-armed: exact text equality anywhere, or on the first assistant message alone a strip of exactly the leading summary part. The compaction seam now coalesces your summary into the tail’s first assistant message, so the recognition reads the coalesced head.

The split delegates to TokenSplitter’s budgeted variant, which is pair-safe and snaps to an assistant then user transition. Its durable pull is budget-bounded since 0.4.1, so Durable receipts ride verbatim only while your landing’s budget allows, newest first.

Evicted content rides CompactionOutcome::evicted for your demotion sink, and the summarizer’s own token spend reports per pass through loopctl.compaction.summarizer.tokens events.

For the deep dive, see Compaction and the compactor family.